webgpu: address review @yomaytk, name the mat-vec row threshold

The flash attention vec path and the MMVQ path append a scratch area
after dst, but get_alloc_size only accounted for it when the current
batch takes that path. The size was therefore not monotonic in the
batch length: a graph reserved at n_ubatch (tiled path, no scratch)
or at one token did not cover batches of 2 to 19 query rows, and
GGML_SCHED_NO_REALLOC builds aborted on an unexpected reallocation.

Pass the row count to both path predicates and size the scratch with
the row count clamped to the longest batch the short path serves.

.github/workflows/behavior-check.yml

@@ -13,6 +13,8 @@ on:
       '**/*.hpp',
       '**/*.c',
       '**/*.cpp',
+      '**/*.py',
+      '**/requirements*.txt',
     ]
 
   pull_request:
@@ -24,7 +26,9 @@ on:
       '**/*.h',
       '**/*.hpp',
       '**/*.c',
-      '**/*.cpp'
+      '**/*.cpp',
+      '**/*.py',
+      '**/requirements*.txt',
     ]
 
 cache-mode: none
@@ -69,7 +73,7 @@ jobs:
         run: |
           python3 -m venv .venv
           source .venv/bin/activate
-          pip install jinja2==3.1.6
+          pip install -r requirements/requirements-convert_hf_to_gguf.txt
 
       - name: ccache-buckets-restore
         uses: ./.github/actions/ccache-buckets
@@ -134,7 +138,7 @@ jobs:
         run: |
           python3 -m venv .venv
           source .venv/Scripts/activate
-          pip install jinja2==3.1.6
+          pip install -r requirements/requirements-convert_hf_to_gguf.txt
 
       - name: Build
         id: cmake_build
@@ -156,3 +160,13 @@ jobs:
           source .venv/Scripts/activate
           cd build
           ctest -C Release -L behavior -LE 'model|backend' --verbose --timeout 900
+
+      - name: ccache-clear
+        uses: ./.github/actions/ccache-clear
+        env:
+          GH_TOKEN: ${{ github.token }}
+        with:
+          key: behavior-check-windows-2025
+          older: 5m
+          min: 1
+          dry-run: ${{ github.event_name != 'push' || github.ref != 'refs/heads/master' }}

.github/workflows/build-apple-daily.yml

@@ -0,0 +1,160 @@
+name: CI (apple - daily)
+
+on:
+  workflow_dispatch: # allows manual triggering
+  schedule:
+    - cron: '0 0 * * *'
+
+cache-mode: none
+permissions:
+  contents: read
+
+concurrency:
+  group: ${{ github.workflow }}-${{ github.head_ref && github.ref || github.run_id }}
+  cancel-in-progress: true
+
+env:
+  GGML_NLOOP: 3
+  GGML_N_THREADS: 1
+  LLAMA_ARG_LOG_COLORS: 1
+  LLAMA_ARG_LOG_PREFIX: 1
+  LLAMA_ARG_LOG_TIMESTAMPS: 1
+
+jobs:
+  macos-latest-ios-xcode:
+    runs-on: macos-latest
+
+    permissions:
+      actions: write
+      contents: read
+
+    steps:
+      - name: Checkout code
+        uses: actions/checkout@v6
+
+      - name: Setup Xcode
+        uses: ggml-org/setup-xcode@v1
+        with:
+          xcode-version: latest-stable
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          cmake -B build -G Xcode \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DLLAMA_OPENSSL=OFF \
+            -DLLAMA_BUILD_APP=OFF \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TOOLS=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_SYSTEM_NAME=iOS \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
+            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
+          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
+
+      - name: xcodebuild for swift package
+        id: xcodebuild
+        run: |
+          ./build-xcframework.sh
+
+      - name: Upload xcframework artifact
+        uses: actions/upload-artifact@v6
+        with:
+          name: llama-xcframework
+          path: build-apple/llama.xcframework/
+          retention-days: 1
+
+      - name: Build Xcode project
+        run: |
+          xcodebuild -downloadPlatform iOS
+          xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' FRAMEWORK_FOLDER_PATH=./build-ios build
+
+  macos-latest-tvos:
+    runs-on: macos-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          cmake -B build -G Xcode \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DLLAMA_BUILD_COMMON=OFF \
+            -DLLAMA_BUILD_APP=OFF \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TOOLS=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_SYSTEM_NAME=tvOS \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
+            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
+          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
+
+  macos-latest-visionos:
+    runs-on: macos-latest
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Build
+        id: cmake_build
+        run: |
+          sysctl -a
+          cmake -B build -G Xcode \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DLLAMA_BUILD_COMMON=OFF \
+            -DLLAMA_BUILD_APP=OFF \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TOOLS=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_SYSTEM_NAME=visionOS \
+            -DCMAKE_OSX_DEPLOYMENT_TARGET=1.0 \
+            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
+          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
+
+  macos-latest-swift:
+    runs-on: macos-latest
+    needs: macos-latest-ios-xcode
+
+    permissions:
+      actions: read
+      contents: read
+
+    strategy:
+      matrix:
+        destination: ['generic/platform=macOS', 'generic/platform=iOS', 'generic/platform=tvOS']
+
+    steps:
+      - name: Clone
+        id: checkout
+        uses: actions/checkout@v6
+
+      - name: Download xcframework artifact
+        uses: actions/download-artifact@v7
+        with:
+          name: llama-xcframework
+          path: build-apple/llama.xcframework/
+
+      - name: Build llama.cpp with CMake
+        id: cmake_build
+        run: |
+          sysctl -a
+          cmake -B build -G Xcode \
+            -DGGML_METAL_EMBED_LIBRARY=ON \
+            -DLLAMA_OPENSSL=OFF \
+            -DLLAMA_BUILD_APP=OFF \
+            -DLLAMA_BUILD_EXAMPLES=OFF \
+            -DLLAMA_BUILD_TOOLS=OFF \
+            -DLLAMA_BUILD_TESTS=OFF \
+            -DLLAMA_BUILD_SERVER=OFF \
+            -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64"
+          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)

.github/workflows/build-apple.yml

@@ -107,140 +107,3 @@ jobs:
           # ref: https://github.com/ggml-org/llama.cpp/pull/19802#issuecomment-4013704023
           ctest -L main -E "test-llama-archs|test-llama-context|test-speculative-rejection-sampling" --verbose --timeout 900
 
-  macos-latest-ios-xcode:
-    runs-on: macos-latest
-
-    permissions:
-      actions: write
-      contents: read
-
-    steps:
-      - name: Checkout code
-        uses: actions/checkout@v6
-
-      - name: Setup Xcode
-        uses: ggml-org/setup-xcode@v1
-        with:
-          xcode-version: latest-stable
-
-      - name: Build
-        id: cmake_build
-        run: |
-          sysctl -a
-          cmake -B build -G Xcode \
-            -DGGML_METAL_EMBED_LIBRARY=ON \
-            -DLLAMA_OPENSSL=OFF \
-            -DLLAMA_BUILD_APP=OFF \
-            -DLLAMA_BUILD_EXAMPLES=OFF \
-            -DLLAMA_BUILD_TOOLS=OFF \
-            -DLLAMA_BUILD_TESTS=OFF \
-            -DLLAMA_BUILD_SERVER=OFF \
-            -DCMAKE_SYSTEM_NAME=iOS \
-            -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
-            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
-          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
-
-      - name: xcodebuild for swift package
-        id: xcodebuild
-        run: |
-          ./build-xcframework.sh
-
-      - name: Upload xcframework artifact
-        uses: actions/upload-artifact@v6
-        with:
-          name: llama-xcframework
-          path: build-apple/llama.xcframework/
-          retention-days: 1
-
-      - name: Build Xcode project
-        run: |
-          xcodebuild -downloadPlatform iOS
-          xcodebuild -project examples/llama.swiftui/llama.swiftui.xcodeproj -scheme llama.swiftui -sdk iphoneos CODE_SIGNING_REQUIRED=NO CODE_SIGN_IDENTITY= -destination 'generic/platform=iOS' FRAMEWORK_FOLDER_PATH=./build-ios build
-
-  macos-latest-tvos:
-    runs-on: macos-latest
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Build
-        id: cmake_build
-        run: |
-          sysctl -a
-          cmake -B build -G Xcode \
-            -DGGML_METAL_EMBED_LIBRARY=ON \
-            -DLLAMA_BUILD_COMMON=OFF \
-            -DLLAMA_BUILD_APP=OFF \
-            -DLLAMA_BUILD_EXAMPLES=OFF \
-            -DLLAMA_BUILD_TOOLS=OFF \
-            -DLLAMA_BUILD_TESTS=OFF \
-            -DLLAMA_BUILD_SERVER=OFF \
-            -DCMAKE_SYSTEM_NAME=tvOS \
-            -DCMAKE_OSX_DEPLOYMENT_TARGET=14.0 \
-            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
-          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
-
-  macos-latest-visionos:
-    runs-on: macos-latest
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Build
-        id: cmake_build
-        run: |
-          sysctl -a
-          cmake -B build -G Xcode \
-            -DGGML_METAL_EMBED_LIBRARY=ON \
-            -DLLAMA_BUILD_COMMON=OFF \
-            -DLLAMA_BUILD_APP=OFF \
-            -DLLAMA_BUILD_EXAMPLES=OFF \
-            -DLLAMA_BUILD_TOOLS=OFF \
-            -DLLAMA_BUILD_TESTS=OFF \
-            -DLLAMA_BUILD_SERVER=OFF \
-            -DCMAKE_SYSTEM_NAME=visionOS \
-            -DCMAKE_OSX_DEPLOYMENT_TARGET=1.0 \
-            -DCMAKE_XCODE_ATTRIBUTE_DEVELOPMENT_TEAM=ggml
-          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu) -- CODE_SIGNING_ALLOWED=NO
-
-  macos-latest-swift:
-    runs-on: macos-latest
-    needs: macos-latest-ios-xcode
-
-    permissions:
-      actions: read
-      contents: read
-
-    strategy:
-      matrix:
-        destination: ['generic/platform=macOS', 'generic/platform=iOS', 'generic/platform=tvOS']
-
-    steps:
-      - name: Clone
-        id: checkout
-        uses: actions/checkout@v6
-
-      - name: Download xcframework artifact
-        uses: actions/download-artifact@v7
-        with:
-          name: llama-xcframework
-          path: build-apple/llama.xcframework/
-
-      - name: Build llama.cpp with CMake
-        id: cmake_build
-        run: |
-          sysctl -a
-          cmake -B build -G Xcode \
-            -DGGML_METAL_EMBED_LIBRARY=ON \
-            -DLLAMA_OPENSSL=OFF \
-            -DLLAMA_BUILD_APP=OFF \
-            -DLLAMA_BUILD_EXAMPLES=OFF \
-            -DLLAMA_BUILD_TOOLS=OFF \
-            -DLLAMA_BUILD_TESTS=OFF \
-            -DLLAMA_BUILD_SERVER=OFF \
-            -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64"
-          cmake --build build --config Release -j $(sysctl -n hw.logicalcpu)

CMakeLists.txt

@@ -150,6 +150,7 @@ option(LLAMA_LLGUIDANCE "llama-common: include LLGuidance library for structured
 # Required for relocatable CMake package
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/build-info.cmake)
 include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/common.cmake)
+include(${CMAKE_CURRENT_SOURCE_DIR}/cmake/win-app.cmake)
 
 if (NOT DEFINED LLAMA_BUILD_NUMBER)
     set(LLAMA_BUILD_NUMBER        ${BUILD_NUMBER})
@@ -269,6 +270,8 @@ if (LLAMA_BUILD_MTMD AND NOT (LLAMA_BUILD_COMMON AND LLAMA_BUILD_TOOLS))
     add_subdirectory(tools/mtmd)
 endif()
 
+llama_apply_app_manifests()
+
 #
 # install
 #

cmake/win-app.cmake

@@ -0,0 +1,66 @@
+set(LLAMA_WIN_APP_MANIFEST "${CMAKE_CURRENT_LIST_DIR}/../common/win-app.manifest")
+
+if(WIN32)
+    if(MSVC OR CMAKE_CXX_SIMULATE_ID STREQUAL "MSVC")
+        set(LLAMA_WIN_APP_MSVC_ABI TRUE)
+    else()
+        set(LLAMA_WIN_APP_MSVC_ABI FALSE)
+    endif()
+
+    if(NOT LLAMA_WIN_APP_MSVC_ABI)
+        if(CMAKE_VERSION VERSION_LESS 3.21)
+            message(FATAL_ERROR "CMake 3.21 or newer is required to build for Windows with the GNU ABI")
+        endif()
+        enable_language(RC)
+        set(LLAMA_WIN_APP_MANIFEST_RC "${CMAKE_CURRENT_BINARY_DIR}/win-app-manifest.rc")
+        configure_file("${CMAKE_CURRENT_LIST_DIR}/win-app.rc.in" "${LLAMA_WIN_APP_MANIFEST_RC}" @ONLY)
+        set_property(DIRECTORY APPEND PROPERTY CMAKE_CONFIGURE_DEPENDS "${LLAMA_WIN_APP_MANIFEST}")
+    endif()
+endif()
+
+function(llama_add_app_manifest target)
+    if(NOT WIN32)
+        return()
+    endif()
+    get_target_property(_aliased ${target} ALIASED_TARGET)
+    if(_aliased)
+        return()
+    endif()
+    get_target_property(_type ${target} TYPE)
+    if(NOT _type STREQUAL "EXECUTABLE")
+        return()
+    endif()
+    if(LLAMA_WIN_APP_MSVC_ABI)
+        target_sources(${target} PRIVATE "${LLAMA_WIN_APP_MANIFEST}")
+    else()
+        if(NOT TARGET llama-win-app-manifest-res)
+            add_library(llama-win-app-manifest-res OBJECT ${LLAMA_WIN_APP_MANIFEST_RC})
+            set_source_files_properties(${LLAMA_WIN_APP_MANIFEST_RC} PROPERTIES OBJECT_DEPENDS "${LLAMA_WIN_APP_MANIFEST}")
+        endif()
+        target_link_libraries(${target} PRIVATE "$<TARGET_OBJECTS:llama-win-app-manifest-res>")
+    endif()
+endfunction()
+
+function(llama_apply_app_manifests_dir dir)
+    get_property(_targets DIRECTORY "${dir}" PROPERTY BUILDSYSTEM_TARGETS)
+    foreach(_target IN LISTS _targets)
+        llama_add_app_manifest(${_target})
+    endforeach()
+    get_property(_subdirs DIRECTORY "${dir}" PROPERTY SUBDIRECTORIES)
+    foreach(_subdir IN LISTS _subdirs)
+        llama_apply_app_manifests_dir("${_subdir}")
+    endforeach()
+endfunction()
+
+function(llama_apply_app_manifests)
+    if(NOT WIN32)
+        return()
+    endif()
+    get_property(_subdirs DIRECTORY "${CMAKE_CURRENT_SOURCE_DIR}" PROPERTY SUBDIRECTORIES)
+    foreach(_dir IN LISTS _subdirs)
+        get_filename_component(_name "${_dir}" NAME)
+        if(_name STREQUAL "app" OR _name STREQUAL "examples" OR _name STREQUAL "pocs" OR _name STREQUAL "tests" OR _name STREQUAL "tools")
+            llama_apply_app_manifests_dir("${_dir}")
+        endif()
+    endforeach()
+endfunction()

cmake/win-app.rc.in

@@ -0,0 +1,3 @@
+// generated from cmake/win-app.rc.in - do not edit
+// RT_MANIFEST (24), ID 1: the application manifest shared by all llama.cpp executables
+1 24 "@LLAMA_WIN_APP_MANIFEST@"

common/common.cpp

@@ -219,6 +219,23 @@ int32_t common_cpu_get_num_math() {
             }
         }
     }
+#elif defined(__riscv) && defined(__linux__)
+    std::ifstream cpuinfo("/proc/cpuinfo");
+    std::string line;
+    int32_t n_a100 = 0;
+    while (std::getline(cpuinfo, line)) {
+        std::istringstream fields(line);
+        std::string key;
+        std::string separator;
+        uint64_t marchid;
+        if (fields >> key >> separator >> std::hex >> marchid &&
+            key == "marchid" && separator == ":" && marchid == UINT64_C(0x8000000041000002)) {
+            ++n_a100;
+        }
+    }
+    if (n_a100 > 0) {
+        return n_a100;
+    }
 #elif defined(__powerpc64__) || defined(__powerpc__)
     int32_t smt_factor = 1;
     int phy_cpus = common_cpu_get_num_physical_cores();

common/reasoning-budget.cpp

@@ -136,7 +136,8 @@ static void common_reasoning_budget_accept(struct llama_sampler * smpl, llama_to
             // track the end sequence within forced_tokens so it is also reported on DONE
             const int32_t match = ctx->end_matcher.advance(token);
             ctx->force_pos++;
-            if (ctx->force_pos >= ctx->forced_tokens.size()) {
+            // an end tag that is already in the prompt (e.g. "<think></think>") also ends forcing
+            if (match >= 0 || ctx->force_pos >= ctx->forced_tokens.size()) {
                 ctx->state = REASONING_BUDGET_DONE;
                 ctx->end_match = match;
                 COM_TRC("%s", "forced sequence complete, done\n");

common/win-app.manifest

@@ -0,0 +1,9 @@
+<?xml version="1.0" encoding="UTF-8" standalone="yes"?>
+<assembly xmlns="urn:schemas-microsoft-com:asm.v1" manifestVersion="1.0">
+  <application xmlns="urn:schemas-microsoft-com:asm.v3">
+    <windowsSettings>
+      <activeCodePage xmlns="http://schemas.microsoft.com/SMI/2019/WindowsSettings">UTF-8</activeCodePage>
+      <longPathAware xmlns="http://schemas.microsoft.com/SMI/2016/WindowsSettings">true</longPathAware>
+    </windowsSettings>
+  </application>
+</assembly>

conversion/__init__.py

@@ -176,6 +176,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "Lfm3ForCausalLM": "lfm3",
     "Lfm3OmniForConditionalGeneration": "lfm3",
     "Lfm3VlForConditionalGeneration": "lfm3",
+    "LimiteForCausalLM": "limite",
     "Llama4ForCausalLM": "llama",
     "Llama4ForConditionalGeneration": "llama",
     "LlamaBidirectionalModel": "llama",
@@ -212,6 +213,7 @@ TEXT_MODEL_MAP: dict[str, str] = {
     "Ministral3Model": "mistral3",
     "Mistral3ForConditionalGeneration": "mistral3",
     "Mistral3Model": "mistral3",
+    "MistralLarge4ForConditionalGeneration": "mistral4",
     "VoxtralRealtimeForConditionalGeneration": "mistral3",
     "MistralForCausalLM": "llama",
     "MixtralForCausalLM": "llama",

conversion/base.py

@@ -16,6 +16,7 @@ from pathlib import Path
 from hashlib import sha256
 from typing import TYPE_CHECKING, Any, Callable, ContextManager, Iterable, Iterator, Literal, Sequence, TypeVar, cast
 from itertools import chain
+from math import prod
 from transformers import AutoConfig
 
 import numpy as np
@@ -155,6 +156,7 @@ class ModelBase:
         self.fuse_gate_up_exps = fuse_gate_up_exps
         self._gate_exp_buffer: dict[int, Tensor] = {}
         self._up_exp_buffer: dict[int, Tensor] = {}
+        self._unfusable_gate_up_layers: set[int] = set()
         self.fuse_qkv = fuse_qkv
         self._q_buffer: dict[int, Tensor] = {}
         self._k_buffer: dict[int, Tensor] = {}
@@ -175,7 +177,9 @@ class ModelBase:
         self._nvfp4_layer_algo: dict[str, str | None] = {} # per-layer quant_algo, keyed by HF module path
         self._prec_a4: dict[str, bool] = {} # gguf tensor name -> can use 4-bit (A4) activations
         self._fp8_as_q8 = fp8_as_q8
+        self._preserve_fp8 = ftype == gguf.LlamaFileType.GUESSED and not fp8_as_q8
         self._fp8_dequantized: set[str] = set()
+        self._fp8_e4m3_preserved: set[str] = set()
 
         # Apply heuristics to figure out typical tensor encoding based on first tensor's dtype
         # NOTE: can't use field "torch_dtype" in config.json, because some finetunes lie.
@@ -326,7 +330,7 @@ class ModelBase:
 
     def dequant_model(self):
         # If all quantized tensors were already handled (e.g. pure NVFP4), skip
-        if self._is_nvfp4 and not any(k.endswith((".weight_scale", ".weight_scale_inv")) for k in self.model_tensors):
+        if self._is_nvfp4 and not any(k.endswith((".weight_scale", ".weight_scale_inv", ".input_scale", ".activation_scale", "_activation_scale", ".k_scale", ".v_scale")) for k in self.model_tensors):
             return
 
         tensors_to_remove: list[str] = []
@@ -350,6 +354,12 @@ class ModelBase:
             def dequant_simple(weight: Tensor, scale: Tensor, block_size: Sequence[int] | None = None) -> Tensor:
                 scale = scale.float()
 
+                # ModelOpt block scales are pre-shaped for broadcasting, e.g. [out/128, 1, in/128, 1]
+                if block_size is None and scale.ndim == 2 * weight.ndim and all(n == 1 for n in scale.shape[1::2]):
+                    assert all(w % n == 0 for w, n in zip(weight.shape, scale.shape[0::2]))
+                    block_size = [w // n for w, n in zip(weight.shape, scale.shape[0::2])]
+                    scale = scale.reshape(tuple(scale.shape[0::2]))
+
                 if block_size is not None:
                     dim_offset = scale.ndim - len(block_size)
                     for i, size in enumerate(block_size):
@@ -440,7 +450,7 @@ class ModelBase:
 
                 return (unpacked * scale.unsqueeze(-1).float()).reshape(shape)
 
-            def dequant_fp8() -> None:
+            def dequant_fp8(fp8_configs: list[dict[str, Any]] | None = None) -> None:
                 for name in self.model_tensors.keys():
                     if name.endswith(".weight_scale"):
                         weight_name = name.removesuffix("_scale")
@@ -449,14 +459,29 @@ class ModelBase:
                             continue
                         w = self.model_tensors[weight_name]
                         s = self.model_tensors[name]
+                        block_size = None
                         is_fp8_weight = False
-                        if self._fp8_as_q8:
-                            is_fp8_weight = w().dtype in (torch.float8_e4m3fn, torch.float8_e5m2)
-                        self.model_tensors[weight_name] = lambda w=w, s=s: dequant_simple(w(), s(), None)
+                        if self._fp8_as_q8 or fp8_configs is not None:
+                            weight = w()
+                            is_fp8_weight = weight.dtype in (torch.float8_e4m3fn, torch.float8_e5m2)
+                        if fp8_configs is not None:
+                            if not is_fp8_weight:
+                                raise ValueError(f"Expected remaining weight {weight_name!r} to be FP8, got {weight.dtype}")
+                            scale = s()
+                            is_channel_scale = scale.ndim >= 1 and scale.ndim <= weight.ndim and scale.shape[0] == weight.shape[0] and all(size == 1 for size in scale.shape[1:])
+                            if prod(scale.shape) != 1 and not is_channel_scale:
+                                block_configs = [c for c in fp8_configs if c.get("strategy") == "block"]
+                                if not block_configs or any(not c.get("block_structure") or c.get("group_size") is not None for c in block_configs):
+                                    raise ValueError(f"FP8 weight {weight_name!r} with scale shape {list(scale.shape)} requires a block layout")
+                                block_sizes = {tuple(c["block_structure"]) for c in block_configs}
+                                if len(block_sizes) != 1:
+                                    raise ValueError(f"FP8 weight {weight_name!r} has ambiguous block layouts")
+                                block_size = next(iter(block_sizes))
+                        self.model_tensors[weight_name] = lambda w=w, s=s, bs=block_size: dequant_simple(w(), s(), bs)
                         tensors_to_remove.append(name)
-                        if is_fp8_weight:
+                        if self._fp8_as_q8 and is_fp8_weight:
                             self._fp8_dequantized.add(weight_name)
-                    if name.endswith((".input_scale", ".k_scale", ".v_scale")):
+                    if name.endswith((".input_scale", ".activation_scale", "_activation_scale", ".k_scale", ".v_scale")):
                         tensors_to_remove.append(name)
 
             if quant_method == "bitnet":
@@ -521,29 +546,37 @@ class ModelBase:
                 nvfp4_compressed_tensors = self._is_nvfp4_compressed_tensors(quant_method, quant_format, groups)
 
                 if nvfp4_compressed_tensors:
-                    dequant_fp8()
+                    fp8_configs = [
+                        g["weights"] for g in groups.values()
+                        if isinstance(g, dict) and g.get("format") == "float-quantized"
+                        and g.get("weights", {}).get("type") == "float"
+                        and g.get("weights", {}).get("num_bits") == 8
+                    ]
+                    dequant_fp8(fp8_configs)
                 elif len(groups) > 1:
                     raise NotImplementedError("Can't handle multiple config groups for compressed-tensors yet")
                 elif quant_format == "float-quantized" or quant_format == "int-quantized" or quant_format == "naive-quantized":
                     weight_config = tuple(groups.values())[0]["weights"]
                     block_size = weight_config.get("block_structure", None)
                     strategy = weight_config.get("strategy")
-                    assert strategy == "channel" or strategy == "block"
-                    assert weight_config.get("group_size") is None  # didn't find a model using this yet
                     is_fp8 = (
                         quant_format == "float-quantized"
                         and weight_config.get("type") == "float"
                         and weight_config.get("num_bits") == 8
                     )
+                    assert strategy in ("channel", "block") or (is_fp8 and strategy == "tensor")
+                    assert weight_config.get("group_size") is None  # didn't find a model using this yet
                     for name in self.model_tensors.keys():
                         if name.endswith(".weight_scale"):
                             weight_name = name.removesuffix("_scale")
                             w = self.model_tensors[weight_name]
                             s = self.model_tensors[name]
-                            self.model_tensors[weight_name] = lambda w=w, s=s: dequant_simple(w(), s(), block_size)
+                            self.model_tensors[weight_name] = lambda w=w, s=s, bs=block_size: dequant_simple(w(), s(), bs)
                             tensors_to_remove.append(name)
                             if self._fp8_as_q8 and is_fp8:
                                 self._fp8_dequantized.add(weight_name)
+                        if is_fp8 and name.endswith((".input_scale", ".activation_scale", "_activation_scale", ".k_scale", ".v_scale")):
+                            tensors_to_remove.append(name)
                 elif quant_format == "pack-quantized":
                     weight_config = tuple(groups.values())[0]["weights"]
                     assert weight_config.get("strategy") == "group"
@@ -581,6 +614,148 @@ class ModelBase:
         for name, value in new_tensors.items():
             self.model_tensors[name] = value
 
+    def _map_fp8_weight_names(self, name: str) -> tuple[str, ...]:
+        new_name = self.tensor_map.get_name(name, try_suffixes=(".weight",))
+        return (new_name,) if new_name is not None else ()
+
+    def _prepare_fp8_e4m3_tensors(self):
+        if not self._preserve_fp8:
+            return
+
+        scale_tensors: dict[str, list[tuple[int, float]] | np.ndarray] = {}
+        input_scale_tensors: dict[str, list[tuple[int, float]] | np.ndarray] = {}
+        consumed: list[str] = []
+        candidates: dict[str, tuple[str, Tensor, tuple[str, ...], torch.Size]] = {}
+        weight_dtypes: dict[str, torch.dtype] = {}
+        scaled_weights: set[str] = set()
+        expert_pattern = re.compile(r"(\.[^.]*experts)\.(\d+)\.")
+        scale_types = {
+            gguf.MODEL_TENSOR.ATTN_Q, gguf.MODEL_TENSOR.ATTN_K, gguf.MODEL_TENSOR.ATTN_V,
+            gguf.MODEL_TENSOR.ATTN_OUT, gguf.MODEL_TENSOR.ATTN_QKV, gguf.MODEL_TENSOR.ATTN_GATE,
+            gguf.MODEL_TENSOR.FFN_GATE, gguf.MODEL_TENSOR.FFN_UP, gguf.MODEL_TENSOR.FFN_DOWN,
+            gguf.MODEL_TENSOR.FFN_GATE_SHEXP, gguf.MODEL_TENSOR.FFN_UP_SHEXP, gguf.MODEL_TENSOR.FFN_DOWN_SHEXP,
+            gguf.MODEL_TENSOR.FFN_GATE_EXP, gguf.MODEL_TENSOR.FFN_UP_EXP, gguf.MODEL_TENSOR.FFN_DOWN_EXP,
+            gguf.MODEL_TENSOR.SSM_IN, gguf.MODEL_TENSOR.SSM_OUT, gguf.MODEL_TENSOR.SSM_ALPHA, gguf.MODEL_TENSOR.SSM_BETA,
+            gguf.MODEL_TENSOR.NEXTN_EH_PROJ, gguf.MODEL_TENSOR.NEXTN_SHARED_HEAD_HEAD, gguf.MODEL_TENSOR.OUTPUT,
+        }
+
+        for scale_name in list(self.model_tensors.keys()):
+            weight_name = None
+            if scale_name.endswith(".weight_scale"):
+                weight_name = scale_name.removesuffix("_scale")
+            elif scale_name.endswith("_scale_inv"):
+                weight_name = scale_name.removesuffix("_scale_inv")
+            elif scale_name.endswith(".qscale_weight"):
+                weight_name = scale_name.removesuffix("qscale_weight") + "weight"
+
+            if weight_name is None or weight_name not in self.model_tensors:
+                continue
+
+            scaled_weights.add(weight_name)
+            weight = self.model_tensors[weight_name]()
+            weight_dtypes[weight_name] = weight.dtype
+            if weight.dtype != torch.float8_e4m3fn or weight.ndim < 2:
+                continue
+
+            scale = self.model_tensors[scale_name]()
+            if prod(scale.shape) != 1:
+                continue
+
+            merged_name = expert_pattern.sub(r"\1.", weight_name)
+            new_names = self._map_fp8_weight_names(merged_name)
+            if not new_names:
+                continue
+            if any((mapped := self.tensor_map.get_type_and_name(new_name, try_suffixes=(".weight",))) is None or mapped[0] not in scale_types for new_name in new_names):
+                continue
+            bid = next((int(part) for part in new_names[0].split(".") if part.isdecimal()), None)
+            self._fp8_e4m3_preserved.add(weight_name)
+            if any(new_name.endswith("_norm.weight") or self.tensor_force_quant(weight_name, new_name, bid, weight.ndim) != gguf.GGMLQuantizationType.F8_E4M3 for new_name in new_names):
+                self._fp8_e4m3_preserved.discard(weight_name)
+                continue
+            candidates[weight_name] = (scale_name, scale, new_names, weight.shape)
+
+        expert_groups: dict[str, list[str]] = {}
+        for name in self.model_tensors:
+            if (match := expert_pattern.search(name)) and name.endswith(".weight"):
+                expert_groups.setdefault(name[:match.end(1)], []).append(name)
+        for names in expert_groups.values():
+            for name in names:
+                if name not in weight_dtypes:
+                    weight_dtypes[name] = self.model_tensors[name]().dtype
+            dtypes = {weight_dtypes[name] for name in names}
+            if torch.float8_e4m3fn not in dtypes:
+                continue
+            if dtypes == {torch.float8_e4m3fn} and not any(name in scaled_weights for name in names):
+                continue
+            # The converter merges expert projections together.
+            if not all(name in candidates for name in names):
+                for name in names:
+                    candidates.pop(name, None)
+                    self._fp8_e4m3_preserved.discard(name)
+                    if name not in scaled_weights:
+                        gen = self.model_tensors[name]
+                        self.model_tensors[name] = lambda gen=gen: gen().float()
+
+        for weight_name, (scale_name, scale, new_names, shape) in candidates.items():
+            scale = LazyTorchTensor.to_eager(scale).float().flatten()
+
+            weight_prefix = weight_name.removesuffix(".weight")
+            # Transformers fine-grained FP8 uses activation_scale while ModelOpt uses input_scale.
+            # Ref: https://github.com/vllm-project/vllm/blob/main/vllm/model_executor/models/mistral3.py#L357-L361
+            input_scale_name = next((name for name in (
+                weight_prefix + ".input_scale",
+                weight_prefix + ".activation_scale",
+                weight_prefix + "_activation_scale",
+                weight_prefix + ".qscale_act",
+            ) if name in self.model_tensors), None)
+            input_scale = None
+            if input_scale_name is not None:
+                input_scale = self.model_tensors[input_scale_name]()
+                if prod(input_scale.shape) != 1:
+                    raise ValueError(f"FP8 input scale {input_scale_name!r} must be a scalar")
+                input_scale = LazyTorchTensor.to_eager(input_scale).float().flatten()
+                consumed.append(input_scale_name)
+
+            consumed.append(scale_name)
+
+            expert_match = expert_pattern.search(weight_name)
+            if expert_match:
+                expert_id = int(expert_match.group(2))
+                for new_name in new_names:
+                    target_name = new_name.replace(".weight", ".scale")
+                    entries = scale_tensors.setdefault(target_name, [])
+                    assert isinstance(entries, list)
+                    cast(list[tuple[int, float]], entries).append((expert_id, float(scale[0])))
+                    if input_scale is not None:
+                        target_name = new_name.replace(".weight", ".input_scale")
+                        entries = input_scale_tensors.setdefault(target_name, [])
+                        assert isinstance(entries, list)
+                        cast(list[tuple[int, float]], entries).append((expert_id, float(input_scale[0])))
+            else:
+                for new_name in new_names:
+                    mapped = self.tensor_map.get_type_and_name(new_name, try_suffixes=(".weight",)) if len(shape) == 3 else None
+                    is_packed_expert = mapped is not None and mapped[0] in (
+                        gguf.MODEL_TENSOR.FFN_GATE_EXP, gguf.MODEL_TENSOR.FFN_UP_EXP, gguf.MODEL_TENSOR.FFN_DOWN_EXP,
+                    )
+                    # Packed expert weights need one scale value per expert.
+                    n_scales = shape[0] if is_packed_expert else 1
+                    scale_tensors[new_name.replace(".weight", ".scale")] = np.repeat(scale.numpy(), n_scales)
+                    if input_scale is not None:
+                        input_scale_tensors[new_name.replace(".weight", ".input_scale")] = np.repeat(input_scale.numpy(), n_scales)
+
+        for name in consumed:
+            self.model_tensors.pop(name, None)
+
+        for name, values in chain(scale_tensors.items(), input_scale_tensors.items()):
+            if isinstance(values, list):
+                values.sort(key=lambda item: item[0])
+                scale = np.array([item[1] for item in values], dtype=np.float32)
+            else:
+                scale = values.astype(np.float32)
+            if not np.allclose(scale, 1.0, atol=1e-6):
+                logger.info(f"  + {name} (FP8 scale, shape [{scale.size}])")
+                self.gguf_writer.add_tensor(name, scale)
+
     @classmethod
     def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
         name, gen = item
@@ -625,6 +800,26 @@ class ModelBase:
             raise ValueError(f"Can not map tensor {name!r}")
         return new_name
 
+    def prepare_gate_up_fusion(self) -> None:
+        self._unfusable_gate_up_layers.clear()
+        if not self.fuse_gate_up_exps or self._fp8_as_q8:
+            return
+
+        for name, gen in self.model_tensors.items():
+            merged_name = re.sub(r"(\.experts)\.\d+\.", r"\1.", name)
+            mapped = self.tensor_map.get_type_and_name(merged_name, try_suffixes=(".weight",))
+            if mapped is None:
+                continue
+            tensor_type, new_name = mapped
+            if tensor_type not in (gguf.MODEL_TENSOR.FFN_GATE_EXP, gguf.MODEL_TENSOR.FFN_UP_EXP, gguf.MODEL_TENSOR.FFN_GATE_UP_EXP) or not new_name.endswith(".weight"):
+                continue
+            if gen().dtype == torch.float8_e4m3fn:
+                bid = next(int(part) for part in new_name.split(".") if part.isdecimal())
+                self._unfusable_gate_up_layers.add(bid)
+
+        for bid in sorted(self._unfusable_gate_up_layers):
+            logger.warning(f"Skipping --fuse-gate-up-exps for layer {bid}: fusion of preserved FP8 expert projections is not supported")
+
     def prepare_qkv_fusion(self) -> None:
         self._fusable_qkv_weight_layers.clear()
         self._fusable_qkv_bias_layers.clear()
@@ -638,6 +833,7 @@ class ModelBase:
         }
         weights: dict[int, set[gguf.MODEL_TENSOR]] = {}
         biases: dict[int, set[gguf.MODEL_TENSOR]] = {}
+        fp8_layers: set[int] = set()
 
         for name in self.model_tensors:
             mapped = self.tensor_map.get_type_and_name(name, try_suffixes=(".weight", ".bias"))
@@ -652,12 +848,19 @@ class ModelBase:
                 continue
             if new_name.endswith(".weight"):
                 weights.setdefault(bid, set()).add(tensor_type)
+                if not self._fp8_as_q8 and self.model_tensors[name]().dtype == torch.float8_e4m3fn:
+                    fp8_layers.add(bid)
             elif new_name.endswith(".bias"):
                 biases.setdefault(bid, set()).add(tensor_type)
 
         for bid, weight_types in weights.items():
             bias_types = biases.get(bid, set())
             if weight_types == qkv_types and (not bias_types or bias_types == qkv_types):
+                # NVFP4 weights use separate reordering and repacking hooks and are exported before this fusion pass.
+                # Separate FP8 scales cannot be represented by one fused QKV scale.
+                if bid in fp8_layers:
+                    logger.warning(f"Skipping --fuse-qkv for layer {bid}: fusion of preserved FP8 projections is not supported")
+                    continue
                 self._fusable_qkv_weight_layers.add(bid)
                 if bias_types:
                     self._fusable_qkv_bias_layers.add(bid)
@@ -846,7 +1049,8 @@ class ModelBase:
         new_name = self.map_tensor_name(name)
 
         # Handle gate/up expert tensor fusion if enabled
-        if self.fuse_gate_up_exps and bid is not None:
+        # Preserved FP8 gate/up scales must stay with their separate projections.
+        if self.fuse_gate_up_exps and bid is not None and bid not in self._unfusable_gate_up_layers:
             if self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.FFN_GATE_EXP, bid):
                 self._gate_exp_buffer[bid] = data_torch
             elif self.match_model_tensor_name(new_name, gguf.MODEL_TENSOR.FFN_UP_EXP, bid):
@@ -903,8 +1107,46 @@ class ModelBase:
 
         return [(new_name, data_torch)]
 
+    def _tensor_requires_f32(self, new_name: str, bid: int | None, n_dims: int) -> bool:
+        # Conditions should closely match those in llama_model_quantize_internal in llama.cpp.
+        return n_dims <= 1 or new_name.endswith("_norm.weight") or (
+            any(
+                self.match_model_tensor_name(new_name, key, bid)
+                for key in (
+                    gguf.MODEL_TENSOR.FFN_GATE_INP,
+                    gguf.MODEL_TENSOR.FFN_GATE_INP_SHEXP,
+                    gguf.MODEL_TENSOR.POS_EMBD,
+                    gguf.MODEL_TENSOR.TOKEN_TYPES,
+                    gguf.MODEL_TENSOR.SSM_CONV1D,
+                    gguf.MODEL_TENSOR.SHORTCONV_CONV,
+                    gguf.MODEL_TENSOR.TIME_MIX_FIRST,
+                    gguf.MODEL_TENSOR.TIME_MIX_W1,
+                    gguf.MODEL_TENSOR.TIME_MIX_W2,
+                    gguf.MODEL_TENSOR.TIME_MIX_DECAY_W1,
+                    gguf.MODEL_TENSOR.TIME_MIX_DECAY_W2,
+                    gguf.MODEL_TENSOR.TIME_MIX_LERP_FUSED,
+                    gguf.MODEL_TENSOR.POSNET_NORM1,
+                    gguf.MODEL_TENSOR.POSNET_NORM2,
+                    gguf.MODEL_TENSOR.V_ENC_EMBD_POS,
+                    gguf.MODEL_TENSOR.A_ENC_EMBD_POS,
+                    gguf.MODEL_TENSOR.ALTUP_CORRECT_COEF,
+                    gguf.MODEL_TENSOR.ALTUP_PREDICT_COEF,
+                    # Kimi KDA conv weights should be F32
+                    gguf.MODEL_TENSOR.SSM_CONV1D_Q,
+                    gguf.MODEL_TENSOR.SSM_CONV1D_K,
+                    gguf.MODEL_TENSOR.SSM_CONV1D_V,
+                    # DSA indexer weights should be F32
+                    gguf.MODEL_TENSOR.INDEXER_PROJ,
+                )
+            )
+            or new_name[-7:] not in (".weight", ".lora_a", ".lora_b")
+        )
+
     def tensor_force_quant(self, name: str, new_name: str, bid: int | None, n_dims: int) -> gguf.GGMLQuantizationType | bool:
-        del new_name, bid  # unused
+        if name in self._fp8_e4m3_preserved and n_dims >= 2:
+            if self._tensor_requires_f32(new_name, bid, n_dims):
+                return gguf.GGMLQuantizationType.F32
+            return gguf.GGMLQuantizationType.F8_E4M3
         # Force FP8-original tensors to Q8_0 when requested; Q8_0 is faster than F16/BF16.
         if self._fp8_as_q8 and name in self._fp8_dequantized and n_dims >= 2:
             return gguf.GGMLQuantizationType.Q8_0
@@ -1035,7 +1277,7 @@ class ModelBase:
         expert_scales: dict[tuple[int, str], list[tuple[int, float]]] = {}
         expert_input_scales: dict[tuple[int, str], list[tuple[int, float]]] = {}
         expert_shapes: dict[tuple[int, str], list[int]] = {}
-        n_experts = self.find_hparam(["num_local_experts", "num_experts"], optional=True) or 0
+        n_experts = self.find_hparam(["num_local_experts", "num_experts", "n_routed_experts"], optional=True) or 0
         consumed: list[str] = []
 
         for name in self.model_tensors.keys():
@@ -1046,14 +1288,17 @@ class ModelBase:
             input_scale_name = name.replace(".weight", ".input_scale")
             if scale_name not in self.model_tensors:
                 continue
-            # Force eager materialization of lazy tensors
-            weight = LazyTorchTensor.to_eager(self.model_tensors[name]())
-            scale = LazyTorchTensor.to_eager(self.model_tensors[scale_name]())
+            weight = self.model_tensors[name]()
+            scale = self.model_tensors[scale_name]()
 
-            # Skip non-NVFP4 tensors(e.g. 1D scale, or float8 weight)
-            if scale.ndim < 2 or weight.dtype in (torch.float8_e4m3fn, torch.float8_e5m2):
+            # Leave FP8 weights and their scales for _prepare_fp8_e4m3_tensors.
+            if weight.dtype in (torch.float8_e4m3fn, torch.float8_e5m2) or scale.ndim < 2:
                 continue
+            if weight.ndim != 2 or scale.shape[0] != weight.shape[0] or scale.shape[1] * 8 != weight.shape[1]:
+                raise ValueError(f"NVFP4 weight {name!r} has incompatible shapes: weight {list(weight.shape)}, scale {list(scale.shape)}")
 
+            weight = LazyTorchTensor.to_eager(weight)
+            scale = LazyTorchTensor.to_eager(scale)
             scale2 = LazyTorchTensor.to_eager(self.model_tensors.get(scale2_name, lambda: torch.tensor(1.0))())
             input_scale = LazyTorchTensor.to_eager(self.model_tensors.get(input_scale_name, lambda: torch.tensor(1.0))())
 
@@ -1209,8 +1454,10 @@ class ModelBase:
                             self.model_tensors[input_scale_name] = inverse_scale(self.model_tensors.pop(name))
             self._generate_nvfp4_tensors()
 
+        self._prepare_fp8_e4m3_tensors()
         self.dequant_model()
 
+        self.prepare_gate_up_fusion()
         self.prepare_qkv_fusion()
 
         # Handle empty tensor_map for models with block_count=0 (like MobileNetV5)
@@ -1224,10 +1471,21 @@ class ModelBase:
             if name.endswith((".attention.masked_bias", ".attention.bias", ".rotary_emb.inv_freq")):
                 continue
 
+            # FP8 transforms must only reorder values; norms and vectors use numeric F32 transforms.
+            preserve_fp8_e4m3 = self._preserve_fp8 and data_torch.dtype == torch.float8_e4m3fn and data_torch.ndim >= 2 and not name.endswith(("norm.weight", "norm_f.weight"))
+            if preserve_fp8_e4m3:
+                for new_name in self._map_fp8_weight_names(name):
+                    mapped_bid = next((int(part) for part in new_name.split(".") if part.isdecimal()), None)
+                    if self._tensor_requires_f32(new_name, mapped_bid, data_torch.ndim):
+                        preserve_fp8_e4m3 = False
+                        break
+                if preserve_fp8_e4m3:
+                    self._fp8_e4m3_preserved.add(name)
+
             old_dtype = data_torch.dtype
 
             # convert any unsupported data types to float32
-            if data_torch.dtype not in (torch.float16, torch.float32):
+            if data_torch.dtype not in (torch.float16, torch.float32) and name not in self._fp8_e4m3_preserved:
                 data_torch = data_torch.to(torch.float32)
 
             # use the first number-like part of the tensor name as the block id
@@ -1238,51 +1496,14 @@ class ModelBase:
                     break
 
             for new_name, data_torch in (self.modify_tensors(data_torch, name, bid)):
-                # TODO: why do we squeeze here?
-                # data = data_torch.squeeze().numpy()
-                data = data_torch.numpy()
-
-                n_dims = len(data.shape)
+                n_dims = len(data_torch.shape)
                 data_qtype: gguf.GGMLQuantizationType | bool = self.tensor_force_quant(name, new_name, bid, n_dims)
 
                 # Most of the codebase that takes in 1D tensors or norms only handles F32 tensors
                 if n_dims <= 1 or new_name.endswith("_norm.weight"):
                     data_qtype = gguf.GGMLQuantizationType.F32
 
-                # Conditions should closely match those in llama_model_quantize_internal in llama.cpp
-                # Some tensor types are always in float32
-                if data_qtype is False and (
-                    any(
-                        self.match_model_tensor_name(new_name, key, bid)
-                        for key in (
-                            gguf.MODEL_TENSOR.FFN_GATE_INP,
-                            gguf.MODEL_TENSOR.FFN_GATE_INP_SHEXP,
-                            gguf.MODEL_TENSOR.POS_EMBD,
-                            gguf.MODEL_TENSOR.TOKEN_TYPES,
-                            gguf.MODEL_TENSOR.SSM_CONV1D,
-                            gguf.MODEL_TENSOR.SHORTCONV_CONV,
-                            gguf.MODEL_TENSOR.TIME_MIX_FIRST,
-                            gguf.MODEL_TENSOR.TIME_MIX_W1,
-                            gguf.MODEL_TENSOR.TIME_MIX_W2,
-                            gguf.MODEL_TENSOR.TIME_MIX_DECAY_W1,
-                            gguf.MODEL_TENSOR.TIME_MIX_DECAY_W2,
-                            gguf.MODEL_TENSOR.TIME_MIX_LERP_FUSED,
-                            gguf.MODEL_TENSOR.POSNET_NORM1,
-                            gguf.MODEL_TENSOR.POSNET_NORM2,
-                            gguf.MODEL_TENSOR.V_ENC_EMBD_POS,
-                            gguf.MODEL_TENSOR.A_ENC_EMBD_POS,
-                            gguf.MODEL_TENSOR.ALTUP_CORRECT_COEF,
-                            gguf.MODEL_TENSOR.ALTUP_PREDICT_COEF,
-                            # Kimi KDA conv weights should be F32
-                            gguf.MODEL_TENSOR.SSM_CONV1D_Q,
-                            gguf.MODEL_TENSOR.SSM_CONV1D_K,
-                            gguf.MODEL_TENSOR.SSM_CONV1D_V,
-                            # DSA indexer weights should be F32
-                            gguf.MODEL_TENSOR.INDEXER_PROJ,
-                        )
-                    )
-                    or new_name[-7:] not in (".weight", ".lora_a", ".lora_b")
-                ):
+                if data_qtype is False and self._tensor_requires_f32(new_name, bid, n_dims):
                     data_qtype = gguf.GGMLQuantizationType.F32
 
                 if data_qtype is False and any(
@@ -1320,16 +1541,22 @@ class ModelBase:
                     else:
                         raise ValueError(f"Unknown file type: {self.ftype.name}")
 
+                if data_torch.dtype == torch.float8_e4m3fn:
+                    data = data_torch.view(torch.uint8).numpy() if data_qtype == gguf.GGMLQuantizationType.F8_E4M3 else data_torch.float().numpy()
+                else:
+                    data = data_torch.numpy()
+
                 # a chunked tensor quantizes as one chunk at a time, while it is written
                 quantize = data.quantize if isinstance(data, gguf.LazyChunkedTensor) else (
                     lambda qtype, d=data: gguf.quants.quantize(d, qtype))
 
