llama.cpp : bump version to 0.4.0 (#28386 )

sync : ggml (#28379 )
* ggml : rename and make private ggml_op_alloc_size_may_expand() (ggml/0) cont https://github.com/ggml-org/llama.cpp/pull/27960 * ggml : bump version to 0.23.0 (ggml/1618) * sync : ggml
2026-10-26 22:47:35 -05:00 · 2026-09-04 15:22:38 +03:00 · 2026-09-04 14:39:19 +03:00 · 2026-09-04 12:23:27 +02:00 · 2026-09-04 11:46:31 +02:00 · 2026-09-04 10:28:23 +02:00
227 changed files with 31427 additions and 1469 deletions
@@ -24,7 +24,7 @@ runs:

        write-host "Installing ROCm wheels for multi-arch support"
        # Install ROCm wheels for multi-arch support (this may take several minutes)
-        python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,devel]==${{ inputs.version }}"
+        python -m pip install --index-url https://stable.repo.amd.com/rocm/whl-next/ "rocm[libraries,devel]==${{ inputs.version }}"

        # Pre-expand the devel tree so it is included in the cache
        write-host "Initializing ROCm devel tree"
@@ -66,7 +66,13 @@ jobs:
            -DGGML_RPC=ON \
            -DCMAKE_OSX_DEPLOYMENT_TARGET=13.3
          time cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)
-          leaks -atExit -- ./build/bin/test-thread-safety -hf ggml-org/gemma-3-270m-qat-GGUF -ngl 99 -p "$(printf 'hello %.0s' {1..128})" -n 16 -c 512 -ub 32 -np 2 -t 2 -lv 1
+
+      - name: Check for leaks
+        run: |
+          cmd=(./build/bin/test-thread-safety -hf ggml-org/gemma-3-270m-qat-GGUF -ngl 99 -p "$(printf 'hello %.0s' {1..128})" -n 16 -c 512 -ub 32 -np 2 -t 2 -lv 1)
+          leaks -atExit -- "${cmd[@]}"
+          # Graphics devices are leaked by Metal in Apple code sometimes, so we ignore those leaks
+          OBJC_DEBUG_MISSING_POOLS=YES "${cmd[@]}" 2>&1 | awk '{ print } index($0, "autoreleased with no pool in place") && !/class [a-zA-Z0-9]+Device autoreleased/ { found = 1 } END { exit found }'

      - name: Test
        id: cmake_test
@@ -32,8 +32,8 @@ env:
  LLAMA_ARG_LOG_COLORS: 1
  LLAMA_ARG_LOG_PREFIX: 1
  LLAMA_ARG_LOG_TIMESTAMPS: 1
-  # TODO: fix and re-enable the `test-llama-archs` and `test-recurrent-state-rollback`
-  CTEST_EXCLUDE: "test-llama-archs|^test-recurrent-state-rollback"
+  # TODO: fix failing tests on OpenVINO backend
+  CTEST_EXCLUDE: "test-llama-archs|^test-recurrent-state-|test-backend-ops|test-save-load-state"

 jobs:
  ubuntu-24-openvino:
@@ -725,7 +725,7 @@ jobs:
    strategy:
      matrix:
        include:
-          - ROCM_VERSION: "7.14.0"
+          - ROCM_VERSION: "10.0.0"
            gpu_targets: "gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1103;gfx1150;gfx1151;gfx1152;gfx1153;gfx1200;gfx1201"
            build: x64

@@ -1279,7 +1279,7 @@ jobs:
    strategy:
      matrix:
        include:
-          - ROCM_VERSION: "7.14.0"
+          - ROCM_VERSION: "10.0.0"
            gpu_targets: "gfx908;gfx90a;gfx942;gfx950;gfx1010;gfx1011;gfx1012;gfx1030;gfx1031;gfx1032;gfx1033;gfx1034;gfx1035;gfx1036;gfx1100;gfx1101;gfx1102;gfx1150;gfx1151;gfx1152;gfx1200;gfx1201"
            build: 'x64'

@@ -1333,7 +1333,7 @@ jobs:
          # libraries = HIP runtime and CMake configs needed for linking
          # devel = compilers, headers, static libs
          python -m pip install --upgrade pip
-          python -m pip install --index-url https://repo.amd.com/rocm/whl-multi-arch/ "rocm[libraries,devel]==${{ matrix.ROCM_VERSION }}"
+          python -m pip install --index-url https://stable.repo.amd.com/rocm/whl-next/ "rocm[libraries,devel]==${{ matrix.ROCM_VERSION }}"

          # Get ROCm installation paths using the rocm-sdk CLI tool
          ROCM_PATH=$(rocm-sdk path --root)
@@ -1703,7 +1703,7 @@ jobs:
            - [Ubuntu s390x (CPU)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-s390x.tar.gz)
            - [Ubuntu x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-x64.tar.gz)
            - [Ubuntu arm64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-vulkan-arm64.tar.gz)
-            - [Ubuntu x64 (ROCm 7.14)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-7.14-x64.tar.gz)
+            - [Ubuntu x64 (ROCm 10.0)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-rocm-10.0-x64.tar.gz)
            - [Ubuntu x64 (OpenVINO)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-openvino-${{ needs.ubuntu-24-openvino.outputs.openvino_version }}-x64.tar.gz)
            - [Ubuntu x64 (SYCL FP32)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-fp32-x64.tar.gz)
            - [Ubuntu x64 (SYCL FP16)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-ubuntu-sycl-fp16-x64.tar.gz)
@@ -1721,7 +1721,7 @@ jobs:
            - [Windows x64 (Vulkan)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-vulkan-x64.zip)
            - [Windows x64 (OpenVINO)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-openvino-${{ needs.windows-openvino.outputs.openvino_version }}-x64.zip)
            - [Windows x64 (SYCL)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-sycl-x64.zip)
-            - [Windows x64 (ROCm 7.14)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-rocm-7.14-x64.zip)
+            - [Windows x64 (ROCm 10.0)](https://github.com/ggml-org/llama.cpp/releases/download/${{ steps.tag.outputs.name }}/llama-${{ steps.tag.outputs.name }}-bin-win-rocm-10.0-x64.zip)

            **openEuler:**
            - [DISABLED](https://github.com/ggml-org/llama.cpp/pull/23705)
@@ -103,7 +103,7 @@ jobs:
          source .venv/bin/activate
          cd tools/server/tests
          export ${{ matrix.extra_args }}
-          ./tests.sh
+          PYTEST_WORKERS=1 ./tests.sh

      - name: Slow tests
        id: server_integration_tests_slow
@@ -112,4 +112,4 @@ jobs:
          source .venv/bin/activate
          cd tools/server/tests
          export ${{ matrix.extra_args }}
-          SLOW_TESTS=1 ./tests.sh
+          PYTEST_WORKERS=1 SLOW_TESTS=1 ./tests.sh
@@ -102,7 +102,7 @@ jobs:
          ./tests.sh

  server-cuda:
-    runs-on: [self-hosted, llama-server, Linux, NVIDIA]
+    runs-on: "hf-jobs-t4-small:cuda13"

    steps:
      - name: Clone
@@ -112,12 +112,42 @@ jobs:
          fetch-depth: 0
          ref: ${{ github.event.inputs.sha || github.event.pull_request.head.sha || github.sha || github.head_ref || github.ref_name }}

+      - name: Install dependencies
+        run: |
+          sudo apt update
+          sudo apt install -y cmake libssl-dev python3 python3-venv python3-pip
+
+      - name: ccache
+        uses: ggml-org/ccache-action@v1.2.24
+        with:
+          restore: false
+          save: false
+
+      - name: ccache-buckets-restore
+        uses: ./.github/actions/ccache-buckets
+        with:
+          key: self-hosted-server-cuda
+          folder: llama.cpp
+          hf_bucket: ggml-org/cache
+
      - name: Build
        id: cmake_build
        run: |
-          cmake -B build -DGGML_CUDA=ON -DGGML_SCHED_NO_REALLOC=ON
+          cmake -B build -DGGML_CUDA=ON -DGGML_SCHED_NO_REALLOC=ON -DCMAKE_CUDA_COMPILER=/usr/local/cuda/bin/nvcc
          cmake --build build --config Release -j $(nproc) --target llama-server

+      - name: ccache-buckets-save
+        if: ${{ github.event_name == 'push' && github.ref == 'refs/heads/master' }}
+        uses: ./.github/actions/ccache-buckets
+        env:
+          HF_TOKEN: ${{ secrets.HF_TOKEN_CACHE_OUTPUT }}
+        with:
+          key: self-hosted-server-cuda
+          folder: llama.cpp
+          evict-old-files: 1d
+          hf_bucket: ggml-org/cache
+          save: true
+
      - name: Python setup
        id: setup_python
        run: |
@@ -1,4 +1,4 @@
-# date: Tue Aug 18 14:32:43 EEST 2026
+# date: Fri Sep  4 10:06:46 EEST 2026
 # this file is auto-generated by scripts/gen-authors.sh

 Нияз Гарифзянов <112617865+garrnizon@users.noreply.github.com>
@@ -46,6 +46,7 @@ Abhijit Ramesh <abhijitramesh2k@gmail.com>
 abhijitb11 <113058133+abhijitb11@users.noreply.github.com>
 Abhilash Majumder <30946547+abhilash1910@users.noreply.github.com>
 Abhinay Krishna <abhinaykrishna60@gmail.com>
+Abhiram <78226909+geckguy@users.noreply.github.com>
 Abhishek Gopinath K <31348521+overtunned@users.noreply.github.com>
 abotsis <github@bots.is>
 Abraham Gonzalez <theabecaster0@gmail.com>
@@ -87,6 +88,7 @@ akleine <alb.kleine@gmx.de>
 Al G <toasting@gmail.com>
 Al Mochkin <14274697+amochkin@users.noreply.github.com>
 Alan Gray <agray3@users.noreply.github.com>
+Alan Tseng <alanhc.tseng1999@gmail.com>
 Alawode Oluwandabira <dabiraalawode@yahoo.com>
 Albert Jin <albert.jin@gmail.com>
 Alberto <57916483+albbus-stack@users.noreply.github.com>
@@ -136,7 +138,9 @@ alonfaraj <alonfaraj@gmail.com>
 AlpinDale <52078762+AlpinDale@users.noreply.github.com>
 alwqx <kenan3015@gmail.com>
 Aman <amangupta052@gmail.com>
+Aman Chadha(IVIXMMI) <79802170+ac-mmi@users.noreply.github.com>
 Aman Gupta <amangupta052@gmail.com>
+Aman Karki <itsamankarki@gmail.com>
 amd-dwang <dong.wang@amd.com>
 amd-lalithnc <lalithnc@amd.com>
 Amir <amir_zia@outlook.com>
@@ -187,6 +191,7 @@ Anton Mitkov <anton.mitkov@codeplay.com>
 Antonis Makropoulos <benuix@gmail.com>
 Anudit Nagar <nagaranudit@gmail.com>
 Anuj Attri <anujattri01@gmail.com>
+anujj <ajalota@nvidia.com>
 anzz1 <anzz1@live.com>
 Aparna M P <aparmp@qti.qualcomm.com>
 Aparna M P <quic_aparmp@quicinc.com>
@@ -196,6 +201,7 @@ arch-btw <57669023+arch-btw@users.noreply.github.com>
 arcrank <arcrank@gmail.com>
 ardfork <134447697+ardfork@users.noreply.github.com>
 Arik Poznanski <arikpoz@users.noreply.github.com>
+Aritro Bandyopadhyay <71339004+AriBandyo@users.noreply.github.com>
 arlo-phoenix <140345165+arlo-phoenix@users.noreply.github.com>
 Armen Kaleshian <kriation@users.noreply.github.com>
 Arsen Arutunan <58118221+limloop@users.noreply.github.com>
@@ -230,6 +236,7 @@ bandoti <141645996+bandoti@users.noreply.github.com>
 Bar Haim <barvhaim@gmail.com>
 BarfingLemurs <128182951+BarfingLemurs@users.noreply.github.com>
 Bart Louwers <bart.louwers@gmail.com>
+Bartosz Taudul <wolf@nereid.pl>
 Bartowski <3266127+bartowski1182@users.noreply.github.com>
 Bartowski <ckealty1182@gmail.com>
 Bas Nijholt <basnijholt@gmail.com>
@@ -277,6 +284,7 @@ Bono Lv <lvscar@users.noreply.github.com>
 Borislav Stanimirov <b.stanimirov@abv.bg>
 Borislav Stanimirov <b@ibob.bg>
 Bowen Han <fancycode@gmail.com>
+Brad Smith <1472326+infinitewarp@users.noreply.github.com>
 Branden Butler <bwtbutler@hotmail.com>
 Brandon Squizzato <35474886+bsquizz@users.noreply.github.com>
 Brian <mofosyne@gmail.com>
@@ -287,6 +295,7 @@ Bryan Honof <bryanhonof@gmail.com>
 bryanSwk <93190252+bryanSwk@users.noreply.github.com>
 bsilvereagle <bsilvereagle@users.noreply.github.com>
 bssrdf <merlintiger@hotmail.com>
+Buğra Özgürsoy <13810383+ozgursoy@users.noreply.github.com>
 byte-6174 <88070277+byte-6174@users.noreply.github.com>
 Caleb DeLeeuw <143902425+SolshineCode@users.noreply.github.com>
 Calvin Laurenson <calvin@laurenson.dev>
@@ -326,6 +335,7 @@ Chenguang Li <757486878@qq.com>
 Chenguang Li <87689256+noemotiovon@users.noreply.github.com>
 Chipmunk <101038159+CHIPMUNK-T0T@users.noreply.github.com>
 chiranko <96988916+chiranko@users.noreply.github.com>
+Chris Danis <cdanis@gmail.com>
 Chris Elrod <elrodc@gmail.com>
 Chris Kuehl <ckuehl@ckuehl.me>
 Chris Lee <clee@mg8.org>
@@ -356,6 +366,7 @@ clyang <clyang@clyang.net>
 cmdr2 <secondary.cmdr2@gmail.com>
 cmdr2 <shashank.shekhar.global@gmail.com>
 cocktailpeanut <121128867+cocktailpeanut@users.noreply.github.com>
+codemonkey <441345965@qq.com>
 codezjx <code.zjx@gmail.com>
 coezbek <c.oezbek@gmail.com>
 comex <comexk@gmail.com>
@@ -367,6 +378,8 @@ Copilot <198982749+Copilot@users.noreply.github.com>
 Corentin REGAL <corentin.regal@gmail.com>
 cphlipot <9103367+cphlipot@users.noreply.github.com>
 cpumaxx <163466046+cpumaxx@users.noreply.github.com>
+cqderek <cqderek@gmail.com>
+cqderek <cqiang@qti.qualcomm.com>
 crasm <crasm@git.vczf.net>
 crasm <crasm@git.vczf.us>
 crat0z <11581854+crat0z@users.noreply.github.com>
@@ -427,6 +440,7 @@ DavidKorczynski <david@adalogics.com>
 davidrhodus <david@vacovideo.com>
 Dawid Potocki <github@dawidpotocki.com>
 Dawid Wysocki <62249621+TortillaZHawaii@users.noreply.github.com>
+Daya Adianto <addianto@users.noreply.github.com>
 ddh0 <chemist-mulches-39@icloud.com>
 ddh0 <dylanhalladay02@icloud.com>
 ddpasa <112642920+ddpasa@users.noreply.github.com>
@@ -463,6 +477,7 @@ Dmytro Romanov <casteldazur@gmail.com>
 Dobri Danchev <12420863+danchev@users.noreply.github.com>
 DocShotgun <126566557+DocShotgun@users.noreply.github.com>
 Doctor Shotgun <126566557+DocShotgun@users.noreply.github.com>
+Dominik Pantaleoni <95251853+dpantaleoni@users.noreply.github.com>
 Don Mahurin <dmahurin@users.noreply.github.com>
 Dong Won Kim <63934649+ddwkim@users.noreply.github.com>
 Donghyeon Jeong <54725479+djeong20@users.noreply.github.com>
@@ -504,6 +519,7 @@ Emmanuel Ferdman <emmanuelferdman@gmail.com>
 Emreerdog <34742675+Emreerdog@users.noreply.github.com>
 Engininja2 <139037756+Engininja2@users.noreply.github.com>
 Equim <sayaka@ekyu.moe>
+Eric A Stalee <87948564+Eric-A-Stalee@users.noreply.github.com>
 Eric Curtin <ecurtin@redhat.com>
 Eric Curtin <eric.curtin@docker.com>
 Eric Curtin <ericcurtin17@gmail.com>
@@ -519,6 +535,7 @@ Esko Toivonen <eskot98@gmail.com>
 Ethan Turner <eturner64@gmail.com>
 Ettore Di Giacinto <mudler@users.noreply.github.com>
 EugeoSynthesisThirtyTwo <gabriel.dhimoila@gmail.com>
+Eurekatic <eurekatic@eurekatic.eu>
 Evan Huus <eapache@gmail.com>
 Evan Jones <evan.q.jones@gmail.com>
 Evan Miller <emmiller@gmail.com>
@@ -677,6 +694,7 @@ HimariO <dsfhe49854@gmail.com>
 hipudding <huafengchun@gmail.com>
 Hitesh Chopra <34310832+hiteshchopra11@users.noreply.github.com>
 hksdpc255 <43977088+hksdpc255@users.noreply.github.com>
+hmirin <hmirin@users.noreply.github.com>
 hmscider <201289679+hmscider@users.noreply.github.com>
 Hoang Nguyen <hugo53@users.noreply.github.com>
 hoangmit <hoangmit@users.noreply.github.com>
@@ -701,6 +719,7 @@ Huawei Lin <huaweilin.cs@gmail.com>
 Hugo <hugo@whynothugo.nl>
 Hugo Roussel <hugo.rous@gmail.com>
 Huifeng Ou <79071290+ho2103@users.noreply.github.com>
+HumerousGorgon <31957201+HumerousGorgon@users.noreply.github.com>
 hutli <6594598+hutli@users.noreply.github.com>
 hutli <hutli@hutli.hu>
 hutli <jensstaermose@hotmail.com>
@@ -738,12 +757,15 @@ intelmatt <61025942+intelmatt@users.noreply.github.com>
 iohub <rickyang.pro@gmail.com>
 Ionoclast Laboratories <brigham@ionoclast.com>
 iron <lizhenneng@gmail.com>
+Isaac <34376531+init-22@users.noreply.github.com>
 Isaac McFadyen <isaac@imcf.me>
 IsaacDynamo <61521674+IsaacDynamo@users.noreply.github.com>
 Ishaan Gandhi <Ishaangandhi@gmail.com>
 iSma <ismail.senhaji@gmail.com>
 Ismail <115064057+AlrIsmail@users.noreply.github.com>
 issixx <46835150+issixx@users.noreply.github.com>
+itsnotoger <19309683+itsnotoger@users.noreply.github.com>
+itterative <190138728+itterative@users.noreply.github.com>
 Ivan <nekotekina@gmail.com>
 Ivan Chikish <nekotekina@gmail.com>
 Ivan Filipov <159561759+vanaka11@users.noreply.github.com>
@@ -768,6 +790,7 @@ Jakkala Mahesh <155058658+MaheshJakkala@users.noreply.github.com>
 Jakub N <jakubniemczyk97@gmail.com>
 JamePeng <jame_peng@sina.com>
 James A Capozzoli <157492257+jac-jim@users.noreply.github.com>
+James Francis <6763899+JamesFranc@users.noreply.github.com>
 James O'Leary <65884233+jpohhhh@users.noreply.github.com>
 James Reynolds <magnusviri@users.noreply.github.com>
 jameswu2014 <545426914@qq.com>
@@ -798,6 +821,7 @@ Jed Fox <git@jedfox.com>
 Jeff Bolz <jbolz@nvidia.com>
 Jeffrey Morgan <jmorganca@gmail.com>
 Jeffrey Quesnelle <emozilla@nousresearch.com>
+Jeremie Miller <jeremie.miller@gmail.com>
 Jeremy Demeule <jdemeule@users.noreply.github.com>
 Jeremy Rand <244188+JeremyRand@users.noreply.github.com>
 Jeroen Mostert <jeroen.mostert@cm.com>
@@ -809,6 +833,7 @@ Jesse Jojo Johnson <williamsaintgeorge@gmail.com>
 Jesse LaRose <jesse@taey.ai>
 Jesse Posner <jesse.posner@gmail.com>
 Jesus Talavera <145992175+jesus-talavera-ibm@users.noreply.github.com>
+Jetson Tan <tanzongyouyi@outlook.com>
 Jett Janiak <jettjaniak@gmail.com>
 Jeximo <jeximo@gmail.com>
 JFLFY2255 <JFLFY2255@163.com>
@@ -825,6 +850,7 @@ Jie Fu (傅杰) <jiefu@tencent.com>
 jiez <373447296@qq.com>
 Jillis ter Hove <j.terhove@gmail.com>
 Jim Wu <jimw567@users.noreply.github.com>
+Jingxin (Philip) Li <philipaslee@gmail.com>
 Jinwoo Jeong <33892306+williamjeong2@users.noreply.github.com>
 Jinyang He <hejinyang@loongson.cn>
 jinzihao <jinzihao1996@gmail.com>
@@ -850,11 +876,13 @@ John Balis <phobossystems@gmail.com>
 John Bean <113509988+johnbean393@users.noreply.github.com>
 John Eismeier <42679190+jeis4wpi@users.noreply.github.com>
 John Smith <67539080+kingsidelee@users.noreply.github.com>
+John-Henry Lim <42513874+Interpause@users.noreply.github.com>
 Johnathan Craig Maudlin <13183098+jcmdln@users.noreply.github.com>
 JohnnyB <jboero@users.noreply.github.com>
 johnson442 <56517414+johnson442@users.noreply.github.com>
 jojorne <jojorne@users.noreply.github.com>
 jon-chuang <9093549+jon-chuang@users.noreply.github.com>
+Jonas J <111707981+John-194@users.noreply.github.com>
 Jonas Jankaitis <111707981+John-194@users.noreply.github.com>
 Jonas Wunderlich <32615971+jonas-w@users.noreply.github.com>
 Jonathan <47618606+jbuchananr@users.noreply.github.com>
@@ -924,6 +952,7 @@ Karsten Weiss <knweiss@gmail.com>
 Karthick <j.karthic2004@gmail.com>
 Karthik Kumar Viswanathan <195178+guilt@users.noreply.github.com>
 Karthik Sethuraman <k.seth1993@gmail.com>
+Kartik Gulia <kgulia@nvidia.com>
 Kartik Sirohi <99896785+sirohikartik@users.noreply.github.com>
 Kashif Rasul <kashif.rasul@gmail.com>
 KASR <karim.asrih@gmail.com>
@@ -931,6 +960,7 @@ Kasumi <90275229+kasumi-1@users.noreply.github.com>
 Katostrofik <georgiopapairo@gmail.com>
 katsu560 <118887472+katsu560@users.noreply.github.com>
 Kawrakow <48489457+ikawrakow@users.noreply.github.com>
+kbenkhaled <khalilbenkhaled01@gmail.com>
 kchro3 <62481661+kchro3@users.noreply.github.com>
 kdkd <2569413+kdkd@users.noreply.github.com>
 Keiichi Tabata <keiichi.tabata@outlook.com>
@@ -939,6 +969,7 @@ Kenvix ⭐ <kenvixzure@live.com>
 Kerfuffle <44031344+KerfuffleV2@users.noreply.github.com>
 Kevin Gibbons <bakkot@gmail.com>
 Kevin Hannon <kehannon@redhat.com>
+Kevin Hopper <93635715+kh0pper@users.noreply.github.com>
 Kevin Ji <1146876+kevinji@users.noreply.github.com>
 Kevin Kwok <antimatter15@gmail.com>
 Kevin Liu <4396kevinliu@gmail.com>
@@ -964,12 +995,14 @@ Konstantin Herud <konstantin.herud@denkbares.com>
 Konstantin Zhuravlyov <konstantin.zhuravlyov@amd.com>
 Krishna Sridhar <99914379+srikris-sridhar@users.noreply.github.com>
 krystiancha <krystian@krystianch.com>
+krzsztf <krzysztof@witkowscy.org>
 kubawoo <k-wach@o2.pl>
 kumaal <44551860+kumaal@users.noreply.github.com>
 kunal-vaishnavi <115581922+kunal-vaishnavi@users.noreply.github.com>
 kunnis <kunnis@users.noreply.github.com>
 Kunshang Ji <kunshang.ji@intel.com>
 kuronekosaiko <EvanChanJ@163.com>
+kurquhar <kurquhar@qti.qualcomm.com>
 Kusha Gharahi <3326002+kushagharahi@users.noreply.github.com>
 kustaaya <58045274+kustaaya@users.noreply.github.com>
 kuvaus <22169537+kuvaus@users.noreply.github.com>
@@ -981,6 +1014,7 @@ Kyle Liang <liangmanlai@gmail.com>
 Kyle Mistele <kyle@mistele.com>
 KyleHagy <59183061+KyleHagy@users.noreply.github.com>
 Kylin <56434533+KyL0N@users.noreply.github.com>
+Kyozzz <1147385157@qq.com>
 l-austenfeld <53152202+l-austenfeld@users.noreply.github.com>
 l3utterfly <gc.pthzfoldr@gmail.com>
 l8bloom <l8bloomapi@gmail.com>
@@ -992,6 +1026,7 @@ Lars Sonchocky-Helldorf <lars.sonchocky-helldorf@hamburg.de>
 las7 <98077186+las7@users.noreply.github.com>
 Lasse Lauwerys <65569591+Iemand005@users.noreply.github.com>
 Laura <Tijntje_7@msn.com>
+Laurent Zuijdwijk <laurent.zuijdwijk@gmail.com>
 Law Po Ying <30721578+yingying0906@users.noreply.github.com>
 lcy <lcy0321@users.noreply.github.com>
 ldwang <ftgreat@163.com>
@@ -1039,6 +1074,8 @@ Ludovic Henry <git@ludovic.dev>
 Ludovic Henry <ludovic@rivosinc.com>
 Lukas Straub <lukasstraub2@web.de>
 Łukasz Ślusarczyk <112692748+lslusarczyk@users.noreply.github.com>
+Lukasz Stolcman <4583553+lstolcman@users.noreply.github.com>
+LunalFresh <165352784+LunalFresh@users.noreply.github.com>
 Luo Tian <lt@basecity.com>
 luoyu-intel <yu.luo@intel.com>
 luyhcsu <110711054+luyhcsu@users.noreply.github.com>
@@ -1054,6 +1091,7 @@ Maarten ter Huurne <maarten@treewalker.org>
 Maciej Lisowski <39798354+MaciejDromin@users.noreply.github.com>
 Mack Straight <eiz@users.noreply.github.com>
 maddes8cht <55592906+maddes8cht@users.noreply.github.com>
+Mads Marquart <mads@marquart.dk>
 Maël Kerbiriou <m431.kerbiriou@gmail.com>
 MaggotHATE <clay1326@gmail.com>
 MagicExists <106458387+gugugiyu@users.noreply.github.com>
@@ -1215,6 +1253,8 @@ Naco Siren <naco-siren@users.noreply.github.com>
 Nam D. Tran <42194884+namtranase@users.noreply.github.com>
 nanahi <130121847+na-na-hi@users.noreply.github.com>
 Nathan Epstein <nate2@umbc.edu>
+Nathan Wilson <67372905+Nathanw1014@users.noreply.github.com>
+Nathanw1014 <67372905+Nathanw1014@users.noreply.github.com>
 Natsu <chino@hotococoa.moe>
 Nauful Shaikh <nauful@gmail.com>
 NawafAlansari <72708095+NawafAlansari@users.noreply.github.com>
@@ -1237,6 +1277,7 @@ niansa/tuxifan <tuxifan@posteo.de>
 Nicholai Tukanov <nicholaitukanov@gmail.com>
 Nicholas Sparks <157740354+nisparks@users.noreply.github.com>
 Nick <0x0b4ac@gmail.com>
+Nick Farrell <nick.farrell@aiven.io>
 nick huang <nickhuang99@hotmail.com>
 Nick Lafleur <55208706+nicklafleur@users.noreply.github.com>
 Nick Towle <ntowle@gmail.com>
@@ -1259,6 +1300,7 @@ NikolaiLyssogor <59844691+NikolaiLyssogor@users.noreply.github.com>
 Nikolaos Pothitos <pothitos@di.uoa.gr>
 Nikolas <127742645+nneubacher@users.noreply.github.com>
 Nikolay Popov <131475237+npopov-vst@users.noreply.github.com>
+Nils Gladitz <nilsgladitz@gmail.com>
 Nindaleth <Nindaleth@users.noreply.github.com>
 ningshanwutuobang <ningshanwutuobang@gmail.com>
 Noah <99681487+NoahOksuz@users.noreply.github.com>
@@ -1355,6 +1397,7 @@ Pop Flamingo <trevor.annedenise@icloud.com>
 postmasters <namnguyen@google.com>
 Pouya <PooyaGhahramanian@Gmail.com>
 pqnet <119850+pqnet@users.noreply.github.com>
+Prabhsimran Singh <pskrunner14@gmail.com>
 Prabod <prabod@maincode.com>
 Prajwal B Mehendarkar <prajwal.b.mehendarkar@ibm.com>
 Pranav Dhinakar <pdhinaka@qti.qualcomm.com>
@@ -1378,6 +1421,7 @@ qouoq <qouoq@fastmail.com>
 Qu Zongfu <43257352+yancaoweidaode@users.noreply.github.com>
 quei <56998528+quei4r@users.noreply.github.com>
 Quentin Bramas <quentin.bramas@gmail.com>
+QuintinShaw <github@xyt.email>
 QuintinShaw <yx6f20@soton.ac.uk>
 qunash <anzoria@gmail.com>
 quyentonndbs <raynaedgar8677@outlook.com>
@@ -1462,6 +1506,7 @@ robertomeroni <150194833+robertomeroni@users.noreply.github.com>
 Robey Holderith <robey@flaminglunchbox.net>
 Robin Davidsson <40024429+R-Dson@users.noreply.github.com>
 Robyn <robyngraf@users.noreply.github.com>
+Rock Chen <rockchen.tw@gmail.com>
 Rőczey Barnabás <31726601+An0nie@users.noreply.github.com>
 RodriMora <bullerwins@gmail.com>
 Roger Chen <chenrui@gmail.com>
@@ -1499,17 +1544,21 @@ runfuture <runfuture@users.noreply.github.com>
 RunningLeon <maningsheng@sensetime.com>
 RunningLeon <mnsheng@yeah.net>
 Russyyds <161207317+Russyyds@users.noreply.github.com>
+Ryan C <ryan5rdx@users.noreply.github.com>
 Ryan Goulden <percontation@gmail.com>
 Ryan Landay <rlanday@gmail.com>
 Ryan Mangeno <160974989+ryan-mangeno@users.noreply.github.com>
 Ryder Wishart <ryderwishart@gmail.com>
 Ryuei <louixs@users.noreply.github.com>
 s-goto-11 <206795233+s-goto-11@users.noreply.github.com>
+s0mecode <213953308+s0mecode@users.noreply.github.com>
 s8322 <s0527684199@gmail.com>
+Saad Ali <NIXKnight@users.noreply.github.com>
 Saba Fallah <10401143+sfallah@users.noreply.github.com>
 Saba Fallah <sabafallah@gmail.com>
 Sachin Desai <smdesai@gmail.com>
 Sachin Sharma <sachin@zettabolt.com>
+Safi Ullah <safiullah.3915@gmail.com>
 safranowith <bsh155762@gmail.com>
 SakuraUmi <yukinon244@gmail.com>
 Salvador E. Tropea <stropea@inti.gob.ar>
@@ -1552,6 +1601,7 @@ Sergey Alirzaev <l29ah@riseup.net>
 Sergey Alirzaev <zl29ah@gmail.com>
 Sergey Fedorov <vital.had@gmail.com>
 Sergey Malinin <sergmalinin@gmail.com>
+Sergey Sklyarov <sergey.sklyarov@gmail.com>
 Sergio López <slp@redhat.com>
 Sergio López <slp@sinrega.org>
 Sergiu <8598216+mzsergiu@users.noreply.github.com>
@@ -1582,11 +1632,13 @@ Shawn Gu <shawngu@qti.qualcomm.com>
 Shawn yang <137684499+Yangxiaoz@users.noreply.github.com>
 Shelby Jenkins <47464908+ShelbyJenkins@users.noreply.github.com>
 Sheldon Robinson <sheldon.robinson@live.com>
+Shenghan Yang <ysharke@sjtu.edu.cn>
 shibe2 <shibe@tuta.io>
 Shijie <821898965@qq.com>
 Shin-myoung-serp <relent95@naver.com>
 Shintarou Okada <kokuzen@gmail.com>
 shivamkumard-ctrl <shivamkumard@nvidia.com>
+Shobhit <sobhit.me@gmail.com>
 Shouyu <65317431+joeldushouyu@users.noreply.github.com>
 Shouzheng Liu <61452103+lshzh-ww@users.noreply.github.com>
 Shouzheng Liu <lshzh.hi@gmail.com>
@@ -1607,6 +1659,7 @@ Sigbjørn Skjæret <sigbjorn.skjaeret@huggingface.co>
 Sigbjørn Skjæret <sigbjorn.skjaeret@scala.com>
 simevo <github@simevo.com>
 Simon Redman <simon@ergotech.com>
+Simon Teixidor <simon@flaskpost.me>
 Simon Willison <swillison@gmail.com>
 simon886212 <37953122+simon886212@users.noreply.github.com>
 Simranjeet Singh <105192966+simrnsingh@users.noreply.github.com>
@@ -1663,6 +1716,7 @@ stevenkuang <stevenkuang@tencent.com>
 Steward Garcia <57494570+FSSRepo@users.noreply.github.com>
 StrangeBytesDev <141275258+StrangeBytesDev@users.noreply.github.com>
 strawberrymelonpanda <152940198+strawberrymelonpanda@users.noreply.github.com>
+Strongtut <Strongtut@users.noreply.github.com>
 Suaj Carrot <72162667+SuajCarrot@users.noreply.github.com>
 sudhiarm <sudhi.sathyavathy@arm.com>
 Sukriti Sharma <Ssukriti@users.noreply.github.com>
@@ -1687,6 +1741,7 @@ Tamar <Tamar0812@outlook.co.il>
 tamarPal <tamarp3385@gmail.com>
 Tameem <113388789+AhmadTameem@users.noreply.github.com>
 Tamotsu Takahashi <ttakah+github@gmail.com>
+Tanner Bruhn <66120666+tannerbruhn@users.noreply.github.com>
 tarcey <cey.tarik@gmail.com>
 Tarek Dakhran <t.dakhran@gmail.com>
 Tarek Dakhran <tarek@liquid.ai>
@@ -1696,6 +1751,7 @@ Taylor <quantumtraveling@gmail.com>
 tc-mb <157115220+tc-mb@users.noreply.github.com>
 TecJesh <qdvm5gl@163.com>
 Tei Home <taiteitonghome@proton.me>
+Tekin Ertekin <tekin.ertekin@gmail.com>
 Tekin Ertekin <tekinertekin@gmail.com>
 tempstudio <49735574+tempstudio@users.noreply.github.com>
 teo <TeoZosa@users.noreply.github.com>
@@ -1737,6 +1793,7 @@ Ting Lou <louting@189.cn>
 Ting Lou <ting.lou@gmail.com>
 Ting Sun <suntcrick@gmail.com>
 Titaniumtown <titaniumtown@proton.me>
+Tiwei Bie <tiwei.btw@antgroup.com>
 tjohnman <tjohnman@users.noreply.github.com>
 Tobias Lütke <tobi@shopify.com>
 Toby <25832191+aetherbird@users.noreply.github.com>
@@ -1813,6 +1870,7 @@ Vishal Agarwal <vishalagarwal.jss@gmail.com>
 Vishal Singh <vishal@zettabolt.com>
 Vitali Lovich <vlovich+github@gmail.com>
 Vivian <vynride@gmail.com>
+vk <89937361+itsvedantkumar@users.noreply.github.com>
 Vlad <spitfireage@gmail.com>
 Vladimir <bogdad@gmail.com>
 Vladimir Malyutin <first-leon@yandex.ru>
@@ -1897,6 +1955,7 @@ Yaiko <elyaiko@hotmail.com>
 Yakine Tahtah <96926916+ReinforcedKnowledge@users.noreply.github.com>
 YangLe <smilingpoplar@gmail.com>
 yangli2 <yangli2@gmail.com>
+Yaniss Amazouz <yaniss91600@gmail.com>
 Yann Follet <131855179+YannFollet@users.noreply.github.com>
 Yanzhao Wang <yanzhaow@qti.qualcomm.com>
 Yarden Tal <yardent@qti.qualcomm.com>
@@ -4,7 +4,7 @@ include(CheckIncludeFileCXX)

 ### llama.cpp version
 set(LLAMA_VERSION_MAJOR 0)
-set(LLAMA_VERSION_MINOR 3)
+set(LLAMA_VERSION_MINOR 4)
 set(LLAMA_VERSION_PATCH 0)
 set(LLAMA_VERSION_BASE "${LLAMA_VERSION_MAJOR}.${LLAMA_VERSION_MINOR}.${LLAMA_VERSION_PATCH}")

@@ -13,7 +13,7 @@
 [![Docker](https://img.shields.io/github/actions/workflow/status/ggml-org/llama.cpp/docker.yml?label=Docker)](https://github.com/ggml-org/llama.cpp/actions/workflows/docker.yml)
 [![Winget](https://img.shields.io/github/actions/workflow/status/ggml-org/llama.cpp/winget.yml?label=Winget)](https://github.com/ggml-org/llama.cpp/actions/workflows/winget.yml)

-[ggml](https://github.com/ggml-org/ggml) / [ops](https://github.com/ggml-org/llama.cpp/blob/master/docs/ops.md) / [maintainer PRs](https://github.com/ggml-org/llama.cpp/issues?q=is%3Apr%20is%3Aopen%20draft%3AFalse%20(author%3Argerganov%20OR%20author%3AKitaitiMakoto%20OR%20author%3Adanbev%20OR%20author%3Aaldehir%20OR%20author%3Amax-krasnyansky%20OR%20author%3ACISC%20OR%20author%3Aggerganov%20OR%20author%3Aam17an%20OR%20author%3Abartowski1182%20OR%20author%3Anikwen%20OR%20author%3Ahipudding%20OR%20author%3AServeurpersoCom%20OR%20author%3Apwilkin%20OR%20author%3Areeselevine%20OR%20author%3Angxson%20OR%20author%3Ajeffbolznv%20OR%20author%3Amarty1885%20OR%20author%3A0cc4m%20OR%20author%3ATitaniumtown%20OR%20author%3Aangt%20OR%20author%3AIMbackK%20OR%20author%3Aarthw%20OR%20author%3AJohannesGaessler%20OR%20author%3AORippler%20OR%20author%3Aruixiang63%20OR%20author%3Axctan%20OR%20author%3Aallozaur%20OR%20author%3Ayomaytk%20OR%20author%3Aaendk%20OR%20author%3Agaugarg-nv%20OR%20author%3Ataronaeo%20OR%20author%3Aforforever73%20OR%20author%3Alhez%20OR%20author%3Anetrunnereve%20OR%20author%3Afairydreaming)%20sort%3Aupdated-desc) / [dev stats](https://github.com/ggml-org/llama.cpp-dev) / [lib llama API](https://github.com/ggml-org/llama.cpp/issues/9289) / [llama-server REST API](https://github.com/ggml-org/llama.cpp/issues/9291)
+[ggml](https://github.com/ggml-org/ggml) / [ops](https://github.com/ggml-org/llama.cpp/blob/master/docs/ops.md) / [maintainer PRs](https://github.com/ggml-org/llama.cpp/issues?q=is%3Apr%20is%3Aopen%20draft%3AFalse%20(author%3Argerganov%20OR%20author%3AKitaitiMakoto%20OR%20author%3Adanbev%20OR%20author%3Aaldehir%20OR%20author%3Amax-krasnyansky%20OR%20author%3ACISC%20OR%20author%3Aggerganov%20OR%20author%3Aam17an%20OR%20author%3Ajhen0409%20OR%20author%3Abartowski1182%20OR%20author%3Anikwen%20OR%20author%3Ahipudding%20OR%20author%3Aravi9%20OR%20author%3AServeurpersoCom%20OR%20author%3Apwilkin%20OR%20author%3Areeselevine%20OR%20author%3Angxson%20OR%20author%3Ajeffbolznv%20OR%20author%3Amarty1885%20OR%20author%3A0cc4m%20OR%20author%3ATitaniumtown%20OR%20author%3Aangt%20OR%20author%3AIMbackK%20OR%20author%3Aarthw%20OR%20author%3AJohannesGaessler%20OR%20author%3AORippler%20OR%20author%3Aruixiang63%20OR%20author%3Axctan%20OR%20author%3Aallozaur%20OR%20author%3Ayomaytk%20OR%20author%3Aaendk%20OR%20author%3Awine99%20OR%20author%3Agaugarg-nv%20OR%20author%3Ataronaeo%20OR%20author%3Aforforever73%20OR%20author%3Alhez%20OR%20author%3Anetrunnereve%20OR%20author%3Afairydreaming)%20sort%3Aupdated-desc) / [dev stats](https://github.com/ggml-org/llama.cpp-dev) / [lib llama API](https://github.com/ggml-org/llama.cpp/issues/9289) / [llama-server REST API](https://github.com/ggml-org/llama.cpp/issues/9291)

 </div>

@@ -74,7 +74,7 @@ The `llama.cpp` project is build on top of the [ggml](https://github.com/ggml-or
 | [CANN](docs/build.md#cann) | Ascend NPU |
 | [CUDA](docs/build.md#cuda) | Nvidia GPU |
 | [HIP](docs/build.md#hip) | AMD GPU |
-| [Hexagon [In Progress]](docs/backend/snapdragon/README.md) | Snapdragon |
+| [Hexagon](docs/backend/snapdragon/README.md) | Snapdragon |
 | [IBM zDNN](docs/backend/zDNN.md) | IBM Z & LinuxONE |
 | [MUSA](docs/build.md#musa) | Moore Threads GPU |
 | [Metal](docs/build.md#metal-build) | Apple Silicon |
@@ -80,7 +80,7 @@ static const command cmds[] = {
 #undef UPDATE_HIDDEN

 static int version(int /*argc*/, char ** /*argv*/) {
-    llama_print_build_info(llama_version());
+    llama_print_build_info(llama_version(), stdout);
    return 0;
 }

@@ -18,7 +18,7 @@ LLAMA_BUILD_TESTS=OFF
 LLAMA_BUILD_SERVER=OFF
 LLAMA_BUILD_MTMD=ON
 GGML_METAL=ON
-GGML_METAL_EMBED_LIBRARY=ON
+GGML_METAL_EMBED_LIBRARY=${GGML_METAL_EMBED_LIBRARY:-ON}
 GGML_BLAS_DEFAULT=ON
 GGML_OPENMP=OFF

@@ -169,6 +169,14 @@ setup_framework_structure() {
    cp tools/mtmd/mtmd.h           ${header_path}
    cp tools/mtmd/mtmd-helper.h    ${header_path}

+    if [[ "$GGML_METAL_EMBED_LIBRARY" == "OFF" ]]; then
+        if [[ "$platform" == "macos" ]]; then
+            cp ${build_dir}/bin/*.metallib ${build_dir}/framework/${framework_name}.framework/Versions/A/Resources/
+        else
+            cp ${build_dir}/bin/*.metallib ${build_dir}/framework/${framework_name}.framework/
+        fi
+    fi
+
    # Create module map (common for all platforms)
    cat > ${module_path}module.modulemap << EOF
 framework module llama {
@@ -450,6 +458,7 @@ build_ios_sim() {
        -DIOS=ON \
        -DCMAKE_SYSTEM_NAME=iOS \
        -DCMAKE_OSX_SYSROOT=iphonesimulator \
+        -DGGML_METAL_TARGET_OS=ios \
        -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
        -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphonesimulator \
        -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
@@ -467,6 +476,7 @@ build_ios_device() {
        -DCMAKE_OSX_DEPLOYMENT_TARGET=${IOS_MIN_OS_VERSION} \
        -DCMAKE_SYSTEM_NAME=iOS \
        -DCMAKE_OSX_SYSROOT=iphoneos \
+        -DGGML_METAL_TARGET_OS=ios \
        -DCMAKE_OSX_ARCHITECTURES="arm64" \
        -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=iphoneos \
        -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
@@ -498,6 +508,7 @@ build_visionos() {
        -DCMAKE_OSX_ARCHITECTURES="arm64" \
        -DCMAKE_SYSTEM_NAME=visionOS \
        -DCMAKE_OSX_SYSROOT=xros \
+        -DGGML_METAL_TARGET_OS=xros \
        -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xros \
        -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
        -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
@@ -516,6 +527,7 @@ build_visionos_sim() {
        -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
        -DCMAKE_SYSTEM_NAME=visionOS \
        -DCMAKE_OSX_SYSROOT=xrsimulator \
+        -DGGML_METAL_TARGET_OS=xros \
        -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=xrsimulator \
        -DCMAKE_C_FLAGS="${COMMON_C_FLAGS}" \
        -DCMAKE_CXX_FLAGS="${COMMON_CXX_FLAGS}" \
@@ -534,6 +546,7 @@ build_tvos_sim() {
        -DCMAKE_OSX_DEPLOYMENT_TARGET=${TVOS_MIN_OS_VERSION} \
        -DCMAKE_SYSTEM_NAME=tvOS \
        -DCMAKE_OSX_SYSROOT=appletvsimulator \
+        -DGGML_METAL_TARGET_OS=tvos \
        -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64" \
        -DGGML_METAL=ON \
        -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvsimulator \
@@ -552,6 +565,7 @@ build_tvos_device() {
        -DCMAKE_OSX_DEPLOYMENT_TARGET=${TVOS_MIN_OS_VERSION} \
        -DCMAKE_SYSTEM_NAME=tvOS \
        -DCMAKE_OSX_SYSROOT=appletvos \
+        -DGGML_METAL_TARGET_OS=tvos \
        -DCMAKE_OSX_ARCHITECTURES="arm64" \
        -DGGML_METAL=ON \
        -DCMAKE_XCODE_ATTRIBUTE_SUPPORTED_PLATFORMS=appletvos \
@@ -189,8 +189,8 @@ if [ ! -z ${GG_BUILD_OPENVINO} ]; then
    fi
    CMAKE_EXTRA="${CMAKE_EXTRA} -DGGML_OPENVINO=ON"

-    # TODO: fix and re-enable the `test-llama-archs` and `test-recurrent-state-rollback*`
-    CTEST_EXTRA="-E test-llama-archs|^test-recurrent-state-rollback"
+    # TODO: fix failing tests on OpenVINO backend
+    CTEST_EXTRA="-E test-llama-archs|^test-recurrent-state-|test-backend-ops|test-save-load-state"
 fi

 ## helpers
@@ -960,6 +960,11 @@ static bool common_params_parse_ex(int argc, char ** argv, common_params_context
        ));
    }

+    // if the preserve_reasoning kwarg was not specified explicitly, enable it by default
+    if (!params.default_template_kwargs.count("preserve_reasoning")) {
+        params.default_template_kwargs["preserve_reasoning"] = "true";
+    }
+
    return true;
 }

@@ -3553,6 +3558,10 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
                    LOG_WRN("Setting 'enable_thinking' via --chat-template-kwargs is deprecated. "
                            "Use --reasoning on / --reasoning off instead.\n");
                }
+                if (item.key() == "preserve_reasoning") {
+                    LOG_WRN("Setting 'preserve_reasoning' via --chat-template-kwargs is deprecated. "
+                            "Use --reasoning-preserve / --no-reasoning-preserve instead.\n");
+                }
                params.default_template_kwargs[item.key()] = item.value().dump();
            }
        }
@@ -3743,7 +3752,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
    add_opt(common_arg(
        {"--reasoning-preserve"},
        {"--no-reasoning-preserve"},
-        "preserve reasoning trace in the full history, not just the last assistant message (default: template default)\n"
+        "preserve reasoning trace in the full history, not just the last assistant message (default: enabled)\n"
        "compatible with certain templates having 'supports_preserve_reasoning' capability\n"
        "example: https://docs.z.ai/guides/capabilities/thinking-mode#preserved-thinking",
        [](common_params & params, bool value) {
@@ -3752,6 +3761,7 @@ common_params_context common_params_parser_init(common_params & params, llama_ex
            } else {
                params.default_template_kwargs["preserve_reasoning"] = "false";
            }
+            params.preserve_reasoning_specified = true;
        }
    ).set_examples({LLAMA_EXAMPLE_SERVER, LLAMA_EXAMPLE_COMPLETION, LLAMA_EXAMPLE_CLI}).set_env("LLAMA_ARG_REASONING_PRESERVE"));
    add_opt(common_arg(
@@ -29,7 +29,7 @@ const char * llama_build_info(void) {
    return s.c_str();
 }

-void llama_print_build_info(const char * llama_version) {
-    fprintf(stderr, "version: %s (build %d, commit %s)\n", llama_version, llama_build_number(), llama_commit());
-    fprintf(stderr, "built with %s for %s\n", llama_compiler(), llama_build_target());
+void llama_print_build_info(const char * llama_version, FILE * stream) {
+    fprintf(stream, "version: %s (build %d, commit %s)\n", llama_version, llama_build_number(), llama_commit());
+    fprintf(stream, "built with %s for %s\n", llama_compiler(), llama_build_target());
 }
@@ -1,5 +1,7 @@
 #pragma once

+#include <cstdio>
+
 int llama_build_number(void);

 const char * llama_commit(void);
@@ -8,4 +10,4 @@ const char * llama_compiler(void);
 const char * llama_build_target(void);
 const char * llama_build_info(void);

-void llama_print_build_info(const char *);
+void llama_print_build_info(const char *, FILE * = stderr);
@@ -270,7 +270,7 @@ struct common_params_sampling {
        COMMON_SAMPLER_TYPE_TEMPERATURE,
    };

-    common_grammar              grammar;      // optional grammar constraint (user / output-format / tool-calls)
+    common_grammar                      grammar;          // optional grammar constraint (user / output-format / tool-calls)
    bool                                grammar_lazy = false;
    std::vector<common_grammar_trigger> grammar_triggers; // optional triggers (for lazy grammars)
    std::set<llama_token>               preserved_tokens;
@@ -657,6 +657,7 @@ struct common_params {
    std::string ssl_file_cert = "";                                                                         // NOLINT

    std::map<std::string, std::string> default_template_kwargs;
+    bool preserve_reasoning_specified = false;

    // CLI params
    std::string server_base; // if set, connect to this server instead of starting a new one
@@ -748,6 +748,10 @@ private:
            optional_props.push_back("*");
        }

+        if (required_props.empty() && optional_props.empty()) {
+            return "\"{\" space \"}\"";
+        }
+
        std::string rule = "\"{\" space ";
        for (size_t i = 0; i < required_props.size(); i++) {
            if (i > 0) {
@@ -188,6 +188,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
    "NanbeigeForCausalLM": "nanbeige",
    "NemotronForCausalLM": "nemotron",
    "NemotronHForCausalLM": "nemotron",
+    "NemotronHPuzzleForCausalLM": "nemotron",
    "NeoBERT": "bert",
    "NeoBERTForSequenceClassification": "bert",
    "NeoBERTLMHead": "bert",
@@ -286,6 +287,7 @@ MMPROJ_MODEL_MAP: dict[str, str] = {
    "CogVLMForCausalLM": "cogvlm",
    "DeepseekOCR2ForCausalLM": "deepseek",
    "DeepseekOCRForCausalLM": "deepseek",
+    "DeepseekV4ForCausalLM": "deepseek",
    "Dots3NoteForCausalLM": "dots3",
    "Dots3NoteForConditionalGeneration": "dots3",
    "DotsOCRForCausalLM": "dotsocr",
@@ -578,6 +578,8 @@ class DeepseekV4Model(TextModel):
    @classmethod
    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
        name, gen = item
+        if name.startswith(("aligner.", "image_")):
+            return None
        if name.startswith("mtp."):
            if not cls.mtp_only:
                cls._skipped_mtp_tensors += 1
@@ -853,6 +855,7 @@ class DeepseekV4Model(TextModel):
            "ffn_norm.weight": (gguf.MODEL_TENSOR.FFN_NORM, ".weight"),
            "ffn.gate.weight": (gguf.MODEL_TENSOR.FFN_GATE_INP, ".weight"),
            "ffn.gate.bias": (gguf.MODEL_TENSOR.FFN_EXP_PROBS_B, ".bias"),
+            "ffn.gate.bias_vl": (gguf.MODEL_TENSOR.FFN_EXP_PROBS_B_VL, ".bias"),
            "ffn.gate.tid2eid": (gguf.MODEL_TENSOR.FFN_GATE_TID2EID, ".weight"),
            "ffn.shared_experts.w1.weight": (gguf.MODEL_TENSOR.FFN_GATE_SHEXP, ".weight"),
            "ffn.shared_experts.w2.weight": (gguf.MODEL_TENSOR.FFN_DOWN_SHEXP, ".weight"),
@@ -878,6 +881,10 @@ class DeepseekV4Model(TextModel):
        if re.match(r"layers\.\d+\.ffn\.experts\.\d+\.w[123]\.(weight|scale)$", name):
            return []

+        # hash layers route text tokens via tid2eid and image tokens via bias_vl; gate.bias is unused
+        if name.endswith(".ffn.gate.bias") and bid is not None and bid < self.hparams["num_hash_layers"]:
+            return []
+
        tensor_key, suffix = self._map_dsv4_tensor_name(name, bid)
        if tensor_key == gguf.MODEL_TENSOR.FFN_GATE_TID2EID:
            return []
@@ -1000,6 +1007,13 @@ class DeepseekV4DSparkModel(DeepseekV4Model):
            return self._DSPARK_ROOT_MAP[name]
        return super()._map_dsv4_tensor_name(name, bid)

+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # the DFlash draft uses the plain exp-probs bias (ffn.gate.bias -> FFN_EXP_PROBS_B);
+        # the mtmd-only hash routing tensors (bias_vl, tid2eid) are not part of the DFLASH arch
+        if name.endswith(".ffn.gate.bias_vl"):
+            return
+        yield from super().modify_tensors(data_torch, name, bid)
+
    def set_vocab(self):
        if self.target_model_dir is None:
            raise ValueError("DeepSeek-V4 DSpark requires --target-model-dir with the target tokenizer")
@@ -1018,3 +1032,73 @@ class DeepseekV4DSparkModel(DeepseekV4Model):

        self.gguf_writer.add_block_size(self.hparams["dspark_block_size"])
        self.gguf_writer.add_target_layers([layer + 1 for layer in self.hparams["dspark_target_layer_ids"]])
+
+
+@ModelBase.register("DeepseekV4ForCausalLM")
+@ModelBase.example("deepseek-ai/DeepSeek-V4-Flash-Vision-Exp")
+class DeepseekV4FlashVisionModel(MmprojModel):
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        assert self.hparams_vision is not None
+        # no preprocessor_config.json in the repo; normalization is (x/255 - 0.5) / 0.5
+        # ref: inference/image_processor.py (load_image)
+        self.preprocessor_config = {
+            "image_mean": [0.5, 0.5, 0.5],
+            "image_std":  [0.5, 0.5, 0.5],
+            **self.preprocessor_config,
+        }
+
+    def get_vision_config(self) -> dict[str, Any] | None:
+        cfg = self.global_config
+        if cfg.get("vision_n_layers", 0) == 0:
+            raise ValueError("DeepseekV4FlashVisionModel requires vision_n_layers > 0 in the model config")
+        return {
+            "num_hidden_layers":   cfg["vision_n_layers"],
+            "hidden_size":         cfg["vision_dim"],
+            "num_attention_heads": cfg["vision_n_heads"],
+            "intermediate_size":   cfg["vision_inter_dim"],
+            "patch_size":          cfg["vision_patch_size"],
+            # dynamic resolution; only used for compat / warmup
+            "image_size":          cfg["vision_patch_size"] * cfg["vision_downsample_ratio"] * 16,
+            "rope_theta":          cfg.get("vision_rope_theta", 10000.0),
+            "downsample_ratio":    cfg["vision_downsample_ratio"],
+            "min_pixels":          cfg["vision_min_pixels"],
+        }
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        assert self.hparams_vision is not None
+        self.gguf_writer.add_clip_projector_type(gguf.VisionProjectorType.DEEPSEEK4V)
+        # vision RMSNorm eps is the pytorch default, NOT the LLM's rms_norm_eps (1e-20)
+        # ref: inference/vision.py (RMSNorm)
+        self.gguf_writer.add_vision_attention_layernorm_eps(1e-6)
+        self.gguf_writer.add_vision_use_silu(True) # SwiGLU MLP
+        self.gguf_writer.add_vision_projector_scale_factor(self.hparams_vision["downsample_ratio"])
+        self.gguf_writer.add_vision_min_pixels(self.hparams_vision["min_pixels"])
+        # hardcoded on the C++ side (see PROJECTOR_TYPE_DEEPSEEK4V in clip.cpp)
+        # if future models use different values, add GGUF keys for those
+        assert self.global_config["vision_max_n_token"] == 384
+        assert self.global_config["vision_max_wh_ratio"] == 8
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, _ = item
+        if not (name.startswith(("vision.", "aligner.", "image_"))):
+            return None
+        return super().filter_tensors(item)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        assert self.hparams_vision is not None
+        if name == "vision.patch_embed.proj.weight":
+            # nn.Linear over flattened (3, p, p) patches == conv2d weight
+            p = self.hparams_vision["patch_size"]
+            data_torch = data_torch.reshape(data_torch.shape[0], 3, p, p)
+
+        if ".mlp.w1." in name:
+            # fused SwiGLU gate+up
+            gate, up = data_torch.chunk(2, dim=0)
+            yield from super().modify_tensors(gate, name.replace("w1", "w1_gate"), bid)
+            yield from super().modify_tensors(up, name.replace("w1", "w1_up"), bid)
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
@@ -5,6 +5,7 @@ from typing import Any, Callable, Iterable, TYPE_CHECKING
 import torch

 if TYPE_CHECKING:
+    from pathlib import Path
    from torch import Tensor

 from .base import MmprojModel, ModelBase, TextModel, gguf, logger
@@ -201,6 +202,7 @@ class NemotronHModel(GraniteHybridModel):
    model_arch = gguf.MODEL_ARCH.NEMOTRON_H
    is_moe: bool = False
    supports_mtp_export = True
+    _experts: list[dict[str, Tensor]] | None = None

    _SSM_LAYER_TYPES = {"mamba", "linear_attention"}
    _ATTN_LAYER_TYPES = {"attention", "full_attention"}
@@ -513,3 +515,88 @@ class NemotronHModel(GraniteHybridModel):
            experts = [k for d in self._experts for k in d.keys()]
            if len(experts) > 0:
                raise ValueError(f"Unprocessed experts: {experts}")
+
+
+@ModelBase.register("NemotronHPuzzleForCausalLM")
+@ModelBase.example("nvidia/NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16")
+class NemotronHPuzzleModel(NemotronHModel):
+    """NVIDIA Puzzle: NemotronH with a per-block MoE config (block_configs).
+
+    The checkpoint also ships an MTP draft head (mtp.safetensors). It is skipped
+    here: there is no Puzzle MTP inference path in tree, and the head is laid out
+    by mtp_block_configs rather than the mtp.layers.* form NemotronHModel maps."""
+
+    model_arch = gguf.MODEL_ARCH.NEMOTRON_H_MOE
+    is_moe: bool = True
+    supports_mtp_export = False
+
+    def __init__(self, dir_model: "Path", *args, **kwargs):
+        hparams = dict(kwargs.pop("hparams", None) or ModelBase.load_hparams(dir_model, self.is_mistral_format))
+
+        self.block_configs: list[dict] = hparams["block_configs"]
+        self.n_layer_trunk = len(self.block_configs)
+
+        # block_configs carries the per-block MoE shape, and is the authority on the
+        # block pattern too: the layers_block_type the HF config wrapper computes is
+        # not sized to it.
+        hparams["num_hidden_layers"] = self.n_layer_trunk
+        hparams["layers_block_type"] = [bc["block_type"] for bc in self.block_configs]
+
+        self.model_arch = gguf.MODEL_ARCH.NEMOTRON_H_MOE
+
+        # Bypass NemotronHModel.__init__: it assumes a flat num_experts_per_tok /
+        # moe_intermediate_size and a layers_block_type sized to block_count, neither
+        # of which hold for Puzzle's per-block config.
+        GraniteHybridModel.__init__(self, dir_model, *args, hparams=hparams, **kwargs)
+
+        self.head_dim = self.find_hparam(["head_dim", "attention_head_dim"])
+        self.d_inner = self.find_hparam(["num_heads"]) * self.d_model
+
+        # NemotronHModel.__init__ folds an MTP block into block_count when the
+        # config carries num_nextn_predict_layers; Puzzle's config does, but its
+        # head has a different layout and no inference path, so stay opted out.
+        self._mtp_bid = None
+
+    def set_gguf_parameters(self):
+        GraniteHybridModel.set_gguf_parameters(self)
+
+        head_dim = self.head_dim
+        if head_dim is None:
+            raise ValueError("Could not find the attention head dim in config")
+        self.gguf_writer.add_key_length(head_dim)
+        self.gguf_writer.add_value_length(head_dim)
+
+        ffn_lengths = [bc.get("moe_intermediate_size") or 0 for bc in self.block_configs]
+        experts_used = [bc.get("num_experts_per_tok") or 0 for bc in self.block_configs]
+
+        self.gguf_writer.add_feed_forward_length(ffn_lengths)
+        self.gguf_writer.add_expert_feed_forward_length(ffn_lengths)
+        self.gguf_writer.add_expert_used_count(experts_used)
+
+        self.gguf_writer.add_expert_shared_feed_forward_length(self.hparams["moe_shared_expert_intermediate_size"])
+        self.gguf_writer.add_expert_count(self.hparams["n_routed_experts"])
+        self.gguf_writer.add_expert_shared_count(self.hparams["n_shared_experts"])
+        self.gguf_writer.add_expert_weights_norm(self.hparams["norm_topk_prob"])
+        self.gguf_writer.add_expert_weights_scale(self.hparams["routed_scaling_factor"])
+        self.gguf_writer.add_expert_group_count(self.hparams["n_group"])
+        self.gguf_writer.add_moe_latent_size(self.hparams["moe_latent_size"])
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        # The official BF16 checkpoint (NVIDIA-Nemotron-Labs-3-Puzzle-75B-A9B-BF16)
+        # names the trunk "model.*" (model.layers.*, model.embeddings, model.norm_f)
+        # where the original release used the NemotronH-style "backbone.*", and spells
+        # the router bias "e_score_correction_bias" instead of "e_score_correction.bias";
+        # normalize so both convert identically.
+        if name.startswith("model."):
+            name = "backbone." + name[len("model."):]
+        if name.endswith("mixer.gate.e_score_correction_bias"):
+            name = name[: -len("e_score_correction_bias")] + "e_score_correction.bias"
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        # Drop the MTP head unconditionally; see the class docstring.
+        if item[0].startswith("mtp."):
+            return None
+        return super().filter_tensors(item)
@@ -276,6 +276,10 @@ class Qwen3TTSSpeakerEncoderModel(MmprojModel):
        # ConvTranspose1d kernels: only F16/F32 are implemented, no BF16
        if new_name.endswith(".conv.weight") and (".up.blk." in new_name or ".dac.blk." in new_name):
            return gguf.GGMLQuantizationType.F32
+        # the code predictor FFN intermediate peaks around 1.5e5, above the F16 range, and mul_mat
+        # casts its input to the weight type
+        if new_name.startswith("a.gen.code.blk.") and new_name.endswith(".ffn_down.weight"):
+            return gguf.GGMLQuantizationType.F32
        return super().tensor_force_quant(name, new_name, bid, n_dims)

    @classmethod
@@ -806,7 +806,7 @@ User can use the device management in [docs/multi-gpu.md](https://github.com/ggm
 | GGML_SYCL_ENABLE_MKL_FA | 1 (default) or 0 | Enable oneMKL GEMM flash attention for XMX-accelerated prompt processing with quantized KV cache. Automatically activates during prefill (prompt processing) when all conditions are met: (1) flash-attn enabled (`-fa` or `--flash-attn on`), (2) KV cache quantized (`--cache-type-k q8_0 --cache-type-v q8_0` or other `*_0/*_1` types), (3) batch size ≥ 1024 (`--batch-size 1024`), (4) prompt length ≥ 1024 tokens. Set to 0 to force the TILE kernel for A/B testing. Example minimum command: `llama-cli -m model.gguf -fa -ngl 99 --cache-type-k q8_0 --cache-type-v q8_0 --batch-size 1024 -p "your prompt"` |
 | GGML_SYCL_MKL_FA_DEBUG | 0 (default) or 1 | Enable per-call diagnostic logging for MKL flash attention: GEMM/softmax timings, interleaved-head detection, and buffer memory usage. |
 | GGML_SYCL_MKL_FA_DIAG | 0 (default) or 1 | Enable output fingerprinting for MKL flash attention. Dumps the first 64 float output values for the first 6 FA calls with n_kv ≥ 1024, labeled with kernel type (MKL/TILE/VEC) for cross-kernel comparison. |
-| GGML_SYCL_ENABLE_FUSION | 0 or 1 (default) | Enable fused-kernel dispatch in graph compute. |
+| GGML_SYCL_ENABLE_FUSION | 0 or 1 (default) | Enable fused-kernel dispatch in graph compute. Unsupported types and layouts fall back to the standalone op kernels. See `ggml_sycl_can_fuse()`. |
 | GGML_SYCL_ENABLE_ESIMD | 0 or 1 (default)| Enable ESIMD kernels when available. |
 | ZES_ENABLE_SYSMAN | 0 (default) or 1 | Support to get free memory of GPU by sycl::aspect::ext_intel_free_memory.<br>Recommended to use when --split-mode = layer |
 | UR_L0_ENABLE_RELAXED_ALLOCATION_LIMITS | 0 (default) or 1 | Allow SYCL/Unified Runtime Level Zero device allocations larger than 4 GiB. llama.cpp's direct Level Zero allocation path requests the relaxed maximum-size limit itself when GGML_SYCL_ENABLE_LEVEL_ZERO=1. |
@@ -27,6 +27,7 @@ The following sections describe how to build with different backends and options
 * [OpenCL](#opencl)
 * [Android](#android-1)
 * [OpenVINO](#openvino)
+* [Hexagon](#hexagon)
 * [Notes about GPU-accelerated backends](#notes-about-gpu-accelerated-backends)

 ## CPU Build
@@ -299,7 +300,6 @@ The following compilation options are also available to tweak performance:
 |-------------------------------|------------------------|---------|----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
 | GGML_CUDA_FORCE_MMQ           | Boolean                | false   | Force the use of custom matrix multiplication kernels for quantized models instead of FP16 cuBLAS even if there is no int8 tensor core implementation available (affects V100, CDNA and RDNA3+). MMQ kernels are enabled by default on GPUs with int8 tensor core support. With MMQ force enabled, speed for large batch sizes will be worse but VRAM consumption will be lower. |
 | GGML_CUDA_FORCE_CUBLAS        | Boolean                | false   | Force the use of FP16 cuBLAS instead of custom matrix multiplication kernels for quantized models. There may be issues with numerical overflows (except for V100, CDNA and RDNA4 which use FP32 compute type by default) and memory use will be higher. Prompt processing may become faster on recent datacenter GPUs (the custom kernels were tuned primarily for RTX 3000/4000).   |
-| GGML_CUDA_PEER_MAX_BATCH_SIZE | Positive integer       | 128     | Maximum batch size for which to enable peer access between multiple GPUs. Peer access requires either Linux or NVLink. When using NVLink enabling peer access for larger batch sizes is potentially beneficial.                                                                                                                                                                  |
 | GGML_CUDA_FA_ALL_QUANTS       | Boolean                | false   | Compile support for all KV cache quantization type (combinations) for the FlashAttention CUDA kernels. More fine-grained control over KV cache size but compilation takes much longer.                                                                                                                                                                                           |

 ## MUSA
@@ -830,6 +830,9 @@ To read documentation for how to build on IBM Z & LinuxONE, [click here](./build

 For build instructions and usage examples, refer to [OPENVINO.md](backend/OPENVINO.md).

+### Hexagon
+
+Check [README.md](./backend/snapdragon/README.md) for target specific build and run info.

 ---
 ## Notes about GPU-accelerated backends
@@ -12,116 +12,117 @@ Legend:
 - 🟡 Partially supported by this backend
 - ❌ Not supported by this backend

-| Operation | BLAS | CANN | CPU | CUDA | ET | MTL | OpenCL | SYCL | Vulkan | WebGPU | ZenDNN | zDNN |
-|-----------|------|------|------|------|------|------|------|------|------|------|------|------|
-|                              ABS | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                              ACC | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                              ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             ADD1 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
-|                           ADD_ID | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                           ARANGE | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             CEIL | ❌ | ❌ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                        COL2IM_1D | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                           CONCAT | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
-|                             CONT | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
-|                          CONV_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                          CONV_3D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                              COS | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
-|               CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|          CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                           CUMSUM | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             DIAG | ❌ | ❌ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                    DIAG_MASK_INF | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                              DIV | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                     DSV4_HC_COMB | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                     DSV4_HC_POST | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                      DSV4_HC_PRE | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                              DUP | ❌ | ✅ | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                            EXPM1 | ❌ | ❌ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             FILL | ❌ | ❌ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
-|                            FLOOR | ❌ | ❌ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                  GATED_DELTA_NET | ❌ | ❌ | ✅ | ❌ | ✅ | 🟡 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                GATED_LINEAR_ATTN | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                      GEGLU_QUICK | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             GELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                         GELU_ERF | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                       GELU_QUICK | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ | ❌ |
-|                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
-|                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                           IM2COL | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                        IM2COL_3D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                          L2_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ✅ | ✅ | 🟡 | ❌ | ❌ |
-|                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                LIGHTNING_INDEXER | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                              LOG | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                              MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                          MUL_MAT | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
-|                 MUL_MAT_HADAMARD | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | 🟡 | 🟡 | ❌ |
-|                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | 🟡 | ❌ | ❌ |
-|                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                     OPT_STEP_SGD | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                         OUT_PROD | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 |
-|                              PAD | ❌ | 🟡 | ✅ | 🟡 | ❌ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ |
-|                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                          POOL_1D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             RELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                           REPEAT | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
-|                      REPEAT_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                         RMS_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                             ROLL | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                             ROPE | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                        ROPE_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                            ROUND | ❌ | ❌ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                        RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                        RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                            SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                              SET | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ |
-|                         SET_ROWS | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ |
-|                              SGN | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                          SIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
-|                              SIN | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                         SOFTPLUS | ❌ | ❌ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ |
-|                        SOLVE_TRI | ❌ | ❌ | ✅ | 🟡 | 🟡 | ✅ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ |
-|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             SQRT | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                         SSM_CONV | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
-|                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                              SUM | ❌ | 🟡 | ✅ | 🟡 | ❌ | 🟡 | ❌ | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
-|                         SUM_ROWS | ❌ | ✅ | ✅ | 🟡 | ❌ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ |
-|                           SWIGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                       SWIGLU_OAI | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                            TOP_K | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
-|                              TRI | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                            TRUNC | ❌ | ❌ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                            XIELU | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+| Operation | BLAS | CANN | CPU | CUDA | ET | HTP | MTL | OpenCL | SYCL | Vulkan | WebGPU | ZenDNN | zDNN |
+|-----------|------|------|------|------|------|------|------|------|------|------|------|------|------|
+|                              ABS | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              ACC | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                              ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             ADD1 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                           ADD_ID | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                           ARANGE | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             CEIL | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                        COL2IM_1D | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                           CONCAT | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
+|                             CONT | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
+|                          CONV_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                          CONV_3D | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                              COS | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                      COUNT_EQUAL | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
+|               CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|          CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                           CUMSUM | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             DIAG | ❌ | ❌ | ✅ | ✅ | 🟡 | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                    DIAG_MASK_INF | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                              DIV | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                     DSV4_HC_COMB | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                     DSV4_HC_POST | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                      DSV4_HC_PRE | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                              DUP | ❌ | ✅ | ✅ | 🟡 | ❌ | ❌ | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                            EXPM1 | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             FILL | ❌ | ❌ | ✅ | ✅ | 🟡 | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
+|                            FLOOR | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                  GATED_DELTA_NET | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | 🟡 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                GATED_LINEAR_ATTN | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                      GEGLU_QUICK | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             GELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                         GELU_ERF | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                       GELU_QUICK | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ | ❌ |
+|                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ |
+|                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                           IM2COL | ❌ | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                        IM2COL_3D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                          L2_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | 🟡 | ❌ | ❌ |
+|                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | 🟡 | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                LIGHTNING_INDEXER | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                              LOG | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                              MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                          MUL_MAT | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
+|                 MUL_MAT_HADAMARD | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | ✅ | 🟡 | 🟡 | ❌ |
+|                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | 🟡 | ❌ | ❌ |
+|                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                     OPT_STEP_SGD | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                         OUT_PROD | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | 🟡 |
+|                              PAD | ❌ | 🟡 | ✅ | 🟡 | ❌ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ |
+|                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                          POOL_1D | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             RELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                           REPEAT | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
+|                      REPEAT_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                         RMS_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                             ROLL | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                             ROPE | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                        ROPE_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                            ROUND | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                        RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                        RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                            SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              SET | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ |
+|                         SET_ROWS | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | 🟡 | ❌ | ❌ |
+|                              SGN | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                          SIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
+|                              SIN | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                         SOFTPLUS | ❌ | ❌ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ |
+|                        SOLVE_TRI | ❌ | ❌ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ |
+|                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             SQRT | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                         SSM_CONV | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
+|                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              SUM | ❌ | 🟡 | ✅ | 🟡 | ❌ | ❌ | 🟡 | ❌ | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
+|                         SUM_ROWS | ❌ | ✅ | ✅ | 🟡 | ❌ | 🟡 | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ |
+|                           SWIGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                     SWIGLU_CLAMP | ❌ | ❌ | ❌ | ❌ | ❌ | 🟡 | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                       SWIGLU_OAI | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                            TOP_K | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
+|                              TRI | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                            TRUNC | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                            XIELU | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
@@ -734,6 +734,9 @@ class SchemaConverter:
            )
            optional_props.append("*")

+        if not required_props and not optional_props:
+            return '"{" space "}"'
+
        rule = '"{" space '
        rule += ' "," space '.join(prop_kv_rule_names[k] for k in required_props)

@@ -6,6 +6,8 @@ Finetuning of Stories 260K and LLaMA 3.2 1b seems to work with 24 GB of memory.
 **For CPU training, compile llama.cpp without any additional backends such as CUDA.**
 **For CUDA training, use the maximum number of GPU layers.**