-                try:
-                    data = quantize(data_qtype)
-                except gguf.QuantError as e:
-                    logger.warning("%s, %s", e, "falling back to F16")
-                    data_qtype = gguf.GGMLQuantizationType.F16
-                    data = quantize(data_qtype)
+                if not (data_qtype == gguf.GGMLQuantizationType.F8_E4M3 and data.dtype == np.uint8):
+                    try:
+                        data = quantize(data_qtype)
+                    except gguf.QuantError as e:
+                        logger.warning("%s, %s", e, "falling back to F16")
+                        data_qtype = gguf.GGMLQuantizationType.F16
+                        data = quantize(data_qtype)
 
                 shape = gguf.quant_shape_from_byte_shape(data.shape, data_qtype) if data.dtype == np.uint8 else data.shape
 
@@ -1370,6 +1597,8 @@ class ModelBase:
                 self.ftype = gguf.LlamaFileType.MOSTLY_NVFP4
             elif self._is_mxfp4:
                 self.ftype = gguf.LlamaFileType.MOSTLY_MXFP4_MOE
+            elif self._fp8_e4m3_preserved:
+                self.ftype = gguf.LlamaFileType.MOSTLY_F8_E4M3
 
         # Generate parameter weight class (useful for leader boards) if not yet determined
         if self.metadata.size_label is None and total_params > 0:
@@ -2672,7 +2901,13 @@ class TextModel(ModelBase):
         )
 
         if vocab.tokenizer_type == MistralTokenizerType.tekken:
-            self.gguf_writer.add_tokenizer_pre("tekken")
+            with open(vocab.fname_tokenizer, encoding="utf-8") as f:
+                pattern = json.load(f)["config"]["pattern"]
+            # newer tekken.json split numbers into right-aligned groups of 3 digits instead of single digits
+            if "\\p{N}{1,3}(?=(?:\\p{N}{3})*(?!\\p{N}))" in pattern:
+                self.gguf_writer.add_tokenizer_pre("tekken-num3")
+            else:
+                self.gguf_writer.add_tokenizer_pre("tekken")
             self.gguf_writer.add_token_merges(
                 vocab.extract_vocab_merges_from_model()
             )

conversion/deepseek.py

@@ -236,6 +236,12 @@ class DeepseekModel(TextModel):
 class DeepseekV2Model(TextModel):
     model_arch = gguf.MODEL_ARCH.DEEPSEEK2
 
+    def _map_fp8_weight_names(self, name: str) -> tuple[str, ...]:
+        # MLA projections do not have runtime weight-scale support.
+        if name.endswith((".q_a_proj.weight", ".q_b_proj.weight", ".kv_a_proj_with_mqa.weight", ".kv_b_proj.weight")):
+            return ()
+        return super()._map_fp8_weight_names(name)
+
     # TODO @ngxson : remove this when we support MTP for deepseek models
     skip_mtp = True
 

conversion/lfm3.py

@@ -453,9 +453,6 @@ class LFM3AudioMmprojModel(MmprojModel):
     def tensor_force_quant(self, name, new_name, bid, n_dims):
         if new_name.startswith("a.conv1d.") or "conv_dw" in new_name or new_name.startswith("a.gen.wav.quant."):
             return gguf.GGMLQuantizationType.F32
-        # the codec decoder is small and its log-magnitude/phase output does not survive quantization
-        if new_name.startswith("a.gen.wav."):
-            return gguf.GGMLQuantizationType.F16
         return super().tensor_force_quant(name, new_name, bid, n_dims)
 
     def generate_extra_tensors(self) -> Iterable[tuple[str, Tensor]]:

conversion/limite.py

@@ -0,0 +1,184 @@
+from __future__ import annotations
+
+from typing import Iterable, TYPE_CHECKING
+
+import torch
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+from .base import ModelBase, TextModel, gguf
+
+
+@ModelBase.register("LimiteForCausalLM")
+@ModelBase.example("paradigma-inc/limite-1b-violetto")
+class LimiteModel(TextModel):
+    model_arch = gguf.MODEL_ARCH.LIMITE
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        hp = self.hparams
+
+        expected = {
+            "pos_mode": "rope",
+            "qk_norm": "rms_pre_rope",
+            "rope_style": "interleaved_pairs_odd_lane_sign_flip",
+            "rope_per_layer": False,
+            "rms_norm_has_weight": False,
+            "rms_norm_eps_mode": "torch_finfo_default",
+            "mlp_type": "swiglu",
+            "attn_gate_applied": "per_head_before_o_proj",
+            "ve_head_slice": "first_num_key_value_heads",
+            "ve_applied_before_qk_norm": True,
+            "global_nope": True,
+            "global_window": -1,
+            "mudd_r_site": "resid",
+            "tie_word_embeddings": True,
+            "final_softcap": 0.0,
+        }
+        for key, value in expected.items():
+            if hp.get(key) != value:
+                raise NotImplementedError(f"Limite: unsupported {key}={hp.get(key)!r}, expected {value!r}")
+        if hp["softcap_logits"]["kind"] != "sigmoid":
+            raise NotImplementedError(f"Limite: unsupported softcap_logits kind {hp['softcap_logits']['kind']!r}")
+
+        self.n_head    = hp["num_attention_heads"]
+        self.n_head_kv = hp["num_key_value_heads"]
+        self.head_dim  = hp["head_dim"]
+        self.n_pairs   = hp["rope_n_pairs"]
+
+        self.mudd_taps: dict[int, list[int]] = {}
+        if hp.get("mudd"):
+            if not hp.get("mudd_mlp"):
+                raise NotImplementedError("Limite: mudd without mudd_mlp is not supported")
+            self.mudd_taps = {int(il): [int(t) for t in taps] for il, taps in hp["mudd_tap_idx"].items()}
+            assert all(len(taps) == hp["mudd_taps"] for taps in self.mudd_taps.values())
+
+    def set_vocab(self):
+        self._set_vocab_gpt2()
+
+    def set_gguf_parameters(self):
+        super().set_gguf_parameters()
+        hp = self.hparams
+
+        self.gguf_writer.add_key_length(self.head_dim)
+        self.gguf_writer.add_value_length(self.head_dim)
+
+        # torch rms_norm with eps=None uses the fp32 epsilon, even for bf16 inputs
+        self.gguf_writer.add_layer_norm_rms_eps(torch.finfo(torch.float32).eps)
+        self.gguf_writer.add_attention_scale(hp["attention_softmax_scale"])
+        self.gguf_writer.add_xsa_eps(hp["xsa_normalize_eps"] if hp.get("xsa") else 0.0)
+
+        # HF freq_i = base^(-i/(n_pairs-1)), ggml uses freq_base^(-i/n_pairs)
+        self.gguf_writer.add_rope_dimension_count(2 * self.n_pairs)
+        self.gguf_writer.add_rope_freq_base(hp["rope_base_local"] ** (self.n_pairs / (self.n_pairs - 1)))
+
+        # HF keeps keys k >= q - window, so a window of W spans W + 1 tokens
+        global_layers = set(hp["global_layers"])
+        self.gguf_writer.add_sliding_window(hp["sliding_window"] + 1)
+        self.gguf_writer.add_sliding_window_pattern([il not in global_layers for il in range(self.block_count)])
+
+        cap = hp["softcap_logits"]
+        self.gguf_writer.add_final_logit_sigmoid_capping(cap["a"], cap["b"], cap["c"])
+
+        if self.mudd_taps:
+            n_taps = hp["mudd_taps"]
+            tap_indices = [-1] * (self.block_count * n_taps)
+            for il, taps in self.mudd_taps.items():
+                tap_indices[il * n_taps:(il + 1) * n_taps] = taps
+            self.gguf_writer.add_mudd_feed_forward_length(hp["mudd_inter"])
+            self.gguf_writer.add_mudd_tap_count(n_taps)
+            self.gguf_writer.add_mudd_tap_indices(tap_indices)
+
+    def tensor_force_quant(self, name, new_name, bid, n_dims):
+        # tiny mixing weights, kept at full precision as in the reference
+        # the mudd tensors come from a global source tensor, so take the block id from the new name
+        if new_name.startswith("blk."):
+            bid = int(new_name.split(".")[1])
+        for key in (gguf.MODEL_TENSOR.ATTN_GATE, gguf.MODEL_TENSOR.ATTN_VE_GATE, gguf.MODEL_TENSOR.MUDD_DOWN, gguf.MODEL_TENSOR.MUDD_UP):
+            if self.match_model_tensor_name(new_name, key, bid):
+                return gguf.GGMLQuantizationType.F32
+        return super().tensor_force_quant(name, new_name, bid, n_dims)
+
+    def _swap_rope_pairs(self, w: Tensor, n_head: int) -> Tensor:
+        # HF rotates each pair by -theta; swapping the pair lanes gives the ggml +theta rotation.
+        # q and k are both swapped, so the attention scores do not change.
+        idx = torch.arange(self.head_dim)
+        idx[:2 * self.n_pairs] = idx[:2 * self.n_pairs].view(-1, 2).flip(-1).reshape(-1)
+        w = w.reshape(n_head, self.head_dim, -1)
+        return w[:, idx, :].reshape(n_head * self.head_dim, -1)
+
+    def _pad_cols(self, w: Tensor) -> Tensor:
+        # gates read only the first channels of the input, pad them to n_embd
+        return torch.nn.functional.pad(w, (0, self.hparams["hidden_size"] - w.shape[-1]))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        hp = self.hparams
+        T = gguf.MODEL_TENSOR
+
+        def load(n: str) -> Tensor:
+            return self.model_tensors[n]().float()
+
+        if name == "model.value_embeds.weight":
+            ve = data_torch.reshape(data_torch.shape[0], -1, hp["ve_dim"])[:, :self.n_head_kv]
+            ve = torch.nn.functional.pad(ve, (0, self.head_dim - hp["ve_dim"]))
+            yield self.format_tensor_name(T.VALUE_EMBD), ve.reshape(ve.shape[0], -1) * hp["ve_gate_scale"]
+            return
+
+        if name == "model.mudd.dense1":
+            for il in self.mudd_taps:
+                yield self.format_tensor_name(T.MUDD_DOWN, il), data_torch
+            return
+        if name == "model.mudd.dense2":
+            # rows [0, n) mix the attention input, rows [n, 2n) mix the residual
+            dense2_mlp = load("model.mudd.dense2_mlp")
+            bias       = load("model.mudd.bias")
+            bias_mlp   = load("model.mudd.bias_mlp")
+            for il, taps in self.mudd_taps.items():
+                n = len(taps)
+                w = torch.cat([data_torch[il, :n], dense2_mlp[il, :n]], 0)
+                b = torch.cat([bias[il, :n], bias_mlp[il, :n]], 0)
+                yield self.format_tensor_name(T.MUDD_UP, il), w
+                yield self.format_tensor_name(T.MUDD_UP, il, ".bias"), b
+            return
+        if name.startswith("model.mudd."):
+            return
+
+        if bid is None:
+            yield from super().modify_tensors(data_torch, name, bid)
+            return
+
+        prefix = f"model.layers.{bid}."
+        name = name.removeprefix(prefix)
+
+        if name in ("self_attn.q_proj.weight", "self_attn.k_proj.weight", "self_attn.v_proj.weight"):
+            data_torch = data_torch * load(prefix + "self_attn.qkv_scale")
+            if name == "self_attn.q_proj.weight":
+                data_torch = self._swap_rope_pairs(data_torch, self.n_head)
+            elif name == "self_attn.k_proj.weight":
+                data_torch = self._swap_rope_pairs(data_torch, self.n_head_kv)
+        elif name == "self_attn.o_proj.weight":
+            data_torch = data_torch * load(prefix + "self_attn.o_scale") * load(prefix + "post_lambda_attn") * hp["attn_gate_scale"]
+        elif name == "mlp.down_proj.weight":
+            data_torch = data_torch * load(prefix + "post_lambda_mlp")
+        elif name == "self_attn.attn_gate":
+            yield self.format_tensor_name(T.ATTN_GATE, bid), self._pad_cols(data_torch)
+            return
+        elif name == "self_attn.ve_gate":
+            yield self.format_tensor_name(T.ATTN_VE_GATE, bid), self._pad_cols(data_torch[:self.n_head_kv])
+            return
+        elif name == "self_attn.xsa_alpha":
+            if hp.get("xsa") and bid in hp["xsa_layers"]:
+                yield self.format_tensor_name(T.ATTN_XSA_ALPHA, bid), torch.tanh(data_torch)
+            return
+        elif name == "resid_lambda_attn":
+            yield self.format_tensor_name(T.ATTN_RESID_SCALE, bid), data_torch.reshape(1)
+            return
+        elif name == "resid_lambda_mlp":
+            yield self.format_tensor_name(T.FFN_RESID_SCALE, bid), data_torch.reshape(1)
+            return
+        elif name in ("self_attn.qkv_scale", "self_attn.o_scale", "post_lambda_attn", "post_lambda_mlp"):
+            # folded into the projections above
+            return
+
+        yield from super().modify_tensors(data_torch, prefix + name, bid)

conversion/mistral4.py

@@ -0,0 +1,217 @@
+from __future__ import annotations
+
+from typing import Callable, Iterable, TYPE_CHECKING
+
+import torch
+
+if TYPE_CHECKING:
+    from torch import Tensor
+
+from .base import ModelBase, gguf
+
+from .mistral import MistralModel, MistralMoeModel
+
+
+@ModelBase.register("MistralLarge4ForConditionalGeneration")
+class MistralLarge4Model(MistralMoeModel):
+    model_arch = gguf.MODEL_ARCH.MISTRAL_LARGE4
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        n_layer = self.hparams["num_hidden_layers"]
+        dsa = self.hparams["dsa"]
+
+        def expand(pattern: str) -> list[str]:
+            values = pattern.split(",")
+            return [values[i % len(values)] for i in range(n_layer)]
+
+        self.layer_has_rope = [v == "rope" for v in expand(self.hparams.get("nope") or "rope")]
+        self.layer_indexer_full = [v == "full" for v in expand(dsa.get("sharing") or "full")]
+
+        kv_sharing = self.hparams.get("kv_sharing") or {}
+        if "source_layers_str" in kv_sharing:
+            self.layer_kv_source = [int(v) for v in kv_sharing["source_layers_str"].split(",")]
+        else:
+            assert not kv_sharing, f"unsupported kv_sharing config: {kv_sharing}"
+            self.layer_kv_source = list(range(n_layer))
+        assert len(self.layer_kv_source) == n_layer
+
+        # note: the rope dim of the indexer must match the main attention
+        assert dsa.get("rope_head_dim", -1) in (-1, self.hparams["qk_rope_head_dim"])
+        assert dsa.get("match_nope", False), "indexer must follow the NoPE pattern of the main attention"
+        assert dsa.get("ssmax_scaling", False), "indexer without SSMax is not supported"
+        assert self.hparams.get("q_lora_rank") is None
+        assert not self.hparams["yarn"].get("apply_scale", True), "only YaRN without magnitude scaling is supported"
+
+    def dequant_model(self):
+        quant_config = self.hparams.get("quantization")
+        if quant_config is not None:
+            assert quant_config["qformat_weight"] == "fp8_e4m3"
+            assert quant_config.get("qscheme_weight") == "BLOCK"
+            self.hparams["quantization_config"] = {
+                "quant_method": "fp8",
+                "weight_block_size": [quant_config["block_size_out"], quant_config["block_size_in"]],
+            }
+        return super().dequant_model()
+
+    def set_gguf_parameters(self):
+        # skip the MistralMoeModel overrides, they apply the deepseek2 YaRN scaling
+        super(MistralMoeModel, self).set_gguf_parameters()
+        MistralModel.set_mistral_config(self.gguf_writer, self.hparams)
+
+        dsa = self.hparams["dsa"]
+        self.gguf_writer.add_indexer_head_count(dsa["n_heads"])
+        self.gguf_writer.add_indexer_key_length(dsa["head_dim"])
+        self.gguf_writer.add_indexer_top_k(dsa["topk"])
+        self.gguf_writer.add_indexer_types(self.layer_indexer_full)
+        self.gguf_writer.add_layer_norm_eps(dsa.get("norm_eps", 1e-6))
+
+        self.gguf_writer.add_rope_pattern(self.layer_has_rope)
+        self.gguf_writer.add_kv_source_map(self.layer_kv_source)
+        self.gguf_writer.add_expert_gating_func(gguf.ExpertGatingFuncType.SOFTMAX)
+        self.gguf_writer.add_final_logit_softcapping(self.hparams["softmax_tanh_gating"])
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, gen = item
+
+        # exported separately with --dspark
+        if name.startswith("mtp_"):
+            return None
+
+        # experts need HF names for the merging logic in DeepseekV2Model
+        # note: bare "layers.N.experts" comes from the NVFP4 quantized_layers config
+        if ".experts." in name or name.endswith(".experts"):
+            name = "model." + name.replace(".experts", ".mlp.experts")
+            name = name.replace(".w1.", ".gate_proj.").replace(".w2.", ".down_proj.").replace(".w3.", ".up_proj.")
+
+        # skip the MistralMoeModel renaming, the qscale names are needed for block FP8 dequant
+        return super(MistralMoeModel, cls).filter_tensors((name, gen))
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name == "lerope.alpha":
+            # LeRoPE scales inv_freq by exp(alpha), ggml divides theta by the freq factors
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.ROPE_FREQS), torch.exp(-data_torch.float()))
+            return
+
+        if bid is not None and name.endswith((".ssmax_scaling.s", ".ssmax_scaling.b")):
+            tensor = gguf.MODEL_TENSOR.INDEXER_SSMAX if ".indexer." in name else gguf.MODEL_TENSOR.ATTN_SSMAX
+            suffix = ".weight" if name.endswith(".s") else ".bias"
+            yield (self.format_tensor_name(tensor, bid, suffix), data_torch.float())
+            return
+
+        if bid is not None and ".indexer.q_input_norm." in name:
+            suffix = ".weight" if name.endswith(".weight") else ".bias"
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.INDEXER_Q_NORM, bid, suffix), data_torch)
+            return
+
+        if bid is not None and ".indexer." in name:
+            name = "model." + name.replace(".attention.", ".self_attn.")
+
+        # NoPE layers have a larger non-rotary head, so wkv_b is split per layer
+        if bid is not None and name.endswith(".wkv_b.weight"):
+            n_head = self.hparams["num_attention_heads"]
+            v_head_dim = self.hparams["v_head_dim"]
+            qk_nope_head_dim = self.hparams["qk_nope_head_dim"]
+            if not self.layer_has_rope[bid]:
+                qk_nope_head_dim += self.hparams["qk_rope_head_dim"]
+
+            assert data_torch.shape[0] == n_head * (qk_nope_head_dim + v_head_dim)
+
+            kv_b = data_torch.view(n_head, qk_nope_head_dim + v_head_dim, data_torch.shape[-1])
+            k_b, v_b = torch.split(kv_b, [qk_nope_head_dim, v_head_dim], dim=1)
+            k_b = k_b.transpose(1, 2)
+
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_K_B, bid), k_b)
+            yield (self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_V_B, bid), v_b)
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)
+
+
+class MistralLarge4DSparkModel(MistralLarge4Model):
+    """Mistral Large 4 DSpark draft (mtp_dspark.*) as a DFLASH GGUF, export with --dspark"""
+    model_arch = gguf.MODEL_ARCH.DFLASH
+
+    _DSPARK_ROOT: dict[str, gguf.MODEL_TENSOR] = {
+        "backbone.target_projection.weight": gguf.MODEL_TENSOR.FC,
+        "backbone.target_norm.weight":       gguf.MODEL_TENSOR.ENC_OUTPUT_NORM,
+        "backbone.final_norm.weight":        gguf.MODEL_TENSOR.OUTPUT_NORM,
+        "backbone.mask_embedding":           gguf.MODEL_TENSOR.DSPARK_MASK_EMBD,
+        "autoregressive_head.embedding.weight": gguf.MODEL_TENSOR.DSPARK_MARKOV_W1,
+        "autoregressive_head.output.weight":    gguf.MODEL_TENSOR.DSPARK_MARKOV_W2,
+        "autoregressive_head.joint_proj.weight": gguf.MODEL_TENSOR.DSPARK_JOINT_PROJ,
+        "autoregressive_head.joint_proj.bias":   gguf.MODEL_TENSOR.DSPARK_JOINT_PROJ,
+    }
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        mtp = self.hparams.get("mtp") or {}
+        dspark = mtp.get("dspark")
+        if dspark is None:
+            raise ValueError("Mistral Large 4 DSpark export needs the mtp.dspark config")
+        assert dspark.get("head_type") == "rnn", f"unsupported DSpark head type: {dspark.get('head_type')}"
+        assert dspark.get("draft_input") == "position_specific", f"unsupported DSpark draft input: {dspark.get('draft_input')}"
+
+        windows = {int(v) for v in (mtp.get("ragged_attention") or "").split(",") if v.strip()}
+        assert len(windows) == 1, f"DSpark draft layers must share one sliding window, got {mtp.get('ragged_attention')}"
+
+        self._n_stage = int(mtp["n_layers"])
+        self._depth = int(mtp["depth"])
+        self._window = windows.pop()
+        self._target_layers = [int(v) for v in mtp["target_activation_layers"].split(",")]
+
+        self.layer_has_rope = [True] * self._n_stage
+
+    @classmethod
+    def filter_tensors(cls, item: tuple[str, Callable[[], Tensor]]) -> tuple[str, Callable[[], Tensor]] | None:
+        name, gen = item
+        if not name.startswith("mtp_dspark."):
+            return None
+        name = name.removeprefix("mtp_dspark.")
+        if name.startswith("backbone.layers."):
+            return name.removeprefix("backbone."), gen
+        return f"dspark.{name}", gen
+
+    def set_vocab(self):
+        super().set_vocab()
+        # block uses mask_embd, mask token id is not used
+        self.gguf_writer.add_mask_token_id(0)
+
+    def set_gguf_parameters(self):
+        hp = self.hparams
+        w = self.gguf_writer
+        w.add_block_count(self._n_stage)
+        w.add_context_length(hp["max_position_embeddings"])
+        w.add_embedding_length(hp["hidden_size"])
+        w.add_feed_forward_length(hp["intermediate_size"])
+        w.add_head_count(hp["num_attention_heads"])
+        w.add_head_count_kv(1)
+        w.add_key_length(hp["kv_lora_rank"] + hp["qk_rope_head_dim"])
+        w.add_value_length(hp["kv_lora_rank"])
+        w.add_key_length_mla(hp["qk_nope_head_dim"] + hp["qk_rope_head_dim"])
+        w.add_value_length_mla(hp["v_head_dim"])
+        w.add_kv_lora_rank(hp["kv_lora_rank"])
+        w.add_rope_dimension_count(hp["qk_rope_head_dim"])
+        w.add_rope_freq_base(float(hp["rope_theta"]))
+        w.add_layer_norm_rms_eps(hp["rms_norm_eps"])
+        MistralModel.set_mistral_config(w, hp)
+        w.add_sliding_window(self._window)
+        w.add_final_logit_softcapping(hp["softmax_tanh_gating"])
+        w.add_block_size(self._depth)
+        # draft reads layer output = input of next layer
+        w.add_target_layers([layer + 1 for layer in self._target_layers])
+        w.add_sample_from_anchor(True)
+        w.add_has_confidence_head(False)
+        w.add_file_type(self.ftype)
+
+    def modify_tensors(self, data_torch: Tensor, name: str, bid: int | None) -> Iterable[tuple[str, Tensor]]:
+        if name.startswith("dspark."):
+            key = name.removeprefix("dspark.")
+            if key not in self._DSPARK_ROOT:
+                raise ValueError(f"Unexpected DSpark tensor {name!r}")
+            suffix = ".bias" if key.endswith(".bias") else ".weight"
+            yield (self.format_tensor_name(self._DSPARK_ROOT[key], None, suffix), data_torch)
+            return
+
+        yield from super().modify_tensors(data_torch, name, bid)

conversion/nemotron.py

@@ -433,6 +433,11 @@ class NemotronHModel(GraniteHybridModel):
     _ATTN_LAYER_TYPES = {"attention", "full_attention"}
     _MLP_LAYER_TYPES = {"moe"}
 
+    def _map_fp8_weight_names(self, name: str) -> tuple[str, ...]:
+        if ".mixer.experts." in name:
+            name = name.replace("backbone.", "model.").replace(".mixer.experts.", ".mlp.experts.")
+        return super()._map_fp8_weight_names(name)
+
     def __init__(self, *args, **kwargs):
         # We have to determine the correct model architecture (MoE vs non-MoE) before
         # calling the parent __init__. This is because the parent constructor

conversion/qwen.py

@@ -83,6 +83,15 @@ class Qwen2Model(TextModel):
 class Qwen2MoeModel(TextModel):
     model_arch = gguf.MODEL_ARCH.QWEN2MOE
 
+    def _map_fp8_weight_names(self, name: str) -> tuple[str, ...]:
+        if name.removesuffix(".weight").endswith(".mlp.experts.gate_up_proj"):
+            bid = next(int(part) for part in self.map_tensor_name(name).split(".") if part.isdecimal())
+            return (
+                self.format_tensor_name(gguf.MODEL_TENSOR.FFN_GATE_EXP, bid),
+                self.format_tensor_name(gguf.MODEL_TENSOR.FFN_UP_EXP, bid),
+            )
+        return super()._map_fp8_weight_names(name)
+
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         if (moe_intermediate_size := self.hparams.get("moe_intermediate_size")) is not None:
@@ -382,6 +391,16 @@ class _QwenMtpMixin:
 class Qwen3NextModel(_QwenMtpMixin, Qwen2MoeModel):
     model_arch = gguf.MODEL_ARCH.QWEN3NEXT
 
+    def _map_fp8_weight_names(self, name: str) -> tuple[str, ...]:
+        if name.endswith(".linear_attn.in_proj_qkvz.weight"):
+            bid = next(int(part) for part in self.map_tensor_name(name).split(".") if part.isdecimal())
+            # Both split projections keep the original per-tensor scales.
+            return (
+                self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_QKV, bid),
+                self.format_tensor_name(gguf.MODEL_TENSOR.ATTN_GATE, bid),
+            )
+        return super()._map_fp8_weight_names(name)
+
     def set_gguf_parameters(self):
         super().set_gguf_parameters()
         self.gguf_writer.add_ssm_conv_kernel(self.hparams["linear_conv_kernel_dim"])

convert_hf_to_gguf.py

@@ -58,7 +58,7 @@ def parse_args() -> argparse.Namespace:
     )
     parser.add_argument(
         "--outtype", type=str, choices=["f32", "f16", "bf16", "q8_0", "tq1_0", "tq2_0", "auto"], default="auto",
-        help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, q8_0 for Q8_0, tq1_0 or tq2_0 for ternary, and auto for the highest-fidelity 16-bit float type",
+        help="output format - use f32 for float32, f16 for float16, bf16 for bfloat16, q8_0 for Q8_0, tq1_0 or tq2_0 for ternary, and auto to preserve supported FP8 weights and otherwise select the highest-fidelity 16-bit float type",
     )
     parser.add_argument(
         "--bigendian", action="store_true",
@@ -127,7 +127,7 @@ def parse_args() -> argparse.Namespace:
     )
     parser.add_argument(
         "--dspark", action="store_true",
-        help="Export only the DeepSeek-V4 DSpark draft tensors as a separate GGUF.",
+        help="Export only the DeepSeek-V4 or Mistral Large 4 DSpark draft tensors as a separate GGUF.",
     )
     parser.add_argument(
         "--mistral-format", action="store_true",
@@ -255,6 +255,9 @@ def main() -> None:
             assert hparams.get("vision_encoder") is not None, "This model does not support multimodal"
             from conversion.pixtral import PixtralModel
             model_class = PixtralModel
+        elif hparams.get("moe") is not None and hparams.get("dsa") is not None:
+            from conversion.mistral4 import MistralLarge4Model
+            model_class = MistralLarge4Model
         elif hparams.get("moe") is not None:
             from conversion.mistral import MistralMoeModel
             model_class = MistralMoeModel
@@ -267,10 +270,16 @@ def main() -> None:
             sys.exit(1)
 
         if args.dspark:
-            if is_mistral_format or model_architecture not in ("DeepseekV4ForCausalLM", "DeepseekV41ForCausalLM"):
-                logger.error("--dspark is only supported for DeepseekV4ForCausalLM and DeepseekV41ForCausalLM")
+            if is_mistral_format or model_architecture == "MistralLarge4ForConditionalGeneration":
+                if model_class.__name__ != "MistralLarge4Model":
+                    logger.error("--dspark is only supported for Mistral Large 4 among Mistral-format models")
+                    sys.exit(1)
+                from conversion.mistral4 import MistralLarge4DSparkModel
+                model_class = MistralLarge4DSparkModel
+            elif model_architecture not in ("DeepseekV4ForCausalLM", "DeepseekV41ForCausalLM"):
+                logger.error("--dspark is only supported for DeepseekV4ForCausalLM, DeepseekV41ForCausalLM and Mistral Large 4")
                 sys.exit(1)
-            if model_architecture == "DeepseekV41ForCausalLM":
+            elif model_architecture == "DeepseekV41ForCausalLM":
                 from conversion.deepseek41 import DeepseekV41DSparkModel
                 model_class = DeepseekV41DSparkModel
             else:

docs/ops.md

@@ -14,22 +14,22 @@ Legend:
 
 | Operation | BLAS | CANN | CPU | CUDA | ET | HTP | MTL | OpenCL | SYCL | Vulkan | WebGPU | ZenDNN | zDNN |
 |-----------|------|------|------|------|------|------|------|------|------|------|------|------|------|
-|                              ABS | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              ABS | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                              ACC | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                              ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              ADD | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             ADD1 | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
 |                           ADD_ID | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                           ARANGE | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                           ARGMAX | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          ARGSORT | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             CEIL | ❌ | ❌ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             CEIL | ❌ | ❌ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                            CLAMP | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        COL2IM_1D | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                           CONCAT | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
 |                             CONT | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
-|                       CONV_1D_DW | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ✅ |
+|                       CONV_1D_DW | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ✅ |
 |                          CONV_2D | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                       CONV_2D_DW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          CONV_3D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_1D | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                CONV_TRANSPOSE_2D | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
@@ -38,49 +38,49 @@ Legend:
 |                              CPY | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
 |               CROSS_ENTROPY_LOSS | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |          CROSS_ENTROPY_LOSS_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                           CUMSUM | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             DIAG | ❌ | ❌ | ✅ | ✅ | 🟡 | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                           CUMSUM | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             DIAG | ❌ | ❌ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    DIAG_MASK_INF | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                              DIV | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              DIV | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                     DSV4_HC_COMB | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                     DSV4_HC_POST | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                      DSV4_HC_PRE | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                              DUP | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                            EXPM1 | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             FILL | ❌ | ❌ | ✅ | ✅ | 🟡 | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              ELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              EXP | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                            EXPM1 | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             FILL | ❌ | ❌ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                   FLASH_ATTN_EXT | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
-|                            FLOOR | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                  GATED_DELTA_NET | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | 🟡 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                            FLOOR | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                  GATED_DELTA_NET | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                GATED_LINEAR_ATTN | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                      GEGLU_QUICK | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                            GEGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                        GEGLU_ERF | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                      GEGLU_QUICK | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             GELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         GELU_ERF | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                       GELU_QUICK | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         GET_ROWS | ❌ | 🟡 | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ❌ | ❌ |
 |                    GET_ROWS_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ |
 |                       GROUP_NORM | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                      HARDSIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                        HARDSWISH | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                           IM2COL | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                        IM2COL_3D | ❌ | ❌ | ✅ | ✅ | ❌ | 🟡 | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                          L2_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | 🟡 | ❌ | ❌ |
+|                        IM2COL_3D | ❌ | ❌ | ✅ | ✅ | ❌ | 🟡 | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
+|                          L2_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
 |                       LEAKY_RELU | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 | 🟡 | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                LIGHTNING_INDEXER | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                LIGHTNING_INDEXER | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                              LOG | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             MEAN | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ | ❌ |
-|                              MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              MUL | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          MUL_MAT | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 |
-|                 MUL_MAT_HADAMARD | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | 🟡 |
+|                 MUL_MAT_HADAMARD | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | 🟡 | ✅ | ✅ | ✅ | ❌ | 🟡 |
 |                       MUL_MAT_ID | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | 🟡 | ❌ |
-|                  MUL_MAT_ID_W4A4 | ❌ | ❌ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                  MUL_MAT_ID_W4A8 | ❌ | ❌ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                     MUL_MAT_W4A4 | ❌ | ❌ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                     MUL_MAT_W4A8 | ❌ | ❌ | ✅ | ❌ | ❌ | 🟡 | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
-|                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                  MUL_MAT_ID_W4A4 | ❌ | ❌ | ✅ | ❌ | ❌ | 🟡 | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                  MUL_MAT_ID_W4A8 | ❌ | ❌ | ✅ | ❌ | ❌ | 🟡 | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                     MUL_MAT_W4A4 | ❌ | ❌ | ✅ | ❌ | ❌ | 🟡 | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                     MUL_MAT_W4A8 | ❌ | ❌ | ✅ | ❌ | ❌ | 🟡 | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                              NEG | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | 🟡 | ❌ | ❌ |
 |                   OPT_STEP_ADAMW | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                     OPT_STEP_SGD | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
@@ -89,45 +89,46 @@ Legend:
 |                   PAD_REFLECT_1D | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                          POOL_1D | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                          POOL_2D | ❌ | 🟡 | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                            REGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             RELU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                           REPEAT | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | 🟡 | ❌ | ❌ |
 |                      REPEAT_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                         RMS_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                         RMS_NORM | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    RMS_NORM_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                             ROLL | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                             ROPE | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        ROPE_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                            ROUND | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                            ROUND | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        RWKV_WKV6 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
 |                        RWKV_WKV7 | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                            SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                              SET | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ |
+|                            SCALE | ❌ | 🟡 | ✅ | ✅ | ✅ | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              SET | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | 🟡 | ✅ | ✅ | ❌ | ❌ |
 |                         SET_ROWS | ❌ | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
-|                              SGN | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              SGN | ❌ | ✅ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          SIGMOID | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             SILU | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                        SILU_BACK | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ |
 |                              SIN | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                         SOFTPLUS | ❌ | ❌ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                         SOFTPLUS | ❌ | ❌ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         SOFT_MAX | ❌ | 🟡 | ✅ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                    SOFT_MAX_BACK | ❌ | ❌ | 🟡 | 🟡 | ❌ | ❌ | ❌ | ❌ | 🟡 | ✅ | ❌ | ❌ | ❌ |
-|                        SOLVE_TRI | ❌ | ❌ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ |
+|                        SOLVE_TRI | ❌ | ❌ | ✅ | 🟡 | 🟡 | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ❌ | ❌ |
 |                              SQR | ❌ | ✅ | ✅ | ✅ | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                             SQRT | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                         SSM_CONV | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | ❌ | 🟡 | 🟡 | ✅ | ❌ | ❌ |
-|                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                         SSM_SCAN | ❌ | ❌ | ✅ | ✅ | ✅ | ❌ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ❌ | ❌ |
+|                             STEP | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                              SUB | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                              SUM | ❌ | 🟡 | ✅ | 🟡 | ❌ | 🟡 | 🟡 | ❌ | 🟡 | 🟡 | 🟡 | ❌ | ❌ |
 |                         SUM_ROWS | ❌ | ✅ | ✅ | 🟡 | ❌ | 🟡 | ✅ | 🟡 | 🟡 | ✅ | ✅ | ❌ | ❌ |
-|                           SWIGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                     SWIGLU_CLAMP | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ❌ | ❌ |
+|                           SWIGLU | ❌ | ✅ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                     SWIGLU_CLAMP | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ❌ | ❌ | ❌ | ❌ |
 |                       SWIGLU_OAI | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                             TANH | ❌ | ✅ | ✅ | 🟡 | 🟡 | 🟡 | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |               TIMESTEP_EMBEDDING | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ | ❌ |
-|                            TOP_K | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ✅ | 🟡 | ✅ | ❌ | ❌ |
-|                              TRI | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
-|                            TRUNC | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                            TOP_K | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | 🟡 | ✅ | ❌ | ❌ |
+|                       TOP_K_MASK | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ | ❌ |
+|                              TRI | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ | ❌ |
+|                            TRUNC | ❌ | ❌ | ✅ | 🟡 | 🟡 | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                          UPSCALE | ❌ | 🟡 | ✅ | ✅ | ❌ | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |
 |                            XIELU | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |

ggml/include/ggml-backend.h

@@ -70,7 +70,8 @@ extern "C" {
     GGML_API void                           ggml_backend_buffer_reset         (ggml_backend_buffer_t buffer);
 
     // tensor copy between different backends
-    GGML_API void ggml_backend_tensor_copy(const struct ggml_tensor * src, struct ggml_tensor * dst);
+    GGML_API void ggml_backend_tensor_copy  (const struct ggml_tensor * src, struct ggml_tensor * dst);
+    GGML_API void ggml_backend_tensor_copy_n(const struct ggml_tensor * const * srcs, struct ggml_tensor * const * dsts, size_t n_tensors);
 
     //
     // Backend (stream)
@@ -95,6 +96,8 @@ extern "C" {
     GGML_API void ggml_backend_tensor_get   (const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size);
     GGML_API void ggml_backend_tensor_set_2d(      struct ggml_tensor * tensor, const void * data, size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data);
     GGML_API void ggml_backend_tensor_get_2d(const struct ggml_tensor * tensor,       void * data, size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data);
+    GGML_API void ggml_backend_tensor_set_n (      struct ggml_tensor * const * tensors, const void * data, const size_t * offsets, const size_t * sizes, const size_t * data_offsets, size_t n_tensors);
+    GGML_API void ggml_backend_tensor_get_n (const struct ggml_tensor * const * tensors,       void * data, const size_t * offsets, const size_t * sizes, const size_t * data_offsets, size_t n_tensors);
     GGML_API void ggml_backend_tensor_memset(      struct ggml_tensor * tensor,     uint8_t value, size_t offset, size_t size);
 
     GGML_API void ggml_backend_synchronize(ggml_backend_t backend);

ggml/include/ggml.h

@@ -381,6 +381,8 @@ extern "C" {
     GGML_API void        ggml_fp32_to_bf16_row_ref(const float *, ggml_bf16_t *, int64_t);
     GGML_API void        ggml_fp32_to_bf16_row(const float *, ggml_bf16_t *, int64_t);
 
+    typedef struct { uint8_t bits; } ggml_fp8_e4m3_t;
+
     struct ggml_object;
     struct ggml_context;
     struct ggml_cgraph;
@@ -430,7 +432,8 @@ extern "C" {
         GGML_TYPE_NVFP4   = 40, // NVFP4 (4 blocks, E4M3 scale)
         GGML_TYPE_Q1_0    = 41,
         GGML_TYPE_Q2_0    = 42,
-        GGML_TYPE_COUNT   = 43,
+        GGML_TYPE_F8_E4M3 = 43,
+        GGML_TYPE_COUNT   = 44,
     };
 
     // [TAG_GGML_PREC]
@@ -486,6 +489,7 @@ extern "C" {
         GGML_FTYPE_MOSTLY_NVFP4   = 26, // except 1d tensors
         GGML_FTYPE_MOSTLY_Q1_0    = 27, // except 1d tensors
         GGML_FTYPE_MOSTLY_Q2_0    = 28, // except 1d tensors
+        GGML_FTYPE_MOSTLY_F8_E4M3 = 29, // except 1d tensors
     };
 
     // available tensor operations:

ggml/src/ggml-backend-impl.h

@@ -47,6 +47,9 @@ extern "C" {
     // Backend buffer
     //
 
+    // TODO: remove `set_tensor`, `get_tensor`, `set_tensor_2d`, `get_tensor_2d` and `cpy_tensor` from the buffer interface
+    //       they are superseded by the `set_tensor_n`, `get_tensor_n`, `cpy_tensor_n` interface
+    //       ref: https://github.com/ggml-org/llama.cpp/pull/30881#discussion_r4400299490
     struct ggml_backend_buffer_i {
         // (optional) free the buffer
         void         (*free_buffer)  (ggml_backend_buffer_t buffer);
@@ -68,6 +71,11 @@ extern "C" {
         void         (*clear)        (ggml_backend_buffer_t buffer, uint8_t value);
         // (optional) reset any internal state due to tensor initialization, such as tensor extras
         void         (*reset)        (ggml_backend_buffer_t buffer);
+        // (optional) batched tensor data copies
+        void         (*set_tensor_n) (ggml_backend_buffer_t buffer,       struct ggml_tensor * const * tensors, const void * const * data, const size_t * offsets, const size_t * sizes, size_t n_tensors);
+        void         (*get_tensor_n) (ggml_backend_buffer_t buffer, const struct ggml_tensor * const * tensors,       void * const * data, const size_t * offsets, const size_t * sizes, size_t n_tensors);
+        // (optional) batched tensor copies: returns true if all copies were performed, false if unsupported and no copies were performed
+        bool         (*cpy_tensor_n) (ggml_backend_buffer_t buffer, const struct ggml_tensor * const * srcs, struct ggml_tensor * const * dsts, size_t n_tensors);
     };
 
     struct ggml_backend_buffer {

ggml/src/ggml-backend-meta.cpp

@@ -1699,6 +1699,9 @@ static const ggml_backend_buffer_i ggml_backend_meta_buffer_iface = {
     /* .cpy_tensor      = */ nullptr,
     /* .clear           = */ ggml_backend_meta_buffer_clear,
     /* .reset           = */ ggml_backend_meta_buffer_reset,
+    /* .set_tensor_n    = */ NULL,
+    /* .get_tensor_n    = */ NULL,
+    /* .cpy_tensor_n    = */ NULL,
 };
 
 bool ggml_backend_buffer_is_meta(ggml_backend_buffer_t buf) {

ggml/src/ggml-backend.cpp

@@ -402,6 +402,62 @@ size_t ggml_backend_get_max_size(ggml_backend_t backend) {
     return ggml_backend_buft_get_max_size(ggml_backend_get_default_buffer_type(backend));
 }
 
+template <typename cb_get_buffer>
+static void ggml_backend_tensor_buft_groups(size_t n_tensors, cb_get_buffer && get_buffer, std::vector<std::vector<size_t>> & groups) {
+    std::unordered_map<ggml_backend_buffer_type_t, size_t> buft_idx;
+    groups.clear();
+    for (size_t i = 0; i < n_tensors; i++) {
+        ggml_backend_buffer_t buf = get_buffer(i);
+        auto it = buft_idx.find(buf->buft);
+        if (it == buft_idx.end()) {
+            buft_idx[buf->buft] = groups.size();
+            groups.emplace_back();
+        }
+        groups[buft_idx[buf->buft]].push_back(i);
+    }
+}
+
+static bool ggml_backend_tensor_batch_is_2d(
+        const struct ggml_tensor * const * tensors,
+        const size_t * offsets,
+        const size_t * sizes,
+        const size_t * data_offsets,
+        size_t n_tensors,
+        size_t * base_offset,
+        size_t * size,
+        size_t * stride_tensor,
+        size_t * stride_data) {
+    if (n_tensors <= 1) {
+        return false;
+    }
+
+    const struct ggml_tensor * tensor = tensors[0];
+    const size_t size0 = sizes[0];
+
+    if (size0 == 0 || offsets[1] < offsets[0] || data_offsets[1] < data_offsets[0]) {
+        return false;
+    }
+
+    const size_t stride_tensor0 = offsets[1] - offsets[0];
+    const size_t stride_data0   = data_offsets[1] - data_offsets[0];
+
+    for (size_t i = 1; i < n_tensors; i++) {
+        if (tensors[i] != tensor || sizes[i] != size0) {
+            return false;
+        }
+        if (offsets[i] != offsets[0] + i*stride_tensor0 || data_offsets[i] != data_offsets[0] + i*stride_data0) {
+            return false;
+        }
+    }
+
+    *base_offset   = offsets[0];
+    *size          = size0;
+    *stride_tensor = stride_tensor0;
+    *stride_data   = stride_data0;
+
+    return true;
+}
+
 void ggml_backend_tensor_set_async(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size) {
     GGML_ASSERT(backend);
     GGML_ASSERT(tensor);
@@ -546,6 +602,154 @@ void ggml_backend_tensor_get_2d(const struct ggml_tensor * tensor, void * data,
     buf->iface.get_tensor_2d(buf, tensor, data, offset, size, n_copies, stride_tensor, stride_data);
 }
 
+void ggml_backend_tensor_set_n(struct ggml_tensor * const * tensors, const void * data,
+        const size_t * offsets, const size_t * sizes, const size_t * data_offsets, size_t n_tensors) {
+    if (n_tensors == 0) {
+        return;
+    }
+
+    std::vector<const void *> data_ptrs(n_tensors);
+    for (size_t i = 0; i < n_tensors; i++) {
+        data_ptrs[i] = (const char *) data + data_offsets[i];
+    }
+
+    ggml_backend_buffer_t buf = NULL;
+    bool same_buft = true;
+
+    for (size_t i = 0; i < n_tensors; i++) {
+        GGML_ASSERT(tensors[i] != NULL);
+        ggml_backend_buffer_t tensor_buf = tensors[i]->view_src ? tensors[i]->view_src->buffer : tensors[i]->buffer;
+        GGML_ASSERT(tensor_buf != NULL && "tensor buffer not set");
+        GGML_ASSERT(tensors[i]->data != NULL && "tensor not allocated");
+        GGML_ASSERT(offsets[i] + sizes[i] <= ggml_nbytes(tensors[i]) && "tensor write out of bounds");
+
+        if (buf == NULL) {
+            buf = tensor_buf;
+        } else if (buf->buft != tensor_buf->buft) {
+            same_buft = false;
+        }
+    }
+
+    if (same_buft && buf->iface.set_tensor_n != NULL) {
+        buf->iface.set_tensor_n(buf, tensors, data_ptrs.data(), offsets, sizes, n_tensors);
+        return;
+    }
+
+    if (!same_buft) {
+        std::vector<std::vector<size_t>> groups;
+        ggml_backend_tensor_buft_groups(n_tensors, [&](size_t i) {
+            return tensors[i]->view_src ? tensors[i]->view_src->buffer : tensors[i]->buffer;
+        }, groups);
+
+        for (size_t g = 0; g < groups.size(); g++) {
+            std::vector<ggml_tensor *> g_tensors;
+            std::vector<size_t> g_offsets;
+            std::vector<size_t> g_sizes;
+            std::vector<size_t> g_data_offsets;
+            g_tensors.reserve(groups[g].size());
+            g_offsets.reserve(groups[g].size());
+            g_sizes.reserve(groups[g].size());
+            g_data_offsets.reserve(groups[g].size());
+
+            for (size_t i : groups[g]) {
+                g_tensors.push_back(tensors[i]);
+                g_offsets.push_back(offsets[i]);
+                g_sizes.push_back(sizes[i]);
+                g_data_offsets.push_back(data_offsets[i]);
+            }
+
+            ggml_backend_tensor_set_n(g_tensors.data(), data, g_offsets.data(), g_sizes.data(), g_data_offsets.data(), g_tensors.size());
+        }
+        return;
+    }
+
+    size_t base_offset = 0, size = 0, stride_tensor = 0, stride_data = 0;
+    if (ggml_backend_tensor_batch_is_2d(tensors, offsets, sizes, data_offsets, n_tensors,
+                &base_offset, &size, &stride_tensor, &stride_data)) {
+        ggml_backend_tensor_set_2d(tensors[0], data_ptrs[0], base_offset, size,
+                n_tensors, stride_tensor, stride_data);
+        return;
+    }
+
+    for (size_t i = 0; i < n_tensors; i++) {
+        ggml_backend_tensor_set(tensors[i], data_ptrs[i], offsets[i], sizes[i]);
+    }
+}
+
+void ggml_backend_tensor_get_n(const struct ggml_tensor * const * tensors, void * data,
+        const size_t * offsets, const size_t * sizes, const size_t * data_offsets, size_t n_tensors) {
+    if (n_tensors == 0) {
+        return;
+    }
+
+    std::vector<void *> data_ptrs(n_tensors);
+    for (size_t i = 0; i < n_tensors; i++) {
+        data_ptrs[i] = (char *) data + data_offsets[i];
+    }
+
+    ggml_backend_buffer_t buf = NULL;
+    bool same_buft = true;
+
+    for (size_t i = 0; i < n_tensors; i++) {
+        GGML_ASSERT(tensors[i] != NULL);
+        ggml_backend_buffer_t tensor_buf = tensors[i]->view_src ? tensors[i]->view_src->buffer : tensors[i]->buffer;
+        GGML_ASSERT(tensor_buf != NULL && "tensor buffer not set");
+        GGML_ASSERT(tensors[i]->data != NULL && "tensor not allocated");
+        GGML_ASSERT(offsets[i] + sizes[i] <= ggml_nbytes(tensors[i]) && "tensor read out of bounds");
+
+        if (buf == NULL) {
+            buf = tensor_buf;
+        } else if (buf->buft != tensor_buf->buft) {
+            same_buft = false;
+        }
+    }
+
+    if (same_buft && buf->iface.get_tensor_n != NULL) {
+        buf->iface.get_tensor_n(buf, tensors, data_ptrs.data(), offsets, sizes, n_tensors);
+        return;
+    }
+
+    if (!same_buft) {
+        std::vector<std::vector<size_t>> groups;
+        ggml_backend_tensor_buft_groups(n_tensors, [&](size_t i) {
+            return tensors[i]->view_src ? tensors[i]->view_src->buffer : tensors[i]->buffer;
+        }, groups);
+
+        for (size_t g = 0; g < groups.size(); g++) {
+            std::vector<const ggml_tensor *> g_tensors;
+            std::vector<size_t> g_offsets;
+            std::vector<size_t> g_sizes;
+            std::vector<size_t> g_data_offsets;
+            g_tensors.reserve(groups[g].size());
+            g_offsets.reserve(groups[g].size());
+            g_sizes.reserve(groups[g].size());
+            g_data_offsets.reserve(groups[g].size());
+
+            for (size_t i : groups[g]) {
+                g_tensors.push_back(tensors[i]);
+                g_offsets.push_back(offsets[i]);
+                g_sizes.push_back(sizes[i]);
+                g_data_offsets.push_back(data_offsets[i]);
+            }
+
+            ggml_backend_tensor_get_n(g_tensors.data(), data, g_offsets.data(), g_sizes.data(), g_data_offsets.data(), g_tensors.size());
+        }
+        return;
+    }
+
+    size_t base_offset = 0, size = 0, stride_tensor = 0, stride_data = 0;
+    if (ggml_backend_tensor_batch_is_2d(tensors, offsets, sizes, data_offsets, n_tensors,
+                &base_offset, &size, &stride_tensor, &stride_data)) {
+        ggml_backend_tensor_get_2d(tensors[0], data_ptrs[0], base_offset, size,
+                n_tensors, stride_tensor, stride_data);
+        return;
+    }
+
+    for (size_t i = 0; i < n_tensors; i++) {
+        ggml_backend_tensor_get(tensors[i], data_ptrs[i], offsets[i], sizes[i]);
+    }
+}
+
 void ggml_backend_tensor_memset(struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size) {
     GGML_ASSERT(tensor);
     ggml_backend_buffer_t buf = tensor->view_src ? tensor->view_src->buffer : tensor->buffer;
@@ -648,6 +852,59 @@ void ggml_backend_tensor_copy(const struct ggml_tensor * src, struct ggml_tensor
     }
 }
 