+Flash attention is disabled during training because `FLASH_ATTN_EXT` has no backward pass.
+
 Proof of concept:

 ``` sh
@@ -4,7 +4,7 @@ project("ggml" C CXX ASM)

 ### GGML Version
 set(GGML_VERSION_MAJOR 0)
-set(GGML_VERSION_MINOR 22)
+set(GGML_VERSION_MINOR 23)
 set(GGML_VERSION_PATCH 0)
 set(GGML_VERSION_BASE "${GGML_VERSION_MAJOR}.${GGML_VERSION_MINOR}.${GGML_VERSION_PATCH}")

@@ -200,8 +200,6 @@ option(GGML_CUDA                            "ggml: use CUDA"
 option(GGML_MUSA                            "ggml: use MUSA"                                  OFF)
 option(GGML_CUDA_FORCE_MMQ                  "ggml: use mmq kernels instead of cuBLAS"         OFF)
 option(GGML_CUDA_FORCE_CUBLAS               "ggml: always use cuBLAS instead of mmq kernels"  OFF)
-set   (GGML_CUDA_PEER_MAX_BATCH_SIZE "128" CACHE STRING
-                                            "ggml: max. batch size for using peer access")
 option(GGML_CUDA_NO_PEER_COPY               "ggml: do not use peer to peer copies"            OFF)
 option(GGML_CUDA_NO_VMM                     "ggml: do not try to use CUDA VMM"                OFF)
 option(GGML_CUDA_FA                         "ggml: compile ggml FlashAttention CUDA kernels"  ON)
@@ -242,6 +240,8 @@ option(GGML_METAL_EMBED_LIBRARY             "ggml: embed Metal library"
 set   (GGML_METAL_MACOSX_VERSION_MIN "" CACHE STRING
                                            "ggml: metal minimum macOS version")
 set   (GGML_METAL_STD "" CACHE STRING       "ggml: metal standard version (-std flag)")
+set   (GGML_METAL_TARGET_OS "macos" CACHE STRING
+                                            "ggml: metal -mtargetos OS name (macos, ios, xros, tvos)")
 option(GGML_OPENMP                          "ggml: use OpenMP"                                ON)
 option(GGML_OPENMP_FETCH                    "ggml: fetch LLVM OpenMP"                         OFF)
 option(GGML_RPC                             "ggml: use RPC"                                   OFF)
@@ -404,10 +404,6 @@ write_basic_package_version_file(
    VERSION ${GGML_INSTALL_VERSION}
    COMPATIBILITY SameMajorVersion)

-target_compile_definitions(ggml-base PRIVATE
-    GGML_VERSION="${GGML_INSTALL_VERSION}"
-    GGML_COMMIT="${GGML_BUILD_COMMIT}"
-)
 message(STATUS "ggml version: ${GGML_INSTALL_VERSION}")
 message(STATUS "ggml commit:  ${GGML_BUILD_COMMIT}")

@@ -424,10 +424,6 @@ extern "C" {
    // Compare the output of two backends
    GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data, struct ggml_tensor const * const * test_nodes, size_t num_test_nodes);

-    // returns true for ops that may require additional memory for fleeting data on some backends,
-    // i.e. the backend's get_alloc_size may return more than ggml_nbytes for the output tensor
-    GGML_API bool ggml_backend_op_alloc_size_may_expand(enum ggml_op op);
-
    // Tensor initialization
    GGML_API enum ggml_status ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);
    GGML_API enum ggml_status ggml_backend_view_init(struct ggml_tensor * tensor);
@@ -2453,6 +2453,12 @@ extern "C" {
    GGML_API enum ggml_prec ggml_flash_attn_ext_get_prec(
            const struct ggml_tensor * a);

+    // Use finite mask entries as a sparse K/V set. Set 0 to disable.
+    // n_kv_max must bound the number of finite entries in every mask row.
+    GGML_API void ggml_flash_attn_ext_set_n_kv_max(
+            struct ggml_tensor * a,
+            int32_t              n_kv_max);
+
    GGML_API void ggml_flash_attn_ext_add_sinks(
            struct ggml_tensor * a,
            struct ggml_tensor * sinks);
@@ -213,7 +213,9 @@ set_target_properties(ggml-base PROPERTIES
    SOVERSION ${GGML_VERSION_MAJOR}
 )

-target_include_directories(ggml-base PRIVATE .)
+configure_file(ggml-version.h.in ${CMAKE_CURRENT_BINARY_DIR}/ggml-version.h @ONLY)
+
+target_include_directories(ggml-base PRIVATE . ${CMAKE_CURRENT_BINARY_DIR})
 if (GGML_BACKEND_DL)
    target_compile_definitions(ggml-base PUBLIC GGML_BACKEND_DL)
 endif()
@@ -34,6 +34,11 @@ extern "C" {
        void * context;
    };

+    // [TAG_ALLOC_SIZE_EXPAND]
+    // returns true for ops that may require additional memory for fleeting data on some backends,
+    // i.e. the backend buffer type's get_alloc_size may return more than ggml_nbytes for the output tensor
+    GGML_API bool ggml_op_alloc_size_may_expand(enum ggml_op op);
+
    //
    // Backend buffer
    //
@@ -490,7 +490,13 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
 #endif
        // default search paths: executable directory, current directory
        search_paths.push_back(get_executable_path());
-        search_paths.push_back(fs::current_path());
+        std::error_code cwd_ec;
+        const fs::path cwd = fs::current_path(cwd_ec);
+        if (cwd_ec) {
+            GGML_LOG_DEBUG("%s: current_path() failure, error-message: %s\n", __func__, cwd_ec.message().c_str());
+        } else {
+            search_paths.push_back(cwd);
+        }
    } else {
        search_paths.push_back(fs::u8path(user_search_path));
    }
@@ -508,8 +514,14 @@ static ggml_backend_reg_t ggml_backend_load_best(const char * name, bool silent,
            }
            continue;
        }
-        fs::directory_iterator dir_it(search_path, fs::directory_options::skip_permission_denied);
-        for (const auto & entry : dir_it) {
+        std::error_code dir_ec;
+        fs::directory_iterator dir_it(search_path, fs::directory_options::skip_permission_denied, dir_ec);
+        if (dir_ec) {
+            GGML_LOG_DEBUG("%s: failed to enumerate %s: %s\n", __func__, path_str(search_path).c_str(), dir_ec.message().c_str());
+            continue;
+        }
+        for (const fs::directory_iterator end; dir_it != end; dir_it.increment(dir_ec)) {
+            const auto & entry = *dir_it;
            if (entry.is_regular_file(ec)) {
                auto filename = entry.path().filename();
                auto ext = entry.path().extension();
@@ -71,7 +71,7 @@ size_t ggml_backend_buft_get_alloc_size(ggml_backend_buffer_type_t buft, const s
        GGML_ASSERT(size <= ggml_nbytes(tensor) ||
                    ggml_op_is_empty(tensor->op) ||
                    ggml_is_quantized(tensor->type) || // [TAG_ALLOC_SIZE_EXPAND]
-                    ggml_backend_op_alloc_size_may_expand(tensor->op));
+                    ggml_op_alloc_size_may_expand(tensor->op));

        return size;
    }
@@ -2109,10 +2109,7 @@ ggml_backend_t ggml_backend_sched_get_tensor_backend(ggml_backend_sched_t sched,

 // utils

-// [TAG_ALLOC_SIZE_EXPAND]
-// returns true for ops that may require additional memory for fleeting data on some backends,
-// i.e. the backend's get_alloc_size may return more than ggml_nbytes for the output tensor
-bool ggml_backend_op_alloc_size_may_expand(enum ggml_op op) {
+bool ggml_op_alloc_size_may_expand(enum ggml_op op) {
    switch (op) {
        case GGML_OP_FLASH_ATTN_EXT:
        case GGML_OP_MUL_MAT:
@@ -455,12 +455,16 @@ function(ggml_add_cpu_backend_variant_impl tag_name)
                ggml-cpu/spacemit/repack.h
                ggml-cpu/spacemit/ime_env.cpp
                ggml-cpu/spacemit/ime_env.h
-                ggml-cpu/spacemit/ime1_kernels.cpp
-                ggml-cpu/spacemit/ime2_kernels.cpp
                ggml-cpu/spacemit/ime_kernels.h
                ggml-cpu/spacemit/rvv_kernels.cpp
                ggml-cpu/spacemit/rvv_kernels.h
            )
+            if ("RISCV64_SPACEMIT_IME1" IN_LIST RISCV64_SPACEMIT_IME_SPEC)
+                list(APPEND GGML_CPU_SOURCES ggml-cpu/spacemit/ime1_kernels.cpp)
+            endif()
+            if ("RISCV64_SPACEMIT_IME2" IN_LIST RISCV64_SPACEMIT_IME_SPEC)
+                list(APPEND GGML_CPU_SOURCES ggml-cpu/spacemit/ime2_kernels.cpp)
+            endif()
        endif()
        if(NOT GGML_CPU_ALL_VARIANTS)
            set(MARCH_STR "rv64gc")
@@ -636,7 +636,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const voi
        const float32x4_t v_xyf = vec_float(v_xy);

        const float32x4_t v_d = vec_splats(GGML_CPU_FP16_TO_FP32(x0->d) * GGML_CPU_FP16_TO_FP32(y0->d));
-        const float32x4_t v_acc = vec_madd(v_xyf, v_d, v_acc);
+        const float32x4_t v_acc = vec_madd(v_xyf, v_d, vec_splats(0.0f));

        sumf += vec_hsum_f32x4(v_acc) + summs;
    }
@@ -1823,7 +1823,7 @@ class extra_buffer_type : ggml::cpu::extra_buffer_type {
        const bool src0_is_kleidiai =
            op->src[0]->buffer &&
            (ggml_n_dims(op->src[0]) == 2) &&
-            op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type() &&
+            op->src[0]->buffer->buft->context == this &&
            slot_total > 0;

        if ((op->op == GGML_OP_MUL_MAT || op->op == GGML_OP_GET_ROWS) &&
@@ -1862,7 +1862,7 @@ class extra_buffer_type : ggml::cpu::extra_buffer_type {

    ggml::cpu::tensor_traits * get_tensor_traits(const struct ggml_tensor * op) override {
        if (op->op == GGML_OP_MUL_MAT || op->op == GGML_OP_GET_ROWS) {
-            if (op->src[0]->buffer && op->src[0]->buffer->buft == ggml_backend_cpu_kleidiai_buffer_type()) {
+            if (op->src[0]->buffer && op->src[0]->buffer->buft->context == this) {
                return (ggml::cpu::tensor_traits *) op->src[0]->extra;
            } else {
                // KleidiAI only has kernels for Q4_0 and Q8_0. For a quantized weight of any
@@ -129,8 +129,6 @@ if (CUDAToolkit_FOUND)
                             ${GGML_SOURCES_CUDA}
                            )

-    add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})
-
    if (GGML_CUDA_GRAPHS)
        add_compile_definitions(GGML_CUDA_USE_GRAPHS)
    endif()
@@ -52,6 +52,7 @@
 #define GGML_CUDA_CC_VOLTA           700
 #define GGML_CUDA_CC_TURING          750
 #define GGML_CUDA_CC_AMPERE          800
+#define GGML_CUDA_CC_ORIN            870
 #define GGML_CUDA_CC_ADA_LOVELACE    890
 #define GGML_CUDA_CC_HOPPER          900
 // While BW spans CC 1000, 1100 & 1200, we are integrating Tensor Core instructions available to 1200 family, see
@@ -718,6 +718,9 @@ static __global__ void flash_attn_mask_to_KV_max(
    KV_max[sequence*ne31 + jt] = KV_max_sj;
 }

+void ggml_cuda_flash_attn_ext_compact_mask(
+        const ggml_tensor * mask, int32_t * indices, int32_t n_kv_max, cudaStream_t stream);
+
 template<int D, int ncols1, int ncols2> // D == head size
 __launch_bounds__(D, 1)
 static __global__ void flash_attn_stream_k_fixup_uniform(
@@ -972,7 +975,8 @@ static __global__ void flash_attn_combine_results(
 template <int DV, int ncols1, int ncols2>
 void launch_fattn(
    ggml_backend_cuda_context & ctx, ggml_tensor * dst, fattn_kernel_t fattn_kernel, const int nwarps, const size_t nbytes_shared,
-    const int nbatch_fa, const bool need_f16_K, const bool need_f16_V, const bool stream_k, const int warp_size = WARP_SIZE
+    const int nbatch_fa, const bool need_f16_K, const bool need_f16_V, const bool stream_k, const bool use_sparse,
+    const int warp_size = WARP_SIZE
 ) {
    constexpr int ncols = ncols1 * ncols2;

@@ -1088,10 +1092,20 @@ void launch_fattn(
    const int ntiles_z_gqa = ((gqa_ratio + ncols2 - 1) / ncols2);
    const int ntiles_dst   = ntiles_x * ntiles_z_gqa * K->ne[2] * Q->ne[3];

+    const int32_t n_kv_max = use_sparse ? ggml_get_op_params_i32(KQV, 4) : 0;
+    if (use_sparse) {
+        GGML_ASSERT(mask != nullptr);
+        GGML_ASSERT(n_kv_max > 0);
+        const size_t mask_rows = size_t(mask->ne[1]) * mask->ne[3];
+
+        KV_max.alloc(size_t(n_kv_max) * mask_rows);
+        ggml_cuda_flash_attn_ext_compact_mask(mask, KV_max.ptr, n_kv_max, main_stream);
+    }
+
    // Optional optimization where the mask is scanned to determine whether part of the calculation can be skipped.
    // Only worth the overhead if there is at lease one FATTN_KQ_STRIDE x FATTN_KQ_STRIDE square to be skipped or
    //     multiple sequences of possibly different lengths.
-    if (mask && K->ne[1] % FATTN_KQ_STRIDE == 0 && (Q->ne[1] >= 1024 || Q->ne[3] > 1)) {
+    if (!use_sparse && mask && K->ne[1] % FATTN_KQ_STRIDE == 0 && (Q->ne[1] >= 1024 || Q->ne[3] > 1)) {
        const int64_t s31 = mask->nb[1] / sizeof(half2);
        const int64_t s33 = mask->nb[3] / sizeof(half2);

@@ -1114,7 +1128,8 @@ void launch_fattn(
    GGML_ASSERT(max_blocks_per_sm > 0);
    int parallel_blocks = max_blocks_per_sm;

-    const int ntiles_KV = (K->ne[1] + nbatch_fa - 1) / nbatch_fa; // Max. number of parallel blocks limited by KV cache length.
+    const int64_t n_kv = use_sparse ? n_kv_max : K->ne[1];
+    const int ntiles_KV = (n_kv + nbatch_fa - 1) / nbatch_fa; // Max. number of parallel blocks limited by KV cache length.

    dim3 blocks_num;
    if (stream_k) {
@@ -1218,7 +1233,7 @@ void launch_fattn(
        !stream_k && parallel_blocks > 1 ? dst_tmp.ptr : (float *) KQV->data, dst_tmp_meta.ptr,
        scale, max_bias, m0, m1, n_head_log2, logit_softcap,
        Q->ne[0], ne01,     Q->ne[2], Q->ne[3], Q->nb[1], Q->nb[2], Q->nb[3],
-        K->ne[0], K->ne[1], K->ne[2], K->ne[3], nb11, nb12, nb13,
+        K->ne[0], n_kv, K->ne[2], K->ne[3], nb11, nb12, nb13,
        nb21, nb22, nb23,
        mask ? mask->ne[1] : 0, mask ? mask->ne[2] : 0, mask ? mask->ne[3] : 0,
        mask ? mask->nb[1] : 0, mask ? mask->nb[2] : 0, mask ? mask->nb[3] : 0
@@ -350,20 +350,24 @@ static __host__ int ggml_cuda_fattn_mma_get_nstages(const int DKQ, const int DV,
    return cp_async_available(cc) && ncols2 >= 2 ? ggml_cuda_fattn_mma_get_nstages_target(DKQ, DV, ncols1*ncols2, cc) : 0;
 }

-static constexpr __device__ int ggml_cuda_fattn_mma_get_nstages(const int DKQ, const int DV, const int ncols1, const int ncols2) {
+static constexpr __device__ int ggml_cuda_fattn_mma_get_nstages(
+        const int DKQ, const int DV, const int ncols1, const int ncols2, const bool use_sparse) {
 #ifdef CP_ASYNC_AVAILABLE
-    return ncols2 >= 2 ? ggml_cuda_fattn_mma_get_nstages_target(DKQ, DV, ncols1*ncols2) : 0;
+    const int nstages_target = ncols2 >= 2 ? ggml_cuda_fattn_mma_get_nstages_target(DKQ, DV, ncols1*ncols2) : 0;
+    // sparse gather is not implemented for multi-stage loading
+    return use_sparse && nstages_target > 1 ? 1 : nstages_target;
 #else
-    GGML_UNUSED_VARS(DKQ, DV, ncols1, ncols2);
+    GGML_UNUSED_VARS(DKQ, DV, ncols1, ncols2, use_sparse);
    return 0;
 #endif // CP_ASYNC_AVAILABLE
 }

 // ------------------------------------------------------------------------------------------------------------------

-template<int stride_tile, bool swz, int nwarps, int nbatch_fa, bool use_cp_async, bool oob_check>
+template<int stride_tile, bool swz, int nwarps, int nbatch_fa, bool use_cp_async, bool oob_check, bool use_sparse>
 static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
-        const half2 * const __restrict__ KV, half2 * const __restrict__ tile_KV, const int D2, const int stride_KV, const int i_sup) {
+        const half2 * const __restrict__ KV, half2 * const __restrict__ tile_KV, const int D2, const int stride_KV,
+        const int k_VKQ_0, const int i_sup, const int32_t * const __restrict__ indices) {
    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
    // K/V data is loaded with decreasing granularity for D for better memory bandwidth.
    // The minimum granularity is 16 bytes.
@@ -371,7 +375,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
    const int chunks_per_row = D2 / h2_per_chunk;
    if constexpr (use_cp_async) {
        static_assert(warp_size == 32, "bad warp_size");
-        static_assert(!oob_check, "OOB check not compatible with cp_async");
+        static_assert(!oob_check || use_sparse, "OOB check not compatible with cp_async");
        constexpr int preload = 64;

        const unsigned int tile_KV_32 = ggml_cuda_cvta_generic_to_shared(tile_KV);
@@ -394,15 +398,24 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
                    break;
                }

+                int64_t i_KV;
+                if constexpr (use_sparse) {
+                    // padded slots gather row 0, the -inf mask removes their contribution
+                    const int32_t index = i < i_sup ? indices[k_VKQ_0 + i] : 0;
+                    i_KV = index >= 0 ? index : 0;
+                } else {
+                    i_KV = k_VKQ_0 + i;
+                }
+
 #pragma unroll
                for (int k0 = k0_start; k0 < k0_stop; k0 += stride_k) {
                    const int k = k0 + (stride_k == warp_size ? threadIdx.x : threadIdx.x % stride_k);

                    if constexpr (swz) {
                        const int smem_offs_b = ggml_cuda_fattn_smem_swizzle::bytes_rc<stride_tile>(i, k*h2_per_chunk);
-                        cp_async_cg_16<preload>(tile_KV_32 + smem_offs_b, KV + i*stride_KV + k*h2_per_chunk);
+                        cp_async_cg_16<preload>(tile_KV_32 + smem_offs_b, KV + i_KV*stride_KV + k*h2_per_chunk);
                    } else {
-                        cp_async_cg_16<preload>(tile_KV_32 + i*(stride_tile*sizeof(half2)) + k*16, KV + i*stride_KV + k*h2_per_chunk);
+                        cp_async_cg_16<preload>(tile_KV_32 + i*(stride_tile*sizeof(half2)) + k*16, KV + i_KV*stride_KV + k*h2_per_chunk);
                    }
                }
            }
@@ -438,12 +451,17 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
                for (int k0 = k0_start; k0 < k0_stop; k0 += stride_k) {
                    const int k = k0 + (stride_k == warp_size ? threadIdx.x : threadIdx.x % stride_k);

-                    if constexpr (swz) {
-                        ggml_cuda_memcpy_1<16>((char *) tile_KV + ggml_cuda_fattn_smem_swizzle::bytes_rc<stride_tile>(i, k*h2_per_chunk),
-                            !oob_check || i < i_sup ? KV + i*stride_KV + k*h2_per_chunk : zero);
+                    const half2 * src;
+                    if constexpr (use_sparse) {
+                        const int32_t index = i < i_sup ? indices[k_VKQ_0 + i] : -1;
+                        src = index >= 0 ? KV + int64_t(index)*stride_KV + k*h2_per_chunk : zero;
                    } else {
-                        ggml_cuda_memcpy_1<16>(tile_KV + i*stride_tile + k*4,
-                            !oob_check || i < i_sup ? KV + i*stride_KV + k*h2_per_chunk : zero);
+                        src = !oob_check || i < i_sup ? KV + int64_t(k_VKQ_0 + i)*stride_KV + k*h2_per_chunk : zero;
+                    }
+                    if constexpr (swz) {
+                        ggml_cuda_memcpy_1<16>((char *) tile_KV + ggml_cuda_fattn_smem_swizzle::bytes_rc<stride_tile>(i, k*h2_per_chunk), src);
+                    } else {
+                        ggml_cuda_memcpy_1<16>(tile_KV + i*stride_tile + k*4, src);
                    }
                }
            }
@@ -458,14 +476,16 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_tile(
    }
 }

-template<int ncols1, int nwarps, int nbatch_fa, bool use_cp_async, bool oob_check>
+template<int ncols1, int nwarps, int nbatch_fa, bool use_cp_async, bool oob_check, bool use_sparse>
 static __device__ __forceinline__ void flash_attn_ext_f16_load_mask(
        const half * const __restrict__ mask_h, half * const __restrict__ tile_mask,
-        const int stride_mask, const int i_sup, const int j0, const uint3 ne01) {
+        const int stride_mask, const int k_VKQ_0, const int i_sup, const int j0, const uint3 ne01,
+        const int32_t * const __restrict__ indices) {
    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
    if constexpr (use_cp_async) {
        static_assert(nbatch_fa <= 8*warp_size && nbatch_fa % 8 == 0, "bad nbatch_fa");
        static_assert(!oob_check, "OOB check incompatible with cp_async");
+        static_assert(!use_sparse, "sparse gather incompatible with cp_async");
        constexpr int preload = nbatch_fa >= 32 ? nbatch_fa * sizeof(half) : 64;
        constexpr int cols_per_warp = 8*warp_size/nbatch_fa;
        constexpr int stride_j = nwarps * cols_per_warp;
@@ -483,9 +503,9 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_mask(

            const int i = 8 * (threadIdx.x % (nbatch_fa/8));

-            cp_async_cg_16<preload>(tile_mask_32 + j_sram*(nbatch_fa*sizeof(half) + 16) + i*sizeof(half), mask_h + int64_t(j_vram)*stride_mask + i);
+            cp_async_cg_16<preload>(tile_mask_32 + j_sram*(nbatch_fa*sizeof(half) + 16) + i*sizeof(half), mask_h + int64_t(j_vram)*stride_mask + k_VKQ_0 + i);
        }
-    } else if constexpr (oob_check) {
+    } else if constexpr (oob_check || use_sparse) {
 #pragma unroll
        for (int j1 = 0; j1 < ncols1; j1 += nwarps) {
            const int j_sram = j1 + threadIdx.y;
@@ -499,7 +519,12 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_mask(
            for (int i0 = 0; i0 < nbatch_fa; i0 += warp_size) {
                const int i = i0 + threadIdx.x;

-                tile_mask[j_sram*(nbatch_fa + 8) + i] = i < i_sup ? mask_h[int64_t(j_vram)*stride_mask + i] : half(0.0f);
+                if constexpr (use_sparse) {
+                    const int32_t index = i < i_sup ? indices[k_VKQ_0 + i] : -1;
+                    tile_mask[j_sram*(nbatch_fa + 8) + i] = index >= 0 ? mask_h[int64_t(j_vram)*stride_mask + index] : half(-INFINITY);
+                } else {
+                    tile_mask[j_sram*(nbatch_fa + 8) + i] = i < i_sup ? mask_h[int64_t(j_vram)*stride_mask + k_VKQ_0 + i] : half(0.0f);
+                }
            }
        }
    } else if constexpr (nbatch_fa < 2*warp_size) {
@@ -516,7 +541,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_mask(

            const int i = threadIdx.x % (warp_size/cols_per_warp);

-            ggml_cuda_memcpy_1<sizeof(half2)>(tile_mask + j_sram*(nbatch_fa + 8) + 2*i, mask_h + int64_t(j_vram)*stride_mask + 2*i);
+            ggml_cuda_memcpy_1<sizeof(half2)>(tile_mask + j_sram*(nbatch_fa + 8) + 2*i, mask_h + int64_t(j_vram)*stride_mask + k_VKQ_0 + 2*i);
        }
    } else {
 #pragma unroll
@@ -532,20 +557,21 @@ static __device__ __forceinline__ void flash_attn_ext_f16_load_mask(
            for (int i0 = 0; i0 < nbatch_fa; i0 += 2*warp_size) {
                const int i = i0 + 2*threadIdx.x;

-                ggml_cuda_memcpy_1<sizeof(half2)>(tile_mask + j_sram*(nbatch_fa + 8) + i, mask_h + int64_t(j_vram)*stride_mask + i);
+                ggml_cuda_memcpy_1<sizeof(half2)>(tile_mask + j_sram*(nbatch_fa + 8) + i, mask_h + int64_t(j_vram)*stride_mask + k_VKQ_0 + i);
            }
        }
    }
 }

 template<int DKQ, int DV, int ncols1, int ncols2, int nwarps,
-    bool use_logit_softcap, bool V_is_K_view, bool needs_fixup, bool is_fixup, bool last_iter, bool oob_check,
+    bool use_logit_softcap, bool V_is_K_view, bool use_sparse, bool needs_fixup, bool is_fixup, bool last_iter, bool oob_check,
    typename T_A_KQ, typename T_B_KQ, typename T_C_KQ, typename T_A_VKQ, typename T_B_VKQ, typename T_C_VKQ>
 static __device__ __forceinline__ void flash_attn_ext_f16_iter(
        const float2 * const __restrict__ Q_f2,
        const half2  * const __restrict__ K_h2,
        const half2  * const __restrict__ V_h2,
        const half   * const __restrict__ mask_h,
+        const int32_t * const __restrict__ indices,
        float2       * const __restrict__ dstk,
        float2       * const __restrict__ dstk_fixup,
        const float scale,
@@ -577,7 +603,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
    constexpr int  nbatch_K2       = ggml_cuda_fattn_mma_get_nbatch_K2(DKQ, DV, ncols);
    constexpr int  nbatch_V2       = ggml_cuda_fattn_mma_get_nbatch_V2(DKQ, DV, ncols);
    constexpr bool Q_in_reg        = ggml_cuda_fattn_mma_get_Q_in_reg (DKQ, DV, ncols);
-    constexpr int  nstages         = ggml_cuda_fattn_mma_get_nstages  (DKQ, DV, ncols1, ncols2);
+    constexpr int  nstages         = ggml_cuda_fattn_mma_get_nstages  (DKQ, DV, ncols1, ncols2, use_sparse);

    // swizzle the tile stride for K and V based on the batch size.
    constexpr int stride_tile_K = ggml_cuda_fattn_smem_swizzle::tile_stride(nbatch_K2);
@@ -601,13 +627,14 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
        constexpr bool use_cp_async = true;
        cp_async_wait_all();
        __syncthreads();
-        flash_attn_ext_f16_load_tile<stride_tile_V, swz_V, nwarps, nbatch_fa, use_cp_async, oob_check>
-            (V_h2 + int64_t(k_VKQ_0)*stride_V, tile_V, nbatch_V2, stride_V, k_VKQ_sup);
+        flash_attn_ext_f16_load_tile<stride_tile_V, swz_V, nwarps, nbatch_fa, use_cp_async, oob_check, use_sparse>
+            (V_h2, tile_V, nbatch_V2, stride_V, k_VKQ_0, k_VKQ_sup, nullptr);
    } else {
-        constexpr bool use_cp_async = nstages == 1;
+        // the sparse mask values are gathered per element, always load them synchronously
+        constexpr bool use_cp_async = nstages == 1 && !use_sparse;
        if (ncols2 > 1 || mask_h) {
-            flash_attn_ext_f16_load_mask<ncols1, nwarps, nbatch_fa, use_cp_async, oob_check>
-                (mask_h + k_VKQ_0, tile_mask, stride_mask, k_VKQ_sup, jt*ncols1, ne01);
+            flash_attn_ext_f16_load_mask<ncols1, nwarps, nbatch_fa, use_cp_async, oob_check, use_sparse>
+                (mask_h, tile_mask, stride_mask, k_VKQ_0, k_VKQ_sup, jt*ncols1, ne01, indices);
        }
    }

@@ -620,8 +647,8 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
        if constexpr (nstages <= 1) {
            const int k0_diff = k0_stop - k0_start;
            constexpr bool use_cp_async = nstages == 1;
-            flash_attn_ext_f16_load_tile<stride_tile_K, swz_K, nwarps, nbatch_fa, use_cp_async, oob_check>
-                (K_h2 + int64_t(k_VKQ_0)*stride_K + k0_start, tile_K, k0_diff, stride_K, k_VKQ_sup);
+            flash_attn_ext_f16_load_tile<stride_tile_K, swz_K, nwarps, nbatch_fa, use_cp_async, oob_check, use_sparse>
+                (K_h2 + k0_start, tile_K, k0_diff, stride_K, k_VKQ_0, k_VKQ_sup, indices);
            if (use_cp_async) {
                cp_async_wait_all();
            }
@@ -946,6 +973,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
    }

    if constexpr (nstages > 1) {
+        static_assert(!use_sparse, "sparse gather not implemented for multi-stage loading");
        static_assert(!V_is_K_view, "K data reuse not implemented multi-stage loading");
        // Preload K tile for next iteration:
        constexpr bool use_cp_async = true;
@@ -953,11 +981,11 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
        __syncthreads();
        if (!last_iter) {
            if (ncols2 > 1 || mask_h) {
-                flash_attn_ext_f16_load_mask<ncols1, nwarps, nbatch_fa, use_cp_async, oob_check>
-                    (mask_h + k_VKQ_0 + nbatch_fa, tile_mask, stride_mask, k_VKQ_sup, jt*ncols1, ne01);
+                flash_attn_ext_f16_load_mask<ncols1, nwarps, nbatch_fa, use_cp_async, oob_check, use_sparse>
+                    (mask_h, tile_mask, stride_mask, k_VKQ_0 + nbatch_fa, k_VKQ_sup, jt*ncols1, ne01, nullptr);
            }
-            flash_attn_ext_f16_load_tile<stride_tile_K, swz_K, nwarps, nbatch_fa, use_cp_async, oob_check>
-                (K_h2 + int64_t(k_VKQ_0 + nbatch_fa)*stride_K, tile_K, nbatch_K2, stride_K, k_VKQ_sup);
+            flash_attn_ext_f16_load_tile<stride_tile_K, swz_K, nwarps, nbatch_fa, use_cp_async, oob_check, use_sparse>
+                (K_h2, tile_K, nbatch_K2, stride_K, k_VKQ_0 + nbatch_fa, k_VKQ_sup, nullptr);
        }
    }

@@ -972,8 +1000,8 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
            const int i0_diff = i0_stop - i0_start;
            if (!V_is_K_view || i0_stop > 2*nbatch_K2) {
                constexpr bool use_cp_async = nstages == 1;
-                flash_attn_ext_f16_load_tile<stride_tile_V, swz_V, nwarps, nbatch_fa, use_cp_async, oob_check>
-                    (V_h2 + int64_t(k_VKQ_0)*stride_V + i0_start/2, tile_V, i0_diff/2, stride_V, k_VKQ_sup);
+                flash_attn_ext_f16_load_tile<stride_tile_V, swz_V, nwarps, nbatch_fa, use_cp_async, oob_check, use_sparse>
+                    (V_h2 + i0_start/2, tile_V, i0_diff/2, stride_V, k_VKQ_0, k_VKQ_sup, indices);
                if (use_cp_async) {
                    cp_async_wait_all();
                }
@@ -1028,7 +1056,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_iter(
        }
    }
 #else
-    GGML_UNUSED_VARS(Q_f2, K_h2, V_h2, mask_h, dstk, dstk_fixup,
+    GGML_UNUSED_VARS(Q_f2, K_h2, V_h2, mask_h, indices, dstk, dstk_fixup,
        scale, slope, logit_softcap, ne01, ne02,
        stride_K, stride_V, stride_mask,
        tile_Q, tile_K, tile_V, tile_mask,
@@ -1126,12 +1154,13 @@ template<int DV, int ncols> struct mma_tile_sizes {
 };
 #endif // defined(TURING_MMA_AVAILABLE)

-template<int DKQ, int DV, int ncols1, int ncols2, int nwarps, bool use_logit_softcap, bool V_is_K_view, bool needs_fixup, bool is_fixup>
+template<int DKQ, int DV, int ncols1, int ncols2, int nwarps, bool use_logit_softcap, bool V_is_K_view, bool use_sparse, bool needs_fixup, bool is_fixup>
 static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
        const float2 * const __restrict__ Q_f2,
        const half2  * const __restrict__ K_h2,
        const half2  * const __restrict__ V_h2,
        const half   * const __restrict__ mask_h,
+        const int32_t * const __restrict__ indices,
        const float  * const __restrict__ sinks_f,
        float2       * const __restrict__ dstk,
        float2       * const __restrict__ dstk_fixup,
@@ -1171,7 +1200,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
    constexpr int  nbatch_V2       = ggml_cuda_fattn_mma_get_nbatch_V2     (DKQ, DV, ncols);
    constexpr int  nbatch_combine  = ggml_cuda_fattn_mma_get_nbatch_combine(DKQ, DV, ncols);
    constexpr bool Q_in_reg        = ggml_cuda_fattn_mma_get_Q_in_reg      (DKQ, DV, ncols);
-    constexpr int  nstages         = ggml_cuda_fattn_mma_get_nstages       (DKQ, DV, ncols1, ncols2);
+    constexpr int  nstages         = ggml_cuda_fattn_mma_get_nstages       (DKQ, DV, ncols1, ncols2, use_sparse);

    if (cols_per_warp > ncols) {
        NO_DEVICE_CODE;
@@ -1272,37 +1301,38 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(

    // Preload mask and K data for first iteration when using cp_async with multiple stages:
    if constexpr (nstages > 1) {
+        static_assert(!use_sparse, "sparse gather not implemented for multi-stage loading");
        static_assert(nbatch_K2 == DKQ/2, "batching not implemented for multi-stage pipeline");
        constexpr bool use_cp_async = true;
        constexpr bool oob_check    = false;
        constexpr int  k_VKQ_sup    = nbatch_fa;
        if (ncols2 > 1 || mask_h) {
-            flash_attn_ext_f16_load_mask<ncols1, nwarps, nbatch_fa, use_cp_async, oob_check>
-                (mask_h + kb0*nbatch_fa, tile_mask, stride_mask, k_VKQ_sup, jt*ncols1, ne01);
+            flash_attn_ext_f16_load_mask<ncols1, nwarps, nbatch_fa, use_cp_async, oob_check, use_sparse>
+                (mask_h, tile_mask, stride_mask, kb0*nbatch_fa, k_VKQ_sup, jt*ncols1, ne01, nullptr);
        }
-        flash_attn_ext_f16_load_tile<stride_tile_K, swz_K, nwarps, nbatch_fa, use_cp_async, oob_check>
-            (K_h2 + int64_t(kb0)*nbatch_fa*stride_K, tile_K, nbatch_K2, stride_K, k_VKQ_sup);
+        flash_attn_ext_f16_load_tile<stride_tile_K, swz_K, nwarps, nbatch_fa, use_cp_async, oob_check, use_sparse>
+            (K_h2, tile_K, nbatch_K2, stride_K, kb0*nbatch_fa, k_VKQ_sup, nullptr);
    }

    // kb0_start is always < kb0_stop so the last iter can be executed unconditionally.
-    if constexpr (ncols2 == 1) {
+    if constexpr (ncols2 == 1 || use_sparse) {
        constexpr bool oob_check = true;
        for (; kb0 < kb0_stop-1; ++kb0) {
            constexpr bool last_iter = false;
            constexpr int  k_VKQ_sup = nbatch_fa;
            flash_attn_ext_f16_iter
-                <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, needs_fixup, is_fixup, last_iter, oob_check,
+                <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, use_sparse, needs_fixup, is_fixup, last_iter, oob_check,
                 T_A_KQ, T_B_KQ, T_C_KQ, T_A_VKQ, T_B_VKQ, T_C_VKQ>
-                (Q_f2, K_h2, V_h2, mask_h, dstk, dstk_fixup, scale, slope, logit_softcap,
+                (Q_f2, K_h2, V_h2, mask_h, indices, dstk, dstk_fixup, scale, slope, logit_softcap,
                 ne01, ne02, stride_K, stride_V, stride_mask, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C,
                 KQ_max, KQ_rowsum, jt, kb0, k_VKQ_sup);
        }
        constexpr bool last_iter = true;
        const     int  k_VKQ_sup = ne11 - kb0*nbatch_fa;
        flash_attn_ext_f16_iter
-            <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, needs_fixup, is_fixup, last_iter, oob_check,
+            <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, use_sparse, needs_fixup, is_fixup, last_iter, oob_check,
              T_A_KQ, T_B_KQ, T_C_KQ, T_A_VKQ, T_B_VKQ, T_C_VKQ>
-            (Q_f2, K_h2, V_h2, mask_h, dstk, dstk_fixup, scale, slope, logit_softcap,
+            (Q_f2, K_h2, V_h2, mask_h, indices, dstk, dstk_fixup, scale, slope, logit_softcap,
             ne01, ne02, stride_K, stride_V, stride_mask, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C,
             KQ_max, KQ_rowsum, jt, kb0, k_VKQ_sup);
    } else {
@@ -1311,18 +1341,18 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
            constexpr bool last_iter = false;
            constexpr int  k_VKQ_sup = nbatch_fa;
            flash_attn_ext_f16_iter
-                <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, needs_fixup, is_fixup, last_iter, oob_check,
+                <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, use_sparse, needs_fixup, is_fixup, last_iter, oob_check,
                 T_A_KQ, T_B_KQ, T_C_KQ, T_A_VKQ, T_B_VKQ, T_C_VKQ>
-                (Q_f2, K_h2, V_h2, mask_h, dstk, dstk_fixup, scale, slope, logit_softcap,
+                (Q_f2, K_h2, V_h2, mask_h, indices, dstk, dstk_fixup, scale, slope, logit_softcap,
                 ne01, ne02, stride_K, stride_V, stride_mask, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C,
                 KQ_max, KQ_rowsum, jt, kb0, k_VKQ_sup);
        }
        constexpr bool last_iter = true;
        constexpr int  k_VKQ_sup = nbatch_fa;
        flash_attn_ext_f16_iter
-            <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, needs_fixup, is_fixup, last_iter, oob_check,
+            <DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, use_sparse, needs_fixup, is_fixup, last_iter, oob_check,
             T_A_KQ, T_B_KQ, T_C_KQ, T_A_VKQ, T_B_VKQ, T_C_VKQ>
-            (Q_f2, K_h2, V_h2, mask_h, dstk, dstk_fixup, scale, slope, logit_softcap,
+            (Q_f2, K_h2, V_h2, mask_h, indices, dstk, dstk_fixup, scale, slope, logit_softcap,
             ne01, ne02, stride_K, stride_V, stride_mask, tile_Q, tile_K, tile_V, tile_mask, Q_B, VKQ_C,
             KQ_max, KQ_rowsum, jt, kb0, k_VKQ_sup);
    }
@@ -1717,7 +1747,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
        }
    }
 #else
-    GGML_UNUSED_VARS(Q_f2, K_h2, V_h2, mask_h, sinks_f, dstk, dstk_fixup,
+    GGML_UNUSED_VARS(Q_f2, K_h2, V_h2, mask_h, indices, sinks_f, dstk, dstk_fixup,
        scale, slope, logit_softcap, ne01, ne02, gqa_ratio,
        stride_Q1, stride_Q2, stride_K, stride_V, stride_mask,
        jt, kb0_start, kb0_stop);
@@ -1725,7 +1755,13 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
 #endif // defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE)
 }

-template<int DKQ, int DV, int ncols1, int ncols2, bool use_logit_softcap, bool V_is_K_view>
+static constexpr __host__ __device__ bool ggml_cuda_flash_attn_ext_mma_f16_may_use_sparse(
+        const int DKQ, const int DV, const int ncols1, const int ncols2) {
+    return (DKQ == 512 && DV == 512 && ncols1 == 1 && ncols2 == 8) ||
+           (DKQ == 576 && DV == 512 && ncols1 == 1 && ncols2 == 16);
+}
+
+template<int DKQ, int DV, int ncols1, int ncols2, bool use_logit_softcap, bool V_is_K_view, bool use_sparse>
 __launch_bounds__(ggml_cuda_fattn_mma_get_nthreads(DKQ, DV, ncols1*ncols2), ggml_cuda_fattn_mma_get_occupancy(DKQ, DV, ncols1*ncols2))
 static __global__ void flash_attn_ext_f16(
        const char * Q_ptr,
@@ -1751,14 +1787,15 @@ static __global__ void flash_attn_ext_f16(
                            const int32_t nb31, const int32_t nb32, const int64_t nb33) {
    ggml_cuda_pdl_sync(); // TODO optimize placement
 #if defined(FLASH_ATTN_AVAILABLE) && (defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE))
-    const char * GGML_CUDA_RESTRICT Q        = Q_ptr;
-    const char * GGML_CUDA_RESTRICT K        = K_ptr;
-    const char * GGML_CUDA_RESTRICT V        = V_ptr;
-    const char * GGML_CUDA_RESTRICT mask     = mask_ptr;
-    const char * GGML_CUDA_RESTRICT sinks    = sinks_ptr;
-    const int  * GGML_CUDA_RESTRICT KV_max   = KV_max_ptr;
-    float      * GGML_CUDA_RESTRICT dst      = dst_ptr;
-    float2     * GGML_CUDA_RESTRICT dst_meta = dst_meta_ptr;
+    const char * GGML_CUDA_RESTRICT Q              = Q_ptr;
+    const char * GGML_CUDA_RESTRICT K              = K_ptr;
+    const char * GGML_CUDA_RESTRICT V              = V_ptr;
+    const char * GGML_CUDA_RESTRICT mask           = mask_ptr;
+    const char * GGML_CUDA_RESTRICT sinks          = sinks_ptr;
+    const int  * GGML_CUDA_RESTRICT KV_max         = use_sparse ? nullptr : KV_max_ptr;
+    const int  * GGML_CUDA_RESTRICT sparse_indices = use_sparse ? KV_max_ptr : nullptr;
+    float      * GGML_CUDA_RESTRICT dst            = dst_ptr;
+    float2     * GGML_CUDA_RESTRICT dst_meta       = dst_meta_ptr;

    // Skip unused kernel variants for faster compilation:
    if (use_logit_softcap && !(DKQ == 128 || DKQ == 256 || DKQ == 512)) {
@@ -1769,6 +1806,11 @@ static __global__ void flash_attn_ext_f16(
        NO_DEVICE_CODE;
        return;
    }
+
+    if (!ggml_cuda_flash_attn_ext_mma_f16_may_use_sparse(DKQ, DV, ncols1, ncols2) && use_sparse) {
+        NO_DEVICE_CODE;
+        return;
+    }
 #ifdef VOLTA_MMA_AVAILABLE
    if (ncols1*ncols2 < 32) {
        NO_DEVICE_CODE;
@@ -1845,6 +1887,7 @@ static __global__ void flash_attn_ext_f16(

        const half2 * V_h2 = V_is_K_view ? K_h2 : (const half2 *) (V + nb23*sequence + nb22*z_KV);
        const float * sinks_f = sinks ? (const float *) sinks + zt_Q : nullptr;
+        const int32_t * indices = use_sparse ? sparse_indices + (int64_t(sequence % ne33)*ne31 + jt*ncols1)*ne11 : nullptr;

        const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, zt_Q, n_head_log2, m0, m1) : 1.0f;

@@ -1854,13 +1897,13 @@ static __global__ void flash_attn_ext_f16(
        constexpr bool is_fixup = false; // All but (potentially) the last iterations write their data to dst rather than the fixup buffer.
        if (kb0_start == 0) {
            constexpr bool needs_fixup = false; // CUDA block is working on an entire tile.
-            flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, needs_fixup, is_fixup>
-                (Q_f2, K_h2, V_h2, mask_h, sinks_f, dstk, dst_meta, scale, slope, logit_softcap,
+            flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, use_sparse, needs_fixup, is_fixup>
+                (Q_f2, K_h2, V_h2, mask_h, indices, sinks_f, dstk, dst_meta, scale, slope, logit_softcap,
                 ne01, ne02, gqa_ratio, ne11, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, zt_gqa, kb0_start, kb0_stop);
        } else {
            constexpr bool needs_fixup = true; // CUDA block is missing the beginning of a tile.
-            flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, needs_fixup, is_fixup>
-                (Q_f2, K_h2, V_h2, mask_h, sinks_f, dstk, dst_meta, scale, slope, logit_softcap,
+            flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, use_sparse, needs_fixup, is_fixup>
+                (Q_f2, K_h2, V_h2, mask_h, indices, sinks_f, dstk, dst_meta, scale, slope, logit_softcap,
                 ne01, ne02, gqa_ratio, ne11, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, zt_gqa, kb0_start, kb0_stop);
        }

@@ -1891,6 +1934,7 @@ static __global__ void flash_attn_ext_f16(

    const half2 * V_h2 = V_is_K_view ? K_h2 : (const half2 *) (V + nb23*sequence + nb22*z_KV);
    const float * sinks_f = sinks ? (const float *) sinks + zt_Q : nullptr;
+    const int32_t * indices = use_sparse ? sparse_indices + (int64_t(sequence % ne33)*ne31 + jt*ncols1)*ne11 : nullptr;

    const float slope = ncols2 == 1 ? get_alibi_slope(max_bias, zt_Q, n_head_log2, m0, m1) : 1.0f;

@@ -1900,8 +1944,8 @@ static __global__ void flash_attn_ext_f16(

    constexpr bool is_fixup = true; // Last index writes its data to fixup buffer to avoid data races with other blocks.
    constexpr bool needs_fixup = false;
-    flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, needs_fixup, is_fixup>
-        (Q_f2, K_h2, V_h2, mask_h, sinks_f, dstk, dst_meta, scale, slope, logit_softcap,
+    flash_attn_ext_f16_process_tile<DKQ, DV, ncols1, ncols2, nwarps, use_logit_softcap, V_is_K_view, use_sparse, needs_fixup, is_fixup>
+        (Q_f2, K_h2, V_h2, mask_h, indices, sinks_f, dstk, dst_meta, scale, slope, logit_softcap,
         ne01, ne02, gqa_ratio, ne11, stride_Q1, stride_Q2, stride_K, stride_V, stride_mask, jt, zt_gqa, kb0_start, kb0_stop);
 #else
    GGML_UNUSED_VARS(Q_ptr, K_ptr, V_ptr, mask_ptr, sinks_ptr, KV_max_ptr, dst_ptr, dst_meta_ptr, scale,
@@ -1917,6 +1961,8 @@ static __global__ void flash_attn_ext_f16(
 #endif // defined(FLASH_ATTN_AVAILABLE) && (defined(VOLTA_MMA_AVAILABLE) || defined(TURING_MMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) || defined(AMD_MFMA_AVAILABLE))
 }

+bool ggml_cuda_flash_attn_ext_mma_f16_shall_use_sparse(ggml_backend_cuda_context & ctx, ggml_tensor * dst);
+
 template <int DKQ, int DV, int ncols1, int ncols2>
 void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    const ggml_tensor * KQV = dst;
@@ -1963,20 +2009,49 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
    using fattn_kernel_ptr_t = fattn_kernel_t;
 #endif // defined(GGML_USE_HIP)
    fattn_kernel_t fattn_kernel;
+    bool use_sparse = false;
    if (logit_softcap == 0.0f) {
        constexpr bool use_logit_softcap = false;
-        fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, use_logit_softcap, V_is_K_view>;
+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+        if constexpr (ggml_cuda_flash_attn_ext_mma_f16_may_use_sparse(DKQ, DV, ncols1, ncols2)) {
+            if (ggml_cuda_flash_attn_ext_mma_f16_shall_use_sparse(ctx, dst)) {
+                constexpr bool use_sparse_kernel = true;
+                fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, use_logit_softcap, V_is_K_view, use_sparse_kernel>;
+                use_sparse = true;
+
+                static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
+                if (!shared_memory_limit_raised[id]) {
+                    CUDA_CHECK(cudaFuncSetAttribute(reinterpret_cast<fattn_kernel_ptr_t>(fattn_kernel), cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared_total));
+                    shared_memory_limit_raised[id] = true;
+                }
+            } else {
+                constexpr bool use_sparse_kernel = false;
+                fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, use_logit_softcap, V_is_K_view, use_sparse_kernel>;
+
+                static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
+                if (!shared_memory_limit_raised[id]) {
+                    CUDA_CHECK(cudaFuncSetAttribute(reinterpret_cast<fattn_kernel_ptr_t>(fattn_kernel), cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared_total));
+                    shared_memory_limit_raised[id] = true;
+                }
+            }
+        } else
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+        {
+            constexpr bool use_sparse_kernel = false;
+            fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, use_logit_softcap, V_is_K_view, use_sparse_kernel>;

 #if !defined(GGML_USE_MUSA)
-        static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
-        if (!shared_memory_limit_raised[id]) {
-            CUDA_CHECK(cudaFuncSetAttribute(reinterpret_cast<fattn_kernel_ptr_t>(fattn_kernel), cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared_total));
-            shared_memory_limit_raised[id] = true;
-        }
+            static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
+            if (!shared_memory_limit_raised[id]) {
+                CUDA_CHECK(cudaFuncSetAttribute(reinterpret_cast<fattn_kernel_ptr_t>(fattn_kernel), cudaFuncAttributeMaxDynamicSharedMemorySize, nbytes_shared_total));
+                shared_memory_limit_raised[id] = true;
+            }
 #endif // !defined(GGML_USE_MUSA)
+        }
    } else {
        constexpr bool use_logit_softcap = true;
-        fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, use_logit_softcap, V_is_K_view>;
+        constexpr bool use_sparse_kernel = false;
+        fattn_kernel = flash_attn_ext_f16<DKQ, DV, ncols1, ncols2, use_logit_softcap, V_is_K_view, use_sparse_kernel>;

 #if !defined(GGML_USE_MUSA)
        static bool shared_memory_limit_raised[GGML_CUDA_MAX_DEVICES] = {false};
@@ -1988,7 +2063,7 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
    }

    launch_fattn<DV, ncols1, ncols2>
-        (ctx, dst, fattn_kernel, nwarps, nbytes_shared_total, nbatch_fa, true, true, true, warp_size_host);
+        (ctx, dst, fattn_kernel, nwarps, nbytes_shared_total, nbatch_fa, true, true, true, use_sparse, warp_size_host);
 }


@@ -1163,7 +1163,7 @@ static void launch_fattn_tile_switch_ncols1(ggml_backend_cuda_context & ctx, ggm
            const int nbatch_fa = ggml_cuda_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
            fattn_kernel_t fattn_kernel = flash_attn_tile<DKQ, DV, cols_per_block/ncols2, ncols2, use_logit_softcap>;
            launch_fattn<DV, cols_per_block/ncols2, ncols2>
-                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, nbatch_fa, true, true, false, warp_size);
+                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, nbatch_fa, true, true, false, false, warp_size);
            return;
        }
    }
@@ -1179,7 +1179,7 @@ static void launch_fattn_tile_switch_ncols1(ggml_backend_cuda_context & ctx, ggm
            const int nbatch_fa = ggml_cuda_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
            fattn_kernel_t fattn_kernel = flash_attn_tile<DKQ, DV, cols_per_block/ncols2, ncols2, use_logit_softcap>;
            launch_fattn<DV, cols_per_block/ncols2, ncols2>
-                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, nbatch_fa, true, true, false, warp_size);
+                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, nbatch_fa, true, true, false, false, warp_size);
            return;
        }
    }
@@ -1191,7 +1191,7 @@ static void launch_fattn_tile_switch_ncols1(ggml_backend_cuda_context & ctx, ggm
            const int nbatch_fa = ggml_cuda_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
            fattn_kernel_t fattn_kernel = flash_attn_tile<DKQ, DV, cols_per_block/ncols2, ncols2, use_logit_softcap>;
            launch_fattn<DV, cols_per_block/ncols2, ncols2>
-                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, nbatch_fa, true, true, false, warp_size);
+                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, nbatch_fa, true, true, false, false, warp_size);
            return;
        }
    }
@@ -1203,7 +1203,7 @@ static void launch_fattn_tile_switch_ncols1(ggml_backend_cuda_context & ctx, ggm
            const int nbatch_fa = ggml_cuda_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
            fattn_kernel_t fattn_kernel = flash_attn_tile<DKQ, DV, cols_per_block/ncols2, ncols2, use_logit_softcap>;
            launch_fattn<DV, cols_per_block/ncols2, ncols2>
-                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, nbatch_fa, true, true, false, warp_size);
+                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, nbatch_fa, true, true, false, false, warp_size);
            return;
        }
    }
@@ -1215,7 +1215,7 @@ static void launch_fattn_tile_switch_ncols1(ggml_backend_cuda_context & ctx, ggm
            const int nbatch_fa = ggml_cuda_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
            fattn_kernel_t fattn_kernel = flash_attn_tile<DKQ, DV, cols_per_block/ncols2, ncols2, use_logit_softcap>;
            launch_fattn<DV, cols_per_block/ncols2, ncols2>
-                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, nbatch_fa, true, true, false, warp_size);
+                (ctx, dst, fattn_kernel, nwarps, nbytes_shared, nbatch_fa, true, true, false, false, warp_size);
            return;
        }
    }
@@ -1226,7 +1226,7 @@ static void launch_fattn_tile_switch_ncols1(ggml_backend_cuda_context & ctx, ggm
        const int nbatch_fa = ggml_cuda_fattn_tile_get_nbatch_fa(DKQ, DV, cols_per_block, cc);
        fattn_kernel_t fattn_kernel = flash_attn_tile<DKQ, DV, cols_per_block/ncols2, ncols2, use_logit_softcap>;
        launch_fattn<DV, cols_per_block/ncols2, ncols2>
-            (ctx, dst, fattn_kernel, nwarps, nbytes_shared, nbatch_fa, true, true, false, warp_size);
+            (ctx, dst, fattn_kernel, nwarps, nbytes_shared, nbatch_fa, true, true, false, false, warp_size);
        return;
    }

@@ -540,7 +540,7 @@ void ggml_cuda_flash_attn_ext_vec_case_impl(ggml_backend_cuda_context & ctx, ggm
    const bool need_f16_K = type_K == GGML_TYPE_F16;
    const bool need_f16_V = type_V == GGML_TYPE_F16;
    constexpr size_t nbytes_shared = 0;
-    launch_fattn<D, cols_per_block, 1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false);
+    launch_fattn<D, cols_per_block, 1>(ctx, dst, fattn_kernel, nwarps, nbytes_shared, D, need_f16_K, need_f16_V, false, false);
 }

 template <int D, ggml_type type_K, ggml_type type_V>
@@ -5,11 +5,144 @@
 #include "fattn-vec.cuh"
 #include "fattn.cuh"

+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+__launch_bounds__(256, 1)
+static __global__ void flash_attn_mask_to_sparse_indices(
+        const half * mask_ptr, int32_t * indices_ptr, const int ne30, const int n_kv_max,
+        const int64_t s31, const int64_t s33) {
+    ggml_cuda_pdl_sync();
+
+    constexpr int values_per_lane = 8;
+    const int tid      = threadIdx.x;
+    const int warp     = tid / WARP_SIZE;
+    const int lane     = tid % WARP_SIZE;
+    const int sequence = blockIdx.y;
+    const int query    = blockIdx.x;
+
+    const half * mask = mask_ptr + sequence*s33 + query*s31;
+    int32_t * indices = indices_ptr + (int64_t(sequence)*gridDim.x + query)*n_kv_max;
+
+    __shared__ int warp_offsets[256/WARP_SIZE];
+    __shared__ int row_count;
+    __shared__ int chunk_count;
+
+    if (tid == 0) {
+        row_count = 0;
+    }
+    __syncthreads();
+
+    for (int i0 = 0; i0 < ne30; i0 += blockDim.x*values_per_lane) {
+        uint32_t selected_warp[values_per_lane];
+        int warp_count = 0;
+#pragma unroll
+        for (int item = 0; item < values_per_lane; ++item) {
+            const int i = i0 + (warp*values_per_lane + item)*WARP_SIZE + lane;
+            const bool selected = i < ne30 && isfinite(__half2float(mask[i]));
+            selected_warp[item] = __ballot_sync(0xFFFFFFFF, selected);
+            warp_count += __popc(selected_warp[item]);
+        }
+
+        if (lane == 0) {
+            warp_offsets[warp] = warp_count;
+        }
+        __syncthreads();
+
+        if (tid == 0) {
+            int offset = 0;
+#pragma unroll
+            for (int iw = 0; iw < 256/WARP_SIZE; ++iw) {
+                const int count = warp_offsets[iw];
+                warp_offsets[iw] = offset;
+                offset += count;
+            }
+            chunk_count = offset;
+        }
+        __syncthreads();
+
+        const uint32_t lane_mask = lane == 0 ? 0 : (1u << lane) - 1;
+        int warp_item_offset = 0;
+#pragma unroll
+        for (int item = 0; item < values_per_lane; ++item) {
+            const int i = i0 + (warp*values_per_lane + item)*WARP_SIZE + lane;
+            const int dst = row_count + warp_offsets[warp] + warp_item_offset + __popc(selected_warp[item] & lane_mask);
+            if ((selected_warp[item] & (uint32_t(1) << lane)) && dst < n_kv_max) {
+                indices[dst] = i;
+            }
+            warp_item_offset += __popc(selected_warp[item]);
+        }
+        __syncthreads();
+
+        if (tid == 0) {
+            row_count += chunk_count;
+        }
+        __syncthreads();
+    }
+
+    const int count = row_count;
+    for (int i = count + tid; i < n_kv_max; i += blockDim.x) {
+        indices[i] = -1;
+    }
+    __syncthreads();
+
+    // the dependent grid reads indices, signal once the row is complete
+    ggml_cuda_pdl_lc();
+}
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+
+void ggml_cuda_flash_attn_ext_compact_mask(
+        const ggml_tensor * mask, int32_t * indices, int32_t n_kv_max, cudaStream_t stream) {
+#if defined(GGML_USE_HIP) || defined(GGML_USE_MUSA)
+    GGML_UNUSED_VARS(mask, indices, n_kv_max, stream);
+    GGML_ABORT("sparse flash attention is only supported on NVIDIA CUDA");
+#else
+    const int64_t s31 = mask->nb[1] / sizeof(half);
+    const int64_t s33 = mask->nb[3] / sizeof(half);
+    const dim3 blocks_num(mask->ne[1], mask->ne[3], 1);
+    const dim3 block_dim(256, 1, 1);
+    const ggml_cuda_kernel_launch_params launch_params(blocks_num, block_dim, 0, stream);
+    ggml_cuda_kernel_launch(flash_attn_mask_to_sparse_indices, launch_params,
+        (const half *) mask->data, indices, int(mask->ne[0]), n_kv_max, s31, s33);
+    CUDA_CHECK(cudaGetLastError());
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+}
+
+bool ggml_cuda_flash_attn_ext_mma_f16_shall_use_sparse(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
+#if defined(GGML_USE_HIP) || defined(GGML_USE_MUSA)
+    GGML_UNUSED_VARS(ctx, dst);
+    return false;
+#else
+    const ggml_tensor * Q    = dst->src[0];
+    const ggml_tensor * K    = dst->src[1];
+    const ggml_tensor * mask = dst->src[3];
+    const int cc = ggml_cuda_info().devices[ctx.device].cc;
+
+    float max_bias = 0.0f;
+    float logit_softcap = 0.0f;
+    memcpy(&max_bias,      (const float *) dst->op_params + 1, sizeof(float));
+    memcpy(&logit_softcap, (const float *) dst->op_params + 2, sizeof(float));
+
+    const int32_t n_kv_max = ggml_get_op_params_i32(dst, 4);
+    return GGML_CUDA_CC_IS_NVIDIA(cc) && turing_mma_available(cc) &&
+        mask != nullptr && n_kv_max > 0 && max_bias == 0.0f && logit_softcap == 0.0f &&
+        mask->ne[0] == K->ne[1] && mask->ne[1] >= Q->ne[1] && mask->ne[2] == 1 &&
+        K->ne[1] >= std::max<int64_t>(4096, 2LL*n_kv_max);
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+}
+
 template <int DKQ, int DV, int ncols2>
 static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
    const int cc = ggml_cuda_info().devices[ggml_cuda_get_device()].cc;
    const ggml_tensor * Q = dst->src[0];

+#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+    if constexpr (ggml_cuda_flash_attn_ext_mma_f16_may_use_sparse(DKQ, DV, 1, ncols2)) {
+        if (ggml_cuda_flash_attn_ext_mma_f16_shall_use_sparse(ctx, dst)) {
+            ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 1, ncols2>(ctx, dst);
+            return;
+        }
+    }
+#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+
    if constexpr (ncols2 <= 8) {
        if (turing_mma_available(cc) && Q->ne[1] <= 8/ncols2) {
            ggml_cuda_flash_attn_ext_mma_f16_case<DKQ, DV, 8/ncols2, ncols2>(ctx, dst);
@@ -32,6 +32,7 @@
 #include "ggml-cuda/mmq.cuh"
 #include "ggml-cuda/mmvf.cuh"
 #include "ggml-cuda/mmvq.cuh"
+#include "ggml-cuda/moe-weighted-reduction.cuh"
 #include "ggml-cuda/norm.cuh"
 #include "ggml-cuda/opt-step-adamw.cuh"
 #include "ggml-cuda/opt-step-sgd.cuh"
@@ -3026,6 +3027,150 @@ static bool ggml_cuda_check_fusion_memory_ranges(const ggml_cgraph * cgraph,
    return is_ok;
 }

+// The long form spans 2*k + 1 nodes. ggml_can_fuse_subgraph() accepts at most
+// 31 nodes, so k <= 15; larger values use the per-operation path.
+static constexpr int MOE_WEIGHTED_REDUCTION_MAX_EXPERTS = 15;
+
+struct ggml_cuda_moe_weighted_reduction_match {
+    const ggml_tensor * experts      = nullptr;
+    const ggml_tensor * expert_scale = nullptr;
+    const ggml_tensor * weights      = nullptr;
+    ggml_tensor *       dst          = nullptr;
+    int                 node_count   = 0;
+};
+
+static bool ggml_cuda_match_moe_weighted_reduction(
+        const ggml_cgraph * cgraph,
+        int node_idx,
+        ggml_cuda_moe_weighted_reduction_match & match) {
+    const ggml_tensor * first = cgraph->nodes[node_idx];
+    if (first->op != GGML_OP_MUL || first->type != GGML_TYPE_F32 || !ggml_is_contiguous(first)) {
+        return false;
+    }
+
+    auto split_mul = [](const ggml_tensor * mul, const ggml_tensor *& full, const ggml_tensor *& broadcast) {
+        auto is_weights = [mul](const ggml_tensor * tensor) {
+            return tensor && tensor->type == GGML_TYPE_F32 && ggml_is_contiguous(tensor) && tensor->ne[0] == 1 &&
+                tensor->ne[1] == mul->ne[1] && tensor->ne[2] == mul->ne[2] && tensor->ne[3] == mul->ne[3];
+        };
+        auto is_experts = [mul](const ggml_tensor * tensor) {
+            return tensor && tensor->type == GGML_TYPE_F32 && ggml_is_contiguous(tensor) &&
+                ggml_are_same_shape(tensor, mul);
+        };
+
+        if (is_experts(mul->src[0]) && is_weights(mul->src[1])) {
+            full      = mul->src[0];
+            broadcast = mul->src[1];
+            return true;
+        }
+        if (is_experts(mul->src[1]) && is_weights(mul->src[0])) {
+            full      = mul->src[1];
+            broadcast = mul->src[0];
+            return true;
+        }
+        return false;
+    };
+
+    const ggml_tensor * weighted     = first;
+    const ggml_tensor * experts      = nullptr;
+    const ggml_tensor * expert_scale = nullptr;
+    const ggml_tensor * weights      = nullptr;
+    int                 mul_count    = 1;
+
+    // Match both structural forms:
+    //   (experts * expert_scale) * router_weight
+    //   experts * router_weight
+    // The matcher does not depend on the model or quantization type.
+    if (node_idx + 1 < cgraph->n_nodes) {
+        const ggml_tensor * second = cgraph->nodes[node_idx + 1];
+        const ggml_tensor * scaled = nullptr;
+        const ggml_tensor * route  = nullptr;
+        const ggml_tensor * raw    = nullptr;
+        const ggml_tensor * scale  = nullptr;
+        if (second->op == GGML_OP_MUL && second->type == GGML_TYPE_F32 && ggml_is_contiguous(second) &&
+                split_mul(second, scaled, route) && scaled == first && split_mul(first, raw, scale)) {
+            weighted     = second;
+            experts      = raw;
+            expert_scale = scale;
+            weights      = route;
+            mul_count    = 2;
+        }
+    }
+
+    if (experts == nullptr && !split_mul(first, experts, weights)) {
+        return false;
+    }
+
+    const int     n_expert_used = (int) weighted->ne[1];
+    const int64_t n_tokens      = weighted->ne[2] * weighted->ne[3];
+    if (n_expert_used < 2 || n_expert_used > MOE_WEIGHTED_REDUCTION_MAX_EXPERTS || n_tokens <= 0) {
+        return false;
+    }
+
+    const int node_count = 2 * n_expert_used + mul_count - 1;
+    if (node_idx + node_count > cgraph->n_nodes) {
+        return false;
+    }
+
+    std::vector<ggml_op> ops(node_count, GGML_OP_VIEW);
+    ops[0] = GGML_OP_MUL;
+    if (mul_count == 2) {
+        ops[1] = GGML_OP_MUL;
+    }
+    std::vector<const ggml_tensor *> views;
+    views.reserve(n_expert_used);
+    const ggml_tensor * previous = nullptr;
+    int n_adds = 0;
+    for (int offset = mul_count; offset < node_count; ++offset) {
+        const ggml_tensor * candidate = cgraph->nodes[node_idx + offset];
+        ops[offset] = candidate->op;
+
+        if (candidate->op == GGML_OP_VIEW) {
+            const int expert = (int) views.size();
+            if (expert >= n_expert_used || candidate->src[0] != weighted || candidate->view_src != weighted ||
+                    candidate->type != GGML_TYPE_F32 || candidate->ne[0] != weighted->ne[0] ||
+                    candidate->ne[1] != n_tokens || candidate->ne[2] != 1 || candidate->ne[3] != 1 ||
+                    candidate->nb[0] != weighted->nb[0] || candidate->nb[1] != weighted->nb[2] ||
+                    candidate->view_offs != (size_t) expert * weighted->nb[1]) {
+                return false;
+            }
+            views.push_back(candidate);
+            continue;
+        }
+
+        if (candidate->op != GGML_OP_ADD || views.size() < 2 || n_adds + 1 >= (int) views.size()) {
+            return false;
+        }
+        const ggml_tensor * lhs = n_adds == 0 ? views[0] : previous;
+        const ggml_tensor * rhs = views[n_adds + 1];
+        if (candidate->src[0] != lhs || candidate->src[1] != rhs || candidate->type != GGML_TYPE_F32) {
+            return false;
+        }
+        previous = candidate;
+        ++n_adds;
+    }
+
+    if ((int) views.size() != n_expert_used || n_adds != n_expert_used - 1 || previous == nullptr) {
+        return false;
+    }
+    if (!ggml_is_contiguous(previous) || previous->ne[0] != weighted->ne[0] ||
+            previous->ne[1] != n_tokens || previous->ne[2] != 1 || previous->ne[3] != 1) {
+        return false;
+    }
+
+    const int output_idx = node_idx + node_count - 1;
+    if (!ggml_can_fuse_subgraph(cgraph, node_idx, node_count, ops.data(), &output_idx, 1)) {
+        return false;
+    }
+
+    match.experts      = experts;
+    match.expert_scale = expert_scale;
+    match.weights      = weights;
+    match.dst          = cgraph->nodes[output_idx];
+    match.node_count   = node_count;
+    return true;
+}
+

 static bool ggml_cuda_can_fuse(const struct ggml_cgraph *                cgraph,
                               int                                       node_idx,
@@ -3288,6 +3433,18 @@ static int ggml_cuda_try_fuse(ggml_backend_cuda_context * cuda_ctx, ggml_cgraph

    ggml_tensor * node = cgraph->nodes[i];

+    if (node->op == GGML_OP_MUL) {
+        ggml_cuda_moe_weighted_reduction_match match;
+        if (ggml_cuda_match_moe_weighted_reduction(cgraph, i, match)) {
+            const int output_idx = i + match.node_count - 1;
+            if (ggml_cuda_check_fusion_memory_ranges(cgraph, i, match.node_count, &output_idx, 1)) {
+                ggml_cuda_op_moe_weighted_reduction(
+                    *cuda_ctx, match.experts, match.expert_scale, match.weights, match.dst);
+                return match.node_count - 1;
+            }
+        }
+    }
+
    // gated_delta_net -> cpy: scatter recurrent-state snapshots into the cache
    if (node->op == GGML_OP_GATED_DELTA_NET) {
        ggml_cuda_gated_delta_net_fused_cache fused_state_cpy;
@@ -4340,10 +4497,30 @@ static void ggml_backend_cuda_event_wait(ggml_backend_t backend, ggml_backend_ev
 }

 static void ggml_backend_cuda_graph_optimize(ggml_backend_t backend, ggml_cgraph * cgraph, ggml_backend_graph_optimize_params * params) {
-    GGML_UNUSED(params);
-
    ggml_backend_cuda_context * cuda_ctx = (ggml_backend_cuda_context *) backend->context;

+    static const bool disable_fusion = getenv("GGML_CUDA_DISABLE_FUSION") != nullptr && std::atoi(getenv("GGML_CUDA_DISABLE_FUSION"));
+    if (!disable_fusion) {
+        for (int i = 0; i < cgraph->n_nodes; ++i) {
+            if (cgraph->nodes[i]->op != GGML_OP_MUL) {
+                continue;
+            }
+
+            ggml_cuda_moe_weighted_reduction_match match;
+            if (!ggml_cuda_match_moe_weighted_reduction(cgraph, i, match)) {
+                continue;
+            }
+
+            params->add_alloc_dep(params->user_data, const_cast<ggml_tensor *>(match.experts), match.dst);
+            params->add_alloc_dep(params->user_data, const_cast<ggml_tensor *>(match.weights), match.dst);
+            if (match.expert_scale != nullptr) {
+                params->add_alloc_dep(
+                    params->user_data, const_cast<ggml_tensor *>(match.expert_scale), match.dst);
+            }
+            i += match.node_count - 1;
+        }
+    }
+
 #ifdef USE_CUDA_GRAPH
    const void * graph_key = ggml_cuda_graph_get_key(cgraph);
    const bool use_cuda_graph = ggml_cuda_graph_set_enabled(cuda_ctx, graph_key);
@@ -4365,10 +4542,12 @@ static void ggml_backend_cuda_graph_optimize(ggml_backend_t backend, ggml_cgraph
    ggml_cuda_stream_context & stream_context = cuda_ctx->stream_context();
    stream_context.reset();

-    if (!use_cuda_graph || ggml_backend_cuda_get_device_count() != 1) {
+    if (!use_cuda_graph) {
        return;
    }

+    ggml_cuda_set_device(cuda_ctx->device);
+
    // number of out-degrees for a particular node
    std::unordered_map<const ggml_tensor *, int> fan_out;
    // reverse mapping of node to index in the cgraph
@@ -148,7 +148,6 @@ static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma(
    typedef tile<16,  8, int, input_layout>        tile_B;
    typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C;

-    constexpr int I             = ggml_cuda_mmq_get_I(type, J, fallback);
    constexpr int sram_stride   = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
    constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback);
    constexpr int ntx           = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
@@ -204,7 +203,6 @@ static __device__ __forceinline__ void ggml_cuda_mmq_vec_dot_q8_0_q8_1_mma(
    typedef tile< 8, 8, int> tile_B;
    typedef tile<16, 8, int> tile_C;

-    constexpr int I             = ggml_cuda_mmq_get_I(type, J, fallback);
    constexpr int sram_stride   = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
    constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback);
    constexpr int ntx           = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
@@ -320,7 +318,6 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
    typedef tile<16,  8, int, input_layout>        tile_B;
    typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C;

-    constexpr int I             = ggml_cuda_mmq_get_I(type, J, fallback);
    constexpr int sram_stride   = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
    constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback);
    constexpr int ntx           = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
@@ -371,7 +368,6 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
    typedef tile< 8,  8, int> tile_B;
    typedef tile<16,  8, int> tile_C;

-    constexpr int I             = ggml_cuda_mmq_get_I(type, J, fallback);
    constexpr int sram_stride   = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
    constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback);
    constexpr int ntx           = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
@@ -486,7 +482,6 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
    typedef tile<16,  4, int, input_layout>        tile_B;
    typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C;

-    constexpr int I             = ggml_cuda_mmq_get_I(type, J, fallback);
    constexpr int sram_stride   = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
    constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback);
    constexpr int ntx           = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
@@ -537,7 +532,6 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
    typedef tile< 8, 4, int> tile_B;
    typedef tile<16, 8, int> tile_C;

-    constexpr int I             = ggml_cuda_mmq_get_I(type, J, fallback);
    constexpr int sram_stride   = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
    constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback);
    constexpr int ntx           = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
@@ -686,7 +680,6 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
    typedef tile<16,  4, int, input_layout>        tile_B;
    typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C;

-    constexpr int I             = ggml_cuda_mmq_get_I(type, J, fallback);
    constexpr int sram_stride   = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
    constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback);
    constexpr int ntx           = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
@@ -756,7 +749,6 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
    typedef tile< 8, 4, int> tile_B;
    typedef tile<16, 8, int> tile_C;

-    constexpr int I             = ggml_cuda_mmq_get_I(type, J, fallback);
    constexpr int sram_stride   = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
    constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback);
    constexpr int ntx           = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
@@ -1023,7 +1015,6 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
    typedef tile<16,  4, int, input_layout>        tile_B;
    typedef tile<16, 16, int, DATA_LAYOUT_J_MAJOR> tile_C;

-    constexpr int I             = ggml_cuda_mmq_get_I(type, J, fallback);
    constexpr int sram_stride   = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
    constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback);
    constexpr int ntx           = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
@@ -1075,7 +1066,6 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
    typedef tile< 8, 4, int> tile_B;
    typedef tile<16, 8, int> tile_C;

-    constexpr int I             = ggml_cuda_mmq_get_I(type, J, fallback);
    constexpr int sram_stride   = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
    constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback);
    constexpr int ntx           = rows_per_warp/tile_C::I; // Number of x minitiles per warp.
@@ -1190,7 +1180,6 @@ template <ggml_type type, int J, bool fallback> static __device__ __forceinline_
    typedef tile<8,  8, int>   tile_B;
    typedef tile<16, 8, float> tile_C;

-    constexpr int I             = ggml_cuda_mmq_get_I(type, J, fallback);
    constexpr int sram_stride   = ggml_cuda_mmq_get_sram_stride(type, J, fallback);
    constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback);
    constexpr int ntx           = rows_per_warp / tile_C::I;
@@ -481,9 +481,6 @@ static __device__ __forceinline__ void ggml_cuda_mmq_write_back_mma(
    typedef tile<16,  8, int> tile_C;
 #endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE)

-    constexpr int warp_size     = ggml_cuda_get_physical_warp_size();
-    constexpr int nwarps        = ggml_cuda_mmq_get_nthreads(type, J, fallback) / warp_size;
-    constexpr int I             = ggml_cuda_mmq_get_I(type, J, fallback);
    constexpr int rows_per_warp = ggml_cuda_mmq_get_rows_per_warp(type, J, fallback);
    constexpr int ntx           = rows_per_warp/tile_C::I; // Number of x minitiles per warp.

@@ -540,8 +537,6 @@ struct ggml_cuda_mmq_util_funcs {

 template <ggml_type type, int J, bool fallback>
 static constexpr __device__ ggml_cuda_mmq_util_funcs ggml_cuda_mmq_get_util_funcs() {
-    constexpr int I = ggml_cuda_mmq_get_I(type, J, fallback);
-
    if (!ggml_cuda_mmq_get_config(type, J, fallback).use_mma_data_layout()) {
        switch (type) {
            case GGML_TYPE_Q1_0:
@@ -326,6 +326,18 @@ bool ggml_cuda_should_use_mmvq(enum ggml_type type, int cc, int64_t ne11) {
                return ne11 <= MMVQ_MAX_BATCH_SIZE;
        }
    }
+    if (GGML_CUDA_CC_IS_NVIDIA(cc) && cc == GGML_CUDA_CC_ORIN) {
+        switch (type) { // tuned for Jetson Orin
+            case GGML_TYPE_Q2_K:
+            case GGML_TYPE_Q3_K:
+            case GGML_TYPE_Q4_K:
+            case GGML_TYPE_Q5_K:
+            case GGML_TYPE_Q6_K:
+                return ne11 <= 1;
+            default:
+                return ne11 <= MMVQ_MAX_BATCH_SIZE;
+        }
+    }
    if (GGML_CUDA_CC_IS_CDNA(cc)) {
        if (GGML_CUDA_CC_IS_CDNA1(cc)) {
            switch (type) {
@@ -0,0 +1,65 @@
+#include "moe-weighted-reduction.cuh"
+
+static __global__ void moe_weighted_reduction_f32(const float * __restrict__ experts,
+                                                  const float * __restrict__ expert_scale,
+                                                  const float * __restrict__ weights,
+                                                  float * __restrict__ dst,
+                                                  const int64_t n_embd,
+                                                  const int     n_expert_used) {
+    const int64_t token = blockIdx.x;
+    const int64_t col   = (int64_t) blockIdx.y * blockDim.x + threadIdx.x;
+    if (col >= n_embd) {
+        return;
+    }
+
+    const uint64_t first_row   = (uint64_t) token * n_expert_used;
+    const float    first_scale = expert_scale != nullptr ? expert_scale[first_row] : 1.0f;
+    float          sum         = (experts[first_row * n_embd + col] * first_scale) * weights[first_row];
+
+    for (int expert = 1; expert < n_expert_used; ++expert) {
+        const uint64_t row   = first_row + expert;
+        const float   scale = expert_scale != nullptr ? expert_scale[row] : 1.0f;
+        sum += (experts[row * n_embd + col] * scale) * weights[row];
+    }
+    dst[token * n_embd + col] = sum;
+}
+
+static void launch_moe_weighted_reduction(const float * experts,
+                                          const float * expert_scale,
+                                          const float * weights,
+                                          float *       dst,
+                                          int64_t       n_embd,
+                                          int64_t       n_tokens,
+                                          int           n_expert_used,
+                                          cudaStream_t  stream) {
+    constexpr int threads = 256;
+    const dim3 blocks(n_tokens, (n_embd + threads - 1) / threads, 1);
+    moe_weighted_reduction_f32
+        <<<blocks, threads, 0, stream>>>(experts, expert_scale, weights, dst, n_embd, n_expert_used);
+}
+
+void ggml_cuda_op_moe_weighted_reduction(ggml_backend_cuda_context & ctx,
+                                         const ggml_tensor *         experts,
+                                         const ggml_tensor *         expert_scale,
+                                         const ggml_tensor *         weights,
+                                         ggml_tensor *               dst) {
+    GGML_ASSERT(experts->type == GGML_TYPE_F32);
+    GGML_ASSERT(weights->type == GGML_TYPE_F32);
+    GGML_ASSERT(expert_scale == nullptr || expert_scale->type == GGML_TYPE_F32);
+    GGML_ASSERT(dst->type == GGML_TYPE_F32);
+    GGML_ASSERT(ggml_is_contiguous(experts));
+    GGML_ASSERT(ggml_is_contiguous(weights));
+    GGML_ASSERT(expert_scale == nullptr || ggml_is_contiguous(expert_scale));
+    GGML_ASSERT(ggml_is_contiguous(dst));
+
+    const int64_t n_embd        = experts->ne[0];
+    const int64_t n_expert_used = experts->ne[1];
+    const int64_t n_tokens      = experts->ne[2] * experts->ne[3];
+    cudaStream_t  stream        = ctx.stream();
+
+    launch_moe_weighted_reduction((const float *) experts->data,
+                                  expert_scale ? (const float *) expert_scale->data : nullptr,
+                                  (const float *) weights->data,
+                                  (float *) dst->data, n_embd, n_tokens, (int) n_expert_used, stream);
+    CUDA_CHECK(cudaGetLastError());
+}
@@ -0,0 +1,7 @@
+#include "common.cuh"
+
+void ggml_cuda_op_moe_weighted_reduction(ggml_backend_cuda_context & ctx,
+                                         const ggml_tensor *         experts,
+                                         const ggml_tensor *         expert_scale,
+                                         const ggml_tensor *         weights,
+                                         ggml_tensor *               dst);
@@ -98,12 +98,26 @@ static int    opt_ar_select = 2; // 2 = fused ALLREDUCE+ADD (DMA, default), 1 =
 //     https://docs.qualcomm.com/doc/80-N2040-61/topic/hvx-pmu-events.html
 static u32vec opt_pmu_evt { 0x3, 0x111, 0x100, 0x105, 0x240, 0x256, 0x7D, 0x8C };

-static int opt_opbatch  = 1024; // max number of ops in a batch
-static int opt_opqueue  = 64;   // max number of pending batches
+static int opt_opbatch  = 1280; // max number of ops in a batch
+static int opt_opqueue  = 32;   // max number of pending batches
 static int opt_optrace  = 0;    // trace buffer size per thread (0 means default)
 static int opt_oppoll   = 0;    // polling for batch completions
 static int opt_opfusion = 1;    // enable/disable op fusion

+enum ggml_hexagon_fusion_flags {
+    GGML_HEXAGON_FUSE_ALLREDUCE_ADD = (1 << 1), // 2
+    GGML_HEXAGON_FUSE_RMS_NORM_MUL  = (1 << 2), // 4
+    GGML_HEXAGON_FUSE_MUL_MAT_ADD   = (1 << 3), // 8
+    GGML_HEXAGON_FUSE_MUL_MAT_NX    = (1 << 4), // 16
+    GGML_HEXAGON_FUSE_MUL_MAT_ID_NX = (1 << 5), // 32
+};
+
+static inline bool ggml_hexagon_is_fusion_enabled(int flag) {
+    if (opt_opfusion <= 0) return false;
+    if (opt_opfusion == 1) return true; // 1 enables all
+    return (opt_opfusion & flag) != 0;
+}
+
 static std::regex* opt_opfilter = NULL; // regex of ops to not claim

 #define HEX_VERBOSE(...) \
@@ -293,6 +307,15 @@ static void ggml_hexagon_precompute_fused_mmnx_params(
    struct htp_mm_kernel_params * kparams
 );

+static void ggml_hexagon_precompute_fused_mmidnx_params(
+    const struct ggml_hexagon_session * sess,
+    const struct ggml_tensor * src0,
+    const struct ggml_tensor * src1,
+    const struct ggml_tensor * dst,
+    int32_t n_weights,
+    struct htp_mm_kernel_params * kparams
+);
+
 static bool ggml_hexagon_precompute_allreduce_params(
    const struct ggml_hexagon_session * sess,
    const struct ggml_tensor * dst,
@@ -304,8 +327,12 @@ static bool ggml_hexagon_precompute_allreduce_params(
 );

 static bool mm_is_hmx_eligible(const ggml_tensor * t);
+static bool is_supported_mul_mat_nx_kernel(const ggml_tensor * src0, const struct htp_mm_kernel_params * kparams);
+static bool is_supported_mul_mat_id_nx_kernel(const ggml_tensor * src0, const struct htp_mm_kernel_params * kparams);
 static bool is_mergeable_mul_mat(const ggml_tensor * t);
 static bool is_mergeable_mul_mat_pair(const ggml_tensor * n1, const ggml_tensor * n2);
+static bool is_mergeable_mul_mat_id(const ggml_tensor * t);
+static bool is_mergeable_mul_mat_id_pair(const ggml_tensor * n1, const ggml_tensor * n2);

 // ** backend sessions

@@ -1832,6 +1859,42 @@ struct ggml_hexagon_opbatch {
        }
    }

+    void sort_buffers() {
+        if (n_bufs <= 1) return;
+
+        std::vector<int> order(n_bufs);
+        for (unsigned int i = 0; i < n_bufs; i++) { order[i] = (int) i; }
+
+        std::stable_sort(order.begin(), order.end(), [&](int a, int b) {
+            return h_bufs[a].size > h_bufs[b].size;
+        });
+
+        bool already_sorted = true;
+        for (unsigned int i = 0; i < n_bufs; i++) {
+            if (order[i] != (int) i) {
+                already_sorted = false;
+                break;
+            }
+        }
+        if (already_sorted) return;
+
+        std::vector<uint16_t> remap(n_bufs);
+        std::vector<htp_buf_desc> sorted_bufs(n_bufs);
+        for (unsigned int new_bi = 0; new_bi < n_bufs; new_bi++) {
+            int old_bi = order[new_bi];
+            remap[old_bi] = (uint16_t) new_bi;
+            sorted_bufs[new_bi] = h_bufs[old_bi];
+        }
+
+        for (unsigned int i = 0; i < n_bufs; i++) {
+            h_bufs[i] = sorted_bufs[i];
+        }
+
+        for (unsigned int i = 0; i < n_tens; i++) {
+            h_tens[i].bi = remap[h_tens[i].bi];
+        }
+    }
+
    bool try_fuse_allreduce_add(const htp_opnode & node) {
        if (n_ops == 0 || opt_ar_select != 2) return false;
        if (node.opcode != HTP_OP_ADD) return false;
@@ -2144,9 +2207,15 @@ struct ggml_hexagon_opbatch {
            if (x_in != x || w_in->type != w0->type || w_in->ne[0] != w0->ne[0]) {
                return false;
            }
+            if (!last_node.fused.empty() && (mm_is_hmx_eligible(last_node.fused[0]) != mm_is_hmx_eligible(node.node))) {
+                return false;
+            }

            struct htp_mm_kernel_params kparams;
            ggml_hexagon_precompute_fused_mmnx_params(sess, w0, x, curr_n + 1, &kparams);
+            if (!is_supported_mul_mat_nx_kernel(w0, &kparams)) {
+                return false;
+            }
            if ((size_t) kparams.vtcm_size > sess->vtcm_size) {
                HEX_VERBOSE("ggml-hex: %s skip NX fusion: VTCM needed (%d) > budget (%zu)\n",
                            sess->c_name(), kparams.vtcm_size, sess->vtcm_size);
@@ -2210,6 +2279,9 @@ struct ggml_hexagon_opbatch {

            struct htp_mm_kernel_params kparams;
            ggml_hexagon_precompute_fused_mmnx_params(sess, w0, x, 2, &kparams);
+            if (!is_supported_mul_mat_nx_kernel(w0, &kparams)) {
+                return false;
+            }
            if ((size_t) kparams.vtcm_size > sess->vtcm_size) {
                HEX_VERBOSE("ggml-hex: %s skip NX fusion: VTCM needed (%d) > budget (%zu)\n",
                            sess->c_name(), kparams.vtcm_size, sess->vtcm_size);
@@ -2272,18 +2344,172 @@ struct ggml_hexagon_opbatch {
        return false;
    }

-enum ggml_hexagon_fusion_flags {
-    GGML_HEXAGON_FUSE_ALLREDUCE_ADD = (1 << 1), // 2
-    GGML_HEXAGON_FUSE_RMS_NORM_MUL  = (1 << 2), // 4
-    GGML_HEXAGON_FUSE_MUL_MAT_ADD   = (1 << 3), // 8
-    GGML_HEXAGON_FUSE_MUL_MAT_NX    = (1 << 4), // 16
-};
+    bool try_fuse_mul_mat_id_nx(const htp_opnode & node) {
+        if (n_ops == 0 || node.opcode != HTP_OP_MUL_MAT_ID) return false;
+        if (!is_mergeable_mul_mat_id(node.node)) return false;

-static inline bool ggml_hexagon_is_fusion_enabled(int flag) {
-    if (opt_opfusion <= 0) return false;
-    if (opt_opfusion == 1) return true; // 1 enables all
-    return (opt_opfusion & flag) != 0;
-}
+        const ggml_tensor * w_in   = node.src0();
+        const ggml_tensor * x_in   = node.src1();
+        const ggml_tensor * ids_in = node.node->src[2];
+        const ggml_tensor * d_in   = node.dst();
+        if (!w_in || !x_in || !ids_in || !d_in) return false;
+
+        htp_opnode & last_node = ops[n_ops - 1];
+
+        // Case 1: last_node is already MUL_MAT_ID_NX
+        if (last_node.opcode == HTP_OP_MUL_MAT_ID_NX) {
+            const uint32_t curr_n = (uint32_t) last_node.outputs.size();
+            if (curr_n >= HTP_OP_MAX_OUTPUTS || curr_n + 2 >= HTP_OP_MAX_INPUTS) {
+                return false;
+            }
+
+            const ggml_tensor * w0  = last_node.inputs[0];
+            const ggml_tensor * x   = last_node.inputs[curr_n];
+            const ggml_tensor * ids = last_node.inputs[curr_n + 1];
+
+            if (x_in != x || ids_in != ids || w_in->type != w0->type || w_in->ne[0] != w0->ne[0] || w_in->ne[2] != w0->ne[2]) {
+                return false;
+            }
+            if (!last_node.fused.empty() && (mm_is_hmx_eligible(last_node.fused[0]) != mm_is_hmx_eligible(node.node))) {
+                return false;
+            }
+
+            struct htp_mm_kernel_params kparams;
+            ggml_hexagon_precompute_fused_mmidnx_params(sess, w0, x, d_in, curr_n + 1, &kparams);
+            if (!is_supported_mul_mat_id_nx_kernel(w0, &kparams)) {
+                return false;
+            }
+            if ((size_t) kparams.vtcm_size > sess->vtcm_size) {
+                HEX_VERBOSE("ggml-hex: %s skip ID NX fusion: VTCM needed (%d) > budget (%zu)\n",
+                            sess->c_name(), kparams.vtcm_size, sess->vtcm_size);
+                return false;
+            }
+
+            size_t extra_bufs = 0, extra_vmem = 0, extra_tens = 0;
+            auto fit_t = [&](const ggml_tensor * t) {
+                if (!t) return;
+                if (!t_map.count(t)) {
+                    extra_tens++;
+                    auto sbuf = static_cast<ggml_hexagon_shared_buffer *>(t->buffer->context);
+                    if (!b_map.count(sbuf->fd())) {
+                        extra_vmem += sbuf->size();
+                        extra_bufs += 1;
+                    }
+                }
+            };
+            fit_t(w_in);
+            fit_t(d_in);
+            if ((extra_bufs + n_bufs) > n_bufs_max || (extra_tens + n_tens) > n_tens_max || (extra_vmem + b_vmem) > b_vmem_max) {
+                return false;
+            }
+
+            last_node.inputs[curr_n] = w_in;
+            last_node.inputs[curr_n + 1] = x;
+            last_node.inputs.push_back(ids);
+            last_node.outputs.push_back(d_in);
+            last_node.fused.push_back(node.node);
+            memcpy(last_node.kernel_params, &kparams, sizeof(kparams));
+
+            htp_op_desc & o = h_ops[n_ops - 1];
+            memcpy(o.kernel_params, &kparams, sizeof(kparams));
+
+            for (uint32_t s = 0; s <= curr_n + 2; s++) {
+                o.src[s] = add_tensor(last_node.inputs[s]);
+            }
+            for (uint32_t s = curr_n + 3; s < HTP_OP_MAX_INPUTS; s++) {
+                o.src[s] = 0xffff;
+            }
+            for (uint32_t d = 0; d <= curr_n; d++) {
+                o.dst[d] = add_tensor(last_node.outputs[d]);
+            }
+            for (uint32_t d = curr_n + 1; d < HTP_OP_MAX_OUTPUTS; d++) {
+                o.dst[d] = 0xffff;
+            }
+
+            HEX_VERBOSE("ggml-hex: %s fused MUL_MAT_ID_NX (N=%u, #%u)\n", sess->c_name(), curr_n + 1, n_ops - 1);
+            return true;
+        }
+
+        // Case 2: last_node is single MUL_MAT_ID
+        if (last_node.opcode == HTP_OP_MUL_MAT_ID) {
+            if (!is_mergeable_mul_mat_id_pair(last_node.node, node.node)) {
+                return false;
+            }
+
+            const ggml_tensor * w0  = last_node.src0();
+            const ggml_tensor * x   = last_node.src1();
+            const ggml_tensor * ids = last_node.node->src[2];
+            const ggml_tensor * w1  = node.src0();
+            if (!w0 || !x || !ids || !w1) return false;
+
+            struct htp_mm_kernel_params kparams;
+            ggml_hexagon_precompute_fused_mmidnx_params(sess, w0, x, node.dst(), 2, &kparams);
+            if (!is_supported_mul_mat_id_nx_kernel(w0, &kparams)) {
+                return false;
+            }
+            if ((size_t) kparams.vtcm_size > sess->vtcm_size) {
+                HEX_VERBOSE("ggml-hex: %s skip ID NX fusion: VTCM needed (%d) > budget (%zu)\n",
+                            sess->c_name(), kparams.vtcm_size, sess->vtcm_size);
+                return false;
+            }
+
+            size_t extra_bufs = 0, extra_vmem = 0, extra_tens = 0;
+            auto fit_t = [&](const ggml_tensor * t) {
+                if (!t) return;
+                if (!t_map.count(t)) {
+                    extra_tens++;
+                    auto sbuf = static_cast<ggml_hexagon_shared_buffer *>(t->buffer->context);
+                    if (!b_map.count(sbuf->fd())) {
+                        extra_vmem += sbuf->size();
+                        extra_bufs += 1;
+                    }
+                }
+            };
+            fit_t(w1);
+            fit_t(node.dst());
+            if ((extra_bufs + n_bufs) > n_bufs_max || (extra_tens + n_tens) > n_tens_max || (extra_vmem + b_vmem) > b_vmem_max) {
+                return false;
+            }
+
+            const ggml_tensor * dst_0 = last_node.dst();
+            const ggml_tensor * dst_1 = node.dst();
+
+            last_node.opcode = HTP_OP_MUL_MAT_ID_NX;
+            last_node.name   = "MUL_MAT_ID_NX";
+            last_node.inputs.clear();
+            last_node.inputs.push_back(w0);
+            last_node.inputs.push_back(w1);
+            last_node.inputs.push_back(x);
+            last_node.inputs.push_back(ids);
+            last_node.outputs.clear();
+            last_node.outputs.push_back(dst_0);
+            last_node.outputs.push_back(dst_1);
+            last_node.fused.push_back(node.node);
+            memcpy(last_node.kernel_params, &kparams, sizeof(kparams));
+
+            htp_op_desc & o = h_ops[n_ops - 1];
+            o.opcode = HTP_OP_MUL_MAT_ID_NX;
+            memcpy(o.kernel_params, &kparams, sizeof(kparams));
+
+            o.src[0] = add_tensor(w0);
+            o.src[1] = add_tensor(w1);
+            o.src[2] = add_tensor(x);
+            o.src[3] = add_tensor(ids);
+            for (uint32_t s = 4; s < HTP_OP_MAX_INPUTS; s++) {
+                o.src[s] = 0xffff;
+            }
+            o.dst[0] = add_tensor(dst_0);
+            o.dst[1] = add_tensor(dst_1);
+            for (uint32_t d = 2; d < HTP_OP_MAX_OUTPUTS; d++) {
+                o.dst[d] = 0xffff;
+            }
+
+            HEX_VERBOSE("ggml-hex: %s fused MUL_MAT_ID_NX (N=2, #%u)\n", sess->c_name(), n_ops - 1);
+            return true;
+        }
+
+        return false;
+    }

    bool try_fuse(const htp_opnode & node) {
        if (!opt_opfusion) return false;
@@ -2291,6 +2517,7 @@ static inline bool ggml_hexagon_is_fusion_enabled(int flag) {
        if (ggml_hexagon_is_fusion_enabled(GGML_HEXAGON_FUSE_RMS_NORM_MUL)  && try_fuse_rms_norm_mul(node))  return true;
        if (ggml_hexagon_is_fusion_enabled(GGML_HEXAGON_FUSE_MUL_MAT_ADD)   && try_fuse_mul_mat_add(node))   return true;
        if (ggml_hexagon_is_fusion_enabled(GGML_HEXAGON_FUSE_MUL_MAT_NX)    && try_fuse_mul_mat_nx(node))    return true;
+        if (ggml_hexagon_is_fusion_enabled(GGML_HEXAGON_FUSE_MUL_MAT_ID_NX) && try_fuse_mul_mat_id_nx(node)) return true;
        return false;
    }
 };
@@ -2350,6 +2577,8 @@ struct ggml_hexagon_opqueue {
        delete shm_buf;
    }

+    size_t shm_size() const { return shm_buf ? shm_buf->size() : 0; }
+
    // push new batch
    bool push(htp_opbatch_req& req, dspqueue_buffer& dbuf, ggml_hexagon_opbatch* op_batch) {
        static_assert(sizeof(htp_opbatch_req) % 8 == 0, "sizeof(htp_opbatch_req) must be multiple of 8");
@@ -2396,6 +2625,8 @@ struct ggml_hexagon_opqueue {
        uint8_t * t_ptr = m_ptr; m_ptr += t_size;
        uint8_t * o_ptr = m_ptr;

+        op_batch->sort_buffers();
+
        memcpy(b_ptr, (void *) op_batch->h_bufs.data(), b_size);
        memcpy(t_ptr, (void *) op_batch->h_tens.data(), t_size);
        memcpy(o_ptr, (void *) op_batch->h_ops.data(),  o_size);
@@ -3018,7 +3249,8 @@ void ggml_hexagon_session::allocate(const ggml_hexagon_device_config & config) n
        opt_vmem = ggml_hexagon_measure_max_vmem(this);
        GGML_LOG_INFO("ggml-hex: %s measured max vmem %zu\n", this->c_name(), opt_vmem);
    }
-    this->max_vmem = opt_vmem;
+    const size_t shm_size = this->op_queue->shm_size();
+    this->max_vmem = (opt_vmem > shm_size) ? (opt_vmem - shm_size) : opt_vmem;

    this->op_batch = new ggml_hexagon_opbatch(this, opt_opbatch, this->max_vmem);

@@ -3378,6 +3610,10 @@ static bool ggml_hexagon_matmul_is_hmx_eligible(
    bool is_matmul_id,
    bool is_batched
 ) {
+    if (src1->type != GGML_TYPE_F32) {
+        return false;
+    }
+
    const int ne00  = src0->ne[0];
    const int ne11  = src1->ne[1];
    const int ne12  = src1->ne[2];
@@ -3408,7 +3644,8 @@ static bool ggml_hexagon_matmul_is_hmx_eligible(
        return false;
    }

-    // M alignment: Use HMX when M > HTP_MM_HMX_MIN_NROWS
+    // M alignment: Use HMX when M > HTP_MM_HMX_MIN_NROWS.
+    // For MUL_MAT_ID, src1 shape is [K, n_expert_used, n_tokens, 1], so n_tokens is ne12.
    const int m = is_matmul_id ? ne12 : ne11;
    if (m <= HTP_MM_HMX_MIN_NROWS) {
        return false;
@@ -3460,7 +3697,7 @@ static bool ggml_hexagon_precompute_hmx_mm_params(

    if (!use_grouped) {
        // Fallback to simple 2D path (group_size = 1)
-        const int m_id_rows = (int) ((size_t) dst->ne[1] * dst->ne[2]);
+        const int m_id_rows = (dst && is_matmul_id) ? (int) ((size_t) dst->ne[1] * dst->ne[2]) : 0;
        if (!htp_mm_hmx_solve_2d_params(wtype, ne00_padded, m_id_rows, ne01_padded, ne11_padded, ne11, n_threads, pipeline, is_matmul_id, aligned_tile_size, vtcm_budget, &m_chunk, &n_chunk, &act_threads_selected, &vtcm_size)) {
            return false;
        }
@@ -3768,8 +4005,10 @@ static void ggml_hexagon_precompute_unary_params(

    kparams->n_threads = n_threads;

-    const size_t src0_data_row_size = src0->ne[0] * sizeof(float);
-    const size_t dst_data_row_size  = dst->ne[0]  * sizeof(float);
+    const size_t elem_size = ggml_type_size(src0->type);
+
+    const size_t src0_data_row_size = src0->ne[0] * elem_size;
+    const size_t dst_data_row_size  = dst->ne[0]  * ggml_type_size(dst->type);

    const size_t src0_row_size_aligned = hex_round_up(src0_data_row_size, 128);
    const size_t dst_row_size_aligned  = hex_round_up(dst_data_row_size,  128);
@@ -3783,7 +4022,7 @@ static void ggml_hexagon_precompute_unary_params(

    if (op == HTP_OP_RMS_NORM_MUL) {
        GGML_ASSERT(src1 != nullptr);
-        src1_data_row_size = src1->ne[0] * sizeof(float);
+        src1_data_row_size = src1->ne[0] * ggml_type_size(src1->type);
        src1_row_size_aligned = hex_round_up(src1_data_row_size, 128);
        broadcast_weight = (src1->ne[1] * src1->ne[2] * src1->ne[3] == 1);
    }
@@ -3797,7 +4036,7 @@ static void ggml_hexagon_precompute_unary_params(

    htp_unary_vtcm_layout_build(&L, op, src0->ne[0], dst->ne[0],
                                op == HTP_OP_RMS_NORM_MUL ? src1->ne[0] : 0,
-                                broadcast_weight, n_threads, sess->vtcm_size,
+                                broadcast_weight, n_threads, sess->vtcm_size, elem_size,
                                &col_tile, &vtcm_row_per_thread);

    kparams->col_tile = col_tile;
@@ -3918,64 +4157,113 @@ static void ggml_hexagon_precompute_fused_mmnx_params(
 ) {
    memset(kparams, 0, sizeof(*kparams));

-    const int wtype = src0->type;
-    const bool is_repack = ggml_hexagon_is_repack_type((ggml_type) wtype);
+    const int ne00 = src0->ne[0];
+    const int ne01 = src0->ne[1];
+    const int ne02 = src0->ne[2];
+    const int ne03 = src0->ne[3];

    const int ne10 = src1->ne[0];
-    const int src1_nrows = src1->ne[1] * src1->ne[2] * src1->ne[3];
-    const size_t src1_row_size = (wtype == GGML_TYPE_Q4_1) ? htp_mm_q8_1_tiled_row_size(ne10) : htp_mm_q8_0_tiled_row_size(ne10);
-    const size_t src0_row_size = src0->nb[1];
+    const int ne11 = src1->ne[1];
+    const int ne12 = src1->ne[2];
+    const int ne13 = src1->ne[3];

-    uint32_t best_n_prefetch = 16;
+    const int wtype = src0->type;
+    const bool is_repack = ggml_hexagon_is_repack_type((ggml_type) wtype);
+    const int ne00_padded = is_repack ? hex_round_up(ne00, 32) : ne00;
+    const int ne01_padded = is_repack ? hex_round_up(ne01, 32) : ne01;
+    const int ne11_padded = hex_round_up(ne11, 32);

-    if (is_repack) {
-        const uint32_t max_prefetch = (src1_nrows > HTP_MM_HMX_MIN_NROWS) ? 2 : 16;
-        best_n_prefetch = 2;
-        for (uint32_t d = max_prefetch; d >= 2; d /= 2) {
-            struct htp_mm_hvx_vtcm_layout L;
-            htp_mm_hvx_vtcm_layout_build(
-                &L, HTP_MM_KERNEL_HVX_QUANT_ROW, wtype, ne10, src1_nrows, sess->n_threads,
-                0, src0_row_size, src1_row_size, 0, d, false, true
-            );
-            if (L.total_bytes <= sess->vtcm_size) {
-                best_n_prefetch = d;
-                break;
-            }
+    const size_t vtcm_budget = sess->vtcm_size;
+    const bool is_batched = (ne02 * ne03 > 1 || ne12 * ne13 > 1);
+
+    bool hmx_enabled = (sess->n_hmx > 0) && (opt_mm_select >= 3);
+    if (hmx_enabled && ggml_hexagon_matmul_is_hmx_eligible(src0, src1, nullptr, ne01_padded, false, is_batched)) {
+        if (ggml_hexagon_precompute_hmx_mm_params(sess, src0, src1, nullptr, wtype, ne00_padded, ne01_padded, ne02, ne11, ne12, ne11_padded, false, is_batched, vtcm_budget, kparams)) {
+            kparams->n_weights = n_weights;
+            goto finalize;
        }
    }

-    struct htp_mm_hvx_vtcm_layout L;
-    bool try_tiled = (opt_mm_select >= 2);
-
-    // Test tiled first
-    htp_mm_hvx_vtcm_layout_build(
-        &L, HTP_MM_KERNEL_HVX_QUANT_ROW, wtype, ne10, src1_nrows, sess->n_threads,
-        0, src0_row_size, src1_row_size, 0, best_n_prefetch, false, true
-    );
-
-    if (try_tiled && L.total_bytes <= sess->vtcm_size) {
-        kparams->kernel_type = HTP_MM_KERNEL_HVX_QUANT_ROW;
-        kparams->vtcm_src0_size = L.src0_bytes;
-        kparams->vtcm_src1_size = L.src1_bytes;
-        kparams->vtcm_dst_size  = L.dst_bytes;
-        kparams->vtcm_size      = L.total_bytes;
-        kparams->n_prefetch     = best_n_prefetch;
-        kparams->n_weights      = n_weights;
-    } else {
-        kparams->kernel_type = HTP_MM_KERNEL_HVX_QUANT_ROW_FLAT;
-        size_t flat_src1_row_size = (wtype == GGML_TYPE_Q4_1) ? htp_mm_q8_1_flat_row_size(ne10) : htp_mm_q8_0_flat_row_size(ne10);
-
-        htp_mm_hvx_vtcm_layout_build(
-            &L, HTP_MM_KERNEL_HVX_QUANT_ROW_FLAT, wtype, ne10, src1_nrows, sess->n_threads,
-            0, src0_row_size, flat_src1_row_size, 0, best_n_prefetch, false, true
-        );
-        kparams->vtcm_src0_size = L.src0_bytes;
-        kparams->vtcm_src1_size = L.src1_bytes;
-        kparams->vtcm_dst_size  = L.dst_bytes;
-        kparams->vtcm_size      = L.total_bytes;
-        kparams->n_prefetch     = best_n_prefetch;
-        kparams->n_weights      = n_weights;
+    if (!is_repack) {
+        kparams->kernel_type = HTP_MM_KERNEL_UNSUPPORTED;
+        return;
    }
+
+    {
+        const int src1_nrows = ne11 * ne12 * ne13;
+        const size_t src1_row_size = (wtype == GGML_TYPE_Q4_1) ? htp_mm_q8_1_tiled_row_size(ne10) : htp_mm_q8_0_tiled_row_size(ne10);
+        const size_t src0_row_size = src0->nb[1];
+
+        uint32_t best_n_prefetch = 16;
+
+        if (is_repack) {
+            const uint32_t max_prefetch = (src1_nrows > HTP_MM_HMX_MIN_NROWS) ? 2 : 16;
+            best_n_prefetch = 2;
+            for (uint32_t d = max_prefetch; d >= 2; d /= 2) {
+                struct htp_mm_hvx_vtcm_layout L;
+                htp_mm_hvx_vtcm_layout_build(
+                    &L, HTP_MM_KERNEL_HVX_QUANT_ROW, wtype, ne10, src1_nrows, sess->n_threads,
+                    0, src0_row_size, src1_row_size, 0, d, false, true
+                );
+                if (L.total_bytes <= sess->vtcm_size) {
+                    best_n_prefetch = d;
+                    break;
+                }
+            }
+        }
+
+        struct htp_mm_hvx_vtcm_layout L;
+        bool try_tiled = (opt_mm_select >= 2);
+
+        // Test tiled first
+        htp_mm_hvx_vtcm_layout_build(
+            &L, HTP_MM_KERNEL_HVX_QUANT_ROW, wtype, ne10, src1_nrows, sess->n_threads,
+            0, src0_row_size, src1_row_size, 0, best_n_prefetch, false, true
+        );
+
+        if (try_tiled && L.total_bytes <= sess->vtcm_size) {
+            kparams->kernel_type = HTP_MM_KERNEL_HVX_QUANT_ROW;
+            kparams->vtcm_src0_size = L.src0_bytes;
+            kparams->vtcm_src1_size = L.src1_bytes;
+            kparams->vtcm_dst_size  = L.dst_bytes;
+            kparams->vtcm_size      = L.total_bytes;
+            kparams->n_prefetch     = best_n_prefetch;
+            kparams->n_weights      = n_weights;
+        } else {
+            kparams->kernel_type = HTP_MM_KERNEL_HVX_QUANT_ROW_FLAT;
+            size_t flat_src1_row_size = (wtype == GGML_TYPE_Q4_1) ? htp_mm_q8_1_flat_row_size(ne10) : htp_mm_q8_0_flat_row_size(ne10);
+
+            htp_mm_hvx_vtcm_layout_build(
+                &L, HTP_MM_KERNEL_HVX_QUANT_ROW_FLAT, wtype, ne10, src1_nrows, sess->n_threads,
+                0, src0_row_size, flat_src1_row_size, 0, best_n_prefetch, false, true
+            );
+            kparams->vtcm_src0_size = L.src0_bytes;
+            kparams->vtcm_src1_size = L.src1_bytes;
+            kparams->vtcm_dst_size  = L.dst_bytes;
+            kparams->vtcm_size      = L.total_bytes;
+            kparams->n_prefetch     = best_n_prefetch;
+            kparams->n_weights      = n_weights;
+        }
+    }
+
+finalize:
+    kparams->div_ne12_ne1 = init_fastdiv_values(ne12 * ne11);
+    kparams->div_ne1      = init_fastdiv_values(ne11);
+    kparams->div_r2       = init_fastdiv_values(ne02 > 0 ? ne12 / ne02 : 1);
+    kparams->div_r3       = init_fastdiv_values(ne03 > 0 ? ne13 / ne03 : 1);
+    kparams->div_ne11     = init_fastdiv_values(ne11);
+}
+
+static void ggml_hexagon_precompute_fused_mmidnx_params(
+    const struct ggml_hexagon_session * sess,
+    const struct ggml_tensor * src0, // W0
+    const struct ggml_tensor * src1, // x
+    const struct ggml_tensor * dst,  // dst0
+    int32_t n_weights,
+    struct htp_mm_kernel_params * kparams
+) {
+    ggml_hexagon_precompute_matmul_params_impl(sess, src0, src1, dst, 0, kparams);
+    kparams->n_weights = n_weights;
 }

 static bool ggml_hexagon_tensor_is_host(const struct ggml_hexagon_session * sess, const struct ggml_tensor * t) {
@@ -4010,11 +4298,6 @@ static bool ggml_hexagon_supported_mul_mat(const struct ggml_hexagon_session * s
                return false;
            }

-            // hardcoded limit to refuse the lm-head for now
-            if (src0->ne[1] > 32768) {
-                return false;
-            }
-
            if (src1->ne[2] != 1 || src1->ne[3] != 1) {
                return false;  // no broadcasting (for now)
            }
@@ -4170,15 +4453,39 @@ static bool ggml_hexagon_supported_unary(const struct ggml_hexagon_session * ses
    const struct ggml_tensor * src0 = op->src[0];
    const struct ggml_tensor * dst  = op;

-    if (src0->type != GGML_TYPE_F32) {
+    if (src0->type != GGML_TYPE_F32 && src0->type != GGML_TYPE_F16) {
        return false;
    }
-    if (dst->type != GGML_TYPE_F32) {
+    if (dst->type != src0->type) {
        return false;
    }
    if (!ggml_is_contiguous_rows(src0)) {
        return false;
    }
+
+    // F16 device kernels only cover this explicit whitelist (must stay in sync with
+    // the is_f16 whitelist in execute_op_unary(), unary-ops.c).
+    if (src0->type == GGML_TYPE_F16) {
+        switch (op->op) {
+            case GGML_OP_NORM:
+            case GGML_OP_RMS_NORM:
+            case GGML_OP_L2_NORM:
+            case GGML_OP_SCALE:
+            case GGML_OP_CLAMP:
+            case GGML_OP_SQR:
+            case GGML_OP_SQRT:
+            case GGML_OP_LOG:
+                break;
+            case GGML_OP_UNARY:
+                if (ggml_get_unary_op(op) != GGML_UNARY_OP_ABS) {
+                    return false;
+                }
+                break;
+            default:
+                return false;
+        }
+    }
+
    if (!ggml_are_same_shape(src0, dst)) {
        return false;
    }
@@ -4348,6 +4655,13 @@ static bool ggml_hexagon_supported_get_rows(const struct ggml_hexagon_session *
    const struct ggml_tensor * src1 = op->src[1]; // indices
    const struct ggml_tensor * dst  = op;

+    if (src0->extra) {
+        const auto * extra = (const ggml_hexagon_tensor_extra *) src0->extra;
+        if (extra->flags & GGML_HEXAGON_TENSOR_REPACK) {
+            return false;
+        }
+    }
+
    if (src0->type != GGML_TYPE_F32 && src0->ne[0] < 32) {
        return false;
    }
@@ -4734,10 +5048,43 @@ static bool mm_is_hmx_eligible(const ggml_tensor * t) {
    return ggml_hexagon_matmul_is_hmx_eligible(src0, src1, t, ne01_padded, is_matmul_id, is_batched);
 }

+static bool is_supported_mul_mat_nx_kernel(const ggml_tensor * src0, const struct htp_mm_kernel_params * kparams) {
+    if (kparams->n_hmx) {
+        return kparams->kernel_type == HTP_MM_KERNEL_HMX_2D;
+    }
+
+    if (!ggml_hexagon_is_repack_type(src0->type)) {
+        return false;
+    }
+
+    return kparams->kernel_type == HTP_MM_KERNEL_HVX_QUANT_ROW || kparams->kernel_type == HTP_MM_KERNEL_HVX_QUANT_ROW_FLAT;
+}
+
+static bool is_supported_mul_mat_id_nx_kernel(const ggml_tensor * src0, const struct htp_mm_kernel_params * kparams) {
+    if (kparams->n_hmx) {
+        return kparams->kernel_type == HTP_MM_KERNEL_HMX_2D;
+    }
+
+    if (!ggml_hexagon_is_repack_type(src0->type)) {
+        return false;
+    }
+
+    return kparams->kernel_type == HTP_MM_KERNEL_HVX_QUANT_ROW || kparams->kernel_type == HTP_MM_KERNEL_HVX_QUANT_BLOCK;
+}
+
 static bool is_mergeable_mul_mat(const ggml_tensor * t) {
-    if (!t || t->op != GGML_OP_MUL_MAT)   return false;
-    if (t->src[1]->type != GGML_TYPE_F32) return false;
-    return ggml_is_quantized(t->src[0]->type) && !mm_is_hmx_eligible(t);
+    if (!t || t->op != GGML_OP_MUL_MAT) return false;
+
+    const ggml_tensor * src0 = t->src[0];
+    const ggml_tensor * src1 = t->src[1];
+    if (src1->type != GGML_TYPE_F32) return false;
+    if (src0->ne[2] != 1 || src0->ne[3] != 1) return false;
+
+    if (mm_is_hmx_eligible(t)) {
+        return ggml_hexagon_is_hmx_weight_type(src0->type);
+    }
+
+    return ggml_hexagon_is_repack_type(src0->type);
 }

 static bool is_mergeable_mul_mat_pair(const ggml_tensor * n1, const ggml_tensor * n2) {
@@ -4753,6 +5100,41 @@ static bool is_mergeable_mul_mat_pair(const ggml_tensor * n1, const ggml_tensor
    if (n1->src[0]->type != n2->src[0]->type) {
        return false;
    }
+    if (mm_is_hmx_eligible(n1) != mm_is_hmx_eligible(n2)) {
+        return false;
+    }
+    return true;
+}
+
+static bool is_mergeable_mul_mat_id(const ggml_tensor * t) {
+    if (!t || t->op != GGML_OP_MUL_MAT_ID) return false;
+
+    const ggml_tensor * src0 = t->src[0];
+    return ggml_hexagon_is_repack_type(src0->type);
+}
+
+static bool is_mergeable_mul_mat_id_pair(const ggml_tensor * n1, const ggml_tensor * n2) {
+    if (!is_mergeable_mul_mat_id(n1) || !is_mergeable_mul_mat_id(n2)) {
+        return false;
+    }
+    if (n1->src[1] != n2->src[1]) {
+        return false;
+    }
+    if (n1->src[2] != n2->src[2]) {
+        return false;
+    }
+    if (n1->src[0]->ne[0] != n2->src[0]->ne[0]) {
+        return false;
+    }
+    if (n1->src[0]->ne[2] != n2->src[0]->ne[2]) {
+        return false;
+    }
+    if (n1->src[0]->type != n2->src[0]->type) {
+        return false;
+    }
+    if (mm_is_hmx_eligible(n1) != mm_is_hmx_eligible(n2)) {
+        return false;
+    }
    return true;
 }

@@ -4776,8 +5158,8 @@ static ggml_status ggml_backend_hexagon_graph_compute(ggml_backend_t backend, gg

            if (graph->nodes[i]->op == GGML_OP_RMS_NORM && ggml_can_fuse(graph, i, { GGML_OP_RMS_NORM, GGML_OP_MUL })) {
                extra->flags |= GGML_HEXAGON_TENSOR_FUSEABLE;
-            } else if (graph->nodes[i]->op == GGML_OP_MUL_MAT) {
-                if ((i + 1 < graph->n_nodes && graph->nodes[i + 1]->op == GGML_OP_ADD && ggml_can_fuse(graph, i, { GGML_OP_MUL_MAT, GGML_OP_ADD })) ||
+            } else if (graph->nodes[i]->op == GGML_OP_MUL_MAT || graph->nodes[i]->op == GGML_OP_MUL_MAT_ID) {
+                if ((i + 1 < graph->n_nodes && graph->nodes[i + 1]->op == GGML_OP_ADD && ggml_can_fuse(graph, i, { graph->nodes[i]->op, GGML_OP_ADD })) ||
                    ggml_node_has_n_uses(graph, i, 1)) {
                    extra->flags |= GGML_HEXAGON_TENSOR_FUSEABLE;
                }
@@ -315,7 +315,8 @@ struct htp_opformat {
    }
    void format_kernel_params(char * str, size_t max_size, const htp_opnode & node) {
        if (node.opcode == HTP_OP_MUL_MAT || node.opcode == HTP_OP_MUL_MAT_ID ||
-            node.opcode == HTP_OP_MUL_MAT_NX || node.opcode == HTP_OP_MUL_MAT_ADD) {
+            node.opcode == HTP_OP_MUL_MAT_NX || node.opcode == HTP_OP_MUL_MAT_ID_NX ||
+            node.opcode == HTP_OP_MUL_MAT_ADD) {
            const auto * kparams = (const struct htp_mm_kernel_params *) node.kernel_params;
            const char * path = "unknown";
            int32_t type = kparams->kernel_type;
@@ -323,6 +323,10 @@ int op_cpy(struct htp_ops_context * octx) {
        return HTP_STATUS_NO_SUPPORT;
    }

+    FARF(HIGH, "cpy-%s-%s: (%ux%ux%ux%u) -> (%ux%ux%ux%u) : use_dma=%d n_threads %u\n",
+         src0->type == HTP_TYPE_F32 ? "f32" : "f16", dst->type == HTP_TYPE_F32 ? "f32" : "f16",
+         ne00, ne01, ne02, ne03, ne0, ne1, ne2, ne3, use_dma, n_threads);
+
    if (use_dma) {
        cpy_dma_sametype_sameshape(octx, dst, src0, ct.src0_type_size, ne00, ne01, ne02, ne03, nb01, nb02, nb03, nb1, nb2, nb3);
    } else {
@@ -1138,6 +1138,15 @@ int op_gated_delta_net(struct htp_ops_context * octx) {
    gctx.vtcm_base = octx->ctx->vtcm_base;
    gctx.vtcm_per_thread = 2 * state_aligned;

+    FARF(HIGH, "gated-delta-net-f32: q(%ux%ux%ux%u) k(%ux%ux%ux%u) v(%ux%ux%ux%u) state(%ux%ux%ux%u) -> (%ux%ux%ux%u) : "
+         "vtcm-size %zu n_threads %u\n",
+         q->ne[0], q->ne[1], q->ne[2], q->ne[3],
+         k->ne[0], k->ne[1], k->ne[2], k->ne[3],
+         v->ne[0], v->ne[1], v->ne[2], v->ne[3],
+         state->ne[0], state->ne[1], state->ne[2], state->ne[3],
+         dst->ne[0], dst->ne[1], dst->ne[2], dst->ne[3],
+         gctx.vtcm_per_thread * octx->n_threads, octx->n_threads);
+
    if (n_tokens == 1) {
        worker_pool_run_func(octx->ctx->worker_pool, gated_delta_net_f32_tg_thread, &gctx, octx->n_threads);
    } else {
@@ -247,6 +247,14 @@ int op_get_rows(struct htp_ops_context * octx) {
        }
    }

+    FARF(HIGH, "get-rows: (%ux%ux%ux%u) x (%ux%ux%ux%u) -> (%ux%ux%ux%u) : src0-vtcm-size %zu dst-vtcm-size %zu use_dma=%d n_threads %d\n",
+         octx->src[0]->ne[0], octx->src[0]->ne[1], octx->src[0]->ne[2], octx->src[0]->ne[3],
+         octx->src[1]->ne[0], octx->src[1]->ne[1], octx->src[1]->ne[2], octx->src[1]->ne[3],
+         octx->dst->ne[0], octx->dst->ne[1], octx->dst->ne[2], octx->dst->ne[3],
+         grctx.vtcm_layout.src0_bytes_per_thread * kparams->n_threads,
+         grctx.vtcm_layout.dst_bytes_per_thread  * kparams->n_threads,
+         kparams->use_dma, kparams->n_threads);
+
    work_queue_run(octx->ctx->work_queue, q_func, &grctx, kparams->n_threads);
    return HTP_STATUS_OK;
 }
@@ -118,6 +118,7 @@ struct htp_context {
 int op_matmul(struct htp_ops_context * octx);
 int op_matmul_id(struct htp_ops_context * octx);
 int op_matmul_nx(struct htp_ops_context * octx);
+int op_matmul_id_nx(struct htp_ops_context * octx);
 int op_binary(struct htp_ops_context * octx);
 int op_unary(struct htp_ops_context * octx);
 int op_sum_rows(struct htp_ops_context * octx);
@@ -52,6 +52,7 @@ enum htp_op_code {
    HTP_OP_MUL_MAT,
    HTP_OP_MUL_MAT_ID,
    HTP_OP_MUL_MAT_NX,
+    HTP_OP_MUL_MAT_ID_NX,
    HTP_OP_MUL_MAT_ADD,
    HTP_OP_RMS_NORM,
    HTP_OP_RMS_NORM_MUL,
@@ -358,6 +358,54 @@ static inline void hvx_clamp_scalar_f32(uint8_t * restrict dst, const uint8_t *
    }
 }

+#define HVX_OP_CLAMP_SCALAR_F16(v)                                       \
+    ({                                                                   \
+        HVX_VectorPred pred_cap_right = Q6_Q_vcmp_gt_VhfVhf(v, max_vec); \
+        HVX_VectorPred pred_cap_left  = Q6_Q_vcmp_gt_VhfVhf(min_vec, v); \
+        HVX_Vector tmp = Q6_V_vmux_QVV(pred_cap_right, max_vec, v);      \
+        Q6_V_vmux_QVV(pred_cap_left, min_vec, tmp);                      \
+    })
+
+static inline void hvx_clamp_scalar_f16_aa(uint8_t * restrict dst, const uint8_t * restrict src, const _Float16 min, const _Float16 max, uint32_t n) {
+    const HVX_Vector min_vec = hvx_vec_splat_f16(min);
+    const HVX_Vector max_vec = hvx_vec_splat_f16(max);
+    assert((unsigned long) dst % 128 == 0);
+    assert((unsigned long) src % 128 == 0);
+    hvx_scalar_loop_body(HVX_Vector, HVX_Vector, sizeof(_Float16), hvx_vec_store_a, HVX_OP_CLAMP_SCALAR_F16);
+}
+
+static inline void hvx_clamp_scalar_f16_au(uint8_t * restrict dst, const uint8_t * restrict src, const _Float16 min, const _Float16 max, uint32_t n) {
+    const HVX_Vector min_vec = hvx_vec_splat_f16(min);
+    const HVX_Vector max_vec = hvx_vec_splat_f16(max);
+    assert((unsigned long) dst % 128 == 0);
+    hvx_scalar_loop_body(HVX_Vector, HVX_UVector, sizeof(_Float16), hvx_vec_store_a, HVX_OP_CLAMP_SCALAR_F16);
+}
+
+static inline void hvx_clamp_scalar_f16_ua(uint8_t * restrict dst, const uint8_t * restrict src, const _Float16 min, const _Float16 max, uint32_t n) {
+    const HVX_Vector min_vec = hvx_vec_splat_f16(min);
+    const HVX_Vector max_vec = hvx_vec_splat_f16(max);
+    assert((unsigned long) src % 128 == 0);
+    hvx_scalar_loop_body(HVX_UVector, HVX_Vector, sizeof(_Float16), hvx_vec_store_u, HVX_OP_CLAMP_SCALAR_F16);
+}
+
+static inline void hvx_clamp_scalar_f16_uu(uint8_t * restrict dst, const uint8_t * restrict src, const _Float16 min, const _Float16 max, uint32_t n) {
+    const HVX_Vector min_vec = hvx_vec_splat_f16(min);
+    const HVX_Vector max_vec = hvx_vec_splat_f16(max);
+    hvx_scalar_loop_body(HVX_UVector, HVX_UVector, sizeof(_Float16), hvx_vec_store_u, HVX_OP_CLAMP_SCALAR_F16);
+}
+
+static inline void hvx_clamp_scalar_f16(uint8_t * restrict dst, const uint8_t * restrict src, const _Float16 min, const _Float16 max, const int num_elems) {
+    if (hex_is_aligned((void *) dst, 128) && hex_is_aligned((void *) src, 128)) {
+        hvx_clamp_scalar_f16_aa(dst, src, min, max, num_elems);
+    } else if (hex_is_aligned((void *) dst, 128)) {
+        hvx_clamp_scalar_f16_au(dst, src, min, max, num_elems);
+    } else if (hex_is_aligned((void *) src, 128)) {
+        hvx_clamp_scalar_f16_ua(dst, src, min, max, num_elems);
+    } else {
+        hvx_clamp_scalar_f16_uu(dst, src, min, max, num_elems);
+    }
+}
+
 //
 // Abs
 //
@@ -386,11 +434,69 @@ static inline void hvx_abs_f32_aa(uint8_t * restrict dst, const uint8_t * restri
    }
 }

+#define hvx_abs_f16_loop_body(dst_type, src_type, vec_store)               \
+    do {                                                                   \
+        dst_type * restrict vdst  = (dst_type *) dst;                      \
+        src_type * restrict vsrc = (src_type *) src;                       \
+                                                                           \
+        const uint32_t elem_size = sizeof(_Float16);                       \
+        const uint32_t epv  = 128 / elem_size;                             \
+        const uint32_t nvec = n / epv;                                     \
+        const uint32_t nloe = n % epv;                                     \
+                                                                           \
+        uint32_t i = 0;                                                    \
+                                                                           \
+        _Pragma("unroll(4)")                                               \
+        for (; i < nvec; i++) {                                            \
+            vdst[i] = hvx_vec_abs_f16(vsrc[i]);                            \
+        }                                                                  \
+        if (nloe) {                                                        \
+            HVX_Vector v = hvx_vec_abs_f16(vsrc[i]);                       \
+            vec_store((void *) &vdst[i], nloe * elem_size, v);             \
+        }                                                                  \
+    } while(0)
+
+static inline void hvx_abs_f16_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    assert((unsigned long) src % 128 == 0);
+    hvx_abs_f16_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+}
+
+static inline void hvx_abs_f16_au(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    hvx_abs_f16_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a);
+}
+
+static inline void hvx_abs_f16_ua(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) src % 128 == 0);
+    hvx_abs_f16_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
+}
+
+static inline void hvx_abs_f16_uu(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    hvx_abs_f16_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
+}
+
+static inline void hvx_abs_f16(uint8_t * restrict dst, const uint8_t * restrict src, const uint32_t num_elems) {
+    if (hex_is_aligned((void *) dst, 128)) {
+        if (hex_is_aligned((void *) src, 128)) {
+            hvx_abs_f16_aa(dst, src, num_elems);
+        } else {
+            hvx_abs_f16_au(dst, src, num_elems);
+        }
+    } else {
+        if (hex_is_aligned((void *) src, 128)) {
+            hvx_abs_f16_ua(dst, src, num_elems);
+        } else {
+            hvx_abs_f16_uu(dst, src, num_elems);
+        }
+    }
+}
+
 //
 // Square
 //

-#define hvx_sqr_f32_loop_body(dst_type, src_type, vec_store)           \
+#define hvx_sqr_f32_loop_body(dst_type, src_type, vec_store)               \
    do {                                                                   \
        dst_type * restrict vdst  = (dst_type *) dst;                      \
        src_type * restrict vsrc = (src_type *) src;                       \
@@ -404,10 +510,10 @@ static inline void hvx_abs_f32_aa(uint8_t * restrict dst, const uint8_t * restri
                                                                           \
        _Pragma("unroll(4)")                                               \
        for (; i < nvec; i++) {                                            \
-            vdst[i] = HVX_OP_MUL_F32(vsrc[i], vsrc[i]);                        \
+            vdst[i] = HVX_OP_MUL_F32(vsrc[i], vsrc[i]);                    \
        }                                                                  \
        if (nloe) {                                                        \
-            HVX_Vector v = HVX_OP_MUL_F32(vsrc[i], vsrc[i]);                   \
+            HVX_Vector v = HVX_OP_MUL_F32(vsrc[i], vsrc[i]);               \
            vec_store((void *) &vdst[i], nloe * elem_size, v);             \
        }                                                                  \
    } while(0)
@@ -448,6 +554,64 @@ static inline void hvx_sqr_f32(uint8_t * restrict dst, const uint8_t * restrict
    }
 }

+#define hvx_sqr_f16_loop_body(dst_type, src_type, vec_store)               \
+    do {                                                                   \
+        dst_type * restrict vdst  = (dst_type *) dst;                      \
+        src_type * restrict vsrc = (src_type *) src;                       \
+                                                                           \
+        const uint32_t elem_size = sizeof(_Float16);                       \
+        const uint32_t epv  = 128 / elem_size;                             \
+        const uint32_t nvec = n / epv;                                     \
+        const uint32_t nloe = n % epv;                                     \
+                                                                           \
+        uint32_t i = 0;                                                    \
+                                                                           \
+        _Pragma("unroll(4)")                                               \
+        for (; i < nvec; i++) {                                            \
+            vdst[i] = HVX_OP_MUL_F16(vsrc[i], vsrc[i]);                    \
+        }                                                                  \
+        if (nloe) {                                                        \
+            HVX_Vector v = HVX_OP_MUL_F16(vsrc[i], vsrc[i]);               \
+            vec_store((void *) &vdst[i], nloe * elem_size, v);             \
+        }                                                                  \
+    } while(0)
+
+static inline void hvx_sqr_f16_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    assert((unsigned long) src % 128 == 0);
+    hvx_sqr_f16_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+}
+
+static inline void hvx_sqr_f16_au(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    hvx_sqr_f16_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a);
+}
+
+static inline void hvx_sqr_f16_ua(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) src % 128 == 0);
+    hvx_sqr_f16_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
+}
+
+static inline void hvx_sqr_f16_uu(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    hvx_sqr_f16_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
+}
+
+static inline void hvx_sqr_f16(uint8_t * restrict dst, const uint8_t * restrict src, const uint32_t num_elems) {
+    if (hex_is_aligned((void *) dst, 128)) {
+        if (hex_is_aligned((void *) src, 128)) {
+            hvx_sqr_f16_aa(dst, src, num_elems);
+        } else {
+            hvx_sqr_f16_au(dst, src, num_elems);
+        }
+    } else {
+        if (hex_is_aligned((void *) src, 128)) {
+            hvx_sqr_f16_ua(dst, src, num_elems);
+        } else {
+            hvx_sqr_f16_uu(dst, src, num_elems);
+        }
+    }
+}
+
 #undef HVX_OP_ADD_F32
 #undef HVX_OP_SUB_F32
 #undef HVX_OP_MUL_F32
@@ -464,6 +628,7 @@ static inline void hvx_sqr_f32(uint8_t * restrict dst, const uint8_t * restrict
 #undef hvx_scalar_loop_body
 #undef HVX_OP_MIN_SCALAR
 #undef HVX_OP_CLAMP_SCALAR
+#undef HVX_OP_CLAMP_SCALAR_F16
 #undef DEFINE_HVX_BINARY_OP_VARIANTS
 #undef HVX_BINARY_DISPATCHER
 #undef UNUSED
@@ -86,4 +86,33 @@ static inline void hvx_log_f32_aa(uint8_t * restrict dst, const uint8_t * restri
    }
 }

+// Compute log(x) for f16 by promoting to f32, applying hvx_vec_log_f32, and narrowing back.
+static inline void hvx_log_f16_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    assert((unsigned long) src % 128 == 0);
+
+    HVX_Vector * restrict vdst = (HVX_Vector *) dst;
+    HVX_Vector * restrict vsrc = (HVX_Vector *) src;
+
+    const uint32_t nvec = n / VLEN_FP16;
+    const uint32_t nloe = n % VLEN_FP16;
+
+    uint32_t i = 0;
+
+    _Pragma("unroll(4)")
+    for (; i < nvec; i++) {
+        HVX_VectorPair p = hvx_vec_f16_to_f32(vsrc[i]);
+        HVX_Vector r0 = hvx_vec_log_f32(Q6_V_lo_W(p));
+        HVX_Vector r1 = hvx_vec_log_f32(Q6_V_hi_W(p));
+        vdst[i] = hvx_vec_f32_to_f16(r0, r1);
+    }
+    if (nloe) {
+        HVX_VectorPair p = hvx_vec_f16_to_f32(vsrc[i]);
+        HVX_Vector r0 = hvx_vec_log_f32(Q6_V_lo_W(p));
+        HVX_Vector r1 = hvx_vec_log_f32(Q6_V_hi_W(p));
+        HVX_Vector v = hvx_vec_f32_to_f16(r0, r1);
+        hvx_vec_store_a((void *) &vdst[i], nloe * SIZEOF_FP16, v);
+    }
+}
+
 #endif /* HVX_LOG_H */
@@ -254,4 +254,201 @@ static inline void hvx_fast_l2_norm_f32(const uint8_t * restrict src,
    }
 }

+// F16 norm kernels: reduce and scale in f32 (via promote/narrow), matching the
+// precision-preserving pattern used by the flash-attn f16 kernels.
+
+static inline void hvx_fast_rms_norm_f16(const uint8_t * restrict src,
+                                         uint8_t * restrict dst,
+                                         const int num_elems,
+                                         float     epsilon) {
+
+    const HVX_Vector * restrict v_src = (HVX_Vector *) src;
+    HVX_Vector * restrict v_dst       = (HVX_Vector *) dst;
+
+    const int nvec = num_elems / VLEN_FP16;    // number of full f16 vectors
+    const int nloe = num_elems % VLEN_FP16;    // leftover elements
+
+    HVX_Vector sum_v = Q6_V_vsplat_R(0x00000000);
+    HVX_Vector epsilon_v = hvx_vec_splat_f32(epsilon);
+
+    #pragma unroll(4)
+    for (int i = 0; i < nvec; i++) {
+        HVX_VectorPair p = hvx_vec_f16_to_f32(v_src[i]);
+        HVX_Vector p0 = Q6_V_lo_W(p);
+        HVX_Vector p1 = Q6_V_hi_W(p);
+        sum_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, Q6_Vqf32_vmpy_VsfVsf(p0, p0));
+        sum_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, Q6_Vqf32_vmpy_VsfVsf(p1, p1));
+    }
+
+    if (nloe > 0) {
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * SIZEOF_FP16);
+        HVX_Vector v1 = Q6_V_vand_QV(bmask, v_src[nvec]);
+        HVX_VectorPair p = hvx_vec_f16_to_f32(v1);
+        HVX_Vector p0 = Q6_V_lo_W(p);
+        HVX_Vector p1 = Q6_V_hi_W(p);
+        sum_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, Q6_Vqf32_vmpy_VsfVsf(p0, p0));
+        sum_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, Q6_Vqf32_vmpy_VsfVsf(p1, p1));
+    }
+
+    sum_v = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(sum_v));
+
+    HVX_Vector t_v            = hvx_vec_splat_f32((float) num_elems);
+    HVX_Vector denom_v        = hvx_vec_inverse_f32(t_v);
+    HVX_Vector mean_v         = Q6_Vqf32_vmpy_VsfVsf(sum_v, denom_v);
+    HVX_Vector mean_epsilon_v = Q6_Vqf32_vadd_Vqf32Vsf(mean_v, epsilon_v);
+
+    HVX_Vector scale_v = hvx_vec_rsqrt_f32(Q6_Vsf_equals_Vqf32(mean_epsilon_v));
+
+    #pragma unroll(4)
+    for (int i = 0; i < nvec; i++) {
+        HVX_VectorPair p = hvx_vec_f16_to_f32(v_src[i]);
+        HVX_Vector r0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(p), scale_v));
+        HVX_Vector r1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_V_hi_W(p), scale_v));
+        v_dst[i] = hvx_vec_f32_to_f16(r0, r1);
+    }
+
+    if (nloe > 0) {
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * SIZEOF_FP16);
+        HVX_Vector v1 = Q6_V_vand_QV(bmask, v_src[nvec]);
+        HVX_VectorPair p = hvx_vec_f16_to_f32(v1);
+        HVX_Vector r0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(p), scale_v));
+        HVX_Vector r1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_V_hi_W(p), scale_v));
+        HVX_Vector result = hvx_vec_f32_to_f16(r0, r1);
+        hvx_vec_store_a(&v_dst[nvec], nloe * SIZEOF_FP16, result);
+    }
+}
+
+static inline void hvx_fast_norm_f16(const uint8_t * restrict src,
+                                     uint8_t * restrict dst,
+                                     const int num_elems,
+                                     float     epsilon) {
+
+    const HVX_Vector * restrict v_src = (HVX_Vector *) src;
+    HVX_Vector * restrict v_dst       = (HVX_Vector *) dst;
+
+    const int nvec = num_elems / VLEN_FP16;
+    const int nloe = num_elems % VLEN_FP16;
+
+    HVX_Vector sum_sq_v = Q6_V_vsplat_R(0x00000000);
+    HVX_Vector sum_x_v  = Q6_V_vsplat_R(0x00000000);
+    HVX_Vector epsilon_v = hvx_vec_splat_f32(epsilon);
+
+    #pragma unroll(4)
+    for (int i = 0; i < nvec; i++) {
+        HVX_VectorPair p = hvx_vec_f16_to_f32(v_src[i]);
+        HVX_Vector p0 = Q6_V_lo_W(p);
+        HVX_Vector p1 = Q6_V_hi_W(p);
+        sum_sq_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_sq_v, Q6_Vqf32_vmpy_VsfVsf(p0, p0));
+        sum_sq_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_sq_v, Q6_Vqf32_vmpy_VsfVsf(p1, p1));
+        sum_x_v  = Q6_Vqf32_vadd_Vqf32Vqf32(sum_x_v,  Q6_Vqf32_vadd_VsfVsf(p0, Q6_V_vzero()));
+        sum_x_v  = Q6_Vqf32_vadd_Vqf32Vqf32(sum_x_v,  Q6_Vqf32_vadd_VsfVsf(p1, Q6_V_vzero()));
+    }
+
+    if (nloe > 0) {
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * SIZEOF_FP16);
+        HVX_Vector v1 = Q6_V_vand_QV(bmask, v_src[nvec]);
+        HVX_VectorPair p = hvx_vec_f16_to_f32(v1);
+        HVX_Vector p0 = Q6_V_lo_W(p);
+        HVX_Vector p1 = Q6_V_hi_W(p);
+        sum_sq_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_sq_v, Q6_Vqf32_vmpy_VsfVsf(p0, p0));
+        sum_sq_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_sq_v, Q6_Vqf32_vmpy_VsfVsf(p1, p1));
+        sum_x_v  = Q6_Vqf32_vadd_Vqf32Vqf32(sum_x_v,  Q6_Vqf32_vadd_VsfVsf(p0, Q6_V_vzero()));
+        sum_x_v  = Q6_Vqf32_vadd_Vqf32Vqf32(sum_x_v,  Q6_Vqf32_vadd_VsfVsf(p1, Q6_V_vzero()));
+    }
+
+    sum_sq_v = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(sum_sq_v));
+    sum_x_v  = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(sum_x_v));
+
+    HVX_Vector t_v            = hvx_vec_splat_f32((float) num_elems);
+    HVX_Vector denom_v        = hvx_vec_inverse_f32(t_v);
+    HVX_Vector mean_sq_v      = Q6_Vqf32_vmpy_VsfVsf(sum_sq_v, denom_v);
+    HVX_Vector mean_x_v       = Q6_Vqf32_vmpy_VsfVsf(sum_x_v,  denom_v);
+    HVX_Vector mean_x_sq_v    = Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(mean_x_v), Q6_Vsf_equals_Vqf32(mean_x_v));
+    HVX_Vector var_v          = Q6_Vqf32_vsub_Vqf32Vqf32(mean_sq_v, mean_x_sq_v);
+    HVX_Vector var_epsilon_v  = Q6_Vqf32_vadd_Vqf32Vsf(var_v, epsilon_v);
+
+    HVX_Vector scale_v  = hvx_vec_rsqrt_f32(Q6_Vsf_equals_Vqf32(var_epsilon_v));
+    HVX_Vector mean_x_b = hvx_vec_repl_f32(Q6_Vsf_equals_Vqf32(mean_x_v));
+
+    #pragma unroll(4)
+    for (int i = 0; i < nvec; i++) {
+        HVX_VectorPair p = hvx_vec_f16_to_f32(v_src[i]);
+        HVX_Vector d0 = Q6_Vqf32_vsub_VsfVsf(Q6_V_lo_W(p), mean_x_b);
+        HVX_Vector d1 = Q6_Vqf32_vsub_VsfVsf(Q6_V_hi_W(p), mean_x_b);
+        HVX_Vector r0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(d0), scale_v));
+        HVX_Vector r1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(d1), scale_v));
+        v_dst[i] = hvx_vec_f32_to_f16(r0, r1);
+    }
+
+    if (nloe > 0) {
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * SIZEOF_FP16);
+        HVX_Vector v1 = Q6_V_vand_QV(bmask, v_src[nvec]);
+        HVX_VectorPair p = hvx_vec_f16_to_f32(v1);
+        HVX_Vector d0 = Q6_Vqf32_vsub_VsfVsf(Q6_V_lo_W(p), mean_x_b);
+        HVX_Vector d1 = Q6_Vqf32_vsub_VsfVsf(Q6_V_hi_W(p), mean_x_b);
+        HVX_Vector r0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(d0), scale_v));
+        HVX_Vector r1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_Vsf_equals_Vqf32(d1), scale_v));
+        HVX_Vector result = hvx_vec_f32_to_f16(r0, r1);
+        hvx_vec_store_a(&v_dst[nvec], nloe * SIZEOF_FP16, result);
+    }
+}
+
+static inline void hvx_fast_l2_norm_f16(const uint8_t * restrict src,
+                                        uint8_t * restrict dst,
+                                        const int num_elems,
+                                        float     epsilon) {
+
+    const HVX_Vector * restrict v_src = (HVX_Vector *) src;
+    HVX_Vector * restrict v_dst       = (HVX_Vector *) dst;
+
+    const int nvec = num_elems / VLEN_FP16;
+    const int nloe = num_elems % VLEN_FP16;
+
+    HVX_Vector sum_v = hvx_vec_splat_f32(0.0f);
+
+    #pragma unroll(4)
+    for (int i = 0; i < nvec; i++) {
+        HVX_VectorPair p = hvx_vec_f16_to_f32(v_src[i]);
+        HVX_Vector p0 = Q6_V_lo_W(p);
+        HVX_Vector p1 = Q6_V_hi_W(p);
+        sum_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, Q6_Vqf32_vmpy_VsfVsf(p0, p0));
+        sum_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, Q6_Vqf32_vmpy_VsfVsf(p1, p1));
+    }
+
+    if (nloe > 0) {
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * SIZEOF_FP16);
+        HVX_Vector v1 = Q6_V_vand_QV(bmask, v_src[nvec]);
+        HVX_VectorPair p = hvx_vec_f16_to_f32(v1);
+        HVX_Vector p0 = Q6_V_lo_W(p);
+        HVX_Vector p1 = Q6_V_hi_W(p);
+        sum_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, Q6_Vqf32_vmpy_VsfVsf(p0, p0));
+        sum_v = Q6_Vqf32_vadd_Vqf32Vqf32(sum_v, Q6_Vqf32_vmpy_VsfVsf(p1, p1));
+    }
+
+    HVX_Vector sum_sf    = hvx_vec_reduce_sum_f32(Q6_Vsf_equals_Vqf32(sum_v));
+    HVX_Vector rsqrt_v   = hvx_vec_rsqrt_f32(sum_sf);
+    HVX_Vector sqrt_v    = hvx_vec_inverse_f32(rsqrt_v);
+    HVX_Vector epsilon_v = hvx_vec_splat_f32(epsilon);
+    HVX_Vector denom_v   = Q6_Vsf_vmax_VsfVsf(sqrt_v, epsilon_v);
+    HVX_Vector scale_v   = hvx_vec_inverse_f32(denom_v);
+
+    #pragma unroll(4)
+    for (int i = 0; i < nvec; i++) {
+        HVX_VectorPair p = hvx_vec_f16_to_f32(v_src[i]);
+        HVX_Vector r0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(p), scale_v));
+        HVX_Vector r1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_V_hi_W(p), scale_v));
+        v_dst[i] = hvx_vec_f32_to_f16(r0, r1);
+    }
+
+    if (nloe > 0) {
+        HVX_VectorPred bmask = Q6_Q_vsetq_R(nloe * SIZEOF_FP16);
+        HVX_Vector v1 = Q6_V_vand_QV(bmask, v_src[nvec]);
+        HVX_VectorPair p = hvx_vec_f16_to_f32(v1);
+        HVX_Vector r0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(p), scale_v));
+        HVX_Vector r1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vmpy_VsfVsf(Q6_V_hi_W(p), scale_v));
+        HVX_Vector result = hvx_vec_f32_to_f16(r0, r1);
+        hvx_vec_store_a(&v_dst[nvec], nloe * SIZEOF_FP16, result);
+    }
+}
+
 #endif // HVX_NORM_H
@@ -130,4 +130,70 @@ static inline void hvx_scale_offset_f32(uint8_t * restrict dst, const uint8_t *
    }
 }

+// Scale+offset computed by promoting f16 -> f32, then narrowing the result back to f16.
+#define hvx_scale_offset_f16_loop_body(dst_type, src_type, vec_store)                                                \
+    do {                                                                                                             \
+        dst_type * restrict vdst = (dst_type *) dst;                                                                 \
+        src_type * restrict vsrc = (src_type *) src;                                                                 \
+                                                                                                                     \
+        HVX_Vector vs = hvx_vec_splat_f32(scale);                                                                    \
+        HVX_Vector vo = hvx_vec_splat_f32(offset);                                                                   \
+                                                                                                                     \
+        const uint32_t nvec = n / VLEN_FP16;                                                                         \
+        const uint32_t nloe = n % VLEN_FP16;                                                                         \
+                                                                                                                     \
+        uint32_t i = 0;                                                                                              \
+                                                                                                                     \
+        _Pragma("unroll(4)")                                                                                         \
+        for (; i < nvec; ++i) {                                                                                      \
+            HVX_VectorPair p = hvx_vec_f16_to_f32(vsrc[i]);                                                          \
+            HVX_Vector r0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(p), vs), vo)); \
+            HVX_Vector r1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vmpy_VsfVsf(Q6_V_hi_W(p), vs), vo)); \
+            vdst[i] = hvx_vec_f32_to_f16(r0, r1);                                                                    \
+        }                                                                                                            \
+        if (nloe) {                                                                                                  \
+            HVX_VectorPair p = hvx_vec_f16_to_f32(vsrc[i]);                                                          \
+            HVX_Vector r0 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vmpy_VsfVsf(Q6_V_lo_W(p), vs), vo)); \
+            HVX_Vector r1 = Q6_Vsf_equals_Vqf32(Q6_Vqf32_vadd_Vqf32Vsf(Q6_Vqf32_vmpy_VsfVsf(Q6_V_hi_W(p), vs), vo)); \
+            HVX_Vector v = hvx_vec_f32_to_f16(r0, r1);                                                               \
+            vec_store((void *) &vdst[i], nloe * SIZEOF_FP16, v);                                                     \
+        }                                                                                                            \
+    } while(0)
+
+static inline void hvx_scale_offset_f16_aa(uint8_t * restrict dst, const uint8_t * restrict src, const int n, const float scale, const float offset) {
+    assert((size_t) dst % 128 == 0);
+    assert((size_t) src % 128 == 0);
+    hvx_scale_offset_f16_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+}
+
+static inline void hvx_scale_offset_f16_au(uint8_t * restrict dst, const uint8_t * restrict src, const int n, const float scale, const float offset) {
+    assert((size_t) dst % 128 == 0);
+    hvx_scale_offset_f16_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a);
+}
+
+static inline void hvx_scale_offset_f16_ua(uint8_t * restrict dst, const uint8_t * restrict src, const int n, const float scale, const float offset) {
+    assert((size_t) src % 128 == 0);
+    hvx_scale_offset_f16_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
+}
+
+static inline void hvx_scale_offset_f16_uu(uint8_t * restrict dst, const uint8_t * restrict src, const int n, const float scale, const float offset) {
+    hvx_scale_offset_f16_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
+}
+
+static inline void hvx_scale_offset_f16(uint8_t * restrict dst, const uint8_t * restrict src, const int n, const float scale, const float offset) {
+    if (((size_t) dst & 127) == 0) {
+        if (((size_t) src & 127) == 0) {
+            hvx_scale_offset_f16_aa(dst, src, n, scale, offset);
+        } else {
+            hvx_scale_offset_f16_au(dst, src, n, scale, offset);
+        }
+    } else {
+        if (((size_t) src & 127) == 0) {
+            hvx_scale_offset_f16_ua(dst, src, n, scale, offset);
+        } else {
+            hvx_scale_offset_f16_uu(dst, src, n, scale, offset);
+        }
+    }
+}
+
 #endif // HVX_SCALE_H
@@ -123,4 +123,67 @@ static inline void hvx_sqrt_f32(uint8_t * restrict dst, const uint8_t * restrict
    }
 }

+// Compute sqrt(x) for f16 by promoting to f32, applying hvx_vec_rsqrt_f32, and narrowing back.
+#define hvx_sqrt_f16_loop_body(dst_type, src_type, vec_store)                          \
+    do {                                                                               \
+        dst_type * restrict vdst = (dst_type *) dst;                                   \
+        src_type * restrict vsrc = (src_type *) src;                                   \
+                                                                                       \
+        const uint32_t nvec = n / VLEN_FP16;                                           \
+        const uint32_t nloe = n % VLEN_FP16;                                           \
+                                                                                       \
+        uint32_t i = 0;                                                                \
+                                                                                       \
+        _Pragma("unroll(4)")                                                           \
+        for (; i < nvec; i++) {                                                        \
+            HVX_VectorPair p = hvx_vec_f16_to_f32(vsrc[i]);                            \
+            HVX_Vector r0 = HVX_OP_MUL(hvx_vec_rsqrt_f32(Q6_V_lo_W(p)), Q6_V_lo_W(p)); \
+            HVX_Vector r1 = HVX_OP_MUL(hvx_vec_rsqrt_f32(Q6_V_hi_W(p)), Q6_V_hi_W(p)); \
+            vdst[i] = hvx_vec_f32_to_f16(r0, r1);                                      \
+        }                                                                              \
+        if (nloe) {                                                                    \
+            HVX_VectorPair p = hvx_vec_f16_to_f32(vsrc[i]);                            \
+            HVX_Vector r0 = HVX_OP_MUL(hvx_vec_rsqrt_f32(Q6_V_lo_W(p)), Q6_V_lo_W(p)); \
+            HVX_Vector r1 = HVX_OP_MUL(hvx_vec_rsqrt_f32(Q6_V_hi_W(p)), Q6_V_hi_W(p)); \
+            HVX_Vector v = hvx_vec_f32_to_f16(r0, r1);                                 \
+            vec_store((void *) &vdst[i], nloe * SIZEOF_FP16, v);                       \
+        }                                                                              \
+    } while(0)
+
+static inline void hvx_sqrt_f16_aa(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    assert((unsigned long) src % 128 == 0);
+    hvx_sqrt_f16_loop_body(HVX_Vector, HVX_Vector, hvx_vec_store_a);
+}
+
+static inline void hvx_sqrt_f16_au(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) dst % 128 == 0);
+    hvx_sqrt_f16_loop_body(HVX_Vector, HVX_UVector, hvx_vec_store_a);
+}
+
+static inline void hvx_sqrt_f16_ua(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    assert((unsigned long) src % 128 == 0);
+    hvx_sqrt_f16_loop_body(HVX_UVector, HVX_Vector, hvx_vec_store_u);
+}
+