+void ggml_backend_tensor_copy_n(const struct ggml_tensor * const * srcs, struct ggml_tensor * const * dsts, size_t n_tensors) {
+    if (n_tensors == 0) {
+        return;
+    }
+
+    for (size_t i = 0; i < n_tensors; ++i) {
+        GGML_ASSERT(ggml_are_same_layout(srcs[i], dsts[i]) && "cannot copy tensors with different layouts");
+    }
+
+    ggml_backend_buffer_t buf = dsts[0]->view_src ? dsts[0]->view_src->buffer : dsts[0]->buffer;
+    bool same_buft = true;
+    for (size_t i = 0; i < n_tensors; ++i) {
+        ggml_backend_buffer_t dst_buf = dsts[i]->view_src ? dsts[i]->view_src->buffer : dsts[i]->buffer;
+        if (dst_buf->buft != buf->buft) {
+            same_buft = false;
+            break;
+        }
+    }
+
+    // if cpy_tensor_n returns false, it must not have performed any copies
+    if (same_buft && buf->iface.cpy_tensor_n != NULL) {
+        if (buf->iface.cpy_tensor_n(buf, srcs, dsts, n_tensors)) {
+            return;
+        }
+    }
+
+    if (!same_buft) {
+        std::vector<std::vector<size_t>> groups;
+        ggml_backend_tensor_buft_groups(n_tensors, [&](size_t i) {
+            return dsts[i]->view_src ? dsts[i]->view_src->buffer : dsts[i]->buffer;
+        }, groups);
+
+        for (size_t g = 0; g < groups.size(); g++) {
+            std::vector<const ggml_tensor *> g_srcs;
+            std::vector<ggml_tensor *> g_dsts;
+            g_srcs.reserve(groups[g].size());
+            g_dsts.reserve(groups[g].size());
+
+            for (size_t i : groups[g]) {
+                g_srcs.push_back(srcs[i]);
+                g_dsts.push_back(dsts[i]);
+            }
+
+            ggml_backend_tensor_copy_n(g_srcs.data(), g_dsts.data(), g_srcs.size());
+        }
+        return;
+    }
+
+    for (size_t i = 0; i < n_tensors; ++i) {
+        ggml_backend_tensor_copy(srcs[i], dsts[i]);
+    }
+}
+
 void ggml_backend_tensor_copy_async(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst) {
     GGML_ASSERT(ggml_are_same_layout(src, dst) && "cannot copy tensors with different layouts");
 
@@ -853,6 +1110,9 @@ static const struct ggml_backend_buffer_i ggml_backend_multi_buffer_i = {
     /* .cpy_tensor      = */ NULL,
     /* .clear           = */ ggml_backend_multi_buffer_clear,
     /* .reset           = */ NULL,
+    /* .set_tensor_n    = */ NULL,
+    /* .get_tensor_n    = */ NULL,
+    /* .cpy_tensor_n    = */ NULL,
 };
 
 ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers) {
@@ -2532,6 +2792,9 @@ static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_i = {
     /* .cpy_tensor      = */ ggml_backend_cpu_buffer_cpy_tensor,
     /* .clear           = */ ggml_backend_cpu_buffer_clear,
     /* .reset           = */ NULL,
+    /* .set_tensor_n    = */ NULL,
+    /* .get_tensor_n    = */ NULL,
+    /* .cpy_tensor_n    = */ NULL,
 };
 
 static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_from_ptr_i = {
@@ -2546,6 +2809,9 @@ static const struct ggml_backend_buffer_i ggml_backend_cpu_buffer_from_ptr_i = {
     /* .cpy_tensor      = */ ggml_backend_cpu_buffer_cpy_tensor,
     /* .clear           = */ ggml_backend_cpu_buffer_clear,
     /* .reset           = */ NULL,
+    /* .set_tensor_n    = */ NULL,
+    /* .get_tensor_n    = */ NULL,
+    /* .cpy_tensor_n    = */ NULL,
 };
 
 // CPU backend buffer type

ggml/src/ggml-cann/ggml-cann.cpp

@@ -1478,6 +1478,9 @@ static const ggml_backend_buffer_i ggml_backend_cann_buffer_interface = {
     /* .cpy_tensor      = */ ggml_backend_cann_buffer_cpy_tensor,
     /* .clear           = */ ggml_backend_cann_buffer_clear,
     /* .reset           = */ NULL,
+    /* .set_tensor_n    = */ NULL,
+    /* .get_tensor_n    = */ NULL,
+    /* .cpy_tensor_n    = */ NULL,
 };
 
 /**

ggml/src/ggml-cpu/amx/amx.cpp

@@ -116,6 +116,9 @@ static ggml_backend_buffer_i ggml_backend_amx_buffer_interface = {
     /* .cpy_tensor      = */ nullptr,
     /* .clear           = */ ggml_backend_amx_buffer_clear,
     /* .reset           = */ nullptr,
+    /* .set_tensor_n    = */ NULL,
+    /* .get_tensor_n    = */ NULL,
+    /* .cpy_tensor_n    = */ NULL,
 };
 
 static const char * ggml_backend_amx_buffer_type_get_name(ggml_backend_buffer_type_t buft) {

ggml/src/ggml-cpu/common.h

@@ -43,6 +43,14 @@ static inline float f32_to_f32(float x) {
     return x;
 }
 
+static inline float f8_e4m3_to_f32(ggml_fp8_e4m3_t x) {
+    return ggml_f8_e4m3_to_fp32(x.bits);
+}
+
+static inline ggml_fp8_e4m3_t f32_to_f8_e4m3(float x) {
+    return { ggml_fp32_to_f8_e4m3(x) };
+}
+
 // TODO - merge this into the traits table, after using row-based conversions
 template <class T>
 struct type_conversion_table;
@@ -65,6 +73,12 @@ struct type_conversion_table<ggml_bf16_t> {
     static constexpr ggml_bf16_t (*from_f32)(float) = f32_to_bf16;
 };
 
+template <>
+struct type_conversion_table<ggml_fp8_e4m3_t> {
+    static constexpr float (*to_f32)(ggml_fp8_e4m3_t) = f8_e4m3_to_f32;
+    static constexpr ggml_fp8_e4m3_t (*from_f32)(float) = f32_to_f8_e4m3;
+};
+
 template <>
 struct type_conversion_table<int32_t> {
     static constexpr float (*to_f32)(int32_t) = i32_to_f32;

ggml/src/ggml-cpu/ggml-cpu.c

@@ -86,6 +86,9 @@ float ggml_table_f32_e8m0_half[1 << 8];
 // precomputed f32 table for ue4m3 (1 KB) (simd-mappings.h)
 float ggml_table_f32_ue4m3[1 << 8];
 
+// Signed FP8 needs its own table: UE4M3 ignores the sign, halves values for NVFP4, and maps NaN to zero.
+float ggml_table_f32_f8_e4m3[1 << 8];
+
 #if defined(__ARM_ARCH)
 struct ggml_arm_arch_features_type {
     int sve_cnt;
@@ -296,6 +299,12 @@ static const struct ggml_type_traits_cpu type_traits_cpu[GGML_TYPE_COUNT] = {
         .vec_dot_type             = GGML_TYPE_Q8_0,
         .nrows                    = 1,
     },
+    [GGML_TYPE_F8_E4M3] = {
+        .from_float               = quantize_row_f8_e4m3,
+        .vec_dot                  = ggml_vec_dot_f8_e4m3_f32,
+        .vec_dot_type             = GGML_TYPE_F32,
+        .nrows                    = 1,
+    },
     [GGML_TYPE_Q2_K] = {
         .from_float               = quantize_row_q2_K,
         .vec_dot                  = ggml_vec_dot_q2_K_q8_K,
@@ -2876,7 +2885,7 @@ struct ggml_cplan ggml_graph_plan(
                 case GGML_OP_ADD_ID:
                 case GGML_OP_ADD1:
                     {
-                        if (ggml_is_quantized(node->src[0]->type)) {
+                        if (ggml_is_quantized(node->src[0]->type) || node->src[0]->type == GGML_TYPE_F8_E4M3) {
                             cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
                         }
                     } break;
@@ -2924,7 +2933,7 @@ struct ggml_cplan ggml_graph_plan(
                     } break;
                 case GGML_OP_OUT_PROD:
                     {
-                        if (ggml_is_quantized(node->src[0]->type) ||
+                        if (ggml_is_quantized(node->src[0]->type) || node->src[0]->type == GGML_TYPE_F8_E4M3 ||
                             node->src[0]->type == GGML_TYPE_F16) {
                             cur = ggml_type_size(GGML_TYPE_F32) * node->src[0]->ne[0] * n_tasks;
                         }
@@ -3908,6 +3917,7 @@ void ggml_cpu_init(void) {
             // initialize UE4M3 table (256 entries)
             for (int i = 0; i < (1 << 8); ++i) {
                 ggml_table_f32_ue4m3[i] = ggml_ue4m3_to_fp32(i);
+                ggml_table_f32_f8_e4m3[i] = ggml_f8_e4m3_to_fp32(i);
             }
 
             const uint64_t t_end = ggml_time_us(); UNUSED(t_end);

ggml/src/ggml-cpu/ops.cpp

@@ -537,26 +537,36 @@ void ggml_compute_forward_dup(
     switch (src0->type) {
         case GGML_TYPE_F16:
             {
-                /**/ if (dst->type == GGML_TYPE_F16)  ggml_compute_forward_dup_flt<ggml_fp16_t, ggml_fp16_t>(params, dst);
-                else if (dst->type == GGML_TYPE_BF16) ggml_compute_forward_dup_flt<ggml_fp16_t, ggml_bf16_t>(params, dst);
-                else if (dst->type == GGML_TYPE_F32)  ggml_compute_forward_dup_flt<ggml_fp16_t, float      >(params, dst);
+                /**/ if (dst->type == GGML_TYPE_F16)     ggml_compute_forward_dup_flt<ggml_fp16_t, ggml_fp16_t>(params, dst);
+                else if (dst->type == GGML_TYPE_BF16)    ggml_compute_forward_dup_flt<ggml_fp16_t, ggml_bf16_t>(params, dst);
+                else if (dst->type == GGML_TYPE_F32)     ggml_compute_forward_dup_flt<ggml_fp16_t, float      >(params, dst);
+                else if (dst->type == GGML_TYPE_F8_E4M3) ggml_compute_forward_dup_flt<ggml_fp16_t, ggml_fp8_e4m3_t>(params, dst);
                 else ggml_compute_forward_dup_to_q<ggml_fp16_t>(params, dst);
             } break;
         case GGML_TYPE_BF16:
             {
-                /**/ if (dst->type == GGML_TYPE_F16)  ggml_compute_forward_dup_flt<ggml_bf16_t, ggml_fp16_t>(params, dst);
-                else if (dst->type == GGML_TYPE_BF16) ggml_compute_forward_dup_flt<ggml_bf16_t, ggml_bf16_t>(params, dst);
-                else if (dst->type == GGML_TYPE_F32)  ggml_compute_forward_dup_flt<ggml_bf16_t, float      >(params, dst);
+                /**/ if (dst->type == GGML_TYPE_F16)     ggml_compute_forward_dup_flt<ggml_bf16_t, ggml_fp16_t>(params, dst);
+                else if (dst->type == GGML_TYPE_BF16)    ggml_compute_forward_dup_flt<ggml_bf16_t, ggml_bf16_t>(params, dst);
+                else if (dst->type == GGML_TYPE_F32)     ggml_compute_forward_dup_flt<ggml_bf16_t, float      >(params, dst);
+                else if (dst->type == GGML_TYPE_F8_E4M3) ggml_compute_forward_dup_flt<ggml_bf16_t, ggml_fp8_e4m3_t>(params, dst);
                 else ggml_compute_forward_dup_to_q<ggml_bf16_t>(params, dst);
             } break;
         case GGML_TYPE_F32:
             {
-                /**/ if (dst->type == GGML_TYPE_F16)  ggml_compute_forward_dup_flt<float, ggml_fp16_t>(params, dst);
-                else if (dst->type == GGML_TYPE_BF16) ggml_compute_forward_dup_flt<float, ggml_bf16_t>(params, dst);
-                else if (dst->type == GGML_TYPE_F32)  ggml_compute_forward_dup_flt<float, float      >(params, dst);
-                else if (dst->type == GGML_TYPE_I32)  ggml_compute_forward_dup_flt<float, int32_t    >(params, dst);
+                /**/ if (dst->type == GGML_TYPE_F16)     ggml_compute_forward_dup_flt<float, ggml_fp16_t>(params, dst);
+                else if (dst->type == GGML_TYPE_BF16)    ggml_compute_forward_dup_flt<float, ggml_bf16_t>(params, dst);
+                else if (dst->type == GGML_TYPE_F32)     ggml_compute_forward_dup_flt<float, float      >(params, dst);
+                else if (dst->type == GGML_TYPE_F8_E4M3) ggml_compute_forward_dup_flt<float, ggml_fp8_e4m3_t>(params, dst);
+                else if (dst->type == GGML_TYPE_I32)     ggml_compute_forward_dup_flt<float, int32_t    >(params, dst);
                 else ggml_compute_forward_dup_to_q<float>(params, dst);
             } break;
+        case GGML_TYPE_F8_E4M3:
+            {
+                /**/ if (dst->type == GGML_TYPE_F16)  ggml_compute_forward_dup_flt<ggml_fp8_e4m3_t, ggml_fp16_t>(params, dst);
+                else if (dst->type == GGML_TYPE_BF16) ggml_compute_forward_dup_flt<ggml_fp8_e4m3_t, ggml_bf16_t>(params, dst);
+                else if (dst->type == GGML_TYPE_F32)  ggml_compute_forward_dup_flt<ggml_fp8_e4m3_t, float      >(params, dst);
+                else GGML_ABORT("not implemented");
+            } break;
         case GGML_TYPE_I32:
             {
                 if (dst->type == GGML_TYPE_F32) ggml_compute_forward_dup_flt<int32_t, float>(params, dst);
@@ -605,8 +615,9 @@ static void ggml_compute_forward_add_q_f32(
     GGML_ASSERT(nb1 <= nb2);
     GGML_ASSERT(nb2 <= nb3);
 
-    GGML_ASSERT(ggml_is_quantized(src0->type));
+    GGML_ASSERT(ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F8_E4M3);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
+    GGML_ASSERT(ne00 % ggml_blck_size(type) == 0);
 
     // rows per thread
     const int dr = (nr + nth - 1)/nth;
@@ -636,8 +647,6 @@ static void ggml_compute_forward_add_q_f32(
         float * src1_row = (float *)((char *) src1->data + (i11*nb11 + i12*nb12 + i13*nb13));
         void  * dst_row  = (void *) ((char *)  dst->data + ( i1*nb1  +  i2*nb2  +  i3*nb3));
 
-        assert(ne00 % 32 == 0);
-
         // unquantize row from src0 to temp buffer
         dequantize_row_q(src0_row, wdata, ne00);
         // add src1
@@ -673,6 +682,7 @@ void ggml_compute_forward_add(
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_MXFP4:
         case GGML_TYPE_NVFP4:
+        case GGML_TYPE_F8_E4M3:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -955,7 +965,7 @@ static void ggml_compute_forward_add1_q_f32(
     GGML_ASSERT(nb1 <= nb2);
     GGML_ASSERT(nb2 <= nb3);
 
-    GGML_ASSERT(ggml_is_quantized(src0->type));
+    GGML_ASSERT(ggml_is_quantized(src0->type) || src0->type == GGML_TYPE_F8_E4M3);
     GGML_ASSERT(dst->type == src0->type);
     GGML_ASSERT(src1->type == GGML_TYPE_F32);
 
@@ -1130,6 +1140,7 @@ void ggml_compute_forward_add1(
         case GGML_TYPE_Q8_1:
         case GGML_TYPE_MXFP4:
         case GGML_TYPE_NVFP4:
+        case GGML_TYPE_F8_E4M3:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -1261,6 +1272,7 @@ void ggml_compute_forward_acc(
         case GGML_TYPE_Q8_1:
         case GGML_TYPE_MXFP4:
         case GGML_TYPE_NVFP4:
+        case GGML_TYPE_F8_E4M3:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -4804,6 +4816,7 @@ void ggml_compute_forward_out_prod(
         case GGML_TYPE_Q8_0:
         case GGML_TYPE_MXFP4:
         case GGML_TYPE_NVFP4:
+        case GGML_TYPE_F8_E4M3:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -5129,6 +5142,7 @@ void ggml_compute_forward_set(
         case GGML_TYPE_Q8_1:
         case GGML_TYPE_MXFP4:
         case GGML_TYPE_NVFP4:
+        case GGML_TYPE_F8_E4M3:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -5354,6 +5368,7 @@ void ggml_compute_forward_get_rows(
         case GGML_TYPE_Q8_1:
         case GGML_TYPE_MXFP4:
         case GGML_TYPE_NVFP4:
+        case GGML_TYPE_F8_E4M3:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
@@ -6114,6 +6129,7 @@ void ggml_compute_forward_clamp(
         case GGML_TYPE_Q8_1:
         case GGML_TYPE_MXFP4:
         case GGML_TYPE_NVFP4:
+        case GGML_TYPE_F8_E4M3:
         case GGML_TYPE_Q2_K:
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:

ggml/src/ggml-cpu/quants.c

@@ -3,6 +3,7 @@
 
 #include "ggml-cpu-impl.h"
 #include "simd-mappings.h"
+#include "vec.h"
 #include "ggml-quants.h"
 #include "quants.h"
 
@@ -62,6 +63,10 @@ void quantize_row_nvfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, i
     quantize_row_nvfp4_ref(x, y, k);
 }
 
+void quantize_row_f8_e4m3(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+    quantize_row_f8_e4m3_ref(x, y, k);
+}
+
 //
 // 2-6 bit quantization in super-blocks
 //
@@ -362,6 +367,49 @@ void ggml_vec_dot_nvfp4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs,
     *s = sumf;
 }
 
+void ggml_vec_dot_f8_e4m3_f32(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+    assert(nrc == 1);
+    UNUSED(nrc);
+    UNUSED(bx);
+    UNUSED(by);
+    UNUSED(bs);
+
+    const ggml_fp8_e4m3_t * GGML_RESTRICT x = vx;
+    const float * GGML_RESTRICT y = vy;
+
+    float sumf = 0.0f;
+    int pos = 0;
+#if defined(GGML_SIMD) && !defined(__ARM_FEATURE_SVE) && !defined(__riscv_v_intrinsic)
+    GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
+    const int np = n - n % GGML_F32_STEP;
+    for (; pos < np; pos += GGML_F32_STEP) {
+        for (int j = 0; j < GGML_F32_ARR; ++j) {
+            const int offset = pos + j*GGML_F32_EPR;
+            GGML_F32_VEC ax;
+#if defined(__AVX512F__) && defined(__AVX512BW__)
+            const __m128i codes = _mm_loadu_si128((const __m128i *) (x + offset));
+            ax = _mm512_i32gather_ps(_mm512_cvtepu8_epi32(codes), ggml_table_f32_f8_e4m3, sizeof(float));
+#elif defined(__AVX2__) && !defined(__AVX512F__)
+            const __m128i codes = _mm_loadl_epi64((const __m128i *) (x + offset));
+            ax = _mm256_i32gather_ps(ggml_table_f32_f8_e4m3, _mm256_cvtepu8_epi32(codes), sizeof(float));
+#else
+            float values[GGML_F32_EPR];
+            for (int k = 0; k < GGML_F32_EPR; ++k) {
+                values[k] = ggml_table_f32_f8_e4m3[x[offset + k].bits];
+            }
+            ax = GGML_F32_VEC_LOAD(values);
+#endif
+            sum[j] = GGML_F32_VEC_FMA(sum[j], ax, GGML_F32_VEC_LOAD(y + offset));
+        }
+    }
+    GGML_F32_VEC_REDUCE(sumf, sum);
+#endif
+    for (; pos < n; ++pos) {
+        sumf += ggml_table_f32_f8_e4m3[x[pos].bits] * y[pos];
+    }
+    *s = sumf;
+}
+
 void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
     const int qk = QK8_0;
     const int nb = n / qk;

ggml/src/ggml-cpu/quants.h

@@ -23,6 +23,7 @@ void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, in
 
 void quantize_row_mxfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 void quantize_row_nvfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_f8_e4m3(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 
 void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
 void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
@@ -40,6 +41,7 @@ void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y,
 // Dot product
 void ggml_vec_dot_q1_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q2_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_f8_e4m3_f32(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
 void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);

ggml/src/ggml-cpu/simd-mappings.h

@@ -127,6 +127,8 @@ extern float ggml_table_f32_e8m0_half[1 << 8];
 // defined in ggml-cpu.c, initialized in ggml_cpu_init()
 extern float ggml_table_f32_ue4m3[1 << 8];
 
+extern float ggml_table_f32_f8_e4m3[1 << 8];
+
 // Use lookup table for E8M0 on x86 (faster than bit manipulation)
 #if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
 #define GGML_CPU_E8M0_TO_FP32_HALF(x) ggml_table_f32_e8m0_half[(uint8_t)(x)]

ggml/src/ggml-cpu/spacemit/ime.cpp

@@ -1489,6 +1489,9 @@ static const ggml_backend_buffer_i ggml_backend_riscv64_spacemit_buffer_i = {
     /* .cpy_tensor      = */ nullptr,
     /* .clear           = */ ggml_backend_riscv64_spacemit_buffer_clear,
     /* .reset           = */ nullptr,
+    /* .set_tensor_n    = */ NULL,
+    /* .get_tensor_n    = */ NULL,
+    /* .cpy_tensor_n    = */ NULL,
 };
 
 static const char * ggml_backend_cpu_riscv64_spacemit_buffer_type_get_name(ggml_backend_buffer_type_t buft) {

ggml/src/ggml-cuda/common.cuh

@@ -82,6 +82,7 @@
 #define GGML_CUDA_CC_RDNA2      (GGML_CUDA_CC_OFFSET_AMD + 0x1030) // RX 6000, minimum for dp4a
 #define GGML_CUDA_CC_RDNA3      (GGML_CUDA_CC_OFFSET_AMD + 0x1100) // RX 7000, minimum for WMMA
 #define GGML_CUDA_CC_RDNA3_5    (GGML_CUDA_CC_OFFSET_AMD + 0x1150) // AI 370, AI Max 395 laptops.
+#define GGML_CUDA_CC_RDNA4_M    (GGML_CUDA_CC_OFFSET_AMD + 0x1170) // gfx117X
 #define GGML_CUDA_CC_RDNA4      (GGML_CUDA_CC_OFFSET_AMD + 0x1200) // RX 9000
 
 #define GGML_CUDA_CC_IS_AMD(cc)     (cc >= GGML_CUDA_CC_OFFSET_AMD)
@@ -89,8 +90,9 @@
 #define GGML_CUDA_CC_IS_RDNA1(cc)   (cc >= GGML_CUDA_CC_RDNA1 && cc < GGML_CUDA_CC_RDNA2)
 #define GGML_CUDA_CC_IS_RDNA2(cc)   (cc >= GGML_CUDA_CC_RDNA2 && cc < GGML_CUDA_CC_RDNA3)
 #define GGML_CUDA_CC_IS_RDNA3_0(cc) (cc >= GGML_CUDA_CC_RDNA3 && cc < GGML_CUDA_CC_RDNA3_5)
-#define GGML_CUDA_CC_IS_RDNA3_5(cc) (cc >= GGML_CUDA_CC_RDNA3_5 && cc < GGML_CUDA_CC_RDNA4)
+#define GGML_CUDA_CC_IS_RDNA3_5(cc) (cc >= GGML_CUDA_CC_RDNA3_5 && cc < GGML_CUDA_CC_RDNA4_M)
 #define GGML_CUDA_CC_IS_RDNA3(cc)   (GGML_CUDA_CC_IS_RDNA3_0(cc) || GGML_CUDA_CC_IS_RDNA3_5(cc))
+#define GGML_CUDA_CC_IS_RDNA4_M(cc) (cc >= GGML_CUDA_CC_RDNA4_M && cc < GGML_CUDA_CC_RDNA4)
 #define GGML_CUDA_CC_IS_RDNA4(cc)   (cc >= GGML_CUDA_CC_RDNA4)
 #define GGML_CUDA_CC_IS_GCN_APU(cc) ((cc) == GGML_CUDA_CC_GFX909 || (cc) == GGML_CUDA_CC_GFX90C)
 #define GGML_CUDA_CC_IS_GCN(cc)     ((cc > GGML_CUDA_CC_OFFSET_AMD && cc < GGML_CUDA_CC_CDNA1) || GGML_CUDA_CC_IS_GCN_APU(cc))
@@ -277,9 +279,17 @@ static const char * cu_get_error_str(CUresult err) {
 #define AMD_MFMA_AVAILABLE
 #endif // defined(GGML_USE_HIP) && defined(CDNA) && !defined(GGML_HIP_NO_MMQ_MFMA)
 
-#if defined(GGML_USE_HIP) && (defined(RDNA4) || defined(RDNA3))
+#if defined(GGML_USE_HIP) && (defined(RDNA4) || defined(RDNA4_M))
+#define AMD_WMMA_128B_AVAILABLE
+#endif // defined(GGML_USE_HIP) && (defined(RDNA4) || defined(RDNA4_M))
+
+#if defined(GGML_USE_HIP) && defined(RDNA3)
+#define AMD_WMMA_256B_AVAILABLE
+#endif // defined(GGML_USE_HIP) && defined(RDNA3)
+
+#if defined(AMD_WMMA_128B_AVAILABLE) || defined(AMD_WMMA_256B_AVAILABLE)
 #define AMD_WMMA_AVAILABLE
-#endif // defined(GGML_USE_HIP) && defined(RDNA4)
+#endif // defined(AMD_WMMA_128B_AVAILABLE) || defined(AMD_WMMA_256B_AVAILABLE)
 
 // The Volta instructions are in principle available on Turing or newer but they are effectively unusable:
 #if !defined(GGML_USE_HIP) && __CUDA_ARCH__ == GGML_CUDA_CC_VOLTA
@@ -326,7 +336,7 @@ static bool fast_fp16_hardware_available(const int cc) {
 // To be used for feature selection of external libraries, e.g. cuBLAS.
 static bool fp16_mma_hardware_available(const int cc) {
     return (GGML_CUDA_CC_IS_NVIDIA(cc) && cc >= GGML_CUDA_CC_VOLTA) ||
-        GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc) ||
+        GGML_CUDA_CC_IS_CDNA(cc) || GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4_M(cc) || GGML_CUDA_CC_IS_RDNA4(cc) ||
         (GGML_CUDA_CC_IS_MTHREADS(cc) && cc >= GGML_CUDA_CC_QY2);
 }
 
@@ -355,7 +365,7 @@ static bool amd_mfma_available(const int cc) {
 }
 
 static bool amd_wmma_available(const int cc) {
-    return (GGML_CUDA_CC_IS_RDNA4(cc) || GGML_CUDA_CC_IS_RDNA3(cc));
+    return GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4_M(cc) || GGML_CUDA_CC_IS_RDNA4(cc);
 }
 
 static bool volta_mma_available(const int cc) {
@@ -713,7 +723,7 @@ static __device__ __forceinline__ int ggml_cuda_dp4a(const int a, const int b, i
 #if defined(GGML_USE_HIP)
 #if defined(CDNA) || defined(RDNA2) || defined(__gfx906__)
     c = __builtin_amdgcn_sdot4(a, b, c, false);
-#elif defined(RDNA3) || defined(RDNA4)
+#elif defined(RDNA3) || defined(RDNA4_M) || defined(RDNA4)
     c = __builtin_amdgcn_sudot4( true, a, true, b, c, false);
 #elif defined(RDNA1) || defined(__gfx900__)
     int tmp1;
@@ -758,9 +768,9 @@ static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const float2 v
     acc += v.y*u.y;
 }
 
-#if defined(GGML_USE_HIP) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4) || defined(__gfx906__) || defined(CDNA))
+#if defined(GGML_USE_HIP) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4_M) || defined(RDNA4) || defined(__gfx906__) || defined(CDNA))
 #define V_DOT2_F32_F16_AVAILABLE
-#endif // defined(GGML_USE_HIP) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4) || defined(__gfx906__) || defined(CDNA))
+#endif // defined(GGML_USE_HIP) && (defined(RDNA2) || defined(RDNA3) || defined(RDNA4_M) || defined(RDNA4) || defined(__gfx906__) || defined(CDNA))
 
 static __device__ __forceinline__ void ggml_cuda_mad(float & acc, const half2 v, const half2 u) {
 #ifdef V_DOT2_F32_F16_AVAILABLE

ggml/src/ggml-cuda/fattn-mma-f16.cuh

@@ -1255,7 +1255,7 @@ static __device__ __forceinline__ void flash_attn_ext_f16_process_tile(
     T_B_KQ    Q_B[(Q_in_reg ? DKQ/(2*T_B_KQ::J) : 1)];
 #if defined(TURING_MMA_AVAILABLE)
     T_C_VKQ VKQ_C[cols_per_warp == 8 ? DV/T_C_VKQ::I : DV/(2*T_C_VKQ::J)];
-#elif defined(AMD_WMMA_AVAILABLE) && defined(RDNA3)
+#elif defined(AMD_WMMA_256B_AVAILABLE)
     T_C_VKQ VKQ_C[DV % 32 != 0       ? DV/T_C_VKQ::J : DV/(2*T_C_VKQ::J)];
 #elif defined(AMD_MFMA_AVAILABLE)
     T_C_VKQ VKQ_C[                                     DV/T_C_VKQ::J];
@@ -1866,7 +1866,7 @@ static __global__ void flash_attn_ext_f16(
 #endif // defined(AMD_WMMA_AVAILABLE)
 
 #if defined(AMD_MFMA_AVAILABLE)
-    if (ncols1*ncols2 < 16 || (DKQ > 256 && ncols1*ncols2 < 32)) {
+    if (ncols1*ncols2 < 16 || (DKQ > 256 && ncols1*ncols2 < 32 && !use_sparse)) {
         NO_DEVICE_CODE;
         return;
     }
@@ -2054,7 +2054,7 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
     bool use_sparse = false;
     if (logit_softcap == 0.0f) {
         constexpr bool use_logit_softcap = false;
-#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+#if !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(cc, dst, ncols1, ncols2)) {
                 constexpr bool use_sparse_kernel = true;
@@ -2077,7 +2077,7 @@ void ggml_cuda_flash_attn_ext_mma_f16_case(ggml_backend_cuda_context & ctx, ggml
                 }
             }
         } else
-#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+#endif // !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>;

ggml/src/ggml-cuda/fattn.cu

@@ -5,7 +5,8 @@
 #include "fattn-vec.cuh"
 #include "fattn.cuh"
 
-#if !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+#if !defined(GGML_USE_MUSA)
+
 // one list per group of ncols1 queries: a column is selected if any query of the group can see it
 template <int ncols1, bool oob>
 __launch_bounds__(256, 1)
@@ -14,10 +15,20 @@ static __global__ void flash_attn_mask_to_sparse_indices(
         const int n_kv_max, const int64_t s31, const int64_t s33) {
     ggml_cuda_pdl_sync();
 
+    constexpr int warp_size = ggml_cuda_get_physical_warp_size();
     constexpr int values_per_lane = 8;
+    using lane_mask_t = std::conditional_t<warp_size == 64, uint64_t, uint32_t>;
+    constexpr uint64_t ballot_mask = ~(lane_mask_t)0;
+    const auto popc = [](const lane_mask_t mask) {
+        if constexpr (warp_size == 64) {
+            return __popcll(mask);
+        } else {
+            return __popc(mask);
+        }
+    };
     const int tid      = threadIdx.x;
-    const int warp     = tid / WARP_SIZE;
-    const int lane     = tid % WARP_SIZE;
+    const int warp     = tid / warp_size;
+    const int lane     = tid % warp_size;
     const int sequence = blockIdx.y;
     const int group    = blockIdx.x;
 
@@ -27,7 +38,7 @@ static __global__ void flash_attn_mask_to_sparse_indices(
     const half * mask = mask_ptr + sequence*s33 + q0*s31;
     int32_t * indices = indices_ptr + (int64_t(sequence)*gridDim.x + group)*n_kv_max;
 
-    __shared__ int warp_offsets[256/WARP_SIZE];
+    __shared__ int warp_offsets[256/warp_size];
     __shared__ int row_count;
     __shared__ int chunk_count;
 
@@ -37,11 +48,11 @@ static __global__ void flash_attn_mask_to_sparse_indices(
     __syncthreads();
 
     for (int i0 = 0; i0 < ne30; i0 += blockDim.x*values_per_lane) {
-        uint32_t selected_warp[values_per_lane];
+        lane_mask_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 int i = i0 + (warp*values_per_lane + item)*warp_size + lane;
             bool selected = false;
             if (i < ne30) {
 #pragma unroll
@@ -49,8 +60,8 @@ static __global__ void flash_attn_mask_to_sparse_indices(
                     selected |= (!oob || q < q1 - q0) && isfinite(__half2float(mask[q*s31 + i]));
                 }
             }
-            selected_warp[item] = __ballot_sync(0xFFFFFFFF, selected);
-            warp_count += __popc(selected_warp[item]);
+            selected_warp[item] = (lane_mask_t)__ballot_sync(ballot_mask, (int) selected);
+            warp_count += popc(selected_warp[item]);
         }
 
         if (lane == 0) {
@@ -61,7 +72,7 @@ static __global__ void flash_attn_mask_to_sparse_indices(
         if (tid == 0) {
             int offset = 0;
 #pragma unroll
-            for (int iw = 0; iw < 256/WARP_SIZE; ++iw) {
+            for (int iw = 0; iw < 256/warp_size; ++iw) {
                 const int count = warp_offsets[iw];
                 warp_offsets[iw] = offset;
                 offset += count;
@@ -70,16 +81,16 @@ static __global__ void flash_attn_mask_to_sparse_indices(
         }
         __syncthreads();
 
-        const uint32_t lane_mask = lane == 0 ? 0 : (1u << lane) - 1;
+        const lane_mask_t lane_mask = lane == 0 ? 0 : ((lane_mask_t)1 << 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) {
+            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] & ((lane_mask_t)1 << lane)) && dst < n_kv_max) {
                 indices[dst] = i;
             }
-            warp_item_offset += __popc(selected_warp[item]);
+            warp_item_offset += popc(selected_warp[item]);
         }
         __syncthreads();
 
@@ -101,13 +112,13 @@ static __global__ void flash_attn_mask_to_sparse_indices(
     // the dependent grid reads indices, signal once the row is complete
     ggml_cuda_pdl_lc();
 }
-#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+#endif // !defined(GGML_USE_MUSA)
 
 void ggml_cuda_flash_attn_ext_compact_mask(
         const ggml_tensor * mask, int32_t * indices, int32_t * counts, int32_t n_queries, int32_t ncols1, int32_t n_kv_max, cudaStream_t stream) {
-#if defined(GGML_USE_HIP) || defined(GGML_USE_MUSA)
+#ifdef GGML_USE_MUSA
     GGML_UNUSED_VARS(mask, indices, counts, n_queries, ncols1, n_kv_max, stream);
-    GGML_ABORT("sparse flash attention is only supported on NVIDIA CUDA");
+    GGML_ABORT("sparse flash attention is not supported on MUSA");
 #else
     const int64_t s31 = mask->nb[1] / sizeof(half);
     const int64_t s33 = mask->nb[3] / sizeof(half);
@@ -122,16 +133,17 @@ void ggml_cuda_flash_attn_ext_compact_mask(
     ggml_cuda_kernel_launch(kernel, launch_params,
         (const half *) mask->data, indices, counts, int(mask->ne[0]), n_queries, n_kv_max, s31, s33);
     CUDA_CHECK(cudaGetLastError());
-#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+#endif // !defined(GGML_USE_MUSA)
 }
 
 bool ggml_cuda_flash_attn_ext_mma_f16_shall_use_sparse(const int cc, const ggml_tensor * dst, const int ncols1, const int ncols2) {
-#if defined(GGML_USE_HIP) || defined(GGML_USE_MUSA)
+#ifdef GGML_USE_MUSA
     GGML_UNUSED_VARS(cc, dst, ncols1, ncols2);
     return false;
 #else
     const ggml_tensor * Q    = dst->src[0];
     const ggml_tensor * K    = dst->src[1];
+    const ggml_tensor * V    = dst->src[2];
     const ggml_tensor * mask = dst->src[3];
 
     float max_bias = 0.0f;
@@ -139,16 +151,21 @@ bool ggml_cuda_flash_attn_ext_mma_f16_shall_use_sparse(const int cc, const ggml_
     memcpy(&max_bias,      (const float *) dst->op_params + 1, sizeof(float));
     memcpy(&logit_softcap, (const float *) dst->op_params + 2, sizeof(float));
 
+    // On CDNA only the DKQ=576 kernel variant with ncols2=16 is compiled with sparse support.
+    if (amd_mfma_available(cc) && !ggml_cuda_flash_attn_ext_mma_f16_may_use_sparse(int(Q->ne[0]), int(V->ne[0]), 1, 16)) {
+        return false;
+    }
+
     const int32_t n_kv_max = ggml_get_op_params_i32(dst, 4);
 
     // the dense kernel handles up to 64/ncols2 queries per K/V pass, the single-query gather has to beat that
     const int64_t n_gather = (ncols1 == 1 ? std::min<int64_t>(Q->ne[1], 64/ncols2) : ncols1) * (int64_t) n_kv_max;
 
-    return GGML_CUDA_CC_IS_NVIDIA(cc) && turing_mma_available(cc) &&
+    return ((GGML_CUDA_CC_IS_NVIDIA(cc) && turing_mma_available(cc)) || amd_mfma_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, 2*n_gather);
-#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+        K->ne[1] >= std::max<int64_t>(4096, GGML_CUDA_CC_IS_NVIDIA(cc) ? 2*n_gather : 2LL*n_kv_max);
+#endif // !defined(GGML_USE_MUSA)
 }
 
 template <int DKQ, int DV, int ncols2>
@@ -156,7 +173,7 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ggml_backend_cuda_con
     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 !defined(GGML_USE_MUSA)
     if constexpr (ggml_cuda_flash_attn_ext_mma_f16_may_use_sparse(DKQ, DV, 1, ncols2)) {
         // a sparse variant at the full tile width gathers the union of its queries once, prefer it for large batches
         constexpr bool has_wide_sparse = ggml_cuda_flash_attn_ext_mma_f16_may_use_sparse(DKQ, DV, 64/ncols2, ncols2);
@@ -165,7 +182,7 @@ static void ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1(ggml_backend_cuda_con
             return;
         }
     }
-#endif // !defined(GGML_USE_HIP) && !defined(GGML_USE_MUSA)
+#endif // !defined(GGML_USE_MUSA)
 
     if constexpr (ncols2 <= 8) {
         if (turing_mma_available(cc) && Q->ne[1] <= 8/ncols2) {
@@ -365,7 +382,9 @@ static void ggml_cuda_flash_attn_ext_mma_f16(ggml_backend_cuda_context & ctx, gg
 
             GGML_ASSERT(Q->ne[2] % K->ne[2] == 0);
             const int gqa_ratio = Q->ne[2] / K->ne[2];
-            if (gqa_ratio == 20 && GGML_CUDA_CC_IS_NVIDIA(cc)) { // GLM 4.7 Flash
+            if (amd_mfma_available(cc) && ggml_cuda_flash_attn_ext_mma_f16_shall_use_sparse(cc, dst, 1, 16)) {
+                ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst);
+            } else if (gqa_ratio == 20 && GGML_CUDA_CC_IS_NVIDIA(cc)) { // GLM 4.7 Flash
                 if (cc >= GGML_CUDA_CC_DGX_SPARK) {
                     if (Q->ne[1] <= 8) {
                         ggml_cuda_flash_attn_ext_mma_f16_switch_ncols1<576, 512, 16>(ctx, dst);
@@ -679,6 +698,11 @@ static best_fattn_kernel ggml_cuda_get_best_fattn_kernel(const int device, const
         return BEST_FATTN_KERNEL_MMA_F16;
     }
 
+    // On CDNA, the sparse MMA kernel is used for DKQ=576 regardless of the head size heuristics.
+    if (amd_mfma_available(cc) && ggml_cuda_flash_attn_ext_mma_f16_shall_use_sparse(cc, dst, 1, 16)) {
+        return BEST_FATTN_KERNEL_MMA_F16;
+    }
+
     // AMD MFMA needs a certain minimum batch size to outscale the tile kernel for large head sizes.
     if (amd_mfma_available(cc) && Q->ne[0] != 40 && Q->ne[0] != 72) {
         if ((Q->ne[0] <= 64 && Q->ne[1] * gqa_ratio_eff > 8)) {

ggml/src/ggml-cuda/ggml-cuda.cu

@@ -861,6 +861,46 @@ static void ggml_backend_cuda_buffer_get_tensor_2d(ggml_backend_buffer_t buffer,
     CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
 }
 
+static void ggml_backend_cuda_buffer_set_tensor_n(ggml_backend_buffer_t buffer, struct ggml_tensor * const * tensors, const void * const * data, const size_t * offsets, const size_t * sizes, size_t n_tensors) {
+    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
+
+    ggml_cuda_set_device(ctx->device);
+
+    for (size_t i = 0; i < n_tensors; i++) {
+        if (sizes[i] == 0) {
+            continue;
+        }
+        CUDA_CHECK(cudaMemcpyAsync(
+            (char *) tensors[i]->data + offsets[i],
+            data[i],
+            sizes[i],
+            cudaMemcpyHostToDevice,
+            cudaStreamPerThread));
+    }
+
+    CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
+}
+
+static void ggml_backend_cuda_buffer_get_tensor_n(ggml_backend_buffer_t buffer, const struct ggml_tensor * const * tensors, void * const * data, const size_t * offsets, const size_t * sizes, size_t n_tensors) {
+    ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
+
+    ggml_cuda_set_device(ctx->device);
+
+    for (size_t i = 0; i < n_tensors; i++) {
+        if (sizes[i] == 0) {
+            continue;
+        }
+        CUDA_CHECK(cudaMemcpyAsync(
+            data[i],
+            (const char *) tensors[i]->data + offsets[i],
+            sizes[i],
+            cudaMemcpyDeviceToHost,
+            cudaStreamPerThread));
+    }
+
+    CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
+}
+
 static bool ggml_backend_cuda_buffer_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
     if (ggml_backend_buffer_is_cuda(src->buffer)) {
         ggml_backend_cuda_buffer_context * src_ctx = (ggml_backend_cuda_buffer_context *)src->buffer->context;
@@ -886,6 +926,54 @@ static bool ggml_backend_cuda_buffer_cpy_tensor(ggml_backend_buffer_t buffer, co
     GGML_UNUSED(buffer);
 }
 
+static bool ggml_backend_cuda_buffer_cpy_tensor_n(ggml_backend_buffer_t buffer, const struct ggml_tensor * const * srcs, struct ggml_tensor * const * dsts, size_t n_tensors) {
+    ggml_backend_cuda_buffer_context * dst_ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
+
+    for (size_t i = 0; i < n_tensors; i++) {
+        ggml_backend_buffer_t src_buf = srcs[i]->view_src ? srcs[i]->view_src->buffer : srcs[i]->buffer;
+        if (!ggml_backend_buffer_is_cuda(src_buf)) {
+            return false;
+        }
+    }
+
+#ifdef GGML_CUDA_NO_PEER_COPY
+    for (size_t i = 0; i < n_tensors; i++) {
+        ggml_backend_buffer_t src_buf = srcs[i]->view_src ? srcs[i]->view_src->buffer : srcs[i]->buffer;
+        ggml_backend_buffer_t dst_buf = dsts[i]->view_src ? dsts[i]->view_src->buffer : dsts[i]->buffer;
+        ggml_backend_cuda_buffer_context * src_ctx = (ggml_backend_cuda_buffer_context *)src_buf->context;
+        ggml_backend_cuda_buffer_context * dst_ctx_i = (ggml_backend_cuda_buffer_context *)dst_buf->context;
+        const int src_physical = ggml_cuda_get_physical_device(src_ctx->device);
+        const int dst_physical = ggml_cuda_get_physical_device(dst_ctx_i->device);
+        if (src_physical != dst_physical) {
+            return false;
+        }
+    }
+#endif
+
+    ggml_cuda_set_device(dst_ctx->device);
+
+    for (size_t i = 0; i < n_tensors; i++) {
+        ggml_backend_buffer_t src_buf = srcs[i]->view_src ? srcs[i]->view_src->buffer : srcs[i]->buffer;
+        ggml_backend_buffer_t dst_buf = dsts[i]->view_src ? dsts[i]->view_src->buffer : dsts[i]->buffer;
+        ggml_backend_cuda_buffer_context * src_ctx = (ggml_backend_cuda_buffer_context *)src_buf->context;
+        ggml_backend_cuda_buffer_context * dst_ctx_i = (ggml_backend_cuda_buffer_context *)dst_buf->context;
+        const int src_physical = ggml_cuda_get_physical_device(src_ctx->device);
+        const int dst_physical = ggml_cuda_get_physical_device(dst_ctx_i->device);
+        if (src_physical == dst_physical) {
+            CUDA_CHECK(cudaMemcpyAsync(dsts[i]->data, srcs[i]->data, ggml_nbytes(srcs[i]), cudaMemcpyDeviceToDevice, cudaStreamPerThread));
+        } else {
+#ifdef GGML_CUDA_NO_PEER_COPY
+            GGML_ABORT("peer copy not supported");
+#else
+            CUDA_CHECK(cudaMemcpyPeerAsync(dsts[i]->data, dst_physical, srcs[i]->data, src_physical, ggml_nbytes(srcs[i]), cudaStreamPerThread));
+#endif
+        }
+    }
+
+    CUDA_CHECK(cudaStreamSynchronize(cudaStreamPerThread));
+    return true;
+}
+
 static void ggml_backend_cuda_buffer_clear(ggml_backend_buffer_t buffer, uint8_t value) {
     ggml_backend_cuda_buffer_context * ctx = (ggml_backend_cuda_buffer_context *)buffer->context;
 