+static inline void hvx_sqrt_f16_uu(uint8_t * restrict dst, const uint8_t * restrict src, uint32_t n) {
+    hvx_sqrt_f16_loop_body(HVX_UVector, HVX_UVector, hvx_vec_store_u);
+}
+
+static inline void hvx_sqrt_f16(uint8_t * restrict dst, const uint8_t * restrict src, const int num_elems) {
+    if ((unsigned long) dst % 128 == 0) {
+        if ((unsigned long) src % 128 == 0) {
+            hvx_sqrt_f16_aa(dst, src, num_elems);
+        } else {
+            hvx_sqrt_f16_au(dst, src, num_elems);
+        }
+    } else {
+        if ((unsigned long) src % 128 == 0) {
+            hvx_sqrt_f16_ua(dst, src, num_elems);
+        } else {
+            hvx_sqrt_f16_uu(dst, src, num_elems);
+        }
+    }
+}
+
 #endif /* HVX_SQRT_H */
@@ -753,6 +753,9 @@ static int execute_op(struct htp_ops_context * octx) {
        case HTP_OP_MUL_MAT_ID:
            return op_matmul_id(octx);

+        case HTP_OP_MUL_MAT_ID_NX:
+            return op_matmul_id_nx(octx);
+
        case HTP_OP_MUL_MAT_NX:
            return op_matmul_nx(octx);

@@ -878,8 +881,8 @@ static inline void drop_mmap(struct htp_context *ctx, struct htp_mmap *m) {
    }
 }

-static inline void mmap_buf(struct htp_context *ctx, struct htp_buf_desc *b) {
-    if (b->base) return; // already mapped
+static inline bool mmap_buf(struct htp_context *ctx, struct htp_buf_desc *b) {
+    if (b->base) return true; // already mapped

    // find unused mapping
    for (uint32_t i=0; i < HTP_MAX_MMAPS; i++) {
@@ -887,8 +890,8 @@ static inline void mmap_buf(struct htp_context *ctx, struct htp_buf_desc *b) {
        if (!m->size) {
            void *va = htp_mmap(b->fd, b->size);
            if (va == NULL) {
-                FARF(ERROR, "mmap failed : fd %u size %u", b->fd, (uint32_t) b->size);
-                abort(); // can't do much else at this point
+                FARF(HIGH, "mmap failed (will attempt defrag) : fd %u size %u", b->fd, (uint32_t) b->size);
+                return false;
            }

            m->base   = b->base = (uint64_t) va;
@@ -896,12 +899,12 @@ static inline void mmap_buf(struct htp_context *ctx, struct htp_buf_desc *b) {
            m->size   = b->size;

            FARF(ALWAYS, "mmap : fd %u base %p size %u", m->fd, (void*) m->base, (uint32_t) m->size);
-            return;
+            return true;
        }
    }

    FARF(ERROR, "mmap failed : exceeded mapping capacity limit of %u", HTP_MAX_MMAPS);
-    abort();
+    return false;
 }

 static void prep_op_bufs(struct htp_context *ctx, struct htp_buf_desc *bufs, uint32_t n_bufs) {
@@ -934,12 +937,32 @@ static void prep_op_bufs(struct htp_context *ctx, struct htp_buf_desc *bufs, uin
        }
    }

-    // Create missing mappings
+    // Create missing mappings (pass 1)
+    bool mmap_ok = true;
    for (uint32_t i=0; i < n_bufs; i++) {
        struct htp_buf_desc *b = bufs + i;
-        mmap_buf(ctx, b);
+        if (!mmap_buf(ctx, b)) {
+            mmap_ok = false;
+            break;
+        }
        FARF(HIGH, "prep-buf #%u : pass1 fd %u base %p size %u flags 0x%x", i, b->fd, (void*) b->base, (uint32_t) b->size, b->flags);
    }
+
+    if (!mmap_ok) {
+        // Attempt clean defragmentation: drop all mappings and remap (pass 2)
+        FARF(HIGH, "prep-bufs : dropping all mappings to defragment address space");
+        for (uint32_t i=0; i < HTP_MAX_MMAPS; i++) { drop_mmap(ctx, ctx->mmap + i); }
+
+        for (uint32_t i=0; i < n_bufs; i++) {
+            struct htp_buf_desc *b = bufs + i;
+            b->base = 0;
+            if (!mmap_buf(ctx, b)) {
+                FARF(ERROR, "prep-bufs : mmap failed after defragmentation (fd %u size %u)", b->fd, (uint32_t) b->size);
+                abort();
+            }
+            FARF(HIGH, "prep-buf #%u : pass2 fd %u base %p size %u flags 0x%x", i, b->fd, (void*) b->base, (uint32_t) b->size, b->flags);
+        }
+    }
 }

 static void prep_tensor(struct htp_context *ctx, struct htp_buf_desc *bufs, struct htp_tensor *tens, uint32_t idx, struct htp_tensor *t) {
@@ -981,7 +1004,7 @@ static int proc_op_req(struct htp_ops_context * octx, struct htp_tensor *tens, u
        octx->src_dma[i] = octx->ctx->dma; // FIXME: ? octx->ctx->dma_cached : octx->ctx->dma;

        FARF(HIGH, "prep-src #%u: data %p size %u : %u:%u:%u:%u", op->src[i], (void*) src->data, src->size,
-            src->ne[0], src->ne[1], src->ne[3], src->ne[3]);
+            src->ne[0], src->ne[1], src->ne[2], src->ne[3]);
    }

    htp_tensor_flush_all(octx->ctx, octx->src, HTP_OP_MAX_INPUTS);
@@ -134,7 +134,8 @@ static inline int htp_mm_hmx_compute_chunks(size_t   vtcm_total,
    size_t best_mn   = 0;
    size_t best_m = 0, best_n = 0;

-    const size_t n_max = hex_align_down((size_t)n, HTP_MM_HMX_TILE_N_COLS);
+    const size_t max_nc_budget = (usable / per_n_cost);
+    const size_t n_max = hex_align_down(hex_smin((size_t)n, max_nc_budget), HTP_MM_HMX_TILE_N_COLS);
    for (size_t nc = n_max; nc >= HTP_MM_HMX_TILE_N_COLS; nc -= HTP_MM_HMX_TILE_N_COLS) {
        size_t n_fixed = 0, ncmn = 0, mc_denom = 0;
        if (hex_mul_overflow(nc, per_n_cost, &n_fixed)) continue;
@@ -299,6 +300,15 @@ static inline void htp_mm_hmx_get_batched_chunk_costs(
    *size_per_mn_out = sizeof(uint16_t);
 }

+static inline size_t htp_mm_hmx_get_2d_overhead(bool pipeline, bool is_matmul_id) {
+    size_t num_regions = pipeline ? 7 : (is_matmul_id ? 4 : 5);
+    return num_regions * HTP_MM_HMX_TILE_SIZE + 256;
+}
+
+static inline size_t htp_mm_hmx_get_batched_overhead(void) {
+    return 5 * HTP_MM_HMX_TILE_SIZE + 256;
+}
+
 struct htp_mm_hmx_vtcm_layout {
    // Byte offsets from vtcm_base for each region
    size_t off_weight[2];     // [1] is only used when pipelined
@@ -568,10 +578,8 @@ static inline void htp_mm_hvx_vtcm_layout_build(
                }

                size_t quant_scratch_size_per_thread = htp_mm_round_up(ne10 * sizeof(float), QK_Q8_0_TILED * sizeof(float));
-                size_t dst_size_per_thread = dst_nrows > 0 ? htp_mm_round_up(dst_row_size, 128) : 0;
-                if (dst_size_per_thread < quant_scratch_size_per_thread) {
-                    dst_size_per_thread = quant_scratch_size_per_thread;
-                }
+                size_t dst_slice_per_thread = (dst_nrows > 0 && src1_nrows == 1) ? htp_mm_round_up((dst_row_size + n_threads - 1) / n_threads, 128) : 0;
+                size_t dst_size_per_thread = (dst_slice_per_thread > quant_scratch_size_per_thread) ? dst_slice_per_thread : quant_scratch_size_per_thread;
                dst_sz = dst_size_per_thread * n_threads;
                break;
            }
@@ -592,10 +600,8 @@ static inline void htp_mm_hvx_vtcm_layout_build(
                }

                size_t quant_scratch_size_per_thread = htp_mm_round_up(ne10 * sizeof(float), QK_Q8_0_TILED * sizeof(float));
-                size_t dst_size_per_thread = dst_nrows > 0 ? htp_mm_round_up(dst_row_size, 128) : 0;
-                if (dst_size_per_thread < quant_scratch_size_per_thread) {
-                    dst_size_per_thread = quant_scratch_size_per_thread;
-                }
+                size_t dst_slice_per_thread = dst_nrows > 0 ? htp_mm_round_up((dst_row_size + n_threads - 1) / n_threads, 128) : 0;
+                size_t dst_size_per_thread = (dst_slice_per_thread > quant_scratch_size_per_thread) ? dst_slice_per_thread : quant_scratch_size_per_thread;
                dst_sz = dst_size_per_thread * n_threads;
                break;
            }
@@ -658,7 +664,7 @@ static inline bool htp_mm_hmx_solve_batched_params(

    int act_threads = n_threads;
    while (act_threads >= 1) {
-        size_t group_overhead = 256;
+        size_t group_overhead = htp_mm_hmx_get_batched_overhead();
        size_t group_size_per_n, group_size_per_m, group_size_per_mn;
        htp_mm_hmx_get_batched_chunk_costs(k, group_size, &group_size_per_n, &group_size_per_m, &group_size_per_mn);

@@ -725,7 +731,7 @@ static inline bool htp_mm_hmx_solve_2d_params(

    int act_threads = n_threads;
    while (act_threads >= 1) {
-        size_t simple_2d_overhead = 256;
+        size_t simple_2d_overhead = htp_mm_hmx_get_2d_overhead(pipeline, is_matmul_id);
        size_t simple_2d_size_per_n, simple_2d_size_per_m, simple_2d_size_per_mn;
        htp_mm_hmx_get_2d_chunk_costs(wtype, k, pipeline, aligned_tile_size, &simple_2d_size_per_n, &simple_2d_size_per_m, &simple_2d_size_per_mn);

@@ -216,6 +216,14 @@ int op_set_rows(struct htp_ops_context * octx) {
        default:            return HTP_STATUS_NO_SUPPORT;
    }

+    FARF(HIGH, "set-rows: (%ux%ux%ux%u) x (%ux%ux%ux%u) -> (%ux%ux%ux%u) : src0-vtcm-size %zu dst-vtcm-size %zu n_threads %d\n",
+         octx->src[0]->ne[0], octx->src[0]->ne[1], octx->src[0]->ne[2], octx->src[0]->ne[3],
+         octx->src[1]->ne[0], octx->src[1]->ne[1], octx->src[1]->ne[2], octx->src[1]->ne[3],
+         octx->dst->ne[0], octx->dst->ne[1], octx->dst->ne[2], octx->dst->ne[3],
+         srctx.vtcm_layout.src0_bytes_per_thread * kparams->n_threads,
+         srctx.vtcm_layout.dst_bytes_per_thread  * kparams->n_threads,
+         kparams->n_threads);
+
    work_queue_run(octx->ctx->work_queue, q_func, &srctx, kparams->n_threads);

    return HTP_STATUS_OK;
@@ -234,6 +234,146 @@ static void sqrt_f32(const float * restrict src,
    }
 }

+static void scale_f16(const _Float16 * restrict src,
+                      _Float16 * restrict dst,
+                      const uint32_t num_rows,
+                      const struct htp_unary_context * uctx) {
+    htp_unary_op_preamble;
+    float scale = 0.f;
+    float bias  = 0.f;
+    memcpy(&scale, &op_params[0], sizeof(float));
+    memcpy(&bias,  &op_params[1], sizeof(float));
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * src0_row_size_aligned);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * dst_row_size_aligned);
+
+        hvx_scale_offset_f16_aa((uint8_t *) dst_local, (const uint8_t *) src_local, ne0, scale, bias);
+    }
+}
+
+static void clamp_f16(const _Float16 * restrict src,
+                      _Float16 * restrict dst,
+                      const uint32_t num_rows,
+                      const struct htp_unary_context * uctx) {
+    htp_unary_op_preamble;
+    float min = 0.f;
+    float max = 0.f;
+    memcpy(&min, &op_params[0], sizeof(float));
+    memcpy(&max, &op_params[1], sizeof(float));
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * src0_row_size_aligned);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * dst_row_size_aligned);
+
+        hvx_clamp_scalar_f16(dst_local, src_local, (_Float16) min, (_Float16) max, ne0);
+    }
+}
+
+static void rms_norm_f16(const _Float16 * restrict src,
+                         _Float16 * restrict dst,
+                         const uint32_t num_rows,
+                         const struct htp_unary_context * uctx) {
+    htp_unary_op_preamble;
+    float epsilon = 0.f;
+    memcpy(&epsilon, op_params, sizeof(float));
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * src0_row_size_aligned);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * dst_row_size_aligned);
+
+        hvx_fast_rms_norm_f16((const uint8_t *) src_local, (uint8_t *) dst_local, ne0, epsilon);
+    }
+}
+
+static void norm_f16(const _Float16 * restrict src,
+                     _Float16 * restrict dst,
+                     const uint32_t num_rows,
+                     const struct htp_unary_context * uctx) {
+    htp_unary_op_preamble;
+    float epsilon = 0.f;
+    memcpy(&epsilon, op_params, sizeof(float));
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * src0_row_size_aligned);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * dst_row_size_aligned);
+
+        hvx_fast_norm_f16((const uint8_t *) src_local, (uint8_t *) dst_local, ne0, epsilon);
+    }
+}
+
+static void sqr_f16(const _Float16 * restrict src,
+                    _Float16 * restrict dst,
+                    const uint32_t num_rows,
+                    const struct htp_unary_context * uctx) {
+    htp_unary_op_preamble;
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * src0_row_size_aligned);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * dst_row_size_aligned);
+
+        hvx_sqr_f16_aa((uint8_t *) dst_local, (const uint8_t *) src_local, ne0);
+    }
+}
+
+static void sqrt_f16(const _Float16 * restrict src,
+                     _Float16 * restrict dst,
+                     const uint32_t num_rows,
+                     const struct htp_unary_context * uctx) {
+    htp_unary_op_preamble;
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * src0_row_size_aligned);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * dst_row_size_aligned);
+
+        hvx_sqrt_f16_aa((uint8_t *) dst_local, (const uint8_t *) src_local, ne0);
+    }
+}
+
+static void abs_f16(const _Float16 * restrict src,
+                    _Float16 * restrict dst,
+                    const uint32_t num_rows,
+                    const struct htp_unary_context * uctx) {
+    htp_unary_op_preamble;
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * src0_row_size_aligned);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * dst_row_size_aligned);
+
+        hvx_abs_f16_aa((uint8_t *) dst_local, (const uint8_t *) src_local, ne0);
+    }
+}
+
+static void log_f16(const _Float16 * restrict src,
+                    _Float16 * restrict dst,
+                    const uint32_t num_rows,
+                    const struct htp_unary_context * uctx) {
+    htp_unary_op_preamble;
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_local = (const uint8_t *)src + (ir * src0_row_size_aligned);
+        uint8_t * restrict dst_local       = (uint8_t *)dst + (ir * dst_row_size_aligned);
+
+        hvx_log_f16_aa((uint8_t *) dst_local, (const uint8_t *) src_local, ne0);
+    }
+}
+
+static void l2_norm_f16(const _Float16 * restrict src,
+                        _Float16 * restrict dst,
+                        const uint32_t num_rows,
+                        const struct htp_unary_context * uctx) {
+    htp_unary_op_preamble;
+    float epsilon = 0.f;
+    memcpy(&epsilon, op_params, sizeof(float));
+
+    for (uint32_t ir = 0; ir < num_rows; ir++) {
+        const uint8_t * restrict src_f = (const uint8_t *)src + (ir * src0_row_size_aligned);
+        uint8_t * restrict dst_f       = (uint8_t *)dst + (ir * dst_row_size_aligned);
+
+        hvx_fast_l2_norm_f16((const uint8_t *)src_f, (uint8_t *)dst_f, ne0, epsilon);
+    }
+}
+
 static void neg_f32(const float * restrict src,
                    float * restrict dst,
                    const uint32_t num_rows,
@@ -471,8 +611,8 @@ static void log_f32(const float * restrict src,
    }
 }

-#define DEFINE_UNARY_TASK(NAME, IS_RMS_NORM_MUL, IS_TRI, CORE_EXPR)                                                 \
-static void unary_task_f32_##NAME(unsigned int nth, unsigned int ith, void * data) {                                \
+#define DEFINE_UNARY_TASK_IMPL(NAME, TYPE, SUFFIX, IS_RMS_NORM_MUL, IS_TRI, CORE_EXPR)                              \
+static void unary_task_##SUFFIX##_##NAME(unsigned int nth, unsigned int ith, void * data) {                         \
    const struct htp_unary_context * uctx = (const struct htp_unary_context *) data;                                \
    struct htp_ops_context * octx = uctx->octx;                                                                     \
    const struct htp_tensor * src = octx->src[0];                                                                   \
@@ -536,7 +676,7 @@ static void unary_task_f32_##NAME(unsigned int nth, unsigned int ith, void * dat
    const uint32_t dst_max_block  = block_dst_contig  ? uctx->block : MIN((uint32_t)uctx->block, ne1);              \
    const uint32_t BLOCK = MIN(src0_max_block, dst_max_block);                                                      \
    if (BLOCK == 0) {                                                                                               \
-        FARF(ERROR, "unary-f32 : current VTCM reservation %zu is too small, needed at least %zu\n",                 \
+        FARF(ERROR, "unary-" #SUFFIX " : current VTCM reservation %zu is too small, needed at least %zu\n",         \
             uctx->vtcm_src0_size_per_thread, src0_row_size_aligned);                                               \
        return;                                                                                                     \
    }                                                                                                               \
@@ -578,11 +718,11 @@ static void unary_task_f32_##NAME(unsigned int nth, unsigned int ith, void * dat
        const uint32_t block_size = unary_block_size(ir, src0_end_row, BLOCK, block_src0_contig, block_dst_contig,  \
                                                     ne01, div_ne01);                                               \
                                                                                                                    \
-        float * dst_vtcm  = (float *) dma_queue_pop(dma_queue).src;                                                 \
-        float * src0_vtcm = (float *) dma_queue_pop(dma_queue).dst;                                                 \
-        float * src1_vtcm = NULL;                                                                                   \
+        TYPE * dst_vtcm  = (TYPE *) dma_queue_pop(dma_queue).src;                                                   \
+        TYPE * src0_vtcm = (TYPE *) dma_queue_pop(dma_queue).dst;                                                   \
+        TYPE * src1_vtcm = NULL;                                                                                    \
        if ((IS_RMS_NORM_MUL) && !uctx->broadcast_weight) {                                                         \
-            src1_vtcm = (float *) dma_queue_pop(dma_queue).dst;                                                     \
+            src1_vtcm = (TYPE *) dma_queue_pop(dma_queue).dst;                                                      \
        }                                                                                                           \
                                                                                                                    \
        htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, ir);                                                      \
@@ -625,6 +765,10 @@ static void unary_task_f32_##NAME(unsigned int nth, unsigned int ith, void * dat
    dma_queue_flush(dma_queue);                                                                                     \
 }

+// F32 unary task: row-block DMA/VTCM plumbing, float-typed VTCM buffers.
+#define DEFINE_UNARY_TASK(NAME, IS_RMS_NORM_MUL, IS_TRI, CORE_EXPR) \
+    DEFINE_UNARY_TASK_IMPL(NAME, float, f32, IS_RMS_NORM_MUL, IS_TRI, CORE_EXPR)
+
 DEFINE_UNARY_TASK(norm,           false, false, norm_f32(src0_vtcm, dst_vtcm, block_size, uctx))
 DEFINE_UNARY_TASK(rms_norm,       false, false, rms_norm_f32(src0_vtcm, dst_vtcm, block_size, uctx))
 DEFINE_UNARY_TASK(rms_norm_mul,   true,  false, rms_norm_mul_f32(src0_vtcm, uctx->broadcast_weight ? (const float *) src1_vtcm_data : src1_vtcm, dst_vtcm, block_size, uctx))
@@ -644,6 +788,18 @@ DEFINE_UNARY_TASK(unary_log,      false, false, log_f32(src0_vtcm, dst_vtcm, blo
 DEFINE_UNARY_TASK(l2_norm,        false, false, l2_norm_f32(src0_vtcm, dst_vtcm, block_size, uctx))
 DEFINE_UNARY_TASK(tri,            false, true,  tri_f32(src0_vtcm, dst_vtcm, block_size, ir, uctx))

+// F16 unary tasks: same DMA/VTCM plumbing as DEFINE_UNARY_TASK, but VTCM buffers are
+// _Float16-typed. None of the current F16 ops need RMS_NORM_MUL or TRI support.
+DEFINE_UNARY_TASK_IMPL(norm,      _Float16, f16, false, false, norm_f16(src0_vtcm, dst_vtcm, block_size, uctx))
+DEFINE_UNARY_TASK_IMPL(rms_norm,  _Float16, f16, false, false, rms_norm_f16(src0_vtcm, dst_vtcm, block_size, uctx))
+DEFINE_UNARY_TASK_IMPL(scale,     _Float16, f16, false, false, scale_f16(src0_vtcm, dst_vtcm, block_size, uctx))
+DEFINE_UNARY_TASK_IMPL(clamp,     _Float16, f16, false, false, clamp_f16(src0_vtcm, dst_vtcm, block_size, uctx))
+DEFINE_UNARY_TASK_IMPL(sqr,       _Float16, f16, false, false, sqr_f16(src0_vtcm, dst_vtcm, block_size, uctx))
+DEFINE_UNARY_TASK_IMPL(sqrt,      _Float16, f16, false, false, sqrt_f16(src0_vtcm, dst_vtcm, block_size, uctx))
+DEFINE_UNARY_TASK_IMPL(l2_norm,   _Float16, f16, false, false, l2_norm_f16(src0_vtcm, dst_vtcm, block_size, uctx))
+DEFINE_UNARY_TASK_IMPL(unary_abs, _Float16, f16, false, false, abs_f16(src0_vtcm, dst_vtcm, block_size, uctx))
+DEFINE_UNARY_TASK_IMPL(unary_log, _Float16, f16, false, false, log_f16(src0_vtcm, dst_vtcm, block_size, uctx))
+
 // Apply a pointwise unary op to one column tile that is already in VTCM.
 #define DEFINE_UNARY_TILED_TASK(NAME, IS_TRI, CORE_TILE_EXPR)                                                       \
 static void unary_task_f32_tiled_##NAME(unsigned int nth, unsigned int ith, void * data) {                          \
@@ -892,50 +1048,76 @@ DEFINE_UNARY_TILED_TASK(unary_abs,      false, hvx_abs_f32_aa(dst_vtcm, src_vtcm
 DEFINE_UNARY_TILED_TASK(unary_log,      false, hvx_log_f32_aa(dst_vtcm, src_vtcm, tw))
 DEFINE_UNARY_TILED_TASK(tri,            true,  tri_apply_tile_f32(src_vtcm, dst_vtcm, tw, col, i01, ne0, tri_ttype))

-static int execute_op_unary_f32(struct htp_ops_context * octx) {
+static int execute_op_unary(struct htp_ops_context * octx) {
    int err = HTP_STATUS_OK;

    const struct htp_tensor * src0 = octx->src[0];
    const struct htp_tensor * dst  = octx->dst;

+    const bool is_f16 = (src0->type == HTP_TYPE_F16);
+
    const char * op_type = NULL;

    switch (octx->op) {
-        case HTP_OP_NORM:            op_type = "norm-f32";         break;
-        case HTP_OP_RMS_NORM:        op_type = "rmsnorm-f32";      break;
-        case HTP_OP_RMS_NORM_MUL:    op_type = "rmsnorm-mul-f32";  break;
-        case HTP_OP_SCALE:           op_type = "scale-f32";        break;
-        case HTP_OP_CLAMP:           op_type = "clamp-f32";        break;
-        case HTP_OP_SQR:             op_type = "sqr-f32";          break;
-        case HTP_OP_SQRT:            op_type = "sqrt-f32";         break;
-        case HTP_OP_UNARY_NEG:       op_type = "neg-f32";          break;
-        case HTP_OP_UNARY_EXP:       op_type = "exp-f32";          break;
-        case HTP_OP_UNARY_SIGMOID:   op_type = "sigmoid-f32";      break;
-        case HTP_OP_UNARY_SILU:      op_type = "silu-f32";         break;
-        case HTP_OP_UNARY_GELU:      op_type = "gelu-f32";         break;
-        case HTP_OP_UNARY_SOFTPLUS:  op_type = "softplus-f32";     break;
-        case HTP_OP_UNARY_TANH:      op_type = "tanh-f32";         break;
-        case HTP_OP_UNARY_ABS:       op_type = "abs-f32";          break;
-        case HTP_OP_UNARY_LOG:       op_type = "log-f32";          break;
-        case HTP_OP_L2_NORM:         op_type = "l2norm-f32";       break;
-        case HTP_OP_TRI:             op_type = "tri-f32";          break;
+        case HTP_OP_NORM:            op_type = is_f16 ? "norm-f16"     : "norm-f32";         break;
+        case HTP_OP_RMS_NORM:        op_type = is_f16 ? "rmsnorm-f16"  : "rmsnorm-f32";      break;
+        case HTP_OP_RMS_NORM_MUL:    op_type = "rmsnorm-mul-f32";                            break;
+        case HTP_OP_SCALE:           op_type = is_f16 ? "scale-f16"    : "scale-f32";        break;
+        case HTP_OP_CLAMP:           op_type = is_f16 ? "clamp-f16"    : "clamp-f32";        break;
+        case HTP_OP_SQR:             op_type = is_f16 ? "sqr-f16"      : "sqr-f32";          break;
+        case HTP_OP_SQRT:            op_type = is_f16 ? "sqrt-f16"     : "sqrt-f32";         break;
+        case HTP_OP_UNARY_NEG:       op_type = "neg-f32";                                    break;
+        case HTP_OP_UNARY_EXP:       op_type = "exp-f32";                                    break;
+        case HTP_OP_UNARY_SIGMOID:   op_type = "sigmoid-f32";                                break;
+        case HTP_OP_UNARY_SILU:      op_type = "silu-f32";                                   break;
+        case HTP_OP_UNARY_GELU:      op_type = "gelu-f32";                                   break;
+        case HTP_OP_UNARY_SOFTPLUS:  op_type = "softplus-f32";                               break;
+        case HTP_OP_UNARY_TANH:      op_type = "tanh-f32";                                   break;
+        case HTP_OP_UNARY_ABS:       op_type = is_f16 ? "abs-f16"      : "abs-f32";          break;
+        case HTP_OP_UNARY_LOG:       op_type = is_f16 ? "log-f16"      : "log-f32";          break;
+        case HTP_OP_L2_NORM:         op_type = is_f16 ? "l2norm-f16"   : "l2norm-f32";       break;
+        case HTP_OP_TRI:             op_type = "tri-f32";                                    break;

        default:
            FARF(ERROR, "Unsupported unary Op %u\n", octx->op);
            return HTP_STATUS_NO_SUPPORT;
    }

+    // F16 only has row-block kernels for this subset of ops (see the dispatch switch
+    // below) - reject everything else up front, before touching kparams/VTCM.
+    if (is_f16) {
+        switch (octx->op) {
+            case HTP_OP_NORM:
+            case HTP_OP_RMS_NORM:
+            case HTP_OP_SCALE:
+            case HTP_OP_CLAMP:
+            case HTP_OP_SQR:
+            case HTP_OP_SQRT:
+            case HTP_OP_L2_NORM:
+            case HTP_OP_UNARY_ABS:
+            case HTP_OP_UNARY_LOG:
+                break;
+            default:
+                FARF(ERROR, "unary-%s: not supported for F16\n", op_type);
+                return HTP_STATUS_NO_SUPPORT;
+        }
+    }
+
    const struct htp_unary_kernel_params * kparams = (const struct htp_unary_kernel_params *) octx->kernel_params;

    const uint32_t src0_nrows = src0->ne[1] * src0->ne[2] * src0->ne[3];
    const uint32_t n_threads  = kparams->n_threads;

-    const size_t src0_data_row_size = src0->ne[0] * sizeof(float);
-    const size_t dst_data_row_size  = dst->ne[0]  * sizeof(float);
+    const size_t elem_size = is_f16 ? sizeof(_Float16) : sizeof(float);
+
+    const size_t src0_data_row_size = src0->ne[0] * elem_size;
+    const size_t dst_data_row_size  = dst->ne[0]  * elem_size;

    const size_t src0_row_size_aligned = kparams->src0_row_size_aligned;
    const size_t dst_row_size_aligned  = kparams->dst_row_size_aligned;

+    // Always 0 for F16 - htp_unary_vtcm_layout_build() keeps F16 on the row-block path,
+    // since only F32 has unary_task_f32_tiled_* kernels.
    const uint32_t col_tile = kparams->col_tile;

    size_t src1_data_row_size = 0;
@@ -943,6 +1125,8 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
    bool broadcast_weight = kparams->broadcast_weight;
    const struct htp_tensor * src1 = NULL;

+    // RMS_NORM_MUL fusion is F32-only (its weight tensor is always F32; see
+    // try_fuse_node()'s type guard), so this never triggers when is_f16 is true.
    if (octx->op == HTP_OP_RMS_NORM_MUL) {
        src1 = octx->src[1];
        src1_data_row_size = src1->ne[0] * sizeof(float);
@@ -987,7 +1171,7 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {

            .block                 = kparams->block,
            .nc                    = src0->ne[0],
-            .col_tile              = (uint32_t) kparams->col_tile,
+            .col_tile              = col_tile,
            .broadcast_weight      = broadcast_weight,

            .vtcm_src0             = VTCM_LAYOUT_PTR(uint8_t, base, 0),
@@ -1020,6 +1204,19 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
                case HTP_OP_TRI:             task_func = unary_task_f32_tiled_tri;            break;
                default:                     break;
            }
+        } else if (is_f16) {
+            switch (octx->op) {
+                case HTP_OP_NORM:            task_func = unary_task_f16_norm;                 break;
+                case HTP_OP_RMS_NORM:        task_func = unary_task_f16_rms_norm;             break;
+                case HTP_OP_SCALE:           task_func = unary_task_f16_scale;                break;
+                case HTP_OP_CLAMP:           task_func = unary_task_f16_clamp;                break;
+                case HTP_OP_SQR:             task_func = unary_task_f16_sqr;                  break;
+                case HTP_OP_SQRT:            task_func = unary_task_f16_sqrt;                 break;
+                case HTP_OP_L2_NORM:         task_func = unary_task_f16_l2_norm;              break;
+                case HTP_OP_UNARY_ABS:       task_func = unary_task_f16_unary_abs;            break;
+                case HTP_OP_UNARY_LOG:       task_func = unary_task_f16_unary_log;            break;
+                default:                     break;
+            }
        } else {
            switch (octx->op) {
                case HTP_OP_NORM:            task_func = unary_task_f32_norm;                 break;
@@ -1047,7 +1244,7 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
        if (task_func) {
            worker_pool_run_func(octx->ctx->worker_pool, task_func, &uctx, n_threads);
        } else {
-            FARF(ERROR, "execute_op_unary_f32: task function is NULL for op %d\n", octx->op);
+            FARF(ERROR, "execute_op_unary: task function is NULL for op %d\n", octx->op);
            err = HTP_STATUS_NO_SUPPORT;
        }
    }
@@ -1058,7 +1255,8 @@ static int execute_op_unary_f32(struct htp_ops_context * octx) {
 int op_unary(struct htp_ops_context * octx) {
    switch (octx->src[0]->type) {
        case HTP_TYPE_F32:
-            return execute_op_unary_f32(octx);
+        case HTP_TYPE_F16:
+            return execute_op_unary(octx);

        default:
            return HTP_STATUS_NO_SUPPORT;
@@ -85,17 +85,19 @@ static inline void htp_unary_vtcm_layout_build(
    bool broadcast_weight,
    uint32_t n_threads,
    size_t vtcm_size,
+    size_t elem_size,
    uint32_t * out_col_tile,
    uint32_t * out_vtcm_row_per_thread
 ) {
-    const size_t src0_data_row_size = ne00 * sizeof(float);
-    const size_t dst_data_row_size  = ne10 * sizeof(float);
+    const size_t src0_data_row_size = ne00 * elem_size;
+    const size_t dst_data_row_size  = ne10 * elem_size;

    const size_t src0_row_size_aligned = hex_round_up(src0_data_row_size, 128);
    const size_t dst_row_size_aligned  = hex_round_up(dst_data_row_size,  128);

    size_t src1_row_size_aligned = 0;
    if (op == HTP_OP_RMS_NORM_MUL) {
+        // RMS_NORM_MUL fusion is F32-only; its weight tensor is always F32.
        const size_t src1_data_row_size = ne11 * sizeof(float);
        src1_row_size_aligned = hex_round_up(src1_data_row_size, 128);
    }
@@ -125,12 +127,19 @@ static inline void htp_unary_vtcm_layout_build(

    const bool is_reduction = (op == HTP_OP_NORM || op == HTP_OP_RMS_NORM ||
                               op == HTP_OP_RMS_NORM_MUL || op == HTP_OP_L2_NORM);
+    // The tiled fallback path below only has F32 task functions (unary_task_f32_tiled_*);
+    // F16 has no tiled kernels, so it must stay on the row-block path like reduction ops.
+    // NOTE: if F16 ends up with vtcm_row_per_thread == 0 here (row too large for the VTCM
+    // budget), execute_op_unary() will see BLOCK == 0 and skip computation for that op
+    // (logged via FARF(ERROR, ...)) since there is no F16 tiled fallback. This is a known
+    // limitation; supporting it would require adding F16 tiled kernels.
+    const bool is_f16 = (elem_size == sizeof(_Float16));
    uint32_t col_tile = 0;

-    if (vtcm_row_per_thread == 0 && !is_reduction) {
+    if (vtcm_row_per_thread == 0 && !is_reduction && !is_f16) {
        const size_t per_thread_budget = vtcm_size / n_threads;
        const size_t col_tile_bytes = hex_align_down(per_thread_budget / 4, 128);
-        col_tile = (uint32_t) (col_tile_bytes / sizeof(float));
+        col_tile = (uint32_t) (col_tile_bytes / elem_size);

        L->src0_bytes = col_tile_bytes * 2;
        L->dst_bytes  = col_tile_bytes * 2;
@@ -127,6 +127,18 @@ else()
        configure_file(${src} ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${src} COPYONLY)
    endforeach()

+    # CMAKE_OSX_SYSROOT is an SDK name or path - xcrun accepts both
+    set(METAL_SDK ${CMAKE_OSX_SYSROOT})
+    if (NOT METAL_SDK)
+        set(METAL_SDK macosx)
+    endif()
+
+    if (CMAKE_OSX_SYSROOT MATCHES "[Ss]imulator")
+        set(METAL_TARGET_SIM "-simulator")
+    else()
+        set(METAL_TARGET_SIM "")
+    endif()
+
    if (GGML_METAL_SHADER_DEBUG)
        # note: disabling fast math is needed in order to pass tests/test-backend-ops
        # note: adding -fno-inline fixes the tests when using MTL_SHADER_VALIDATION=1
@@ -138,9 +150,19 @@ else()
        set(XC_FLAGS -O3)
    endif()

+    execute_process(COMMAND xcrun -sdk ${METAL_SDK} --show-sdk-version OUTPUT_VARIABLE METAL_SDK_VERSION OUTPUT_STRIP_TRAILING_WHITESPACE)
+    if (METAL_SDK_VERSION VERSION_GREATER_EQUAL 26.0)
+        set(GGML_METAL_HAS_TENSOR_LIB ON)
+    else()
+        message(STATUS "Metal SDK ${METAL_SDK_VERSION} does not support the tensor API, skipping ggml-tensor.metallib")
+    endif()
+
    if (GGML_METAL_MACOSX_VERSION_MIN)
        message(STATUS "Adding  -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN} flag to metal compilation")
        list   (APPEND XC_FLAGS -mmacosx-version-min=${GGML_METAL_MACOSX_VERSION_MIN})
+    elseif (NOT GGML_METAL_TARGET_OS STREQUAL "macos" AND CMAKE_OSX_DEPLOYMENT_TARGET)
+        message(STATUS "Adding  -mtargetos=${GGML_METAL_TARGET_OS}${CMAKE_OSX_DEPLOYMENT_TARGET}${METAL_TARGET_SIM} flag to metal compilation")
+        list   (APPEND XC_FLAGS -mtargetos=${GGML_METAL_TARGET_OS}${CMAKE_OSX_DEPLOYMENT_TARGET}${METAL_TARGET_SIM})
    endif()

    if (GGML_METAL_STD)
@@ -156,33 +178,41 @@ else()
        list(APPEND AIR_FILES ${AIR})
        add_custom_command(
            OUTPUT ${AIR}
-            COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -I ${CMAKE_RUNTIME_OUTPUT_DIRECTORY} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${src} -o ${AIR}
+            COMMAND xcrun -sdk ${METAL_SDK} metal ${XC_FLAGS} -I ${CMAKE_RUNTIME_OUTPUT_DIRECTORY} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${src} -o ${AIR}
            DEPENDS ${src} kernels/common.h kernels/dequantize.h kernels/quantize.h ${METALLIB_COMMON} ggml-metal-impl.h
            COMMENT "Compiling ${src}"
            VERBATIM
        )
    endforeach()

-    # the tensor API kernels go in a separate metallib, loaded only where supported
-    set(AIR_MM_TENSOR "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/mul_mm_tensor.air")
-    add_custom_command(
-        OUTPUT ${AIR_MM_TENSOR}
-        COMMAND xcrun -sdk macosx metal ${XC_FLAGS} -DGGML_METAL_HAS_TENSOR -I ${CMAKE_RUNTIME_OUTPUT_DIRECTORY} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/mul_mm.metal -o ${AIR_MM_TENSOR}
-        DEPENDS kernels/mul_mm.metal kernels/common.h kernels/dequantize.h ${METALLIB_COMMON} ggml-metal-impl.h
-        COMMENT "Compiling kernels/mul_mm.metal (tensor API)"
-        VERBATIM
-    )
+    set(METALLIB_FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib)

-    add_custom_command(
-        OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-tensor.metallib
-        COMMAND xcrun -sdk macosx metallib ${AIR_MM_TENSOR} -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-tensor.metallib
-        DEPENDS ${AIR_MM_TENSOR}
-        COMMENT "Linking tensor API Metal kernels into ggml-tensor.metallib"
-    )
+    # the tensor API kernels go in a separate metallib, loaded only where supported
+    if (GGML_METAL_HAS_TENSOR_LIB)
+        set(AIR_MM_TENSOR "${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/mul_mm_tensor.air")
+        # the tensor API needs OS 26+
+        set(XC_FLAGS_TENSOR ${XC_FLAGS} -mtargetos=${GGML_METAL_TARGET_OS}26.0${METAL_TARGET_SIM})
+        add_custom_command(
+            OUTPUT ${AIR_MM_TENSOR}
+            COMMAND xcrun -sdk ${METAL_SDK} metal ${XC_FLAGS_TENSOR} -DGGML_METAL_HAS_TENSOR -I ${CMAKE_RUNTIME_OUTPUT_DIRECTORY} -c ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels/mul_mm.metal -o ${AIR_MM_TENSOR}
+            DEPENDS kernels/mul_mm.metal kernels/common.h kernels/dequantize.h ${METALLIB_COMMON} ggml-metal-impl.h
+            COMMENT "Compiling kernels/mul_mm.metal (tensor API)"
+            VERBATIM
+        )
+
+        add_custom_command(
+            OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-tensor.metallib
+            COMMAND xcrun -sdk ${METAL_SDK} metallib ${AIR_MM_TENSOR} -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-tensor.metallib
+            DEPENDS ${AIR_MM_TENSOR}
+            COMMENT "Linking tensor API Metal kernels into ggml-tensor.metallib"
+        )
+
+        list(APPEND METALLIB_FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-tensor.metallib)
+    endif()

    add_custom_command(
        OUTPUT ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-        COMMAND xcrun -sdk macosx metallib ${AIR_FILES} -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
+        COMMAND xcrun -sdk ${METAL_SDK} metallib ${AIR_FILES} -o ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
        COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-common.h
        COMMAND rm -f ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-metal-impl.h
        COMMAND rm -rf ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/kernels
@@ -192,8 +222,7 @@ else()

    add_custom_target(
        ggml-metal-lib ALL
-        DEPENDS ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-                ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-tensor.metallib
+        DEPENDS ${METALLIB_FILES}
    )
 endif() # GGML_METAL_EMBED_LIBRARY

@@ -205,8 +234,7 @@ if (NOT GGML_METAL_EMBED_LIBRARY)
    )

    install(
-        FILES ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/default.metallib
-              ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/ggml-tensor.metallib
+        FILES ${METALLIB_FILES}
        DESTINATION ${CMAKE_INSTALL_BINDIR}
    )
 endif()
@@ -69,6 +69,10 @@ struct ggml_metal {
    // extra command buffers for things like getting, setting and copying tensors
    NSMutableArray * cmd_bufs_ext;

+    // buffers to release after async Metal operations complete
+    // if Metal released them, it would do so on a Metal-internal thread without an autorelease pool, which could cause leaks
+    NSMutableArray * buf_refs;
+
    // the last command buffer queued into the Metal queue with operations relevant to the current Metal backend
    id<MTLCommandBuffer> cmd_buf_last;

@@ -179,6 +183,7 @@ ggml_metal_t ggml_metal_init(ggml_metal_device_t dev) {
        }

        res->cmd_bufs_ext = [[NSMutableArray alloc] init];
+        res->buf_refs     = [[NSMutableArray alloc] init];

        res->cmd_buf_last = nil;

@@ -206,6 +211,11 @@ void ggml_metal_free(ggml_metal_t ctx) {
    [ctx->cmd_bufs_ext removeAllObjects];
    [ctx->cmd_bufs_ext release];

+    @autoreleasepool {
+        [ctx->buf_refs removeAllObjects];
+        [ctx->buf_refs release];
+    }
+
    if (ctx->pipelines_ext) {
        ggml_metal_pipelines_free(ctx->pipelines_ext);
        ctx->pipelines_ext = nil;
@@ -294,6 +304,10 @@ void ggml_metal_synchronize(ggml_metal_t ctx) {

        [ctx->cmd_bufs_ext removeAllObjects];
    }
+
+    @autoreleasepool {
+        [ctx->buf_refs removeAllObjects];
+    }
 }

 static struct ggml_metal_buffer_id ggml_metal_get_buffer_id(const struct ggml_tensor * t) {
@@ -337,6 +351,8 @@ void ggml_metal_set_tensor_async(ggml_metal_t ctx, struct ggml_tensor * tensor,

        [encoder endEncoding];
        [cmd_buf commit];
+
+        [ctx->buf_refs addObject:buf_src];
        [buf_src release];

        // do not wait here for completion
@@ -381,6 +397,8 @@ void ggml_metal_get_tensor_async(ggml_metal_t ctx, const struct ggml_tensor * te

        [encoder endEncoding];
        [cmd_buf commit];
+
+        [ctx->buf_refs addObject:buf_dst];
        [buf_dst release];

        // do not wait here for completion
@@ -1577,6 +1577,26 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext(
    return res;
 }

+ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_vec_idx(
+        ggml_metal_library_t lib,
+        const ggml_tensor * op) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+    assert(op->src[3]);
+
+    char name[256];
+
+    snprintf(name, 256, "kernel_flash_attn_ext_vec_idx");
+
+    ggml_metal_pipeline_with_params res = ggml_metal_library_get_pipeline(lib, name);
+    if (!res.pipeline) {
+        res = ggml_metal_library_compile_pipeline(lib, name, name, nullptr);
+    }
+
+    GGML_UNUSED(op);
+
+    return res;
+}
+
 ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_vec(
        ggml_metal_library_t lib,
        const ggml_tensor * op,
@@ -1585,6 +1605,7 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_v
        bool    has_bias,
        bool    has_scap,
        bool    has_kvpad,
+        bool    has_sparse,
        int32_t nqpsg,
        int32_t ne,
        int32_t nsg,
@@ -1614,13 +1635,14 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_v
            dv,
            qne_suffix);

-    snprintf(name, 256, "%s_mask=%d_sink=%d_bias=%d_scap=%d_kvpad=%d_ns10=%d_ns20=%d_nsg=%d_nwg=%d",
+    snprintf(name, 256, "%s_mask=%d_sink=%d_bias=%d_scap=%d_kvpad=%d_sparse=%d_ns10=%d_ns20=%d_nsg=%d_nwg=%d",
            base,
            has_mask,
            has_sinks,
            has_bias,
            has_scap,
            has_kvpad,
+            has_sparse,
            ns10,
            ns20,
            nsg, nwg);
@@ -1633,7 +1655,8 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_v
        ggml_metal_cv_set_bool(cv, has_sinks, FC_FLASH_ATTN_EXT_VEC + 1);
        ggml_metal_cv_set_bool(cv, has_bias,  FC_FLASH_ATTN_EXT_VEC + 2);
        ggml_metal_cv_set_bool(cv, has_scap,  FC_FLASH_ATTN_EXT_VEC + 3);
-        ggml_metal_cv_set_bool(cv, has_kvpad, FC_FLASH_ATTN_EXT_VEC + 4);
+        ggml_metal_cv_set_bool(cv, has_kvpad,  FC_FLASH_ATTN_EXT_VEC + 4);
+        ggml_metal_cv_set_bool(cv, has_sparse, FC_FLASH_ATTN_EXT_VEC + 5);

        ggml_metal_cv_set_int32(cv, ns10, FC_FLASH_ATTN_EXT_VEC + 20);
        ggml_metal_cv_set_int32(cv, ns20, FC_FLASH_ATTN_EXT_VEC + 21);
@@ -201,6 +201,10 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_att
        int32_t ns10,
        int32_t ns20);

+struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_vec_idx(
+        ggml_metal_library_t lib,
+        const struct ggml_tensor * op);
+
 struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_attn_ext_vec(
        ggml_metal_library_t lib,
        const struct ggml_tensor * op,
@@ -209,6 +213,7 @@ struct ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_flash_att
        bool    has_bias,
        bool    has_scap,
        bool    has_kvpad,
+        bool    has_sparse,
        int32_t nqpsg,
        int32_t ne,
        int32_t nsg,
@@ -259,6 +264,7 @@ enum ggml_metal_device_id {
    GGML_METAL_DEVICE_M5_PRO,
    GGML_METAL_DEVICE_M5_MAX,
    GGML_METAL_DEVICE_M5_ULTRA,
+    GGML_METAL_DEVICE_A18_PRO,
 };

 const char * ggml_metal_device_id_token(enum ggml_metal_device_id id);
@@ -1056,6 +1056,7 @@ static const struct {
    DEV("M5 Pro",   GGML_METAL_DEVICE_M5_PRO),
    DEV("M5 Max",   GGML_METAL_DEVICE_M5_MAX),
    DEV("M5 Ultra", GGML_METAL_DEVICE_M5_ULTRA),
+    DEV("A18 Pro",  GGML_METAL_DEVICE_A18_PRO),
 #undef DEV
 };

@@ -1346,19 +1347,21 @@ ggml_metal_device_t ggml_metal_device_init(int device, int n_devices) {
 void ggml_metal_device_free(ggml_metal_device_t dev) {
    assert(dev != NULL);

-    ggml_metal_rsets_free(dev->rsets);
+    @autoreleasepool {
+        ggml_metal_rsets_free(dev->rsets);

-    ggml_metal_library_free(dev->library);
-    dev->library = NULL;
+        ggml_metal_library_free(dev->library);
+        dev->library = NULL;

-    if (dev->mtl_queue) {
-        [dev->mtl_queue release];
-        dev->mtl_queue = nil;
-    }
+        if (dev->mtl_queue) {
+            [dev->mtl_queue release];
+            dev->mtl_queue = nil;
+        }

-    if (dev->mtl_device) {
-        [dev->mtl_device release];
-        dev->mtl_device = nil;
+        if (dev->mtl_device) {
+            [dev->mtl_device release];
+            dev->mtl_device = nil;
+        }
    }

    free(dev);
@@ -1446,12 +1449,14 @@ ggml_metal_event_t ggml_metal_device_event_init(ggml_metal_device_t dev) {
 }

 void ggml_metal_device_event_free(ggml_metal_device_t dev, ggml_metal_event_t ev) {
-    id<MTLSharedEvent> event = ev->obj;
-    [event release];
+    @autoreleasepool {
+        id<MTLSharedEvent> event = ev->obj;
+        [event release];

-    free(ev);
+        free(ev);

-    GGML_UNUSED(dev);
+        GGML_UNUSED(dev);
+    }
 }

 void ggml_metal_device_event_synchronize(ggml_metal_device_t dev, ggml_metal_event_t ev) {
@@ -1466,8 +1471,10 @@ void ggml_metal_device_event_synchronize(ggml_metal_device_t dev, ggml_metal_eve

 void ggml_metal_device_get_memory(ggml_metal_device_t dev, size_t * free, size_t * total) {
    if (@available(macOS 10.12, iOS 16.0, *)) {
-        *total = dev->mtl_device.recommendedMaxWorkingSetSize;
-        *free  = *total - dev->mtl_device.currentAllocatedSize;
+        *total     = dev->mtl_device.recommendedMaxWorkingSetSize;
+        size_t cur = dev->mtl_device.currentAllocatedSize;
+        // it's possible to allocate more than `recommendedMaxWorkingSetSize`
+        *free      = *total > cur ? *total - cur : 0;
    } else {
        *free = 0;
        *total = 0;
@@ -2226,14 +2233,16 @@ ggml_metal_buffer_t ggml_metal_buffer_map(ggml_metal_device_t dev, void * ptr, s
 }

 void ggml_metal_buffer_free(ggml_metal_buffer_t buf) {
-    ggml_metal_device_rsets_rm(buf->dev, buf->rset);
+    @autoreleasepool {
+        ggml_metal_device_rsets_rm(buf->dev, buf->rset);

-    for (int i = 0; i < buf->n_buffers; i++) {
-        [buf->buffers[i].metal release];
+        for (int i = 0; i < buf->n_buffers; i++) {
+            [buf->buffers[i].metal release];
+        }
+
+        ggml_metal_buffer_rset_free(buf);
    }

-    ggml_metal_buffer_rset_free(buf);
-
    if (buf->is_shared && buf->owned) {
 #if TARGET_OS_OSX
        vm_deallocate((vm_map_t)mach_task_self(), (vm_address_t)buf->all_data, buf->all_size);
@@ -458,8 +458,21 @@ typedef struct {
    float    m1;
    int32_t  n_head_log2;
    float    logit_softcap;
+    int32_t  n_kv_max_padded;
 } ggml_metal_kargs_flash_attn_ext_vec;

+typedef struct {
+    int32_t  ne30;
+    int32_t  ne31;
+    int32_t  ne32;
+    int32_t  ne33;
+    uint64_t nb31;
+    uint64_t nb32;
+    uint64_t nb33;
+    int32_t  n_kv_max;
+    int32_t  n_kv_max_padded;
+} ggml_metal_kargs_flash_attn_ext_vec_idx;
+
 typedef struct {
    int32_t  nrows;
 } ggml_metal_kargs_flash_attn_ext_vec_reduce;
@@ -917,7 +917,7 @@ int ggml_metal_op_glu(ggml_metal_op_t ctx, int idx) {

    const int64_t nrows = ggml_nrows(op->src[0]);

-    const int32_t nth = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne00/2);
+    const int32_t nth = std::max(1, std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne00/2));

    ggml_metal_encoder_set_pipeline(enc, pipeline);
    ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
@@ -2857,6 +2857,65 @@ static bool ggml_metal_op_flash_attn_ext_use_kv_f16(const ggml_tensor * op) {
    }
 }

+// returns the n_kv_max hint if the sparse path is available for this op, or 0 otherwise
+// the mask (src[3]) remains the single source of truth: finite entries are the valid KV positions,
+// n_kv_max is only an upper bound on their number per mask row, used to size the index lists
+static int ggml_metal_op_flash_attn_ext_n_kv_max_sparse(const ggml_tensor * op) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+
+    int32_t n_kv_max = 0;
+    memcpy(&n_kv_max, ((const int32_t *) op->op_params) + 4, sizeof(n_kv_max));
+
+    if (n_kv_max <= 0) {
+        return 0;
+    }
+
+    // the sparse indices are gathered from the mask
+    if (!op->src[3]) {
+        return 0;
+    }
+
+    // bound the size of the index lists
+    if (n_kv_max > 4096) {
+        return 0;
+    }
+
+    // vec kernel instantiations exist for these (type, dk, dv) combinations only
+    const int64_t dk = op->src[1]->ne[0];
+    const int64_t dv = op->src[2]->ne[0];
+
+    const bool dk_dv_ok = (dk == 32  && dv == 32)  ||
+                          (dk == 64  && dv == 64)  ||
+                          (dk == 96  && dv == 96)  ||
+                          (dk == 128 && dv == 128) ||
+                          (dk == 192 && dv == 128) ||
+                          (dk == 192 && dv == 192) ||
+                          (dk == 256 && dv == 256) ||
+                          (dk == 320 && dv == 256) ||
+                          (dk == 512 && dv == 512) ||
+                          (dk == 576 && dv == 512);
+
+    if (!dk_dv_ok) {
+        return 0;
+    }
+
+    switch (op->src[1]->type) {
+        case GGML_TYPE_F16:
+        case GGML_TYPE_BF16:
+        case GGML_TYPE_F32:
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+            break;
+        default:
+            return 0;
+    }
+
+    return n_kv_max;
+}
+
 // in some models (e.g. MLA-based), V is a view of K (the first ne20 elements of each K row);
 // the dequantized V is then a view of the dequantized K and does not need its own dequant or scratch
 // - ref: https://github.com/ggml-org/llama.cpp/pull/13435
@@ -3027,6 +3086,24 @@ size_t ggml_metal_op_flash_attn_ext_extra_kv_f16(const ggml_tensor * op) {
    return k_size + v_size;
 }

+// size of the sparse index lists: one list of KV indices per mask row,
+// padded with -1 up to a multiple of OP_FLASH_ATTN_EXT_VEC_NCPSG
+size_t ggml_metal_op_flash_attn_ext_extra_idx(const ggml_tensor * op) {
+    assert(op->op == GGML_OP_FLASH_ATTN_EXT);
+
+    GGML_TENSOR_LOCALS( int32_t, ne3, op->src[3], ne);
+
+    const int n_kv_max = ggml_metal_op_flash_attn_ext_n_kv_max_sparse(op);
+
+    if (n_kv_max <= 0) {
+        return 0;
+    }
+
+    const int n_kv_max_padded = GGML_PAD(n_kv_max, OP_FLASH_ATTN_EXT_VEC_NCPSG);
+
+    return GGML_PAD(sizeof(int32_t)*(size_t) n_kv_max_padded*ne31*ne32*ne33, 16);
+}
+
 int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
    ggml_tensor * op = ctx->node(idx);

@@ -3104,7 +3181,16 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
    ggml_metal_buffer_id bid_kv_f16 = bid_tmp;
    bid_kv_f16.offs += ggml_metal_op_flash_attn_ext_extra_tmp(op);

-    const bool use_kv_f16 = ggml_metal_op_flash_attn_ext_use_kv_f16(op);
+    // sparse path: gather the finite mask entries into index lists and run the vec kernels over them
+    const int n_kv_max_sparse = ggml_metal_op_flash_attn_ext_n_kv_max_sparse(op);
+    const bool use_sparse = n_kv_max_sparse > 0;
+    const int n_kv_max_padded = use_sparse ? GGML_PAD(n_kv_max_sparse, OP_FLASH_ATTN_EXT_VEC_NCPSG) : 0;
+
+    // the vec kernels dequantize the KV inline; no need for the F16 dequant pass in the sparse path
+    const bool use_kv_f16 = !use_sparse && ggml_metal_op_flash_attn_ext_use_kv_f16(op);
+
+    ggml_metal_buffer_id bid_idx = bid_kv_f16;
+    bid_idx.offs += ggml_metal_op_flash_attn_ext_extra_kv_f16(op);

    ggml_metal_buffer_id bid_k = bid_src1;
    ggml_metal_buffer_id bid_v = bid_src2;
@@ -3206,7 +3292,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
        }
    }

-    if (!ggml_metal_op_flash_attn_ext_use_vec(op)) {
+    if (!use_sparse && !ggml_metal_op_flash_attn_ext_use_vec(op)) {
        // half8x8 kernel
        const int nqptg = OP_FLASH_ATTN_EXT_NQPSG; // queries per threadgroup
        const int ncpsg = OP_FLASH_ATTN_EXT_NCPSG; // cache values per simdgroup
@@ -3378,13 +3464,18 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
 #undef FATTN_SMEM
    } else {
        // half4x4 kernel
-        auto cfg = ggml_metal_tuning::fa_vec_pick(
-                props_dev->device_id,
-                props_dev->gpu_family,
-                (int) op->src[1]->type,
-                (int) ne00, (int) ne20,   // dk, dv (ne00 == dk for FA)
-                ne11, ne01);
-        int nqptg = cfg.Q;                             // queries per threadgroup
+        // sparse: the index lists are per query row, so a threadgroup can share KV with Q == 1 only
+        auto cfg = use_sparse
+                ? ggml_metal_tuning::fa_vec_baseline_cfg((int) ne00, (int) ne20)
+                : ggml_metal_tuning::fa_vec_pick(
+                          props_dev->device_id,
+                          props_dev->gpu_family,
+                          (int) op->src[1]->type,
+                          (int) ne00, (int) ne20,   // dk, dv (ne00 == dk for FA)
+                          ne11, ne01);
+
+        int nqptg = cfg.Q; // queries per threadgroup
+
        const int ncpsg = OP_FLASH_ATTN_EXT_VEC_NCPSG; // cache values per simdgroup !! sync with kernel template arguments !!
        const int nhptg = 1;                           // heads per threadgroup

@@ -3394,7 +3485,39 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {

        bool need_sync = false;

-        const bool has_kvpad = ne11 % ncpsg != 0;
+        const bool has_kvpad = !use_sparse && ne11 % ncpsg != 0;
+
+        if (use_sparse) {
+            assert(ggml_metal_op_flash_attn_ext_extra_idx(op) != 0);
+
+            GGML_ASSERT(ne30 == ne11);
+
+            ggml_metal_kargs_flash_attn_ext_vec_idx args0 = {
+                /*.ne30              =*/ ne30,
+                /*.ne31              =*/ ne31,
+                /*.ne32              =*/ ne32,
+                /*.ne33              =*/ ne33,
+                /*.nb31              =*/ nb31,
+                /*.nb32              =*/ nb32,
+                /*.nb33              =*/ nb33,
+                /*.n_kv_max          =*/ n_kv_max_sparse,
+                /*.n_kv_max_padded   =*/ n_kv_max_padded,
+            };
+
+            auto pipeline0 = ggml_metal_library_get_pipeline_flash_attn_ext_vec_idx(lib, op);
+
+            ggml_metal_encoder_set_pipeline(enc, pipeline0);
+            ggml_metal_encoder_set_bytes   (enc, &args0, sizeof(args0), 0);
+            ggml_metal_encoder_set_buffer  (enc, bid_src3, 1);
+            ggml_metal_encoder_set_buffer  (enc, bid_idx,  2);
+
+            int nth = std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline0), 256);
+            nth = std::max(32, (nth/32)*32);
+
+            ggml_metal_encoder_dispatch_threadgroups(enc, ne31, ne32, ne33, nth, 1, 1);
+
+            need_sync = true;
+        }

        if (has_kvpad) {
            assert(ggml_metal_op_flash_attn_ext_extra_pad(op) != 0);
@@ -3455,11 +3578,26 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
        // workgroups
        // each workgroup handles nsg*nkpsg cache values
        int32_t nwg = 1;
-        if (false) {
-            // for small KV caches, we could launch a single workgroup and write the results directly to dst/
-            // however, this does not lead to significant improvement, so disabled
-            nwg = 1;
-            nsg = 4;
+        if (use_sparse) {
+            if (ne01 > 32) {
+                // large sparse batch
+                nwg = 1;
+                nsg = 1;
+                if (n_kv_max_padded == 640) {
+                    nsg = 4; // 640 % (4*32) == 0
+                } else {
+                    while (2*nwg*nsg*ncpsg < n_kv_max_padded && nsg < 4) {
+                        nsg *= 2;
+                    }
+                }
+            } else {
+                // small sparse batch
+                nwg = 32;
+                nsg = 1;
+                while (2*nwg*nsg*ncpsg < n_kv_max_padded && nsg < 4) {
+                    nsg *= 2;
+                }
+            }
        } else {
            nwg = 32;
            nsg = 1;
@@ -3484,7 +3622,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
            /*.nb01          =*/ nb01,
            /*.nb02          =*/ nb02,
            /*.nb03          =*/ nb03,
-            /*.ne11          =*/ ne11,
+            /*.ne11          =*/ use_sparse ? n_kv_max_padded : ne11,
            /*.ne_12_2       =*/ ne12,
            /*.ne_12_3       =*/ ne13,
            /*.ns10          =*/ ns10,
@@ -3510,9 +3648,10 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
            /*.m1            =*/ m1,
            /*.n_head_log2   =*/ n_head_log2,
            /*.logit_softcap =*/ logit_softcap,
+            /*.n_kv_max_padded =*/ n_kv_max_padded,
        };

-        auto pipeline = ggml_metal_library_get_pipeline_flash_attn_ext_vec(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, nqptg, cfg.NE, nsg, nwg, use_kv_f16, ns10, ns20);
+        auto pipeline = ggml_metal_library_get_pipeline_flash_attn_ext_vec(lib, op, has_mask, has_sinks, has_bias, has_scap, has_kvpad, use_sparse, nqptg, cfg.NE, nsg, nwg, use_kv_f16, ns10, ns20);

        GGML_ASSERT(nsg*32 <= ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));

@@ -3523,6 +3662,7 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
        ggml_metal_encoder_set_buffer  (enc, bid_v,    3);
        ggml_metal_encoder_set_buffer  (enc, bid_src3, 4);
        ggml_metal_encoder_set_buffer  (enc, bid_src4, 5);
+        ggml_metal_encoder_set_buffer  (enc, use_sparse ? bid_idx : bid_src0, 8);

        const size_t smem = FATTN_SMEM(nsg);

@@ -3530,8 +3670,6 @@ int ggml_metal_op_flash_attn_ext(ggml_metal_op_t ctx, int idx) {
        GGML_ASSERT(smem <= props_dev->max_theadgroup_memory_size);

        if (nwg == 1) {
-            assert(ggml_metal_op_flash_attn_ext_extra_tmp(op) == 0);
-
            // using 1 workgroup -> write the result directly into dst
            ggml_metal_encoder_set_buffer(enc, bid_pad, 6);
            ggml_metal_encoder_set_buffer(enc, bid_dst, 7);
@@ -43,6 +43,7 @@ size_t ggml_metal_op_flash_attn_ext_extra_pad(const struct ggml_tensor * op);
 size_t ggml_metal_op_flash_attn_ext_extra_blk(const struct ggml_tensor * op);
 size_t ggml_metal_op_flash_attn_ext_extra_tmp(const struct ggml_tensor * op);
 size_t ggml_metal_op_flash_attn_ext_extra_kv_f16(const struct ggml_tensor * op);
+size_t ggml_metal_op_flash_attn_ext_extra_idx(const struct ggml_tensor * op);

 int ggml_metal_op_concat            (ggml_metal_op_t ctx, int idx);
 int ggml_metal_op_repeat            (ggml_metal_op_t ctx, int idx);
@@ -1029,6 +1029,197 @@ constexpr fa_vec_entry_t fa_vec_tuned_table[] = {
    { { GGML_METAL_DEVICE_M2, GGML_TYPE_Q8_0, 576, 512, -1, 0 }, { 1, 4 } },
    { { GGML_METAL_DEVICE_M2, GGML_TYPE_Q8_0, 576, 512, -1, 1 }, { 1, 4 } },

+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_F16, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_F16, 64, 64, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_F16, 128, 128, 2, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_F16, 128, 128, 2, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_F16, 128, 128, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_F16, 128, 128, 2, 3 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_F16, 128, 128, 2, 4 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_F16, 128, 128, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_F16, 128, 128, 3, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_F16, 192, 128, 1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_F16, 192, 128, 1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_F16, 192, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_F16, 192, 128, 1, 3 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_F16, 192, 128, 1, 4 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_F16, 320, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_F16, 320, 256, 1, 1 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_F16, 320, 256, 1, 3 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_F16, 320, 256, 1, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 32, 32, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 32, 32, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 32, 32, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 32, 32, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 32, 32, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 64, 64, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 64, 64, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 64, 64, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 64, 64, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 96, 96, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 96, 96, 2, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 96, 96, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 96, 96, 3, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 96, 96, 3, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 128, 128, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 192, 192, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 192, 128, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 256, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 320, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 320, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 320, 256, 3, 3 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 320, 256, 3, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 576, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_0, 576, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 32, 32, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 32, 32, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 32, 32, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 32, 32, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 32, 32, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 64, 64, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 64, 64, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 96, 96, 2, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 96, 96, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 96, 96, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 128, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 128, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 128, 128, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 128, 128, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 192, 192, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 192, 192, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 192, 192, 1, 4 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 192, 192, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 192, 192, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 192, 128, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 256, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 320, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 320, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q4_1, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 32, 32, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 32, 32, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 32, 32, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 64, 64, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 64, 64, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 64, 64, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 64, 64, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 96, 96, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 96, 96, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 96, 96, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 96, 96, 2, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 96, 96, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 96, 96, 3, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 128, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 128, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 128, 128, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 128, 128, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 128, 128, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 192, 192, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 192, 192, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 192, 128, -1, 1 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 192, 128, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 192, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 192, 128, 1, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 192, 128, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 192, 128, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 256, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 320, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 320, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 576, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_0, 576, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 32, 32, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 32, 32, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 32, 32, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 64, 64, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 64, 64, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 64, 64, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 64, 64, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 96, 96, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 96, 96, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 96, 96, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 96, 96, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 128, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 128, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 128, 128, 2, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 128, 128, 2, 3 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 128, 128, 3, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 128, 128, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 192, 192, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 192, 192, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 192, 192, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 192, 192, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 192, 128, -1, 1 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 192, 128, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 192, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 192, 128, 1, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 192, 128, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 192, 128, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 256, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 320, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 320, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 576, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q5_1, 576, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 32, 32, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 32, 32, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 32, 32, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 32, 32, 2, 4 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 32, 32, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 32, 32, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 64, 64, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 64, 64, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 64, 64, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 64, 64, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 96, 96, 2, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 96, 96, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 96, 96, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 128, 128, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 192, 192, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 192, 128, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 256, 256, -1, 0 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 256, 256, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 320, 256, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 320, 256, 1, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 320, 256, 1, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 320, 256, 2, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 320, 256, 2, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 320, 256, 3, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 576, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M2_PRO, GGML_TYPE_Q8_0, 576, 512, -1, 1 }, { 1, 4 } },
+
    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 32, 32, 1, 3 }, { 2, 4 } },
    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 32, 32, 2, 3 }, { 2, 4 } },
    { { GGML_METAL_DEVICE_M2_MAX, GGML_TYPE_F16, 64, 64, 1, 0 }, { 1, 4 } },
@@ -1277,6 +1468,107 @@ constexpr fa_vec_entry_t fa_vec_tuned_table[] = {
    { { GGML_METAL_DEVICE_M2_ULTRA, GGML_TYPE_Q8_0, 576, 512, -1, 0 }, { 1, 4 } },
    { { GGML_METAL_DEVICE_M2_ULTRA, GGML_TYPE_Q8_0, 576, 512, -1, 1 }, { 1, 4 } },

+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 32, 32, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 32, 32, 1, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 32, 32, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 32, 32, 2, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 32, 32, 2, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 32, 32, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 64, 64, 3, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 64, 64, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 64, 64, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 64, 64, 1, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 64, 64, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 64, 64, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 96, 96, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 96, 96, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 96, 96, 1, 4 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 96, 96, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 96, 96, 2, 4 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 96, 96, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 128, 128, 1, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 128, 128, 2, 3 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 128, 128, 3, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 128, 128, 3, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 192, 3, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 192, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 192, 1, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 192, 1, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 192, 2, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 128, 2, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 128, 3, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 128, 1, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 128, 1, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 128, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 128, 2, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 128, 3, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 192, 128, 3, 2 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 256, 256, 1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 256, 256, 3, 0 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 256, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 256, 256, 1, 1 }, { 1, 1 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 256, 256, 1, 2 }, { 1, 1 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 256, 256, 1, 4 }, { 1, 1 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 256, 256, 2, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 320, 256, 2, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 320, 256, 3, 0 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 320, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 320, 256, 1, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 320, 256, 1, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 512, 512, 2, 0 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 512, 512, 3, 0 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 512, 512, 3, 1 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 512, 512, 3, 3 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 576, 512, 2, 0 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 576, 512, 2, 1 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 576, 512, 2, 2 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_F16, 576, 512, 3, 1 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 32, 32, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 32, 32, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 32, 32, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 64, 64, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 64, 64, 2, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 64, 64, 2, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 64, 64, 3, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 64, 64, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 96, 96, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 96, 96, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 96, 96, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 96, 96, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 128, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 128, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 128, 128, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 128, 128, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 128, 128, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 128, 128, 3, 4 }, { 1, 1 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 192, 192, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 192, 192, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 192, 192, 2, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 192, 192, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 192, 192, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 192, 128, 1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 192, 128, 2, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 192, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 192, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 192, 128, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 256, 256, 3, 0 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 256, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 320, 256, 1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 320, 256, 3, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 320, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 320, 256, 1, 1 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 576, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3, GGML_TYPE_Q8_0, 576, 512, -1, 1 }, { 1, 4 } },
+
    { { GGML_METAL_DEVICE_M3_PRO, GGML_TYPE_F16, 32, 32, 1, 1 }, { 2, 4 } },
    { { GGML_METAL_DEVICE_M3_PRO, GGML_TYPE_F16, 32, 32, 1, 3 }, { 2, 4 } },
    { { GGML_METAL_DEVICE_M3_PRO, GGML_TYPE_F16, 32, 32, 2, 1 }, { 2, 4 } },
@@ -1534,6 +1826,151 @@ constexpr fa_vec_entry_t fa_vec_tuned_table[] = {
    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_F16, 576, 512, 2, 2 }, { 4, 4 } },
    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_F16, 576, 512, 2, 3 }, { 4, 2 } },
    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_F16, 576, 512, 3, 1 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 32, 32, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 32, 32, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 32, 32, 3, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 32, 32, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 32, 32, 3, 4 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 64, 64, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 96, 96, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 96, 96, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 96, 96, 1, 4 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 96, 96, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 96, 96, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 128, 128, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 192, 192, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 192, 128, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 192, 128, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 192, 128, 3, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 256, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 320, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 320, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 576, 512, 2, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 576, 512, 3, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 576, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 576, 512, 1, 1 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 576, 512, 1, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 576, 512, 1, 3 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_0, 576, 512, 1, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 32, 32, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 32, 32, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 32, 32, 3, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 32, 32, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 32, 32, 3, 4 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 64, 64, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 96, 96, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 96, 96, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 96, 96, 1, 4 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 96, 96, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 96, 96, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 128, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 128, 128, 1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 128, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 128, 128, 1, 4 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 128, 128, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 128, 128, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 192, 192, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 192, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 192, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 192, 128, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 192, 128, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 256, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 320, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 576, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q4_1, 576, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 32, 32, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 32, 32, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 32, 32, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 32, 32, 2, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 64, 64, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 64, 64, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 64, 64, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 64, 64, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 96, 96, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 96, 96, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 96, 96, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 96, 96, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 96, 96, 3, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 128, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 192, 192, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 192, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 256, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 256, 256, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 256, 256, 1, 4 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 256, 256, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 256, 256, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 320, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 320, 256, 1, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 320, 256, 2, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 320, 256, 3, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 320, 256, 3, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 576, 512, 2, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 576, 512, 2, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 576, 512, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 576, 512, 2, 3 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_0, 576, 512, 2, 4 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 32, 32, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 32, 32, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 32, 32, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 32, 32, 2, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 64, 64, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 64, 64, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 64, 64, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 64, 64, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 96, 96, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 96, 96, 1, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 96, 96, 1, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 96, 96, 2, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 96, 96, 2, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 96, 96, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 128, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 128, 128, 2, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 192, 192, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 192, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 256, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 256, 256, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 256, 256, 1, 4 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 256, 256, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 256, 256, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 320, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 320, 256, 1, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 320, 256, 2, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 320, 256, 2, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 320, 256, 3, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 320, 256, 3, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 576, 512, 2, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 576, 512, 2, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 576, 512, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 576, 512, 2, 3 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q5_1, 576, 512, 2, 4 }, { 1, 4 } },
    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q8_0, 32, 32, -1, 1 }, { 4, 4 } },
    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q8_0, 32, 32, 1, 1 }, { 2, 4 } },
    { { GGML_METAL_DEVICE_M3_MAX, GGML_TYPE_Q8_0, 32, 32, 1, 4 }, { 2, 4 } },
@@ -2981,6 +3418,239 @@ constexpr fa_vec_entry_t fa_vec_tuned_table[] = {
    { { GGML_METAL_DEVICE_M5_MAX, GGML_TYPE_Q8_0, 576, 512, -1, 1 }, { 1, 4 } },
    { { GGML_METAL_DEVICE_M5_MAX, GGML_TYPE_Q8_0, 576, 512, 1, 4 }, { 1, 2 } },
    { { GGML_METAL_DEVICE_M5_MAX, GGML_TYPE_Q8_0, 576, 512, 2, 1 }, { 4, 4 } },
+
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 32, 32, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 32, 32, 1, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 32, 32, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 32, 32, 2, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 32, 32, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 64, 64, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 64, 64, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 64, 64, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 64, 64, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 64, 64, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 96, 96, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 96, 96, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 96, 96, 1, 4 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 96, 96, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 96, 96, 2, 4 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 128, 128, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 128, 128, 1, 2 }, { 1, 1 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 128, 128, 1, 4 }, { 1, 1 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 128, 128, 2, 2 }, { 1, 1 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 128, 128, 2, 4 }, { 1, 1 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 192, 192, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 192, 192, 1, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 192, 192, 1, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 192, 128, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 192, 128, 1, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 192, 128, 1, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 192, 128, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 192, 128, 2, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 256, 256, 2, 0 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 256, 256, 3, 0 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 256, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 256, 256, 1, 2 }, { 1, 1 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 256, 256, 1, 4 }, { 1, 1 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 320, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 512, 512, 3, 0 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 512, 512, 3, 2 }, { 4, 1 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 512, 512, 3, 3 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 576, 512, 2, 0 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 576, 512, 2, 2 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_F16, 576, 512, 3, 2 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 32, 32, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 32, 32, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 32, 32, 1, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 32, 32, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 64, 64, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 64, 64, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 64, 64, 1, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 64, 64, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 64, 64, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 96, 96, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 96, 96, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 96, 96, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 96, 96, 2, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 96, 96, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 128, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 128, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 192, 192, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 192, 192, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 192, 192, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 192, 192, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 192, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 192, 128, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 256, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 320, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 320, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 576, 512, 2, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 576, 512, 3, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 576, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 576, 512, 1, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_0, 576, 512, 1, 3 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 32, 32, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 32, 32, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 32, 32, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 64, 64, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 64, 64, 1, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 64, 64, 1, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 64, 64, 2, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 64, 64, 2, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 64, 64, 3, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 64, 64, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 96, 96, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 96, 96, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 96, 96, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 128, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 192, 192, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 192, 192, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 192, 192, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 192, 192, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 192, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 256, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 320, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 320, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 320, 256, 3, 1 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 512, 512, 2, 0 }, { 4, 1 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 512, 512, 3, 1 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 512, 512, 3, 2 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 576, 512, 2, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 576, 512, 3, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q4_1, 576, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 32, 32, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 32, 32, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 32, 32, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 64, 64, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 64, 64, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 64, 64, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 64, 64, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 96, 96, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 96, 96, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 96, 96, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 96, 96, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 128, 128, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 128, 128, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 128, 128, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 128, 128, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 192, 192, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 192, 128, -1, 1 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 192, 128, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 192, 128, 1, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 192, 128, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 192, 128, 2, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 192, 128, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 256, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 256, 256, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 320, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 320, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 320, 256, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 512, 512, 2, 3 }, { 2, 1 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 512, 512, 3, 1 }, { 2, 1 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 512, 512, 3, 3 }, { 2, 1 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 576, 512, 3, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 576, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 576, 512, 1, 2 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 576, 512, 1, 3 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_0, 576, 512, 1, 4 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 32, 32, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 32, 32, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 32, 32, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 64, 64, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 64, 64, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 64, 64, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 64, 64, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 96, 96, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 96, 96, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 96, 96, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 96, 96, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 128, 128, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 128, 128, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 128, 128, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 128, 128, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 192, 192, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 192, 128, -1, 1 }, { 4, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 192, 128, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 192, 128, 1, 4 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 192, 128, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 192, 128, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 256, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 320, 256, 1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 320, 256, -1, 1 }, { 2, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 320, 256, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 512, 512, 3, 3 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 576, 512, 2, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q5_1, 576, 512, 3, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 32, 32, -1, 1 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 32, 32, 1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 32, 32, 2, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 32, 32, 3, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 64, 64, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 64, 64, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 64, 64, 1, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 64, 64, 1, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 64, 64, 2, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 64, 64, 2, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 64, 64, 3, 2 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 64, 64, 3, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 96, 96, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 96, 96, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 96, 96, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 96, 96, 2, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 128, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 128, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 128, 128, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 128, 128, 2, 3 }, { 4, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 192, 192, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 192, 192, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 192, 192, 1, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 192, 192, 2, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 192, 192, 3, 2 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 192, 128, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 192, 128, -1, 1 }, { 2, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 256, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 256, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 320, 256, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 320, 256, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 512, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 512, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 576, 512, -1, 0 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 576, 512, -1, 1 }, { 1, 4 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 576, 512, 1, 1 }, { 1, 2 } },
+    { { GGML_METAL_DEVICE_A18_PRO, GGML_TYPE_Q8_0, 576, 512, 1, 4 }, { 1, 2 } },
 };

 static enum ggml_metal_device_id fa_vec_family_representative(int gpu_family) {
@@ -232,6 +232,7 @@ static size_t ggml_backend_metal_buffer_type_get_alloc_size(ggml_backend_buffer_
                res += ggml_metal_op_flash_attn_ext_extra_blk(tensor);
                res += ggml_metal_op_flash_attn_ext_extra_tmp(tensor);
                res += ggml_metal_op_flash_attn_ext_extra_kv_f16(tensor);
+                res += ggml_metal_op_flash_attn_ext_extra_idx(tensor);
            } break;
        case GGML_OP_CUMSUM:
        case GGML_OP_ARGSORT:
@@ -1071,6 +1071,112 @@ constant int32_t FC_flash_attn_ext_vec_ns10 [[function_constant(FC_FLASH_ATTN_EX
 constant int32_t FC_flash_attn_ext_vec_ns20 [[function_constant(FC_FLASH_ATTN_EXT_VEC + 21)]];
 constant int32_t FC_flash_attn_ext_vec_nsg  [[function_constant(FC_FLASH_ATTN_EXT_VEC + 22)]];
 constant int32_t FC_flash_attn_ext_vec_nwg  [[function_constant(FC_FLASH_ATTN_EXT_VEC + 23)]];
+constant bool    FC_flash_attn_ext_vec_has_sparse [[function_constant(FC_FLASH_ATTN_EXT_VEC + 5)]];
+
+// compress the finite entries of each KQ mask row into a list of KV indices (ascending order),
+// padded with -1 up to n_kv_max_padded (a multiple of OP_FLASH_ATTN_EXT_VEC_NCPSG)
+// one threadgroup per mask row; the mask remains the single source of truth for the values
+kernel void kernel_flash_attn_ext_vec_idx(
+        constant ggml_metal_kargs_flash_attn_ext_vec_idx & args,
+        device const half * mask,
+        device       int  * idx,
+        uint3   tgpig[[threadgroup_position_in_grid]],
+        ushort  tiitg[[thread_index_in_threadgroup]],
+        ushort3 ntg[[threads_per_threadgroup]]) {
+    constexpr short NW = N_SIMDWIDTH;
+    constexpr short NLOCAL = 32; // max finite positions kept in registers per thread
+
+    const int i1 = tgpig[0];
+    const int i2 = tgpig[1];
+    const int i3 = tgpig[2];
+
+    device const half * pm  = (device const half *) ((device const char *) mask + i1*args.nb31 + i2*args.nb32 + i3*args.nb33);
+    device int * pidx = idx + (((int64_t)i3*args.ne32 + i2)*args.ne31 + i1)*args.n_kv_max_padded;
+
+    const int n  = args.ne30;
+    const int q  = n/ntg.x;
+    const int r  = n%ntg.x;
+
+    // each thread handles a contiguous slice of the mask row
+    const int r0 = q*tiitg + min((int) tiitg, r);
+    const int r1 = r0 + q + (tiitg < r ? 1 : 0);
+
+    // count the finite entries in the slice and keep their positions in registers (single mask read)
+    int cnt = 0;  // total finite entries in the slice
+    int nloc = 0; // finite entries kept in registers
+    int local[NLOCAL];
+    for (int i = r0; i < r1; ++i) {
+        if (isfinite((float) pm[i])) {
+            if (nloc < NLOCAL) {
+                local[nloc] = i;
+                nloc++;
+            }
+            cnt++;
+        }
+    }
+
+    const short sgitg = tiitg/NW;
+    const short tiisg = tiitg%NW;
+
+    threadgroup int tcount[8];
+
+    // simd_sum is a collective: all lanes must evaluate it
+    const int sg_sum = simd_sum(cnt);
+    if (tiisg == 0) {
+        tcount[sgitg] = sg_sum;
+    }
+
+    threadgroup_barrier(mem_flags::mem_threadgroup);
+
+    int total = 0;
+    for (short s = 0; s < ntg.x/NW; ++s) {
+        total += tcount[s];
+    }
+
+    // base offset of this thread's slice in the output list (exclusive scan within the simdgroup)
+    int sg_base = 0;
+    for (short s = 0; s < sgitg; ++s) {
+        sg_base += tcount[s];
+    }
+
+    // exclusive prefix scan of the per-thread counts within the simdgroup
+    int incl = cnt;
+    for (int d = 1; d < NW; d <<= 1) {
+        const int v = simd_shuffle_up(incl, d);
+        if (tiisg >= d) {
+            incl += v;
+        }
+    }
+    const int base = sg_base + (incl - cnt);
+
+    // write the finite positions in order; if the hint is violated, keep only the first n_kv_max entries
+    int j = 0;
+    for (; j < nloc && base + j < args.n_kv_max; ++j) {
+        pidx[base + j] = local[j];
+    }
+
+    // a dense mask may have more than NLOCAL finite entries in a slice; re-read the mask to write the rest
+    if (cnt > nloc && base + nloc < args.n_kv_max) {
+        int j2 = 0;
+        for (int i = r0; i < r1; ++i) {
+            if (isfinite((float) pm[i])) {
+                if (j2 >= nloc) {
+                    pidx[base + j2] = i;
+                }
+                j2++;
+                if (base + j2 >= args.n_kv_max) {
+                    break;
+                }
+            }
+        }
+    }
+
+    // pad the tail of the list with -1
+    const int count = min(total, args.n_kv_max);
+    for (int i = count + tiitg; i < args.n_kv_max_padded; i += ntg.x) {
+        pidx[i] = -1;
+    }
+}

 template<
    typename q4_t,  // query types in shared memory
@@ -1091,6 +1197,7 @@ template<
    short NE = 4,   // head elements per thread
    short Q  = OP_FLASH_ATTN_EXT_VEC_NQPSG,  // queries per threadgroup
    short C  = OP_FLASH_ATTN_EXT_VEC_NCPSG>  // cache items per threadgroup
+
 kernel void kernel_flash_attn_ext_vec(
        constant ggml_metal_kargs_flash_attn_ext_vec & args,
        device const char * q,
@@ -1100,6 +1207,7 @@ kernel void kernel_flash_attn_ext_vec(
        device const char * sinks,
        device const char * pad,
        device       char * dst,
+        device const char * idx,
        threadgroup  half * shmem_f16 [[threadgroup(0)]],
        uint3   tgpig[[threadgroup_position_in_grid]],
        ushort  tiisg[[thread_index_in_simdgroup]],
@@ -1137,8 +1245,8 @@ kernel void kernel_flash_attn_ext_vec(

  //const short T = PK + NSG*SH; // shared memory size per query in (half)

-  //threadgroup q_t   * sq  = (threadgroup q_t   *) (shmem_f16 +                            0*PK); // holds the query data
-    threadgroup q4_t  * sq4 = (threadgroup q4_t  *) (shmem_f16 +                            0*PK); // same as above but in q4_t
+  //threadgroup q_t   * sq  = (threadgroup q_t   *) (shmem_f16 +                          0*PK); // holds the query data
+    threadgroup q4_t  * sq4 = (threadgroup q4_t  *) (shmem_f16 +                          0*PK); // same as above but in q4_t
    threadgroup s_t   * ss  = (threadgroup s_t   *) (shmem_f16 +   sgitg*SH         + Q*NSG*PK); // scratch buffer for attention
    threadgroup s4_t  * ss4 = (threadgroup s4_t  *) (shmem_f16 +   sgitg*SH         + Q*NSG*PK); // same as above but in s4_t
    threadgroup half  * sm  = (threadgroup half  *) (shmem_f16 +   sgitg*SH + 2*Q*C + Q*NSG*PK); // scratch buffer for mask
@@ -1207,6 +1315,14 @@ kernel void kernel_flash_attn_ext_vec(
        // pointer to the mask
        device const half * pm_base = (device const half *) (mask + iq1*Q*args.nb31 + (iq2%args.ne32)*args.nb32 + (iq3%args.ne33)*args.nb33);

+        // sparse indices: the list of finite mask entries per query row
+        // the sparse path requires Q == 1 (enforced by the host)
+        device const int * pidx = nullptr;
+        if (FC_flash_attn_ext_vec_has_sparse) {
+            pidx = (device const int *) idx +
+                ((int64_t)(iq3%args.ne33)*args.ne32 + (iq2%args.ne32))*args.ne31*args.n_kv_max_padded + (iq1%args.ne31)*args.n_kv_max_padded;
+        }
+
        float slope = 1.0f;

        // ALiBi
@@ -1265,11 +1381,22 @@ kernel void kernel_flash_attn_ext_vec(
            }

            if (FC_flash_attn_ext_vec_has_mask) {
-                FOR_UNROLL (short qq = 0; qq < Q; ++qq) {
-                    if ((iq1*Q + qq) < args.ne01) {
-                        sm[qq*C + tiisg] = pm[qq][ic + tiisg];
-                    } else {
-                        sm[qq*C + tiisg] = -MAXHALF;
+                if (FC_flash_attn_ext_vec_has_sparse) {
+                    FOR_UNROLL (short qq = 0; qq < Q; ++qq) {
+                        const int i11 = pidx[ic + tiisg];
+                        if ((iq1*Q + qq) < args.ne01 && i11 >= 0) {
+                            sm[qq*C + tiisg] = pm[qq][i11];
+                        } else {
+                            sm[qq*C + tiisg] = -MAXHALF;
+                        }
+                    }
+                } else {
+                    FOR_UNROLL (short qq = 0; qq < Q; ++qq) {
+                        if ((iq1*Q + qq) < args.ne01) {
+                            sm[qq*C + tiisg] = pm[qq][ic + tiisg];
+                        } else {
+                            sm[qq*C + tiisg] = -MAXHALF;
+                        }
                    }
                }
            } else {
@@ -1280,6 +1407,7 @@ kernel void kernel_flash_attn_ext_vec(
                }
            }

+            // skip -INF mask
            {
                bool any_finite = false;
                FOR_UNROLL (short qq = 0; qq < Q; ++qq) {
@@ -1294,9 +1422,13 @@ kernel void kernel_flash_attn_ext_vec(

            // Q*K^T
            {
-                device      const k4_t * pk4 = (device const k4_t *) (k + ic*args.nb11);
+                device      const k4_t * pk4 = nullptr;

-                pk4 += ty*NS10/4 + tx;
+                if (!FC_flash_attn_ext_vec_has_sparse) {
+                    pk4 = (device const k4_t *) (k + ic*args.nb11);
+
+                    pk4 += ty*NS10/4 + tx;
+                }

                qk_t mqk[Q][C/NE];
                FOR_UNROLL (short qq = 0; qq < Q; ++qq) {
@@ -1307,7 +1439,35 @@ kernel void kernel_flash_attn_ext_vec(

                // each simdgroup processes Q queries and NE (NW/NL) cache elements
                FOR_UNROLL (short cc = 0; cc < C/NE; ++cc) {
-                    if (is_same<kd4_t, k4_t>::value) {
+                    if (FC_flash_attn_ext_vec_has_sparse) {
+                        // the KV rows are gathered from the index list; -1 entries are padding
+                        const int i11 = pidx[ic + NE*cc + ty];
+                        if (i11 >= 0) {
+                            if (is_same<kd4_t, k4_t>::value) {
+                                device const k4_t * pk4s = (device const k4_t *) (k + i11*args.nb11) + tx;
+                                FOR_UNROLL (short ii = 0; ii < DK4/NL; ++ii) {
+                                    const k4_t k_elem = pk4s[ii*NL];
+                                    FOR_UNROLL (short qq = 0; qq < Q; ++qq) {
+                                        mqk[qq][cc] += dot((float4) k_elem, (float4) sq4[qq*PK4 + ii*NL + tx]);
+                                    }
+                                }
+                            } else {
+                                device const kd4_t * pk = (device const kd4_t *) (k + i11*args.nb11);
+
+                                k4_t mk;
+
+                                FOR_UNROLL (short ii = 0; ii < DK4/NL; ++ii) {
+                                    const short i = ii*NL + tx;
+
+                                    deq_k_t4(pk + i/nl_k, i%nl_k, mk);
+
+                                    FOR_UNROLL (short qq = 0; qq < Q; ++qq) {
+                                        mqk[qq][cc] += dot((float4) mk, (float4) sq4[qq*PK4 + i]);
+                                    }
+                                }
+                            }
+                        }
+                    } else if (is_same<kd4_t, k4_t>::value) {
                        FOR_UNROLL (short ii = 0; ii < DK4/NL; ++ii) {
                            const k4_t k_elem = pk4[cc*NE*NS10/4 + ii*NL];
                            FOR_UNROLL (short qq = 0; qq < Q; ++qq) {
@@ -1422,7 +1582,40 @@ kernel void kernel_flash_attn_ext_vec(
                    }
                }

-                if (is_same<vd4_t, v4_t>::value) {
+                if (FC_flash_attn_ext_vec_has_sparse) {
+                    FOR_UNROLL (short cc = 0; cc < C/NE; ++cc) {
+                        // the KV rows are gathered from the index list; -1 entries are padding
+                        const int i11 = pidx[ic + NE*cc + ty];
+                        if (i11 >= 0) {
+                            if (is_same<vd4_t, v4_t>::value) {
+                                device const v4_t * pv4 = (device const v4_t *) (v + i11*args.nb21);
+
+                                pv4 += tx;
+
+                                FOR_UNROLL (short ii = 0; ii < DV4/NL; ++ii) {
+                                    const v4_t v_elem = pv4[ii*NL];
+                                    FOR_UNROLL (short qq = 0; qq < Q; ++qq) {
+                                        lo[qq][ii] += o4_t(float4(v_elem)*float4(ss[qq*C + cc*NE + ty]));
+                                    }
+                                }
+                            } else {
+                                device const vd4_t * pv4 = (device const vd4_t *) (v + i11*args.nb21);
+
+                                FOR_UNROLL (short ii = 0; ii < DV4/NL; ++ii) {
+                                    const short i = ii*NL + tx;
+
+                                    v4_t mv;
+
+                                    deq_v_t4(pv4 + i/nl_v, i%nl_v, mv);
+
+                                    FOR_UNROLL (short qq = 0; qq < Q; ++qq) {
+                                        lo[qq][ii] += o4_t(float4(mv)*float4(ss[qq*C + cc*NE + ty]));
+                                    }
+                                }
+                            }
+                        }
+                    }
+                } else if (is_same<vd4_t, v4_t>::value) {
                    device const v4_t * pv4 = (device const v4_t *) (v + ic*args.nb21);

                    pv4 += ty*NS20/4 + tx;
@@ -75,7 +75,6 @@ if (MUSAToolkit_FOUND)
    endif()

    add_compile_definitions(GGML_USE_MUSA)
-    add_compile_definitions(GGML_CUDA_PEER_MAX_BATCH_SIZE=${GGML_CUDA_PEER_MAX_BATCH_SIZE})

    if (GGML_MUSA_GRAPHS)
        add_compile_definitions(GGML_MUSA_GRAPHS)
@@ -85,6 +85,7 @@ set(GGML_OPENCL_KERNELS
    mul_mv_f16_f32_1row
    mul_mv_f16_f32_l4
    mul_mv_f16_f32
+    mul_mv_f16_f32_mrow
    mul_mv_f32_f32
    mul_mv_q1_0_f32
    mul_mv_q1_0_f32_flat
@@ -180,9 +181,14 @@ set(GGML_OPENCL_KERNELS
    gemv_noshuffle_q8_0_f32
    gemm_noshuffle_q8_0_f32
    gemv_noshuffle_q4_k_f32
+    gemv_noshuffle_q4_k_f32_o4
+    gemv_noshuffle_q4_k_f32_tiled
    gemm_noshuffle_q4_k_f32
    gemv_noshuffle_q6_k_f32
+    gemv_noshuffle_q6_k_f32_o4
+    gemv_noshuffle_q6_k_f32_tiled
    gemm_noshuffle_q6_k_f32
+    gemm_noshuffle_q6_k_f32_tiled
    gemv_noshuffle_q5_k_f32
    gemm_noshuffle_q5_k_f32
    mul
@@ -1110,6 +1110,78 @@ kernel void kernel_restore_block_q4_k_trans4_ns(
    }
 }

+//------------------------------------------------------------------------------
+// kernel_convert_block_q4_k_tiled_ns
+//
+// Tiled-wide layout for the long-vocab q4_K lm_head/embed GEMV (decode path).
+// Mirror of kernel_convert_block_q6_k_tiled_ns: recovers each weight's 4-bit
+// code in CANONICAL ggml element order (e in [0,256)) and re-packs into 32 uints
+// (8 codes/uint), stored TILED by 64 output rows so the matching GEMV
+// (gemv_noshuffle_q4_k_f32_tiled) coalesces every weight load. The 12-byte
+// packed scale block `s` and d/dm are stored per (row, K-block) tiled; the GEMV
+// re-derives the 8 (scale,min) pairs via get_scale_min_k4, exactly like the o4
+// kernel. Both ends owned here -> correct by construction vs the reference q4_K
+// dequant. Requires ne01 % 64 == 0 (gated host-side). Buffer sizes identical to
+// the trans4_ns layout.
+//
+//   q  uint4 granule g of (row r, K-block sb): idx = ((rt*ne00_blk+sb)*8 + g)*64 + rit
+//   s  (12 bytes) of (r, sb):                  idx = (rt*ne00_blk+sb)*64 + rit, *12
+//   d/dm (half)  of (r, sb):                   idx = (rt*ne00_blk+sb)*64 + rit
+//   where rt = r/64, rit = r%64.
+//------------------------------------------------------------------------------
+kernel void kernel_convert_block_q4_k_tiled_ns(
+    __global struct block_q4_K * src0,
+    __global uint  * dst_q,    // 32 uints / superblock (4-bit codes, 8 codes/uint)
+    __global half  * dst_d,    // 1 half  / superblock
+    __global half  * dst_dm,   // 1 half  / superblock
+    __global uchar * dst_s,    // K_SCALE_SIZE (12) bytes / superblock
+    uint ne00,
+    uint ne01
+) {
+    uint i00 = get_global_id(1);   // K-block index (superblock along ne00)
+    uint i01 = get_global_id(0);   // output row index (along ne01)
+    uint i02 = get_global_id(2);   // batch
+
+    uint ne00_blk = ne00 / QK_K;
+
+    uint src_blk_offset = i00 + i01 * ne00_blk + i02 * ne00_blk * ne01;
+    __global struct block_q4_K * b = src0 + src_blk_offset;
+
+    uint rt  = i01 / 64;
+    uint rit = i01 % 64;
+    uint tile_blk = (i02 * (ne01 / 64) + rt) * ne00_blk + i00;
+
+    // --- recover canonical 4-bit codes in e-order, pack 8 codes/uint ---
+    uint qw[32] = {0};
+    for (uint e = 0; e < 256; ++e) {
+        uint g    = e >> 6;           // group 0..3 (q advances 32 bytes/group)
+        uint within = e & 63u;
+        uint hlf  = within >> 5;      // 0 = low nibble, 1 = high nibble
+        uint l    = within & 31u;     // 0..31
+        uchar byte = b->q[g * 32u + l];
+        uint code = (hlf == 0u) ? (uint)(byte & 0x0F) : (uint)(byte >> 4);
+        qw[e >> 3] |= code << ((e & 7u) * 4u);
+    }
+
+    for (uint gr = 0; gr < 8; ++gr) {
+        uint base = (tile_blk * 8u + gr) * 64u + rit;   // uint4 index
+        dst_q[base * 4u + 0u] = qw[gr * 4u + 0u];
+        dst_q[base * 4u + 1u] = qw[gr * 4u + 1u];
+        dst_q[base * 4u + 2u] = qw[gr * 4u + 2u];
+        dst_q[base * 4u + 3u] = qw[gr * 4u + 3u];
+    }
+
+    // packed scales (12 bytes), tiled per (row, block)
+    __global uchar * s_dst = dst_s + (tile_blk * 64u + rit) * K_SCALE_SIZE;
+    #pragma unroll
+    for (int i = 0; i < K_SCALE_SIZE; ++i) {
+        s_dst[i] = b->s[i];
+    }
+
+    dst_d [tile_blk * 64u + rit] = b->d;
+    dst_dm[tile_blk * 64u + rit] = b->dm;
+}
+
 kernel void kernel_convert_block_q5_k_trans4_ns(
    __global struct block_q5_K * src0,
    __global uint  * dst_qs,
@@ -1494,6 +1566,105 @@ kernel void kernel_restore_block_mxfp4_trans(
    b->e = src_e[src_blk_offset];
 }

+//------------------------------------------------------------------------------
+// kernel_convert_block_q6_k_tiled_ns
+//
+// Tiled-wide layout for the long-vocab q6_K lm_head/embed GEMV (decode path).
+// Unlike *_trans4_ns (which mirrors the bit-interleave the legacy 2-output GEMV
+// consumes), this kernel is correct-by-construction against the CANONICAL ggml
+// q6_K dequant: it recovers each weight's 6-bit code in element order e in
+// [0,256), then re-packs low-4-bits into 32 uints (8 codes/uint) and high-2-bits
+// into 16 uints (16 codes/uint). The matching GEMV (gemv_noshuffle_q6_k_f32_tiled)
+// unpacks the same order, so both ends are owned here.
+//
+// Storage is TILED by 64 output rows so the GEMV's 64-thread tile coalesces:
+//   ql uint4 granule g of (row r, K-block sb): idx = ((rt*ne00_blk + sb)*8 + g)*64 + rit
+//   qh uint4 granule g:                         idx = ((rt*ne00_blk + sb)*4 + g)*64 + rit
+//   scales (char16) of (r, sb):                 idx = (rt*ne00_blk + sb)*64 + rit
+//   d (half) of (r, sb):                        idx = (rt*ne00_blk + sb)*64 + rit
+// where rt = r/64, rit = r%64. Requires ne01 % 64 == 0 (gated host-side).
+// Buffer sizes are byte-identical to the trans4_ns layout.
+//------------------------------------------------------------------------------
+kernel void kernel_convert_block_q6_k_tiled_ns(
+    __global struct block_q6_K * src0,
+    __global uint  * dst_ql,   // 32 uints / superblock (low 4 bits, 8 codes/uint)
+    __global uint  * dst_qh,   // 16 uints / superblock (high 2 bits, 16 codes/uint)
+    __global half  * dst_d,    // 1 half  / superblock
+    __global char  * dst_s,    // 16 chars/ superblock
+    uint ne00,
+    uint ne01
+) {
+    uint i00 = get_global_id(1);   // K-block index (superblock along ne00)
+    uint i01 = get_global_id(0);   // output row index (along ne01)
+    uint i02 = get_global_id(2);   // batch
+
+    uint ne00_blk = ne00 / QK_K;
+
+    // Source block: row-major over (i02, i01, i00).
+    uint src_blk_offset = i00 + i01 * ne00_blk + i02 * ne00_blk * ne01;
+    __global struct block_q6_K * b = src0 + src_blk_offset;
+
+    uint rt  = i01 / 64;
+    uint rit = i01 % 64;
+    uint tile_blk = (i02 * (ne01 / 64) + rt) * ne00_blk + i00;  // tile-major (row-tile, K-block)
+
+    // --- recover canonical 6-bit codes, pack into ql (4b) + qh (2b) in e-order ---
+    // 32 ql-uints (8 low-nibbles each) + 16 qh-uints (16 2-bit slots each).
+    uint qlw[32] = {0};
+    uint qhw[16] = {0};
+
+    for (uint e = 0; e < 256; ++e) {
+        uint n   = (e >= 128) ? 1u : 0u;       // which 128-half
+        uint within = e - n * 128u;
+        uint q   = within / 32u;               // quadrant 0..3
+        uint l   = within % 32u;               // 0..31
+
+        uint off_ql = n * 64u;                 // raw ql byte base for this half
+        uint off_qh = n * 32u;                 // raw qh byte base for this half
+
+        uchar low4;
+        uchar qlb0 = b->ql[off_ql + l];
+        uchar qlb1 = b->ql[off_ql + l + 32];
+        if (q == 0)      low4 = qlb0 & 0x0F;
+        else if (q == 1) low4 = qlb1 & 0x0F;
+        else if (q == 2) low4 = (qlb0 >> 4) & 0x0F;
+        else             low4 = (qlb1 >> 4) & 0x0F;
+
+        uchar hi2 = (b->qh[off_qh + l] >> (q * 2u)) & 0x03;
+
+        // pack low4 (e-order): uint e/8, nibble (e%8)
+        qlw[e >> 3] |= ((uint)low4) << ((e & 7u) * 4u);
+        // pack hi2 (e-order): uint e/16, 2-bit slot (e%16)
+        qhw[e >> 4] |= ((uint)hi2)  << ((e & 15u) * 2u);
+    }
+
+    // --- write tiled ---
+    for (uint g = 0; g < 8; ++g) {
+        uint base = (tile_blk * 8u + g) * 64u + rit;   // uint4 index
+        dst_ql[base * 4u + 0u] = qlw[g * 4u + 0u];
+        dst_ql[base * 4u + 1u] = qlw[g * 4u + 1u];
+        dst_ql[base * 4u + 2u] = qlw[g * 4u + 2u];
+        dst_ql[base * 4u + 3u] = qlw[g * 4u + 3u];
+    }
+    for (uint g = 0; g < 4; ++g) {
+        uint base = (tile_blk * 4u + g) * 64u + rit;   // uint4 index
+        dst_qh[base * 4u + 0u] = qhw[g * 4u + 0u];
+        dst_qh[base * 4u + 1u] = qhw[g * 4u + 1u];
+        dst_qh[base * 4u + 2u] = qhw[g * 4u + 2u];
+        dst_qh[base * 4u + 3u] = qhw[g * 4u + 3u];
+    }
+
+    // scales: 16 chars contiguous per (row, block), tiled
+    __global char * s_dst = dst_s + (tile_blk * 64u + rit) * 16u;
+    #pragma unroll
+    for (int i = 0; i < 16; ++i) {
+        s_dst[i] = b->scales[i];
+    }
+
+    // super-block scale
+    dst_d[tile_blk * 64u + rit] = b->d;
+}
+
 kernel void kernel_convert_block_mxfp4_trans4_ns(
    global struct block_mxfp4 * src0,
    __global uint * dst_q,
@@ -4,6 +4,7 @@
 #pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
 #define ADRENO_GPU 1
 #define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
 #endif
 #define QK_K         256
 #define K_SCALE_SIZE 12
@@ -171,3 +172,319 @@ kernel void kernel_gemm_noshuffle_q4_k_f32(
        vstore4((float4)(c0.s7, c1.s7, c2.s7, c3.s7), 0, dst + idx);
    }
 }
+
+// 1x8 per-WI tile (1 output row x 8 output cols). For the small-batch
+// (medium n_q, e.g. MTP/spec verify) path where the 2x8 kernel is starved:
+// at ne1<=8 the grid is (1, ceil(M/2)) -> only ~M/256 workgroups, leaving
+// the SP under-occupied. 1 row per WI doubles the M-axis workgroup count
+// (ceil(M/1)/128 vs ceil(M/2)/128) AND collapses the accumulators to a
+// single half8 (16 regs, no spill), so more waves co-reside. Same weight
+// traffic as 2x8 (rows never share weights); the win is pure occupancy.
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_128
+#endif
+kernel void kernel_gemm_noshuffle_q4_k_f32_r1(
+    global const ushort * src0_q,
+    global const uchar  * src0_s,
+    global const half   * src0_d,
+    global const half   * src0_dm,
+    read_only image1d_buffer_t src1,
+    global float * dst,
+    ulong offsetd,
+    int m,
+    int n,
+    int k,
+    int n_no_padding,
+    uchar mask_d6,
+    uchar mask_d4,
+    uchar mask_hi2
+) {
+    dst = (global float *)((global char *)dst + offsetd);
+    int n_4 = n >> 2;
+    int gy = get_global_id(0);
+    int gx = get_global_id(1);       // 1 row per WI
+
+    half8 c0 = 0;
+    half8 B;
+    half dq;
+
+    int num_blocks_K = k / QK_K;
+
+    global const ushort * weight_ptr = src0_q + gx;
+    global const half   * d_ptr      = src0_d  + gx;
+    global const half   * dm_ptr     = src0_dm + gx;
+
+    for (int i = 0; i < k; i += 32) {
+        int sb_idx  = i / QK_K;
+        int sub_idx = (i / 32) % 8;
+
+        half dd  = d_ptr [sb_idx * m];
+        half dmm = dm_ptr[sb_idx * m];
+
+        global const uchar * sc0 = src0_s + sb_idx * K_SCALE_SIZE * m + gx;
+
+        uchar sv0, mn0;
+        get_scale_min_k4(sub_idx, sc0, m, &sv0, &mn0, mask_d6, mask_d4, mask_hi2);
+
+        half scale = convert_half(convert_float(dd)  * (float)sv0);
+        half mval  = convert_half(convert_float(dmm) * (float)mn0);
+
+        for (int l = 0; l < 32; l += 4) {
+            int ki = i + l;
+            ushort bits = weight_ptr[(ki/4) * m];
+
+            B.s0123 = read_imageh(src1, gy*2   + (ki+0) * n_4);
+            B.s4567 = read_imageh(src1, gy*2+1 + (ki+0) * n_4);
+            dq = (bits & 0x000F) * scale - mval;
+            c0 += B * dq;
+
+            B.s0123 = read_imageh(src1, gy*2   + (ki+1) * n_4);
+            B.s4567 = read_imageh(src1, gy*2+1 + (ki+1) * n_4);
+            dq = ((bits & 0x00F0) >> 4) * scale - mval;
+            c0 += B * dq;
+
+            B.s0123 = read_imageh(src1, gy*2   + (ki+2) * n_4);
+            B.s4567 = read_imageh(src1, gy*2+1 + (ki+2) * n_4);
+            dq = ((bits & 0x0F00) >> 8) * scale - mval;
+            c0 += B * dq;
+
+            B.s0123 = read_imageh(src1, gy*2   + (ki+3) * n_4);
+            B.s4567 = read_imageh(src1, gy*2+1 + (ki+3) * n_4);
+            dq = ((bits & 0xF000) >> 12) * scale - mval;
+            c0 += B * dq;
+        }
+    }
+
+    // Output: 8 cols, 1 row per col-step. Scalar store, coalesced across
+    // neighbouring WIs (consecutive gx -> consecutive dst addresses).
+    int idx = (gy<<3)*m + gx;
+    if (idx < m*n_no_padding) { dst[idx] = c0.s0; idx += m; }
+    if (idx < m*n_no_padding) { dst[idx] = c0.s1; idx += m; }
+    if (idx < m*n_no_padding) { dst[idx] = c0.s2; idx += m; }
+    if (idx < m*n_no_padding) { dst[idx] = c0.s3; idx += m; }
+    if (idx < m*n_no_padding) { dst[idx] = c0.s4; idx += m; }
+    if (idx < m*n_no_padding) { dst[idx] = c0.s5; idx += m; }
+    if (idx < m*n_no_padding) { dst[idx] = c0.s6; idx += m; }
+    if (idx < m*n_no_padding) { dst[idx] = c0.s7; }
+}
+
+// 2x8 tile, but weights read through an image1d_buffer (CL_R/UINT32 over the
+// same packed-q buffer) instead of a plain global buffer. The ne1==1 GEMV
+// already does this and is much faster per weight byte than this GEMM at
+// small n_q; the structural difference is the image path hits the dedicated
+// TPL1 weight cache (L1) while the global path only reaches L2. At small n_q
+// the forward is weight-read-bound, so L1-cached weights is the lever.
+// The 2 adjacent rows the 2x8 tile reads as a ushort2 are exactly one uint32,
+// so the vload2 becomes a single read_imageui at index gx + (ki/4)*(m/2).
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_128
+#endif
+kernel void kernel_gemm_noshuffle_q4_k_f32_kimg(
+    read_only image1d_buffer_t src0_q_img,
+    global const uchar  * src0_s,
+    global const half   * src0_d,
+    global const half   * src0_dm,
+    read_only image1d_buffer_t src1,
+    global float * dst,
+    ulong offsetd,
+    int m,
+    int n,
+    int k,
+    int n_no_padding,
+    uchar mask_d6,
+    uchar mask_d4,
+    uchar mask_hi2
+) {
+    dst = (global float *)((global char *)dst + offsetd);
+    int n_4 = n >> 2;
+    int m_2 = m >> 1;
+    int gy = get_global_id(0);
+    int gx = get_global_id(1);
+    int gx_2 = gx << 1;
+
+    half8 c0 = 0, c1 = 0;
+    half8 B;
+    half2 dequantized_weights;
+
+    int num_blocks_K = k / QK_K;
+
+    global const half * d_ptr  = src0_d  + gx_2;
+    global const half * dm_ptr = src0_dm + gx_2;
+
+    for (int i = 0; i < k; i += 32) {
+        int sb_idx  = i / QK_K;
+        int sub_idx = (i / 32) % 8;
+
+        half2 d  = vload2(0, d_ptr  + sb_idx * m);
+        half2 dm = vload2(0, dm_ptr + sb_idx * m);
+
+        global const uchar * sc0 = src0_s + sb_idx * K_SCALE_SIZE * m + (gx_2+0);
+        global const uchar * sc1 = sc0 + 1;
+
+        uchar sv0, mn0, sv1, mn1;
+        get_scale_min_k4(sub_idx, sc0, m, &sv0, &mn0, mask_d6, mask_d4, mask_hi2);
+        get_scale_min_k4(sub_idx, sc1, m, &sv1, &mn1, mask_d6, mask_d4, mask_hi2);
+
+        half2 scale = convert_half2(convert_float2(d)  * convert_float2((uchar2)(sv0, sv1)));
+        half2 mval  = convert_half2(convert_float2(dm) * convert_float2((uchar2)(mn0, mn1)));
+
+        for (int l = 0; l < 32; l += 4) {
+            int ki = i + l;
+            uint wpacked = read_imageui(src0_q_img, gx + (ki/4) * m_2).x;
+            ushort2 bits2 = (ushort2)((ushort)(wpacked & 0xFFFFu), (ushort)(wpacked >> 16));
+
+            // j=0
+            B.s0123 = read_imageh(src1, gy*2   + (ki+0) * n_4);
+            B.s4567 = read_imageh(src1, gy*2+1 + (ki+0) * n_4);
+            dequantized_weights.s0 = (bits2.s0 & 0x000F) * scale.s0 - mval.s0;
+            dequantized_weights.s1 = (bits2.s1 & 0x000F) * scale.s1 - mval.s1;
+            c0 += B * dequantized_weights.s0;
+            c1 += B * dequantized_weights.s1;
+
+            // j=1
+            B.s0123 = read_imageh(src1, gy*2   + (ki+1) * n_4);
+            B.s4567 = read_imageh(src1, gy*2+1 + (ki+1) * n_4);
+            dequantized_weights.s0 = ((bits2.s0 & 0x00F0) >> 4) * scale.s0 - mval.s0;
+            dequantized_weights.s1 = ((bits2.s1 & 0x00F0) >> 4) * scale.s1 - mval.s1;
+            c0 += B * dequantized_weights.s0;
+            c1 += B * dequantized_weights.s1;
+
+            // j=2
+            B.s0123 = read_imageh(src1, gy*2   + (ki+2) * n_4);
+            B.s4567 = read_imageh(src1, gy*2+1 + (ki+2) * n_4);
+            dequantized_weights.s0 = ((bits2.s0 & 0x0F00) >> 8) * scale.s0 - mval.s0;
+            dequantized_weights.s1 = ((bits2.s1 & 0x0F00) >> 8) * scale.s1 - mval.s1;
+            c0 += B * dequantized_weights.s0;
+            c1 += B * dequantized_weights.s1;
+
+            // j=3
+            B.s0123 = read_imageh(src1, gy*2   + (ki+3) * n_4);
+            B.s4567 = read_imageh(src1, gy*2+1 + (ki+3) * n_4);
+            dequantized_weights.s0 = ((bits2.s0 & 0xF000) >> 12) * scale.s0 - mval.s0;
+            dequantized_weights.s1 = ((bits2.s1 & 0xF000) >> 12) * scale.s1 - mval.s1;
+            c0 += B * dequantized_weights.s0;
+            c1 += B * dequantized_weights.s1;
+        }
+    }
+
+    int idx = (gy<<3)*m + (gx<<1);
+    if (idx+1 < m*n_no_padding) { vstore2((float2)(c0.s0, c1.s0), 0, dst + idx); idx += m; }
+    if (idx+1 < m*n_no_padding) { vstore2((float2)(c0.s1, c1.s1), 0, dst + idx); idx += m; }
+    if (idx+1 < m*n_no_padding) { vstore2((float2)(c0.s2, c1.s2), 0, dst + idx); idx += m; }
+    if (idx+1 < m*n_no_padding) { vstore2((float2)(c0.s3, c1.s3), 0, dst + idx); idx += m; }
+    if (idx+1 < m*n_no_padding) { vstore2((float2)(c0.s4, c1.s4), 0, dst + idx); idx += m; }
+    if (idx+1 < m*n_no_padding) { vstore2((float2)(c0.s5, c1.s5), 0, dst + idx); idx += m; }
+    if (idx+1 < m*n_no_padding) { vstore2((float2)(c0.s6, c1.s6), 0, dst + idx); idx += m; }
+    if (idx+1 < m*n_no_padding) { vstore2((float2)(c0.s7, c1.s7), 0, dst + idx); }
+}
+
+// Cooperative-K GEMM for the small-batch (n_q in [2..8]) path. Mirrors the
+// ne1==1 GEMV's structure: a WG is (COK_SG lanes x COK_NSG subgroups); each
+// lane owns ONE output row and computes its 8 (padded) columns, and the
+// COK_NSG subgroups SPLIT the K reduction round-robin, combining via a
+// __local reduction. This is the thing the per-WI GEMM lacked — at small n_q
+// the old kernel had ~M/256 workgroups each walking all of K serially; this
+// has M/64 workgroups AND COK_NSG-way K parallelism. Uses REQD_SUBGROUP_SIZE_64
+// + barrier (same safe reduction pattern as the GEMV; never sub_group_reduce
+// at full width on X2 per the GDN miscompile note).
+#define COK_NSG 8
+#define COK_SG  64
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_gemm_noshuffle_q4_k_f32_cok(
+    global const ushort * src0_q,
+    global const uchar  * src0_s,
+    global const half   * src0_d,
+    global const half   * src0_dm,
+    read_only image1d_buffer_t src1,
+    global float * dst,
+    ulong offsetd,
+    int m,
+    int n,
+    int k,
+    int n_no_padding,
+    uchar mask_d6,
+    uchar mask_d4,
+    uchar mask_hi2
+) {
+    dst = (global float *)((global char *)dst + offsetd);
+    int n_4  = n >> 2;
+    int gx   = get_global_id(0);     // output row
+    int sg   = get_local_id(1);      // subgroup index (K-split lane)
+    int lane = get_local_id(0);      // lane within subgroup (0..COK_SG-1)
+
+    int num_blocks_K = k / QK_K;
+    int num_32blk    = k / 32;
+
+    global const ushort * weight_ptr = src0_q + gx;
+    global const half   * d_ptr      = src0_d  + gx;
+    global const half   * dm_ptr     = src0_dm + gx;
+
+    half8 acc = 0;
+    half8 B;
+    half  dq;
+
+    for (int blk = sg; blk < num_32blk; blk += COK_NSG) {
+        int i       = blk << 5;       // blk * 32
+        int sb_idx  = blk >> 3;       // (blk*32) / QK_K  (QK_K = 256 = 32*8)
+        int sub_idx = blk & 7;        // (i/32) % 8
+
+        half dd  = d_ptr [sb_idx * m];
+        half dmm = dm_ptr[sb_idx * m];
+
+        global const uchar * sc0 = src0_s + sb_idx * K_SCALE_SIZE * m + gx;
+        uchar sv0, mn0;
+        get_scale_min_k4(sub_idx, sc0, m, &sv0, &mn0, mask_d6, mask_d4, mask_hi2);
+        half scale = convert_half(convert_float(dd)  * (float)sv0);
+        half mval  = convert_half(convert_float(dmm) * (float)mn0);
+
+        for (int l = 0; l < 32; l += 4) {
+            int ki = i + l;
+            ushort bits = weight_ptr[(ki>>2) * m];
+
+            B.s0123 = read_imageh(src1,     (ki+0) * n_4);
+            B.s4567 = read_imageh(src1, 1 + (ki+0) * n_4);
+            dq = (bits & 0x000F) * scale - mval;
+            acc += B * dq;
+
+            B.s0123 = read_imageh(src1,     (ki+1) * n_4);
+            B.s4567 = read_imageh(src1, 1 + (ki+1) * n_4);
+            dq = ((bits & 0x00F0) >> 4) * scale - mval;
+            acc += B * dq;
+
+            B.s0123 = read_imageh(src1,     (ki+2) * n_4);
+            B.s4567 = read_imageh(src1, 1 + (ki+2) * n_4);
+            dq = ((bits & 0x0F00) >> 8) * scale - mval;
+            acc += B * dq;
+
+            B.s0123 = read_imageh(src1,     (ki+3) * n_4);
+            B.s4567 = read_imageh(src1, 1 + (ki+3) * n_4);
+            dq = ((bits & 0xF000) >> 12) * scale - mval;
+            acc += B * dq;
+        }
+    }
+
+    // cross-subgroup reduction over the K-split (float for accuracy)
+    local float8 reduceLM[COK_SG * (COK_NSG - 1)];
+    if (sg > 0) {
+        reduceLM[(sg - 1) * COK_SG + lane] = convert_float8(acc);
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (sg == 0) {
+        float8 sum = convert_float8(acc);
+        for (int s = 0; s < COK_NSG - 1; s++) {
+            sum += reduceLM[s * COK_SG + lane];
+        }
+        int idx = gx;
+        if (idx < m*n_no_padding) { dst[idx] = sum.s0; idx += m; }
+        if (idx < m*n_no_padding) { dst[idx] = sum.s1; idx += m; }
+        if (idx < m*n_no_padding) { dst[idx] = sum.s2; idx += m; }
+        if (idx < m*n_no_padding) { dst[idx] = sum.s3; idx += m; }
+        if (idx < m*n_no_padding) { dst[idx] = sum.s4; idx += m; }
+        if (idx < m*n_no_padding) { dst[idx] = sum.s5; idx += m; }
+        if (idx < m*n_no_padding) { dst[idx] = sum.s6; idx += m; }
+        if (idx < m*n_no_padding) { dst[idx] = sum.s7; }
+    }
+}
@@ -5,6 +5,7 @@
 #pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
 #define ADRENO_GPU 1
 #define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
 #endif

 #ifdef ADRENO_GPU
@@ -138,3 +139,107 @@ kernel void kernel_gemm_noshuffle_q6_K_f32(
        vstore4((float4)(c0.s7, c1.s7, c2.s7, c3.s7), 0, dst + idx);
    }
 }
+
+// Cooperative-K q6_K GEMM for the small-batch (n_q in [2..8]) path. Same idea
+// as the q4_K _cok kernel: WG = (COK_SG lanes x COK_NSG subgroups), each lane
+// owns ONE output row (half8 over the 8 padded cols), and the COK_NSG
+// subgroups split the K iterations round-robin and combine via a __local
+// reduction. Replaces the default 4-row-per-WI tile that walked all of K alone
+// (~M/512 WGs + serial reduction) at small n_q. REQD_SUBGROUP_SIZE_64 +
+// barrier (never sub_group_reduce at full width on X2).
+#define COK_NSG 8
+#define COK_SG  64
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_gemm_noshuffle_q6_K_f32_cok(
+        global const ushort * src0_ql,
+        global const uchar  * src0_qh,
+        global const ushort * src0_s,
+        global const half   * src0_d,
+        read_only image1d_buffer_t src1,
+        global float * dst,
+        ulong offsetd,
+        int m,
+        int n,
+        int k,
+        int n_no_padding,
+        ushort mask_f000,
+        uchar  mask_c0
+) {
+    dst = (global float *)( (global char *)dst + offsetd );
+
+    int n_4  = n >> 2;
+    int gx   = get_global_id(0);     // output row
+    int sg   = get_local_id(1);      // subgroup index (K-split)
+    int lane = get_local_id(0);      // lane within subgroup
+
+    global const ushort * ptr_ql = src0_ql + gx;
+    global const uchar  * ptr_qh = src0_qh + gx;
+    global const ushort * ptr_s  = src0_s  + gx;
+    global const half   * ptr_d  = src0_d  + gx;
+
+    half8 acc = 0;
+    half8 B;
+    half  dq;
+
+    int num_iter = k >> 2;   // k/4 iterations, 4 k-values each
+
+    for (int ib = sg; ib < num_iter; ib += COK_NSG) {
+        int i = ib << 2;     // ib * 4
+
+        ushort bits4 = ptr_ql[ib * m];          // ql for row gx at this 4-block
+        uchar  bits2 = ptr_qh[ib * m];          // qh
+
+        ushort s_packed = ptr_s[(i >> 5) * m];  // (i/16/2) = i/32
+        char2  sc2      = as_char2(s_packed);
+        char   scale_s  = (((i >> 4) & 1) == 0) ? sc2.s0 : sc2.s1; // (i/16)%2
+        half   scale_d  = ptr_d[(i >> 8) * m];  // i/256
+
+        // j=0
+        B.s0123 = read_imageh(src1, (i + 0)*n_4 + 0);
+        B.s4567 = read_imageh(src1, (i + 0)*n_4 + 1);
+        dq = (convert_half((bits4 & 0x000F) | ((bits2 & 0x03) << 4)) - 32.f) * scale_s * scale_d;
+        acc += B * dq;
+
+        // j=1
+        B.s0123 = read_imageh(src1, (i + 1)*n_4 + 0);
+        B.s4567 = read_imageh(src1, (i + 1)*n_4 + 1);
+        dq = (convert_half(((bits4 & 0x00F0) >> 4) | ((bits2 & 0x0C) << 2)) - 32.f) * scale_s * scale_d;
+        acc += B * dq;
+
+        // j=2
+        B.s0123 = read_imageh(src1, (i + 2)*n_4 + 0);
+        B.s4567 = read_imageh(src1, (i + 2)*n_4 + 1);
+        dq = (convert_half(((bits4 & 0x0F00) >> 8) | (bits2 & 0x30)) - 32.f) * scale_s * scale_d;
+        acc += B * dq;
+
+        // j=3
+        B.s0123 = read_imageh(src1, (i + 3)*n_4 + 0);
+        B.s4567 = read_imageh(src1, (i + 3)*n_4 + 1);
+        dq = (convert_half(((bits4 & mask_f000) >> 12) | ((bits2 & mask_c0) >> 2)) - 32.f) * scale_s * scale_d;
+        acc += B * dq;
+    }
+
+    local float8 reduceLM[COK_SG * (COK_NSG - 1)];
+    if (sg > 0) {
+        reduceLM[(sg - 1) * COK_SG + lane] = convert_float8(acc);
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (sg == 0) {
+        float8 sum = convert_float8(acc);
+        for (int s = 0; s < COK_NSG - 1; s++) {
+            sum += reduceLM[s * COK_SG + lane];
+        }
+        int idx = gx;
+        if (idx < m*n_no_padding) { dst[idx] = sum.s0; idx += m; }
+        if (idx < m*n_no_padding) { dst[idx] = sum.s1; idx += m; }
+        if (idx < m*n_no_padding) { dst[idx] = sum.s2; idx += m; }
+        if (idx < m*n_no_padding) { dst[idx] = sum.s3; idx += m; }
+        if (idx < m*n_no_padding) { dst[idx] = sum.s4; idx += m; }
+        if (idx < m*n_no_padding) { dst[idx] = sum.s5; idx += m; }
+        if (idx < m*n_no_padding) { dst[idx] = sum.s6; idx += m; }
+        if (idx < m*n_no_padding) { dst[idx] = sum.s7; }
+    }
+}
@@ -0,0 +1,136 @@
+// Batched (N>1) q6_K GEMM over the 64-row-TILED canonical layout produced by
+// kernel_convert_block_q6_k_tiled_ns (cvt.cl). Companion to the decode kernel
+// kernel_gemv_noshuffle_q6_K_f32_tiled: SAME pack, SAME canonical e-order
+// dequant (correct by construction vs reference ggml q6_K), extended to N output
+// columns. Makes the batched lm_head/embed (perplexity, spec-decode verify,
+// batched serving) correct on GPU while keeping the tiled convert the fast decode
+// GEMV depends on.
+//
+// One work-item owns one output ROW for a block of BN columns. A work-group is
+// {64 lanes, NTILES subgroups} = NTILES*64 rows; the global z dimension tiles the
+// N columns by BN. Each work-item computes its row's FULL K (no K-split, so no
+// cross-subgroup reduction), which lets the whole work-group share one staged
+// activation block:
+//
+//   __local activation staging — the BN columns of the current superblock (BN*256
+//   floats) are loaded into __local once per superblock, cooperatively by all
+//   NTILES*64 work-items, then every row reads its activation from __local. This
+//   removes the ~Nrows-fold redundant image reads of the first version (each lane
+//   re-read the activation), which made the batched GEMM ~2x slower than the plain
+//   noshuffle GEMM.
+//
+// Weights are read from __global (coalesced) — matching the decode kernel; the
+// lm_head weight is streamed with little reuse where coalesced global beats the
+// Adreno texture cache.
+
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#endif
+
+#define NTILES     4        // 64-row tiles per work-group (NTILES*64 = 256 rows)
+#define TILE_ROWS  64
+#define BN         16       // output columns handled per work-group (global z step)
+#define WG_THREADS (NTILES * TILE_ROWS)
+
+#if defined(ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_gemm_noshuffle_q6_K_f32_tiled(
+    __global uint4 * src0_ql,   // tiled: 8 uint4 granules / superblock
+    __global uint4 * src0_qh,   // tiled: 4 uint4 granules / superblock
+    __global char  * src0_s,    // tiled: 16 chars / superblock
+    __global half  * src0_d,    // tiled: 1 half  / superblock
+    read_only image1d_buffer_t src1,   // activation [ne00, ne11] f32 (RGBA), column-major
+    global float * dst,
+    ulong offsetd,
+    int ne00,
+    int ne01,
+    int ne11
+) {
+    int rit = get_local_id(0);          // 0..63  (lane within a tile; coalesces weight loads)
+    int sg  = get_local_id(1);          // 0..NTILES-1
+    int lid = sg * TILE_ROWS + rit;     // 0..WG_THREADS-1 (flat local id)
+    int row = get_group_id(0) * WG_THREADS + lid;
+    int rt  = row / TILE_ROWS;          // global 64-row tile index
+    int col0 = get_global_id(2) * BN;   // first output column of this block
+
+    int nb = ne00 / 256;                // superblocks per row
+    int act_col_stride = ne00 / 4;      // activation float4 pixels per column
+
+    const bool row_ok = row < ne01;
+
+    // staged activation: BN columns x 256 elements for the current superblock
+    __local float lact[BN * 256];
+
+    float acc[BN];
+    #pragma unroll
+    for (int j = 0; j < BN; ++j) acc[j] = 0.0f;
+
+    for (int sb = 0; sb < nb; ++sb) {
+        // cooperatively stage BN columns' 256 activation elements (= BN*64 float4)
+        for (int p = lid; p < BN * 64; p += WG_THREADS) {
+            int j  = p >> 6;            // column within the BN block (p / 64)
+            int e4 = p & 63;            // element-quad within the column (p % 64)
+            int c  = col0 + j;
+            float4 v = (c < ne11)
+                ? read_imagef(src1, c * act_col_stride + sb * 64 + e4)
+                : (float4)(0.0f);
+            lact[p * 4 + 0] = v.x;
+            lact[p * 4 + 1] = v.y;
+            lact[p * 4 + 2] = v.z;
+            lact[p * 4 + 3] = v.w;      // lact[j*256 + e], e = e4*4 + t
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        if (row_ok) {
+            int tile_blk = rt * nb + sb;    // ne02 == 1 for lm_head/embed
+
+            float dval = (float)src0_d[tile_blk * TILE_ROWS + rit];