@@ -906,6 +994,9 @@ static const ggml_backend_buffer_i ggml_backend_cuda_buffer_interface = {
     /* .cpy_tensor      = */ ggml_backend_cuda_buffer_cpy_tensor,
     /* .clear           = */ ggml_backend_cuda_buffer_clear,
     /* .reset           = */ NULL,
+    /* .set_tensor_n    = */ ggml_backend_cuda_buffer_set_tensor_n,
+    /* .get_tensor_n    = */ ggml_backend_cuda_buffer_get_tensor_n,
+    /* .cpy_tensor_n    = */ ggml_backend_cuda_buffer_cpy_tensor_n,
 };
 
 // cuda buffer type
@@ -1565,7 +1656,7 @@ static void ggml_cuda_mul_mat_cublas_impl(ggml_backend_cuda_context & ctx, const
     if (compute_type == GGML_TYPE_F16) {
         prefer_f32_output = cc == GGML_CUDA_CC_VOLTA || GGML_CUDA_CC_IS_RDNA4(cc) || GGML_CUDA_CC_IS_CDNA(cc);
     } else if (compute_type == GGML_TYPE_BF16) {
-        prefer_f32_output = !GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_CDNA(cc);
+        prefer_f32_output = !GGML_CUDA_CC_IS_RDNA3(cc) && !GGML_CUDA_CC_IS_RDNA4_M(cc) && !GGML_CUDA_CC_IS_CDNA(cc);
     }
 
     if (prefer_f32_output) {

ggml/src/ggml-cuda/mma.cuh

@@ -76,11 +76,11 @@ namespace ggml_cuda_mma {
         // For the A/C matrices this means I major == row major, J major == column major.
         // For the B matrix this means I major == column major, J major == row major.
         // MIRRORED == Each data value is held exactly once per thread subgroup.
-        DATA_LAYOUT_I_MAJOR           =  0, // Always used for Turing, Ampere, Ada Lovelace, consumer Blackwell, matrix A&B for RDNA4 and CDNA.
-        DATA_LAYOUT_J_MAJOR           = 10, // Matrix C for CDNA and RDNA4, int and float matrix C for RDNA3.
+        DATA_LAYOUT_I_MAJOR           =  0, // Always used for Turing, Ampere, Ada Lovelace, consumer Blackwell, matrix A&B for WMMA128 and CDNA.
+        DATA_LAYOUT_J_MAJOR           = 10, // Matrix C for CDNA and WMMA128, int and float matrix C for WMMA256.
         DATA_LAYOUT_I_MAJOR_MIRRORED  = 20, // Volta, matrix A&B for RDNA3.
         DATA_LAYOUT_J_MAJOR_MIRRORED  = 30,
-        DATA_LAYOUT_I_MAJOR_SCRAMBLED = 40, // Scrambled matrix C for faster transposition (RDNA4/CDNA), convert to float to unscramble.
+        DATA_LAYOUT_I_MAJOR_SCRAMBLED = 40, // Scrambled matrix C for faster transposition (WMMA128/CDNA), convert to float to unscramble.
     };
     // Implemented mma combinations are:
     //   - (I_MAJOR, I_MAJOR)          -> I_MAJOR
@@ -88,11 +88,11 @@ namespace ggml_cuda_mma {
     //   - (I_MAJOR, J_MAJOR_MIRRORED) -> I_MAJOR
 
     static constexpr __device__ data_layout get_input_data_layout() {
-#if defined(RDNA3) || defined(VOLTA_MMA_AVAILABLE)
+#if defined(AMD_WMMA_256B_AVAILABLE) || defined(VOLTA_MMA_AVAILABLE)
         return DATA_LAYOUT_I_MAJOR_MIRRORED;
 #else
         return DATA_LAYOUT_I_MAJOR;
-#endif // defined(RDNA3) || defined(VOLTA_MMA_AVAILABLE)
+#endif // defined(AMD_WMMA_256B_AVAILABLE) || defined(VOLTA_MMA_AVAILABLE)
     }
 
     template <int I_, int J_, typename T, data_layout ds_=DATA_LAYOUT_I_MAJOR>
@@ -201,7 +201,7 @@ namespace ggml_cuda_mma {
 
         static __device__ __forceinline__ int get_j(const int l) {
             if constexpr (I == 16 && J == 16) {
-#if defined(RDNA3)
+#if defined(AMD_WMMA_256B_AVAILABLE)
                 if constexpr (std::is_same_v<T, float> || std::is_same_v<T, int>) {
                     // matrix C
                     return 2 * l + (threadIdx.x / 16);
@@ -210,11 +210,11 @@ namespace ggml_cuda_mma {
                     return l;
                 }
 #else
-                // matrix C is the transposed matrix A&B on RDNA4
+                // matrix C is the transposed matrix A&B on WMMA128
                 return ne * (threadIdx.x / 16) + l;
-#endif // defined(RDNA3)
+#endif // defined(AMD_WMMA_256B_AVAILABLE)
             } else if constexpr (I == 16 && J == 8) {
-                // mmq input for RDNA4
+                // mmq input for WMMA128
                 return ne * (threadIdx.x / 16) + l;
             } else if constexpr (I == 16 && J == 4) {
                 return ne * (threadIdx.x / 16) + l;
@@ -573,7 +573,7 @@ namespace ggml_cuda_mma {
         static constexpr int         I  = I_;
         static constexpr int         J  = J_;
         static constexpr data_layout dl = DATA_LAYOUT_I_MAJOR_MIRRORED;
-#if defined(RDNA3)
+#if defined(AMD_WMMA_256B_AVAILABLE)
         static constexpr int         ne = tile<I_, J_, float, DATA_LAYOUT_I_MAJOR_MIRRORED>::ne;
 
         half2 x[ne] = {{0.0f, 0.0f}};
@@ -616,7 +616,7 @@ namespace ggml_cuda_mma {
                 return -1;
             }
         }
-#endif // defined(RDNA3)
+#endif // defined(AMD_WMMA_256B_AVAILABLE)
     };
 
     template <int I_, int J_>
@@ -694,7 +694,7 @@ namespace ggml_cuda_mma {
     };
 
     static __device__ __forceinline__ tile<16, 16, half2, DATA_LAYOUT_I_MAJOR> unscramble(const tile<16, 16, half2, DATA_LAYOUT_I_MAJOR_SCRAMBLED> & t) {
-#if defined(AMD_MFMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4))
+#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_128B_AVAILABLE)
         tile<16, 16, half2, DATA_LAYOUT_I_MAJOR> ret;
 #pragma unroll
         for (int l0 = 0; l0 < t.ne/2; ++l0) {
@@ -705,7 +705,7 @@ namespace ggml_cuda_mma {
 #else
         NO_DEVICE_CODE;
         GGML_UNUSED(t);
-#endif // defined(AMD_MFMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4))
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_128B_AVAILABLE)
     }
 
 #if defined(TURING_MMA_AVAILABLE)
@@ -726,7 +726,7 @@ namespace ggml_cuda_mma {
 
         return ret;
     }
-#elif defined(AMD_WMMA_AVAILABLE) && defined(RDNA3)
+#elif defined(AMD_WMMA_256B_AVAILABLE)
     static __device__ __forceinline__ tile<16, 8, half2, DATA_LAYOUT_I_MAJOR_MIRRORED> get_half2(
             const tile<16, 16, float, DATA_LAYOUT_I_MAJOR> & tile_float) {
         tile<16, 8, half2, DATA_LAYOUT_I_MAJOR_MIRRORED> ret;
@@ -1171,7 +1171,7 @@ namespace ggml_cuda_mma {
             : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[3]));
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 #elif defined(AMD_WMMA_AVAILABLE)
-#if defined(RDNA4)
+#if defined(AMD_WMMA_128B_AVAILABLE)
         using halfx8_t = __attribute__((ext_vector_type(8))) _Float16;
         halfx8_t& acc_frag = reinterpret_cast<halfx8_t&>(D.x[0]);
         const halfx8_t& a_frag = reinterpret_cast<const halfx8_t&>(A.x[0]);
@@ -1180,7 +1180,7 @@ namespace ggml_cuda_mma {
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
-#endif // defined(RDNA4)
+#endif // defined(AMD_WMMA_128B_AVAILABLE)
 #elif defined(AMD_MFMA_AVAILABLE)
         // MFMA: FP16 input, FP32 accumulate, convert back to half2.
         using halfx4_t = __attribute__((ext_vector_type(4))) _Float16;
@@ -1218,7 +1218,7 @@ namespace ggml_cuda_mma {
     static __device__ __forceinline__ void mma(
             tile<16, 16, half2, DATA_LAYOUT_I_MAJOR_SCRAMBLED> & D, const tile<32, 8, half2, DATA_LAYOUT_I_MAJOR> & A,
             const tile<16, 8, half2, DATA_LAYOUT_I_MAJOR> & B) {
-#if defined(AMD_MFMA_AVAILABLE) || (defined(AMD_WMMA_AVAILABLE) && defined(RDNA4))
+#if defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_128B_AVAILABLE)
         tile<16, 8, half2>       * D16 = (tile<16, 8, half2>       *) &D;
         const tile<16, 8, half2> * A16 = (const tile<16, 8, half2> *) &A;
         mma(D16[0], A16[0], B);
@@ -1226,7 +1226,7 @@ namespace ggml_cuda_mma {
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
-#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_AVAILABLE) && defined(RDNA4)
+#endif // defined(AMD_MFMA_AVAILABLE) || defined(AMD_WMMA_128B_AVAILABLE)
     }
 
     template <data_layout dl_ab, data_layout dl_d>
@@ -1373,14 +1373,14 @@ namespace ggml_cuda_mma {
             : "r"(Axi[2]), "r"(Axi[3]), "r"(Bxi[3]));
 #endif // __CUDA_ARCH__ >= GGML_CUDA_CC_AMPERE
 #elif defined(AMD_WMMA_AVAILABLE)
-#if defined(RDNA4)
+#if defined(AMD_WMMA_128B_AVAILABLE)
         using halfx8_t = __attribute__((ext_vector_type(8))) _Float16;
         using floatx8_t = __attribute__((ext_vector_type(8))) float;
         floatx8_t& acc_frag = reinterpret_cast<floatx8_t&>(D.x[0]);
         const halfx8_t& a_frag = reinterpret_cast<const halfx8_t&>(A.x[0]);
         const halfx8_t& b_frag = reinterpret_cast<const halfx8_t&>(B.x[0]);
         acc_frag = __builtin_amdgcn_wmma_f32_16x16x16_f16_w32_gfx12(a_frag, b_frag, acc_frag);
-#elif defined(RDNA3)
+#elif defined(AMD_WMMA_256B_AVAILABLE)
         using halfx16_t = __attribute__((ext_vector_type(16))) _Float16;
         using floatx8_t = __attribute__((ext_vector_type(8))) float;
         floatx8_t& acc_frag = reinterpret_cast<floatx8_t&>(D.x[0]);
@@ -1390,7 +1390,7 @@ namespace ggml_cuda_mma {
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
-#endif // RDNA4
+#endif // AMD_WMMA_128B_AVAILABLE
 #elif defined(AMD_MFMA_AVAILABLE)
         using halfx4_t = __attribute__((ext_vector_type(4))) _Float16;
         using floatx4_t = __attribute__((ext_vector_type(4))) float;
@@ -1408,14 +1408,14 @@ namespace ggml_cuda_mma {
     static __device__ __forceinline__ void mma(
             tile<16, 16, float, dl_d> & D, const tile<16, 8, nv_bfloat162, dl_ab> & A, const tile<16, 8, nv_bfloat162, dl_ab> & B) {
 #if defined(AMD_WMMA_AVAILABLE)
-#if defined(RDNA4)
+#if defined(AMD_WMMA_128B_AVAILABLE)
         using bf16x8_t = __attribute__((ext_vector_type(8))) short;
         using floatx8_t = __attribute__((ext_vector_type(8))) float;
         floatx8_t& acc_frag = reinterpret_cast<floatx8_t&>(D.x[0]);
         const bf16x8_t& a_frag = reinterpret_cast<const bf16x8_t&>(A.x[0]);
         const bf16x8_t& b_frag = reinterpret_cast<const bf16x8_t&>(B.x[0]);
         acc_frag = __builtin_amdgcn_wmma_f32_16x16x16_bf16_w32_gfx12(a_frag, b_frag, acc_frag);
-#elif defined(RDNA3)
+#elif defined(AMD_WMMA_256B_AVAILABLE)
         using bf16x16_t = __attribute__((ext_vector_type(16))) short;
         using floatx8_t = __attribute__((ext_vector_type(8))) float;
         floatx8_t& acc_frag = reinterpret_cast<floatx8_t&>(D.x[0]);
@@ -1425,7 +1425,7 @@ namespace ggml_cuda_mma {
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
-#endif // defined(RDNA4)
+#endif // defined(AMD_WMMA_128B_AVAILABLE)
 #elif defined(AMD_MFMA_AVAILABLE)
         using floatx4_t = __attribute__((ext_vector_type(4))) float;
         floatx4_t& acc_frag = reinterpret_cast<floatx4_t&>(D.x[0]);
@@ -1467,19 +1467,19 @@ namespace ggml_cuda_mma {
 #elif defined(AMD_WMMA_AVAILABLE)
         using int32x8_t = __attribute__((__vector_size__(8 * sizeof(int)))) int;
         int32x8_t * acc = (int32x8_t *) D.x;
-#if defined(RDNA4)
+#if defined(AMD_WMMA_128B_AVAILABLE)
         using int32x2_t = __attribute__((__vector_size__(2 * sizeof(int)))) int;
         int32x2_t * a_vec = (int32x2_t *) A.x;
         int32x2_t * b_vec = (int32x2_t *) B.x;
         acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(true, a_vec[0], true, b_vec[0], acc[0], true);
         acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(true, a_vec[1], true, b_vec[1], acc[0], true);
-#elif defined(RDNA3)
+#elif defined(AMD_WMMA_256B_AVAILABLE)
         using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
         int32x4_t * a_vec = (int32x4_t *) A.x;
         int32x4_t * b_vec = (int32x4_t *) B.x;
         acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(true, a_vec[0], true, b_vec[0], acc[0], true);
         acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(true, a_vec[1], true, b_vec[1], acc[0], true);
-#endif // RDNA4
+#endif // AMD_WMMA_128B_AVAILABLE
 #else
         GGML_UNUSED_VARS(D, A, B);
         NO_DEVICE_CODE;
@@ -1556,7 +1556,7 @@ namespace ggml_cuda_mma {
     static __device__ __forceinline__ void mma(
             tile<16, 16, half2, DATA_LAYOUT_I_MAJOR> & D, const tile<32,  8, half2, DATA_LAYOUT_I_MAJOR_MIRRORED> & A,
             const tile<16,  8, half2, DATA_LAYOUT_I_MAJOR_MIRRORED> & B) {
-#if defined(AMD_WMMA_AVAILABLE) && defined(RDNA3)
+#if defined(AMD_WMMA_256B_AVAILABLE)
         using halfx16_t = __attribute__((ext_vector_type(16))) _Float16;
         halfx16_t       * xD = (halfx16_t       *) D.x;
         const halfx16_t * xA = (const halfx16_t *) A.x;
@@ -1585,17 +1585,17 @@ namespace ggml_cuda_mma {
 #elif defined(AMD_WMMA_AVAILABLE)
         using int32x8_t = __attribute__((__vector_size__(8 * sizeof(int)))) int;
         int32x8_t * acc = (int32x8_t *) D.x;
-#if defined(RDNA4)
+#if defined(AMD_WMMA_128B_AVAILABLE)
         using int32x2_t = __attribute__((__vector_size__(2 * sizeof(int)))) int;
         int32x2_t * a_vec = (int32x2_t *) A.x;
         int32x2_t * b_vec = (int32x2_t *) B.x;
         acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(true, a_vec[0], true, b_vec[0], acc[0], false);
-#elif defined(RDNA3)
+#elif defined(AMD_WMMA_256B_AVAILABLE)
         using int32x4_t = __attribute__((__vector_size__(4 * sizeof(int)))) int;
         int32x4_t * a_vec = (int32x4_t *) A.x;
         int32x4_t * b_vec = (int32x4_t *) B.x;
         acc[0] = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32(true, a_vec[0], true, b_vec[0], acc[0], false);
-#endif // RDNA4
+#endif // AMD_WMMA_128B_AVAILABLE
 #else
         GGML_UNUSED(D);
         GGML_UNUSED(A);

ggml/src/ggml-cuda/mmq.cu

@@ -242,7 +242,7 @@ void ggml_cuda_mul_mat_q(
 
             // Each expert only sees ne12*n_expert_used/ne02 tokens on average.
             // On RDNA3 and RDNA4 it is faster to pick the tile size against this value instead of ne12.
-            if (GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {
+            if (GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4_M(cc) || GGML_CUDA_CC_IS_RDNA4(cc)) {
                 ncols_opt = (ne12*n_expert_used + ne02 - 1) / ne02;
             }
         }
@@ -470,7 +470,7 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t
     }
 
     if (amd_wmma_available(cc)) {
-        if (GGML_CUDA_CC_IS_RDNA3(cc)) {
+        if (GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4_M(cc)) {
             // High expert counts are almost always better on MMQ due to
             //     the synchronization overhead in the cuBLAS/hipBLAS path:
             // https://github.com/ggml-org/llama.cpp/pull/18202
@@ -487,7 +487,7 @@ bool ggml_cuda_should_use_mmq(enum ggml_type type, int cc, int64_t ne11, int64_t
                     return ne11 <= (GGML_CUDA_CC_IS_RDNA3_0(cc) ? 128 : 256);
                 case GGML_TYPE_IQ2_XS:
                 case GGML_TYPE_IQ2_S:
-                    return GGML_CUDA_CC_IS_RDNA3_5(cc) || ne11 <= 128;
+                    return GGML_CUDA_CC_IS_RDNA3_5(cc) || GGML_CUDA_CC_IS_RDNA4_M(cc) || ne11 <= 128;
                 default:
                     return true;
             }

ggml/src/ggml-cuda/mmq.cuh

@@ -240,6 +240,9 @@ static __host__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config(const ggml_type ty
         if (GGML_CUDA_CC_IS_RDNA4(cc)) {
             return ggml_cuda_mmq_get_config_rdna4(type, J, fallback);
         }
+        if (GGML_CUDA_CC_IS_RDNA4_M(cc)) {
+            return ggml_cuda_mmq_get_config_rdna3_5(type, J, fallback);
+        }
         if (GGML_CUDA_CC_IS_RDNA3_5(cc)) {
             return ggml_cuda_mmq_get_config_rdna3_5(type, J, fallback);
         }
@@ -272,6 +275,8 @@ static constexpr __device__ ggml_cuda_mmq_config ggml_cuda_mmq_get_config(ggml_t
     return ggml_cuda_mmq_get_config_cdna(type, J, fallback);
 #elif defined(RDNA4)
     return ggml_cuda_mmq_get_config_rdna4(type, J, fallback);
+#elif defined(RDNA4_M)
+    return ggml_cuda_mmq_get_config_rdna3_5(type, J, fallback);
 #elif defined(RDNA3_5)
     return ggml_cuda_mmq_get_config_rdna3_5(type, J, fallback);
 #elif defined(RDNA3)

ggml/src/ggml-cuda/mmvf.cu

@@ -844,7 +844,7 @@ bool ggml_cuda_should_use_mmvf(enum ggml_type type, int cc, int warp_size, const
                     return ne11 <= MMVF_MAX_BATCH_SIZE;
                 }
                 if (fp16_mma_hardware_available(cc)) {
-                    if (GGML_CUDA_CC_IS_RDNA3(cc)) {
+                    if (GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4_M(cc)) {
                         return ne11 <= 3;
                     }
                     if (GGML_CUDA_CC_IS_RDNA4(cc)) {

ggml/src/ggml-cuda/mmvq.cu

@@ -108,7 +108,7 @@ static constexpr __device__ mmvq_parameter_table_id get_device_table_id() {
     return MMVQ_PARAMETERS_RDNA4;
 #elif defined(RDNA3_0)
     return MMVQ_PARAMETERS_RDNA3_0;
-#elif defined(RDNA2) || defined(RDNA3_5)
+#elif defined(RDNA2) || defined(RDNA3_5) || defined(RDNA4_M)
     return MMVQ_PARAMETERS_RDNA2;
 #elif defined(GCN) || defined(CDNA)
     return MMVQ_PARAMETERS_GCN;
@@ -128,7 +128,7 @@ static __host__ mmvq_parameter_table_id get_device_table_id(int cc) {
     if (GGML_CUDA_CC_IS_RDNA3_0(cc)) {
         return MMVQ_PARAMETERS_RDNA3_0;
     }
-    if (GGML_CUDA_CC_IS_RDNA2(cc) || GGML_CUDA_CC_IS_RDNA3_5(cc)) {
+    if (GGML_CUDA_CC_IS_RDNA2(cc) || GGML_CUDA_CC_IS_RDNA3_5(cc) || GGML_CUDA_CC_IS_RDNA4_M(cc)) {
         return MMVQ_PARAMETERS_RDNA2;
     }
     if (GGML_CUDA_CC_IS_GCN(cc) || GGML_CUDA_CC_IS_CDNA(cc)) {
@@ -299,7 +299,7 @@ int get_mmvq_mmid_max_batch(ggml_type type, int cc) {
         if (GGML_CUDA_CC_IS_RDNA4(cc)) {
             return get_mmvq_mmid_max_batch_rdna4(type);
         }
-        if (GGML_CUDA_CC_IS_RDNA3(cc)) {
+        if (GGML_CUDA_CC_IS_RDNA3(cc) || GGML_CUDA_CC_IS_RDNA4_M(cc)) {
             return get_mmvq_mmid_max_batch_rdna3(type);
         }
         if (GGML_CUDA_CC_IS_RDNA1(cc) || GGML_CUDA_CC_IS_RDNA2(cc)) {
@@ -432,7 +432,7 @@ template <ggml_type type>
 static constexpr __device__ int get_mmvq_mmid_max_batch_for_device() {
 #if defined(RDNA4)
     return get_mmvq_mmid_max_batch_rdna4(type);
-#elif defined(RDNA3)
+#elif defined(RDNA3) || defined(RDNA4_M)
     return get_mmvq_mmid_max_batch_rdna3(type);
 #elif defined(RDNA2) || defined(RDNA1)
     return get_mmvq_mmid_max_batch_rdna1_rdna2(type);

ggml/src/ggml-cuda/vendors/hip.h

@@ -216,9 +216,13 @@
 #define RDNA4
 #endif // defined(__GFX12__)
 
-#if defined(__GFX11__)
+#if defined(__gfx1170__) || defined(__gfx1171__)
+#define RDNA4_M
+#endif // defined(__gfx1170__) || defined(__gfx1171__)
+
+#if defined(__GFX11__) && !defined(RDNA4_M)
 #define RDNA3
-#endif // defined(__GFX11__)
+#endif // defined(__GFX11__) && !defined(RDNA4_M)
 
 #if defined(__gfx1150__) || defined(__gfx1151__) || defined(__gfx1152__) || defined(__gfx1153__)
 #define RDNA3_5
@@ -237,9 +241,9 @@
 #define RDNA1
 #endif // defined(__gfx1010__) || defined(__gfx1012__)
 
-#if defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(RDNA1)
+#if defined(RDNA4) || defined(RDNA4_M) || defined(RDNA3) || defined(RDNA2) || defined(RDNA1)
 #define RDNA // For the entire family
-#endif // defined(RDNA4) || defined(RDNA3) || defined(RDNA2) || defined(RDNA1)
+#endif // defined(RDNA4) || defined(RDNA4_M) || defined(RDNA3) || defined(RDNA2) || defined(RDNA1)
 
 #ifndef __has_builtin
     #define __has_builtin(x) 0
@@ -254,6 +258,10 @@ typedef __hip_fp8_e4m3 __nv_fp8_e4m3;
 #define FP8_AVAILABLE
 #endif // HIP_VERSION >= 60300000
 
+#if HIP_VERSION < 70000000
+#define __ballot_sync(mask, var) __ballot(var)
+#endif //HIP_VERSION < 70000000
+
 typedef int8_t int8x4_t __attribute__((ext_vector_type(4)));
 typedef uint8_t uint8x4_t __attribute__((ext_vector_type(4)));
 static __device__ __forceinline__ int __vsubss4(const int a, const int b) {

ggml/src/ggml-et/ggml-et.cpp

@@ -373,6 +373,9 @@ static const struct ggml_backend_buffer_i ggml_backend_et_buffer_i = {
     /* .cpy_tensor      = */ ggml_backend_et_buffer_cpy_tensor,
     /* .clear           = */ ggml_backend_et_buffer_clear,
     /* .reset           = */ NULL,
+    /* .set_tensor_n    = */ NULL,
+    /* .get_tensor_n    = */ NULL,
+    /* .cpy_tensor_n    = */ NULL,
 };
 
 static const char * ggml_backend_et_buffer_type_get_name(ggml_backend_buffer_type_t buft) {

ggml/src/ggml-hexagon/ggml-hexagon.cpp

@@ -843,6 +843,9 @@ static ggml_backend_buffer_i ggml_backend_hexagon_buffer_interface = {
     /* .cpy_tensor      = */ ggml_backend_hexagon_buffer_cpy_tensor,
     /* .clear           = */ ggml_backend_hexagon_buffer_clear,
     /* .reset           = */ NULL,
+    /* .set_tensor_n    = */ NULL,
+    /* .get_tensor_n    = */ NULL,
+    /* .cpy_tensor_n    = */ NULL,
 };
 
 // ** backend buffer type
@@ -877,6 +880,9 @@ static ggml_backend_buffer_i ggml_backend_hexagon_host_buffer_interface = {
     /* .cpy_tensor      = */ ggml_backend_hexagon_buffer_cpy_tensor,
     /* .clear           = */ ggml_backend_hexagon_buffer_clear,
     /* .reset           = */ NULL,
+    /* .set_tensor_n    = */ NULL,
+    /* .get_tensor_n    = */ NULL,
+    /* .cpy_tensor_n    = */ NULL,
 };
 
 // ** backend buffer type

ggml/src/ggml-hexagon/htp-kparams.cpp

@@ -60,6 +60,7 @@ htp_op_code htp_op_remap(const struct ggml_tensor * t) {
         case GGML_OP_CONV_2D:         return HTP_OP_CONV_2D;
         case GGML_OP_CONV_2D_DW:      return HTP_OP_CONV_2D_DW;
         case GGML_OP_ROLL:            return HTP_OP_ROLL;
+        case GGML_OP_UPSCALE:         return HTP_OP_UPSCALE;
 
         case GGML_OP_UNARY:
             switch (ggml_get_unary_op(t)) {
@@ -2185,7 +2186,8 @@ static bool htp_kparams_solve_ssm_conv(
         target_max_tile          = (std::max)(target_max_tile, 32u);
         target_max_tile          = (std::min)(target_max_tile, 128u);
 
-        uint32_t d_inner_tile = (uint32_t) (avail_for_tiles / (2 * (ncs + n_t) * sizeof(float)));
+        const size_t raw_stride_bytes = ncs * sizeof(float) + sizeof(float);
+        uint32_t d_inner_tile = (uint32_t) (avail_for_tiles / (2 * (raw_stride_bytes + n_t * sizeof(float))));
         d_inner_tile          = (d_inner_tile / 32) * 32;
         if (d_inner_tile == 0) {
             d_inner_tile = 32;
@@ -2372,6 +2374,183 @@ static bool htp_kparams_solve_pool_2d(
     return true;
 }
 
+bool htp_kparams_solve_upscale(
+    const struct ggml_hexagon_session * sess,
+    const struct ggml_tensor * src0,
+    const struct ggml_tensor * dst,
+    struct htp_upscale_kernel_params * kparams
+) {
+    memset(kparams, 0, sizeof(*kparams));
+
+    if (src0->type != GGML_TYPE_F32 || dst->type != GGML_TYPE_F32) {
+        return false;
+    }
+
+    const uint32_t src_x = (uint32_t) src0->ne[0];
+    const uint32_t src_y = (uint32_t) src0->ne[1];
+    const uint32_t src_z = (uint32_t) src0->ne[2];
+    const uint32_t src_w = (uint32_t) src0->ne[3];
+
+    const uint32_t dst_x = (uint32_t) dst->ne[0];
+    const uint32_t dst_y = (uint32_t) dst->ne[1];
+    const uint32_t dst_z = (uint32_t) dst->ne[2];
+    const uint32_t dst_w = (uint32_t) dst->ne[3];
+
+    if (src_x == 0 || src_y == 0 || src_z == 0 || src_w == 0 ||
+        dst_x == 0 || dst_y == 0 || dst_z == 0 || dst_w == 0) {
+        return false;
+    }
+
+    const bool is_src_contig_row = (src0->nb[0] == sizeof(float));
+    const bool is_src_strided_2d = (src0->nb[0] > sizeof(float)) &&
+                                   (src0->nb[0] % sizeof(float) == 0) &&
+                                   (src0->nb[1] == (size_t) src_x * src0->nb[0]) &&
+                                   ((size_t) src_x * src_y <= 0xFFFFu);
+
+    if ((!is_src_contig_row && !is_src_strided_2d) || dst->nb[0] != sizeof(float)) {
+        return false;
+    }
+
+    const int32_t mode_flags = dst->op_params[0];
+    const uint32_t mode_raw  = (uint32_t) (mode_flags & 0xFF);
+    if (mode_raw >= GGML_SCALE_MODE_COUNT) {
+        return false;
+    }
+
+    const bool align_corners = (mode_flags & GGML_SCALE_FLAG_ALIGN_CORNERS) != 0;
+    const bool antialias     = (mode_flags & GGML_SCALE_FLAG_ANTIALIAS) != 0;
+
+    float sf0 = (float) dst_x / (float) src_x;
+    float sf1 = (float) dst_y / (float) src_y;
+    float sf2 = (float) dst_z / (float) src_z;
+    float sf3 = (float) dst_w / (float) src_w;
+    float pixel_offset = 0.5f;
+
+    if (align_corners) {
+        pixel_offset = 0.0f;
+        sf0 = (dst_x > 1 && src_x > 1) ? (float) (dst_x - 1) / (float) (src_x - 1) : sf0;
+        sf1 = (dst_y > 1 && src_y > 1) ? (float) (dst_y - 1) / (float) (src_y - 1) : sf1;
+    }
+
+    const float inv_sf0 = 1.0f / sf0;
+    const float inv_sf1 = 1.0f / sf1;
+    const float inv_sf2 = 1.0f / sf2;
+    const float inv_sf3 = 1.0f / sf3;
+
+    const float filter_radius0 = antialias ? (std::max)(1.0f, inv_sf0) : 0.0f;
+    const float filter_radius1 = antialias ? (std::max)(1.0f, inv_sf1) : 0.0f;
+    const float invscale0      = antialias ? (1.0f / filter_radius0) : 0.0f;
+    const float invscale1      = antialias ? (1.0f / filter_radius1) : 0.0f;
+
+    const uint32_t total_slices = dst_z * dst_w;
+    uint32_t       n_threads    = htp_session_n_threads(sess);
+
+    struct htp_upscale_vtcm_layout layout;
+    uint32_t dst_nrows_per_chunk = 0;
+    while (n_threads > 0) {
+        if (htp_upscale_solve_layout(&layout, src_x, src_y, dst_x, dst_y, inv_sf1, filter_radius1, n_threads,
+                                     htp_session_vtcm_size(sess), &dst_nrows_per_chunk)) {
+            break;
+        }
+        n_threads--;
+    }
+    if (n_threads == 0) {
+        return false;
+    }
+
+    size_t lut_x_size = 0;
+    size_t elem_y_sz  = sizeof(uint32_t);
+    if (mode_raw == HTP_UPSCALE_MODE_BICUBIC) {
+        lut_x_size = hex_round_up(dst_x * sizeof(struct upscale_bicubic_x), 128);
+        elem_y_sz  = sizeof(struct upscale_bicubic_y);
+    } else if (antialias) {
+        lut_x_size = hex_round_up(dst_x * sizeof(struct upscale_antialias_x), 128);
+        elem_y_sz  = sizeof(struct upscale_antialias_y);
+    } else if (mode_raw == HTP_UPSCALE_MODE_BILINEAR) {
+        const uint32_t n_chunks = (dst_x + 31) / 32;
+        lut_x_size = hex_round_up(n_chunks * sizeof(struct upscale_bilinear_x_chunk), 128);
+        elem_y_sz  = sizeof(struct upscale_bilinear_y);
+    } else {
+        lut_x_size = hex_round_up(dst_x * sizeof(uint32_t), 128);
+    }
+    const size_t total_lut_size = lut_x_size + hex_round_up(dst_y * elem_y_sz, 128);
+    if (total_lut_size > HTP_DDR_SPAD_SIZE) {
+        return false;
+    }
+
+    const uint32_t chunks_per_slice = (dst_y + dst_nrows_per_chunk - 1) / dst_nrows_per_chunk;
+    const uint32_t total_chunks     = total_slices * chunks_per_slice;
+
+    n_threads = (std::min)(n_threads, total_chunks > 0 ? total_chunks : 1);
+    htp_upscale_vtcm_layout_build(&layout, src_x, src_y, dst_x, dst_y, dst_nrows_per_chunk, inv_sf1, filter_radius1, n_threads);
+
+    const uint32_t slice_dst_bytes = dst_x * dst_y * sizeof(float);
+    const uint32_t chunk_dst_bytes = dst_x * dst_nrows_per_chunk * sizeof(float);
+    const bool     dst_contig      = (dst->nb[1] == (size_t) dst_x * sizeof(float)) &&
+                                     (dst_z <= 1 || dst->nb[2] == (size_t) dst_y * dst->nb[1]) &&
+                                     (dst_w <= 1 || dst->nb[3] == (size_t) dst_z * dst->nb[2]);
+
+    uint32_t chunks_per_group = 0;
+    if (dst_contig) {
+        const uint32_t slices_per_group = (slice_dst_bytes > 0)
+            ? (HEX_L2_LINE_SIZE / hex_gcd_u32(slice_dst_bytes, HEX_L2_LINE_SIZE))
+            : 1;
+        if (slices_per_group == 1 && (chunk_dst_bytes % HEX_L2_LINE_SIZE == 0)) {
+            chunks_per_group = 1;
+        } else {
+            chunks_per_group = slices_per_group * chunks_per_slice;
+        }
+    }
+
+    uint8_t kernel_type = HTP_UPSCALE_KERNEL_BILINEAR;
+    if (mode_raw == HTP_UPSCALE_MODE_NEAREST) {
+        if (dst_x == 2 * src_x && dst_y == 2 * src_y && !align_corners) {
+            kernel_type = HTP_UPSCALE_KERNEL_NEAREST_2X;
+        } else if (dst_x == 4 * src_x && dst_y == 4 * src_y && !align_corners) {
+            kernel_type = HTP_UPSCALE_KERNEL_NEAREST_4X;
+        } else {
+            kernel_type = HTP_UPSCALE_KERNEL_NEAREST;
+        }
+    } else if (mode_raw == HTP_UPSCALE_MODE_BICUBIC) {
+        kernel_type = HTP_UPSCALE_KERNEL_BICUBIC;
+    } else if (antialias) {
+        kernel_type = HTP_UPSCALE_KERNEL_BILINEAR_AA;
+    } else if (dst_x == 2 * src_x && dst_y == 2 * src_y && !align_corners && (src_x % 32 == 0)) {
+        kernel_type = HTP_UPSCALE_KERNEL_BILINEAR_2X;
+    } else {
+        kernel_type = HTP_UPSCALE_KERNEL_BILINEAR;
+    }
+
+    kparams->kernel_type          = kernel_type;
+    kparams->mode                 = (uint8_t) mode_raw;
+    kparams->align_corners        = align_corners ? 1 : 0;
+    kparams->antialias            = antialias ? 1 : 0;
+
+    kparams->inv_sf0              = inv_sf0;
+    kparams->inv_sf1              = inv_sf1;
+    kparams->inv_sf2              = inv_sf2;
+    kparams->inv_sf3              = inv_sf3;
+    kparams->pixel_offset         = pixel_offset;
+
+    kparams->filter_radius0       = filter_radius0;
+    kparams->filter_radius1       = filter_radius1;
+    kparams->invscale0            = invscale0;
+    kparams->invscale1            = invscale1;
+
+    kparams->n_threads            = n_threads;
+    kparams->chunks_per_group     = chunks_per_group;
+
+    kparams->dst_nrows_per_chunk  = dst_nrows_per_chunk;
+    kparams->chunks_per_slice     = chunks_per_slice;
+    kparams->total_chunks         = total_chunks;
+    kparams->vtcm_size            = (uint32_t) layout.total_bytes;
+
+    kparams->div_chunks_per_slice = init_fastdiv_values(chunks_per_slice > 0 ? chunks_per_slice : 1);
+    kparams->div_dst_z            = init_fastdiv_values(dst_z > 0 ? dst_z : 1);
+
+    return true;
+}
+
 static bool htp_kparams_solve_concat(
     const struct ggml_hexagon_session * sess,
     const struct ggml_tensor * op,
@@ -3916,6 +4095,31 @@ void htp_opnode_collapse(const struct ggml_hexagon_session * sess, htp_opnode &
         }
         return;
     }
+
+    if (node.opcode == HTP_OP_UPSCALE) {
+        if (node.inputs.empty() || !node.inputs[0] || node.outputs.empty() || !node.outputs[0]) {
+            return;
+        }
+        const struct ggml_tensor * src0 = node.inputs[0];
+        const struct ggml_tensor * dst  = node.outputs[0];
+
+        if ((src0->ne[3] > 1 || dst->ne[3] > 1) &&
+            src0->ne[2] == dst->ne[2] && src0->ne[3] == dst->ne[3] &&
+            src0->nb[3] == (size_t) src0->ne[2] * src0->nb[2] &&
+            dst->nb[3]  == (size_t) dst->ne[2]  * dst->nb[2]) {
+            struct ggml_tensor src0_c = *src0;
+            struct ggml_tensor dst_c  = *dst;
+            src0_c.ne[2] *= src0_c.ne[3];
+            src0_c.ne[3] = 1;
+            src0_c.nb[3] = (size_t) src0_c.ne[2] * src0_c.nb[2];
+            dst_c.ne[2]  *= dst_c.ne[3];
+            dst_c.ne[3]  = 1;
+            dst_c.nb[3]  = (size_t) dst_c.ne[2] * dst_c.nb[2];
+            node.inputs[0]  = node.add_dummy(src0_c);
+            node.outputs[0] = node.add_dummy(dst_c);
+        }
+        return;
+    }
 }
 
 bool htp_kparams_solve(
@@ -3945,6 +4149,9 @@ bool htp_kparams_solve(
         const size_t src2_size = src2 ? hex_round_up(ggml_nbytes(src2), 128) : 0;
         return htp_kparams_solve_matmul_impl(sess, src0, src1, dst, src2 ? src2->nb[1] : 0, src2_size, node.kparams<htp_mm_kernel_params>());
     }
+    if (node.opcode == HTP_OP_UPSCALE) {
+        return htp_kparams_solve_upscale(sess, src0, dst, node.kparams<htp_upscale_kernel_params>());
+    }
 
     return htp_kparams_solve(sess, node.node, node.kernel_params);
 }
@@ -3955,7 +4162,7 @@ bool htp_kparams_solve(
     void * kparams
 ) {
     const htp_op_code opcode = htp_op_remap(op);
-    if (htp_op_is_unary(opcode) || htp_op_is_glu(opcode)) {
+    if (htp_op_is_unary(opcode) || htp_op_is_glu(opcode) || opcode == HTP_OP_UPSCALE) {
         htp_opnode tmp(opcode, op);
         if (!htp_kparams_solve(sess, tmp)) {
             return false;

ggml/src/ggml-hexagon/htp-kparams.h

@@ -29,6 +29,7 @@
 #include "htp/ssm-conv.h"
 #include "htp/sum-rows-ops.h"
 #include "htp/unary-ops.h"
+#include "htp/upscale-ops.h"
 
 #include <type_traits>
 #include <utility>
@@ -134,4 +135,19 @@ bool htp_kparams_solve_unary_mul_glu(
     struct htp_glu_kernel_params * kparams
 );
 
+bool htp_kparams_solve_upscale(
+    const struct ggml_hexagon_session * sess,
+    const struct ggml_tensor * src0,
+    const struct ggml_tensor * dst,
+    struct htp_upscale_kernel_params * kparams
+);
+
+static inline bool htp_kparams_solve_upscale(
+    const struct ggml_hexagon_session * sess,
+    const struct ggml_tensor * op,
+    struct htp_upscale_kernel_params * kparams
+) {
+    return htp_kparams_solve_upscale(sess, op->src[0], op, kparams);
+}
+
 #endif // HTP_KPARAMS_H

ggml/src/ggml-hexagon/htp-opnode.cpp

@@ -19,6 +19,7 @@
 #include "htp/ssm-conv.h"
 #include "htp/sum-rows-ops.h"
 #include "htp/unary-ops.h"
+#include "htp/upscale-ops.h"
 
 #include <stdio.h>
 
@@ -224,6 +225,19 @@ static std::string format_kernel_params(const htp_opnode & node) {
     } else if (node.opcode == HTP_OP_SOLVE_TRI) {
         const auto * kparams = node.kparams<htp_solve_tri_kernel_params>();
         snprintf(str, max_size, "%s nth %d", kparams->batched ? "batched" : "chunked", (int) kparams->n_threads);
+    } else if (node.opcode == HTP_OP_UPSCALE) {
+        const auto * kparams = node.kparams<htp_upscale_kernel_params>();
+        const char * ktype_str = "unknown";
+        switch (kparams->kernel_type) {
+            case HTP_UPSCALE_KERNEL_NEAREST:     ktype_str = "nearest";     break;
+            case HTP_UPSCALE_KERNEL_NEAREST_2X:  ktype_str = "nearest-2x";  break;
+            case HTP_UPSCALE_KERNEL_NEAREST_4X:  ktype_str = "nearest-4x";  break;
+            case HTP_UPSCALE_KERNEL_BILINEAR:    ktype_str = kparams->align_corners ? "bilinear-ac" : "bilinear"; break;
+            case HTP_UPSCALE_KERNEL_BILINEAR_2X: ktype_str = "bilinear-2x"; break;
+            case HTP_UPSCALE_KERNEL_BILINEAR_AA: ktype_str = "bilinear-aa"; break;
+            case HTP_UPSCALE_KERNEL_BICUBIC:     ktype_str = "bicubic";     break;
+        }
+        snprintf(str, max_size, "%s nth %d chk %d vtcm %u", ktype_str, (int) kparams->n_threads, (int) kparams->dst_nrows_per_chunk, (unsigned int) kparams->vtcm_size);
     } else {
         snprintf(str, max_size, "----");
     }

ggml/src/ggml-hexagon/htp/CMakeLists.txt

@@ -47,6 +47,7 @@ add_library(${HTP_LIB} SHARED
     conv2d-ops.c
     roll-ops.c
     pool-ops.c
+    upscale-ops.c
     allreduce-ops.c
 )
 

ggml/src/ggml-hexagon/htp/concat-ops.c

@@ -47,87 +47,108 @@ static void concat_2d_f32_transposed(unsigned int nth, unsigned int ith, void *
     const uint32_t src0_ne0 = src0->ne[0];
     const uint32_t src1_ne0 = src1->ne[0];
 
-    const uint32_t row_end = cctx->row_start + cctx->nrows;
-    const uint32_t start_i = cctx->row_start + ith * cctx->nrows_per_thread;
-    const uint32_t end_i   = (start_i + cctx->nrows_per_thread < row_end) ? (start_i + cctx->nrows_per_thread) : row_end;
-    if (start_i >= end_i) return;
+    // dst rows [row_begin, row_end) of this thread; a dst row is a src0 row followed by a src1 column
+    const uint32_t row_begin = cctx->row_start + ith * cctx->nrows_per_thread;
+    const uint32_t row_end   = MIN(row_begin + cctx->nrows_per_thread, cctx->row_start + cctx->nrows);
+    if (row_begin >= row_end) return;
 
     dma_queue * dma_q = octx->ctx->dma[ith];
-
-    uint8_t * spad0_base = cctx->spad0_base + ith * cctx->spad0_size_per_thread;
-    uint8_t * spad1_base = cctx->spad1_base + ith * cctx->spad1_size_per_thread;
-
-    const uint32_t block_i = 32;
-    const uint32_t spad1_stride = block_i * sizeof(float);
-
-    const HVX_Vector offsets = hvx_vec_gather_offsets_w(spad1_stride);
-    const uint32_t src1_ne0_padded = hex_round_up(src1_ne0, 32);
-    const uint32_t src0_row_bytes  = src0_ne0 * sizeof(float);
-    const uint32_t src0_row_padded = hex_round_up(src0_row_bytes, VLEN);
-    const uint32_t src0_pre        = src0_row_padded - src0_row_bytes;
-    const uint32_t spad0_row_bytes = src0_row_padded + src1_ne0_padded * sizeof(float);
-
     struct htp_thread_trace * tr = &octx->ctx->trace[ith];
 
     const struct fastdiv_values * div_ne2 = &cctx->div_ne2;
     const uint32_t ne2 = dst->ne[2];
 
-    uint32_t p = 0;
-    uint32_t i = start_i;
+    // tile: block_rows dst rows x block_cols src1 columns
+    const uint32_t block_rows = 128;
+    const uint32_t block_cols = 256;
+    // +1 element spreads the gather lanes across VTCM, so they do not serialize on one path
+    const uint32_t spad1_stride = block_rows * sizeof(float) + sizeof(float);
 
-    const dma_addr_t src1_addr = concat_plane_addr(src1, p, div_ne2, ne2) + i * src1->nb[1];
-    dma_queue_push(dma_q, dma_make_data(spad1_base, src1_addr), spad1_stride, src1->nb[0], MIN(end_i - i, block_i) * sizeof(float), src1_ne0);
+    const HVX_Vector offsets = hvx_vec_gather_offsets_w(spad1_stride);
+    const uint32_t block_cols_padded = MIN(hex_round_up(src1_ne0, 32), block_cols);
+    const uint32_t src0_row_bytes    = src0_ne0 * sizeof(float);
+    const uint32_t src0_row_padded   = hex_round_up(src0_row_bytes, VLEN);
+    const uint32_t src0_pre          = src0_row_padded - src0_row_bytes;
+    const uint32_t spad0_stride      = src0_row_padded + block_cols_padded * sizeof(float);
 
-    const dma_addr_t src0_addr = concat_plane_addr(src0, p, div_ne2, ne2) + i * src0->nb[1];
-    dma_queue_push(dma_q, dma_make_data(spad0_base + src0_pre, src0_addr), spad0_row_bytes, src0->nb[1], src0_row_bytes, MIN(end_i - i, block_i));
+    // two slots per scratchpad: tile b gathers in slot b&1 while dst[b-1] and src1[b+1] DMA in the other
+    uint8_t * spad0_base = cctx->spad0_base + ith * cctx->spad0_size_per_thread;
+    uint8_t * spad1_base = cctx->spad1_base + ith * cctx->spad1_size_per_thread;
+    const uint32_t spad0_slot = block_rows * spad0_stride;
+    const uint32_t spad1_slot = block_cols_padded * spad1_stride;
 
-    dma_queue_pop(dma_q); // src1
+    // tiles walk (plane, row, col) with col fastest; src0 goes out with the col == 0 tile only
+    uint32_t plane = 0;
+    uint32_t row   = row_begin;
+    uint32_t col   = 0;
+    uint32_t slot  = 0;
+    bool dst_pending = false;
 
-    while (p < cctx->nplanes) {
-        const uint32_t current_block_i = MIN(end_i - i, block_i);
+    const dma_addr_t src1_addr = concat_plane_addr(src1, plane, div_ne2, ne2) + row * src1->nb[1];
+    dma_queue_push(dma_q, dma_make_data(spad1_base, src1_addr), spad1_stride, src1->nb[0], MIN(row_end - row, block_rows) * sizeof(float), MIN(src1_ne0, block_cols));
 
-        htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) i);
-        for (uint32_t j = 0; j < src1_ne0; j += 32) {
-            const uint8_t * src_ptr = spad1_base + j * spad1_stride;
-            uint8_t * dst_ptr = spad0_base + src0_row_padded + j * sizeof(float);
-            hvx_transpose_32x32_w_gather(dst_ptr, spad0_row_bytes, src_ptr, spad1_stride, offsets, current_block_i, MIN(src1_ne0 - j, 32));
+    const dma_addr_t src0_addr = concat_plane_addr(src0, plane, div_ne2, ne2) + row * src0->nb[1];
+    dma_queue_push(dma_q, dma_make_data(spad0_base + src0_pre, src0_addr), spad0_stride, src0->nb[1], src0_row_bytes, MIN(row_end - row, block_rows));
+
+    while (plane < cctx->nplanes) {
+        const uint32_t nrows = MIN(row_end - row, block_rows);
+        const uint32_t ncols = MIN(src1_ne0 - col, block_cols);
+
+        uint8_t * spad0 = spad0_base + slot * spad0_slot;
+        uint8_t * spad1 = spad1_base + slot * spad1_slot;
+
+        dma_queue_pop(dma_q); // src1
+        if (col == 0) {
+            dma_queue_pop(dma_q); // src0
         }
-        hvx_gather_sync(spad0_base + src0_row_padded);
-        htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) i);
 