+            __global char * sc = src0_s + (tile_blk * TILE_ROWS + rit) * 16;
+
+            uint ql[32];
+            uint qh[16];
+            #pragma unroll
+            for (int g = 0; g < 8; ++g) {
+                uint4 v = src0_ql[(tile_blk * 8 + g) * TILE_ROWS + rit];
+                ql[g*4+0] = v.x; ql[g*4+1] = v.y; ql[g*4+2] = v.z; ql[g*4+3] = v.w;
+            }
+            #pragma unroll
+            for (int g = 0; g < 4; ++g) {
+                uint4 v = src0_qh[(tile_blk * 4 + g) * TILE_ROWS + rit];
+                qh[g*4+0] = v.x; qh[g*4+1] = v.y; qh[g*4+2] = v.z; qh[g*4+3] = v.w;
+            }
+
+            // NOTE: the e loop (256) is deliberately NOT unrolled. Fully unrolling
+            // 256*BN MACs overflows the in-process Adreno compiler (host stack
+            // overflow at clBuildProgram, same class as the FA DK=512 OOM).
+            for (int e = 0; e < 256; ++e) {
+                uint low4 = (ql[e >> 3] >> ((e & 7) * 4)) & 0xF;
+                uint hi2  = (qh[e >> 4] >> ((e & 15) * 2)) & 0x3;
+                int  code = (int)(low4 | (hi2 << 4)) - 32;
+                int  sidx = ((e >> 7) << 3) + (((e >> 5) & 3) << 1) + ((e >> 4) & 1);
+                float cs  = (float)code * (float)sc[sidx] * dval;
+                #pragma unroll
+                for (int j = 0; j < BN; ++j) {
+                    acc[j] += cs * lact[j * 256 + e];
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    if (row_ok) {
+        dst = (global float*)((global char*)dst + offsetd);
+        #pragma unroll
+        for (int j = 0; j < BN; ++j) {
+            int c = col0 + j;
+            if (c < ne11) {
+                dst[(ulong)c * ne01 + row] = acc[j];
+            }
+        }
+    }
+}
@@ -277,3 +277,107 @@ __kernel void kernel_gemv_noshuffle_q4_0_f32(
    }

 }
+
+// Multi-column (N in [2..4]) variant of the q4_0 decode GEMV, for the speculative
+// / MTP verify batch (n_cols = 2..4 = drafted + bonus positions). Routes the small-
+// batch verify OFF the transposed-GEMM dead-zone (gemm_noshuffle_q4_0) onto the
+// efficient GEMV path. Each K-block's weights (regA hi+lo) are loaded ONCE and
+// reused across the n_cols activation columns. Per-column accumulation is
+// independent and identical to n_cols standalone GEMVs. n_cols==3 is byte-identical
+// to the original mc3 (col3 disabled, slots 6/7 stay zero). Kept the _mc3 name.
+#ifdef VECTOR_SUB_GROUP_BROADCAST
+#define MC_DQ_HI dequantizeBlockAccum_ns_sgbroadcast_8_hi
+#define MC_DQ_LO dequantizeBlockAccum_ns_sgbroadcast_8_lo
+#else
+#define MC_DQ_HI dequantizeBlockAccum_ns_sgbroadcast_1_hi
+#define MC_DQ_LO dequantizeBlockAccum_ns_sgbroadcast_1_lo
+#endif
+// One column c: load this column's activation (own brace scope so the macros'
+// `shared_y` decl is re-scoped), then dequant (hi+lo) against the shared weights.
+#define MC_COL_Q40(ts, c) \
+    { if (slid < 4) { regB.s0123 = read_imagef(src1, (c)*COL_STRIDE     + slid*2 + k*8); \
+                      regB.s4567 = read_imagef(src1, (c)*COL_STRIDE + 1 + slid*2 + k*8); } \
+      MC_DQ_HI(ts, as_ushort8(regA_hi), regS, regB); \
+      MC_DQ_LO(ts, as_ushort8(regA_lo), regS, regB); }
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+__kernel void kernel_gemv_noshuffle_q4_0_f32_mc3(
+        __read_only  image1d_buffer_t src0_q,  // quantized A
+        global half2  * src0_d,                // A scales
+        __read_only  image1d_buffer_t src1,    // B (n_cols columns, col-major image)
+        global float * dst,                    // C (column-major [M x n_cols])
+        ulong offsetd,
+        int ne00,                              // K
+        int ne01,                              // M
+        int n_cols)                            // N (2..4)
+{
+    uint groupId = get_local_id(1);
+    uint gid     = get_global_id(0);
+    ushort slid  = get_sub_group_local_id();
+
+    uint K = ne00;
+    uint M = ne01;
+
+    uint LINE_STRIDE_A  = M / 2;
+    // BLOCK_STRIDE_A is the LAYOUT stride between consecutive K-blocks = 4 uints
+    // per q4_0 block * M (set by the trans4_ns convert). The "4" is uints/block, NOT
+    // the subgroup count — keep it fixed so the K-split count (nsg) can vary.
+    uint BLOCK_STRIDE_A = N_SIMDGROUP * M;   // = 4 * M (N_SIMDGROUP is the #define 4)
+    uint COL_STRIDE     = K / 4;   // float4 pixels per activation column
+    uint nsg            = get_local_size(1);  // runtime K-split (4 default, 8 small-M)
+
+    __private uint4  regA_hi, regA_lo;
+    __private half2  regS;
+    __private float8 regB;
+
+    __private float2 ts0 = (float2)(0.0f);
+    __private float2 ts1 = (float2)(0.0f);
+    __private float2 ts2 = (float2)(0.0f);
+    __private float2 ts3 = (float2)(0.0f);
+
+    for (uint k = groupId; k < (K / QK4_0); k += nsg) {
+        regS = src0_d[gid + k * LINE_STRIDE_A];
+
+        // weights loaded ONCE, reused across the columns
+        regA_hi.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
+        regA_hi.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
+        regA_hi.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
+        regA_hi.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
+        regA_lo.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
+        regA_lo.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
+        regA_lo.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
+        regA_lo.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
+
+        MC_COL_Q40(ts0, 0);
+        MC_COL_Q40(ts1, 1);
+        if (n_cols > 2) MC_COL_Q40(ts2, 2);
+        if (n_cols > 3) MC_COL_Q40(ts3, 3);
+    }
+
+    // cross-subgroup reduce over nsg subgroups: pack the (up to 4) columns' float2
+    // into a float8. Generalized to runtime nsg (4 default, 8 for small-M). Each
+    // subgroup writes its partial; subgroup 0 sums the rest into its own acc. At
+    // nsg==4 this is byte-identical to the original (sums subgroups 1,2,3 in order).
+    __local float8 reduceLM[SIMDGROUP_WIDTH * 8];
+    float8 acc = (float8)(ts0.s0, ts0.s1, ts1.s0, ts1.s1, ts2.s0, ts2.s1, ts3.s0, ts3.s1);
+    reduceLM[groupId * SIMDGROUP_WIDTH + slid] = acc;
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (groupId == 0) {
+        for (uint g = 1; g < nsg; g++) {
+            acc += reduceLM[g * SIMDGROUP_WIDTH + slid];
+        }
+        dst = (global float*)((global char*)dst + offsetd);
+        // dst is column-major [M rows x n_cols cols]: (row, col) at col*M + row
+        vstore2((float2)(acc.s0, acc.s1), 0, &(dst[0 * M + gid * 2]));
+        vstore2((float2)(acc.s2, acc.s3), 0, &(dst[1 * M + gid * 2]));
+        if (n_cols > 2) vstore2((float2)(acc.s4, acc.s5), 0, &(dst[2 * M + gid * 2]));
+        if (n_cols > 3) vstore2((float2)(acc.s6, acc.s7), 0, &(dst[3 * M + gid * 2]));
+    }
+}
+#undef MC_COL_Q40
+#undef MC_DQ_HI
+#undef MC_DQ_LO
@@ -286,3 +286,99 @@ kernel void kernel_gemv_noshuffle_q4_1_f32(
    }

 }
+
+// Multi-column (N in [2..4]) variant of the q4_1 decode GEMV (spec/MTP verify) =
+// q4_0 mc3 + the q4_1 per-block min (regM; dequant = q*scale + minv). n_cols=2..4;
+// routes the small-batch verify OFF the gemm_noshuffle_q4_1 dead-zone. n_cols==3 is
+// byte-identical to the original mc3. NB: this file spells the vec-broadcast define
+// BROADCAT (no S) — match it so the fast _8 path compiles.
+#ifdef VECTOR_SUB_GROUP_BROADCAT
+#define MC_DQ1_HI dequantizeBlockAccum_ns_sgbroadcast_8_hi
+#define MC_DQ1_LO dequantizeBlockAccum_ns_sgbroadcast_8_lo
+#else
+#define MC_DQ1_HI dequantizeBlockAccum_ns_sgbroadcast_1_hi
+#define MC_DQ1_LO dequantizeBlockAccum_ns_sgbroadcast_1_lo
+#endif
+#define MC_COL_Q41(ts, c) \
+    { if (slid < 4) { regB.s0123 = read_imagef(src1, (c)*COL_STRIDE     + slid*2 + k*8); \
+                      regB.s4567 = read_imagef(src1, (c)*COL_STRIDE + 1 + slid*2 + k*8); } \
+      MC_DQ1_HI(ts, as_ushort8(regA_hi), regS, regM, regB); \
+      MC_DQ1_LO(ts, as_ushort8(regA_lo), regS, regM, regB); }
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_gemv_noshuffle_q4_1_f32_mc3(
+        read_only  image1d_buffer_t src0_q,
+        global half2  * src0_d,
+        global half2  * src0_m,
+        read_only  image1d_buffer_t src1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int n_cols)
+{
+    uint groupId = get_local_id(1);
+    uint gid     = get_global_id(0);
+    ushort slid  = get_sub_group_local_id();
+
+    uint K = ne00;
+    uint M = ne01;
+
+    uint LINE_STRIDE_A  = M / 2;
+    uint BLOCK_STRIDE_A = NSUBGROUPS * M;
+    uint COL_STRIDE     = K / 4;   // float4 pixels per activation column
+
+    private uint4  regA_hi, regA_lo;
+    private half2  regS, regM;
+    private float8 regB;
+
+    private float2 ts0 = (float2)(0.0f);
+    private float2 ts1 = (float2)(0.0f);
+    private float2 ts2 = (float2)(0.0f);
+    private float2 ts3 = (float2)(0.0f);
+
+    for (uint k = groupId; k < (K / QK4_0); k += NSUBGROUPS) {
+        regS = src0_d[gid + k * LINE_STRIDE_A];
+        regM = src0_m[gid + k * LINE_STRIDE_A];
+
+        // weights loaded ONCE, reused across the columns
+        regA_hi.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
+        regA_hi.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
+        regA_hi.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
+        regA_hi.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
+        regA_lo.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
+        regA_lo.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
+        regA_lo.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
+        regA_lo.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
+
+        MC_COL_Q41(ts0, 0);
+        MC_COL_Q41(ts1, 1);
+        if (n_cols > 2) MC_COL_Q41(ts2, 2);
+        if (n_cols > 3) MC_COL_Q41(ts3, 3);
+    }
+
+    // cross-subgroup reduce: pack the (up to 4) columns' float2 into a float8.
+    local float8 reduceLM[SUBGROUP_SIZE * 3];
+    float8 acc = (float8)(ts0.s0, ts0.s1, ts1.s0, ts1.s1, ts2.s0, ts2.s1, ts3.s0, ts3.s1);
+    if (groupId == 1) { reduceLM[SUBGROUP_SIZE * 0 + slid] = acc; }
+    if (groupId == 2) { reduceLM[SUBGROUP_SIZE * 1 + slid] = acc; }
+    if (groupId == 3) { reduceLM[SUBGROUP_SIZE * 2 + slid] = acc; }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (groupId == 0) {
+        acc += reduceLM[SUBGROUP_SIZE * 0 + slid];
+        acc += reduceLM[SUBGROUP_SIZE * 1 + slid];
+        acc += reduceLM[SUBGROUP_SIZE * 2 + slid];
+        dst = (global float*)((global char*)dst + offsetd);
+        // dst is column-major [M rows x n_cols cols]: (row, col) at col*M + row
+        vstore2((float2)(acc.s0, acc.s1), 0, &(dst[0 * M + gid * 2]));
+        vstore2((float2)(acc.s2, acc.s3), 0, &(dst[1 * M + gid * 2]));
+        if (n_cols > 2) vstore2((float2)(acc.s4, acc.s5), 0, &(dst[2 * M + gid * 2]));
+        if (n_cols > 3) vstore2((float2)(acc.s6, acc.s7), 0, &(dst[3 * M + gid * 2]));
+    }
+}
+#undef MC_COL_Q41
+#undef MC_DQ1_HI
+#undef MC_DQ1_LO
@@ -228,12 +228,37 @@ kernel void kernel_gemv_noshuffle_q4_k_f32(
    uint groupId = get_local_id(1);
    uint gid     = get_global_id(0);
    ushort slid  = get_sub_group_local_id();
+    // K-split factor = #subgroups in the WG. Read from the launch (NOT a compile
+    // constant) so small-M projections (Kcur/Vcur/Qcur) can dispatch a wider
+    // K-split (more waves/SP -> latency hiding) while large-M keeps 4. The
+    // physical weight layout stride below is INDEPENDENT of this (see BLOCK_STRIDE_A).
+    uint nsg = get_local_size(1);

    uint K = ne00;
    uint M = ne01;

    uint LINE_STRIDE_A  = M / 2;
-    uint BLOCK_STRIDE_A = NSUBGROUPS * M;
+    // Physical per-K-block stride in the packed image: 8 uints/block-row-pair *
+    // (M/2) row-pairs = 4*M uints. This is a layout constant, not tied to nsg.
+    uint BLOCK_STRIDE_A = 4 * M;
+    uint scales_per_row = (K / QK_K) * 12;
+
+    // The x-grid is padded to CEIL_DIV(ne01/2,64)*64, so when ne01 % 128 != 0 the
+    // tail lanes hold gid >= ne01/2. The output stores below are guarded, but the
+    // input fetches are not: src0_d and src0_m are raw global half2 pointers,
+    // src0_s is a raw global uchar pointer, and read_imageui on an
+    // image1d_buffer_t is UNDEFINED out of range -- an image clamps only for
+    // SAMPLER reads, which these are not. Those lanes therefore read past the end
+    // of all three allocations. For a [2816, 2112] weight (2112 % 128 == 64) the
+    // top tail lane is gid = 1087 while only gid < 1056 is backed, and it runs
+    // 32 half2 past src0_d/src0_m, 31 uints past the quant image, and 63 bytes
+    // past src0_s.
+    //
+    // Clamp the row used for every fetch. The lanes stay ACTIVE, which the
+    // sub_group_broadcast in the dequant macros requires, and their results are
+    // still discarded by the existing output guard. No-op and byte-identical
+    // whenever ne01 % 128 == 0.
+    uint gid_s = min(gid, LINE_STRIDE_A - 1);

    private uint4     regA;
    private half2     regS;
@@ -242,6 +267,437 @@ kernel void kernel_gemv_noshuffle_q4_k_f32(

    private float2 totalSum = (float2)(0.0f);

+    for (uint k = groupId; k < (K / 32); k += nsg) {
+        uint sb = k / 8;
+        uint j  = k % 8;
+
+        half2 d   = src0_d[gid_s + sb * LINE_STRIDE_A];
+        half2 dm  = src0_m[gid_s + sb * LINE_STRIDE_A];
+
+        global const uchar * sc0 = src0_s + sb * 12 * M + 2 * gid_s;
+        global const uchar * sc1 = sc0 + 1;
+
+        uchar sv0, mn0, sv1, mn1;
+        get_scale_min_k4(j, sc0, M, &sv0, &mn0, mask_d6, mask_d4, mask_hi2);
+        get_scale_min_k4(j, sc1, M, &sv1, &mn1, mask_d6, mask_d4, mask_hi2);
+
+        regS = convert_half2(convert_float2(d)  * convert_float2((uchar2)(sv0, sv1)));
+        regM = convert_half2(convert_float2(dm) * convert_float2((uchar2)(mn0, mn1)));
+
+        if (slid < 4) {
+            regB.s0123 = read_imagef(src1, (slid * 2 + k * 8));
+            regB.s4567 = read_imagef(src1, (1 + slid * 2 + k * 8));
+        }
+
+        // load half weights for two blocks in consecutive rows
+        regA.s0 = read_imageui(src0_q, (gid_s + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
+        regA.s1 = read_imageui(src0_q, (gid_s + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
+        regA.s2 = read_imageui(src0_q, (gid_s + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
+        regA.s3 = read_imageui(src0_q, (gid_s + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAST
+        dequantizeBlockAccum_ns_sgbroadcast_8_hi(totalSum, as_ushort8(regA), regS, regM, regB);
+#else
+        dequantizeBlockAccum_ns_sgbroadcast_1_hi(totalSum, as_ushort8(regA), regS, regM, regB);
+#endif // VECTOR_SUB_GROUP_BROADCAST
+
+        regA.s0 = read_imageui(src0_q, (gid_s + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
+        regA.s1 = read_imageui(src0_q, (gid_s + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
+        regA.s2 = read_imageui(src0_q, (gid_s + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
+        regA.s3 = read_imageui(src0_q, (gid_s + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAST
+        dequantizeBlockAccum_ns_sgbroadcast_8_lo(totalSum, as_ushort8(regA), regS, regM, regB);
+#else
+        dequantizeBlockAccum_ns_sgbroadcast_1_lo(totalSum, as_ushort8(regA), regS, regM, regB);
+#endif // VECTOR_SUB_GROUP_BROADCAST
+    }
+
+    // Cross-subgroup reduction in local memory. Generalized to nsg subgroups
+    // (was a hard-coded 4-wave unroll). Sized for up to 16 subgroups (the widest
+    // K-split we dispatch for small M). At nsg==4 the accumulation order is
+    // identical to the original unroll -> byte-identical for the large-M path.
+    local float2 reduceLM[SUBGROUP_SIZE * 15];
+    if (groupId > 0) {
+        reduceLM[SUBGROUP_SIZE * (groupId - 1) + slid] = totalSum;
+    }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (groupId == 0) {
+        for (uint i = 0; i < nsg - 1; ++i) {
+            totalSum += reduceLM[SUBGROUP_SIZE * i + slid];
+        }
+    }
+
+    // 2 outputs per fiber in wave 0
+    if (groupId == 0) {
+        dst = (global float*)((global char*)dst + offsetd);
+        // Guard the two output rows. The x-grid is padded to CEIL_DIV(ne01/2,64)*64,
+        // so when ne01 is not a multiple of 128 the tail row-pairs run past row ne01
+        // and would overrun dst into the adjacent tensor. No-op / byte-identical when
+        // ne01 % 128 == 0 (M/2 already a multiple of 64 -> no padding).
+        if (gid * 2 + 0 < M) dst[gid * 2 + 0] = totalSum.s0;
+        if (gid * 2 + 1 < M) dst[gid * 2 + 1] = totalSum.s1;
+    }
+
+}
+
+// --- Fused gate+up GEMV + GLU epilogue (FFN) ------------------------------------
+// Folds the FFN's two decode GEMVs (ffn_gate, ffn_up) and the following GLU into a
+// SINGLE dispatch: {MUL_MAT(Wg,x), MUL_MAT(Wu,x), GLU}. Both matmuls share the same
+// activation x (ffn_norm), so the activation image read is issued ONCE per K-block
+// and reused for the gate and up dot products (the per-op path re-reads it twice and
+// also materializes the two full ffn-wide intermediates to global, which the GLU
+// then re-reads). The gate/up partial sums are accumulated in the SAME per-fiber
+// order and reduced in the SAME cross-subgroup order as the standalone GEMV, and the
+// GLU formula is the exact scalar expression from kernels/glu.cl, so the output is
+// BYTE-IDENTICAL to the per-op matmul+matmul+glu path -> safe to default on.
+//   glu_op: REGLU=0, GEGLU=1, SWIGLU=2, GEGLU_ERF=4, GEGLU_QUICK=5 (ggml_glu_op).
+// Weights: src0g_* = gate (= GLU src[0]); src0u_* = up (= GLU src[1]).
+#define GLU_GEGLU_COEF_A      0.044715f
+#define GLU_SQRT_2_OVER_PI    0.79788456080286535587989211986876f
+#define GLU_SQRT_2_INV        0.70710678118654752440084436210484f
+#define GLU_QUICK_COEF       -1.702f
+
+inline float glu_apply(int glu_op, float g, float u) {
+    float act;
+    if (glu_op == 1) {        // GEGLU (tanh-approx gelu)
+        act = 0.5f*g*(1.0f + tanh(GLU_SQRT_2_OVER_PI*g*(1.0f + GLU_GEGLU_COEF_A*g*g)));
+    } else if (glu_op == 2) { // SWIGLU (silu)
+        act = g / (1.0f + exp(-g));
+    } else if (glu_op == 0) { // REGLU
+        return g*u*(g > 0.0f);
+    } else if (glu_op == 4) { // GEGLU_ERF
+        act = 0.5f*g*(1.0f + erf(g*GLU_SQRT_2_INV));
+    } else {                  // GEGLU_QUICK (glu_op == 5)
+        act = g*(1.0f/(1.0f + exp(GLU_QUICK_COEF*g)));
+    }
+    return act*u;
+}
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_gemv_noshuffle_q4_k_f32_glu(
+        read_only  image1d_buffer_t src0g_q,
+        global half2  * src0g_d,
+        global half2  * src0g_m,
+        global uchar  * src0g_s,
+        read_only  image1d_buffer_t src0u_q,
+        global half2  * src0u_d,
+        global half2  * src0u_m,
+        global uchar  * src0u_s,
+        read_only  image1d_buffer_t src1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int glu_op,
+        uchar mask_d6,
+        uchar mask_d4,
+        uchar mask_hi2)
+{
+    uint groupId = get_local_id(1);
+    uint gid     = get_global_id(0);
+    ushort slid  = get_sub_group_local_id();
+    uint nsg     = get_local_size(1);
+
+    uint K = ne00;
+    uint M = ne01;
+
+    uint LINE_STRIDE_A  = M / 2;
+    uint BLOCK_STRIDE_A = 4 * M;
+
+    private uint4  regA;
+    private half2  regS, regM;
+    private float8 regB;
+
+    private float2 gateSum = (float2)(0.0f);
+    private float2 upSum   = (float2)(0.0f);
+
+    // Two SEQUENTIAL K-loops (gate fully, then up). Keeping only one weight's
+    // working set live at a time holds the kernel's register footprint at ~the
+    // base single-weight GEMV's, so its max WG stays 1024 (16 subgroups) and the
+    // per-subgroup K-split matches the standalone wide GEMV exactly -> the gate
+    // and up partial sums are BYTE-IDENTICAL to the per-op path. The macro body
+    // is the base kernel's inner loop verbatim, parameterized by weight source.
+#define Q4K_GLU_LOOP(SUM, Q, DD, MM, SS)                                                       \
+    for (uint k = groupId; k < (K / 32); k += nsg) {                                           \
+        uint sb = k / 8;                                                                       \
+        uint j  = k % 8;                                                                       \
+        half2 d   = DD[gid + sb * LINE_STRIDE_A];                                              \
+        half2 dm  = MM[gid + sb * LINE_STRIDE_A];                                              \
+        global const uchar * sc0 = SS + sb * 12 * M + 2 * gid;                                 \
+        global const uchar * sc1 = sc0 + 1;                                                    \
+        uchar sv0, mn0, sv1, mn1;                                                              \
+        get_scale_min_k4(j, sc0, M, &sv0, &mn0, mask_d6, mask_d4, mask_hi2);                      \
+        get_scale_min_k4(j, sc1, M, &sv1, &mn1, mask_d6, mask_d4, mask_hi2);                      \
+        regS = convert_half2(convert_float2(d)  * convert_float2((uchar2)(sv0, sv1)));         \
+        regM = convert_half2(convert_float2(dm) * convert_float2((uchar2)(mn0, mn1)));         \
+        if (slid < 4) {                                                                        \
+            regB.s0123 = read_imagef(src1, (slid * 2 + k * 8));                                \
+            regB.s4567 = read_imagef(src1, (1 + slid * 2 + k * 8));                            \
+        }                                                                                      \
+        regA.s0 = read_imageui(Q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;           \
+        regA.s1 = read_imageui(Q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;           \
+        regA.s2 = read_imageui(Q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;           \
+        regA.s3 = read_imageui(Q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;           \
+        DEQ_HI(SUM, as_ushort8(regA), regS, regM, regB);                                       \
+        regA.s0 = read_imageui(Q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;           \
+        regA.s1 = read_imageui(Q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;           \
+        regA.s2 = read_imageui(Q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;           \
+        regA.s3 = read_imageui(Q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;           \
+        DEQ_LO(SUM, as_ushort8(regA), regS, regM, regB);                                       \
+    }
+
+#ifdef VECTOR_SUB_GROUP_BROADCAST
+#define DEQ_HI dequantizeBlockAccum_ns_sgbroadcast_8_hi
+#define DEQ_LO dequantizeBlockAccum_ns_sgbroadcast_8_lo
+#else
+#define DEQ_HI dequantizeBlockAccum_ns_sgbroadcast_1_hi
+#define DEQ_LO dequantizeBlockAccum_ns_sgbroadcast_1_lo
+#endif
+
+    Q4K_GLU_LOOP(gateSum, src0g_q, src0g_d, src0g_m, src0g_s)
+    Q4K_GLU_LOOP(upSum,   src0u_q, src0u_d, src0u_m, src0u_s)
+
+#undef DEQ_HI
+#undef DEQ_LO
+#undef Q4K_GLU_LOOP
+
+    // Cross-subgroup reduction in local memory. Packs gate (xy) + up (zw) into a
+    // float4 so both reduce in one pass; summation order matches the base GEMV's
+    // per-channel loop -> byte-identical partial sums.
+    local float4 reduceLM[SUBGROUP_SIZE * 15];
+    if (groupId > 0) {
+        reduceLM[SUBGROUP_SIZE * (groupId - 1) + slid] = (float4)(gateSum, upSum);
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+    if (groupId == 0) {
+        for (uint i = 0; i < nsg - 1; ++i) {
+            float4 p = reduceLM[SUBGROUP_SIZE * i + slid];
+            gateSum += p.xy;
+            upSum   += p.zw;
+        }
+        dst = (global float*)((global char*)dst + offsetd);
+        dst[gid * 2 + 0] = glu_apply(glu_op, gateSum.s0, upSum.s0);
+        dst[gid * 2 + 1] = glu_apply(glu_op, gateSum.s1, upSum.s1);
+    }
+}
+
+// --- Split-K-across-workgroups decode GEMV (small-M projections) ----------------
+// A single-token GEMV makes only ceil(M/2/64) workgroups; a WG runs on one Adreno
+// compute unit, so for small M (Kcur/Vcur, M=512 -> 4 WGs) most of the 16 CUs sit
+// idle and the matmul is bandwidth-starved even with a wide intra-WG K-split. This
+// variant adds a SECOND grid dimension of `ksplit` workgroups that each reduce a
+// disjoint slice of K and write a per-slice partial; kernel_gemv_splitk_reduce_f32
+// then sums the partials into dst. Identical math/layout to the base kernel
+// (physical block stride 4*M, get_scale_min_k4) -> coherent. Gated host-side to
+// M<=1024 (M>=2048
+// already fills the CUs and the extra reduce dispatch only hurts).
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_gemv_noshuffle_q4_k_f32_splitk(
+        read_only  image1d_buffer_t src0_q,
+        global half2  * src0_d,
+        global half2  * src0_m,
+        global uchar  * src0_s,
+        read_only  image1d_buffer_t src1,
+        global float * partial,          // [ksplit * M], slice-major
+        int ne00,
+        int ne01,
+        uchar mask_d6,
+        uchar mask_d4,
+        uchar mask_hi2)
+{
+    uint groupId = get_local_id(1);
+    uint gid     = get_global_id(0);
+    ushort slid  = get_sub_group_local_id();
+    uint nsg     = get_local_size(1);
+    uint ksplit  = get_num_groups(1);
+    uint kslice  = get_group_id(1);
+
+    uint K = ne00;
+    uint M = ne01;
+    uint LINE_STRIDE_A  = M / 2;
+    uint BLOCK_STRIDE_A = 4 * M;      // physical, independent of the K-split
+
+    private uint4  regA;
+    private half2  regS, regM;
+    private float8 regB;
+    private float2 totalSum = (float2)(0.0f);
+
+    // each (kslice, subgroup) pair owns a disjoint set of K-blocks
+    for (uint k = kslice * nsg + groupId; k < (K / 32); k += ksplit * nsg) {
+        uint sb = k / 8;
+        uint j  = k % 8;
+        half2 d   = src0_d[gid + sb * LINE_STRIDE_A];
+        half2 dm  = src0_m[gid + sb * LINE_STRIDE_A];
+        global const uchar * sc0 = src0_s + sb * 12 * M + 2 * gid;
+        global const uchar * sc1 = sc0 + 1;
+        uchar sv0, mn0, sv1, mn1;
+        get_scale_min_k4(j, sc0, M, &sv0, &mn0, mask_d6, mask_d4, mask_hi2);
+        get_scale_min_k4(j, sc1, M, &sv1, &mn1, mask_d6, mask_d4, mask_hi2);
+        regS = convert_half2(convert_float2(d)  * convert_float2((uchar2)(sv0, sv1)));
+        regM = convert_half2(convert_float2(dm) * convert_float2((uchar2)(mn0, mn1)));
+        if (slid < 4) {
+            regB.s0123 = read_imagef(src1, (slid * 2 + k * 8));
+            regB.s4567 = read_imagef(src1, (1 + slid * 2 + k * 8));
+        }
+        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
+        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
+        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
+        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAST
+        dequantizeBlockAccum_ns_sgbroadcast_8_hi(totalSum, as_ushort8(regA), regS, regM, regB);
+#else
+        dequantizeBlockAccum_ns_sgbroadcast_1_hi(totalSum, as_ushort8(regA), regS, regM, regB);
+#endif
+        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
+        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
+        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
+        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAST
+        dequantizeBlockAccum_ns_sgbroadcast_8_lo(totalSum, as_ushort8(regA), regS, regM, regB);
+#else
+        dequantizeBlockAccum_ns_sgbroadcast_1_lo(totalSum, as_ushort8(regA), regS, regM, regB);
+#endif
+    }
+
+    local float2 reduceLM[SUBGROUP_SIZE * 15];
+    if (groupId > 0) {
+        reduceLM[SUBGROUP_SIZE * (groupId - 1) + slid] = totalSum;
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+    if (groupId == 0) {
+        for (uint i = 0; i < nsg - 1; ++i) {
+            totalSum += reduceLM[SUBGROUP_SIZE * i + slid];
+        }
+        vstore2(totalSum, 0, &(partial[kslice * M + gid * 2]));
+    }
+}
+
+// Sum the per-slice partials [ksplit * M] into dst[M]; applies the dst byte offset.
+kernel void kernel_gemv_splitk_reduce_f32(
+        global float * partial,
+        global float * dst,
+        ulong offsetd,
+        int   ne01,         // M
+        int   ksplit)
+{
+    uint r = get_global_id(0);
+    if (r >= (uint)ne01) return;
+    float acc = 0.0f;
+    for (uint s = 0; s < (uint)ksplit; ++s) {
+        acc += partial[s * (uint)ne01 + r];
+    }
+    dst = (global float*)((global char*)dst + offsetd);
+    dst[r] = acc;
+}
+
+
+// --- Dequant-once macros for the mc3 verify GEMV (Q4K_MC3_DEQUANT_ONCE) ---
+// The inline dequantizeBlockAccum_* macros recompute the dequantized weight
+// ((code & mask)>>shift)*scale - minv ONCE PER COLUMN (3x), and the flat
+// 32-FMA unroll spills ~430 B of temporaries. These macros split the work:
+// DEQUANT_Q4K_BLOCK computes the 16 weights/row of one 32-block ONCE into a
+// half2[] (row0 in .s0, row1 in .s1) — stored as half, the exact type the
+// inline expression yields (int*half-half), so no extra rounding. MAC_Q4K_BLOCK
+// then accumulates them against a column's broadcast activation in the SAME
+// per-accumulator order as the inline macro. Each weight value and each
+// accumulator's add-chain is bit-for-bit identical => byte-identical output,
+// while the dequant ALU drops 3x->1x and the live set shrinks. Requires the
+// Qualcomm vector sub_group_broadcast (float8); enabled opt-in on Adreno.
+#define DEQ_Q4K_HALF2(b0, b1, msk, sh, scale, minv) \
+    (half2)( ((b0 & msk) >> sh) * scale.s0 - minv.s0, \
+             ((b1 & msk) >> sh) * scale.s1 - minv.s1 )
+
+#define DEQUANT_Q4K_BLOCK(wq, bits, scale, minv) \
+    wq[0]  = DEQ_Q4K_HALF2(bits.s0, bits.s1, 0x000F, 0,  scale, minv); \
+    wq[1]  = DEQ_Q4K_HALF2(bits.s0, bits.s1, 0x00F0, 4,  scale, minv); \
+    wq[2]  = DEQ_Q4K_HALF2(bits.s0, bits.s1, 0x0F00, 8,  scale, minv); \
+    wq[3]  = DEQ_Q4K_HALF2(bits.s0, bits.s1, 0xF000, 12, scale, minv); \
+    wq[4]  = DEQ_Q4K_HALF2(bits.s2, bits.s3, 0x000F, 0,  scale, minv); \
+    wq[5]  = DEQ_Q4K_HALF2(bits.s2, bits.s3, 0x00F0, 4,  scale, minv); \
+    wq[6]  = DEQ_Q4K_HALF2(bits.s2, bits.s3, 0x0F00, 8,  scale, minv); \
+    wq[7]  = DEQ_Q4K_HALF2(bits.s2, bits.s3, 0xF000, 12, scale, minv); \
+    wq[8]  = DEQ_Q4K_HALF2(bits.s4, bits.s5, 0x000F, 0,  scale, minv); \
+    wq[9]  = DEQ_Q4K_HALF2(bits.s4, bits.s5, 0x00F0, 4,  scale, minv); \
+    wq[10] = DEQ_Q4K_HALF2(bits.s4, bits.s5, 0x0F00, 8,  scale, minv); \
+    wq[11] = DEQ_Q4K_HALF2(bits.s4, bits.s5, 0xF000, 12, scale, minv); \
+    wq[12] = DEQ_Q4K_HALF2(bits.s6, bits.s7, 0x000F, 0,  scale, minv); \
+    wq[13] = DEQ_Q4K_HALF2(bits.s6, bits.s7, 0x00F0, 4,  scale, minv); \
+    wq[14] = DEQ_Q4K_HALF2(bits.s6, bits.s7, 0x0F00, 8,  scale, minv); \
+    wq[15] = DEQ_Q4K_HALF2(bits.s6, bits.s7, 0xF000, 12, scale, minv);
+
+// ln0/ln1 = the two source lanes whose activation float8 this block consumes
+// (0,1 for the hi block, 2,3 for the lo block — matching the inline _hi/_lo).
+#define MAC_Q4K_BLOCK(ts, wq, y, ln0, ln1) { \
+    float8 sy = sub_group_broadcast(y, ln0); \
+    ts.s0 += wq[0].s0*sy.s0; ts.s0 += wq[1].s0*sy.s1; ts.s0 += wq[2].s0*sy.s2; ts.s0 += wq[3].s0*sy.s3; \
+    ts.s0 += wq[4].s0*sy.s4; ts.s0 += wq[5].s0*sy.s5; ts.s0 += wq[6].s0*sy.s6; ts.s0 += wq[7].s0*sy.s7; \
+    ts.s1 += wq[0].s1*sy.s0; ts.s1 += wq[1].s1*sy.s1; ts.s1 += wq[2].s1*sy.s2; ts.s1 += wq[3].s1*sy.s3; \
+    ts.s1 += wq[4].s1*sy.s4; ts.s1 += wq[5].s1*sy.s5; ts.s1 += wq[6].s1*sy.s6; ts.s1 += wq[7].s1*sy.s7; \
+    sy = sub_group_broadcast(y, ln1); \
+    ts.s0 += wq[8].s0*sy.s0;  ts.s0 += wq[9].s0*sy.s1;  ts.s0 += wq[10].s0*sy.s2; ts.s0 += wq[11].s0*sy.s3; \
+    ts.s0 += wq[12].s0*sy.s4; ts.s0 += wq[13].s0*sy.s5; ts.s0 += wq[14].s0*sy.s6; ts.s0 += wq[15].s0*sy.s7; \
+    ts.s1 += wq[8].s1*sy.s0;  ts.s1 += wq[9].s1*sy.s1;  ts.s1 += wq[10].s1*sy.s2; ts.s1 += wq[11].s1*sy.s3; \
+    ts.s1 += wq[12].s1*sy.s4; ts.s1 += wq[13].s1*sy.s5; ts.s1 += wq[14].s1*sy.s6; ts.s1 += wq[15].s1*sy.s7; \
+}
+
+// Multi-column (N=3) variant of the q4_K decode GEMV, for the speculative /
+// MTP verify batch (ne1=3 = 2 drafts + 1 bonus). Stays on the efficient GEMV
+// path (subgroup-broadcast activation, NSUBGROUPS K-split) instead of the
+// transposed-GEMM dead-zone path. Each K-block's weights (regA_hi/regA_lo) are
+// loaded ONCE and reused across all 3 activation columns — same weight traffic
+// as one decode, ~3x the (cheap) dequant ALU. Per-column accumulation is
+// independent and identical to 3 standalone GEMVs => byte-identical, so it does
+// NOT perturb the lm_head logits / spec accept rate.
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_gemv_noshuffle_q4_k_f32_mc3(
+        read_only  image1d_buffer_t src0_q,
+        global half2  * src0_d,
+        global half2  * src0_m,
+        global uchar  * src0_s,
+        read_only  image1d_buffer_t src1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        uchar mask_d6,
+        uchar mask_d4,
+        uchar mask_hi2)
+{
+    uint groupId = get_local_id(1);
+    uint gid     = get_global_id(0);
+    ushort slid  = get_sub_group_local_id();
+
+    uint K = ne00;
+    uint M = ne01;
+
+    uint LINE_STRIDE_A  = M / 2;
+    uint BLOCK_STRIDE_A = NSUBGROUPS * M;
+    uint COL_STRIDE     = K / 4;   // float4 pixels per activation column
+
+    private uint4  regA_hi, regA_lo;
+    private half2  regS, regM;
+    private float8 regB;
+
+    private float2 ts0 = (float2)(0.0f);
+    private float2 ts1 = (float2)(0.0f);
+    private float2 ts2 = (float2)(0.0f);
+
+#ifdef Q4K_MC3_DEQUANT_LDS
+    // One 16-half2 block buffer per WI (reused hi->lo): forces the dequantized
+    // weights into LDS instead of private arrays (which spill to slow global on
+    // Adreno). 64*NSUBGROUPS WIs * 16 half2 = 16 KB; each WI owns its own slot
+    // range (flat*16) -> no cross-lane sharing, no barrier needed.
+    local half2 wstage[SUBGROUP_SIZE * NSUBGROUPS * 16];
+    local half2 * ws = wstage + (groupId * SUBGROUP_SIZE + slid) * 16;
+#endif
+
    for (uint k = groupId; k < (K / 32); k += NSUBGROUPS) {
        uint sb = k / 8;
        uint j  = k % 8;
@@ -259,66 +715,109 @@ kernel void kernel_gemv_noshuffle_q4_k_f32(
        regS = convert_half2(convert_float2(d)  * convert_float2((uchar2)(sv0, sv1)));
        regM = convert_half2(convert_float2(dm) * convert_float2((uchar2)(mn0, mn1)));

-        if (slid < 4) {
-            regB.s0123 = read_imagef(src1, (slid * 2 + k * 8));
-            regB.s4567 = read_imagef(src1, (1 + slid * 2 + k * 8));
-        }
+        // weights loaded ONCE, reused across the 3 columns
+        regA_hi.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
+        regA_hi.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
+        regA_hi.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
+        regA_hi.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
+        regA_lo.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
+        regA_lo.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
+        regA_lo.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
+        regA_lo.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;

-        // load half weights for two blocks in consecutive rows
-        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
-        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
-        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
-        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
-#ifdef VECTOR_SUB_GROUP_BROADCAST
-        dequantizeBlockAccum_ns_sgbroadcast_8_hi(totalSum, as_ushort8(regA), regS, regM, regB);
+#ifdef Q4K_MC3_DEQUANT_ONCE
+        // Dequant the 32 weights/row (16 hi + 16 lo) ONCE into half2[] (byte-
+        // identical to the inline intermediate), then MAC against each column's
+        // activation. Drops the dequant ALU 3x->1x and the macro-temp spill.
+        half2 wq_hi[16], wq_lo[16];
+        DEQUANT_Q4K_BLOCK(wq_hi, as_ushort8(regA_hi), regS, regM);
+        DEQUANT_Q4K_BLOCK(wq_lo, as_ushort8(regA_lo), regS, regM);
+        { if (slid < 4) { regB.s0123 = read_imagef(src1, 0*COL_STRIDE     + slid*2 + k*8);
+                          regB.s4567 = read_imagef(src1, 0*COL_STRIDE + 1 + slid*2 + k*8); }
+          MAC_Q4K_BLOCK(ts0, wq_hi, regB, 0, 1); MAC_Q4K_BLOCK(ts0, wq_lo, regB, 2, 3); }
+        { if (slid < 4) { regB.s0123 = read_imagef(src1, 1*COL_STRIDE     + slid*2 + k*8);
+                          regB.s4567 = read_imagef(src1, 1*COL_STRIDE + 1 + slid*2 + k*8); }
+          MAC_Q4K_BLOCK(ts1, wq_hi, regB, 0, 1); MAC_Q4K_BLOCK(ts1, wq_lo, regB, 2, 3); }
+        { if (slid < 4) { regB.s0123 = read_imagef(src1, 2*COL_STRIDE     + slid*2 + k*8);
+                          regB.s4567 = read_imagef(src1, 2*COL_STRIDE + 1 + slid*2 + k*8); }
+          MAC_Q4K_BLOCK(ts2, wq_hi, regB, 0, 1); MAC_Q4K_BLOCK(ts2, wq_lo, regB, 2, 3); }
+#elif defined(Q4K_MC3_DEQUANT_LDS)
+        // LDS-staged dequant: dequant a 32-block ONCE into the per-WI LDS slot
+        // (hi pass then lo pass, overwriting), MAC each column from LDS. ts*
+        // receive hi-then-lo in the same order as DEQUANT_ONCE -> byte-identical.
+        // Activations reloaded per pass (cheap, imaged); only one regB + 0 weight
+        // regs live -> the weight working set lives in LDS, not spilled private.
+        DEQUANT_Q4K_BLOCK(ws, as_ushort8(regA_hi), regS, regM);
+        { if (slid < 4) { regB.s0123 = read_imagef(src1, 0*COL_STRIDE     + slid*2 + k*8);
+                          regB.s4567 = read_imagef(src1, 0*COL_STRIDE + 1 + slid*2 + k*8); }
+          MAC_Q4K_BLOCK(ts0, ws, regB, 0, 1); }
+        { if (slid < 4) { regB.s0123 = read_imagef(src1, 1*COL_STRIDE     + slid*2 + k*8);
+                          regB.s4567 = read_imagef(src1, 1*COL_STRIDE + 1 + slid*2 + k*8); }
+          MAC_Q4K_BLOCK(ts1, ws, regB, 0, 1); }
+        { if (slid < 4) { regB.s0123 = read_imagef(src1, 2*COL_STRIDE     + slid*2 + k*8);
+                          regB.s4567 = read_imagef(src1, 2*COL_STRIDE + 1 + slid*2 + k*8); }
+          MAC_Q4K_BLOCK(ts2, ws, regB, 0, 1); }
+        DEQUANT_Q4K_BLOCK(ws, as_ushort8(regA_lo), regS, regM);
+        { if (slid < 4) { regB.s0123 = read_imagef(src1, 0*COL_STRIDE     + slid*2 + k*8);
+                          regB.s4567 = read_imagef(src1, 0*COL_STRIDE + 1 + slid*2 + k*8); }
+          MAC_Q4K_BLOCK(ts0, ws, regB, 2, 3); }
+        { if (slid < 4) { regB.s0123 = read_imagef(src1, 1*COL_STRIDE     + slid*2 + k*8);
+                          regB.s4567 = read_imagef(src1, 1*COL_STRIDE + 1 + slid*2 + k*8); }
+          MAC_Q4K_BLOCK(ts1, ws, regB, 2, 3); }
+        { if (slid < 4) { regB.s0123 = read_imagef(src1, 2*COL_STRIDE     + slid*2 + k*8);
+                          regB.s4567 = read_imagef(src1, 2*COL_STRIDE + 1 + slid*2 + k*8); }
+          MAC_Q4K_BLOCK(ts2, ws, regB, 2, 3); }
 #else
-        dequantizeBlockAccum_ns_sgbroadcast_1_hi(totalSum, as_ushort8(regA), regS, regM, regB);
-#endif // VECTOR_SUB_GROUP_BROADCAST
-
-        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
-        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
-        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
-        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
+        // Per-column: load only this column's activation (single regB live at a
+        // time -> 1/3 the activation register pressure vs holding all 3) then
+        // dequant against the shared weights. Cuts the private-mem spill.
 #ifdef VECTOR_SUB_GROUP_BROADCAST
-        dequantizeBlockAccum_ns_sgbroadcast_8_lo(totalSum, as_ushort8(regA), regS, regM, regB);
+        { if (slid < 4) { regB.s0123 = read_imagef(src1, 0*COL_STRIDE     + slid*2 + k*8);
+                          regB.s4567 = read_imagef(src1, 0*COL_STRIDE + 1 + slid*2 + k*8); }
+          dequantizeBlockAccum_ns_sgbroadcast_8_hi(ts0, as_ushort8(regA_hi), regS, regM, regB);
+          dequantizeBlockAccum_ns_sgbroadcast_8_lo(ts0, as_ushort8(regA_lo), regS, regM, regB); }
+        { if (slid < 4) { regB.s0123 = read_imagef(src1, 1*COL_STRIDE     + slid*2 + k*8);
+                          regB.s4567 = read_imagef(src1, 1*COL_STRIDE + 1 + slid*2 + k*8); }
+          dequantizeBlockAccum_ns_sgbroadcast_8_hi(ts1, as_ushort8(regA_hi), regS, regM, regB);
+          dequantizeBlockAccum_ns_sgbroadcast_8_lo(ts1, as_ushort8(regA_lo), regS, regM, regB); }
+        { if (slid < 4) { regB.s0123 = read_imagef(src1, 2*COL_STRIDE     + slid*2 + k*8);
+                          regB.s4567 = read_imagef(src1, 2*COL_STRIDE + 1 + slid*2 + k*8); }
+          dequantizeBlockAccum_ns_sgbroadcast_8_hi(ts2, as_ushort8(regA_hi), regS, regM, regB);
+          dequantizeBlockAccum_ns_sgbroadcast_8_lo(ts2, as_ushort8(regA_lo), regS, regM, regB); }
 #else
-        dequantizeBlockAccum_ns_sgbroadcast_1_lo(totalSum, as_ushort8(regA), regS, regM, regB);
-#endif // VECTOR_SUB_GROUP_BROADCAST
+        { if (slid < 4) { regB.s0123 = read_imagef(src1, 0*COL_STRIDE     + slid*2 + k*8);
+                          regB.s4567 = read_imagef(src1, 0*COL_STRIDE + 1 + slid*2 + k*8); }
+          dequantizeBlockAccum_ns_sgbroadcast_1_hi(ts0, as_ushort8(regA_hi), regS, regM, regB);
+          dequantizeBlockAccum_ns_sgbroadcast_1_lo(ts0, as_ushort8(regA_lo), regS, regM, regB); }
+        { if (slid < 4) { regB.s0123 = read_imagef(src1, 1*COL_STRIDE     + slid*2 + k*8);
+                          regB.s4567 = read_imagef(src1, 1*COL_STRIDE + 1 + slid*2 + k*8); }
+          dequantizeBlockAccum_ns_sgbroadcast_1_hi(ts1, as_ushort8(regA_hi), regS, regM, regB);
+          dequantizeBlockAccum_ns_sgbroadcast_1_lo(ts1, as_ushort8(regA_lo), regS, regM, regB); }
+        { if (slid < 4) { regB.s0123 = read_imagef(src1, 2*COL_STRIDE     + slid*2 + k*8);
+                          regB.s4567 = read_imagef(src1, 2*COL_STRIDE + 1 + slid*2 + k*8); }
+          dequantizeBlockAccum_ns_sgbroadcast_1_hi(ts2, as_ushort8(regA_hi), regS, regM, regB);
+          dequantizeBlockAccum_ns_sgbroadcast_1_lo(ts2, as_ushort8(regA_lo), regS, regM, regB); }
+#endif
+#endif // Q4K_MC3_DEQUANT_ONCE
    }

-    // reduction in local memory, assumes #wave=4
-    local float2 reduceLM[SUBGROUP_SIZE * 3];
-    if (groupId == 1) {
-        reduceLM[SUBGROUP_SIZE * 0 + slid] = totalSum;
-    }
-    if (groupId == 2) {
-        reduceLM[SUBGROUP_SIZE * 1 + slid] = totalSum;
-    }
-    if (groupId == 3) {
-        reduceLM[SUBGROUP_SIZE * 2 + slid] = totalSum;
-    }
+    // cross-subgroup reduce: pack the 3 columns' float2 into a float8 (6 used).
+    local float8 reduceLM[SUBGROUP_SIZE * 3];
+    float8 acc = (float8)(ts0.s0, ts0.s1, ts1.s0, ts1.s1, ts2.s0, ts2.s1, 0.0f, 0.0f);
+    if (groupId == 1) { reduceLM[SUBGROUP_SIZE * 0 + slid] = acc; }
+    if (groupId == 2) { reduceLM[SUBGROUP_SIZE * 1 + slid] = acc; }
+    if (groupId == 3) { reduceLM[SUBGROUP_SIZE * 2 + slid] = acc; }

    barrier(CLK_LOCAL_MEM_FENCE);

    if (groupId == 0) {
-        totalSum += reduceLM[SUBGROUP_SIZE * 0 + slid];
-    }
-    if (groupId == 0) {
-        totalSum += reduceLM[SUBGROUP_SIZE * 1 + slid];
-    }
-    if (groupId == 0) {
-        totalSum += reduceLM[SUBGROUP_SIZE * 2 + slid];
-    }
-
-    // 2 outputs per fiber in wave 0
-    if (groupId == 0) {
+        acc += reduceLM[SUBGROUP_SIZE * 0 + slid];
+        acc += reduceLM[SUBGROUP_SIZE * 1 + slid];
+        acc += reduceLM[SUBGROUP_SIZE * 2 + slid];
        dst = (global float*)((global char*)dst + offsetd);
-        // Guard the two output rows. The x-grid is padded to CEIL_DIV(ne01/2,64)*64,
-        // so when ne01 is not a multiple of 128 the tail row-pairs run past row ne01
-        // and would overrun dst into the adjacent tensor. No-op / byte-identical when
-        // ne01 % 128 == 0 (M/2 already a multiple of 64 -> no padding).
-        if (gid * 2 + 0 < M) dst[gid * 2 + 0] = totalSum.s0;
-        if (gid * 2 + 1 < M) dst[gid * 2 + 1] = totalSum.s1;
+        // dst is column-major [M rows x 3 cols]: (row, col) at col*M + row
+        vstore2((float2)(acc.s0, acc.s1), 0, &(dst[0 * M + gid * 2]));
+        vstore2((float2)(acc.s2, acc.s3), 0, &(dst[1 * M + gid * 2]));
+        vstore2((float2)(acc.s4, acc.s5), 0, &(dst[2 * M + gid * 2]));
    }
-
 }
@@ -0,0 +1,349 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64 __attribute__((qcom_reqd_sub_group_size("half")))
+#endif
+
+#define QK_K  256
+#define NSUBGROUPS 4
+#define SUBGROUP_SIZE 64
+
+// scales are transposed: consecutive codes of a row are `stride` apart
+inline void get_scale_min_k4(
+    int j,
+    global const uchar * q,
+    uint stride,
+    uchar * d,
+    uchar * m,
+    uchar mask_d6,
+    uchar mask_d4,
+    uchar mask_hi2
+) {
+    if (j < 4) {
+        *d = q[j*stride]     & mask_d6;
+        *m = q[(j+4)*stride] & mask_d6;
+    } else {
+        *d = (q[(j+4)*stride] & mask_d4) | ((q[(j-4)*stride] & mask_hi2) >> 2);
+        *m = ((q[(j+4)*stride] >> 4) & mask_d4) | ((q[j*stride] & mask_hi2) >> 2);
+    }
+}
+
+#define dequantizeBlockAccum_ns_sgbroadcast_1_hi(total_sums, bits4, scale, minv, y) \
+    float shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 0); \
+    total_sums.s0 += ((bits4.s0 & 0x000F) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((bits4.s1 & 0x000F) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 0); \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 0); \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 0); \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 0); \
+    total_sums.s0 += ((bits4.s2 & 0x000F) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((bits4.s3 & 0x000F) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 0); \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 0); \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 0); \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 1); \
+    total_sums.s0 += ((bits4.s4 & 0x000F) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((bits4.s5 & 0x000F) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 1); \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 1); \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 1); \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 1); \
+    total_sums.s0 += ((bits4.s6 & 0x000F) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((bits4.s7 & 0x000F) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 1); \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 1); \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 1); \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_1_lo(total_sums, bits4, scale, minv, y) \
+    shared_y = sub_group_broadcast(y.s0, 2); \
+    total_sums.s0 += ((bits4.s0 & 0x000F) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((bits4.s1 & 0x000F) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 2); \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 2); \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 2); \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 2); \
+    total_sums.s0 += ((bits4.s2 & 0x000F) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((bits4.s3 & 0x000F) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 2); \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 2); \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 2); \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 3); \
+    total_sums.s0 += ((bits4.s4 & 0x000F) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((bits4.s5 & 0x000F) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 3); \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 3); \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 3); \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 3); \
+    total_sums.s0 += ((bits4.s6 & 0x000F) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += ((bits4.s7 & 0x000F) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 3); \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 3); \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8) * scale.s1 - minv.s1) * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 3); \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_8_hi(total_sums, bits4, scale, minv, y) \
+    float8 shared_y; \
+    shared_y = sub_group_broadcast(y, 0); \
+    total_sums.s0 += ((bits4.s0 & 0x000F)         * scale.s0 - minv.s0) * shared_y.s0; \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4)  * scale.s0 - minv.s0) * shared_y.s1; \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8)  * scale.s0 - minv.s0) * shared_y.s2; \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y.s3; \
+    total_sums.s0 += ((bits4.s2 & 0x000F)         * scale.s0 - minv.s0) * shared_y.s4; \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4)  * scale.s0 - minv.s0) * shared_y.s5; \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8)  * scale.s0 - minv.s0) * shared_y.s6; \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y.s7; \
+    total_sums.s1 += ((bits4.s1 & 0x000F)         * scale.s1 - minv.s1) * shared_y.s0; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4)  * scale.s1 - minv.s1) * shared_y.s1; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8)  * scale.s1 - minv.s1) * shared_y.s2; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y.s3; \
+    total_sums.s1 += ((bits4.s3 & 0x000F)         * scale.s1 - minv.s1) * shared_y.s4; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4)  * scale.s1 - minv.s1) * shared_y.s5; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8)  * scale.s1 - minv.s1) * shared_y.s6; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 1); \
+    total_sums.s0 += ((bits4.s4 & 0x000F)         * scale.s0 - minv.s0) * shared_y.s0; \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4)  * scale.s0 - minv.s0) * shared_y.s1; \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8)  * scale.s0 - minv.s0) * shared_y.s2; \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y.s3; \
+    total_sums.s0 += ((bits4.s6 & 0x000F)         * scale.s0 - minv.s0) * shared_y.s4; \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4)  * scale.s0 - minv.s0) * shared_y.s5; \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8)  * scale.s0 - minv.s0) * shared_y.s6; \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y.s7; \
+    total_sums.s1 += ((bits4.s5 & 0x000F)         * scale.s1 - minv.s1) * shared_y.s0; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4)  * scale.s1 - minv.s1) * shared_y.s1; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8)  * scale.s1 - minv.s1) * shared_y.s2; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y.s3; \
+    total_sums.s1 += ((bits4.s7 & 0x000F)         * scale.s1 - minv.s1) * shared_y.s4; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4)  * scale.s1 - minv.s1) * shared_y.s5; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8)  * scale.s1 - minv.s1) * shared_y.s6; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y.s7; \
+
+
+#define dequantizeBlockAccum_ns_sgbroadcast_8_lo(total_sums, bits4, scale, minv, y) \
+    shared_y = sub_group_broadcast(y, 2); \
+    total_sums.s0 += ((bits4.s0 & 0x000F)         * scale.s0 - minv.s0) * shared_y.s0; \
+    total_sums.s0 += (((bits4.s0 & 0x00F0) >> 4)  * scale.s0 - minv.s0) * shared_y.s1; \
+    total_sums.s0 += (((bits4.s0 & 0x0F00) >> 8)  * scale.s0 - minv.s0) * shared_y.s2; \
+    total_sums.s0 += (((bits4.s0 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y.s3; \
+    total_sums.s0 += ((bits4.s2 & 0x000F)         * scale.s0 - minv.s0) * shared_y.s4; \
+    total_sums.s0 += (((bits4.s2 & 0x00F0) >> 4)  * scale.s0 - minv.s0) * shared_y.s5; \
+    total_sums.s0 += (((bits4.s2 & 0x0F00) >> 8)  * scale.s0 - minv.s0) * shared_y.s6; \
+    total_sums.s0 += (((bits4.s2 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y.s7; \
+    total_sums.s1 += ((bits4.s1 & 0x000F)         * scale.s1 - minv.s1) * shared_y.s0; \
+    total_sums.s1 += (((bits4.s1 & 0x00F0) >> 4)  * scale.s1 - minv.s1) * shared_y.s1; \
+    total_sums.s1 += (((bits4.s1 & 0x0F00) >> 8)  * scale.s1 - minv.s1) * shared_y.s2; \
+    total_sums.s1 += (((bits4.s1 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y.s3; \
+    total_sums.s1 += ((bits4.s3 & 0x000F)         * scale.s1 - minv.s1) * shared_y.s4; \
+    total_sums.s1 += (((bits4.s3 & 0x00F0) >> 4)  * scale.s1 - minv.s1) * shared_y.s5; \
+    total_sums.s1 += (((bits4.s3 & 0x0F00) >> 8)  * scale.s1 - minv.s1) * shared_y.s6; \
+    total_sums.s1 += (((bits4.s3 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 3); \
+    total_sums.s0 += ((bits4.s4 & 0x000F)         * scale.s0 - minv.s0) * shared_y.s0; \
+    total_sums.s0 += (((bits4.s4 & 0x00F0) >> 4)  * scale.s0 - minv.s0) * shared_y.s1; \
+    total_sums.s0 += (((bits4.s4 & 0x0F00) >> 8)  * scale.s0 - minv.s0) * shared_y.s2; \
+    total_sums.s0 += (((bits4.s4 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y.s3; \
+    total_sums.s0 += ((bits4.s6 & 0x000F)         * scale.s0 - minv.s0) * shared_y.s4; \
+    total_sums.s0 += (((bits4.s6 & 0x00F0) >> 4)  * scale.s0 - minv.s0) * shared_y.s5; \
+    total_sums.s0 += (((bits4.s6 & 0x0F00) >> 8)  * scale.s0 - minv.s0) * shared_y.s6; \
+    total_sums.s0 += (((bits4.s6 & 0xF000) >> 12) * scale.s0 - minv.s0) * shared_y.s7; \
+    total_sums.s1 += ((bits4.s5 & 0x000F)         * scale.s1 - minv.s1) * shared_y.s0; \
+    total_sums.s1 += (((bits4.s5 & 0x00F0) >> 4)  * scale.s1 - minv.s1) * shared_y.s1; \
+    total_sums.s1 += (((bits4.s5 & 0x0F00) >> 8)  * scale.s1 - minv.s1) * shared_y.s2; \
+    total_sums.s1 += (((bits4.s5 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y.s3; \
+    total_sums.s1 += ((bits4.s7 & 0x000F)         * scale.s1 - minv.s1) * shared_y.s4; \
+    total_sums.s1 += (((bits4.s7 & 0x00F0) >> 4)  * scale.s1 - minv.s1) * shared_y.s5; \
+    total_sums.s1 += (((bits4.s7 & 0x0F00) >> 8)  * scale.s1 - minv.s1) * shared_y.s6; \
+    total_sums.s1 += (((bits4.s7 & 0xF000) >> 12) * scale.s1 - minv.s1) * shared_y.s7; \
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_gemv_noshuffle_q4_k_f32_o4(
+        read_only  image1d_buffer_t src0_q,
+        global half2  * src0_d,
+        global half2  * src0_m,
+        global uchar  * src0_s,
+        read_only  image1d_buffer_t src1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        uchar mask_d6,
+        uchar mask_d4,
+        uchar mask_hi2)
+{
+    uint groupId = get_local_id(1);
+    uint gid     = get_global_id(0);        // 4-output quad index
+    ushort slid  = get_sub_group_local_id();
+
+    // Two consecutive pair-indices (each the same access pattern the 2-output
+    // kernel uses); together they cover 4 consecutive output rows.
+    uint gid_a = gid * 2;
+    uint gid_b = gid * 2 + 1;
+
+    uint K = ne00;
+    uint M = ne01;
+
+    uint LINE_STRIDE_A  = M / 2;
+    uint BLOCK_STRIDE_A = NSUBGROUPS * M;
+
+    private uint4  regA;
+    private half2  regS_a, regS_b;
+    private half2  regM_a, regM_b;
+    private float8 regB;
+
+    private float2 totalSum_a = (float2)(0.0f);
+    private float2 totalSum_b = (float2)(0.0f);
+
+    for (uint k = groupId; k < (K / 32); k += NSUBGROUPS) {
+        uint sb = k / 8;
+        uint j  = k % 8;
+
+        // pair a scales/mins
+        half2 d_a  = src0_d[gid_a + sb * LINE_STRIDE_A];
+        half2 dm_a = src0_m[gid_a + sb * LINE_STRIDE_A];
+        global const uchar * sc0a = src0_s + sb * 12 * M + 2 * gid_a;
+        global const uchar * sc1a = sc0a + 1;
+        uchar sv0a, mn0a, sv1a, mn1a;
+        get_scale_min_k4(j, sc0a, M, &sv0a, &mn0a, mask_d6, mask_d4, mask_hi2);
+        get_scale_min_k4(j, sc1a, M, &sv1a, &mn1a, mask_d6, mask_d4, mask_hi2);
+        regS_a = convert_half2(convert_float2(d_a)  * convert_float2((uchar2)(sv0a, sv1a)));
+        regM_a = convert_half2(convert_float2(dm_a) * convert_float2((uchar2)(mn0a, mn1a)));
+
+        // pair b scales/mins
+        half2 d_b  = src0_d[gid_b + sb * LINE_STRIDE_A];
+        half2 dm_b = src0_m[gid_b + sb * LINE_STRIDE_A];
+        global const uchar * sc0b = src0_s + sb * 12 * M + 2 * gid_b;
+        global const uchar * sc1b = sc0b + 1;
+        uchar sv0b, mn0b, sv1b, mn1b;
+        get_scale_min_k4(j, sc0b, M, &sv0b, &mn0b, mask_d6, mask_d4, mask_hi2);
+        get_scale_min_k4(j, sc1b, M, &sv1b, &mn1b, mask_d6, mask_d4, mask_hi2);
+        regS_b = convert_half2(convert_float2(d_b)  * convert_float2((uchar2)(sv0b, sv1b)));
+        regM_b = convert_half2(convert_float2(dm_b) * convert_float2((uchar2)(mn0b, mn1b)));
+
+        // activation: load once, reuse for both pairs
+        if (slid < 4) {
+            regB.s0123 = read_imagef(src1, (slid * 2 + k * 8));
+            regB.s4567 = read_imagef(src1, (1 + slid * 2 + k * 8));
+        }
+
+        // pair a (own block so _lo sees the shared_y declared by _hi)
+        {
+            regA.s0 = read_imageui(src0_q, (gid_a + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
+            regA.s1 = read_imageui(src0_q, (gid_a + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
+            regA.s2 = read_imageui(src0_q, (gid_a + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
+            regA.s3 = read_imageui(src0_q, (gid_a + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAST
+            dequantizeBlockAccum_ns_sgbroadcast_8_hi(totalSum_a, as_ushort8(regA), regS_a, regM_a, regB);
+#else
+            dequantizeBlockAccum_ns_sgbroadcast_1_hi(totalSum_a, as_ushort8(regA), regS_a, regM_a, regB);
+#endif
+            regA.s0 = read_imageui(src0_q, (gid_a + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
+            regA.s1 = read_imageui(src0_q, (gid_a + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
+            regA.s2 = read_imageui(src0_q, (gid_a + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
+            regA.s3 = read_imageui(src0_q, (gid_a + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAST
+            dequantizeBlockAccum_ns_sgbroadcast_8_lo(totalSum_a, as_ushort8(regA), regS_a, regM_a, regB);
+#else
+            dequantizeBlockAccum_ns_sgbroadcast_1_lo(totalSum_a, as_ushort8(regA), regS_a, regM_a, regB);
+#endif
+        }
+
+        // pair b
+        {
+            regA.s0 = read_imageui(src0_q, (gid_b + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
+            regA.s1 = read_imageui(src0_q, (gid_b + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
+            regA.s2 = read_imageui(src0_q, (gid_b + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
+            regA.s3 = read_imageui(src0_q, (gid_b + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAST
+            dequantizeBlockAccum_ns_sgbroadcast_8_hi(totalSum_b, as_ushort8(regA), regS_b, regM_b, regB);
+#else
+            dequantizeBlockAccum_ns_sgbroadcast_1_hi(totalSum_b, as_ushort8(regA), regS_b, regM_b, regB);
+#endif
+            regA.s0 = read_imageui(src0_q, (gid_b + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
+            regA.s1 = read_imageui(src0_q, (gid_b + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
+            regA.s2 = read_imageui(src0_q, (gid_b + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
+            regA.s3 = read_imageui(src0_q, (gid_b + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
+#ifdef VECTOR_SUB_GROUP_BROADCAST
+            dequantizeBlockAccum_ns_sgbroadcast_8_lo(totalSum_b, as_ushort8(regA), regS_b, regM_b, regB);
+#else
+            dequantizeBlockAccum_ns_sgbroadcast_1_lo(totalSum_b, as_ushort8(regA), regS_b, regM_b, regB);
+#endif
+        }
+    }
+
+    // reduce 4 outputs (a.s0, a.s1, b.s0, b.s1) across the 4 subgroups
+    local float4 reduceLM[SUBGROUP_SIZE * 3];
+    float4 acc = (float4)(totalSum_a.s0, totalSum_a.s1, totalSum_b.s0, totalSum_b.s1);
+    if (groupId == 1) { reduceLM[SUBGROUP_SIZE * 0 + slid] = acc; }
+    if (groupId == 2) { reduceLM[SUBGROUP_SIZE * 1 + slid] = acc; }
+    if (groupId == 3) { reduceLM[SUBGROUP_SIZE * 2 + slid] = acc; }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (groupId == 0) {
+        acc += reduceLM[SUBGROUP_SIZE * 0 + slid];
+        acc += reduceLM[SUBGROUP_SIZE * 1 + slid];
+        acc += reduceLM[SUBGROUP_SIZE * 2 + slid];
+        dst = (global float*)((global char*)dst + offsetd);
+        // The dispatch rounds ne01/4 up to the subgroup width, so the tail
+        // quads past the last row must not store (they wrote 128 rows past
+        // dst on every ne01 % 256 == 128 vocab, e.g. 151936).
+        if (gid * 4 + 3 < (uint)ne01) {
+            vstore4(acc, 0, &(dst[gid * 4]));
+        }
+    }
+}
@@ -0,0 +1,118 @@
+// Tiled-wide q4_K GEMV for the long-vocab lm_head/embed (decode path).
+//
+// Pairs with kernel_convert_block_q4_k_tiled_ns (cvt.cl): the weights are laid
+// out CANONICALLY (4-bit code in element order e in [0,256)) and TILED by 64
+// output rows so the 64-thread lane group coalesces every weight load. Both the
+// pack (convert) and the unpack (here) are owned by us -> correct by
+// construction vs the reference ggml q4_K dequant. Same structure as the q6_K
+// tiled GEMV; the only differences are the 4-bit dequant and the q4_K
+// scale/min decode (get_scale_min_k4 from the packed 12-byte block).
+//
+// One work-item produces one output row. WG = {64 lanes, 4 subgroups}: the 64
+// lanes cover the 64 rows of one tile (coalesced uint4 reads), the 4 subgroups
+// split the K-blocks and reduce through __local at the end. Weights read from
+// __global (lm_head is streamed once per token; texture cache caps it below the
+// coalesced-global rate).
+
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#endif
+
+#define QK_K       256
+#define NSUBGROUPS 4
+#define TILE_ROWS  64
+
+// Decode one q4_K sub-block scale + min from the packed 12-byte block.
+// Identical to the o4 kernel's helper (masks hard-coded: d6=0x3F, d4=0x0F, hi2=0xC0).
+inline void q4k_scale_min(int j, __global const uchar * q, uchar * d, uchar * m) {
+    if (j < 4) {
+        *d = q[j]   & 0x3F;
+        *m = q[j+4] & 0x3F;
+    } else {
+        *d = (q[j+4] & 0x0F) | ((q[j-4] & 0xC0) >> 2);
+        *m = ((q[j+4] >> 4) & 0x0F) | ((q[j] & 0xC0) >> 2);
+    }
+}
+
+#if defined(ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_gemv_noshuffle_q4_k_f32_tiled(
+    __global uint4 * src0_q,    // tiled: 8 uint4 granules / superblock (4-bit codes)
+    __global half  * src0_d,    // tiled: 1 half / superblock
+    __global half  * src0_dm,   // tiled: 1 half / superblock
+    __global uchar * src0_s,    // tiled: 12 bytes / superblock (packed scales)
+    read_only image1d_buffer_t src1,   // activation (RGBA f32)
+    global float * dst,
+    ulong offsetd,
+    int ne00,
+    int ne01
+) {
+    int grp = get_local_id(1);          // subgroup index 0..3 (splits K)
+    int row = get_global_id(0);         // output row along ne01
+    int rt  = row / TILE_ROWS;
+    int rit = row % TILE_ROWS;
+
+    int nb = ne00 / QK_K;               // superblocks per row
+
+    float acc = 0.0f;
+
+    for (int sb = grp; sb < nb; sb += NSUBGROUPS) {
+        int tile_blk = rt * nb + sb;    // ne02 == 1 for lm_head/embed
+
+        float dval  = (float)src0_d [tile_blk * TILE_ROWS + rit];
+        float dmval = (float)src0_dm[tile_blk * TILE_ROWS + rit];
+
+        // decode the 8 sub-block (scale, min) pairs
+        __global uchar * sc = src0_s + (tile_blk * TILE_ROWS + rit) * 12;
+        float scale[8], minv[8];
+        #pragma unroll
+        for (int is = 0; is < 8; ++is) {
+            uchar sd, sm;
+            q4k_scale_min(is, sc, &sd, &sm);
+            scale[is] = dval  * (float)sd;
+            minv[is]  = dmval * (float)sm;
+        }
+
+        // 32 uints of 4-bit codes (8 codes/uint), e-order
+        uint q[32];
+        #pragma unroll
+        for (int g = 0; g < 8; ++g) {
+            uint4 v = src0_q[(tile_blk * 8 + g) * TILE_ROWS + rit];
+            q[g*4+0] = v.x; q[g*4+1] = v.y; q[g*4+2] = v.z; q[g*4+3] = v.w;
+        }
+
+        // dequant 256 codes in canonical e-order, MAC with activation.
+        int act_base = sb * 64;         // activation float4 pixel base (256/4)
+        #pragma unroll
+        for (int e4 = 0; e4 < 64; ++e4) {
+            float4 a = read_imagef(src1, act_base + e4);
+            #pragma unroll
+            for (int t = 0; t < 4; ++t) {
+                int  e    = e4 * 4 + t;
+                uint code = (q[e >> 3] >> ((e & 7) * 4)) & 0xF;
+                int  is   = e >> 5;     // sub-block index = e/32
+                float av  = (t == 0) ? a.x : (t == 1) ? a.y : (t == 2) ? a.z : a.w;
+                acc += ((float)code * scale[is] - minv[is]) * av;
+            }
+        }
+    }
+
+    // reduce across the NSUBGROUPS subgroups (same rit, different K-subset)
+    local float reduce_lm[NSUBGROUPS * TILE_ROWS];
+    reduce_lm[grp * TILE_ROWS + rit] = acc;
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (grp == 0) {
+        float total = reduce_lm[0 * TILE_ROWS + rit]
+                    + reduce_lm[1 * TILE_ROWS + rit]
+                    + reduce_lm[2 * TILE_ROWS + rit]
+                    + reduce_lm[3 * TILE_ROWS + rit];
+        dst = (global float*)((global char*)dst + offsetd);
+        dst[row] = total;
+    }
+}
@@ -329,3 +329,125 @@ kernel void kernel_gemv_noshuffle_q5_k_f32(
        if (gid * 2 + 1 < M) dst[gid * 2 + 1] = totalSum.s1;
    }
 }
+
+// Multi-column (N in [2..4]) variant of the q5_K decode GEMV (spec/MTP verify) =
+// q4_K mc3 + the high-bit qh plane (regH). n_cols = 2..4 (drafted + bonus); routes
+// the small-batch verify OFF the gemm_noshuffle_q5_k dead-zone. n_cols==3 is byte-
+// identical to the original mc3 (col3 disabled, float8 slots 6/7 stay zero).
+#ifdef VECTOR_SUB_GROUP_BROADCAST
+#define MC_DQ5_HI dequantizeBlockAccum_ns_sgbroadcast_8_hi
+#define MC_DQ5_LO dequantizeBlockAccum_ns_sgbroadcast_8_lo
+#else
+#define MC_DQ5_HI dequantizeBlockAccum_ns_sgbroadcast_1_hi
+#define MC_DQ5_LO dequantizeBlockAccum_ns_sgbroadcast_1_lo
+#endif
+#define MC_COL_Q5K(ts, c) \
+    { if (slid < 4) { regB.s0123 = read_imagef(src1, (c)*COL_STRIDE     + slid*2 + k*8); \
+                      regB.s4567 = read_imagef(src1, (c)*COL_STRIDE + 1 + slid*2 + k*8); } \
+      MC_DQ5_HI(ts, as_ushort8(regA_hi), as_uchar8(regH), regS, regM, regB); \
+      MC_DQ5_LO(ts, as_ushort8(regA_lo), as_uchar8(regH), regS, regM, regB); }
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_gemv_noshuffle_q5_k_f32_mc3(
+        read_only  image1d_buffer_t src0_q,
+        read_only  image1d_buffer_t src0_qh,
+        global half2  * src0_d,
+        global half2  * src0_m,
+        global uchar  * src0_s,
+        read_only  image1d_buffer_t src1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        uchar mask_d6,
+        uchar mask_d4,
+        uchar mask_hi2,
+        int n_cols)
+{
+    uint groupId = get_local_id(1);
+    uint gid     = get_global_id(0);
+    ushort slid  = get_sub_group_local_id();
+
+    uint K = ne00;
+    uint M = ne01;
+
+    uint LINE_STRIDE_A     = M / 2;
+    uint BLOCK_STRIDE_A    = NSUBGROUPS * M;
+    uint LINE_STRIDE_A_QH  = M / 2;
+    uint BLOCK_STRIDE_A_QH = NSUBGROUPS * M / 2;
+    uint scales_per_row    = (K / QK_K) * 12;
+    uint COL_STRIDE        = K / 4;   // float4 pixels per activation column
+
+    private uint4   regA_hi, regA_lo;
+    private ushort4 regH;
+    private half2   regS, regM;
+    private float8  regB;
+
+    private float2 ts0 = (float2)(0.0f);
+    private float2 ts1 = (float2)(0.0f);
+    private float2 ts2 = (float2)(0.0f);
+    private float2 ts3 = (float2)(0.0f);
+
+    for (uint k = groupId; k < (K / 32); k += NSUBGROUPS) {
+        uint sb = k / 8;
+        uint j  = k % 8;
+
+        half2 d   = src0_d[gid + sb * LINE_STRIDE_A];
+        half2 dm  = src0_m[gid + sb * LINE_STRIDE_A];
+
+        global const uchar * sc0 = src0_s + 2 * gid * scales_per_row + sb * 12;
+        global const uchar * sc1 = src0_s + (2 * gid + 1) * scales_per_row + sb * 12;
+
+        uchar sv0, mn0, sv1, mn1;
+        get_scale_min_k4(j, sc0, &sv0, &mn0, mask_d6, mask_d4, mask_hi2);
+        get_scale_min_k4(j, sc1, &sv1, &mn1, mask_d6, mask_d4, mask_hi2);
+
+        regS = convert_half2(convert_float2(d)  * convert_float2((uchar2)(sv0, sv1)));
+        regM = convert_half2(convert_float2(dm) * convert_float2((uchar2)(mn0, mn1)));
+
+        // high-bit plane + weights loaded ONCE, reused across the columns
+        regH.s0 = as_ushort(read_imageh(src0_qh, (gid + k * BLOCK_STRIDE_A_QH + LINE_STRIDE_A_QH * 0)).x);
+        regH.s1 = as_ushort(read_imageh(src0_qh, (gid + k * BLOCK_STRIDE_A_QH + LINE_STRIDE_A_QH * 1)).x);
+        regH.s2 = as_ushort(read_imageh(src0_qh, (gid + k * BLOCK_STRIDE_A_QH + LINE_STRIDE_A_QH * 2)).x);
+        regH.s3 = as_ushort(read_imageh(src0_qh, (gid + k * BLOCK_STRIDE_A_QH + LINE_STRIDE_A_QH * 3)).x);
+
+        regA_hi.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
+        regA_hi.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
+        regA_hi.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
+        regA_hi.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
+        regA_lo.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
+        regA_lo.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
+        regA_lo.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
+        regA_lo.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
+
+        MC_COL_Q5K(ts0, 0);
+        MC_COL_Q5K(ts1, 1);
+        if (n_cols > 2) MC_COL_Q5K(ts2, 2);
+        if (n_cols > 3) MC_COL_Q5K(ts3, 3);
+    }
+
+    // cross-subgroup reduce: pack the (up to 4) columns' float2 into a float8.
+    local float8 reduceLM[SUBGROUP_SIZE * 3];
+    float8 acc = (float8)(ts0.s0, ts0.s1, ts1.s0, ts1.s1, ts2.s0, ts2.s1, ts3.s0, ts3.s1);
+    if (groupId == 1) { reduceLM[SUBGROUP_SIZE * 0 + slid] = acc; }
+    if (groupId == 2) { reduceLM[SUBGROUP_SIZE * 1 + slid] = acc; }
+    if (groupId == 3) { reduceLM[SUBGROUP_SIZE * 2 + slid] = acc; }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (groupId == 0) {
+        acc += reduceLM[SUBGROUP_SIZE * 0 + slid];
+        acc += reduceLM[SUBGROUP_SIZE * 1 + slid];