-        uint32_t np = p;
-        uint32_t ni = i + block_i;
-        if (ni >= end_i) {
-            ni = start_i;
-            np++;
+        uint32_t next_plane = plane;
+        uint32_t next_row   = row;
+        uint32_t next_col   = col + block_cols;
+        if (next_col >= src1_ne0) {
+            next_col = 0;
+            next_row = row + block_rows;
+            if (next_row >= row_end) {
+                next_row = row_begin;
+                next_plane++;
+            }
         }
-        const bool has_next = np < cctx->nplanes;
-        const uint32_t next_block_i = MIN(end_i - ni, block_i);
+        const bool has_next = next_plane < cctx->nplanes;
+        const uint32_t next_nrows = MIN(row_end - next_row, block_rows);
+        const uint32_t next_ncols = MIN(src1_ne0 - next_col, block_cols);
 
-        // spad1 is free after the gather sync, prefetch next src1 ahead of the dst write
+        // other spad1 slot is free since the previous gather sync
         if (has_next) {
-            const dma_addr_t nsrc1_addr = concat_plane_addr(src1, np, div_ne2, ne2) + ni * src1->nb[1];
-            dma_queue_push(dma_q, dma_make_data(spad1_base, nsrc1_addr), spad1_stride, src1->nb[0], next_block_i * sizeof(float), src1_ne0);
+            const dma_addr_t addr = concat_plane_addr(src1, next_plane, div_ne2, ne2) + next_row * src1->nb[1] + next_col * src1->nb[0];
+            dma_queue_push(dma_q, dma_make_data(spad1_base + (slot ^ 1) * spad1_slot, addr), spad1_stride, src1->nb[0], next_nrows * sizeof(float), next_ncols);
         }
 
-        dma_queue_pop(dma_q); // src0
-
-        const dma_addr_t dst_addr = concat_plane_addr(dst, p, div_ne2, ne2) + i * dst->nb[1];
-        dma_queue_push(dma_q, dma_make_data(dst_addr, spad0_base + src0_pre), dst->nb[1], spad0_row_bytes, (src0_ne0 + src1_ne0) * sizeof(float), current_block_i);
-
-        if (has_next) {
-            dma_queue_pop(dma_q); // next src1
+        htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) row);
+        for (uint32_t c = 0; c < ncols; c += 32) {
+            hvx_transpose_32x32_w_gather(spad0 + src0_row_padded + c * sizeof(float), spad0_stride, spad1 + c * spad1_stride, spad1_stride, offsets, nrows, MIN(ncols - c, 32));
         }
-        dma_queue_pop(dma_q); // dst
+        hvx_gather_sync(spad0 + src0_row_padded);
+        htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) row);
 
-        // spad0 is free after the dst write
-        if (has_next) {
-            const dma_addr_t nsrc0_addr = concat_plane_addr(src0, np, div_ne2, ne2) + ni * src0->nb[1];
-            dma_queue_push(dma_q, dma_make_data(spad0_base + src0_pre, nsrc0_addr), spad0_row_bytes, src0->nb[1], src0_row_bytes, next_block_i);
+        // other spad0 slot is free once dst[b-1] is written
+        if (dst_pending) {
+            dma_queue_pop(dma_q); // dst
+        }
+        if (has_next && next_col == 0) {
+            const dma_addr_t addr = concat_plane_addr(src0, next_plane, div_ne2, ne2) + next_row * src0->nb[1];
+            dma_queue_push(dma_q, dma_make_data(spad0_base + (slot ^ 1) * spad0_slot + src0_pre, addr), spad0_stride, src0->nb[1], src0_row_bytes, next_nrows);
         }
 
-        p = np;
-        i = ni;
+        // the col == 0 tile also carries the src0 columns in front
+        const uint32_t lead = (col == 0) ? src0_row_bytes : 0;
+        const dma_addr_t dst_addr = concat_plane_addr(dst, plane, div_ne2, ne2) + row * dst->nb[1] + src0_row_bytes + col * sizeof(float) - lead;
+        dma_queue_push(dma_q, dma_make_data(dst_addr, spad0 + src0_row_padded - lead), dst->nb[1], spad0_stride, lead + ncols * sizeof(float), nrows);
+        dst_pending = true;
+
+        plane = next_plane;
+        row   = next_row;
+        col   = next_col;
+        slot ^= 1;
     }
     dma_queue_flush(dma_q);
 }
@@ -143,87 +164,108 @@ static void concat_2d_f16_transposed(unsigned int nth, unsigned int ith, void *
     const uint32_t src0_ne0 = src0->ne[0];
     const uint32_t src1_ne0 = src1->ne[0];
 
-    const uint32_t row_end = cctx->row_start + cctx->nrows;
-    const uint32_t start_i = cctx->row_start + ith * cctx->nrows_per_thread;
-    const uint32_t end_i   = (start_i + cctx->nrows_per_thread < row_end) ? (start_i + cctx->nrows_per_thread) : row_end;
-    if (start_i >= end_i) return;
+    // dst rows [row_begin, row_end) of this thread; a dst row is a src0 row followed by a src1 column
+    const uint32_t row_begin = cctx->row_start + ith * cctx->nrows_per_thread;
+    const uint32_t row_end   = MIN(row_begin + cctx->nrows_per_thread, cctx->row_start + cctx->nrows);
+    if (row_begin >= row_end) return;
 
     dma_queue * dma_q = octx->ctx->dma[ith];
-
-    uint8_t * spad0_base = cctx->spad0_base + ith * cctx->spad0_size_per_thread;
-    uint8_t * spad1_base = cctx->spad1_base + ith * cctx->spad1_size_per_thread;
-
-    const uint32_t block_i = 64;
-    const uint32_t spad1_stride = block_i * sizeof(__fp16);
-
-    const HVX_Vector offsets = hvx_vec_gather_offsets_h(spad1_stride);
-    const uint32_t src1_ne0_padded = hex_round_up(src1_ne0, 64);
-    const uint32_t src0_row_bytes  = src0_ne0 * sizeof(__fp16);
-    const uint32_t src0_row_padded = hex_round_up(src0_row_bytes, VLEN);
-    const uint32_t src0_pre        = src0_row_padded - src0_row_bytes;
-    const uint32_t spad0_row_bytes = src0_row_padded + src1_ne0_padded * sizeof(__fp16);
-
     struct htp_thread_trace * tr = &octx->ctx->trace[ith];
 
     const struct fastdiv_values * div_ne2 = &cctx->div_ne2;
     const uint32_t ne2 = dst->ne[2];
 
-    uint32_t p = 0;
-    uint32_t i = start_i;
+    // tile: block_rows dst rows x block_cols src1 columns
+    const uint32_t block_rows = 128;
+    const uint32_t block_cols = 512;
+    // +1 element spreads the gather lanes across VTCM, so they do not serialize on one path
+    const uint32_t spad1_stride = block_rows * sizeof(__fp16) + sizeof(__fp16);
 
-    const dma_addr_t src1_addr = concat_plane_addr(src1, p, div_ne2, ne2) + i * src1->nb[1];
-    dma_queue_push(dma_q, dma_make_data(spad1_base, src1_addr), spad1_stride, src1->nb[0], MIN(end_i - i, block_i) * sizeof(__fp16), src1_ne0);
+    const HVX_Vector offsets = hvx_vec_gather_offsets_h(spad1_stride);
+    const uint32_t block_cols_padded = MIN(hex_round_up(src1_ne0, 64), block_cols);
+    const uint32_t src0_row_bytes    = src0_ne0 * sizeof(__fp16);
+    const uint32_t src0_row_padded   = hex_round_up(src0_row_bytes, VLEN);
+    const uint32_t src0_pre          = src0_row_padded - src0_row_bytes;
+    const uint32_t spad0_stride      = src0_row_padded + block_cols_padded * sizeof(__fp16);
 
-    const dma_addr_t src0_addr = concat_plane_addr(src0, p, div_ne2, ne2) + i * src0->nb[1];
-    dma_queue_push(dma_q, dma_make_data(spad0_base + src0_pre, src0_addr), spad0_row_bytes, src0->nb[1], src0_row_bytes, MIN(end_i - i, block_i));
+    // two slots per scratchpad: tile b gathers in slot b&1 while dst[b-1] and src1[b+1] DMA in the other
+    uint8_t * spad0_base = cctx->spad0_base + ith * cctx->spad0_size_per_thread;
+    uint8_t * spad1_base = cctx->spad1_base + ith * cctx->spad1_size_per_thread;
+    const uint32_t spad0_slot = block_rows * spad0_stride;
+    const uint32_t spad1_slot = block_cols_padded * spad1_stride;
 
-    dma_queue_pop(dma_q); // src1
+    // tiles walk (plane, row, col) with col fastest; src0 goes out with the col == 0 tile only
+    uint32_t plane = 0;
+    uint32_t row   = row_begin;
+    uint32_t col   = 0;
+    uint32_t slot  = 0;
+    bool dst_pending = false;
 
-    while (p < cctx->nplanes) {
-        const uint32_t current_block_i = MIN(end_i - i, block_i);
+    const dma_addr_t src1_addr = concat_plane_addr(src1, plane, div_ne2, ne2) + row * src1->nb[1];
+    dma_queue_push(dma_q, dma_make_data(spad1_base, src1_addr), spad1_stride, src1->nb[0], MIN(row_end - row, block_rows) * sizeof(__fp16), MIN(src1_ne0, block_cols));
 
-        htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) i);
-        for (uint32_t j = 0; j < src1_ne0; j += 64) {
-            const uint8_t * src_ptr = spad1_base + j * spad1_stride;
-            uint8_t * dst_ptr = spad0_base + src0_row_padded + j * sizeof(__fp16);
-            hvx_transpose_64x64_h_gather(dst_ptr, spad0_row_bytes, src_ptr, spad1_stride, offsets, current_block_i, MIN(src1_ne0 - j, 64));
+    const dma_addr_t src0_addr = concat_plane_addr(src0, plane, div_ne2, ne2) + row * src0->nb[1];
+    dma_queue_push(dma_q, dma_make_data(spad0_base + src0_pre, src0_addr), spad0_stride, src0->nb[1], src0_row_bytes, MIN(row_end - row, block_rows));
+
+    while (plane < cctx->nplanes) {
+        const uint32_t nrows = MIN(row_end - row, block_rows);
+        const uint32_t ncols = MIN(src1_ne0 - col, block_cols);
+
+        uint8_t * spad0 = spad0_base + slot * spad0_slot;
+        uint8_t * spad1 = spad1_base + slot * spad1_slot;
+
+        dma_queue_pop(dma_q); // src1
+        if (col == 0) {
+            dma_queue_pop(dma_q); // src0
         }
-        hvx_gather_sync(spad0_base + src0_row_padded);
-        htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) i);
 
-        uint32_t np = p;
-        uint32_t ni = i + block_i;
-        if (ni >= end_i) {
-            ni = start_i;
-            np++;
+        uint32_t next_plane = plane;
+        uint32_t next_row   = row;
+        uint32_t next_col   = col + block_cols;
+        if (next_col >= src1_ne0) {
+            next_col = 0;
+            next_row = row + block_rows;
+            if (next_row >= row_end) {
+                next_row = row_begin;
+                next_plane++;
+            }
         }
-        const bool has_next = np < cctx->nplanes;
-        const uint32_t next_block_i = MIN(end_i - ni, block_i);
+        const bool has_next = next_plane < cctx->nplanes;
+        const uint32_t next_nrows = MIN(row_end - next_row, block_rows);
+        const uint32_t next_ncols = MIN(src1_ne0 - next_col, block_cols);
 
-        // spad1 is free after the gather sync, prefetch next src1 ahead of the dst write
+        // other spad1 slot is free since the previous gather sync
         if (has_next) {
-            const dma_addr_t nsrc1_addr = concat_plane_addr(src1, np, div_ne2, ne2) + ni * src1->nb[1];
-            dma_queue_push(dma_q, dma_make_data(spad1_base, nsrc1_addr), spad1_stride, src1->nb[0], next_block_i * sizeof(__fp16), src1_ne0);
+            const dma_addr_t addr = concat_plane_addr(src1, next_plane, div_ne2, ne2) + next_row * src1->nb[1] + next_col * src1->nb[0];
+            dma_queue_push(dma_q, dma_make_data(spad1_base + (slot ^ 1) * spad1_slot, addr), spad1_stride, src1->nb[0], next_nrows * sizeof(__fp16), next_ncols);
         }
 
-        dma_queue_pop(dma_q); // src0
-
-        const dma_addr_t dst_addr = concat_plane_addr(dst, p, div_ne2, ne2) + i * dst->nb[1];
-        dma_queue_push(dma_q, dma_make_data(dst_addr, spad0_base + src0_pre), dst->nb[1], spad0_row_bytes, (src0_ne0 + src1_ne0) * sizeof(__fp16), current_block_i);
-
-        if (has_next) {
-            dma_queue_pop(dma_q); // next src1
+        htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) row);
+        for (uint32_t c = 0; c < ncols; c += 64) {
+            hvx_transpose_64x64_h_gather(spad0 + src0_row_padded + c * sizeof(__fp16), spad0_stride, spad1 + c * spad1_stride, spad1_stride, offsets, nrows, MIN(ncols - c, 64));
         }
-        dma_queue_pop(dma_q); // dst
+        hvx_gather_sync(spad0 + src0_row_padded);
+        htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) row);
 
-        // spad0 is free after the dst write
-        if (has_next) {
-            const dma_addr_t nsrc0_addr = concat_plane_addr(src0, np, div_ne2, ne2) + ni * src0->nb[1];
-            dma_queue_push(dma_q, dma_make_data(spad0_base + src0_pre, nsrc0_addr), spad0_row_bytes, src0->nb[1], src0_row_bytes, next_block_i);
+        // other spad0 slot is free once dst[b-1] is written
+        if (dst_pending) {
+            dma_queue_pop(dma_q); // dst
+        }
+        if (has_next && next_col == 0) {
+            const dma_addr_t addr = concat_plane_addr(src0, next_plane, div_ne2, ne2) + next_row * src0->nb[1];
+            dma_queue_push(dma_q, dma_make_data(spad0_base + (slot ^ 1) * spad0_slot + src0_pre, addr), spad0_stride, src0->nb[1], src0_row_bytes, next_nrows);
         }
 
-        p = np;
-        i = ni;
+        // the col == 0 tile also carries the src0 columns in front
+        const uint32_t lead = (col == 0) ? src0_row_bytes : 0;
+        const dma_addr_t dst_addr = concat_plane_addr(dst, plane, div_ne2, ne2) + row * dst->nb[1] + src0_row_bytes + col * sizeof(__fp16) - lead;
+        dma_queue_push(dma_q, dma_make_data(dst_addr, spad0 + src0_row_padded - lead), dst->nb[1], spad0_stride, lead + ncols * sizeof(__fp16), nrows);
+        dst_pending = true;
+
+        plane = next_plane;
+        row   = next_row;
+        col   = next_col;
+        slot ^= 1;
     }
     dma_queue_flush(dma_q);
 }

ggml/src/ggml-hexagon/htp/concat-ops.h

@@ -53,13 +53,17 @@ static inline void htp_concat_transposed_vtcm_layout_build(
     uint32_t type_size,
     uint32_t n_threads) {
 
-    uint32_t block_i = (type_size == 4) ? 32 : 64;
-    uint32_t spad1_stride = block_i * type_size;
-    uint32_t src1_ne0_padded = hex_round_up(src1_ne0, block_i);
-    uint32_t spad0_row_bytes = hex_round_up(src0_ne0 * type_size, 128) + src1_ne0_padded * type_size;
+    // src1_ne0 is processed in chunks of block_cols tokens, so VTCM does not grow with the batch
+    uint32_t block_rows = 128;
+    uint32_t block_cols = (type_size == 4) ? 256 : 512;
+    // +1 element spreads the gather lanes across VTCM, so they do not serialize on one path
+    uint32_t spad1_stride = block_rows * type_size + type_size;
+    uint32_t block_cols_padded = MIN(hex_round_up(src1_ne0, (type_size == 4) ? 32 : 64), block_cols);
+    uint32_t spad0_row_bytes = hex_round_up(src0_ne0 * type_size, 128) + block_cols_padded * type_size;
 
-    L->src0_spad_size_per_thread = block_i * spad0_row_bytes;
-    L->src1_spad_size_per_thread = src1_ne0_padded * spad1_stride;
+    // double-buffered: two slots per thread
+    L->src0_spad_size_per_thread = 2 * block_rows * spad0_row_bytes;
+    L->src1_spad_size_per_thread = 2 * block_cols_padded * spad1_stride;
 
     size_t off = 0;
     VTCM_LAYOUT_ALLOC(off, off_spad0, (size_t) n_threads * L->src0_spad_size_per_thread);

ggml/src/ggml-hexagon/htp/gated-delta-net-ops.c

@@ -1456,9 +1456,10 @@ static inline void gdn_dma_push_chunk_gb(
     const dma_addr_t g_dma    = g->data + (uint64_t) iv3 * g->nb[3] + (uint64_t) t_chunk * g->nb[2] + (uint64_t) iv1 * g->nb[1];
     const dma_addr_t beta_dma = beta->data + (uint64_t) iv3 * beta->nb[3] + (uint64_t) t_chunk * beta->nb[2] + (uint64_t) iv1 * beta->nb[1];
     const uint32_t row_bytes  = n_batch * sizeof(float);
+    const uint32_t dst_stride = (n_batch > 1) ? row_bytes + sizeof(float) : row_bytes;
 
-    dma_queue_push(dma_q, dma_make_data(vtcm_g_raw, g_dma), row_bytes, g->nb[2], row_bytes, chunk_size);
-    dma_queue_push(dma_q, dma_make_data(vtcm_b_raw, beta_dma), row_bytes, beta->nb[2], row_bytes, chunk_size);
+    dma_queue_push(dma_q, dma_make_data(vtcm_g_raw, g_dma), dst_stride, g->nb[2], row_bytes, chunk_size);
+    dma_queue_push(dma_q, dma_make_data(vtcm_b_raw, beta_dma), dst_stride, beta->nb[2], row_bytes, chunk_size);
 }
 
 static inline void gdn_pack_s_col_tiles(
@@ -1707,19 +1708,16 @@ static void gdn_hvx_phase1a_worker(unsigned int n, unsigned int i, void * data)
         hvx_vmem(head->b_f32[curr_buf] + 0)  = hvx_vmem(bctx->vtcm_b_raw + 0);
         hvx_vmem(head->b_f32[curr_buf] + 32) = hvx_vmem(bctx->vtcm_b_raw + 32);
     } else {
-        int32_t offsets[32] __attribute__((aligned(128)));
-        for (int k = 0; k < 32; ++k) {
-            offsets[k] = k * n_batch * sizeof(float);
-        }
-        HVX_Vector vv = *(const HVX_Vector *) offsets;
+        const uint32_t vtcm_stride = (n_batch + 1) * sizeof(float);
+        const HVX_Vector vv = hvx_vec_gather_offsets_w(vtcm_stride);
         const size_t rt_g = (size_t) ((const uint8_t *) bctx->vtcm_g_raw + i * sizeof(float));
         const size_t rt_b = (size_t) ((const uint8_t *) bctx->vtcm_b_raw + i * sizeof(float));
-        const size_t mu   = 64 * n_batch * sizeof(float);
+        const size_t mu   = 64 * vtcm_stride;
 
         Q6_vgather_ARMVw((HVX_Vector *) (head->g_f32[curr_buf] + 0),  rt_g, mu, vv);
-        Q6_vgather_ARMVw((HVX_Vector *) (head->g_f32[curr_buf] + 32), rt_g + 32 * n_batch * sizeof(float), mu, vv);
+        Q6_vgather_ARMVw((HVX_Vector *) (head->g_f32[curr_buf] + 32), rt_g + 32 * vtcm_stride, mu, vv);
         Q6_vgather_ARMVw((HVX_Vector *) (head->b_f32[curr_buf] + 0),  rt_b, mu, vv);
-        Q6_vgather_ARMVw((HVX_Vector *) (head->b_f32[curr_buf] + 32), rt_b + 32 * n_batch * sizeof(float), mu, vv);
+        Q6_vgather_ARMVw((HVX_Vector *) (head->b_f32[curr_buf] + 32), rt_b + 32 * vtcm_stride, mu, vv);
     }
 
     const uint32_t t_chunk = bctx->c * 64;

ggml/src/ggml-hexagon/htp/gated-delta-net-ops.h

@@ -204,7 +204,7 @@ static inline void htp_gdn_hmx_vtcm_layout_build(
     VTCM_LAYOUT_ALLOC_OPTIONAL(off, off_g_f32[1], bh * dma_scalar_sz, pipeline);
     VTCM_LAYOUT_ALLOC(off, off_b_f32[0], bh * dma_scalar_sz);
     VTCM_LAYOUT_ALLOC_OPTIONAL(off, off_b_f32[1], bh * dma_scalar_sz, pipeline);
-    const size_t raw_gb_sz = hex_round_up(bh * chunk_size * sizeof(float), 128);
+    const size_t raw_gb_sz = hex_round_up((bh + 1) * chunk_size * sizeof(float), 128);
     VTCM_LAYOUT_ALLOC(off, off_g_raw[0], raw_gb_sz);
     VTCM_LAYOUT_ALLOC_OPTIONAL(off, off_g_raw[1], raw_gb_sz, pipeline);
     VTCM_LAYOUT_ALLOC(off, off_b_raw[0], raw_gb_sz);

ggml/src/ggml-hexagon/htp/htp-ctx.h

@@ -188,5 +188,6 @@ int op_pool_2d(struct htp_ops_context * octx);
 int op_pool_1d(struct htp_ops_context * octx);
 int op_allreduce(struct htp_ops_context * octx);
 int op_roll(struct htp_ops_context * octx);
+int op_upscale(struct htp_ops_context * octx);
 
 #endif /* HTP_CTX_H */

ggml/src/ggml-hexagon/htp/htp-ops.h

@@ -116,6 +116,7 @@ enum htp_op_code {
     HTP_OP_GLU_SIGMOID,
     HTP_OP_UNARY_CEIL,
     HTP_OP_UNARY_SGN,
+    HTP_OP_UPSCALE,
 
     HTP_OP_INVALID
 };

ggml/src/ggml-hexagon/htp/main.c

@@ -1049,6 +1049,9 @@ static int execute_op(struct htp_ops_context * octx) {
         case HTP_OP_TRI:
             return op_unary(octx);
 
+        case HTP_OP_UPSCALE:
+            return op_upscale(octx);
+
         case HTP_OP_INVALID:
             break;
     }

ggml/src/ggml-hexagon/htp/ssm-conv.c

@@ -173,11 +173,12 @@ static inline void hvx_ssm_conv_decode_4(const float * x, const float * w, float
 // the lanes past the channel tail.
 static inline void hvx_ssm_conv_transpose_block(const float * raw_block,
                                                 float *       T,
+                                                uint32_t      raw_stride_bytes,
                                                 uint32_t      ncs,
                                                 uint32_t      cb_n,
                                                 HVX_Vector    vv) {
     const size_t   base = (size_t) raw_block;
-    const uint32_t mu   = cb_n * ncs * sizeof(float) - 1;
+    const uint32_t mu   = cb_n * raw_stride_bytes - 1;
 
     for (uint32_t t = 0; t < ncs; ++t) {
         Q6_vgather_ARMVw((HVX_Vector *) (T + (size_t) t * VLEN_FP32), base + t * sizeof(float), mu, vv);
@@ -393,7 +394,8 @@ static void ssm_conv_thread_f32_prefill(unsigned int nth, unsigned int ith, void
 
     // src0 spad holds two raw tiles (fetch of tile n+1 overlaps compute of tile n) plus
     // the transposed block. dst spad holds two tiles so a writeback can stay in flight.
-    const size_t src0_tile_raw_bytes = hex_round_up(d_inner_tile * ncs * sizeof(float), 128);
+    const size_t raw_stride_bytes    = ncs * sizeof(float) + sizeof(float);
+    const size_t src0_tile_raw_bytes = hex_round_up(d_inner_tile * raw_stride_bytes, 128);
     const size_t dst_tile_bytes      = hex_round_up(d_inner_tile * n_t * sizeof(float), 128);
 
     float * src0_tile_raw[2] = { (float *) src0_spad_base, (float *) (src0_spad_base + src0_tile_raw_bytes) };
@@ -419,7 +421,7 @@ static void ssm_conv_thread_f32_prefill(unsigned int nth, unsigned int ith, void
                                           (ir0 + tile_off_fetch) * src0_stride_inner_bytes;       \
             dma_queue_push(dma_q,                                                                 \
                            dma_make_data((uint8_t *) src0_tile_raw[(c) & 1], fetch_ddr),          \
-                           row_bytes, src0_stride_inner_bytes, row_bytes, cur_tile_n);            \
+                           raw_stride_bytes, src0_stride_inner_bytes, row_bytes, cur_tile_n);     \
             tile_off_fetch += d_inner_tile;                                                       \
             if (tile_off_fetch >= d_inner_per_thread) {                                           \
                 tile_off_fetch = 0;                                                               \
@@ -446,11 +448,7 @@ static void ssm_conv_thread_f32_prefill(unsigned int nth, unsigned int ith, void
     const uint32_t C_TILE = VLEN_FP32;
 
     // gather offsets: lane r reads channel r of the raw tile block
-    uint32_t __attribute__((aligned(VLEN))) gather_off[VLEN_FP32];
-    for (uint32_t r = 0; r < VLEN_FP32; ++r) {
-        gather_off[r] = r * ncs * sizeof(float);
-    }
-    const HVX_Vector vv = *(const HVX_Vector *) gather_off;
+    const HVX_Vector vv = hvx_vec_gather_offsets_w(raw_stride_bytes);
 
     uint32_t i3       = 0;
     uint32_t tile_off = 0;
@@ -474,7 +472,8 @@ static void ssm_conv_thread_f32_prefill(unsigned int nth, unsigned int ith, void
             const uint32_t cb_n = MIN(C_TILE, tile_n - cb);
 
             htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_A_PREP, (uint16_t) tile_off);
-            hvx_ssm_conv_transpose_block(raw + (size_t) cb * ncs, src0_T, ncs, cb_n, vv);
+            const float * raw_ch = (const float *) ((const uint8_t *) raw + (size_t) cb * raw_stride_bytes);
+            hvx_ssm_conv_transpose_block(raw_ch, src0_T, raw_stride_bytes, ncs, cb_n, vv);
             htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_A_PREP, (uint16_t) tile_off);
 
             const float * restrict wp = src1_T + tile_off + cb;

ggml/src/ggml-hexagon/htp/ssm-conv.h

@@ -54,11 +54,12 @@ static inline void htp_ssm_conv_vtcm_layout_build(
         L->src0_bytes_per_thread         = 2 * src0_tile_raw_bytes + src0_T_bytes;
         L->dst_bytes_per_thread          = 2 * hex_round_up((size_t) d_inner_tile * sizeof(float), 128);
     } else {
-        const uint32_t ncs          = d_conv - 1 + n_t;
-        const size_t   src0_block_T = hex_round_up((size_t) ncs * 32 * sizeof(float), 128);
-        const size_t   src0_tile_raw = hex_round_up((size_t) d_inner_tile * ncs * sizeof(float), 128);
-        L->src0_bytes_per_thread    = 2 * src0_tile_raw + src0_block_T;
-        L->dst_bytes_per_thread     = 2 * hex_round_up((size_t) d_inner_tile * n_t * sizeof(float), 128);
+        const uint32_t ncs              = d_conv - 1 + n_t;
+        const size_t   raw_stride_bytes = ncs * sizeof(float) + sizeof(float);
+        const size_t   src0_block_T     = hex_round_up((size_t) ncs * 32 * sizeof(float), 128);
+        const size_t   src0_tile_raw    = hex_round_up((size_t) d_inner_tile * raw_stride_bytes, 128);
+        L->src0_bytes_per_thread        = 2 * src0_tile_raw + src0_block_T;
+        L->dst_bytes_per_thread         = 2 * hex_round_up((size_t) d_inner_tile * n_t * sizeof(float), 128);
     }
 
     L->src0_bytes = L->src0_bytes_per_thread * n_threads;

ggml/src/ggml-hexagon/htp/upscale-ops.c

@@ -0,0 +1,958 @@
+#pragma clang diagnostic ignored "-Wunused-variable"
+
+#include <float.h>
+#include <math.h>
+#include <HAP_farf.h>
+
+#include "hex-common.h"
+#include "dma-queue.h"
+#include "hex-profile.h"
+#include "htp-ctx.h"
+#include "htp-ops.h"
+#include "htp-tensor.h"
+#include "hvx-types.h"
+#include "hvx-base.h"
+#include "hvx-utils.h"
+#include "upscale-ops.h"
+
+static inline int32_t clamp_i32(int32_t v, int32_t lo, int32_t hi) {
+    return v < lo ? lo : (v > hi ? hi : v);
+}
+
+static inline float clamp_f32(float v, float lo, float hi) {
+    return v < lo ? lo : (v > hi ? hi : v);
+}
+
+static inline float bicubic_weight1(float x) {
+    const float a = -0.75f;
+    return ((a + 2.0f) * x - (a + 3.0f)) * x * x + 1.0f;
+}
+
+static inline float bicubic_weight2(float x) {
+    const float a = -0.75f;
+    return ((a * x - 5.0f * a) * x + 8.0f * a) * x - 4.0f * a;
+}
+
+static inline void bicubic_weights(float x, float w[4]) {
+    w[0] = bicubic_weight2(x + 1.0f);
+    w[1] = bicubic_weight1(x);
+    w[2] = bicubic_weight1(1.0f - x);
+    w[3] = bicubic_weight2(2.0f - x);
+}
+
+static void upscale_init_luts_nearest(
+    uint32_t * lut_x,
+    uint32_t * lut_y,
+    const struct htp_tensor * src0,
+    const struct htp_tensor * dst,
+    const struct htp_upscale_kernel_params * p
+) {
+    const uint32_t src_x = src0->ne[0];
+    const uint32_t src_y = src0->ne[1];
+    const uint32_t dst_x = dst->ne[0];
+    const uint32_t dst_y = dst->ne[1];
+    const uint32_t dst_x_aligned = (dst_x + VLEN_FP32 - 1) & ~(VLEN_FP32 - 1);
+
+    for (uint32_t ox = 0; ox < dst_x_aligned; ++ox) {
+        if (ox < dst_x) {
+            uint32_t x0 = (uint32_t) ((float) ox * p->inv_sf0);
+            lut_x[ox] = MIN(x0, src_x - 1) * sizeof(float);
+        } else {
+            lut_x[ox] = 0;
+        }
+    }
+    for (uint32_t oy = 0; oy < dst_y; ++oy) {
+        uint32_t y0 = (uint32_t) ((float) oy * p->inv_sf1);
+        lut_y[oy] = MIN(y0, src_y - 1);
+    }
+}
+
+static void upscale_init_luts_bilinear(
+    struct upscale_bilinear_x_chunk * lut_x,
+    struct upscale_bilinear_y * lut_y,
+    const struct htp_tensor * src0,
+    const struct htp_tensor * dst,
+    const struct htp_upscale_kernel_params * p
+) {
+    const uint32_t src_x = src0->ne[0];
+    const uint32_t src_y = src0->ne[1];
+    const uint32_t dst_x = dst->ne[0];
+    const uint32_t dst_y = dst->ne[1];
+    const uint32_t n_chunks = (dst_x + VLEN_FP32 - 1) / VLEN_FP32;
+
+    for (uint32_t c = 0; c < n_chunks; ++c) {
+        struct upscale_bilinear_x_chunk * chunk = &lut_x[c];
+        const uint32_t ox_start = c * VLEN_FP32;
+        for (uint32_t l = 0; l < VLEN_FP32; ++l) {
+            const uint32_t ox = ox_start + l;
+            if (ox < dst_x) {
+                float xf = ((float) ox + p->pixel_offset) * p->inv_sf0 - p->pixel_offset;
+                int32_t x0 = (int32_t) floorf(xf);
+                int32_t x1 = x0 + 1;
+                x0 = clamp_i32(x0, 0, (int32_t) src_x - 1);
+                x1 = clamp_i32(x1, 0, (int32_t) src_x - 1);
+                float dx = clamp_f32(xf - (float) x0, 0.0f, 1.0f);
+
+                chunk->off0[l] = x0 * (int32_t) sizeof(float);
+                chunk->off1[l] = x1 * (int32_t) sizeof(float);
+                chunk->w0[l]   = 1.0f - dx;
+                chunk->dx[l]   = dx;
+            } else {
+                chunk->off0[l] = 0;
+                chunk->off1[l] = 0;
+                chunk->w0[l]   = 0.0f;
+                chunk->dx[l]   = 0.0f;
+            }
+        }
+    }
+    for (uint32_t oy = 0; oy < dst_y; ++oy) {
+        float yf = ((float) oy + p->pixel_offset) * p->inv_sf1 - p->pixel_offset;
+        int32_t y0 = (int32_t) floorf(yf);
+        int32_t y1 = y0 + 1;
+        y0 = clamp_i32(y0, 0, (int32_t) src_y - 1);
+        y1 = clamp_i32(y1, 0, (int32_t) src_y - 1);
+        float dy = clamp_f32(yf - (float) y0, 0.0f, 1.0f);
+
+        lut_y[oy].y0  = (uint32_t) y0;
+        lut_y[oy].y1  = (uint32_t) y1;
+        lut_y[oy].wy1 = dy;
+        lut_y[oy].wy0 = 1.0f - dy;
+    }
+}
+
+static void upscale_init_luts_bicubic(
+    struct upscale_bicubic_x * lut_x,
+    struct upscale_bicubic_y * lut_y,
+    const struct htp_tensor * src0,
+    const struct htp_tensor * dst,
+    const struct htp_upscale_kernel_params * p
+) {
+    const uint32_t src_x = src0->ne[0];
+    const uint32_t src_y = src0->ne[1];
+    const uint32_t dst_x = dst->ne[0];
+    const uint32_t dst_y = dst->ne[1];
+
+    for (uint32_t ox = 0; ox < dst_x; ++ox) {
+        float xf = ((float) ox + p->pixel_offset) * p->inv_sf0 - p->pixel_offset;
+        int32_t x0 = (int32_t) floorf(xf);
+        float dx = xf - (float) x0;
+        bicubic_weights(dx, lut_x[ox].w);
+
+        lut_x[ox].x[0] = clamp_i32(x0 - 1, 0, (int32_t) src_x - 1);
+        lut_x[ox].x[1] = clamp_i32(x0 + 0, 0, (int32_t) src_x - 1);
+        lut_x[ox].x[2] = clamp_i32(x0 + 1, 0, (int32_t) src_x - 1);
+        lut_x[ox].x[3] = clamp_i32(x0 + 2, 0, (int32_t) src_x - 1);
+    }
+    for (uint32_t oy = 0; oy < dst_y; ++oy) {
+        float yf = ((float) oy + p->pixel_offset) * p->inv_sf1 - p->pixel_offset;
+        int32_t y0 = (int32_t) floorf(yf);
+        float dy = yf - (float) y0;
+        bicubic_weights(dy, lut_y[oy].w);
+
+        lut_y[oy].y[0] = clamp_i32(y0 - 1, 0, (int32_t) src_y - 1);
+        lut_y[oy].y[1] = clamp_i32(y0 + 0, 0, (int32_t) src_y - 1);
+        lut_y[oy].y[2] = clamp_i32(y0 + 1, 0, (int32_t) src_y - 1);
+        lut_y[oy].y[3] = clamp_i32(y0 + 2, 0, (int32_t) src_y - 1);
+    }
+}
+
+static void upscale_init_luts_antialias(
+    struct upscale_antialias_x * lut_x,
+    struct upscale_antialias_y * lut_y,
+    const struct htp_tensor * src0,
+    const struct htp_tensor * dst,
+    const struct htp_upscale_kernel_params * p
+) {
+    const uint32_t src_x = src0->ne[0];
+    const uint32_t src_y = src0->ne[1];
+    const uint32_t dst_x = dst->ne[0];
+    const uint32_t dst_y = dst->ne[1];
+
+    const float filter_radius0 = p->filter_radius0;
+    const float filter_radius1 = p->filter_radius1;
+
+    for (uint32_t ox = 0; ox < dst_x; ++ox) {
+        const float xf = ((float) ox + p->pixel_offset) * p->inv_sf0;
+        int32_t x_min = (int32_t) MAX(xf - filter_radius0 + p->pixel_offset, 0.0f);
+        int32_t x_max = (int32_t) MIN(xf + filter_radius0 + p->pixel_offset, (float) src_x);
+        lut_x[ox].x_min = clamp_i32(x_min, 0, (int32_t) src_x);
+        lut_x[ox].x_max = clamp_i32(x_max, 0, (int32_t) src_x);
+        lut_x[ox].xf    = xf;
+    }
+    for (uint32_t oy = 0; oy < dst_y; ++oy) {
+        const float yf = ((float) oy + p->pixel_offset) * p->inv_sf1;
+        int32_t y_min = (int32_t) MAX(yf - filter_radius1 + p->pixel_offset, 0.0f);
+        int32_t y_max = (int32_t) MIN(yf + filter_radius1 + p->pixel_offset, (float) src_y);
+        lut_y[oy].y_min = clamp_i32(y_min, 0, (int32_t) src_y);
+        lut_y[oy].y_max = clamp_i32(y_max, 0, (int32_t) src_y);
+        lut_y[oy].yf    = yf;
+    }
+}
+
+static inline void get_src_y_window(
+    const struct htp_tensor * src0,
+    const struct htp_tensor * dst,
+    const struct htp_upscale_kernel_params * p,
+    const void * lut_y,
+    uint32_t dst_y_start,
+    uint32_t dst_y_count,
+    uint32_t * out_src_y_start,
+    uint32_t * out_src_y_count
+) {
+    const uint32_t src_y = src0->ne[1];
+    const uint32_t dst_y = dst->ne[1];
+
+    if (dst_y_count == dst_y) {
+        *out_src_y_start = 0;
+        *out_src_y_count = src_y;
+        return;
+    }
+
+    const uint32_t last_oy = dst_y_start + dst_y_count - 1;
+
+    if (p->mode == HTP_UPSCALE_MODE_NEAREST) {
+        const uint32_t * ly = (const uint32_t *) lut_y;
+        const uint32_t y0 = ly[dst_y_start];
+        const uint32_t y1 = ly[last_oy];
+        *out_src_y_start = y0;
+        *out_src_y_count = y1 - y0 + 1;
+    } else if (p->mode == HTP_UPSCALE_MODE_BICUBIC) {
+        const struct upscale_bicubic_y * ly = (const struct upscale_bicubic_y *) lut_y;
+        const int32_t y0 = ly[dst_y_start].y[0];
+        const int32_t y1 = ly[last_oy].y[3];
+        *out_src_y_start = (uint32_t) y0;
+        *out_src_y_count = (uint32_t) (y1 - y0 + 1);
+    } else if (p->antialias) {
+        const struct upscale_antialias_y * ly = (const struct upscale_antialias_y *) lut_y;
+        const int32_t y0 = ly[dst_y_start].y_min;
+        const int32_t y1 = ly[last_oy].y_max;
+        *out_src_y_start = (uint32_t) y0;
+        *out_src_y_count = (uint32_t) (y1 - y0);
+    } else {
+        const struct upscale_bilinear_y * ly = (const struct upscale_bilinear_y *) lut_y;
+        const uint32_t y0 = ly[dst_y_start].y0;
+        const uint32_t y1 = ly[last_oy].y1;
+        *out_src_y_start = y0;
+        *out_src_y_count = y1 - y0 + 1;
+    }
+}
+
+static inline void resolve_dst_slice(
+    const struct htp_tensor * dst,
+    const struct htp_upscale_kernel_params * p,
+    uint32_t chunk_idx,
+    dma_addr_t * out_slice_dst_off,
+    uint32_t * out_dst_ys,
+    uint32_t * out_dst_yc
+) {
+    const uint32_t s_idx = fastdiv(chunk_idx, &p->div_chunks_per_slice);
+    const uint32_t c_idx = chunk_idx - s_idx * p->chunks_per_slice;
+    const uint32_t d_ys  = c_idx * p->dst_nrows_per_chunk;
+    const uint32_t d_yc  = MIN(p->dst_nrows_per_chunk, dst->ne[1] - d_ys);
+    const uint32_t i3    = fastdiv(s_idx, &p->div_dst_z);
+    const uint32_t i2    = s_idx - i3 * dst->ne[2];
+
+    *out_slice_dst_off = (dma_addr_t) i2 * dst->nb[2] + (dma_addr_t) i3 * dst->nb[3];
+    *out_dst_ys        = d_ys;
+    *out_dst_yc        = d_yc;
+}
+
+static inline void resolve_src_slice(
+    const struct htp_tensor * src0,
+    const struct htp_tensor * dst,
+    const struct htp_upscale_kernel_params * p,
+    const void * lut_y,
+    uint32_t chunk_idx,
+    dma_addr_t * out_slice_src_off,
+    uint32_t * out_src_ys,
+    uint32_t * out_src_yc
+) {
+    const uint32_t s_idx = fastdiv(chunk_idx, &p->div_chunks_per_slice);
+    const uint32_t c_idx = chunk_idx - s_idx * p->chunks_per_slice;
+    const uint32_t d_ys  = c_idx * p->dst_nrows_per_chunk;
+    const uint32_t d_yc  = MIN(p->dst_nrows_per_chunk, dst->ne[1] - d_ys);
+    uint32_t s_ys, s_yc;
+    get_src_y_window(src0, dst, p, lut_y, d_ys, d_yc, &s_ys, &s_yc);
+
+    const uint32_t i3  = fastdiv(s_idx, &p->div_dst_z);
+    const uint32_t i2  = s_idx - i3 * dst->ne[2];
+    const uint32_t i03 = (uint32_t) ((float) i3 * p->inv_sf3);
+    const uint32_t i02 = (uint32_t) ((float) i2 * p->inv_sf2);
+
+    *out_slice_src_off = (dma_addr_t) i02 * src0->nb[2] + (dma_addr_t) i03 * src0->nb[3];
+    *out_src_ys        = s_ys;
+    *out_src_yc        = s_yc;
+}
+
+struct upscale_context;
+
+typedef void (*upscale_chunk_fn_t)(
+    const struct upscale_context * ctx,
+    const float * srcb,
+    float * dstb,
+    float * R_buf,
+    float * gather_tmp,
+    uint32_t dst_y_start,
+    uint32_t dst_y_count,
+    uint32_t src_y_start
+);
+
+struct upscale_context {
+    struct htp_ops_context * octx;
+    const struct htp_upscale_kernel_params * kparams;
+    struct htp_upscale_vtcm_layout layout;
+    upscale_chunk_fn_t compute_chunk;
+    const void * lut_x;
+    const void * lut_y;
+    uint32_t chunk_start;
+    uint32_t chunk_count;
+    uint32_t chunks_per_thread;
+};
+
+static void compute_nearest_chunk(
+    const struct upscale_context * ctx,
+    const float * srcb,
+    float * dstb,
+    float * R_buf,
+    float * gather_tmp,
+    uint32_t dst_y_start,
+    uint32_t dst_y_count,
+    uint32_t src_y_start
+) {
+    const uint32_t src_x = ctx->octx->src[0]->ne[0];
+    const uint32_t dst_x = ctx->octx->dst->ne[0];
+    const uint32_t * lut_x = (const uint32_t *) ctx->lut_x;
+    const uint32_t * lut_y = (const uint32_t *) ctx->lut_y;
+    const uint32_t region_size = hex_round_up(src_x * sizeof(float), 128);
+
+    for (uint32_t oy = dst_y_start; oy < dst_y_start + dst_y_count; ++oy) {
+        const uint32_t y0 = lut_y[oy];
+        const float * r0  = srcb + (y0 - src_y_start) * src_x;
+        float * dst_row   = dstb + (oy - dst_y_start) * dst_x;
+
+        for (uint32_t ix = 0; ix < src_x; ix += VLEN_FP32) {
+            *(HVX_Vector *) (R_buf + ix) = hvx_vmemu(r0 + ix);
+        }
+
+        for (uint32_t ox = 0; ox < dst_x; ox += VLEN_FP32) {
+            const uint32_t rem = dst_x - ox;
+            const uint32_t nbytes = (rem < VLEN_FP32) ? (rem * sizeof(float)) : VLEN;
+            const HVX_Vector voff = hvx_vmem(lut_x + ox);
+            Q6_vgather_ARMVw((HVX_Vector *) gather_tmp, (size_t) R_buf, region_size, voff);
+            hvx_gather_sync(gather_tmp);
+            hvx_vec_store_u(dst_row + ox, nbytes, hvx_vmem(gather_tmp));
+        }
+    }
+}
+
+static void compute_nearest_2x_chunk(
+    const struct upscale_context * ctx,
+    const float * srcb,
+    float * dstb,
+    float * R_buf,
+    float * gather_tmp,
+    uint32_t dst_y_start,
+    uint32_t dst_y_count,
+    uint32_t src_y_start
+) {
+    (void) R_buf;
+    (void) gather_tmp;
+    const uint32_t src_x = ctx->octx->src[0]->ne[0];
+    const uint32_t dst_x = ctx->octx->dst->ne[0];
+
+    for (uint32_t oy = dst_y_start; oy < dst_y_start + dst_y_count; ++oy) {
+        const uint32_t sy = oy >> 1;
+        const float * r0  = srcb + (sy - src_y_start) * src_x;
+        float * dst_row   = dstb + (oy - dst_y_start) * dst_x;
+
+        for (uint32_t ix = 0; ix < src_x; ix += VLEN_FP32) {
+            const HVX_Vector vin = hvx_vmemu(r0 + ix);
+
+            const HVX_VectorPair p = Q6_W_vshuff_VVR(vin, vin, -4);
+            const HVX_Vector v_lo = Q6_V_lo_W(p);
+            const HVX_Vector v_hi = Q6_V_hi_W(p);
+
+            const uint32_t ox0 = 2 * ix;
+            const uint32_t rem_out0 = (dst_x > ox0) ? (dst_x - ox0) : 0;
+            const uint32_t nbytes_out0 = (rem_out0 < VLEN_FP32) ? (rem_out0 * sizeof(float)) : VLEN;
+            hvx_vec_store_u(dst_row + ox0, nbytes_out0, v_lo);
+
+            const uint32_t ox1 = ox0 + VLEN_FP32;
+            if (dst_x > ox1) {
+                const uint32_t rem_out1 = dst_x - ox1;
+                const uint32_t nbytes_out1 = (rem_out1 < VLEN_FP32) ? (rem_out1 * sizeof(float)) : VLEN;
+                hvx_vec_store_u(dst_row + ox1, nbytes_out1, v_hi);
+            }
+        }
+    }
+}
+
+static void compute_nearest_4x_chunk(
+    const struct upscale_context * ctx,
+    const float * srcb,
+    float * dstb,
+    float * R_buf,
+    float * gather_tmp,
+    uint32_t dst_y_start,
+    uint32_t dst_y_count,
+    uint32_t src_y_start
+) {
+    (void) R_buf;
+    (void) gather_tmp;
+    const uint32_t src_x = ctx->octx->src[0]->ne[0];
+    const uint32_t dst_x = ctx->octx->dst->ne[0];
+
+    for (uint32_t oy = dst_y_start; oy < dst_y_start + dst_y_count; ++oy) {
+        const uint32_t sy = oy >> 2;
+        const float * r0  = srcb + (sy - src_y_start) * src_x;
+        float * dst_row   = dstb + (oy - dst_y_start) * dst_x;
+
+        for (uint32_t ix = 0; ix < src_x; ix += VLEN_FP32) {
+            const HVX_Vector vin = hvx_vmemu(r0 + ix);
+
+            const HVX_VectorPair p = Q6_W_vshuff_VVR(vin, vin, -4);
+            const HVX_Vector v_lo = Q6_V_lo_W(p);
+            const HVX_Vector v_hi = Q6_V_hi_W(p);
+
+            const HVX_VectorPair p0 = Q6_W_vshuff_VVR(v_lo, v_lo, -4);
+            const HVX_VectorPair p1 = Q6_W_vshuff_VVR(v_hi, v_hi, -4);
+
+            const HVX_Vector out0 = Q6_V_lo_W(p0);
+            const HVX_Vector out1 = Q6_V_hi_W(p0);
+            const HVX_Vector out2 = Q6_V_lo_W(p1);
+            const HVX_Vector out3 = Q6_V_hi_W(p1);
+
+            const uint32_t ox = 4 * ix;
+            if (dst_x > ox) {
+                const uint32_t rem = dst_x - ox;
+                hvx_vec_store_u(dst_row + ox, (rem < VLEN_FP32) ? (rem * sizeof(float)) : VLEN, out0);
+            }
+            if (dst_x > ox + VLEN_FP32) {
+                const uint32_t rem = dst_x - (ox + VLEN_FP32);
+                hvx_vec_store_u(dst_row + ox + VLEN_FP32, (rem < VLEN_FP32) ? (rem * sizeof(float)) : VLEN, out1);
+            }
+            if (dst_x > ox + 2 * VLEN_FP32) {
+                const uint32_t rem = dst_x - (ox + 2 * VLEN_FP32);
+                hvx_vec_store_u(dst_row + ox + 2 * VLEN_FP32, (rem < VLEN_FP32) ? (rem * sizeof(float)) : VLEN, out2);
+            }
+            if (dst_x > ox + 3 * VLEN_FP32) {
+                const uint32_t rem = dst_x - (ox + 3 * VLEN_FP32);
+                hvx_vec_store_u(dst_row + ox + 3 * VLEN_FP32, (rem < VLEN_FP32) ? (rem * sizeof(float)) : VLEN, out3);
+            }
+        }
+    }
+}
+
+static inline void compute_bilinear_chunk_32(
+    float * restrict dst_chunk,
+    const float * restrict R_buf,
+    float * restrict tmp_a,
+    float * restrict tmp_b,
+    uint32_t region_size,
+    const struct upscale_bilinear_x_chunk * restrict chunk,
+    uint32_t nbytes
+) {
+    const HVX_Vector voff0 = hvx_vmem(chunk->off0);
+    const HVX_Vector voff1 = hvx_vmem(chunk->off1);
+    const HVX_Vector vw0   = hvx_vmem(chunk->w0);
+    const HVX_Vector vdx   = hvx_vmem(chunk->dx);
+
+    Q6_vgather_ARMVw((HVX_Vector *) tmp_a, (size_t) R_buf, region_size, voff0);
+    Q6_vgather_ARMVw((HVX_Vector *) tmp_b, (size_t) R_buf, region_size, voff1);
+    hvx_gather_sync(tmp_b);
+
+    const HVX_Vector va = hvx_vmem(tmp_a);
+    const HVX_Vector vb = hvx_vmem(tmp_b);
+
+    const HVX_Vector val = hvx_vec_add_f32_f32(
+        hvx_vec_mul_f32_f32(vw0, va),
+        hvx_vec_mul_f32_f32(vdx, vb)
+    );
+
+    hvx_vec_store_u(dst_chunk, nbytes, val);
+}
+
+static void compute_bilinear_chunk(
+    const struct upscale_context * ctx,
+    const float * srcb,
+    float * dstb,
+    float * R_buf,
+    float * gather_tmp,
+    uint32_t dst_y_start,
+    uint32_t dst_y_count,
+    uint32_t src_y_start
+) {
+    const uint32_t src_x = ctx->octx->src[0]->ne[0];
+    const uint32_t dst_x = ctx->octx->dst->ne[0];
+    const struct upscale_bilinear_x_chunk * lut_x = (const struct upscale_bilinear_x_chunk *) ctx->lut_x;
+    const struct upscale_bilinear_y * lut_y = (const struct upscale_bilinear_y *) ctx->lut_y;
+    const uint32_t n_chunks = (dst_x + VLEN_FP32 - 1) / VLEN_FP32;
+    const uint32_t region_size = hex_round_up(src_x * sizeof(float), 128);
+    float * tmp_a = gather_tmp;
+    float * tmp_b = gather_tmp + VLEN_FP32;
+
+    for (uint32_t oy = dst_y_start; oy < dst_y_start + dst_y_count; ++oy) {
+        const struct upscale_bilinear_y * ey = &lut_y[oy];
+        const float * r0 = srcb + (ey->y0 - src_y_start) * src_x;
+        const float * r1 = srcb + (ey->y1 - src_y_start) * src_x;
+        float * dst_row  = dstb + (oy - dst_y_start) * dst_x;
+
+        if (ey->y0 == ey->y1) {
+            for (uint32_t ix = 0; ix < src_x; ix += VLEN_FP32) {
+                *(HVX_Vector *) (R_buf + ix) = hvx_vmemu(r0 + ix);
+            }
+        } else {
+            const HVX_Vector vwy0 = hvx_vec_splat_f32(ey->wy0);
+            const HVX_Vector vwy1 = hvx_vec_splat_f32(ey->wy1);
+            for (uint32_t ix = 0; ix < src_x; ix += VLEN_FP32) {
+                const HVX_Vector vr0 = hvx_vmemu(r0 + ix);
+                const HVX_Vector vr1 = hvx_vmemu(r1 + ix);
+                *(HVX_Vector *) (R_buf + ix) = hvx_vec_add_f32_f32(
+                    hvx_vec_mul_f32_f32(vwy0, vr0),
+                    hvx_vec_mul_f32_f32(vwy1, vr1)
+                );
+            }
+        }
+
+        for (uint32_t c = 0; c < n_chunks; ++c) {
+            const uint32_t ox = c * VLEN_FP32;
+            const uint32_t rem = dst_x - ox;
+            const uint32_t nbytes = (rem < VLEN_FP32) ? (rem * sizeof(float)) : VLEN;
+            compute_bilinear_chunk_32(dst_row + ox, R_buf, tmp_a, tmp_b, region_size, &lut_x[c], nbytes);
+        }
+    }
+}
+
+static void compute_bilinear_2x_chunk(
+    const struct upscale_context * ctx,
+    const float * srcb,
+    float * dstb,
+    float * R_buf,
+    float * gather_tmp,
+    uint32_t dst_y_start,
+    uint32_t dst_y_count,
+    uint32_t src_y_start
+) {
+    (void) R_buf;
+    (void) gather_tmp;
+    const uint32_t src_x = ctx->octx->src[0]->ne[0];
+    const uint32_t dst_x = ctx->octx->dst->ne[0];
+    const struct upscale_bilinear_y * lut_y = (const struct upscale_bilinear_y *) ctx->lut_y;
+    const HVX_Vector c_075 = hvx_vec_splat_f32(0.75f);
+    const HVX_Vector c_025 = hvx_vec_splat_f32(0.25f);
+
+    for (uint32_t oy = dst_y_start; oy < dst_y_start + dst_y_count; ++oy) {
+        const struct upscale_bilinear_y * ey = &lut_y[oy];
+        const float * r0 = srcb + (ey->y0 - src_y_start) * src_x;
+        const float * r1 = srcb + (ey->y1 - src_y_start) * src_x;
+        float * dst_row  = dstb + (oy - dst_y_start) * dst_x;
+
+        const bool same_row = (ey->y0 == ey->y1);
+        const HVX_Vector vwy0 = hvx_vec_splat_f32(ey->wy0);
+        const HVX_Vector vwy1 = hvx_vec_splat_f32(ey->wy1);
+
+        HVX_Vector v_curr = hvx_vmem(r0);
+        if (!same_row) {
+            const HVX_Vector v1 = hvx_vmem(r1);
+            v_curr = hvx_vec_add_f32_f32(hvx_vec_mul_f32_f32(vwy0, v_curr), hvx_vec_mul_f32_f32(vwy1, v1));
+        }
+
+        HVX_Vector v_prev = Q6_V_vror_VR(v_curr, 4);
+
+        for (uint32_t ix = 0; ix < src_x; ix += VLEN_FP32) {
+            HVX_Vector v_next_chunk;
+            const uint32_t next_ix = ix + VLEN_FP32;
+            if (next_ix < src_x) {
+                v_next_chunk = hvx_vmem(r0 + next_ix);
+                if (!same_row) {
+                    const HVX_Vector v1 = hvx_vmem(r1 + next_ix);
+                    v_next_chunk = hvx_vec_add_f32_f32(hvx_vec_mul_f32_f32(vwy0, v_next_chunk), hvx_vec_mul_f32_f32(vwy1, v1));
+                }
+            } else {
+                v_next_chunk = Q6_V_vror_VR(v_curr, 124);
+            }
+
+            const HVX_Vector v_shl = Q6_V_valign_VVR(v_curr, v_prev, 124);
+            const HVX_Vector v_shr = Q6_V_valign_VVR(v_next_chunk, v_curr, 4);
+
+            const HVX_Vector ve = hvx_vec_add_f32_f32(hvx_vec_mul_f32_f32(c_075, v_curr), hvx_vec_mul_f32_f32(c_025, v_shl));
+            const HVX_Vector vo = hvx_vec_add_f32_f32(hvx_vec_mul_f32_f32(c_075, v_curr), hvx_vec_mul_f32_f32(c_025, v_shr));
+
+            const HVX_VectorPair p = Q6_W_vshuff_VVR(vo, ve, -4);
+            const HVX_Vector out0 = Q6_V_lo_W(p);
+            const HVX_Vector out1 = Q6_V_hi_W(p);
+
+            const uint32_t ox0 = 2 * ix;
+            hvx_vmem(dst_row + ox0)             = out0;
+            hvx_vmem(dst_row + ox0 + VLEN_FP32) = out1;
+
+            v_prev = v_curr;
+            v_curr = v_next_chunk;
+        }
+    }
+}
+
+static void compute_bilinear_antialias_chunk(
+    const struct upscale_context * ctx,
+    const float * srcb,
+    float * dstb,
+    float * R_buf,
+    float * gather_tmp,
+    uint32_t dst_y_start,
+    uint32_t dst_y_count,
+    uint32_t src_y_start
+) {
+    (void) R_buf;
+    (void) gather_tmp;
+    const uint32_t src_x = ctx->octx->src[0]->ne[0];
+    const uint32_t dst_x = ctx->octx->dst->ne[0];
+    const struct htp_upscale_kernel_params * p = ctx->kparams;
+    const struct upscale_antialias_x * lut_x = (const struct upscale_antialias_x *) ctx->lut_x;
+    const struct upscale_antialias_y * lut_y = (const struct upscale_antialias_y *) ctx->lut_y;
+
+    const float invscale1 = p->invscale1;
+    const float invscale0 = p->invscale0;
+
+    for (uint32_t oy = dst_y_start; oy < dst_y_start + dst_y_count; ++oy) {
+        const struct upscale_antialias_y * ey = &lut_y[oy];
+        const float yf = ey->yf;
+        const int32_t y_min = ey->y_min;
+        const int32_t y_max = ey->y_max;
+
+        float * dst_row = dstb + (oy - dst_y_start) * dst_x;
+
+        for (uint32_t ox = 0; ox < dst_x; ox += VLEN_FP32) {
+            const uint32_t rem = dst_x - ox;
+            const uint32_t nbytes = (rem < VLEN_FP32) ? (rem * sizeof(float)) : VLEN;
+            float tmp[VLEN_FP32] __attribute__((aligned(128)));
+            const uint32_t cnt = rem < VLEN_FP32 ? rem : VLEN_FP32;
+
+            for (uint32_t l = 0; l < cnt; ++l) {
+                const uint32_t cur_ox = ox + l;
+                const struct upscale_antialias_x * ex = &lut_x[cur_ox];
+                const float xf = ex->xf;
+                const int32_t x_min = ex->x_min;
+                const int32_t x_max = ex->x_max;
+
+                float val = 0.0f;
+                float total_weight = 0.0f;
+
+                for (int32_t sy = y_min; sy < y_max; ++sy) {
+                    const float diff_y = fabsf((float) sy - yf + p->pixel_offset) * invscale1;
+                    const float weight_y = MAX(1.0f - diff_y, 0.0f);
+                    if (weight_y <= 0.0f) {
+                        continue;
+                    }
+                    const float * r = srcb + (uint32_t) (sy - (int32_t) src_y_start) * src_x;
+
+                    for (int32_t sx = x_min; sx < x_max; ++sx) {
+                        const float diff_x = fabsf((float) sx - xf + p->pixel_offset) * invscale0;
+                        const float weight_x = MAX(1.0f - diff_x, 0.0f);
+                        const float weight = weight_x * weight_y;
+                        if (weight <= 0.0f) {
+                            continue;
+                        }
+                        val += r[sx] * weight;
+                        total_weight += weight;
+                    }
+                }
+
+                tmp[l] = (total_weight > 0.0f) ? (val / total_weight) : 0.0f;
+            }
+
+            hvx_vec_store_u(dst_row + ox, nbytes, *(const HVX_Vector *) tmp);
+        }
+    }
+}
+
+static void compute_bicubic_chunk(
+    const struct upscale_context * ctx,
+    const float * srcb,
+    float * dstb,
+    float * R_buf,
+    float * gather_tmp,
+    uint32_t dst_y_start,
+    uint32_t dst_y_count,
+    uint32_t src_y_start
+) {
+    (void) R_buf;
+    (void) gather_tmp;
+    const uint32_t src_x = ctx->octx->src[0]->ne[0];
+    const uint32_t dst_x = ctx->octx->dst->ne[0];
+    const struct upscale_bicubic_x * lut_x = (const struct upscale_bicubic_x *) ctx->lut_x;
+    const struct upscale_bicubic_y * lut_y = (const struct upscale_bicubic_y *) ctx->lut_y;
+
+    for (uint32_t oy = dst_y_start; oy < dst_y_start + dst_y_count; ++oy) {
+        const struct upscale_bicubic_y * ey = &lut_y[oy];
+        const float * r0 = srcb + (uint32_t) (ey->y[0] - (int32_t) src_y_start) * src_x;
+        const float * r1 = srcb + (uint32_t) (ey->y[1] - (int32_t) src_y_start) * src_x;
+        const float * r2 = srcb + (uint32_t) (ey->y[2] - (int32_t) src_y_start) * src_x;
+        const float * r3 = srcb + (uint32_t) (ey->y[3] - (int32_t) src_y_start) * src_x;
+
+        float * dst_row = dstb + (oy - dst_y_start) * dst_x;
+
+        const HVX_Vector vwy0 = hvx_vec_splat_f32(ey->w[0]);
+        const HVX_Vector vwy1 = hvx_vec_splat_f32(ey->w[1]);
+        const HVX_Vector vwy2 = hvx_vec_splat_f32(ey->w[2]);
+        const HVX_Vector vwy3 = hvx_vec_splat_f32(ey->w[3]);
+
+        for (uint32_t ox = 0; ox < dst_x; ox += VLEN_FP32) {
+            const uint32_t rem = dst_x - ox;
+            const uint32_t nbytes = (rem < VLEN_FP32) ? (rem * sizeof(float)) : VLEN;
+
+            float v0[VLEN_FP32] __attribute__((aligned(128)));
+            float v1[VLEN_FP32] __attribute__((aligned(128)));
+            float v2[VLEN_FP32] __attribute__((aligned(128)));
+            float v3[VLEN_FP32] __attribute__((aligned(128)));
+
+            const uint32_t cnt = rem < VLEN_FP32 ? rem : VLEN_FP32;
+            for (uint32_t l = 0; l < cnt; ++l) {
+                const struct upscale_bicubic_x * e = &lut_x[ox + l];
+                v0[l] = r0[e->x[0]] * e->w[0] + r0[e->x[1]] * e->w[1] + r0[e->x[2]] * e->w[2] + r0[e->x[3]] * e->w[3];
+                v1[l] = r1[e->x[0]] * e->w[0] + r1[e->x[1]] * e->w[1] + r1[e->x[2]] * e->w[2] + r1[e->x[3]] * e->w[3];
+                v2[l] = r2[e->x[0]] * e->w[0] + r2[e->x[1]] * e->w[1] + r2[e->x[2]] * e->w[2] + r2[e->x[3]] * e->w[3];
+                v3[l] = r3[e->x[0]] * e->w[0] + r3[e->x[1]] * e->w[1] + r3[e->x[2]] * e->w[2] + r3[e->x[3]] * e->w[3];
+            }
+            for (uint32_t l = cnt; l < VLEN_FP32; ++l) {
+                v0[l] = v1[l] = v2[l] = v3[l] = 0.0f;
+            }
+
+            const HVX_Vector vv0 = *(const HVX_Vector *) v0;
+            const HVX_Vector vv1 = *(const HVX_Vector *) v1;
+            const HVX_Vector vv2 = *(const HVX_Vector *) v2;
+            const HVX_Vector vv3 = *(const HVX_Vector *) v3;
+
+            const HVX_Vector acc0 = hvx_vec_add_f32_f32(hvx_vec_mul_f32_f32(vwy0, vv0), hvx_vec_mul_f32_f32(vwy1, vv1));
+            const HVX_Vector acc1 = hvx_vec_add_f32_f32(hvx_vec_mul_f32_f32(vwy2, vv2), hvx_vec_mul_f32_f32(vwy3, vv3));
+            const HVX_Vector val  = hvx_vec_add_f32_f32(acc0, acc1);
+
+            hvx_vec_store_u(dst_row + ox, nbytes, val);
+        }
+    }
+}
+
+static inline void dma_push_src_chunk(
+    dma_queue * dma_q,
+    float * spad,
+    const struct htp_tensor * src0,
+    dma_addr_t slice_src_off,
+    uint32_t src_ys,
+    uint32_t src_yc
+) {
+    const uint32_t src_x = src0->ne[0];
+    const dma_addr_t src_addr = src0->data + slice_src_off + (dma_addr_t) src_ys * src0->nb[1];
+
+    if (src0->nb[0] == sizeof(float)) {
+        const uint32_t row_size   = src_x * sizeof(float);
+        const uint32_t dst_stride = row_size;
+        const uint32_t src_stride = src0->nb[1];
+        dma_queue_push(dma_q, dma_make_data(spad, src_addr), dst_stride, src_stride, row_size, src_yc);
+    } else {
+        const uint32_t row_size   = sizeof(float);
+        const uint32_t dst_stride = sizeof(float);
+        const uint32_t src_stride = src0->nb[0];
+        const uint32_t nrows      = src_x * src_yc;
+        dma_queue_push(dma_q, dma_make_data(spad, src_addr), dst_stride, src_stride, row_size, nrows);
+    }
+}
+
+static void upscale_worker(unsigned int nth, unsigned int ith, void * data) {
+    struct upscale_context * ctx = (struct upscale_context *) data;
+    const struct htp_upscale_kernel_params * p = ctx->kparams;
+    const struct htp_upscale_vtcm_layout * L = &ctx->layout;
+    const struct htp_tensor * src0 = ctx->octx->src[0];
+    const struct htp_tensor * dst  = ctx->octx->dst;
+
+    const uint32_t first = ctx->chunk_start + ith * ctx->chunks_per_thread;
+    const uint32_t last  = MIN(first + ctx->chunks_per_thread, ctx->chunk_start + ctx->chunk_count);
+
+    if (first >= last) {
+        return;
+    }
+
+    struct htp_thread_trace * tr = &ctx->octx->ctx->trace[ith];
+    dma_queue * dma_q            = ctx->octx->ctx->dma[ith];
+
+    const uint32_t src_spad_half = (uint32_t) L->src_spad_half_size;
+    const uint32_t dst_spad_half = (uint32_t) L->dst_spad_half_size;
+
+    uint8_t * vtcm_base  = (uint8_t *) ctx->octx->ctx->vtcm_base;
+    uint8_t * src_spad   = VTCM_LAYOUT_PTR(uint8_t, vtcm_base, L->off_src)    + ith * L->src_bytes_per_thread;
+    uint8_t * dst_spad   = VTCM_LAYOUT_PTR(uint8_t, vtcm_base, L->off_dst)    + ith * L->dst_bytes_per_thread;
+    float   * R_buf      = VTCM_LAYOUT_PTR(float,   vtcm_base, L->off_r)      + ith * (L->r_bytes_per_thread / sizeof(float));
+    float   * gather_tmp = VTCM_LAYOUT_PTR(float,   vtcm_base, L->off_gather) + ith * (L->gather_bytes_per_thread / sizeof(float));
+    const void * lut_y   = ctx->lut_y;
+
+    float * srcb2[2] = { (float *) src_spad, (float *) (src_spad + src_spad_half) };
+    float * dstb2[2] = { (float *) dst_spad, (float *) (dst_spad + dst_spad_half) };
+
+    const uint32_t dst_row_size = dst->ne[0] * sizeof(float);
+    const uint32_t dst_stride   = dst->nb[1];
+
+    const uint32_t total = last - first;
+
+    // Warmup: prime pipeline with initial transfers
+    for (uint32_t i = 0; i < total && i < 2; ++i) {
+        dma_queue_push(dma_q, dma_make_data(dst->data, dstb2[i]), dst_stride, dst_row_size, dst_row_size, 0);
+
+        dma_addr_t slice_src_off;
+        uint32_t src_ys, src_yc;
+        resolve_src_slice(src0, dst, p, lut_y, first + i, &slice_src_off, &src_ys, &src_yc);
+
+        dma_push_src_chunk(dma_q, srcb2[i], src0, slice_src_off, src_ys, src_yc);
+    }
+
+    // Steady state loop
+    for (uint32_t i = 0; i < total; ++i) {
+        const uint32_t buf = i & 1u;
+        float * srcb = srcb2[buf];
+        float * dstb = dstb2[buf];
+
+        dma_queue_pop(dma_q);
+        dma_queue_pop(dma_q);
+
+        dma_addr_t slice_dst_off;
+        uint32_t dst_ys, dst_yc;
+        resolve_dst_slice(dst, p, first + i, &slice_dst_off, &dst_ys, &dst_yc);
+
+        uint32_t src_ys, src_yc;
+        get_src_y_window(src0, dst, p, lut_y, dst_ys, dst_yc, &src_ys, &src_yc);
+
+        htp_trace_event_start(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) (first + i));
+        ctx->compute_chunk(ctx, srcb, dstb, R_buf, gather_tmp, dst_ys, dst_yc, src_ys);
+        htp_trace_event_stop(tr, HTP_TRACE_EVT_HVX_COMP, (uint16_t) (first + i));
+
+        const dma_addr_t dst_addr = dst->data + slice_dst_off + (dma_addr_t) dst_ys * dst_stride;
+        dma_queue_push(dma_q, dma_make_data(dst_addr, dstb), dst_stride, dst_row_size, dst_row_size, dst_yc);
+
+        const uint32_t next = i + 2;
+        if (next < total) {
+            dma_addr_t next_src_off;
+            uint32_t next_src_ys, next_src_yc;
+            resolve_src_slice(src0, dst, p, lut_y, first + next, &next_src_off, &next_src_ys, &next_src_yc);
+
+            dma_push_src_chunk(dma_q, srcb, src0, next_src_off, next_src_ys, next_src_yc);
+        }
+    }
+
+    dma_queue_flush(dma_q);
+
+    (void) nth;
+}
+
+int op_upscale(struct htp_ops_context * octx) {
+    const struct htp_tensor * src0 = octx->src[0];
+    const struct htp_tensor * dst  = octx->dst;
+    const struct htp_upscale_kernel_params * p =
+        (const struct htp_upscale_kernel_params *) octx->kernel_params;
+
+    const uint32_t src_x = src0->ne[0];
+    const uint32_t src_y = src0->ne[1];
+    const uint32_t dst_x = dst->ne[0];
+    const uint32_t dst_y = dst->ne[1];
+
+    uint32_t chunk_start = 0;
+    uint32_t chunk_count = p->total_chunks;
+
+    if (octx->ctx->mdev.count > 1) {
+        const struct htp_tensor_mdev_range range = htp_tensor_mdev_partition(
+            chunk_count, htp_tensor_mdev_data_aligned(dst) ? p->chunks_per_group : 0,
+            octx->ctx->mdev.idx, octx->ctx->mdev.count, &octx->ctx->mdev.count_div);
+        chunk_start = range.start;
+        chunk_count = range.count;
+    }
+
+    if (chunk_count == 0) {
+        return HTP_STATUS_OK;
+    }
+
+    const uint32_t n_threads = MIN(p->n_threads, chunk_count);
+    if (!htp_ops_context_set_n_threads(octx, n_threads)) {
+        return HTP_STATUS_INVAL_PARAMS;
+    }
+
+    const uint32_t chunks_per_thread = fastdiv(chunk_count + n_threads - 1, &octx->n_threads_div);
+
+    size_t lut_x_size = 0;
+    size_t elem_y_sz  = sizeof(uint32_t);
+
+    if (p->mode == HTP_UPSCALE_MODE_BICUBIC) {
+        lut_x_size = hex_round_up(dst_x * sizeof(struct upscale_bicubic_x), 128);
+        elem_y_sz  = sizeof(struct upscale_bicubic_y);
+    } else if (p->antialias) {
+        lut_x_size = hex_round_up(dst_x * sizeof(struct upscale_antialias_x), 128);
+        elem_y_sz  = sizeof(struct upscale_antialias_y);
+    } else if (p->mode == HTP_UPSCALE_MODE_BILINEAR) {
+        const uint32_t n_chunks = (dst_x + VLEN_FP32 - 1) / VLEN_FP32;
+        lut_x_size = hex_round_up(n_chunks * sizeof(struct upscale_bilinear_x_chunk), 128);
+        elem_y_sz  = sizeof(struct upscale_bilinear_y);
+    } else {
+        lut_x_size = hex_round_up(dst_x * sizeof(uint32_t), 128);
+    }
+
+    const size_t lut_y_size = hex_round_up(dst_y * elem_y_sz, 128);
+    const size_t total_lut_size = lut_x_size + lut_y_size;
+
+    if (total_lut_size > octx->ctx->ddr_spad_size) {
+        return HTP_STATUS_NO_SUPPORT;
+    }
+
+    void * base_scratch = octx->ctx->ddr_spad_base;
+    void * lut_x        = base_scratch;
+    void * lut_y        = (uint8_t *) base_scratch + lut_x_size;
+
+    if (p->mode == HTP_UPSCALE_MODE_NEAREST) {
+        upscale_init_luts_nearest((uint32_t *) lut_x, (uint32_t *) lut_y, src0, dst, p);
+    } else if (p->mode == HTP_UPSCALE_MODE_BICUBIC) {
+        upscale_init_luts_bicubic((struct upscale_bicubic_x *) lut_x, (struct upscale_bicubic_y *) lut_y, src0, dst, p);
+    } else if (p->antialias) {
+        upscale_init_luts_antialias((struct upscale_antialias_x *) lut_x, (struct upscale_antialias_y *) lut_y, src0, dst, p);
+    } else {
+        upscale_init_luts_bilinear((struct upscale_bilinear_x_chunk *) lut_x, (struct upscale_bilinear_y *) lut_y, src0, dst, p);
+    }
+
+    upscale_chunk_fn_t compute_chunk;
+    switch (p->kernel_type) {
+        case HTP_UPSCALE_KERNEL_NEAREST:
+            compute_chunk = compute_nearest_chunk;
+            break;
+        case HTP_UPSCALE_KERNEL_NEAREST_2X:
+            compute_chunk = compute_nearest_2x_chunk;
+            break;
+        case HTP_UPSCALE_KERNEL_NEAREST_4X:
+            compute_chunk = compute_nearest_4x_chunk;
+            break;
+        case HTP_UPSCALE_KERNEL_BILINEAR:
+            compute_chunk = compute_bilinear_chunk;
+            break;
+        case HTP_UPSCALE_KERNEL_BILINEAR_2X:
+            compute_chunk = compute_bilinear_2x_chunk;
+            break;
+        case HTP_UPSCALE_KERNEL_BILINEAR_AA:
+            compute_chunk = compute_bilinear_antialias_chunk;
+            break;
+        case HTP_UPSCALE_KERNEL_BICUBIC:
+            compute_chunk = compute_bicubic_chunk;
+            break;
+        default:
+            return HTP_STATUS_INVAL_PARAMS;
+    }
+
+    struct upscale_context ctx = {
+        .octx               = octx,
+        .kparams            = p,
+        .compute_chunk      = compute_chunk,
+        .lut_x              = lut_x,
+        .lut_y              = lut_y,
+        .chunk_start        = chunk_start,
+        .chunk_count        = chunk_count,
+        .chunks_per_thread  = chunks_per_thread,
+    };
+
+    htp_upscale_vtcm_layout_build(&ctx.layout, src_x, src_y, dst_x, dst_y,
+                                  p->dst_nrows_per_chunk, p->inv_sf1, p->filter_radius1, n_threads);
+
+    work_queue_run(octx->ctx->work_queue, upscale_worker, &ctx, n_threads);
+
+    return HTP_STATUS_OK;
+}

ggml/src/ggml-hexagon/htp/upscale-ops.h

@@ -0,0 +1,204 @@
+#ifndef HTP_UPSCALE_OPS_H
+#define HTP_UPSCALE_OPS_H
+
+#include <stdint.h>
+#include <stddef.h>
+#include <stdbool.h>
+#include <math.h>
+
+#include "hex-common.h"
+#include "hex-fastdiv.h"
+#include "htp-vtcm.h"
+
+enum htp_upscale_mode {
+    HTP_UPSCALE_MODE_NEAREST  = 0,
+    HTP_UPSCALE_MODE_BILINEAR = 1,
+    HTP_UPSCALE_MODE_BICUBIC  = 2,
+};
+
+enum htp_upscale_kernel_type {
+    HTP_UPSCALE_KERNEL_NEAREST     = 0,
+    HTP_UPSCALE_KERNEL_NEAREST_2X  = 1,
+    HTP_UPSCALE_KERNEL_NEAREST_4X  = 2,
+    HTP_UPSCALE_KERNEL_BILINEAR    = 3,
+    HTP_UPSCALE_KERNEL_BILINEAR_2X = 4,
+    HTP_UPSCALE_KERNEL_BILINEAR_AA = 5,
+    HTP_UPSCALE_KERNEL_BICUBIC     = 6,
+};
+
+struct htp_upscale_kernel_params {
+    uint8_t  kernel_type;
+    uint8_t  mode;
+    uint8_t  align_corners;
+    uint8_t  antialias;
+
+    float    inv_sf0;
+    float    inv_sf1;
+    float    inv_sf2;
+    float    inv_sf3;
+    float    pixel_offset;
+
+    float    filter_radius0;
+    float    filter_radius1;
+    float    invscale0;
+    float    invscale1;
+
+    uint32_t n_threads;
+    uint32_t chunks_per_group;
+
+    uint32_t dst_nrows_per_chunk;
+    uint32_t chunks_per_slice;
+    uint32_t total_chunks;
+    uint32_t vtcm_size;
+
+    struct fastdiv_values div_chunks_per_slice;
+    struct fastdiv_values div_dst_z;
+};
+
+#if defined(__cplusplus)
+static_assert(sizeof(struct htp_upscale_kernel_params) <= 128, "htp_upscale_kernel_params is too large");
+#else
+_Static_assert(sizeof(struct htp_upscale_kernel_params) <= 128, "htp_upscale_kernel_params is too large");
+#endif
+
+struct __attribute__((aligned(128))) upscale_bilinear_x_chunk {
+    int32_t  off0[32];
+    int32_t  off1[32];
+    float    w0[32];
+    float    dx[32];
+};
+
+struct upscale_bilinear_y {
+    uint32_t y0;
+    uint32_t y1;
+    float    wy0;
+    float    wy1;
+};
+
+struct upscale_bicubic_x {
+    int32_t x[4];
+    float   w[4];
+};
+
+struct upscale_bicubic_y {
+    int32_t y[4];
+    float   w[4];
+};
+
+struct upscale_antialias_x {
+    int32_t x_min;
+    int32_t x_max;
+    float   xf;
+};
+
+struct upscale_antialias_y {
+    int32_t y_min;
+    int32_t y_max;
+    float   yf;
+};
+
+struct htp_upscale_vtcm_layout {
+    size_t   total_bytes;
+    size_t   off_src;
+    size_t   off_dst;
+    size_t   off_r;
+    size_t   off_gather;
+    size_t   src_bytes_per_thread;
+    size_t   dst_bytes_per_thread;
+    size_t   r_bytes_per_thread;
+    size_t   gather_bytes_per_thread;
+    size_t   src_spad_half_size;
+    size_t   dst_spad_half_size;
+    uint32_t max_src_chunk_rows;
+    uint32_t dst_chunk_rows;
+};
+
+static inline void htp_upscale_vtcm_layout_build(
+    struct htp_upscale_vtcm_layout * L,
+    uint32_t src_x,
+    uint32_t src_y,
+    uint32_t dst_x,
+    uint32_t dst_y,
+    uint32_t dst_chunk_rows,
+    float    inv_sf1,
+    float    filter_radius1,
+    uint32_t n_threads
+) {
+    uint32_t max_src_rows;
+    if (dst_chunk_rows >= dst_y) {
+        max_src_rows = src_y;
+    } else {
+        const float est = (float) dst_chunk_rows * inv_sf1 + 2.0f * filter_radius1;
+        max_src_rows = (uint32_t) ceilf(est) + 4;
+        if (max_src_rows > src_y) {
+            max_src_rows = src_y;
+        }
+    }
+
+    const uint32_t src_chunk_bytes = src_x * max_src_rows * sizeof(float);
+    const uint32_t dst_chunk_bytes = dst_x * dst_chunk_rows * sizeof(float);
+
+    L->max_src_chunk_rows      = max_src_rows;
+    L->dst_chunk_rows          = dst_chunk_rows;
+    L->src_spad_half_size      = hex_round_up(src_chunk_bytes + 256, 128);
+    L->dst_spad_half_size      = hex_round_up(dst_chunk_bytes, 128);
+    L->src_bytes_per_thread    = 2 * L->src_spad_half_size;
+    L->dst_bytes_per_thread    = 2 * L->dst_spad_half_size;
+    L->r_bytes_per_thread      = hex_round_up(src_x * sizeof(float), 128);
+    L->gather_bytes_per_thread = 256;
+
+    size_t off = 0;
+    VTCM_LAYOUT_ALLOC(off, off_src,    L->src_bytes_per_thread * n_threads);
+    VTCM_LAYOUT_ALLOC(off, off_dst,    L->dst_bytes_per_thread * n_threads);
+    VTCM_LAYOUT_ALLOC(off, off_r,      L->r_bytes_per_thread   * n_threads);
+    VTCM_LAYOUT_ALLOC(off, off_gather, L->gather_bytes_per_thread * n_threads);
+    L->total_bytes = off;
+}
+
+static inline bool htp_upscale_solve_layout(
+    struct htp_upscale_vtcm_layout * layout,
+    uint32_t src_x,
+    uint32_t src_y,
+    uint32_t dst_x,
+    uint32_t dst_y,
+    float    inv_sf1,
+    float    filter_radius1,
+    uint32_t n_threads,
+    size_t   vtcm_budget,
+    uint32_t * out_dst_chunk_rows
+) {
+    // 1. Try full plane
+    uint32_t chunk_rows = dst_y;
+    htp_upscale_vtcm_layout_build(layout, src_x, src_y, dst_x, dst_y, chunk_rows, inv_sf1, filter_radius1, n_threads);
+    if (layout->total_bytes <= vtcm_budget) {
+        *out_dst_chunk_rows = chunk_rows;
+        return true;
+    }
+
+    // 2. Try smaller chunk bands: 64, 32, 16, 8, 4, 2, 1
+    chunk_rows = 64;
+    while (chunk_rows > 1) {
+        if (chunk_rows < dst_y) {
+            htp_upscale_vtcm_layout_build(layout, src_x, src_y, dst_x, dst_y, chunk_rows, inv_sf1, filter_radius1, n_threads);
+            if (layout->total_bytes <= vtcm_budget) {
+                *out_dst_chunk_rows = chunk_rows;
+                return true;
+            }
+        }
+        chunk_rows /= 2;
+    }
+
+    // 3. Fallback to 1 row
+    chunk_rows = 1;
+    htp_upscale_vtcm_layout_build(layout, src_x, src_y, dst_x, dst_y, chunk_rows, inv_sf1, filter_radius1, n_threads);
+    if (layout->total_bytes <= vtcm_budget) {
+        *out_dst_chunk_rows = chunk_rows;
+        return true;
+    }
+
+    return false;
+}
+
+int op_upscale(struct htp_ops_context * octx);
+
+#endif // HTP_UPSCALE_OPS_H

ggml/src/ggml-impl.h

@@ -512,6 +512,7 @@ static inline float ggml_e8m0_to_fp32_half(uint8_t x) {
 // UE4M3: unsigned, 4 exp bits (bias=7), 3 mantissa bits
 // Returns value * 0.5 to match kvalues_mxfp4 convention (kvalues = 2 * E2M1_float)
 static inline float ggml_ue4m3_to_fp32(uint8_t x) {
+    // TODO: reject invalid unsigned E4M3 scales instead of decoding NaN as zero or ignoring the sign bit.
     if (x == 0 || x == 0x7F) {
         return 0.0f;
     }
@@ -527,6 +528,7 @@ static inline float ggml_ue4m3_to_fp32(uint8_t x) {
 }
 
 static inline uint8_t ggml_fp32_to_ue4m3(float x) {
+    // TODO: use ties-to-even rounding, allow subnormal carry, and preserve finite values with exponent field 15.
     if (!(x > 0.0f)) {
         return 0;
     }
@@ -564,6 +566,62 @@ static inline uint8_t ggml_fp32_to_ue4m3(float x) {
     return (uint8_t) ((ue4m3_exp << 3) | ue4m3_man);
 }
 
+static inline float ggml_f8_e4m3_to_fp32(uint8_t x) {
+    const uint8_t ax = x & 0x7F;
+    if (ax == 0x7F) {
+        // OCP FP8 defines both 0x7F and 0xFF as NaN; the sign need not be preserved.
+        return NAN;
+    }
+
+    const int exp = (ax >> 3) & 0xF;
+    const int man = ax & 0x7;
+    // OCP FP8 rev. 1.1, section 5.1: subnormal magnitude = M * 2^-9; normal magnitude = (1 + M / 8) * 2^(E - 7).
+    const float value = exp == 0
+        ? ldexpf((float) man, -9)
+        : ldexpf(1.0f + (float) man / 8.0f, exp - 7);
+    return x & 0x80 ? -value : value;
+}
+
+static inline int ggml_round_to_nearest_even(float x) {
+    const int value = (int) floorf(x);
+    const float fraction = x - value;
+    return fraction > 0.5f || (fraction == 0.5f && (value & 1)) ? value + 1 : value;
+}
+
+// Supports saturating conversion only; infinities and overflow clamp to +/-448.
+static inline uint8_t ggml_fp32_to_f8_e4m3(float x) {
+    const uint8_t sign = signbit(x) ? 0x80 : 0;
+    x = fabsf(x);
+
+    if (isnan(x)) {
+        // OCP FP8 leaves NaN sign handling to the implementation; use canonical 0x7F.
+        return 0x7F;
+    }
+    if (x == 0.0f) {
+        return sign;
+    }
+    if (isinf(x) || x >= 448.0f) {
+        return sign | 0x7E;
+    }
+    if (x < 0.015625f) {
+        // OCP FP8 rev. 1.1, section 5.1: subnormal magnitude = M * 2^(1 - bias - m) = M / 512.
+        return sign | (uint8_t) ggml_round_to_nearest_even(x * 512.0f);
+    }
+
+    int exp;
+    const float mantissa = frexpf(x, &exp) * 2.0f;
+    int encoded_exp = exp + 6;
+    int encoded_man = ggml_round_to_nearest_even((mantissa - 1.0f) * 8.0f);
+    if (encoded_man == 8) {
+        encoded_man = 0;
+        encoded_exp++;
+    }
+    if (encoded_exp > 15 || (encoded_exp == 15 && encoded_man > 6)) {
+        return sign | 0x7E;
+    }
+    return sign | (uint8_t) (encoded_exp << 3) | (uint8_t) encoded_man;
+}
+
 /**
  * Converts brain16 to float32.
  *

ggml/src/ggml-metal/ggml-metal-device.cpp

@@ -1103,7 +1103,6 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv(ggml_meta
             {
                 nsg = N_SG_MXFP4;
                 nr0 = N_R0_MXFP4;
-                smem = 32*sizeof(float);
             } break;
         case GGML_TYPE_Q2_K:
             {
@@ -1399,7 +1398,6 @@ ggml_metal_pipeline_with_params ggml_metal_library_get_pipeline_mul_mv_id(ggml_m
             {
                 nsg = N_SG_MXFP4;
                 nr0 = N_R0_MXFP4;
-                smem = 32*sizeof(float);
             } break;
         case GGML_TYPE_Q2_K:
             {

ggml/src/ggml-metal/ggml-metal-device.h

@@ -362,7 +362,10 @@ bool   ggml_metal_buffer_is_shared(ggml_metal_buffer_t buf);
 void   ggml_metal_buffer_memset_tensor(ggml_metal_buffer_t buf, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size);
 void   ggml_metal_buffer_set_tensor   (ggml_metal_buffer_t buf, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
 void   ggml_metal_buffer_get_tensor   (ggml_metal_buffer_t buf, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
+void   ggml_metal_buffer_set_tensor_n (ggml_metal_buffer_t buf, struct ggml_tensor * const * tensors, const void * const * data, const size_t * offsets, const size_t * sizes, size_t n_tensors);
+void   ggml_metal_buffer_get_tensor_n (ggml_metal_buffer_t buf, const struct ggml_tensor * const * tensors,       void * const * data, const size_t * offsets, const size_t * sizes, size_t n_tensors);
 bool   ggml_metal_buffer_cpy_tensor   (ggml_metal_buffer_t buf, const struct ggml_tensor * src, struct ggml_tensor * dst);
+bool   ggml_metal_buffer_cpy_tensor_n (ggml_metal_buffer_t buf, const struct ggml_tensor * const * srcs, struct ggml_tensor * const * dsts, size_t n_tensors);
 void   ggml_metal_buffer_clear        (ggml_metal_buffer_t buf, uint8_t value);
 
 // finds the Metal buffer that contains the tensor data on the GPU device

ggml/src/ggml-metal/ggml-metal-device.m

@@ -1516,6 +1516,7 @@ static bool ggml_metal_supports_mul_mat_op(
         bool src0_f16_has_mv,
         bool mm_path) {
     if (!has_simdgroup_reduction ||
+        op->src[0]->type == GGML_TYPE_F8_E4M3 ||
         op->src[0]->type == GGML_TYPE_NVFP4 ||
         op->src[0]->type == GGML_TYPE_TQ1_0) {
         return false;
@@ -1933,7 +1934,8 @@ bool ggml_metal_device_supports_op(ggml_metal_device_t dev, const struct ggml_te
             }
         case GGML_OP_GET_ROWS:
             return op->src[0]->type != GGML_TYPE_NVFP4 &&
-                   op->src[0]->type != GGML_TYPE_TQ1_0;
+                   op->src[0]->type != GGML_TYPE_TQ1_0 &&
+                   op->src[0]->type != GGML_TYPE_F8_E4M3;
         case GGML_OP_SET_ROWS:
             {
                 if (op->src[0]->type == GGML_TYPE_F16) {
@@ -2436,6 +2438,148 @@ void ggml_metal_buffer_get_tensor(ggml_metal_buffer_t buf, const struct ggml_ten
     }
 }
 
+static bool ggml_metal_buffer_data_span(const void * const * data, const size_t * sizes, size_t n_tensors, uintptr_t * min_data, uintptr_t * max_data) {
+    bool found = false;
+
+    for (size_t i = 0; i < n_tensors; i++) {
+        if (sizes[i] == 0) {
+            continue;
+        }
+
+        const uintptr_t start = (uintptr_t) data[i];
+        const uintptr_t end   = start + sizes[i];
+        if (!found) {
+            *min_data = start;
+            *max_data = end;
+            found = true;
+        } else {
+            if (start < *min_data) {
+                *min_data = start;
+            }
+            if (end > *max_data) {
+                *max_data = end;
+            }
+        }
+    }
+
+    return found;
+}
+
+void ggml_metal_buffer_set_tensor_n(ggml_metal_buffer_t buf, struct ggml_tensor * const * tensors, const void * const * data,
+        const size_t * offsets, const size_t * sizes, size_t n_tensors) {
+    if (buf->is_shared) {
+        for (size_t i = 0; i < n_tensors; i++) {
+            if (sizes[i] == 0) {
+                continue;
+            }
+            memcpy((char *) tensors[i]->data + offsets[i], data[i], sizes[i]);
+        }
+        return;
+    }
+
+    uintptr_t min_data = 0, max_data = 0;
+    if (!ggml_metal_buffer_data_span(data, sizes, n_tensors, &min_data, &max_data)) {
+        return;
+    }
+
+    @autoreleasepool {
+        void * data_ptr = (void *) min_data;
+        id<MTLBuffer> buf_src = [buf->dev->mtl_device newBufferWithBytesNoCopy:data_ptr
+                                                               length:max_data - min_data
+                                                              options:MTLResourceStorageModeShared
+                                                          deallocator:nil];
+
+        GGML_ASSERT(buf_src);
+
+        id<MTLCommandBuffer> cmd_buf = [buf->dev->mtl_queue commandBufferWithUnretainedReferences];
+
+        {
+            id<MTLBlitCommandEncoder> encoder = [cmd_buf blitCommandEncoder];
+
+            for (size_t i = 0; i < n_tensors; i++) {
+                if (sizes[i] == 0) {
+                    continue;
+                }
+
+                ggml_backend_buffer_t tensor_buf_ggml = tensors[i]->view_src ? tensors[i]->view_src->buffer : tensors[i]->buffer;
+                ggml_metal_buffer_t tensor_buf = (ggml_metal_buffer_t) tensor_buf_ggml->context;
+                struct ggml_metal_buffer_id bid_dst = ggml_metal_buffer_get_id(tensor_buf, tensors[i]);
+                bid_dst.offs += offsets[i];
+
+                [encoder copyFromBuffer:buf_src
+                           sourceOffset:(uintptr_t) data[i] - min_data
+                               toBuffer:bid_dst.metal
+                      destinationOffset:bid_dst.offs
+                                   size:sizes[i]];
+            }
+
+            [encoder endEncoding];
+        }
+
+        [cmd_buf commit];
+        [cmd_buf waitUntilCompleted];
+
+        [buf_src release];
+    }
+}
+
+void ggml_metal_buffer_get_tensor_n(ggml_metal_buffer_t buf, const struct ggml_tensor * const * tensors, void * const * data,
+        const size_t * offsets, const size_t * sizes, size_t n_tensors) {
+    if (buf->is_shared) {
+        for (size_t i = 0; i < n_tensors; i++) {
+            if (sizes[i] == 0) {
+                continue;
+            }
+            memcpy(data[i], (const char *) tensors[i]->data + offsets[i], sizes[i]);
+        }
+        return;
+    }
+
+    uintptr_t min_data = 0, max_data = 0;
+    if (!ggml_metal_buffer_data_span((const void * const *) data, sizes, n_tensors, &min_data, &max_data)) {
+        return;
+    }
+
+    @autoreleasepool {
+        id<MTLBuffer> buf_dst = [buf->dev->mtl_device newBufferWithBytesNoCopy:(void *) min_data
+                                                               length:max_data - min_data
+                                                              options:MTLResourceStorageModeShared
+                                                          deallocator:nil];
+
+        GGML_ASSERT(buf_dst);
+
+        id<MTLCommandBuffer> cmd_buf = [buf->dev->mtl_queue commandBufferWithUnretainedReferences];
+
+        {
+            id<MTLBlitCommandEncoder> encoder = [cmd_buf blitCommandEncoder];
+
+            for (size_t i = 0; i < n_tensors; i++) {
+                if (sizes[i] == 0) {
+                    continue;
+                }
+
+                ggml_backend_buffer_t tensor_buf_ggml = tensors[i]->view_src ? tensors[i]->view_src->buffer : tensors[i]->buffer;
+                ggml_metal_buffer_t tensor_buf = (ggml_metal_buffer_t) tensor_buf_ggml->context;
+                struct ggml_metal_buffer_id bid_src = ggml_metal_buffer_get_id(tensor_buf, tensors[i]);
+                bid_src.offs += offsets[i];
+
+                [encoder copyFromBuffer:bid_src.metal
+                           sourceOffset:bid_src.offs
+                               toBuffer:buf_dst
+                      destinationOffset:(uintptr_t) data[i] - min_data
+                                   size:sizes[i]];
+            }
+
+            [encoder endEncoding];
+        }
+
+        [cmd_buf commit];
+        [cmd_buf waitUntilCompleted];
+
+        [buf_dst release];
+    }
+}
+
 bool ggml_metal_buffer_cpy_tensor(ggml_metal_buffer_t buf_dst, const struct ggml_tensor * src, struct ggml_tensor * dst) {
     ggml_metal_buffer_t buf_src = (ggml_metal_buffer_t)src->buffer->context;
 
@@ -2477,6 +2621,74 @@ bool ggml_metal_buffer_cpy_tensor(ggml_metal_buffer_t buf_dst, const struct ggml
     return true;
 }
 