+        acc += reduceLM[SUBGROUP_SIZE * 2 + slid];
+        dst = (global float*)((global char*)dst + offsetd);
+        // dst is column-major [M rows x n_cols cols]: (row, col) at col*M + row
+        vstore2((float2)(acc.s0, acc.s1), 0, &(dst[0 * M + gid * 2]));
+        vstore2((float2)(acc.s2, acc.s3), 0, &(dst[1 * M + gid * 2]));
+        if (n_cols > 2) vstore2((float2)(acc.s4, acc.s5), 0, &(dst[2 * M + gid * 2]));
+        if (n_cols > 3) vstore2((float2)(acc.s6, acc.s7), 0, &(dst[3 * M + gid * 2]));
+    }
+}
+#undef MC_COL_Q5K
+#undef MC_DQ5_HI
+#undef MC_DQ5_LO
@@ -296,3 +296,114 @@ kernel void kernel_gemv_noshuffle_q6_K_f32(
        if (gid * 2 + 1 < ne01) dst[gid * 2 + 1] = total_sum.s1;
    }
 }
+
+// Multi-column (N=3) q6_K decode GEMV for the spec/MTP verify batch. Same idea
+// as the q4_K mc3: stay on the efficient GEMV path (subgroup broadcast, no
+// transpose) instead of the transposed-GEMM dead-zone. Each K-block's weights
+// (ql/qh, hi+lo) are loaded ONCE and reused across all 3 activation columns.
+// Per-column accumulation is independent and identical to 3 standalone GEMVs
+// => byte-identical; does NOT perturb the lm_head logits / spec accept rate.
+#if defined(ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_gemv_noshuffle_q6_K_f32_mc3(
+    read_only image1d_buffer_t src0_ql,
+    read_only image1d_buffer_t src0_qh,
+    global half2 * src0_s,
+    global half2 * src0_d,
+    read_only image1d_buffer_t src1,
+    global float * dst,
+    ulong offsetd,
+    int ne00,
+    int ne01
+) {
+    int grp  = get_local_id(1);
+    int gid  = get_global_id(0);
+    ushort slid = get_sub_group_local_id();
+
+    int nb = ne00 / 32;
+    int line_stride_a  = ne01 / 2;
+    int block_stride_a = NSUBGROUPS * ne01;
+    int COL_STRIDE     = ne00 / 4;   // float4 pixels per activation column
+
+    uint4   ql_hi, ql_lo;
+    ushort4 qh_hi, qh_lo;
+    half2   reg_d;
+    char4   reg_s;
+    float8  reg_b;
+
+    float2  ts0 = 0.0f, ts1 = 0.0f, ts2 = 0.0f;
+
+    for (int k = grp; k < nb; k += NSUBGROUPS) {
+        reg_d = src0_d[gid + k/8 * line_stride_a];
+        reg_s = as_char4(src0_s[gid + k * line_stride_a]);
+
+        // weights loaded ONCE (hi: blocks 0-3, lo: blocks 4-7), reused x3 cols
+        ql_hi.s0 = read_imageui(src0_ql, gid + k*block_stride_a + line_stride_a*0).x;
+        ql_hi.s1 = read_imageui(src0_ql, gid + k*block_stride_a + line_stride_a*1).x;
+        ql_hi.s2 = read_imageui(src0_ql, gid + k*block_stride_a + line_stride_a*2).x;
+        ql_hi.s3 = read_imageui(src0_ql, gid + k*block_stride_a + line_stride_a*3).x;
+        qh_hi.s0 = as_ushort(read_imageh(src0_qh, gid + k*block_stride_a + line_stride_a*0).x);
+        qh_hi.s1 = as_ushort(read_imageh(src0_qh, gid + k*block_stride_a + line_stride_a*1).x);
+        qh_hi.s2 = as_ushort(read_imageh(src0_qh, gid + k*block_stride_a + line_stride_a*2).x);
+        qh_hi.s3 = as_ushort(read_imageh(src0_qh, gid + k*block_stride_a + line_stride_a*3).x);
+
+        ql_lo.s0 = read_imageui(src0_ql, gid + k*block_stride_a + line_stride_a*4).x;
+        ql_lo.s1 = read_imageui(src0_ql, gid + k*block_stride_a + line_stride_a*5).x;
+        ql_lo.s2 = read_imageui(src0_ql, gid + k*block_stride_a + line_stride_a*6).x;
+        ql_lo.s3 = read_imageui(src0_ql, gid + k*block_stride_a + line_stride_a*7).x;
+        qh_lo.s0 = as_ushort(read_imageh(src0_qh, gid + k*block_stride_a + line_stride_a*4).x);
+        qh_lo.s1 = as_ushort(read_imageh(src0_qh, gid + k*block_stride_a + line_stride_a*5).x);
+        qh_lo.s2 = as_ushort(read_imageh(src0_qh, gid + k*block_stride_a + line_stride_a*6).x);
+        qh_lo.s3 = as_ushort(read_imageh(src0_qh, gid + k*block_stride_a + line_stride_a*7).x);
+
+        // Per-column: load only this column's activation (single reg_b live) ->
+        // 1/3 the activation register pressure, cutting the private-mem spill.
+#ifdef VECTOR_SUB_GROUP_BROADCAT
+        { if (slid < 4) { reg_b.s0123 = read_imagef(src1, 0*COL_STRIDE + 0 + slid*2 + k*8);
+                          reg_b.s4567 = read_imagef(src1, 0*COL_STRIDE + 1 + slid*2 + k*8); }
+          dequantize_block_acc_bcast_8_hi(ts0, as_ushort8(ql_hi), as_uchar8(qh_hi), reg_d, reg_s, reg_b);
+          dequantize_block_acc_bcast_8_lo(ts0, as_ushort8(ql_lo), as_uchar8(qh_lo), reg_d, reg_s, reg_b); }
+        { if (slid < 4) { reg_b.s0123 = read_imagef(src1, 1*COL_STRIDE + 0 + slid*2 + k*8);
+                          reg_b.s4567 = read_imagef(src1, 1*COL_STRIDE + 1 + slid*2 + k*8); }
+          dequantize_block_acc_bcast_8_hi(ts1, as_ushort8(ql_hi), as_uchar8(qh_hi), reg_d, reg_s, reg_b);
+          dequantize_block_acc_bcast_8_lo(ts1, as_ushort8(ql_lo), as_uchar8(qh_lo), reg_d, reg_s, reg_b); }
+        { if (slid < 4) { reg_b.s0123 = read_imagef(src1, 2*COL_STRIDE + 0 + slid*2 + k*8);
+                          reg_b.s4567 = read_imagef(src1, 2*COL_STRIDE + 1 + slid*2 + k*8); }
+          dequantize_block_acc_bcast_8_hi(ts2, as_ushort8(ql_hi), as_uchar8(qh_hi), reg_d, reg_s, reg_b);
+          dequantize_block_acc_bcast_8_lo(ts2, as_ushort8(ql_lo), as_uchar8(qh_lo), reg_d, reg_s, reg_b); }
+#else
+        { if (slid < 4) { reg_b.s0123 = read_imagef(src1, 0*COL_STRIDE + 0 + slid*2 + k*8);
+                          reg_b.s4567 = read_imagef(src1, 0*COL_STRIDE + 1 + slid*2 + k*8); }
+          dequantize_block_acc_bcast_1_hi(ts0, as_ushort8(ql_hi), as_uchar8(qh_hi), reg_d, reg_s, reg_b);
+          dequantize_block_acc_bcast_1_lo(ts0, as_ushort8(ql_lo), as_uchar8(qh_lo), reg_d, reg_s, reg_b); }
+        { if (slid < 4) { reg_b.s0123 = read_imagef(src1, 1*COL_STRIDE + 0 + slid*2 + k*8);
+                          reg_b.s4567 = read_imagef(src1, 1*COL_STRIDE + 1 + slid*2 + k*8); }
+          dequantize_block_acc_bcast_1_hi(ts1, as_ushort8(ql_hi), as_uchar8(qh_hi), reg_d, reg_s, reg_b);
+          dequantize_block_acc_bcast_1_lo(ts1, as_ushort8(ql_lo), as_uchar8(qh_lo), reg_d, reg_s, reg_b); }
+        { if (slid < 4) { reg_b.s0123 = read_imagef(src1, 2*COL_STRIDE + 0 + slid*2 + k*8);
+                          reg_b.s4567 = read_imagef(src1, 2*COL_STRIDE + 1 + slid*2 + k*8); }
+          dequantize_block_acc_bcast_1_hi(ts2, as_ushort8(ql_hi), as_uchar8(qh_hi), reg_d, reg_s, reg_b);
+          dequantize_block_acc_bcast_1_lo(ts2, as_ushort8(ql_lo), as_uchar8(qh_lo), reg_d, reg_s, reg_b); }
+#endif
+    }
+
+    local float8 reduce_lm[SUBGROUP_SIZE * 3];
+    float8 acc = (float8)(ts0.s0, ts0.s1, ts1.s0, ts1.s1, ts2.s0, ts2.s1, 0.0f, 0.0f);
+    if (grp == 1) { reduce_lm[SUBGROUP_SIZE*0 + slid] = acc; }
+    if (grp == 2) { reduce_lm[SUBGROUP_SIZE*1 + slid] = acc; }
+    if (grp == 3) { reduce_lm[SUBGROUP_SIZE*2 + slid] = acc; }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (grp == 0) {
+        acc += reduce_lm[SUBGROUP_SIZE*0 + slid];
+        acc += reduce_lm[SUBGROUP_SIZE*1 + slid];
+        acc += reduce_lm[SUBGROUP_SIZE*2 + slid];
+        dst = (global float*)((global char*)dst + offsetd);
+        // dst column-major [ne01 rows x 3 cols]: (row, col) at col*ne01 + row
+        vstore2((float2)(acc.s0, acc.s1), 0, &(dst[0*ne01 + gid*2]));
+        vstore2((float2)(acc.s2, acc.s3), 0, &(dst[1*ne01 + gid*2]));
+        vstore2((float2)(acc.s4, acc.s5), 0, &(dst[2*ne01 + gid*2]));
+    }
+}
@@ -0,0 +1,372 @@
+// 4-output-per-WI variant of kernel_gemv_noshuffle_q6_K_f32.
+// Each WI now produces 4 consecutive outputs (output quad). The activation
+// fetch (reg_b) is shared across all 4 outputs, doubling per-WI ALU per
+// activation broadcast and halving the WG count vs the 2-output kernel.
+//
+// Implementation: each K-block we fetch TWO sets of (scales + ql + qh)
+// — one for the low pair (rows 0,1 of the quad) and one for the high pair
+// (rows 2,3) — and invoke the existing 2-output dequant macros twice
+// against the *same* reg_b. Identical data layout to the 2-output kernel,
+// so the host only needs to halve the grid and double the gid-to-output
+// mapping.
+//
+// Opt-in via the host dispatch when GGML_OPENCL_Q6K_GEMV_O4=1.
+
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+#define NSUBGROUPS 4
+#define SUBGROUP_SIZE 64
+
+// Macros are identical to the 2-output kernel — they accept `total_sum` as
+// a parameter so we can call them twice (once per pair) against different
+// accumulators against the same reg_b.
+#define dequantize_block_acc_bcast_8_hi(total_sum, bits4, bits2, cs, y) \
+    float8 shared_y; \
+    shared_y = sub_group_broadcast(y, 0); \
+    total_sum.s0 += ((float)(((bits4.s0 & 0x000F)      ) | ((bits2.s0 & 0x03) << 4)) - 32.f) * cs.s0 * shared_y.s0; \
+    total_sum.s0 += ((float)(((bits4.s0 & 0x00F0) >>  4) | ((bits2.s0 & 0x0C) << 2)) - 32.f) * cs.s0 * shared_y.s1; \
+    total_sum.s0 += ((float)(((bits4.s0 & 0x0F00) >>  8) | ((bits2.s0 & 0x30)     )) - 32.f) * cs.s0 * shared_y.s2; \
+    total_sum.s0 += ((float)(((bits4.s0 & 0xF000) >> 12) | ((bits2.s0 & 0xC0) >> 2)) - 32.f) * cs.s0 * shared_y.s3; \
+    total_sum.s0 += ((float)(((bits4.s2 & 0x000F)      ) | ((bits2.s2 & 0x03) << 4)) - 32.f) * cs.s0 * shared_y.s4; \
+    total_sum.s0 += ((float)(((bits4.s2 & 0x00F0) >>  4) | ((bits2.s2 & 0x0C) << 2)) - 32.f) * cs.s0 * shared_y.s5; \
+    total_sum.s0 += ((float)(((bits4.s2 & 0x0F00) >>  8) | ((bits2.s2 & 0x30)     )) - 32.f) * cs.s0 * shared_y.s6; \
+    total_sum.s0 += ((float)(((bits4.s2 & 0xF000) >> 12) | ((bits2.s2 & 0xC0) >> 2)) - 32.f) * cs.s0 * shared_y.s7; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0x000F)      ) | ((bits2.s1 & 0x03) << 4)) - 32.f) * cs.s2 * shared_y.s0; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0x00F0) >>  4) | ((bits2.s1 & 0x0C) << 2)) - 32.f) * cs.s2 * shared_y.s1; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0x0F00) >>  8) | ((bits2.s1 & 0x30)     )) - 32.f) * cs.s2 * shared_y.s2; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0xF000) >> 12) | ((bits2.s1 & 0xC0) >> 2)) - 32.f) * cs.s2 * shared_y.s3; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0x000F)      ) | ((bits2.s3 & 0x03) << 4)) - 32.f) * cs.s2 * shared_y.s4; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0x00F0) >>  4) | ((bits2.s3 & 0x0C) << 2)) - 32.f) * cs.s2 * shared_y.s5; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0x0F00) >>  8) | ((bits2.s3 & 0x30)     )) - 32.f) * cs.s2 * shared_y.s6; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0xF000) >> 12) | ((bits2.s3 & 0xC0) >> 2)) - 32.f) * cs.s2 * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 1); \
+    total_sum.s0 += ((float)(((bits4.s4 & 0x000F)      ) | ((bits2.s4 & 0x03) << 4)) - 32.f) * cs.s0 * shared_y.s0; \
+    total_sum.s0 += ((float)(((bits4.s4 & 0x00F0) >>  4) | ((bits2.s4 & 0x0C) << 2)) - 32.f) * cs.s0 * shared_y.s1; \
+    total_sum.s0 += ((float)(((bits4.s4 & 0x0F00) >>  8) | ((bits2.s4 & 0x30)     )) - 32.f) * cs.s0 * shared_y.s2; \
+    total_sum.s0 += ((float)(((bits4.s4 & 0xF000) >> 12) | ((bits2.s4 & 0xC0) >> 2)) - 32.f) * cs.s0 * shared_y.s3; \
+    total_sum.s0 += ((float)(((bits4.s6 & 0x000F)      ) | ((bits2.s6 & 0x03) << 4)) - 32.f) * cs.s0 * shared_y.s4; \
+    total_sum.s0 += ((float)(((bits4.s6 & 0x00F0) >>  4) | ((bits2.s6 & 0x0C) << 2)) - 32.f) * cs.s0 * shared_y.s5; \
+    total_sum.s0 += ((float)(((bits4.s6 & 0x0F00) >>  8) | ((bits2.s6 & 0x30)     )) - 32.f) * cs.s0 * shared_y.s6; \
+    total_sum.s0 += ((float)(((bits4.s6 & 0xF000) >> 12) | ((bits2.s6 & 0xC0) >> 2)) - 32.f) * cs.s0 * shared_y.s7; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0x000F)      ) | ((bits2.s5 & 0x03) << 4)) - 32.f) * cs.s2 * shared_y.s0; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0x00F0) >>  4) | ((bits2.s5 & 0x0C) << 2)) - 32.f) * cs.s2 * shared_y.s1; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0x0F00) >>  8) | ((bits2.s5 & 0x30)     )) - 32.f) * cs.s2 * shared_y.s2; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0xF000) >> 12) | ((bits2.s5 & 0xC0) >> 2)) - 32.f) * cs.s2 * shared_y.s3; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0x000F)      ) | ((bits2.s7 & 0x03) << 4)) - 32.f) * cs.s2 * shared_y.s4; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0x00F0) >>  4) | ((bits2.s7 & 0x0C) << 2)) - 32.f) * cs.s2 * shared_y.s5; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0x0F00) >>  8) | ((bits2.s7 & 0x30)     )) - 32.f) * cs.s2 * shared_y.s6; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0xF000) >> 12) | ((bits2.s7 & 0xC0) >> 2)) - 32.f) * cs.s2 * shared_y.s7; \
+
+#define dequantize_block_acc_bcast_8_lo(total_sum, bits4, bits2, cs, y) \
+    shared_y = sub_group_broadcast(y, 2); \
+    total_sum.s0 += ((float)(((bits4.s0 & 0x000F)      ) | ((bits2.s0 & 0x03) << 4)) - 32.f) * cs.s1 * shared_y.s0; \
+    total_sum.s0 += ((float)(((bits4.s0 & 0x00F0) >>  4) | ((bits2.s0 & 0x0C) << 2)) - 32.f) * cs.s1 * shared_y.s1; \
+    total_sum.s0 += ((float)(((bits4.s0 & 0x0F00) >>  8) | ((bits2.s0 & 0x30)     )) - 32.f) * cs.s1 * shared_y.s2; \
+    total_sum.s0 += ((float)(((bits4.s0 & 0xF000) >> 12) | ((bits2.s0 & 0xC0) >> 2)) - 32.f) * cs.s1 * shared_y.s3; \
+    total_sum.s0 += ((float)(((bits4.s2 & 0x000F)      ) | ((bits2.s2 & 0x03) << 4)) - 32.f) * cs.s1 * shared_y.s4; \
+    total_sum.s0 += ((float)(((bits4.s2 & 0x00F0) >>  4) | ((bits2.s2 & 0x0C) << 2)) - 32.f) * cs.s1 * shared_y.s5; \
+    total_sum.s0 += ((float)(((bits4.s2 & 0x0F00) >>  8) | ((bits2.s2 & 0x30)     )) - 32.f) * cs.s1 * shared_y.s6; \
+    total_sum.s0 += ((float)(((bits4.s2 & 0xF000) >> 12) | ((bits2.s2 & 0xC0) >> 2)) - 32.f) * cs.s1 * shared_y.s7; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0x000F)      ) | ((bits2.s1 & 0x03) << 4)) - 32.f) * cs.s3 * shared_y.s0; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0x00F0) >>  4) | ((bits2.s1 & 0x0C) << 2)) - 32.f) * cs.s3 * shared_y.s1; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0x0F00) >>  8) | ((bits2.s1 & 0x30)     )) - 32.f) * cs.s3 * shared_y.s2; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0xF000) >> 12) | ((bits2.s1 & 0xC0) >> 2)) - 32.f) * cs.s3 * shared_y.s3; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0x000F)      ) | ((bits2.s3 & 0x03) << 4)) - 32.f) * cs.s3 * shared_y.s4; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0x00F0) >>  4) | ((bits2.s3 & 0x0C) << 2)) - 32.f) * cs.s3 * shared_y.s5; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0x0F00) >>  8) | ((bits2.s3 & 0x30)     )) - 32.f) * cs.s3 * shared_y.s6; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0xF000) >> 12) | ((bits2.s3 & 0xC0) >> 2)) - 32.f) * cs.s3 * shared_y.s7; \
+    shared_y = sub_group_broadcast(y, 3); \
+    total_sum.s0 += ((float)(((bits4.s4 & 0x000F)      ) | ((bits2.s4 & 0x03) << 4)) - 32.f) * cs.s1 * shared_y.s0; \
+    total_sum.s0 += ((float)(((bits4.s4 & 0x00F0) >>  4) | ((bits2.s4 & 0x0C) << 2)) - 32.f) * cs.s1 * shared_y.s1; \
+    total_sum.s0 += ((float)(((bits4.s4 & 0x0F00) >>  8) | ((bits2.s4 & 0x30)     )) - 32.f) * cs.s1 * shared_y.s2; \
+    total_sum.s0 += ((float)(((bits4.s4 & 0xF000) >> 12) | ((bits2.s4 & 0xC0) >> 2)) - 32.f) * cs.s1 * shared_y.s3; \
+    total_sum.s0 += ((float)(((bits4.s6 & 0x000F)      ) | ((bits2.s6 & 0x03) << 4)) - 32.f) * cs.s1 * shared_y.s4; \
+    total_sum.s0 += ((float)(((bits4.s6 & 0x00F0) >>  4) | ((bits2.s6 & 0x0C) << 2)) - 32.f) * cs.s1 * shared_y.s5; \
+    total_sum.s0 += ((float)(((bits4.s6 & 0x0F00) >>  8) | ((bits2.s6 & 0x30)     )) - 32.f) * cs.s1 * shared_y.s6; \
+    total_sum.s0 += ((float)(((bits4.s6 & 0xF000) >> 12) | ((bits2.s6 & 0xC0) >> 2)) - 32.f) * cs.s1 * shared_y.s7; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0x000F)      ) | ((bits2.s5 & 0x03) << 4)) - 32.f) * cs.s3 * shared_y.s0; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0x00F0) >>  4) | ((bits2.s5 & 0x0C) << 2)) - 32.f) * cs.s3 * shared_y.s1; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0x0F00) >>  8) | ((bits2.s5 & 0x30)     )) - 32.f) * cs.s3 * shared_y.s2; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0xF000) >> 12) | ((bits2.s5 & 0xC0) >> 2)) - 32.f) * cs.s3 * shared_y.s3; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0x000F)      ) | ((bits2.s7 & 0x03) << 4)) - 32.f) * cs.s3 * shared_y.s4; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0x00F0) >>  4) | ((bits2.s7 & 0x0C) << 2)) - 32.f) * cs.s3 * shared_y.s5; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0x0F00) >>  8) | ((bits2.s7 & 0x30)     )) - 32.f) * cs.s3 * shared_y.s6; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0xF000) >> 12) | ((bits2.s7 & 0xC0) >> 2)) - 32.f) * cs.s3 * shared_y.s7; \
+
+#define dequantize_block_acc_bcast_1_hi(total_sum, bits4, bits2, cs, y) \
+    float shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 0); \
+    total_sum.s0 += ((float)(((bits4.s0 & 0x000F)      ) | ((bits2.s0 & 0x03) << 4)) - 32.f) * cs.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0x000F)      ) | ((bits2.s1 & 0x03) << 4)) - 32.f) * cs.s2 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 0); \
+    total_sum.s0 += ((float)(((bits4.s0 & 0x00F0) >>  4) | ((bits2.s0 & 0x0C) << 2)) - 32.f) * cs.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0x00F0) >>  4) | ((bits2.s1 & 0x0C) << 2)) - 32.f) * cs.s2 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 0); \
+    total_sum.s0 += ((float)(((bits4.s0 & 0x0F00) >>  8) | ((bits2.s0 & 0x30)     )) - 32.f) * cs.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0x0F00) >>  8) | ((bits2.s1 & 0x30)     )) - 32.f) * cs.s2 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 0); \
+    total_sum.s0 += ((float)(((bits4.s0 & 0xF000) >> 12) | ((bits2.s0 & 0xC0) >> 2)) - 32.f) * cs.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0xF000) >> 12) | ((bits2.s1 & 0xC0) >> 2)) - 32.f) * cs.s2 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 0); \
+    total_sum.s0 += ((float)(((bits4.s2 & 0x000F)      ) | ((bits2.s2 & 0x03) << 4)) - 32.f) * cs.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0x000F)      ) | ((bits2.s3 & 0x03) << 4)) - 32.f) * cs.s2 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 0); \
+    total_sum.s0 += ((float)(((bits4.s2 & 0x00F0) >>  4) | ((bits2.s2 & 0x0C) << 2)) - 32.f) * cs.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0x00F0) >>  4) | ((bits2.s3 & 0x0C) << 2)) - 32.f) * cs.s2 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 0); \
+    total_sum.s0 += ((float)(((bits4.s2 & 0x0F00) >>  8) | ((bits2.s2 & 0x30)     )) - 32.f) * cs.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0x0F00) >>  8) | ((bits2.s3 & 0x30)     )) - 32.f) * cs.s2 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 0); \
+    total_sum.s0 += ((float)(((bits4.s2 & 0xF000) >> 12) | ((bits2.s2 & 0xC0) >> 2)) - 32.f) * cs.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0xF000) >> 12) | ((bits2.s3 & 0xC0) >> 2)) - 32.f) * cs.s2 * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 1); \
+    total_sum.s0 += ((float)(((bits4.s4 & 0x000F)      ) | ((bits2.s4 & 0x03) << 4)) - 32.f) * cs.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0x000F)      ) | ((bits2.s5 & 0x03) << 4)) - 32.f) * cs.s2 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 1); \
+    total_sum.s0 += ((float)(((bits4.s4 & 0x00F0) >>  4) | ((bits2.s4 & 0x0C) << 2)) - 32.f) * cs.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0x00F0) >>  4) | ((bits2.s5 & 0x0C) << 2)) - 32.f) * cs.s2 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 1); \
+    total_sum.s0 += ((float)(((bits4.s4 & 0x0F00) >>  8) | ((bits2.s4 & 0x30)     )) - 32.f) * cs.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0x0F00) >>  8) | ((bits2.s5 & 0x30)     )) - 32.f) * cs.s2 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 1); \
+    total_sum.s0 += ((float)(((bits4.s4 & 0xF000) >> 12) | ((bits2.s4 & 0xC0) >> 2)) - 32.f) * cs.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0xF000) >> 12) | ((bits2.s5 & 0xC0) >> 2)) - 32.f) * cs.s2 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 1); \
+    total_sum.s0 += ((float)(((bits4.s6 & 0x000F)      ) | ((bits2.s6 & 0x03) << 4)) - 32.f) * cs.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0x000F)      ) | ((bits2.s7 & 0x03) << 4)) - 32.f) * cs.s2 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 1); \
+    total_sum.s0 += ((float)(((bits4.s6 & 0x00F0) >>  4) | ((bits2.s6 & 0x0C) << 2)) - 32.f) * cs.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0x00F0) >>  4) | ((bits2.s7 & 0x0C) << 2)) - 32.f) * cs.s2 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 1); \
+    total_sum.s0 += ((float)(((bits4.s6 & 0x0F00) >>  8) | ((bits2.s6 & 0x30)     )) - 32.f) * cs.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0x0F00) >>  8) | ((bits2.s7 & 0x30)     )) - 32.f) * cs.s2 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 1); \
+    total_sum.s0 += ((float)(((bits4.s6 & 0xF000) >> 12) | ((bits2.s6 & 0xC0) >> 2)) - 32.f) * cs.s0 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0xF000) >> 12) | ((bits2.s7 & 0xC0) >> 2)) - 32.f) * cs.s2 * shared_y; \
+
+#define dequantize_block_acc_bcast_1_lo(total_sum, bits4, bits2, cs, y) \
+    shared_y = sub_group_broadcast(y.s0, 2); \
+    total_sum.s0 += ((float)(((bits4.s0 & 0x000F)      ) | ((bits2.s0 & 0x03) << 4)) - 32.f) * cs.s1 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0x000F)      ) | ((bits2.s1 & 0x03) << 4)) - 32.f) * cs.s3 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 2); \
+    total_sum.s0 += ((float)(((bits4.s0 & 0x00F0) >>  4) | ((bits2.s0 & 0x0C) << 2)) - 32.f) * cs.s1 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0x00F0) >>  4) | ((bits2.s1 & 0x0C) << 2)) - 32.f) * cs.s3 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 2); \
+    total_sum.s0 += ((float)(((bits4.s0 & 0x0F00) >>  8) | ((bits2.s0 & 0x30)     )) - 32.f) * cs.s1 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0x0F00) >>  8) | ((bits2.s1 & 0x30)     )) - 32.f) * cs.s3 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 2); \
+    total_sum.s0 += ((float)(((bits4.s0 & 0xF000) >> 12) | ((bits2.s0 & 0xC0) >> 2)) - 32.f) * cs.s1 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s1 & 0xF000) >> 12) | ((bits2.s1 & 0xC0) >> 2)) - 32.f) * cs.s3 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 2); \
+    total_sum.s0 += ((float)(((bits4.s2 & 0x000F)      ) | ((bits2.s2 & 0x03) << 4)) - 32.f) * cs.s1 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0x000F)      ) | ((bits2.s3 & 0x03) << 4)) - 32.f) * cs.s3 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 2); \
+    total_sum.s0 += ((float)(((bits4.s2 & 0x00F0) >>  4) | ((bits2.s2 & 0x0C) << 2)) - 32.f) * cs.s1 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0x00F0) >>  4) | ((bits2.s3 & 0x0C) << 2)) - 32.f) * cs.s3 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 2); \
+    total_sum.s0 += ((float)(((bits4.s2 & 0x0F00) >>  8) | ((bits2.s2 & 0x30)     )) - 32.f) * cs.s1 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0x0F00) >>  8) | ((bits2.s3 & 0x30)     )) - 32.f) * cs.s3 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 2); \
+    total_sum.s0 += ((float)(((bits4.s2 & 0xF000) >> 12) | ((bits2.s2 & 0xC0) >> 2)) - 32.f) * cs.s1 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s3 & 0xF000) >> 12) | ((bits2.s3 & 0xC0) >> 2)) - 32.f) * cs.s3 * shared_y; \
+    shared_y = sub_group_broadcast(y.s0, 3); \
+    total_sum.s0 += ((float)(((bits4.s4 & 0x000F)      ) | ((bits2.s4 & 0x03) << 4)) - 32.f) * cs.s1 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0x000F)      ) | ((bits2.s5 & 0x03) << 4)) - 32.f) * cs.s3 * shared_y; \
+    shared_y = sub_group_broadcast(y.s1, 3); \
+    total_sum.s0 += ((float)(((bits4.s4 & 0x00F0) >>  4) | ((bits2.s4 & 0x0C) << 2)) - 32.f) * cs.s1 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0x00F0) >>  4) | ((bits2.s5 & 0x0C) << 2)) - 32.f) * cs.s3 * shared_y; \
+    shared_y = sub_group_broadcast(y.s2, 3); \
+    total_sum.s0 += ((float)(((bits4.s4 & 0x0F00) >>  8) | ((bits2.s4 & 0x30)     )) - 32.f) * cs.s1 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0x0F00) >>  8) | ((bits2.s5 & 0x30)     )) - 32.f) * cs.s3 * shared_y; \
+    shared_y = sub_group_broadcast(y.s3, 3); \
+    total_sum.s0 += ((float)(((bits4.s4 & 0xF000) >> 12) | ((bits2.s4 & 0xC0) >> 2)) - 32.f) * cs.s1 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s5 & 0xF000) >> 12) | ((bits2.s5 & 0xC0) >> 2)) - 32.f) * cs.s3 * shared_y; \
+    shared_y = sub_group_broadcast(y.s4, 3); \
+    total_sum.s0 += ((float)(((bits4.s6 & 0x000F)      ) | ((bits2.s6 & 0x03) << 4)) - 32.f) * cs.s1 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0x000F)      ) | ((bits2.s7 & 0x03) << 4)) - 32.f) * cs.s3 * shared_y; \
+    shared_y = sub_group_broadcast(y.s5, 3); \
+    total_sum.s0 += ((float)(((bits4.s6 & 0x00F0) >>  4) | ((bits2.s6 & 0x0C) << 2)) - 32.f) * cs.s1 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0x00F0) >>  4) | ((bits2.s7 & 0x0C) << 2)) - 32.f) * cs.s3 * shared_y; \
+    shared_y = sub_group_broadcast(y.s6, 3); \
+    total_sum.s0 += ((float)(((bits4.s6 & 0x0F00) >>  8) | ((bits2.s6 & 0x30)     )) - 32.f) * cs.s1 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0x0F00) >>  8) | ((bits2.s7 & 0x30)     )) - 32.f) * cs.s3 * shared_y; \
+    shared_y = sub_group_broadcast(y.s7, 3); \
+    total_sum.s0 += ((float)(((bits4.s6 & 0xF000) >> 12) | ((bits2.s6 & 0xC0) >> 2)) - 32.f) * cs.s1 * shared_y; \
+    total_sum.s1 += ((float)(((bits4.s7 & 0xF000) >> 12) | ((bits2.s7 & 0xC0) >> 2)) - 32.f) * cs.s3 * shared_y; \
+
+#if defined(ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+// Q6K_O4_GLOBAL: read the (read-once-per-token, no-reuse) lm_head/embed weights
+// from __global coalesced instead of image1d_buffer. The texture cache caps the
+// streaming (no-reuse) lm_head read bandwidth; global coalesced reaches the
+// higher rate the rest of the model gets. src1 (activation) stays an image (it IS reused via
+// the cross-subgroup broadcast).
+#ifdef Q6K_O4_GLOBAL
+#define Q6K_O4_NAME kernel_gemv_noshuffle_q6_K_f32_o4_global
+#define QL_ARG __global uint * src0_ql
+#define QH_ARG __global half * src0_qh
+#define RD_QL(b,i) (b[i])
+#define RD_QH(b,i) as_ushort(b[i])
+#else
+#define Q6K_O4_NAME kernel_gemv_noshuffle_q6_K_f32_o4
+#define QL_ARG read_only image1d_buffer_t src0_ql
+#define QH_ARG read_only image1d_buffer_t src0_qh
+#define RD_QL(b,i) (read_imageui(b,i).x)
+#define RD_QH(b,i) as_ushort(read_imageh(b,i).x)
+#endif
+kernel void Q6K_O4_NAME(
+    QL_ARG,
+    QH_ARG,
+    global half2 * src0_s,
+    global half2 * src0_d,
+    read_only image1d_buffer_t src1,
+    global float * dst,
+    ulong offsetd,
+    int ne00,
+    int ne01
+) {
+    int grp = get_local_id(1);
+    int gid = get_global_id(0);              // 4-output-quad index
+    ushort slid = get_sub_group_local_id();
+
+    // Map quad index to the two pair-indices the existing 2-output access
+    // pattern uses (consecutive output pairs along ne01). NB: the two pairs are
+    // kept ADJACENT (gid*2, gid*2+1) on purpose -- a "stride-1" split (pairs
+    // ne01/4 apart) is slower because two distant cache-line streams have worse
+    // locality than the adjacent pair whose reads interleave into the same lines
+    // each iteration.
+    int gid_a = gid * 2;
+    int gid_b = gid * 2 + 1;
+
+    int nb = ne00 / 32;
+
+    uint4   reg_a_l_a, reg_a_l_b;
+    ushort4 reg_a_h_a, reg_a_h_b;
+    half2   reg_d_a,   reg_d_b;
+    char4   reg_s_a,   reg_s_b;
+    float8  reg_b;
+
+    float2 total_sum_a = 0.0f;
+    float2 total_sum_b = 0.0f;
+
+    int line_stride_a = ne01 / 2;
+    int block_stride_a = NSUBGROUPS * ne01;
+
+    for (int k = grp; k < nb; k += NSUBGROUPS) {
+        reg_d_a = src0_d[gid_a + k/8 * line_stride_a];
+        reg_d_b = src0_d[gid_b + k/8 * line_stride_a];
+        reg_s_a = as_char4(src0_s[gid_a + k * line_stride_a]);
+        reg_s_b = as_char4(src0_s[gid_b + k * line_stride_a]);
+        // Precompute the loop-invariant combined scale (sub-block scale * super-block d)
+        // once per pair instead of re-multiplying it for every one of the 256 elements.
+        float4 cs_a = (float4)((float)reg_s_a.s0*(float)reg_d_a.s0, (float)reg_s_a.s1*(float)reg_d_a.s0,
+                               (float)reg_s_a.s2*(float)reg_d_a.s1, (float)reg_s_a.s3*(float)reg_d_a.s1);
+        float4 cs_b = (float4)((float)reg_s_b.s0*(float)reg_d_b.s0, (float)reg_s_b.s1*(float)reg_d_b.s0,
+                               (float)reg_s_b.s2*(float)reg_d_b.s1, (float)reg_s_b.s3*(float)reg_d_b.s1);
+
+        if (slid < 4) {
+            reg_b.s0123 = read_imagef(src1, 0 + slid*2 + k*8);
+            reg_b.s4567 = read_imagef(src1, 1 + slid*2 + k*8);
+        }
+
+        // Pair a (output rows gid_a*2, gid_a*2+1): read hi+lo then dequant
+        // both in one block so the `_lo` macro can see the `shared_y` that
+        // `_hi` declared. Pair b follows in its own block — fresh shared_y.
+        {
+            reg_a_l_a.s0 = RD_QL(src0_ql, gid_a + k*block_stride_a + line_stride_a*0);
+            reg_a_l_a.s1 = RD_QL(src0_ql, gid_a + k*block_stride_a + line_stride_a*1);
+            reg_a_l_a.s2 = RD_QL(src0_ql, gid_a + k*block_stride_a + line_stride_a*2);
+            reg_a_l_a.s3 = RD_QL(src0_ql, gid_a + k*block_stride_a + line_stride_a*3);
+            reg_a_h_a.s0 = RD_QH(src0_qh, gid_a + k*block_stride_a + line_stride_a*0);
+            reg_a_h_a.s1 = RD_QH(src0_qh, gid_a + k*block_stride_a + line_stride_a*1);
+            reg_a_h_a.s2 = RD_QH(src0_qh, gid_a + k*block_stride_a + line_stride_a*2);
+            reg_a_h_a.s3 = RD_QH(src0_qh, gid_a + k*block_stride_a + line_stride_a*3);
+#ifdef VECTOR_SUB_GROUP_BROADCAT
+            dequantize_block_acc_bcast_8_hi(total_sum_a, as_ushort8(reg_a_l_a), as_uchar8(reg_a_h_a), cs_a, reg_b);
+#else
+            dequantize_block_acc_bcast_1_hi(total_sum_a, as_ushort8(reg_a_l_a), as_uchar8(reg_a_h_a), cs_a, reg_b);
+#endif
+
+            reg_a_l_a.s0 = RD_QL(src0_ql, gid_a + k*block_stride_a + line_stride_a*4);
+            reg_a_l_a.s1 = RD_QL(src0_ql, gid_a + k*block_stride_a + line_stride_a*5);
+            reg_a_l_a.s2 = RD_QL(src0_ql, gid_a + k*block_stride_a + line_stride_a*6);
+            reg_a_l_a.s3 = RD_QL(src0_ql, gid_a + k*block_stride_a + line_stride_a*7);
+            reg_a_h_a.s0 = RD_QH(src0_qh, gid_a + k*block_stride_a + line_stride_a*4);
+            reg_a_h_a.s1 = RD_QH(src0_qh, gid_a + k*block_stride_a + line_stride_a*5);
+            reg_a_h_a.s2 = RD_QH(src0_qh, gid_a + k*block_stride_a + line_stride_a*6);
+            reg_a_h_a.s3 = RD_QH(src0_qh, gid_a + k*block_stride_a + line_stride_a*7);
+#ifdef VECTOR_SUB_GROUP_BROADCAT
+            dequantize_block_acc_bcast_8_lo(total_sum_a, as_ushort8(reg_a_l_a), as_uchar8(reg_a_h_a), cs_a, reg_b);
+#else
+            dequantize_block_acc_bcast_1_lo(total_sum_a, as_ushort8(reg_a_l_a), as_uchar8(reg_a_h_a), cs_a, reg_b);
+#endif
+        }
+
+        {
+            reg_a_l_b.s0 = RD_QL(src0_ql, gid_b + k*block_stride_a + line_stride_a*0);
+            reg_a_l_b.s1 = RD_QL(src0_ql, gid_b + k*block_stride_a + line_stride_a*1);
+            reg_a_l_b.s2 = RD_QL(src0_ql, gid_b + k*block_stride_a + line_stride_a*2);
+            reg_a_l_b.s3 = RD_QL(src0_ql, gid_b + k*block_stride_a + line_stride_a*3);
+            reg_a_h_b.s0 = RD_QH(src0_qh, gid_b + k*block_stride_a + line_stride_a*0);
+            reg_a_h_b.s1 = RD_QH(src0_qh, gid_b + k*block_stride_a + line_stride_a*1);
+            reg_a_h_b.s2 = RD_QH(src0_qh, gid_b + k*block_stride_a + line_stride_a*2);
+            reg_a_h_b.s3 = RD_QH(src0_qh, gid_b + k*block_stride_a + line_stride_a*3);
+#ifdef VECTOR_SUB_GROUP_BROADCAT
+            dequantize_block_acc_bcast_8_hi(total_sum_b, as_ushort8(reg_a_l_b), as_uchar8(reg_a_h_b), cs_b, reg_b);
+#else
+            dequantize_block_acc_bcast_1_hi(total_sum_b, as_ushort8(reg_a_l_b), as_uchar8(reg_a_h_b), cs_b, reg_b);
+#endif
+
+            reg_a_l_b.s0 = RD_QL(src0_ql, gid_b + k*block_stride_a + line_stride_a*4);
+            reg_a_l_b.s1 = RD_QL(src0_ql, gid_b + k*block_stride_a + line_stride_a*5);
+            reg_a_l_b.s2 = RD_QL(src0_ql, gid_b + k*block_stride_a + line_stride_a*6);
+            reg_a_l_b.s3 = RD_QL(src0_ql, gid_b + k*block_stride_a + line_stride_a*7);
+            reg_a_h_b.s0 = RD_QH(src0_qh, gid_b + k*block_stride_a + line_stride_a*4);
+            reg_a_h_b.s1 = RD_QH(src0_qh, gid_b + k*block_stride_a + line_stride_a*5);
+            reg_a_h_b.s2 = RD_QH(src0_qh, gid_b + k*block_stride_a + line_stride_a*6);
+            reg_a_h_b.s3 = RD_QH(src0_qh, gid_b + k*block_stride_a + line_stride_a*7);
+#ifdef VECTOR_SUB_GROUP_BROADCAT
+            dequantize_block_acc_bcast_8_lo(total_sum_b, as_ushort8(reg_a_l_b), as_uchar8(reg_a_h_b), cs_b, reg_b);
+#else
+            dequantize_block_acc_bcast_1_lo(total_sum_b, as_ushort8(reg_a_l_b), as_uchar8(reg_a_h_b), cs_b, reg_b);
+#endif
+        }
+    }
+
+    // Cross-subgroup reduce. Same shape as the 2-output kernel but with the
+    // pair-a and pair-b accumulators concatenated into a single float4.
+    local float4 reduce_lm[SUBGROUP_SIZE * 3];
+    float4 acc = (float4)(total_sum_a.s0, total_sum_a.s1, total_sum_b.s0, total_sum_b.s1);
+    if (grp == 1) { reduce_lm[SUBGROUP_SIZE*0 + slid] = acc; }
+    if (grp == 2) { reduce_lm[SUBGROUP_SIZE*1 + slid] = acc; }
+    if (grp == 3) { reduce_lm[SUBGROUP_SIZE*2 + slid] = acc; }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (grp == 0) {
+        acc += reduce_lm[SUBGROUP_SIZE*0 + slid];
+        acc += reduce_lm[SUBGROUP_SIZE*1 + slid];
+        acc += reduce_lm[SUBGROUP_SIZE*2 + slid];
+        dst = (global float*)((global char*)dst + offsetd);
+        // The dispatch rounds ne01/4 up to the subgroup width, so the tail
+        // quads past the last row must not store (they wrote 128 rows past
+        // dst on every ne01 % 256 == 128 vocab, e.g. 151936).
+        if (gid * 4 + 3 < (uint)ne01) {
+            vstore4(acc, 0, &(dst[gid * 4]));
+        }
+    }
+}
@@ -0,0 +1,196 @@
+// Tiled-wide q6_K GEMV for the long-vocab lm_head/embed (decode path).
+//
+// Pairs with kernel_convert_block_q6_k_tiled_ns (cvt.cl): the weights are laid
+// out CANONICALLY (6-bit code in element order e in [0,256)) and TILED by 64
+// output rows so the 64-thread lane group coalesces every weight load. Both the
+// pack (convert) and the unpack (here) are owned by us — correct by construction
+// against the reference ggml q6_K dequant, no bit-interleave reverse-engineering.
+//
+// One work-item produces one output row. A work-group is {64 lanes, 4 subgroups}:
+// the 64 lanes cover the 64 rows of one tile (coalesced reads), the 4 subgroups
+// split the K-blocks and reduce through __local at the end.
+//
+// Weights are read from __global (coalesced) rather than image1d_buffer: the
+// lm_head is read once per token with no reuse, and the Adreno texture cache
+// caps such a streaming read well below the coalesced-global rate
+// (see opencl_q6k_gemv_o4_shipped / x2-90 roofline notes).
+
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#endif
+
+#define NSUBGROUPS 4
+#define TILE_ROWS  64
+
+#if defined(ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_gemv_noshuffle_q6_K_f32_tiled(
+    __global uint4 * src0_ql,   // tiled: 8 uint4 granules / superblock
+    __global uint4 * src0_qh,   // tiled: 4 uint4 granules / superblock
+    __global char  * src0_s,    // tiled: 16 chars / superblock
+    __global half  * src0_d,    // tiled: 1 half  / superblock
+    read_only image1d_buffer_t src1,   // activation (RGBA f32)
+    global float * dst,
+    ulong offsetd,
+    int ne00,
+    int ne01
+) {
+    int grp = get_local_id(1);          // subgroup index 0..3 (splits K)
+    int row = get_global_id(0);         // output row along ne01
+    int rt  = row / TILE_ROWS;
+    int rit = row % TILE_ROWS;
+
+    int nb = ne00 / 256;                // superblocks per row
+
+    float acc = 0.0f;
+
+    for (int sb = grp; sb < nb; sb += NSUBGROUPS) {
+        int tile_blk = rt * nb + sb;    // ne02 == 1 for lm_head/embed
+
+        // d + 16 scales for this (row, superblock)
+        float dval = (float)src0_d[tile_blk * TILE_ROWS + rit];
+        __global char * sc = src0_s + (tile_blk * TILE_ROWS + rit) * 16;
+
+        // 32 ql-uints (8 codes/uint) + 16 qh-uints (16 codes/uint)
+        uint ql[32];
+        uint qh[16];
+        #pragma unroll
+        for (int g = 0; g < 8; ++g) {
+            uint4 v = src0_ql[(tile_blk * 8 + g) * TILE_ROWS + rit];
+            ql[g*4+0] = v.x; ql[g*4+1] = v.y; ql[g*4+2] = v.z; ql[g*4+3] = v.w;
+        }
+        #pragma unroll
+        for (int g = 0; g < 4; ++g) {
+            uint4 v = src0_qh[(tile_blk * 4 + g) * TILE_ROWS + rit];
+            qh[g*4+0] = v.x; qh[g*4+1] = v.y; qh[g*4+2] = v.z; qh[g*4+3] = v.w;
+        }
+
+        // dequant 256 codes in canonical e-order, MAC with activation.
+        int act_base = sb * 64;         // activation float4 pixel base (256/4)
+        #pragma unroll
+        for (int e4 = 0; e4 < 64; ++e4) {
+            float4 a = read_imagef(src1, act_base + e4);
+            #pragma unroll
+            for (int t = 0; t < 4; ++t) {
+                int  e    = e4 * 4 + t;
+                uint low4 = (ql[e >> 3] >> ((e & 7) * 4)) & 0xF;
+                uint hi2  = (qh[e >> 4] >> ((e & 15) * 2)) & 0x3;
+                int  code = (int)(low4 | (hi2 << 4)) - 32;
+                int  sidx = ((e >> 7) << 3) + (((e >> 5) & 3) << 1) + ((e >> 4) & 1);
+                float scale = (float)sc[sidx] * dval;
+                float av = (t == 0) ? a.x : (t == 1) ? a.y : (t == 2) ? a.z : a.w;
+                acc += (float)code * scale * av;
+            }
+        }
+    }
+
+    // reduce across the NSUBGROUPS subgroups (same rit, different K-subset)
+    local float reduce_lm[NSUBGROUPS * TILE_ROWS];
+    reduce_lm[grp * TILE_ROWS + rit] = acc;
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (grp == 0) {
+        float total = reduce_lm[0 * TILE_ROWS + rit]
+                    + reduce_lm[1 * TILE_ROWS + rit]
+                    + reduce_lm[2 * TILE_ROWS + rit]
+                    + reduce_lm[3 * TILE_ROWS + rit];
+        dst = (global float*)((global char*)dst + offsetd);
+        dst[row] = total;
+    }
+}
+
+// Multi-column (N=3) variant of the tiled q6_K decode GEMV, for the speculative/
+// MTP VERIFY lm_head/embed (ne1=3 = 2 drafts + 1 bonus). Identical tiled weight
+// layout + unpack as the ne1=1 kernel above; each WI computes 3 output columns,
+// streaming the (large) lm_head weight ONCE per superblock and reusing it across
+// the 3 verify activation columns (dequant once per code, MAC into 3 accs). This
+// is the lm_head analogue of the per-layer mc3 GEMV; the multiply order matches
+// the ne1=1 kernel, so each column is byte-identical to a standalone tiled GEMV.
+#if defined(ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_gemv_noshuffle_q6_K_f32_tiled_mc3(
+    __global uint4 * src0_ql,
+    __global uint4 * src0_qh,
+    __global char  * src0_s,
+    __global half  * src0_d,
+    read_only image1d_buffer_t src1,
+    global float * dst,
+    ulong offsetd,
+    int ne00,
+    int ne01
+) {
+    int grp = get_local_id(1);
+    int row = get_global_id(0);
+    int rt  = row / TILE_ROWS;
+    int rit = row % TILE_ROWS;
+
+    int nb = ne00 / 256;
+    int col_stride = ne00 / 4;          // activation float4 pixels per column
+
+    float acc0 = 0.0f, acc1 = 0.0f, acc2 = 0.0f;
+
+    for (int sb = grp; sb < nb; sb += NSUBGROUPS) {
+        int tile_blk = rt * nb + sb;
+
+        float dval = (float)src0_d[tile_blk * TILE_ROWS + rit];
+        __global char * sc = src0_s + (tile_blk * TILE_ROWS + rit) * 16;
+
+        uint ql[32];
+        uint qh[16];
+        #pragma unroll
+        for (int g = 0; g < 8; ++g) {
+            uint4 v = src0_ql[(tile_blk * 8 + g) * TILE_ROWS + rit];
+            ql[g*4+0] = v.x; ql[g*4+1] = v.y; ql[g*4+2] = v.z; ql[g*4+3] = v.w;
+        }
+        #pragma unroll
+        for (int g = 0; g < 4; ++g) {
+            uint4 v = src0_qh[(tile_blk * 4 + g) * TILE_ROWS + rit];
+            qh[g*4+0] = v.x; qh[g*4+1] = v.y; qh[g*4+2] = v.z; qh[g*4+3] = v.w;
+        }
+
+        int act_base = sb * 64;
+        #pragma unroll
+        for (int e4 = 0; e4 < 64; ++e4) {
+            float4 a0 = read_imagef(src1, 0*col_stride + act_base + e4);
+            float4 a1 = read_imagef(src1, 1*col_stride + act_base + e4);
+            float4 a2 = read_imagef(src1, 2*col_stride + act_base + e4);
+            #pragma unroll
+            for (int t = 0; t < 4; ++t) {
+                int  e    = e4 * 4 + t;
+                uint low4 = (ql[e >> 3] >> ((e & 7) * 4)) & 0xF;
+                uint hi2  = (qh[e >> 4] >> ((e & 15) * 2)) & 0x3;
+                int  code = (int)(low4 | (hi2 << 4)) - 32;
+                int  sidx = ((e >> 7) << 3) + (((e >> 5) & 3) << 1) + ((e >> 4) & 1);
+                float w   = (float)code * ((float)sc[sidx] * dval);  // dequant+scale once
+                float av0 = (t == 0) ? a0.x : (t == 1) ? a0.y : (t == 2) ? a0.z : a0.w;
+                float av1 = (t == 0) ? a1.x : (t == 1) ? a1.y : (t == 2) ? a1.z : a1.w;
+                float av2 = (t == 0) ? a2.x : (t == 1) ? a2.y : (t == 2) ? a2.z : a2.w;
+                acc0 += w * av0;
+                acc1 += w * av1;
+                acc2 += w * av2;
+            }
+        }
+    }
+
+    local float4 reduce_lm[NSUBGROUPS * TILE_ROWS];
+    reduce_lm[grp * TILE_ROWS + rit] = (float4)(acc0, acc1, acc2, 0.0f);
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (grp == 0) {
+        float4 total = reduce_lm[0 * TILE_ROWS + rit]
+                     + reduce_lm[1 * TILE_ROWS + rit]
+                     + reduce_lm[2 * TILE_ROWS + rit]
+                     + reduce_lm[3 * TILE_ROWS + rit];
+        dst = (global float*)((global char*)dst + offsetd);
+        // dst column-major [ne01 rows x 3 cols]: (row, col) at col*ne01 + row
+        dst[0*ne01 + row] = total.x;
+        dst[1*ne01 + row] = total.y;
+        dst[2*ne01 + row] = total.z;
+    }
+}
@@ -118,6 +118,87 @@
    elem = (char)((bits8.s7 & 0xFF000000) >> 24); \
    total_sums += convert_int(elem) * scale * shared_y; \

+// ============================================================================
+// Split-K variant for small-M decode GEMVs.
+// ----------------------------------------------------------------------------
+// The base kernel below puts one output row per lane and splits K only across
+// the N_SIMDGROUP subgroups of a single workgroup, so M=512 yields M/64 = 8
+// workgroups -- half the compute units on a 16-CU X2 sit idle, and the kernel
+// measures ~48 GB/s against the ~122 GB/s the larger projections reach in the
+// same graph. Here each (kslice, subgroup) pair reduces a disjoint set of
+// K-blocks into partial[kslice * M + row]; kernel_gemv_splitk_reduce_f32 (in
+// gemv_noshuffle_q4_k_f32.cl) sums the slices. Same operand order within a
+// slice as the base kernel; only the cross-slice grouping differs.
+//
+// Placed BEFORE the base kernel deliberately: on A6X no kernel may be defined
+// after one that uses a subgroup builtin, or it silently miscompiles.
+// ============================================================================
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+__kernel void kernel_gemv_noshuffle_q8_0_f32_splitk(
+        __read_only  image1d_buffer_t src0_q,   // quantized A (weights)
+        global half  * src0_d,                  // A scales
+        __read_only  image1d_buffer_t src1,     // B (activations)
+        global float * partial,                 // [ksplit * M], slice-major
+        int ne00,                               // K
+        int ne01)                               // M
+{
+    uint groupId = get_local_id(1);
+    uint gid     = get_global_id(0);
+    ushort slid  = get_sub_group_local_id();
+    uint nsg     = get_local_size(1);
+    uint ksplit  = get_num_groups(1);
+    uint kslice  = get_group_id(1);
+
+    uint K = ne00;
+    uint M = ne01;
+
+    uint LINE_STRIDE_A  = M;
+    uint BLOCK_STRIDE_A = 8 * M;   // physical, independent of the K-split
+
+    __private uint8  regA;
+    __private half   regS;
+    __private float8 regB;
+    __private float  totalSum = (float)(0.0f);
+
+    #pragma unroll 1
+    for (uint k = kslice * nsg + groupId; k < (K / QK8_0); k += ksplit * nsg) {
+        regS = src0_d[gid + k * LINE_STRIDE_A];
+        if (slid < 4) {
+            regB.s0123 = read_imagef(src1, (slid * 2 + k * 8));
+            regB.s4567 = read_imagef(src1, (1 + slid * 2 + k * 8));
+        }
+        regA.s0 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 0)).x;
+        regA.s1 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 1)).x;
+        regA.s2 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 2)).x;
+        regA.s3 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 3)).x;
+        regA.s4 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 4)).x;
+        regA.s5 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 5)).x;
+        regA.s6 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 6)).x;
+        regA.s7 = read_imageui(src0_q, (gid + k * BLOCK_STRIDE_A + LINE_STRIDE_A * 7)).x;
+
+        dequantizeBlockAccum_ns_sgbroadcast_1(totalSum, regA, convert_float(regS), regB);
+    }
+
+    // Intra-workgroup reduce across this K-slice's subgroups. Sized for
+    // nsg <= 8; the host never dispatches more.
+    __local float reduceLM[SIMDGROUP_WIDTH * 7];
+    if (groupId > 0) {
+        reduceLM[SIMDGROUP_WIDTH * (groupId - 1) + slid] = totalSum;
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+    if (groupId == 0) {
+        for (uint i = 0; i < nsg - 1; ++i) {
+            totalSum += reduceLM[SIMDGROUP_WIDTH * i + slid];
+        }
+        // x-grid is padded to CEIL_DIV(M,wave)*wave; guard the tail rows.
+        if (gid < M) {
+            partial[kslice * M + gid] = totalSum;
+        }
+    }
+}
+
 #ifdef ADRENO_GPU
 REQD_SUBGROUP_SIZE_64
 #endif
@@ -145,3 +145,52 @@ kernel void kernel_mul_mm_f32_f32_l4_lm(
        }
    }
 }
+
+// Multi-column f32 GEMV for the small-N (spec/MTP verify) batch. The tiled GEMM
+// above always computes a full BM x BN = 64 x 64 output tile, so at ne11=3 with a
+// skinny weight (e.g. GDN ssm_alpha/ssm_beta, M=32) it launches ONE under-occupied
+// workgroup at ~2.3% tile utilization. This kernel assigns one 64-thread workgroup
+// per output element (m,n): the 64 threads split the K reduction (float4) and
+// tree-reduce in __local (no subgroup ops -> portable). ne01*ne11 workgroups.
+// Weight row is re-read per column (N small -> negligible). Summation order differs
+// from the tiled GEMM (lane-strided + tree) -> f32-exact-ish, not bit-identical.
+kernel void kernel_gemv_f32_f32_mc(
+    global float * src0, ulong offset0,   // weight: row m at m*stride_a (elements)
+    global float * src1, ulong offset1,   // activations: col n at n*stride_b
+    global float * dst,  ulong offsetd,   // dst [M x N] col-major: (m,n) at n*stride_d+m
+    int ne00,        // K
+    int ne01,        // M
+    int ne11,        // N
+    int stride_a,    // weight row stride (elements) = K
+    int stride_b,    // activation col stride (elements) = K
+    int stride_d)    // dst column stride (elements) = M
+{
+    src0 = (global float*)((global char*)src0 + offset0);
+    src1 = (global float*)((global char*)src1 + offset1);
+    dst  = (global float*)((global char*)dst  + offsetd);
+
+    uint lane = get_local_id(0);            // 0..63
+    uint out  = get_global_id(1);           // 0 .. ne01*ne11 - 1
+    uint m = out % (uint)ne01;
+    uint n = out / (uint)ne01;
+
+    global float4 * wrow = (global float4*)(src0 + (ulong)m * (uint)stride_a);
+    global float4 * xcol = (global float4*)(src1 + (ulong)n * (uint)stride_b);
+    uint k4 = (uint)ne00 >> 2;
+
+    float acc = 0.0f;
+    for (uint k = lane; k < k4; k += 64) {
+        float4 w = wrow[k];
+        float4 x = xcol[k];
+        acc += w.s0*x.s0 + w.s1*x.s1 + w.s2*x.s2 + w.s3*x.s3;
+    }
+
+    local float red[64];
+    red[lane] = acc;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    for (uint s = 32; s > 0; s >>= 1) {
+        if (lane < s) red[lane] += red[lane + s];
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+    if (lane == 0) dst[(ulong)n * (uint)stride_d + m] = red[0];
+}
@@ -0,0 +1,306 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_subgroups
+#pragma OPENCL EXTENSION cl_intel_subgroups : enable
+#else
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#endif
+
+#ifdef cl_intel_required_subgroup_size
+#pragma OPENCL EXTENSION cl_intel_required_subgroup_size : enable
+#define INTEL_GPU 1
+#define REQD_SUBGROUP_SIZE_16 __attribute__((intel_reqd_sub_group_size(16)))
+#define REQD_SUBGROUP_SIZE_32 __attribute__((intel_reqd_sub_group_size(32)))
+#elif defined(cl_qcom_reqd_sub_group_size)
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+#define ADRENO_GPU 1
+#define REQD_SUBGROUP_SIZE_64  __attribute__((qcom_reqd_sub_group_size("half")))
+#define REQD_SUBGROUP_SIZE_128 __attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+
+// Multi-row f16xf32 GEMV for the DECODE path (single token, ne11*ne12 small).
+// The legacy kernel_mul_mat_f16_f32_1row runs ONE 64-lane subgroup per workgroup =
+// one output row per WG, which caps memory-level parallelism at roughly half of
+// LPDDR5x peak. This variant packs MROW subgroups per workgroup, each
+// computing a distinct output row, so a WG keeps 64*MROW loads in flight. The
+// activation column y (shared by every output row) is staged into __local ONCE per
+// WG and reused across the MROW rows, cutting redundant activation reads. Used for
+// the f16 attention projections (Q/K/V/O) and lm_head, which dominate decode.
+// Numerically equivalent to _1row (same f16->f32 widening, same float4 partial sums,
+// same subgroup-reduce order), so byte-identical to the per-op path.
+
+#define MROW 16
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f32_mrow(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        ulong nb00,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        ulong nb10,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3,
+        __local float * ysh
+) {
+    src0 = (global char*)((global char*)src0 + offset0);
+    src1 = (global char*)((global char*)src1 + offset1);
+    dst  = (global float*)((global char*)dst + offsetd);
+
+    int r0   = get_group_id(0) * MROW + get_local_id(1);  // output row
+    int r1   = get_group_id(1);                            // token (ne11)
+    int im   = get_group_id(2);
+    int lid  = get_sub_group_local_id();                   // 0..63
+    int nsg  = get_local_size(1);                          // == MROW
+
+    int i12 = im % ne12;
+    int i13 = im / ne12;
+
+    ulong offset_src1 = r1*nb11 + (i12)*nb12 + (i13)*nb13;
+    global float * y = (global float *) (src1 + offset_src1);
+
+    // Cooperatively stage the activation column (ne00 floats) into __local once per
+    // WG and reuse across the MROW rows. Staging is the actual win here: dropping it
+    // (each subgroup re-reading y from global) regresses below the 1-row kernel.
+    for (int i = get_local_id(1)*get_sub_group_size() + lid; i < ne00; i += nsg*get_sub_group_size()) {
+        ysh[i] = y[i];
+    }
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (r0 >= ne01) {
+        return;
+    }
+
+    ulong offset_src0 = r0*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+    global half * x = (global half *) (src0 + offset_src0);
+
+    // The vector path below casts the row pointer to half4, which must be 8-byte aligned.
+    // A row address is r0*nb01 + ..., and a permuted or strided src0 leaves nb01/nb02/nb03
+    // unconstrained -- ne00 % 4 == 0 bounds the element count per row, not the byte stride
+    // between rows. Take the vector path only when this work-item's row is actually
+    // aligned; the scalar loop below has no such requirement.
+    const bool row_aligned = (((ulong) x) & 7) == 0;
+
+    float sumf = 0.0f;
+    if (ne00 < 128 || !row_aligned) {
+        for (int i = lid; i < ne00; i += get_sub_group_size()) {
+            sumf += (float) x[i] * ysh[i];
+        }
+        float all_sum = sub_group_reduce_add(sumf);
+        if (lid == 0) {
+            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+        }
+    } else {
+        global half4 * x4 = (global half4 *) x;
+        __local float4 * ysh4 = (__local float4 *) ysh;
+        for (int i = lid; i < ne00/4; i += get_sub_group_size()) {
+            float4 yv = ysh4[i];
+            sumf += (float) x4[i].s0 * yv.s0;
+            sumf += (float) x4[i].s1 * yv.s1;
+            sumf += (float) x4[i].s2 * yv.s2;
+            sumf += (float) x4[i].s3 * yv.s3;
+        }
+        float all_sum = sub_group_reduce_add(sumf);
+        if (lid == 0) {
+            for (int i = 4*(ne00/4); i < ne00; ++i) {
+                all_sum += (float) x[i] * ysh[i];
+            }
+            dst[im*ne1*ne0 + r1*ne0 + r0] = all_sum;
+        }
+    }
+}
+
+// Register-blocked variant: each 64-lane subgroup accumulates RPT consecutive
+// output rows instead of one. The staged activation is reused across all RPT rows,
+// and each lane keeps RPT independent weight loads in flight per column step ->
+// more memory-level parallelism on the streaming f16 weight read (the BW limiter),
+// plus RPT fewer staging barriers per output row. Per-row reduction order is
+// identical to _mrow, so byte-identical to the per-op path. Dispatch guarantees
+// ne00 >= 128 and ne00 % 4 == 0, so only the half4 path is needed (no tail).
+#define MROW_RB_BODY(RPT)                                                                 \
+    src0 = (global char*)((global char*)src0 + offset0);                                  \
+    src1 = (global char*)((global char*)src1 + offset1);                                  \
+    dst  = (global float*)((global char*)dst + offsetd);                                  \
+    int r0b = (get_group_id(0) * get_local_size(1) + get_local_id(1)) * (RPT);            \
+    int r1  = get_group_id(1);                                                            \
+    int im  = get_group_id(2);                                                            \
+    int lid = get_sub_group_local_id();                                                   \
+    int nsg = get_local_size(1);                                                          \
+    int i12 = im % ne12;                                                                  \
+    int i13 = im / ne12;                                                                  \
+    ulong off_y = r1*nb11 + i12*nb12 + i13*nb13;                                          \
+    global float * y = (global float *) (src1 + off_y);                                   \
+    for (int i = get_local_id(1)*get_sub_group_size() + lid; i < ne00;                    \
+         i += nsg*get_sub_group_size()) {                                                 \
+        ysh[i] = y[i];                                                                    \
+    }                                                                                     \
+    barrier(CLK_LOCAL_MEM_FENCE);                                                         \
+    __local float4 * ysh4 = (__local float4 *) ysh;                                       \
+    global half4 * xr[RPT];                                                               \
+    _Pragma("unroll")                                                                     \
+    for (int rr = 0; rr < (RPT); ++rr) {                                                  \
+        int row = r0b + rr;                                                               \
+        if (row > ne01 - 1) row = ne01 - 1;                                               \
+        xr[rr] = (global half4 *) (src0 + (ulong)row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03); \
+    }                                                                                     \
+    float sumf[RPT];                                                                      \
+    _Pragma("unroll")                                                                     \
+    for (int rr = 0; rr < (RPT); ++rr) sumf[rr] = 0.0f;                                   \
+    for (int i = lid; i < ne00/4; i += get_sub_group_size()) {                            \
+        float4 yv = ysh4[i];                                                              \
+        _Pragma("unroll")                                                                 \
+        for (int rr = 0; rr < (RPT); ++rr) {                                              \
+            half4 xv = xr[rr][i];                                                         \
+            sumf[rr] += (float) xv.s0 * yv.s0 + (float) xv.s1 * yv.s1                      \
+                      + (float) xv.s2 * yv.s2 + (float) xv.s3 * yv.s3;                     \
+        }                                                                                 \
+    }                                                                                     \
+    _Pragma("unroll")                                                                     \
+    for (int rr = 0; rr < (RPT); ++rr) {                                                  \
+        float s = sub_group_reduce_add(sumf[rr]);                                         \
+        int row = r0b + rr;                                                               \
+        if (lid == 0 && row < ne01) {                                                     \
+            dst[im*ne1*ne0 + r1*ne0 + row] = s;                                           \
+        }                                                                                 \
+    }
+
+// half8 (128-bit) load variant: Adreno's load/store unit issues 128-bit
+// transactions, so half4 (64-bit) loads may leave the load path half-idle. This
+// processes 8 weight elements per lane per step via half8. Accumulation groups
+// elements in 8s rather than 4s, so it is NOT bit-identical to _1row (float add is
+// non-associative) -- experimental BW probe, gate on ne00 % 8 == 0.
+#define MROW_H8_BODY(RPT)                                                                 \
+    src0 = (global char*)((global char*)src0 + offset0);                                  \
+    src1 = (global char*)((global char*)src1 + offset1);                                  \
+    dst  = (global float*)((global char*)dst + offsetd);                                  \
+    int r0b = (get_group_id(0) * get_local_size(1) + get_local_id(1)) * (RPT);            \
+    int r1  = get_group_id(1);                                                            \
+    int im  = get_group_id(2);                                                            \
+    int lid = get_sub_group_local_id();                                                   \
+    int nsg = get_local_size(1);                                                          \
+    int i12 = im % ne12;                                                                  \
+    int i13 = im / ne12;                                                                  \
+    ulong off_y = r1*nb11 + i12*nb12 + i13*nb13;                                          \
+    global float * y = (global float *) (src1 + off_y);                                   \
+    for (int i = get_local_id(1)*get_sub_group_size() + lid; i < ne00;                    \
+         i += nsg*get_sub_group_size()) {                                                 \
+        ysh[i] = y[i];                                                                    \
+    }                                                                                     \
+    barrier(CLK_LOCAL_MEM_FENCE);                                                         \
+    __local float4 * ysh4 = (__local float4 *) ysh;                                       \
+    global half8 * xr[RPT];                                                               \
+    _Pragma("unroll")                                                                     \
+    for (int rr = 0; rr < (RPT); ++rr) {                                                  \
+        int row = r0b + rr;                                                               \
+        if (row > ne01 - 1) row = ne01 - 1;                                               \
+        xr[rr] = (global half8 *) (src0 + (ulong)row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03); \
+    }                                                                                     \
+    float sumf[RPT];                                                                      \
+    _Pragma("unroll")                                                                     \
+    for (int rr = 0; rr < (RPT); ++rr) sumf[rr] = 0.0f;                                   \
+    for (int i = lid; i < ne00/8; i += get_sub_group_size()) {                            \
+        float4 y0 = ysh4[2*i];                                                            \
+        float4 y1 = ysh4[2*i + 1];                                                        \
+        _Pragma("unroll")                                                                 \
+        for (int rr = 0; rr < (RPT); ++rr) {                                              \
+            half8 xv = xr[rr][i];                                                         \
+            sumf[rr] += (float) xv.s0 * y0.s0 + (float) xv.s1 * y0.s1                      \
+                      + (float) xv.s2 * y0.s2 + (float) xv.s3 * y0.s3                      \
+                      + (float) xv.s4 * y1.s0 + (float) xv.s5 * y1.s1                      \
+                      + (float) xv.s6 * y1.s2 + (float) xv.s7 * y1.s3;                     \
+        }                                                                                 \
+    }                                                                                     \
+    _Pragma("unroll")                                                                     \
+    for (int rr = 0; rr < (RPT); ++rr) {                                                  \
+        float s = sub_group_reduce_add(sumf[rr]);                                         \
+        int row = r0b + rr;                                                               \
+        if (lid == 0 && row < ne01) {                                                     \
+            dst[im*ne1*ne0 + r1*ne0 + row] = s;                                           \
+        }                                                                                 \
+    }
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f32_mrow_h8(
+        global char * src0, ulong offset0,
+        global char * src1, ulong offset1,
+        global float * dst, ulong offsetd,
+        int ne00, int ne01, int ne02,
+        ulong nb00, ulong nb01, ulong nb02, ulong nb03,
+        int ne10, int ne11, int ne12,
+        ulong nb10, ulong nb11, ulong nb12, ulong nb13,
+        int ne0, int ne1, int r2, int r3,
+        __local float * ysh
+) {
+    MROW_H8_BODY(1)
+}
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f32_mrow_h8r2(
+        global char * src0, ulong offset0,
+        global char * src1, ulong offset1,
+        global float * dst, ulong offsetd,
+        int ne00, int ne01, int ne02,
+        ulong nb00, ulong nb01, ulong nb02, ulong nb03,
+        int ne10, int ne11, int ne12,
+        ulong nb10, ulong nb11, ulong nb12, ulong nb13,
+        int ne0, int ne1, int r2, int r3,
+        __local float * ysh
+) {
+    MROW_H8_BODY(2)
+}
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f32_mrow_r2(
+        global char * src0, ulong offset0,
+        global char * src1, ulong offset1,
+        global float * dst, ulong offsetd,
+        int ne00, int ne01, int ne02,
+        ulong nb00, ulong nb01, ulong nb02, ulong nb03,
+        int ne10, int ne11, int ne12,
+        ulong nb10, ulong nb11, ulong nb12, ulong nb13,
+        int ne0, int ne1, int r2, int r3,
+        __local float * ysh
+) {
+    MROW_RB_BODY(2)
+}
+
+#ifdef ADRENO_GPU
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mat_f16_f32_mrow_r4(
+        global char * src0, ulong offset0,
+        global char * src1, ulong offset1,
+        global float * dst, ulong offsetd,
+        int ne00, int ne01, int ne02,
+        ulong nb00, ulong nb01, ulong nb02, ulong nb03,
+        int ne10, int ne11, int ne12,
+        ulong nb10, ulong nb11, ulong nb12, ulong nb13,
+        int ne0, int ne1, int r2, int r3,
+        __local float * ysh
+) {
+    MROW_RB_BODY(4)
+}
@@ -188,3 +188,182 @@ kernel void kernel_rms_norm_mul(
        y[i00] = (x[i00] * scale) * f[i00%(ne10/4)];
    }
 }
+
+//------------------------------------------------------------------------------
+// rms_norm + mul (norm weight) + add (residual), fused. Mirrors
+// kernel_rms_norm_mul with an extra residual operand src2: computes
+//   y = (rmsnorm(x) * w) + g
+// in one dispatch, removing one kernel launch + one global round-trip per
+// residual block (the dominant per-layer adjacency on Gemma matformers).
+//------------------------------------------------------------------------------
+kernel void kernel_rms_norm_mul_add(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * src2,
+        ulong offset2,
+        global char * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne20,
+        int ne21,
+        int ne22,
+        int ne23,
+        ulong nb21,
+        ulong nb22,
+        ulong nb23,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        float eps,
+        local float * sum
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    src2 = src2 + offset2;
+    dst  = dst  + offsetd;
+
+    if (get_sub_group_id() == 0) {
+        sum[get_sub_group_local_id()] = 0.0f;
+    }
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    global float4 * x = (global float4 *) (src0 + i03*nb03 + i02*nb02 + i01*nb01);
+    global float4 * f = (global float4 *) (src1 + (i03%ne13)*nb13 + (i02%ne12)*nb12 + (i01%ne11)*nb11);
+    global float4 * g = (global float4 *) (src2 + (i03%ne23)*nb23 + (i02%ne22)*nb22 + (i01%ne21)*nb21);
+
+    float sumf = 0;
+
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        sumf += dot(x[i00], x[i00]);
+    }
+    sumf = sub_group_reduce_add(sumf);
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (get_sub_group_local_id() == 0) {
+        sum[get_sub_group_id()] = sumf;
+    }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    sumf = sum[get_sub_group_local_id()];
+    sumf = sub_group_reduce_add(sumf);
+
+    float mean  = sumf / ne00;
+    float scale = 1.0f/sqrt(mean + eps);
+
+    global float4 * y = (global float4 *) (dst + i03*nb3 + i02*nb2 + i01*nb1);
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        y[i00] = (x[i00] * scale) * f[i00%(ne10/4)] + g[i00%(ne20/4)];
+    }
+}
+
+//------------------------------------------------------------------------------
+// rms_norm + mul(norm weight) + add(residual) + mul(scalar scale), fused.
+// Computes y = ((rmsnorm(x) * w) + g) * s, where s is a broadcast SCALAR (e.g.
+// Gemma-4 layer_output_scale). Folds the trailing per-layer l_out scale-mul into
+// the residual-norm kernel: one extra dispatch + global round-trip saved per
+// layer. src3 points at the single scale value.
+//------------------------------------------------------------------------------
+kernel void kernel_rms_norm_mul_add_scale(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * src2,
+        ulong offset2,
+        global char * src3,
+        ulong offset3,
+        global char * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne02,
+        int ne03,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne10,
+        int ne11,
+        int ne12,
+        int ne13,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne20,
+        int ne21,
+        int ne22,
+        int ne23,
+        ulong nb21,
+        ulong nb22,
+        ulong nb23,
+        ulong nb1,
+        ulong nb2,
+        ulong nb3,
+        float eps,
+        local float * sum
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    src2 = src2 + offset2;
+    src3 = src3 + offset3;
+    dst  = dst  + offsetd;
+
+    const float sc = *((global float *) src3);
+
+    if (get_sub_group_id() == 0) {
+        sum[get_sub_group_local_id()] = 0.0f;
+    }
+
+    int i03 = get_group_id(2);
+    int i02 = get_group_id(1);
+    int i01 = get_group_id(0);
+
+    global float4 * x = (global float4 *) (src0 + i03*nb03 + i02*nb02 + i01*nb01);
+    global float4 * f = (global float4 *) (src1 + (i03%ne13)*nb13 + (i02%ne12)*nb12 + (i01%ne11)*nb11);
+    global float4 * g = (global float4 *) (src2 + (i03%ne23)*nb23 + (i02%ne22)*nb22 + (i01%ne21)*nb21);
+
+    float sumf = 0;
+
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        sumf += dot(x[i00], x[i00]);
+    }
+    sumf = sub_group_reduce_add(sumf);
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    if (get_sub_group_local_id() == 0) {
+        sum[get_sub_group_id()] = sumf;
+    }
+
+    barrier(CLK_LOCAL_MEM_FENCE);
+
+    sumf = sum[get_sub_group_local_id()];
+    sumf = sub_group_reduce_add(sumf);
+
+    float mean  = sumf / ne00;
+    float scale = 1.0f/sqrt(mean + eps);
+
+    global float4 * y = (global float4 *) (dst + i03*nb3 + i02*nb2 + i01*nb1);
+    for (int i00 = get_local_id(0); i00 < ne00/4; i00 += get_local_size(0)) {
+        y[i00] = ((x[i00] * scale) * f[i00%(ne10/4)] + g[i00%(ne20/4)]) * sc;
+    }
+}
@@ -835,7 +835,7 @@ static size_t ggml_backend_rpc_buffer_type_get_alloc_size(ggml_backend_buffer_ty
    // [TAG_ALLOC_SIZE_EXPAND]
    // ops that may require additional memory for fleeting data on certain backends
    // ref: https://github.com/ggml-org/llama.cpp/pull/15966
-    rpc_get |= ggml_backend_op_alloc_size_may_expand(tensor->op);
+    rpc_get |= ggml_op_alloc_size_may_expand(tensor->op);

    if (rpc_get) {
        ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
--- a/Show More
+++ b/Show More