+bool ggml_metal_buffer_cpy_tensor_n(ggml_metal_buffer_t buf_dst, const struct ggml_tensor * const * srcs, struct ggml_tensor * const * dsts, size_t n_tensors) {
+    if (n_tensors == 0) {
+        return true;
+    }
+
+    bool all_shared = true;
+    for (size_t i = 0; i < n_tensors; ++i) {
+        ggml_backend_buffer_t src_buf = srcs[i]->view_src ? srcs[i]->view_src->buffer : srcs[i]->buffer;
+        ggml_metal_buffer_t buf_src = (ggml_metal_buffer_t) src_buf->context;
+        if (!buf_dst->is_shared || !buf_src->is_shared) {
+            all_shared = false;
+            break;
+        }
+    }
+
+    if (all_shared) {
+        for (size_t i = 0; i < n_tensors; ++i) {
+            memcpy(dsts[i]->data, srcs[i]->data, ggml_nbytes(srcs[i]));
+        }
+        return true;
+    }
+
+    // for private buffers, we need to use Metal blit commands
+    @autoreleasepool {
+        struct ggml_metal_buffer_id * bid_srcs = malloc(n_tensors * sizeof(struct ggml_metal_buffer_id));
+        struct ggml_metal_buffer_id * bid_dsts = malloc(n_tensors * sizeof(struct ggml_metal_buffer_id));
+
+        for (size_t i = 0; i < n_tensors; ++i) {
+            ggml_backend_buffer_t src_buf = srcs[i]->view_src ? srcs[i]->view_src->buffer : srcs[i]->buffer;
+            ggml_backend_buffer_t dst_buf = dsts[i]->view_src ? dsts[i]->view_src->buffer : dsts[i]->buffer;
+            ggml_metal_buffer_t buf_src = (ggml_metal_buffer_t) src_buf->context;
+            ggml_metal_buffer_t buf_dst_i = (ggml_metal_buffer_t) dst_buf->context;
+            bid_srcs[i] = ggml_metal_buffer_get_id(buf_src, srcs[i]);
+            bid_dsts[i] = ggml_metal_buffer_get_id(buf_dst_i, dsts[i]);
+
+            if (bid_srcs[i].metal == nil || bid_dsts[i].metal == nil) {
+                free(bid_srcs);
+                free(bid_dsts);
+                return false;
+            }
+        }
+
+        id<MTLCommandBuffer> cmd_buf = [buf_dst->dev->mtl_queue commandBufferWithUnretainedReferences];
+
+        {
+            id<MTLBlitCommandEncoder> encoder = [cmd_buf blitCommandEncoder];
+
+            for (size_t i = 0; i < n_tensors; ++i) {
+                [encoder copyFromBuffer:bid_srcs[i].metal
+                           sourceOffset:bid_srcs[i].offs
+                               toBuffer:bid_dsts[i].metal
+                      destinationOffset:bid_dsts[i].offs
+                                   size:ggml_nbytes(srcs[i])];
+            }
+
+            [encoder endEncoding];
+        }
+
+        [cmd_buf commit];
+        [cmd_buf waitUntilCompleted];
+
+        free(bid_srcs);
+        free(bid_dsts);
+    }
+
+    return true;
+}
+
 void ggml_metal_buffer_clear(ggml_metal_buffer_t buf, uint8_t value) {
     if (buf->is_shared) {
         memset(buf->all_data, value, buf->all_size);

ggml/src/ggml-metal/ggml-metal-ops.cpp

@@ -899,7 +899,9 @@ int ggml_metal_op_unary(ggml_metal_op_t ctx, int idx) {
     if (pipeline.cnt) {
         const int n = pipeline.c4 ? ggml_nelements(op)/4 : ggml_nelements(op);
 
-        ggml_metal_encoder_dispatch_threadgroups(enc, n, 1, 1, 1, 1, 1);
+        const int nth = MIN(256, n);
+
+        ggml_metal_encoder_dispatch_threadgroups(enc, (n + nth - 1)/nth, 1, 1, nth, 1, 1);
     } else {
         const int nth_max = MIN(256, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
         const int nth = MIN(args.ne00, nth_max);
@@ -952,7 +954,9 @@ 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::max(1, std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), ne00/2));
+    // split each row over several threadgroups: with a single row (decode) one threadgroup would run on one GPU core
+    const int32_t nth = std::max(1, std::min(ggml_metal_pipeline_max_theads_per_threadgroup(pipeline), std::min(256, ne0)));
+    const int32_t nk  = (ne0 + nth - 1)/nth;
 
     ggml_metal_encoder_set_pipeline(enc, pipeline);
     ggml_metal_encoder_set_bytes   (enc, &args, sizeof(args), 0);
@@ -964,7 +968,7 @@ int ggml_metal_op_glu(ggml_metal_op_t ctx, int idx) {
     }
     ggml_metal_encoder_set_buffer  (enc, ggml_metal_get_buffer_id(op),         3);
 
-    ggml_metal_encoder_dispatch_threadgroups(enc, nrows, 1, 1, nth, 1, 1);
+    ggml_metal_encoder_dispatch_threadgroups(enc, nk, nrows, 1, nth, 1, 1);
 
     return 1;
 }
@@ -4328,7 +4332,9 @@ int ggml_metal_op_bin(ggml_metal_op_t ctx, int idx) {
     ggml_metal_encoder_set_buffer  (enc, bid_dst,  3);
 
     if (pipeline.cnt) {
-        ggml_metal_encoder_dispatch_threadgroups(enc, args.ne0, ggml_nrows(op), 1, 1, 1, 1);
+        const int nth = MIN(256, args.ne0);
+
+        ggml_metal_encoder_dispatch_threadgroups(enc, (args.ne0 + nth - 1)/nth, ggml_nrows(op), 1, nth, 1, 1);
     } else {
         const int nth_max = MIN(256, ggml_metal_pipeline_max_theads_per_threadgroup(pipeline));
 

ggml/src/ggml-metal/ggml-metal.cpp

@@ -68,6 +68,22 @@ static void ggml_backend_metal_buffer_shared_get_tensor(ggml_backend_buffer_t bu
     ggml_metal_buffer_get_tensor(ctx, tensor, data, offset, size);
 }
 
+static void ggml_backend_metal_buffer_shared_set_tensor_n(ggml_backend_buffer_t buffer, struct ggml_tensor * const * tensors, const void * const * data, const size_t * offsets, const size_t * sizes, size_t n_tensors) {
+    ggml_metal_buffer_t ctx = (ggml_metal_buffer_t)buffer->context;
+
+    GGML_ASSERT(ggml_metal_buffer_is_shared(ctx));
+
+    ggml_metal_buffer_set_tensor_n(ctx, tensors, data, offsets, sizes, n_tensors);
+}
+
+static void ggml_backend_metal_buffer_shared_get_tensor_n(ggml_backend_buffer_t buffer, const struct ggml_tensor * const * tensors, void * const * data, const size_t * offsets, const size_t * sizes, size_t n_tensors) {
+    ggml_metal_buffer_t ctx = (ggml_metal_buffer_t)buffer->context;
+
+    GGML_ASSERT(ggml_metal_buffer_is_shared(ctx));
+
+    ggml_metal_buffer_get_tensor_n(ctx, tensors, data, offsets, sizes, n_tensors);
+}
+
 static bool ggml_backend_metal_buffer_shared_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
     ggml_metal_buffer_t ctx = (ggml_metal_buffer_t)buffer->context;
 
@@ -80,6 +96,20 @@ static bool ggml_backend_metal_buffer_shared_cpy_tensor(ggml_backend_buffer_t bu
     return ggml_metal_buffer_cpy_tensor(ctx, src, dst);
 }
 
+static bool ggml_backend_metal_buffer_shared_cpy_tensor_n(ggml_backend_buffer_t buffer, const struct ggml_tensor * const * srcs, struct ggml_tensor * const * dsts, size_t n_tensors) {
+    ggml_metal_buffer_t ctx = (ggml_metal_buffer_t)buffer->context;
+
+    GGML_ASSERT(ggml_metal_buffer_is_shared(ctx));
+
+    for (size_t i = 0; i < n_tensors; ++i) {
+        if (!ggml_backend_buffer_is_metal(srcs[i]->buffer)) {
+            return false;
+        }
+    }
+
+    return ggml_metal_buffer_cpy_tensor_n(ctx, srcs, dsts, n_tensors);
+}
+
 static void ggml_backend_metal_buffer_shared_clear(ggml_backend_buffer_t buffer, uint8_t value) {
     ggml_metal_buffer_t ctx = (ggml_metal_buffer_t)buffer->context;
 
@@ -100,6 +130,9 @@ static ggml_backend_buffer_i ggml_backend_metal_buffer_shared_i = {
     /* .cpy_tensor    = */ ggml_backend_metal_buffer_shared_cpy_tensor,
     /* .clear         = */ ggml_backend_metal_buffer_shared_clear,
     /* .reset         = */ NULL,
+    /* .set_tensor_n  = */ ggml_backend_metal_buffer_shared_set_tensor_n,
+    /* .get_tensor_n  = */ ggml_backend_metal_buffer_shared_get_tensor_n,
+    /* .cpy_tensor_n  = */ ggml_backend_metal_buffer_shared_cpy_tensor_n,
 };
 
 // private buffer
@@ -144,6 +177,22 @@ static void ggml_backend_metal_buffer_private_get_tensor(ggml_backend_buffer_t b
     ggml_metal_buffer_get_tensor(ctx, tensor, data, offset, size);
 }
 
+static void ggml_backend_metal_buffer_private_set_tensor_n(ggml_backend_buffer_t buffer, struct ggml_tensor * const * tensors, const void * const * data, const size_t * offsets, const size_t * sizes, size_t n_tensors) {
+    ggml_metal_buffer_t ctx = (ggml_metal_buffer_t)buffer->context;
+
+    GGML_ASSERT(!ggml_metal_buffer_is_shared(ctx));
+
+    ggml_metal_buffer_set_tensor_n(ctx, tensors, data, offsets, sizes, n_tensors);
+}
+
+static void ggml_backend_metal_buffer_private_get_tensor_n(ggml_backend_buffer_t buffer, const struct ggml_tensor * const * tensors, void * const * data, const size_t * offsets, const size_t * sizes, size_t n_tensors) {
+    ggml_metal_buffer_t ctx = (ggml_metal_buffer_t)buffer->context;
+
+    GGML_ASSERT(!ggml_metal_buffer_is_shared(ctx));
+
+    ggml_metal_buffer_get_tensor_n(ctx, tensors, data, offsets, sizes, n_tensors);
+}
+
 static bool ggml_backend_metal_buffer_private_cpy_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * src, ggml_tensor * dst) {
     ggml_metal_buffer_t ctx = (ggml_metal_buffer_t)buffer->context;
 
@@ -156,6 +205,20 @@ static bool ggml_backend_metal_buffer_private_cpy_tensor(ggml_backend_buffer_t b
     return ggml_metal_buffer_cpy_tensor(ctx, src, dst);
 }
 
+static bool ggml_backend_metal_buffer_private_cpy_tensor_n(ggml_backend_buffer_t buffer, const struct ggml_tensor * const * srcs, struct ggml_tensor * const * dsts, size_t n_tensors) {
+    ggml_metal_buffer_t ctx = (ggml_metal_buffer_t)buffer->context;
+
+    GGML_ASSERT(!ggml_metal_buffer_is_shared(ctx));
+
+    for (size_t i = 0; i < n_tensors; ++i) {
+        if (!ggml_backend_buffer_is_metal(srcs[i]->buffer)) {
+            return false;
+        }
+    }
+
+    return ggml_metal_buffer_cpy_tensor_n(ctx, srcs, dsts, n_tensors);
+}
+
 static void ggml_backend_metal_buffer_private_clear(ggml_backend_buffer_t buffer, uint8_t value) {
     ggml_metal_buffer_t ctx = (ggml_metal_buffer_t)buffer->context;
 
@@ -176,6 +239,9 @@ static ggml_backend_buffer_i ggml_backend_metal_buffer_private_i = {
     /* .cpy_tensor    = */ ggml_backend_metal_buffer_private_cpy_tensor,
     /* .clear         = */ ggml_backend_metal_buffer_private_clear,
     /* .reset         = */ NULL,
+    /* .set_tensor_n  = */ ggml_backend_metal_buffer_private_set_tensor_n,
+    /* .get_tensor_n  = */ ggml_backend_metal_buffer_private_get_tensor_n,
+    /* .cpy_tensor_n  = */ ggml_backend_metal_buffer_private_cpy_tensor_n,
 };
 
 static bool ggml_backend_buffer_is_metal(ggml_backend_buffer_t buffer) {

ggml/src/ggml-metal/kernels/binbcast.metal

@@ -22,8 +22,13 @@ kernel void kernel_bin_fuse_impl(
 
     if (FC_RB) {
         // row broadcast
-        const uint i0 = tgpig.y*args.ne00 + tgpig.x;
-        const uint i1 = FC_CB ? tgpig.x%args.ne10 : tgpig.x;
+        const uint i = tgpig.x*ntg.x + tpitg.x;
+        if (i >= (uint) args.ne0) {
+            return;
+        }
+
+        const uint i0 = tgpig.y*args.ne00 + i;
+        const uint i1 = FC_CB ? i%args.ne10 : i;
 
         device const T0 * src0_row = (device const T0 *) (src0);
         device       T  * dst_row  = (device       T  *) (dst);

ggml/src/ggml-metal/kernels/mul_mv.metal

@@ -3030,6 +3030,22 @@ kernel void kernel_mul_mv_iq4_xs_f32(
     kernel_mul_mv_iq4_xs_f32_impl<N_R0_IQ4_XS, constant ggml_metal_kargs_mul_mv &>(args, src0, src1, dst, shmem, tgpig, tiisg, sgitg);
 }
 
+// decode the E2M1 value in the low nibble of n, scaled by 2^-14, without a lookup table: the 3 magnitude bits
+// land in the fp16 exponent/mantissa (e=0 gives the 0.5 subnormal). same decode as MLX's fp4_e2m1
+static inline half mxfp4_dq1(uint32_t n) {
+    const half c = as_type<half>(ushort((n & 7) << 9));
+    return (n & 8) ? -c : c;
+}
+
+// low / high nibbles of the 4 bytes of w, in byte order, scaled by 2^-14
+static inline float4 mxfp4_lo4(uint32_t w) {
+    return float4(mxfp4_dq1(w), mxfp4_dq1(w >> 8), mxfp4_dq1(w >> 16), mxfp4_dq1(w >> 24));
+}
+
+static inline float4 mxfp4_hi4(uint32_t w) {
+    return float4(mxfp4_dq1(w >> 4), mxfp4_dq1(w >> 12), mxfp4_dq1(w >> 20), mxfp4_dq1(w >> 28));
+}
+
 template<int NR0, typename args_t>
 void kernel_mul_mv_mxfp4_f32_impl(
         args_t args,
@@ -3042,8 +3058,6 @@ void kernel_mul_mv_mxfp4_f32_impl(
         ushort sgitg) {
     const short NSG = FC_mul_mv_nsg;
 
-    threadgroup float * shmem_f32 = (threadgroup float *) shmem;
-
     const int r0 = tgpig.x;
     const int r1 = tgpig.y;
     const int im = tgpig.z;
@@ -3065,9 +3079,6 @@ void kernel_mul_mv_mxfp4_f32_impl(
     const short ix = tiisg/2;  // 0...15
     const short it = tiisg%2;  // 0 or 1
 
-    shmem_f32[tiisg] = kvalues_mxfp4_f[tiisg%16];
-    threadgroup_barrier(mem_flags::mem_threadgroup);
-
     float4 yl[4];
     float sumf[NR0]={0.f};
 
@@ -3087,14 +3098,25 @@ void kernel_mul_mv_mxfp4_f32_impl(
             device const block_mxfp4 & xb = x[row*ns01 + ib];
             device const uint8_t     * q2 = (device const uint8_t *)(xb.qs + 8*it);
 
-            float4 acc1 = yl[0]*float4(shmem_f32[q2[0] &  0x0F], shmem_f32[q2[1] &  0x0F], shmem_f32[q2[2] &  0x0F], shmem_f32[q2[3] &  0x0F]);
-            float4 acc2 = yl[1]*float4(shmem_f32[q2[0] >> 4   ], shmem_f32[q2[1] >> 4   ], shmem_f32[q2[2] >> 4   ], shmem_f32[q2[3] >> 4   ]);
-            float4 acc3 = yl[2]*float4(shmem_f32[q2[4] &  0x0F], shmem_f32[q2[5] &  0x0F], shmem_f32[q2[6] &  0x0F], shmem_f32[q2[7] &  0x0F]);
-            float4 acc4 = yl[3]*float4(shmem_f32[q2[4] >> 4   ], shmem_f32[q2[5] >> 4   ], shmem_f32[q2[6] >> 4   ], shmem_f32[q2[7] >> 4   ]);
+            // blocks are 17 bytes, so the 8 quant bytes are only byte aligned: read the aligned words that cover them
+            // and funnel-shift. the third word is read only when needed so no word past the block end is touched
+            device const uint32_t * pw = (device const uint32_t *)((ulong) q2 & ~3ul);
+            const uint sh = ((uint)(ulong) q2 & 3u)*8;
+            const uint32_t a = pw[0];
+            const uint32_t b = pw[1];
+            const uint32_t c = pw[sh ? 2 : 1];
+            const uint32_t w0 = (a >> sh) | ((b << 1) << (31 - sh));
+            const uint32_t w1 = (b >> sh) | ((c << 1) << (31 - sh));
+
+            float4 acc1 = yl[0]*mxfp4_lo4(w0);
+            float4 acc2 = yl[1]*mxfp4_hi4(w0);
+            float4 acc3 = yl[2]*mxfp4_lo4(w1);
+            float4 acc4 = yl[3]*mxfp4_hi4(w1);
 
             acc1 = (acc1 + acc3) + (acc2 + acc4);
 
-            sumf[row] += e8m0_to_fp32(xb.e) * ((acc1[0] + acc1[1]) + (acc1[2] + acc1[3]));
+            // 16384 undoes the 2^-14 of the decode
+            sumf[row] += (e8m0_to_fp32(xb.e)*16384.0f) * ((acc1[0] + acc1[1]) + (acc1[2] + acc1[3]));
         }
 
         yb += 16 * QK_MXFP4;

ggml/src/ggml-metal/kernels/unary.metal

@@ -11,12 +11,12 @@ inline constexpr T ggml_metal_unary_cast(U x) {
 #if defined(GGML_METAL_HAS_BF16)
 // MSL has no C-style cast from float vectors to bfloat vectors
 template <>
-inline constexpr bfloat4 ggml_metal_unary_cast<bfloat4, float4>(float4 x) {
+inline bfloat4 ggml_metal_unary_cast<bfloat4, float4>(float4 x) {
     return bfloat4(bfloat(x[0]), bfloat(x[1]), bfloat(x[2]), bfloat(x[3]));
 }
 
 template <>
-inline constexpr bfloat4 ggml_metal_unary_cast<bfloat4, float>(float x) {
+inline bfloat4 ggml_metal_unary_cast<bfloat4, float>(float x) {
     return bfloat4(bfloat(x), bfloat(x), bfloat(x), bfloat(x));
 }
 #endif
@@ -38,7 +38,7 @@ kernel void kernel_unary_impl(
     int i0;
 
     if (FC_CNT) {
-        i0 = tgpig.x;
+        i0 = tgpig.x*ntg.x + tpitg.x;
 
         src0_ptr = (device const T0 *) (src0);
         dst_ptr  = (device       T  *) (dst);
@@ -61,6 +61,10 @@ kernel void kernel_unary_impl(
             if (i0 >= args.ne0) {
                 return;
             }
+        } else {
+            if (i0 >= args.ne00*args.ne01*args.ne02*args.ne03) {
+                return;
+            }
         }
 
         const TC x = (TC) src0_ptr[i0];
@@ -231,14 +235,15 @@ kernel void kernel_reglu(
         device const char * src0,
         device const char * src1,
         device       char * dst,
-        uint tgpig[[threadgroup_position_in_grid]],
-        uint tpitg[[thread_position_in_threadgroup]],
-        uint   ntg[[threads_per_threadgroup]]) {
-    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
-    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
-    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+        uint2 tgpig[[threadgroup_position_in_grid]],
+        uint2 tpitg[[thread_position_in_threadgroup]],
+        uint2   ntg[[threads_per_threadgroup]],
+        uint2  ntgs[[threadgroups_per_grid]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig.y*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig.y*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig.y*args.nb1);
 
-    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+    for (int i0 = tgpig.x*ntg.x + tpitg.x; i0 < args.ne0; i0 += ntg.x*ntgs.x) {
         const float x0 = src0_row[i0];
         const float x1 = src1_row[i0];
 
@@ -257,14 +262,15 @@ kernel void kernel_geglu(
         device const char * src0,
         device const char * src1,
         device       char * dst,
-        uint tgpig[[threadgroup_position_in_grid]],
-        uint tpitg[[thread_position_in_threadgroup]],
-        uint   ntg[[threads_per_threadgroup]]) {
-    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
-    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
-    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+        uint2 tgpig[[threadgroup_position_in_grid]],
+        uint2 tpitg[[thread_position_in_threadgroup]],
+        uint2   ntg[[threads_per_threadgroup]],
+        uint2  ntgs[[threadgroups_per_grid]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig.y*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig.y*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig.y*args.nb1);
 
-    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+    for (int i0 = tgpig.x*ntg.x + tpitg.x; i0 < args.ne0; i0 += ntg.x*ntgs.x) {
         const float x0 = src0_row[i0];
         const float x1 = src1_row[i0];
 
@@ -285,14 +291,15 @@ kernel void kernel_swiglu(
         device const char * src0,
         device const char * src1,
         device       char * dst,
-        uint tgpig[[threadgroup_position_in_grid]],
-        uint tpitg[[thread_position_in_threadgroup]],
-        uint   ntg[[threads_per_threadgroup]]) {
-    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
-    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
-    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+        uint2 tgpig[[threadgroup_position_in_grid]],
+        uint2 tpitg[[thread_position_in_threadgroup]],
+        uint2   ntg[[threads_per_threadgroup]],
+        uint2  ntgs[[threadgroups_per_grid]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig.y*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig.y*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig.y*args.nb1);
 
-    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+    for (int i0 = tgpig.x*ntg.x + tpitg.x; i0 < args.ne0; i0 += ntg.x*ntgs.x) {
         const float x0 = src0_row[i0];
         const float x1 = src1_row[i0];
 
@@ -313,14 +320,15 @@ kernel void kernel_swiglu_oai(
         device const char * src0,
         device const char * src1,
         device       char * dst,
-        uint tgpig[[threadgroup_position_in_grid]],
-        uint tpitg[[thread_position_in_threadgroup]],
-        uint   ntg[[threads_per_threadgroup]]) {
-    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
-    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
-    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+        uint2 tgpig[[threadgroup_position_in_grid]],
+        uint2 tpitg[[thread_position_in_threadgroup]],
+        uint2   ntg[[threads_per_threadgroup]],
+        uint2  ntgs[[threadgroups_per_grid]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig.y*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig.y*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig.y*args.nb1);
 
-    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+    for (int i0 = tgpig.x*ntg.x + tpitg.x; i0 < args.ne0; i0 += ntg.x*ntgs.x) {
         float x0 = src0_row[i0];
         float x1 = src1_row[i0];
 
@@ -345,14 +353,15 @@ kernel void kernel_swiglu_clamp(
         device const char * src0,
         device const char * src1,
         device       char * dst,
-        uint tgpig[[threadgroup_position_in_grid]],
-        uint tpitg[[thread_position_in_threadgroup]],
-        uint   ntg[[threads_per_threadgroup]]) {
-    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
-    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
-    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+        uint2 tgpig[[threadgroup_position_in_grid]],
+        uint2 tpitg[[thread_position_in_threadgroup]],
+        uint2   ntg[[threads_per_threadgroup]],
+        uint2  ntgs[[threadgroups_per_grid]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig.y*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig.y*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig.y*args.nb1);
 
-    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+    for (int i0 = tgpig.x*ntg.x + tpitg.x; i0 < args.ne0; i0 += ntg.x*ntgs.x) {
         const float gate = min((float) src0_row[i0], args.limit);
         const float up = clamp((float) src1_row[i0], -args.limit, args.limit);
 
@@ -371,14 +380,15 @@ kernel void kernel_geglu_erf(
         device const char * src0,
         device const char * src1,
         device       char * dst,
-        uint tgpig[[threadgroup_position_in_grid]],
-        uint tpitg[[thread_position_in_threadgroup]],
-        uint   ntg[[threads_per_threadgroup]]) {
-    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
-    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
-    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+        uint2 tgpig[[threadgroup_position_in_grid]],
+        uint2 tpitg[[thread_position_in_threadgroup]],
+        uint2   ntg[[threads_per_threadgroup]],
+        uint2  ntgs[[threadgroups_per_grid]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig.y*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig.y*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig.y*args.nb1);
 
-    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+    for (int i0 = tgpig.x*ntg.x + tpitg.x; i0 < args.ne0; i0 += ntg.x*ntgs.x) {
         const float x0 = src0_row[i0];
         const float x1 = src1_row[i0];
 
@@ -399,14 +409,15 @@ kernel void kernel_geglu_quick(
         device const char * src0,
         device const char * src1,
         device       char * dst,
-        uint tgpig[[threadgroup_position_in_grid]],
-        uint tpitg[[thread_position_in_threadgroup]],
-        uint   ntg[[threads_per_threadgroup]]) {
-    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig*args.nb01) + args.i00;
-    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig*args.nb11) + args.i10;
-    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig*args.nb1);
+        uint2 tgpig[[threadgroup_position_in_grid]],
+        uint2 tpitg[[thread_position_in_threadgroup]],
+        uint2   ntg[[threads_per_threadgroup]],
+        uint2  ntgs[[threadgroups_per_grid]]) {
+    device const T * src0_row = (device const T *) ((device const char *) src0 + tgpig.y*args.nb01) + args.i00;
+    device const T * src1_row = (device const T *) ((device const char *) src1 + tgpig.y*args.nb11) + args.i10;
+    device       T * dst_row  = (device       T *) ((device       char *) dst  + tgpig.y*args.nb1);
 
-    for (int i0 = tpitg; i0 < args.ne0; i0 += ntg) {
+    for (int i0 = tgpig.x*ntg.x + tpitg.x; i0 < args.ne0; i0 += ntg.x*ntgs.x) {
         const float x0 = src0_row[i0];
         const float x1 = src1_row[i0];
 

ggml/src/ggml-opencl/CMakeLists.txt

@@ -105,10 +105,37 @@ set(GGML_OPENCL_KERNELS
     gemv_noshuffle_q2_k_f32
     gemm_noshuffle_q2_k_f32
     gemm_noshuffle_q3_k_f32
+    gemm_noshuffle_iq2_s_f32
+    gemm_noshuffle_iq2_xs_f32
+    gemm_noshuffle_iq2_xxs_f32
+    gemm_noshuffle_iq2_xxs_q8_1_dp4a
+    gemv_noshuffle_iq2_xxs_f32
+    gemm_noshuffle_iq2_xs_q8_1_dp4a
+    gemv_noshuffle_iq2_xs_f32
+    gemm_noshuffle_iq2_s_q8_1_dp4a
+    gemv_noshuffle_iq2_s_f32
+    gemm_noshuffle_q2_0_f32
+    gemm_noshuffle_q2_0_q8_1_dp4a
+    gemv_noshuffle_q2_0_f32
     gemm_noshuffle_q3_k_q8_1_dp4a
     gemv_noshuffle_q3_k_f32
     mul_mv_q2_k_f32
     mul_mv_q3_k_f32
+    iq2_s_grid
+    mul_mv_iq2_s_f32
+    iq2_xs_grid
+    mul_mv_iq2_xs_f32
+    iq2_xxs_grid
+    mul_mv_iq2_xxs_f32
+    mul_mv_iq2_xxs_f32_flat
+    mul_mm_iq2_xxs_f32_l4_lm
+    mul_mv_iq2_xs_f32_flat
+    mul_mm_iq2_xs_f32_l4_lm
+    mul_mv_iq2_s_f32_flat
+    mul_mm_iq2_s_f32_l4_lm
+    mul_mv_q2_0_f32
+    mul_mv_q2_0_f32_flat
+    mul_mm_q2_0_f32_l4_lm
     mul_mv_q3_k_f32_flat
     mul_mm_q3_k_f32_l4_lm
     mul_mv_q2_k_f32_flat

ggml/src/ggml-opencl/kernels/cvt.cl

@@ -2926,3 +2926,165 @@ kernel void kernel_restore_block_q3_k_noshuffle(
         b->hmask[ee] = (uchar)v;
     }
 }
+
+struct block_iq2_s {
+    half d;
+    uchar qs[QK_K/4];
+    uchar qh[QK_K/32];
+    uchar scales[QK_K/32];
+};
+
+kernel void kernel_convert_block_iq2_s(
+    global struct block_iq2_s * src0,
+    global uchar * dst_qs,      // QK_K/8 per block
+    global uchar * dst_sg,      // QK_K/8
+    global uchar * dst_qh,      // QK_K/32
+    global uchar * dst_sc,      // QK_K/32
+    global half  * dst_d,       // 1
+    ulong          n_blk
+) {
+    const ulong i = get_global_id(0);
+    if (i >= n_blk) {
+        return;
+    }
+    global struct block_iq2_s * b = src0 + i;
+
+    dst_d[i] = b->d;
+    for (int j = 0; j < QK_K/8;  ++j) { dst_qs[(QK_K/8)  * i + j] = b->qs[j];            }
+    for (int j = 0; j < QK_K/8;  ++j) { dst_sg[(QK_K/8)  * i + j] = b->qs[QK_K/8 + j];   }
+    for (int j = 0; j < QK_K/32; ++j) { dst_qh[(QK_K/32) * i + j] = b->qh[j];            }
+    for (int j = 0; j < QK_K/32; ++j) { dst_sc[(QK_K/32) * i + j] = b->scales[j];        }
+}
+
+kernel void kernel_restore_block_iq2_s(
+    global uchar * src_qs,
+    global uchar * src_sg,
+    global uchar * src_qh,
+    global uchar * src_sc,
+    global half  * src_d,
+    global struct block_iq2_s * dst,
+    ulong          n_blk
+) {
+    const ulong i = get_global_id(0);
+    if (i >= n_blk) {
+        return;
+    }
+    global struct block_iq2_s * b = dst + i;
+
+    b->d = src_d[i];
+    for (int j = 0; j < QK_K/8;  ++j) { b->qs[j]          = src_qs[(QK_K/8)  * i + j]; }
+    for (int j = 0; j < QK_K/8;  ++j) { b->qs[QK_K/8 + j] = src_sg[(QK_K/8)  * i + j]; }
+    for (int j = 0; j < QK_K/32; ++j) { b->qh[j]          = src_qh[(QK_K/32) * i + j]; }
+    for (int j = 0; j < QK_K/32; ++j) { b->scales[j]      = src_sc[(QK_K/32) * i + j]; }
+}
+
+struct block_iq2_xs {
+    half d;
+    ushort qs[QK_K/8];
+    uchar scales[QK_K/32];
+};
+
+kernel void kernel_convert_block_iq2_xs(
+    global struct block_iq2_xs * src0,
+    global ushort * dst_qs,
+    global uchar * dst_sc,
+    global half * dst_d,
+    ulong n_blocks
+) {
+    const ulong i = get_global_id(0);
+    if (i >= n_blocks) { return; }
+    dst_d[i] = src0[i].d;
+    for (int j = 0; j < 32; ++j) { dst_qs[32*i + j] = src0[i].qs[j]; }
+    for (int j = 0; j < 8; ++j) { dst_sc[8*i + j] = src0[i].scales[j]; }
+}
+
+kernel void kernel_restore_block_iq2_xs(
+    global ushort * src_qs,
+    global uchar * src_sc,
+    global half * src_d,
+    global struct block_iq2_xs * dst,
+    ulong n_blocks
+) {
+    const ulong i = get_global_id(0);
+    if (i >= n_blocks) { return; }
+    dst[i].d = src_d[i];
+    for (int j = 0; j < 32; ++j) { dst[i].qs[j] = src_qs[32*i + j]; }
+    for (int j = 0; j < 8; ++j) { dst[i].scales[j] = src_sc[8*i + j]; }
+}
+
+struct block_iq2_xxs {
+    half d;
+    ushort qs[QK_K/8];
+};
+
+kernel void kernel_convert_block_iq2_xxs(
+    global struct block_iq2_xxs * src0,
+    global uchar * dst_qs,
+    global uint * dst_sas,
+    global half * dst_d,
+    ulong n_blocks
+) {
+    const ulong i = get_global_id(0);
+    if (i >= n_blocks) { return; }
+    dst_d[i] = src0[i].d;
+    for (int j = 0; j < 8; ++j) {
+        global ushort * q = src0[i].qs + 4*j;
+        uint indices = (uint)q[0] | ((uint)q[1] << 16);
+        vstore4(as_uchar4(indices), 8*i + j, dst_qs);
+        dst_sas[8*i + j] = (uint)q[2] | ((uint)q[3] << 16);
+    }
+}
+
+kernel void kernel_restore_block_iq2_xxs(
+    global uchar * src_qs,
+    global uint * src_sas,
+    global half * src_d,
+    global struct block_iq2_xxs * dst,
+    ulong n_blocks
+) {
+    const ulong i = get_global_id(0);
+    if (i >= n_blocks) { return; }
+    dst[i].d = src_d[i];
+    for (int j = 0; j < 8; ++j) {
+        global ushort * q = dst[i].qs + 4*j;
+        uint indices = as_uint(vload4(8*i + j, src_qs));
+        uint sas = src_sas[8*i + j];
+        q[0] = (ushort)indices;
+        q[1] = (ushort)(indices >> 16);
+        q[2] = (ushort)sas;
+        q[3] = (ushort)(sas >> 16);
+    }
+}
+
+struct block_q2_0 {
+    half d;
+    uchar qs[16];
+};
+
+kernel void kernel_convert_block_q2_0(
+    global struct block_q2_0 * src0,
+    global uchar * dst_qs,
+    global half * dst_d,
+    ulong n_blocks
+) {
+    const ulong i = get_global_id(0);
+    if (i >= n_blocks) { return; }
+    dst_d[i] = src0[i].d;
+    for (int j = 0; j < 16; ++j) {
+        dst_qs[16*i + j] = src0[i].qs[j];
+    }
+}
+
+kernel void kernel_restore_block_q2_0(
+    global uchar * src_qs,
+    global half * src_d,
+    global struct block_q2_0 * dst,
+    ulong n_blocks
+) {
+    const ulong i = get_global_id(0);
+    if (i >= n_blocks) { return; }
+    dst[i].d = src_d[i];
+    for (int j = 0; j < 16; ++j) {
+        dst[i].qs[j] = src_qs[16*i + j];
+    }
+}

ggml/src/ggml-opencl/kernels/gemm_noshuffle_iq2_s_f32.cl

@@ -0,0 +1,93 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+__attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+kernel void kernel_gemm_noshuffle_iq2_s_f32(
+    global const uchar * src0_qs,
+    global const uchar * src0_sg,
+    global const uchar * src0_qh,
+    global const uchar * src0_sc,
+    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
+) {
+    dst = (global float *)((global char *)dst + offsetd);
+    const int gy = get_global_id(0);
+    const int row = get_global_id(1) * 4;
+    const int n_4 = n / 4;
+    if (row >= m) {
+        return;
+    }
+
+    half8 c0 = 0, c1 = 0, c2 = 0, c3 = 0;
+    for (int i = 0; i < k; i += 16) {
+        const uchar4 sc = vload4(0, src0_sc + row + (i / 32) * m);
+        const half4 d = vload4(0, src0_d + row + (i / 256) * m);
+        const float4 scale_f = convert_float4(d) * 0.25f * (0.5f + convert_float4((sc >> (uchar4)(4*((i / 16) & 1))) & (uchar4)(15)));
+        const half4 scale = convert_half4(scale_f);
+
+        for (int l = 0; l < 16; l += 4) {
+            const int ki = i + l;
+            const uint4 qs = convert_uint4(vload4(0, src0_qs + row + (ki / 8) * m));
+            const uint4 qh = convert_uint4(vload4(0, src0_qh + row + (ki / 32) * m));
+            const uchar4 sg = vload4(0, src0_sg + row + (ki / 8) * m);
+            const uint4 gi = qs | (((qh >> (uint4)(2*((ki / 8) & 3))) & (uint4)(3)) << (uint4)(8));
+            const uint half_grid = (ki & 7) / 4;
+            const uint4 grid = (uint4)(iq2_s_grid[2*gi.s0 + half_grid], iq2_s_grid[2*gi.s1 + half_grid],
+                                      iq2_s_grid[2*gi.s2 + half_grid], iq2_s_grid[2*gi.s3 + half_grid]);
+            #pragma unroll
+            for (int j = 0; j < 4; ++j) {
+                half8 b;
+                b.s0123 = read_imageh(src1, gy * 2     + (ki + j) * n_4);
+                b.s4567 = read_imageh(src1, gy * 2 + 1 + (ki + j) * n_4);
+                const int4 magnitude = convert_int4((grid >> (uint4)(8*j)) & (uint4)(255));
+                const int4 signs = convert_int4((sg >> (uchar4)((ki & 7) + j)) & (uchar4)(1));
+                const half4 w = convert_half4(magnitude * ((int4)(1) - 2*signs)) * scale;
+                c0 += b * w.s0;
+                c1 += b * w.s1;
+                c2 += b * w.s2;
+                c3 += b * w.s3;
+            }
+        }
+    }
+
+    int idx = gy * 8 * m + row;
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s0, c1.s0, c2.s0, c3.s0), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s1, c1.s1, c2.s1, c3.s1), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s2, c1.s2, c2.s2, c3.s2), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s3, c1.s3, c2.s3, c3.s3), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s4, c1.s4, c2.s4, c3.s4), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s5, c1.s5, c2.s5, c3.s5), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s6, c1.s6, c2.s6, c3.s6), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s7, c1.s7, c2.s7, c3.s7), 0, dst + idx);
+    }
+}

ggml/src/ggml-opencl/kernels/gemm_noshuffle_iq2_s_q8_1_dp4a.cl

@@ -0,0 +1,149 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#ifdef cl_khr_integer_dot_product
+#pragma OPENCL EXTENSION cl_khr_integer_dot_product : enable
+#endif
+
+#define QK_K 256
+
+#ifndef TILESIZE_N
+#define TILESIZE_N 32
+#endif
+
+inline uint iq2_s_pack(uint gv, uint sg, uint base) {
+    int v0 = (int)((gv >>  0) & 0xFF); if (sg & (1u << (base + 0))) { v0 = -v0; }
+    int v1 = (int)((gv >>  8) & 0xFF); if (sg & (1u << (base + 1))) { v1 = -v1; }
+    int v2 = (int)((gv >> 16) & 0xFF); if (sg & (1u << (base + 2))) { v2 = -v2; }
+    int v3 = (int)((gv >> 24) & 0xFF); if (sg & (1u << (base + 3))) { v3 = -v3; }
+    return ((uint)v0 & 0xFFu) | (((uint)v1 & 0xFFu) <<  8)
+         | (((uint)v2 & 0xFFu) << 16) | (((uint)v3 & 0xFFu) << 24);
+}
+
+// Pass staged operands by value; local dp4a operands miscompile on X2.
+inline int dot4_q8a_v(uint4 qw, uint4 a) {
+    int r = 0;
+    r = dot_acc_sat_4x8packed_ss_int(qw.s0, a.x, r);
+    r = dot_acc_sat_4x8packed_ss_int(qw.s1, a.y, r);
+    r = dot_acc_sat_4x8packed_ss_int(qw.s2, a.z, r);
+    r = dot_acc_sat_4x8packed_ss_int(qw.s3, a.w, r);
+    return r;
+}
+
+__attribute__((qcom_wave_pair_mode(1)))
+kernel void kernel_gemm_noshuffle_iq2_s_q8_1_dp4a(
+        __global const uchar  * src0_qs,
+        __global const uchar  * src0_sg,
+        __global const uchar  * src0_qh,
+        __global const uchar  * src0_sc,
+        __global const half   * src0_d,
+        __global const uint   * src1_qa,
+        __global const half   * src1_da,
+        __global       float  * dst,
+        ulong  offsetd,
+        int    m,
+        int    n_no_padding,
+        int    k
+) {
+    dst = (global float *)((global char *)dst + offsetd);
+
+    const uint lid = get_local_id(0);
+    const uint block_id_m = get_global_id(1);
+    const uint block_id_n = get_global_id(2);
+
+    const uint row      = block_id_m * 64 + lid;
+    const uint col_base = block_id_n * TILESIZE_N;
+    const bool row_valid = row < (uint)m;
+    const uint rrow     = row_valid ? row : 0;
+
+    const uint k_u = (uint)k >> 2;
+    const uint k_b = (uint)k >> 5;
+
+    __local uint4 sh_qa4[TILESIZE_N][2];
+    __local half sh_d[TILESIZE_N];
+
+#define IQ2_S_GRID(i) iq2_s_grid[(i)]
+
+#define NGROUPS (TILESIZE_N / 4)
+    float4 acc[NGROUPS];
+    #pragma unroll
+    for (int g = 0; g < NGROUPS; ++g) acc[g] = (float4)(0.0f);
+
+    for (uint step = 0; step < (uint)k; step += 32) {
+        const uint sub = step >> 5;
+        const uint ib  = sub >> 3;
+
+        const uint  qhv = (uint)src0_qh[rrow + sub * (uint)m];
+        const uint  scv = (uint)src0_sc[rrow + sub * (uint)m];
+        const float dv  = (float)src0_d[rrow + ib * (uint)m];
+        const float dl0 = dv * (0.5f + (float)( scv       & 0xFu)) * 0.25f;
+        const float dl1 = dv * (0.5f + (float)( scv >> 4)        ) * 0.25f;
+
+        const uint gb = rrow + (step >> 3) * (uint)m;
+
+        uint4 qlo, qhi;
+        {
+            const uint gi0 = (uint)src0_qs[gb + 0u * (uint)m] | ((( qhv       & 3u) << 8));
+            const uint gi1 = (uint)src0_qs[gb + 1u * (uint)m] | ((((qhv >> 2) & 3u) << 8));
+            const uint gi2 = (uint)src0_qs[gb + 2u * (uint)m] | ((((qhv >> 4) & 3u) << 8));
+            const uint gi3 = (uint)src0_qs[gb + 3u * (uint)m] | ((((qhv >> 6) & 3u) << 8));
+            const uint s0  = (uint)src0_sg[gb + 0u * (uint)m];
+            const uint s1  = (uint)src0_sg[gb + 1u * (uint)m];
+            const uint s2  = (uint)src0_sg[gb + 2u * (uint)m];
+            const uint s3  = (uint)src0_sg[gb + 3u * (uint)m];
+            qlo.s0 = iq2_s_pack(IQ2_S_GRID(2u*gi0 + 0u), s0, 0u);
+            qlo.s1 = iq2_s_pack(IQ2_S_GRID(2u*gi0 + 1u), s0, 4u);
+            qlo.s2 = iq2_s_pack(IQ2_S_GRID(2u*gi1 + 0u), s1, 0u);
+            qlo.s3 = iq2_s_pack(IQ2_S_GRID(2u*gi1 + 1u), s1, 4u);
+            qhi.s0 = iq2_s_pack(IQ2_S_GRID(2u*gi2 + 0u), s2, 0u);
+            qhi.s1 = iq2_s_pack(IQ2_S_GRID(2u*gi2 + 1u), s2, 4u);
+            qhi.s2 = iq2_s_pack(IQ2_S_GRID(2u*gi3 + 0u), s3, 0u);
+            qhi.s3 = iq2_s_pack(IQ2_S_GRID(2u*gi3 + 1u), s3, 4u);
+        }
+
+        for (uint idx = lid; idx < TILESIZE_N * 2; idx += 64) {
+            const uint t = idx >> 1;
+            const uint v = idx & 1;
+            const uint c = col_base + t;
+            sh_qa4[t][v] = (c < (uint)n_no_padding)
+                         ? vload4(0, src1_qa + c * k_u + (step >> 2) + (v << 2))
+                         : (uint4)(0u);
+        }
+        if (lid < TILESIZE_N) {
+            const uint c = col_base + lid;
+            sh_d[lid] = (c < (uint)n_no_padding) ? src1_da[c * k_b + sub] : (half)0;
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+#define LD4(arr, b) ((float4)((float)arr[(b)+0], (float)arr[(b)+1], (float)arr[(b)+2], (float)arr[(b)+3]))
+#define IQ2_S_COL(b) (dl0 * (float)dot4_q8a_v(qlo, (uint4)(sh_qa4[b][0]))  \
+                   + dl1 * (float)dot4_q8a_v(qhi, (uint4)(sh_qa4[b][1])))
+        #pragma unroll
+        for (int g = 0; g < NGROUPS; ++g) {
+            const int b = g * 4;
+            float4 rf;
+            rf.s0 = IQ2_S_COL(b+0);  rf.s1 = IQ2_S_COL(b+1);
+            rf.s2 = IQ2_S_COL(b+2);  rf.s3 = IQ2_S_COL(b+3);
+            acc[g] += LD4(sh_d, b) * rf;
+        }
+#undef IQ2_S_COL
+#undef LD4
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    if (!row_valid) {
+        return;
+    }
+
+    #pragma unroll
+    for (int g = 0; g < NGROUPS; ++g) {
+        const uint b = (uint)(g * 4);
+        const float4 a = acc[g];
+        const uint c0 = col_base + b;
+        if (c0 + 0 < (uint)n_no_padding) dst[(c0 + 0) * (uint)m + row] = a.s0;
+        if (c0 + 1 < (uint)n_no_padding) dst[(c0 + 1) * (uint)m + row] = a.s1;
+        if (c0 + 2 < (uint)n_no_padding) dst[(c0 + 2) * (uint)m + row] = a.s2;
+        if (c0 + 3 < (uint)n_no_padding) dst[(c0 + 3) * (uint)m + row] = a.s3;
+    }
+#undef NGROUPS
+#undef IQ2_S_GRID
+}

ggml/src/ggml-opencl/kernels/gemm_noshuffle_iq2_xs_f32.cl

@@ -0,0 +1,91 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+__attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+kernel void kernel_gemm_noshuffle_iq2_xs_f32(
+    global const ushort * src0_qs,
+    global const uchar * src0_sc,
+    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
+) {
+    dst = (global float *)((global char *)dst + offsetd);
+    const int gy = get_global_id(0);
+    const int row = get_global_id(1) * 4;
+    const int n_4 = n / 4;
+    if (row >= m) {
+        return;
+    }
+
+    half8 c0 = 0, c1 = 0, c2 = 0, c3 = 0;
+    for (int i = 0; i < k; i += 16) {
+        const uchar4 sc = vload4(0, src0_sc + row + (i / 32) * m);
+        const half4 d = vload4(0, src0_d + row + (i / 256) * m);
+        const float4 scale_f = convert_float4(d) * 0.25f * (0.5f + convert_float4((sc >> (uchar4)(4*((i / 16) & 1))) & (uchar4)(15)));
+        const half4 scale = convert_half4(scale_f);
+
+        for (int l = 0; l < 16; l += 4) {
+            const int ki = i + l;
+            const uint4 qs = convert_uint4(vload4(0, src0_qs + row + (ki / 8) * m));
+            const uint4 code = qs >> (uint4)(9);
+            const uint4 sg = code | ((popcount(code) & (uint4)(1)) << (uint4)(7));
+            const uint4 gi = qs & (uint4)(511);
+            const uint half_grid = (ki & 7) / 4;
+            const uint4 grid = (uint4)(iq2_xs_grid[2*gi.s0 + half_grid], iq2_xs_grid[2*gi.s1 + half_grid],
+                                      iq2_xs_grid[2*gi.s2 + half_grid], iq2_xs_grid[2*gi.s3 + half_grid]);
+            #pragma unroll
+            for (int j = 0; j < 4; ++j) {
+                half8 b;
+                b.s0123 = read_imageh(src1, gy * 2     + (ki + j) * n_4);
+                b.s4567 = read_imageh(src1, gy * 2 + 1 + (ki + j) * n_4);
+                const int4 magnitude = convert_int4((grid >> (uint4)(8*j)) & (uint4)(255));
+                const int4 signs = convert_int4((sg >> (uint4)((ki & 7) + j)) & (uint4)(1));
+                const half4 w = convert_half4(magnitude * ((int4)(1) - 2*signs)) * scale;
+                c0 += b * w.s0;
+                c1 += b * w.s1;
+                c2 += b * w.s2;
+                c3 += b * w.s3;
+            }
+        }
+    }
+
+    int idx = gy * 8 * m + row;
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s0, c1.s0, c2.s0, c3.s0), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s1, c1.s1, c2.s1, c3.s1), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s2, c1.s2, c2.s2, c3.s2), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s3, c1.s3, c2.s3, c3.s3), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s4, c1.s4, c2.s4, c3.s4), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s5, c1.s5, c2.s5, c3.s5), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s6, c1.s6, c2.s6, c3.s6), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s7, c1.s7, c2.s7, c3.s7), 0, dst + idx);
+    }
+}

ggml/src/ggml-opencl/kernels/gemm_noshuffle_iq2_xs_q8_1_dp4a.cl

@@ -0,0 +1,150 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#ifdef cl_khr_integer_dot_product
+#pragma OPENCL EXTENSION cl_khr_integer_dot_product : enable
+#endif
+
+#define QK_K 256
+
+#ifndef TILESIZE_N
+#define TILESIZE_N 32
+#endif
+
+inline uint iq2_xs_pack(uint gv, uint sg, uint base) {
+    int v0 = (int)((gv >>  0) & 0xFF); if (sg & (1u << (base + 0))) { v0 = -v0; }
+    int v1 = (int)((gv >>  8) & 0xFF); if (sg & (1u << (base + 1))) { v1 = -v1; }
+    int v2 = (int)((gv >> 16) & 0xFF); if (sg & (1u << (base + 2))) { v2 = -v2; }
+    int v3 = (int)((gv >> 24) & 0xFF); if (sg & (1u << (base + 3))) { v3 = -v3; }
+    return ((uint)v0 & 0xFFu) | (((uint)v1 & 0xFFu) <<  8)
+         | (((uint)v2 & 0xFFu) << 16) | (((uint)v3 & 0xFFu) << 24);
+}
+
+// Pass staged operands by value; local dp4a operands miscompile on X2.
+inline int dot4_q8a_v(uint4 qw, uint4 a) {
+    int r = 0;
+    r = dot_acc_sat_4x8packed_ss_int(qw.s0, a.x, r);
+    r = dot_acc_sat_4x8packed_ss_int(qw.s1, a.y, r);
+    r = dot_acc_sat_4x8packed_ss_int(qw.s2, a.z, r);
+    r = dot_acc_sat_4x8packed_ss_int(qw.s3, a.w, r);
+    return r;
+}
+
+__attribute__((qcom_wave_pair_mode(1)))
+kernel void kernel_gemm_noshuffle_iq2_xs_q8_1_dp4a(
+        __global const ushort * src0_qs,
+        __global const uchar  * src0_sc,
+        __global const half   * src0_d,
+        __global const uint   * src1_qa,
+        __global const half   * src1_da,
+        __global       float  * dst,
+        ulong  offsetd,
+        int    m,
+        int    n_no_padding,
+        int    k
+) {
+    dst = (global float *)((global char *)dst + offsetd);
+
+    const uint lid = get_local_id(0);
+    const uint block_id_m = get_global_id(1);
+    const uint block_id_n = get_global_id(2);
+
+    const uint row      = block_id_m * 64 + lid;
+    const uint col_base = block_id_n * TILESIZE_N;
+    const bool row_valid = row < (uint)m;
+    const uint rrow     = row_valid ? row : 0;
+
+    const uint k_u = (uint)k >> 2;
+    const uint k_b = (uint)k >> 5;
+
+    __local uint4 sh_qa4[TILESIZE_N][2];
+    __local half sh_d[TILESIZE_N];
+
+#define IQ2_XS_GRID(i) iq2_xs_grid[(i)]
+
+#define NGROUPS (TILESIZE_N / 4)
+    float4 acc[NGROUPS];
+    #pragma unroll
+    for (int g = 0; g < NGROUPS; ++g) acc[g] = (float4)(0.0f);
+
+    for (uint step = 0; step < (uint)k; step += 32) {
+        const uint sub = step >> 5;
+        const uint ib  = sub >> 3;
+
+        const uint  scv = (uint)src0_sc[rrow + sub * (uint)m];
+        const float dv  = (float)src0_d[rrow + ib * (uint)m];
+        const float dl0 = dv * (0.5f + (float)( scv       & 0xFu)) * 0.25f;
+        const float dl1 = dv * (0.5f + (float)( scv >> 4)        ) * 0.25f;
+
+        const uint gb = rrow + (step >> 3) * (uint)m;
+
+        uint4 qlo, qhi;
+        {
+            const uint q0 = (uint)src0_qs[gb + 0u * (uint)m];
+            const uint q1 = (uint)src0_qs[gb + 1u * (uint)m];
+            const uint q2 = (uint)src0_qs[gb + 2u * (uint)m];
+            const uint q3 = (uint)src0_qs[gb + 3u * (uint)m];
+            const uint gi0 = (q0 & 511u) << 1;
+            const uint gi1 = (q1 & 511u) << 1;
+            const uint gi2 = (q2 & 511u) << 1;
+            const uint gi3 = (q3 & 511u) << 1;
+            const uint s0  = iq2_xs_signs(q0 >> 9);
+            const uint s1  = iq2_xs_signs(q1 >> 9);
+            const uint s2  = iq2_xs_signs(q2 >> 9);
+            const uint s3  = iq2_xs_signs(q3 >> 9);
+            qlo.s0 = iq2_xs_pack(IQ2_XS_GRID(gi0 + 0u), s0, 0u);
+            qlo.s1 = iq2_xs_pack(IQ2_XS_GRID(gi0 + 1u), s0, 4u);
+            qlo.s2 = iq2_xs_pack(IQ2_XS_GRID(gi1 + 0u), s1, 0u);
+            qlo.s3 = iq2_xs_pack(IQ2_XS_GRID(gi1 + 1u), s1, 4u);
+            qhi.s0 = iq2_xs_pack(IQ2_XS_GRID(gi2 + 0u), s2, 0u);
+            qhi.s1 = iq2_xs_pack(IQ2_XS_GRID(gi2 + 1u), s2, 4u);
+            qhi.s2 = iq2_xs_pack(IQ2_XS_GRID(gi3 + 0u), s3, 0u);
+            qhi.s3 = iq2_xs_pack(IQ2_XS_GRID(gi3 + 1u), s3, 4u);
+        }
+
+        for (uint idx = lid; idx < TILESIZE_N * 2; idx += 64) {
+            const uint t = idx >> 1;
+            const uint v = idx & 1;
+            const uint c = col_base + t;
+            sh_qa4[t][v] = (c < (uint)n_no_padding)
+                         ? vload4(0, src1_qa + c * k_u + (step >> 2) + (v << 2))
+                         : (uint4)(0u);
+        }
+        if (lid < TILESIZE_N) {
+            const uint c = col_base + lid;
+            sh_d[lid] = (c < (uint)n_no_padding) ? src1_da[c * k_b + sub] : (half)0;
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+#define LD4(arr, b) ((float4)((float)arr[(b)+0], (float)arr[(b)+1], (float)arr[(b)+2], (float)arr[(b)+3]))
+#define IQ2_XS_COL(b) (dl0 * (float)dot4_q8a_v(qlo, (uint4)(sh_qa4[b][0]))  \
+                    + dl1 * (float)dot4_q8a_v(qhi, (uint4)(sh_qa4[b][1])))
+        #pragma unroll
+        for (int g = 0; g < NGROUPS; ++g) {
+            const int b = g * 4;
+            float4 rf;
+            rf.s0 = IQ2_XS_COL(b+0);  rf.s1 = IQ2_XS_COL(b+1);
+            rf.s2 = IQ2_XS_COL(b+2);  rf.s3 = IQ2_XS_COL(b+3);
+            acc[g] += LD4(sh_d, b) * rf;
+        }
+#undef IQ2_XS_COL
+#undef LD4
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    if (!row_valid) {
+        return;
+    }
+
+    #pragma unroll
+    for (int g = 0; g < NGROUPS; ++g) {
+        const uint b = (uint)(g * 4);
+        const float4 a = acc[g];
+        const uint c0 = col_base + b;
+        if (c0 + 0 < (uint)n_no_padding) dst[(c0 + 0) * (uint)m + row] = a.s0;
+        if (c0 + 1 < (uint)n_no_padding) dst[(c0 + 1) * (uint)m + row] = a.s1;
+        if (c0 + 2 < (uint)n_no_padding) dst[(c0 + 2) * (uint)m + row] = a.s2;
+        if (c0 + 3 < (uint)n_no_padding) dst[(c0 + 3) * (uint)m + row] = a.s3;
+    }
+#undef NGROUPS
+#undef IQ2_XS_GRID
+}

ggml/src/ggml-opencl/kernels/gemm_noshuffle_iq2_xxs_f32.cl

@@ -0,0 +1,90 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+__attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+kernel void kernel_gemm_noshuffle_iq2_xxs_f32(
+    global const uchar * src0_qs,
+    global const uint * src0_sas,
+    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
+) {
+    dst = (global float *)((global char *)dst + offsetd);
+    const int gy = get_global_id(0);
+    const int row = get_global_id(1) * 4;
+    const int n_4 = n / 4;
+    if (row >= m) {
+        return;
+    }
+
+    half8 c0 = 0, c1 = 0, c2 = 0, c3 = 0;
+    for (int i = 0; i < k; i += 16) {
+        const uint4 sas = vload4(0, src0_sas + row + (i / 32) * m);
+        const half4 d = vload4(0, src0_d + row + (i / 256) * m);
+        const float4 scale_f = convert_float4(d) * 0.25f * (0.5f + convert_float4(sas >> (uint4)(28)));
+        const half4 scale = convert_half4(scale_f);
+
+        for (int l = 0; l < 16; l += 4) {
+            const int ki = i + l;
+            const uint4 gi = convert_uint4(vload4(0, src0_qs + row + (ki / 8) * m));
+            const uint4 code = (sas >> (uint4)(7*((ki / 8) & 3))) & (uint4)(127);
+            const uint4 sg = code | ((popcount(code) & (uint4)(1)) << (uint4)(7));
+            const uint half_grid = (ki & 7) / 4;
+            const uint4 grid = (uint4)(iq2_xxs_grid[2*gi.s0 + half_grid], iq2_xxs_grid[2*gi.s1 + half_grid],
+                                      iq2_xxs_grid[2*gi.s2 + half_grid], iq2_xxs_grid[2*gi.s3 + half_grid]);
+            #pragma unroll
+            for (int j = 0; j < 4; ++j) {
+                half8 b;
+                b.s0123 = read_imageh(src1, gy * 2     + (ki + j) * n_4);
+                b.s4567 = read_imageh(src1, gy * 2 + 1 + (ki + j) * n_4);
+                const int4 magnitude = convert_int4((grid >> (uint4)(8*j)) & (uint4)(255));
+                const int4 signs = convert_int4((sg >> (uint4)((ki & 7) + j)) & (uint4)(1));
+                const half4 w = convert_half4(magnitude * ((int4)(1) - 2*signs)) * scale;
+                c0 += b * w.s0;
+                c1 += b * w.s1;
+                c2 += b * w.s2;
+                c3 += b * w.s3;
+            }
+        }
+    }
+
+    int idx = gy * 8 * m + row;
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s0, c1.s0, c2.s0, c3.s0), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s1, c1.s1, c2.s1, c3.s1), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s2, c1.s2, c2.s2, c3.s2), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s3, c1.s3, c2.s3, c3.s3), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s4, c1.s4, c2.s4, c3.s4), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s5, c1.s5, c2.s5, c3.s5), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s6, c1.s6, c2.s6, c3.s6), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s7, c1.s7, c2.s7, c3.s7), 0, dst + idx);
+    }
+}

ggml/src/ggml-opencl/kernels/gemm_noshuffle_iq2_xxs_q8_1_dp4a.cl

@@ -0,0 +1,145 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#ifdef cl_khr_integer_dot_product
+#pragma OPENCL EXTENSION cl_khr_integer_dot_product : enable
+#endif
+
+#define QK_K 256
+
+#ifndef TILESIZE_N
+#define TILESIZE_N 32
+#endif
+
+inline uint iq2_xxs_pack(uint gv, uint sg, uint base) {
+    int v0 = (int)((gv >>  0) & 0xFF); if (sg & (1u << (base + 0))) { v0 = -v0; }
+    int v1 = (int)((gv >>  8) & 0xFF); if (sg & (1u << (base + 1))) { v1 = -v1; }
+    int v2 = (int)((gv >> 16) & 0xFF); if (sg & (1u << (base + 2))) { v2 = -v2; }
+    int v3 = (int)((gv >> 24) & 0xFF); if (sg & (1u << (base + 3))) { v3 = -v3; }
+    return ((uint)v0 & 0xFFu) | (((uint)v1 & 0xFFu) <<  8)
+         | (((uint)v2 & 0xFFu) << 16) | (((uint)v3 & 0xFFu) << 24);
+}
+
+// Pass staged operands by value; local dp4a operands miscompile on X2.
+inline int dot4_q8a_v(uint4 qw, uint4 a) {
+    int r = 0;
+    r = dot_acc_sat_4x8packed_ss_int(qw.s0, a.x, r);
+    r = dot_acc_sat_4x8packed_ss_int(qw.s1, a.y, r);
+    r = dot_acc_sat_4x8packed_ss_int(qw.s2, a.z, r);
+    r = dot_acc_sat_4x8packed_ss_int(qw.s3, a.w, r);
+    return r;
+}
+
+__attribute__((qcom_wave_pair_mode(1)))
+kernel void kernel_gemm_noshuffle_iq2_xxs_q8_1_dp4a(
+        __global const uchar  * src0_qs,
+        __global const uint   * src0_sas,
+        __global const half   * src0_d,
+        __global const uint   * src1_qa,
+        __global const half   * src1_da,
+        __global       float  * dst,
+        ulong  offsetd,
+        int    m,
+        int    n_no_padding,
+        int    k
+) {
+    dst = (global float *)((global char *)dst + offsetd);
+
+    const uint lid = get_local_id(0);
+    const uint block_id_m = get_global_id(1);
+    const uint block_id_n = get_global_id(2);
+
+    const uint row      = block_id_m * 64 + lid;
+    const uint col_base = block_id_n * TILESIZE_N;
+    const bool row_valid = row < (uint)m;
+    const uint rrow     = row_valid ? row : 0;
+
+    const uint k_u = (uint)k >> 2;
+    const uint k_b = (uint)k >> 5;
+
+    __local uint4 sh_qa4[TILESIZE_N][2];
+    __local half sh_d[TILESIZE_N];
+
+#define IQ2_XXS_GRID(i) iq2_xxs_grid[(i)]
+
+#define NGROUPS (TILESIZE_N / 4)
+    float4 acc[NGROUPS];
+    #pragma unroll
+    for (int g = 0; g < NGROUPS; ++g) acc[g] = (float4)(0.0f);
+
+    for (uint step = 0; step < (uint)k; step += 32) {
+        const uint sub = step >> 5;
+        const uint ib  = sub >> 3;
+
+        const uint  aux = src0_sas[rrow + sub * (uint)m];
+        const float dv  = (float)src0_d[rrow + ib * (uint)m];
+        const float dl  = dv * (0.5f + (float)(aux >> 28)) * 0.25f;
+
+        const uint gb = rrow + (step >> 3) * (uint)m;
+
+        uint4 qlo, qhi;
+        {
+            const uint gi0 = (uint)src0_qs[gb + 0u * (uint)m] << 1;
+            const uint gi1 = (uint)src0_qs[gb + 1u * (uint)m] << 1;
+            const uint gi2 = (uint)src0_qs[gb + 2u * (uint)m] << 1;
+            const uint gi3 = (uint)src0_qs[gb + 3u * (uint)m] << 1;
+            const uint s0  = iq2_xxs_signs((aux >>  0) & 127u);
+            const uint s1  = iq2_xxs_signs((aux >>  7) & 127u);
+            const uint s2  = iq2_xxs_signs((aux >> 14) & 127u);
+            const uint s3  = iq2_xxs_signs((aux >> 21) & 127u);
+            qlo.s0 = iq2_xxs_pack(IQ2_XXS_GRID(gi0 + 0u), s0, 0u);
+            qlo.s1 = iq2_xxs_pack(IQ2_XXS_GRID(gi0 + 1u), s0, 4u);
+            qlo.s2 = iq2_xxs_pack(IQ2_XXS_GRID(gi1 + 0u), s1, 0u);
+            qlo.s3 = iq2_xxs_pack(IQ2_XXS_GRID(gi1 + 1u), s1, 4u);
+            qhi.s0 = iq2_xxs_pack(IQ2_XXS_GRID(gi2 + 0u), s2, 0u);
+            qhi.s1 = iq2_xxs_pack(IQ2_XXS_GRID(gi2 + 1u), s2, 4u);
+            qhi.s2 = iq2_xxs_pack(IQ2_XXS_GRID(gi3 + 0u), s3, 0u);
+            qhi.s3 = iq2_xxs_pack(IQ2_XXS_GRID(gi3 + 1u), s3, 4u);
+        }
+
+        for (uint idx = lid; idx < TILESIZE_N * 2; idx += 64) {
+            const uint t = idx >> 1;
+            const uint v = idx & 1;
+            const uint c = col_base + t;
+            sh_qa4[t][v] = (c < (uint)n_no_padding)
+                         ? vload4(0, src1_qa + c * k_u + (step >> 2) + (v << 2))
+                         : (uint4)(0u);
+        }
+        if (lid < TILESIZE_N) {
+            const uint c = col_base + lid;
+            sh_d[lid] = (c < (uint)n_no_padding) ? src1_da[c * k_b + sub] : (half)0;
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+#define LD4(arr, b) ((float4)((float)arr[(b)+0], (float)arr[(b)+1], (float)arr[(b)+2], (float)arr[(b)+3]))
+#define IQ2_XXS_COL(b) (dl * (float)(dot4_q8a_v(qlo, (uint4)(sh_qa4[b][0]))   \
+                                  + dot4_q8a_v(qhi, (uint4)(sh_qa4[b][1]))))
+        #pragma unroll
+        for (int g = 0; g < NGROUPS; ++g) {
+            const int b = g * 4;
+            float4 rf;
+            rf.s0 = IQ2_XXS_COL(b+0);  rf.s1 = IQ2_XXS_COL(b+1);
+            rf.s2 = IQ2_XXS_COL(b+2);  rf.s3 = IQ2_XXS_COL(b+3);
+            acc[g] += LD4(sh_d, b) * rf;
+        }
+#undef IQ2_XXS_COL
+#undef LD4
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    if (!row_valid) {
+        return;
+    }
+
+    #pragma unroll
+    for (int g = 0; g < NGROUPS; ++g) {
+        const uint b = (uint)(g * 4);
+        const float4 a = acc[g];
+        const uint c0 = col_base + b;
+        if (c0 + 0 < (uint)n_no_padding) dst[(c0 + 0) * (uint)m + row] = a.s0;
+        if (c0 + 1 < (uint)n_no_padding) dst[(c0 + 1) * (uint)m + row] = a.s1;
+        if (c0 + 2 < (uint)n_no_padding) dst[(c0 + 2) * (uint)m + row] = a.s2;
+        if (c0 + 3 < (uint)n_no_padding) dst[(c0 + 3) * (uint)m + row] = a.s3;
+    }
+#undef NGROUPS
+#undef IQ2_XXS_GRID
+}

ggml/src/ggml-opencl/kernels/gemm_noshuffle_q2_0_f32.cl

@@ -0,0 +1,80 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_qcom_reqd_sub_group_size
+#pragma OPENCL EXTENSION cl_qcom_reqd_sub_group_size : enable
+__attribute__((qcom_reqd_sub_group_size("full")))
+#endif
+kernel void kernel_gemm_noshuffle_q2_0_f32(
+    global const uchar * src0_qs,
+    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
+) {
+    dst = (global float *)((global char *)dst + offsetd);
+    const int gy = get_global_id(0);
+    const int row = get_global_id(1) * 4;
+    const int n_4 = n / 4;
+    if (row >= m) {
+        return;
+    }
+
+    float8 c0 = 0, c1 = 0, c2 = 0, c3 = 0;
+    for (int i = 0; i < k; i += 16) {
+        const float4 scale = convert_float4(vload4(0, src0_d + row + (i / 64) * m));
+
+        for (int l = 0; l < 16; l += 4) {
+            const int ki = i + l;
+            const uchar4 q = vload4(0, src0_qs + row + (ki / 4) * m);
+            #pragma unroll
+            for (int j = 0; j < 4; ++j) {
+                float8 b;
+                b.s0123 = convert_float4(read_imageh(src1, gy * 2     + (ki + j) * n_4));
+                b.s4567 = convert_float4(read_imageh(src1, gy * 2 + 1 + (ki + j) * n_4));
+                const int4 v = convert_int4((q >> (uchar4)(2*j)) & (uchar4)(3)) - (int4)(1);
+                const float4 w = convert_float4(v) * scale;
+                c0 += b * w.s0;
+                c1 += b * w.s1;
+                c2 += b * w.s2;
+                c3 += b * w.s3;
+            }
+        }
+    }
+
+    int idx = gy * 8 * m + row;
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s0, c1.s0, c2.s0, c3.s0), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s1, c1.s1, c2.s1, c3.s1), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s2, c1.s2, c2.s2, c3.s2), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s3, c1.s3, c2.s3, c3.s3), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s4, c1.s4, c2.s4, c3.s4), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s5, c1.s5, c2.s5, c3.s5), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s6, c1.s6, c2.s6, c3.s6), 0, dst + idx);
+        idx += m;
+    }
+    if (idx + 3 < m * n_no_padding) {
+        vstore4((float4)(c0.s7, c1.s7, c2.s7, c3.s7), 0, dst + idx);
+    }
+}

ggml/src/ggml-opencl/kernels/gemm_noshuffle_q2_0_q8_1_dp4a.cl

@@ -0,0 +1,127 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+#pragma OPENCL EXTENSION cl_khr_subgroups : enable
+#ifdef cl_khr_integer_dot_product
+#pragma OPENCL EXTENSION cl_khr_integer_dot_product : enable
+#endif
+
+#ifndef TILESIZE_N
+#define TILESIZE_N 32
+#endif
+
+inline uint q2_0_pack(uchar q) {
+    return (((uint)(q & 3) - 1u) & 255u)
+         | (((((uint)q >> 2) & 3u) - 1u) & 255u) << 8
+         | (((((uint)q >> 4) & 3u) - 1u) & 255u) << 16
+         | (((((uint)q >> 6) & 3u) - 1u) & 255u) << 24;
+}
+
+// Pass staged operands by value; local dp4a operands miscompile on X2.
+inline int dot8_q8a_v(uint8 qw, uint4 a0, uint4 a1) {
+    int r = 0;
+    r = dot_acc_sat_4x8packed_ss_int(qw.s0, a0.x, r);
+    r = dot_acc_sat_4x8packed_ss_int(qw.s1, a0.y, r);
+    r = dot_acc_sat_4x8packed_ss_int(qw.s2, a0.z, r);
+    r = dot_acc_sat_4x8packed_ss_int(qw.s3, a0.w, r);
+    r = dot_acc_sat_4x8packed_ss_int(qw.s4, a1.x, r);
+    r = dot_acc_sat_4x8packed_ss_int(qw.s5, a1.y, r);
+    r = dot_acc_sat_4x8packed_ss_int(qw.s6, a1.z, r);
+    r = dot_acc_sat_4x8packed_ss_int(qw.s7, a1.w, r);
+    return r;
+}
+
+__attribute__((qcom_wave_pair_mode(1)))
+kernel void kernel_gemm_noshuffle_q2_0_q8_1_dp4a(
+        __global const uchar  * src0_qs,
+        __global const half   * src0_d,
+        __global const uint   * src1_qa,
+        __global const half   * src1_da,
+        __global       float  * dst,
+        ulong  offsetd,
+        int    m,
+        int    n_no_padding,
+        int    k
+) {
+    dst = (global float *)((global char *)dst + offsetd);
+
+    const uint lid = get_local_id(0);
+    const uint block_id_m = get_global_id(1);
+    const uint block_id_n = get_global_id(2);
+
+    const uint row      = block_id_m * 64 + lid;
+    const uint col_base = block_id_n * TILESIZE_N;
+    const bool row_valid = row < (uint)m;
+    const uint rrow     = row_valid ? row : 0;
+
+    const uint k_u = (uint)k >> 2;
+    const uint k_b = (uint)k >> 5;
+
+    __local uint4 sh_qa4[TILESIZE_N][2];
+    __local half sh_d[TILESIZE_N];
+
+#define NGROUPS (TILESIZE_N / 4)
+    float4 acc[NGROUPS];
+    #pragma unroll
+    for (int g = 0; g < NGROUPS; ++g) acc[g] = (float4)(0.0f);
+
+    for (uint step = 0; step < (uint)k; step += 32) {
+        const uint sub = step >> 5;
+
+        const uint ib = sub >> 1;
+        const float d_w = (float)src0_d[rrow + ib*(uint)m];
+        const uint qsb = rrow + sub*8u*(uint)m;
+
+        uint8 qw;
+        qw.s0 = q2_0_pack(src0_qs[qsb + 0u*(uint)m]);
+        qw.s1 = q2_0_pack(src0_qs[qsb + 1u*(uint)m]);
+        qw.s2 = q2_0_pack(src0_qs[qsb + 2u*(uint)m]);
+        qw.s3 = q2_0_pack(src0_qs[qsb + 3u*(uint)m]);
+        qw.s4 = q2_0_pack(src0_qs[qsb + 4u*(uint)m]);
+        qw.s5 = q2_0_pack(src0_qs[qsb + 5u*(uint)m]);
+        qw.s6 = q2_0_pack(src0_qs[qsb + 6u*(uint)m]);
+        qw.s7 = q2_0_pack(src0_qs[qsb + 7u*(uint)m]);
+
+        for (uint idx = lid; idx < TILESIZE_N * 2; idx += 64) {
+            const uint t = idx >> 1;
+            const uint v = idx & 1;
+            const uint c = col_base + t;
+            sh_qa4[t][v] = (c < (uint)n_no_padding)
+                         ? vload4(0, src1_qa + c * k_u + (step >> 2) + (v << 2))
+                         : (uint4)(0u);
+        }
+        if (lid < TILESIZE_N) {
+            const uint c = col_base + lid;
+            sh_d[lid] = (c < (uint)n_no_padding) ? src1_da[c * k_b + sub] : (half)0;
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+#define LD4(arr, b) ((float4)((float)arr[(b)+0], (float)arr[(b)+1], (float)arr[(b)+2], (float)arr[(b)+3]))
+        #pragma unroll
+        for (int g = 0; g < NGROUPS; ++g) {
+            const int b = g * 4;
+            float4 rf;
+#define TDOT4(T) dot8_q8a_v(qw, (uint4)(sh_qa4[T][0]), (uint4)(sh_qa4[T][1]))
+            rf.s0 = (float)TDOT4(b+0);  rf.s1 = (float)TDOT4(b+1);
+            rf.s2 = (float)TDOT4(b+2);  rf.s3 = (float)TDOT4(b+3);
+            acc[g] += d_w * LD4(sh_d, b) * rf;
+        }
+#undef TDOT4
+#undef LD4
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    if (!row_valid) {
+        return;
+    }
+
+    #pragma unroll
+    for (int g = 0; g < NGROUPS; ++g) {
+        const uint b = (uint)(g * 4);
+        const float4 a = acc[g];
+        const uint c0 = col_base + b;
+        if (c0 + 0 < (uint)n_no_padding) dst[(c0 + 0) * (uint)m + row] = a.s0;
+        if (c0 + 1 < (uint)n_no_padding) dst[(c0 + 1) * (uint)m + row] = a.s1;
+        if (c0 + 2 < (uint)n_no_padding) dst[(c0 + 2) * (uint)m + row] = a.s2;
+        if (c0 + 3 < (uint)n_no_padding) dst[(c0 + 3) * (uint)m + row] = a.s3;
+    }
+#undef NGROUPS
+}

ggml/src/ggml-opencl/kernels/gemv_noshuffle_iq2_s_f32.cl

@@ -0,0 +1,122 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define QK_K 256
+
+#ifndef IQ2_S_MV_NSG
+#define IQ2_S_MV_NSG 8
+#endif
+
+inline float4 iq2_s_vals(uint gv, uint sg, uint base) {
+
+    const uint  s   = sg >> base;
+    const uint4 sgn = (uint4)(s << 31, s << 30, s << 29, s << 28) & 0x80000000u;
+    return as_float4(as_uint4(convert_float4(as_uchar4(gv))) ^ sgn);
+}
+
+kernel void kernel_iq2_s_grid_export(global uint * out) {
+    out[get_global_id(0)] = iq2_s_grid[get_global_id(0)];
+}
+
+kernel void kernel_gemv_noshuffle_iq2_s_f32(
+        read_only image1d_buffer_t grid_img,
+        global const uchar * src0_qs,
+        global const uchar * src0_sg,
+        global const uchar * src0_qh,
+        global const uchar * src0_sc,
+        global const half  * src0_d,
+        global const float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne10,
+        int ne0
+) {
+    src1 = (global const float *)((global const char *)src1 + offset1);
+    dst  = (global float       *)((global char       *)dst  + offsetd);
+
+    const uint m   = (uint)ne01;
+    const uint K   = (uint)ne00;
+    const uint nsb = K / QK_K;
+
+    const uint lid = get_local_id(0);
+    const uint sgi = get_local_id(1);
+    const uint col = get_group_id(1);
+
+    global const float * y = src1 + (ulong)col * (uint)ne10;
+
+#define IQ2_S_GRID(i) read_imageui(grid_img, (int)(i)).x
+
+    const uint mh  = m >> 1;
+    const uint j   = get_group_id(0) * 64u + lid;
+    const uint row = j << 1;
+
+    float sumf = 0.f, sumf1 = 0.f;
+
+    if (j < mh) {
+        global const ushort * qsu = (global const ushort *)src0_qs;
+        global const ushort * sgu = (global const ushort *)src0_sg;
+        global const ushort * qhu = (global const ushort *)src0_qh;
+        global const ushort * scu = (global const ushort *)src0_sc;
+
+        for (uint ib = sgi; ib < nsb; ib += IQ2_S_MV_NSG) {
+            const half2 dh = vload2(j + ib * mh, src0_d);
+
+            float acc0 = 0.f, acc1 = 0.f;
+            for (uint sb = 0; sb < 8u; ++sb) {
+                const uint sub = ib * 8u + sb;
+                const uint qhv = (uint)qhu[j + sub * mh];
+                const uint scv = (uint)scu[j + sub * mh];
+                const uint gb  = j + (sub * 4u) * mh;
+
+                for (uint h = 0; h < 2u; ++h) {
+                    const uint n0 = (h == 0u) ? ( scv        & 0xFu) : (( scv        >> 4) & 0xFu);
+                    const uint n1 = (h == 0u) ? ((scv >> 8)  & 0xFu) : (((scv >> 8)  >> 4) & 0xFu);
+
+                    float a0 = 0.f, a1 = 0.f;
+                    for (uint t = 0; t < 2u; ++t) {
+                        const uint l   = 2u*h + t;
+                        const uint qsv = (uint)qsu[gb + l * mh];
+                        const uint sgv = (uint)sgu[gb + l * mh];
+                        const uint gi0 = ( qsv       & 0xFFu) | ((((qhv      ) >> (2u*l)) & 3u) << 8);
+                        const uint gi1 = ((qsv >> 8) & 0xFFu) | ((((qhv >> 8) >> (2u*l)) & 3u) << 8);
+                        const uint s0  =  sgv       & 0xFFu;
+                        const uint s1  = (sgv >> 8) & 0xFFu;
+
+                        const uint grp = (ib * 64u + sb * 8u) + l * 2u;
+                        const float4 y0 = vload4(grp + 0u, y);
+                        const float4 y1 = vload4(grp + 1u, y);
+                        a0 += dot(y0, iq2_s_vals(IQ2_S_GRID(2u*gi0 + 0u), s0, 0u));
+                        a0 += dot(y1, iq2_s_vals(IQ2_S_GRID(2u*gi0 + 1u), s0, 4u));
+                        a1 += dot(y0, iq2_s_vals(IQ2_S_GRID(2u*gi1 + 0u), s1, 0u));
+                        a1 += dot(y1, iq2_s_vals(IQ2_S_GRID(2u*gi1 + 1u), s1, 4u));
+                    }
+                    acc0 += (0.5f + (float)n0) * a0;
+                    acc1 += (0.5f + (float)n1) * a1;
+                }
+            }
+            sumf  += (float)dh.s0 * 0.25f * acc0;
+            sumf1 += (float)dh.s1 * 0.25f * acc1;
+        }
+    }
+
+#if IQ2_S_MV_NSG > 1
+    __local float2 part[IQ2_S_MV_NSG][64];
+    part[sgi][lid] = (float2)(sumf, sumf1);
+    barrier(CLK_LOCAL_MEM_FENCE);
+    if (sgi != 0) {
+        return;
+    }
+    for (uint s = 1; s < IQ2_S_MV_NSG; ++s) {
+        const float2 p = part[s][lid];
+        sumf  += p.s0;
+        sumf1 += p.s1;
+    }
+#endif
+
+    if (j < mh) {
+        vstore2((float2)(sumf, sumf1), 0, dst + (ulong)col * (uint)ne0 + row);
+    }
+#undef IQ2_S_GRID
+}

ggml/src/ggml-opencl/kernels/gemv_noshuffle_iq2_xs_f32.cl

@@ -0,0 +1,128 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define QK_K 256
+
+#ifndef IQ2_XS_MV_NSG
+#define IQ2_XS_MV_NSG 4
+#endif
+
+inline float4 iq2_xs_vals(uint gv, uint sgv, uint base) {
+    const uint  s   = sgv >> base;
+    const uint4 sgn = (uint4)(s << 31, s << 30, s << 29, s << 28) & 0x80000000u;
+    return as_float4(as_uint4(convert_float4(as_uchar4(gv))) ^ sgn);
+}
+
+kernel void kernel_gemv_noshuffle_iq2_xs_f32(
+        __read_only image1d_buffer_t grid_img,
+        global const ushort * src0_qs,
+        global const uchar  * src0_sc,
+        global const half   * src0_d,
+        global const float  * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne10,
+        int ne0
+) {
+    src1 = (global const float *)((global const char *)src1 + offset1);
+    dst  = (global float       *)((global char       *)dst  + offsetd);
+
+    const uint m   = (uint)ne01;
+    const uint K   = (uint)ne00;
+    const uint nsb = K / QK_K;
+
+    const uint lid = get_local_id(0);
+    const uint sgi = get_local_id(1);
+    const uint col = get_group_id(1);
+
+    global const float * y = src1 + (ulong)col * (uint)ne10;
+
+#define IQ2_XS_GRID(i) (read_imageui(grid_img, (int)(i)).x)
+
+    const uint mh  = m >> 1;
+    const uint j   = get_group_id(0) * 64u + lid;
+    const uint row = j << 1;
+
+    float sumf  = 0.f;
+    float sumf1 = 0.f;
+
+    if (j < mh) {
+        global const uint   * qsu = (global const uint   *)src0_qs;
+        global const ushort * scu = (global const ushort *)src0_sc;
+
+        for (uint ib = sgi; ib < nsb; ib += IQ2_XS_MV_NSG) {
+            const half2 dh = vload2(j + ib * mh, src0_d);
+            const float d0 = (float)dh.s0;
+            const float d1 = (float)dh.s1;
+
+            float acc0 = 0.f, acc1 = 0.f;
+            for (uint sb = 0; sb < 8u; ++sb) {
+                const uint sub = ib * 8u + sb;
+                const uint scv = (uint)scu[j + sub * mh];
+
+                const float lo0 = 0.5f + (float)( scv        & 0xFu);
+                const float hi0 = 0.5f + (float)((scv >>  4) & 0xFu);
+                const float lo1 = 0.5f + (float)((scv >>  8) & 0xFu);
+                const float hi1 = 0.5f + (float)((scv >> 12) & 0xFu);
+
+                const uint grp = sub * 4u;
+                const uint qsb = j + grp * mh;
+
+                float alo0 = 0.f, ahi0 = 0.f, alo1 = 0.f, ahi1 = 0.f;
+                for (uint u = 0; u < 4u; ++u) {
+                    const uint qsv = qsu[qsb + u * mh];
+                    const uint q0  = qsv & 0xFFFFu;
+                    const uint q1  = qsv >> 16;
+
+                    const uint g0  = (q0 & 511u) << 1;
+                    const uint g1  = (q1 & 511u) << 1;
+                    const uint sg0 = iq2_xs_signs(q0 >> 9);
+                    const uint sg1 = iq2_xs_signs(q1 >> 9);
+
+                    const float4 yl = vload4(sub * 8u + u * 2u + 0u, y);
+                    const float4 yh = vload4(sub * 8u + u * 2u + 1u, y);
+
+                    const float a0 = dot(yl, iq2_xs_vals(IQ2_XS_GRID(g0    ), sg0, 0u))
+                                   + dot(yh, iq2_xs_vals(IQ2_XS_GRID(g0 + 1), sg0, 4u));
+                    const float a1 = dot(yl, iq2_xs_vals(IQ2_XS_GRID(g1    ), sg1, 0u))
+                                   + dot(yh, iq2_xs_vals(IQ2_XS_GRID(g1 + 1), sg1, 4u));
+
+                    if (u < 2u) { alo0 += a0; alo1 += a1; }
+                    else        { ahi0 += a0; ahi1 += a1; }
+                }
+                acc0 += lo0 * alo0 + hi0 * ahi0;
+                acc1 += lo1 * alo1 + hi1 * ahi1;
+            }
+            sumf  += d0 * 0.25f * acc0;
+            sumf1 += d1 * 0.25f * acc1;
+        }
+    }
+
+#if IQ2_XS_MV_NSG > 1
+    __local float2 part[IQ2_XS_MV_NSG][64];
+    part[sgi][lid] = (float2)(sumf, sumf1);
+    barrier(CLK_LOCAL_MEM_FENCE);
+    if (sgi != 0) {
+        return;
+    }
+    for (uint s = 1; s < IQ2_XS_MV_NSG; ++s) {
+        const float2 p = part[s][lid];
+        sumf  += p.s0;
+        sumf1 += p.s1;
+    }
+#endif
+
+    if (j < mh) {
+        vstore2((float2)(sumf, sumf1), 0, dst + (ulong)col * (uint)ne0 + row);
+    }
+#undef IQ2_XS_GRID
+}
+
+kernel void kernel_iq2_xs_grid_export(global uint * out) {
+    const uint i = get_global_id(0);
+    if (i < 1024u) {
+        out[i] = iq2_xs_grid[i];
+    }
+}

ggml/src/ggml-opencl/kernels/gemv_noshuffle_iq2_xxs_f32.cl

@@ -0,0 +1,117 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#define QK_K 256
+
+#ifndef IQ2_XXS_MV_NSG
+#define IQ2_XXS_MV_NSG 4
+#endif
+
+inline float4 iq2_xxs_vals(uint gv, uint sgv, uint base) {
+    const uint  s   = sgv >> base;
+    const uint4 sgn = (uint4)(s << 31, s << 30, s << 29, s << 28) & 0x80000000u;
+    return as_float4(as_uint4(convert_float4(as_uchar4(gv))) ^ sgn);
+}
+
+kernel void kernel_gemv_noshuffle_iq2_xxs_f32(
+        __read_only image1d_buffer_t grid_img,
+        global const uchar * src0_qs,
+        global const uint  * src0_sas,
+        global const half  * src0_d,
+        global const float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne10,
+        int ne0
+) {
+    src1 = (global const float *)((global const char *)src1 + offset1);
+    dst  = (global float       *)((global char       *)dst  + offsetd);
+
+    const uint m   = (uint)ne01;
+    const uint K   = (uint)ne00;
+    const uint nsb = K / QK_K;
+
+    const uint lid = get_local_id(0);
+    const uint sgi = get_local_id(1);
+    const uint col = get_group_id(1);
+
+    global const float * y = src1 + (ulong)col * (uint)ne10;
+
+#define IQ2_XXS_GRID(i) (read_imageui(grid_img, (int)(i)).x)
+
+    const uint mh  = m >> 1;
+    const uint j   = get_group_id(0) * 64u + lid;
+    const uint row = j << 1;
+
+    float sumf  = 0.f;
+    float sumf1 = 0.f;
+
+    if (j < mh) {
+        global const ushort * qsu = (global const ushort *)src0_qs;
+
+        for (uint ib = sgi; ib < nsb; ib += IQ2_XXS_MV_NSG) {
+            const half2 dh = vload2(j + ib * mh, src0_d);
+            const float d0 = (float)dh.s0;
+            const float d1 = (float)dh.s1;
+
+            float acc0 = 0.f, acc1 = 0.f;
+            for (uint sb = 0; sb < 8u; ++sb) {
+                const uint  sub = ib * 8u + sb;
+                const uint2 aux = vload2(j + sub * mh, src0_sas);
+
+                const uint grp = sub * 4u;
+                const uint qsb = j + grp * mh;
+
+                float a0 = 0.f, a1 = 0.f;
+                for (uint u = 0; u < 4u; ++u) {
+                    const uint qsv = (uint)qsu[qsb + u * mh];
+                    const uint sg0 = iq2_xxs_signs((aux.s0 >> (7u * u)) & 127u);
+                    const uint sg1 = iq2_xxs_signs((aux.s1 >> (7u * u)) & 127u);
+
+                    const uint g0 = ( qsv       & 0xFFu) << 1;
+                    const uint g1 = ((qsv >> 8) & 0xFFu) << 1;
+
+                    const float4 yl = vload4(sub * 8u + u * 2u + 0u, y);
+                    const float4 yh = vload4(sub * 8u + u * 2u + 1u, y);
+
+                    a0 += dot(yl, iq2_xxs_vals(IQ2_XXS_GRID(g0    ), sg0, 0u))
+                        + dot(yh, iq2_xxs_vals(IQ2_XXS_GRID(g0 + 1), sg0, 4u));
+                    a1 += dot(yl, iq2_xxs_vals(IQ2_XXS_GRID(g1    ), sg1, 0u))
+                        + dot(yh, iq2_xxs_vals(IQ2_XXS_GRID(g1 + 1), sg1, 4u));
+                }
+                acc0 += (0.5f + (float)(aux.s0 >> 28)) * a0;
+                acc1 += (0.5f + (float)(aux.s1 >> 28)) * a1;
+            }
+            sumf  += d0 * 0.25f * acc0;
+            sumf1 += d1 * 0.25f * acc1;
+        }
+    }
+
+#if IQ2_XXS_MV_NSG > 1
+    __local float2 part[IQ2_XXS_MV_NSG][64];
+    part[sgi][lid] = (float2)(sumf, sumf1);
+    barrier(CLK_LOCAL_MEM_FENCE);
+    if (sgi != 0) {
+        return;
+    }
+    for (uint s = 1; s < IQ2_XXS_MV_NSG; ++s) {
+        const float2 p = part[s][lid];
+        sumf  += p.s0;
+        sumf1 += p.s1;
+    }
+#endif
+
+    if (j < mh) {
+        vstore2((float2)(sumf, sumf1), 0, dst + (ulong)col * (uint)ne0 + row);
+    }
+#undef IQ2_XXS_GRID
+}
+
+kernel void kernel_iq2_xxs_grid_export(global uint * out) {
+    const uint i = get_global_id(0);
+    if (i < 512u) {
+        out[i] = iq2_xxs_grid[i];
+    }
+}

ggml/src/ggml-opencl/kernels/gemv_noshuffle_q2_0_f32.cl

@@ -0,0 +1,64 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifndef Q2_0_MV_NSG
+#define Q2_0_MV_NSG 8
+#endif
+
+kernel void kernel_gemv_noshuffle_q2_0_f32(
+        global const uchar * src0_qs,
+        global const half * src0_d,
+        global const float * src1,
+        ulong offset1,
+        global float * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        int ne10,
+        int ne0
+) {
+    src1 = (global const float *)((global const char *)src1 + offset1);
+    dst = (global float *)((global char *)dst + offsetd);
+
+    const uint m = (uint)ne01;
+    const uint mh = m >> 1;
+    const uint lid = get_local_id(0);
+    const uint sgi = get_local_id(1);
+    const uint col = get_group_id(1);
+    const uint j = get_group_id(0)*64u + lid;
+    const uint row = j << 1;
+    global const float * y = src1 + (ulong)col * (uint)ne10;
+
+    float2 sum = 0.0f;
+    if (j < mh) {
+        global const ushort * qs = (global const ushort *)src0_qs;
+        for (uint ib = sgi; ib < (uint)ne00/64; ib += Q2_0_MV_NSG) {
+            const float2 d = convert_float2(vload2(j + ib*mh, src0_d));
+            float2 acc = 0.0f;
+            for (uint u = 0; u < 16; ++u) {
+                const uint group = ib*16 + u;
+                const uint q = qs[j + group*mh];
+                const float4 yv = vload4(group, y);
+                const int4 v0 = convert_int4(((uint4)(q) >> (uint4)(0, 2, 4, 6)) & (uint4)(3)) - (int4)(1);
+                const int4 v1 = convert_int4(((uint4)(q) >> (uint4)(8, 10, 12, 14)) & (uint4)(3)) - (int4)(1);
+                acc.s0 += dot(yv, convert_float4(v0));
+                acc.s1 += dot(yv, convert_float4(v1));
+            }
+            sum += d*acc;
+        }
+    }
+
+#if Q2_0_MV_NSG > 1
+    local float2 part[Q2_0_MV_NSG][64];
+    part[sgi][lid] = sum;
+    barrier(CLK_LOCAL_MEM_FENCE);
+    if (sgi != 0) {
+        return;
+    }
+    for (uint s = 1; s < Q2_0_MV_NSG; ++s) {
+        sum += part[s][lid];
+    }
+#endif
+    if (j < mh) {
+        vstore2(sum, 0, dst + (ulong)col*(uint)ne0 + row);
+    }
+}

ggml/src/ggml-opencl/kernels/iq2_s_grid.cl

@@ -0,0 +1,259 @@
+// IQ2_S codebook, stored as pairs of four packed magnitudes.
+constant uint iq2_s_grid[2048] = {
+    0x08080808, 0x08080808, 0x0808082b, 0x08080808, 0x08081919, 0x08080808, 0x08082b08, 0x08080808,
+    0x08082b2b, 0x08080808, 0x08190819, 0x08080808, 0x08191908, 0x08080808, 0x0819192b, 0x08080808,
+    0x08192b19, 0x08080808, 0x082b0808, 0x08080808, 0x082b082b, 0x08080808, 0x082b1919, 0x08080808,
+    0x082b2b08, 0x08080808, 0x19080819, 0x08080808, 0x19081908, 0x08080808, 0x1908192b, 0x08080808,
+    0x19082b19, 0x08080808, 0x19190808, 0x08080808, 0x1919082b, 0x08080808, 0x19191919, 0x08080808,
+    0x19192b08, 0x08080808, 0x192b0819, 0x08080808, 0x192b1908, 0x08080808, 0x192b192b, 0x08080808,
+    0x192b2b19, 0x08080808, 0x2b080808, 0x08080808, 0x2b08082b, 0x08080808, 0x2b081919, 0x08080808,
+    0x2b082b08, 0x08080808, 0x2b190819, 0x08080808, 0x2b191908, 0x08080808, 0x2b2b0808, 0x08080808,
+    0x2b2b1919, 0x08080808, 0x2b2b2b2b, 0x08080808, 0x08080819, 0x08080819, 0x08081908, 0x08080819,
+    0x0808192b, 0x08080819, 0x08082b19, 0x08080819, 0x08190808, 0x08080819, 0x0819082b, 0x08080819,
+    0x08191919, 0x08080819, 0x08192b08, 0x08080819, 0x082b0819, 0x08080819, 0x082b1908, 0x08080819,
+    0x19080808, 0x08080819, 0x1908082b, 0x08080819, 0x19081919, 0x08080819, 0x19082b08, 0x08080819,
+    0x19190819, 0x08080819, 0x19191908, 0x08080819, 0x1919192b, 0x08080819, 0x19192b19, 0x08080819,
+    0x192b0808, 0x08080819, 0x192b1919, 0x08080819, 0x192b2b08, 0x08080819, 0x2b080819, 0x08080819,
+    0x2b081908, 0x08080819, 0x2b190808, 0x08080819, 0x2b19082b, 0x08080819, 0x2b191919, 0x08080819,
+    0x2b2b0819, 0x08080819, 0x2b2b1908, 0x08080819, 0x08080808, 0x0808082b, 0x0808082b, 0x0808082b,
+    0x08081919, 0x0808082b, 0x08082b08, 0x0808082b, 0x08190819, 0x0808082b, 0x08191908, 0x0808082b,
+    0x082b0808, 0x0808082b, 0x082b2b2b, 0x0808082b, 0x19080819, 0x0808082b, 0x19081908, 0x0808082b,
+    0x1908192b, 0x0808082b, 0x19082b19, 0x0808082b, 0x19190808, 0x0808082b, 0x19191919, 0x0808082b,
+    0x2b080808, 0x0808082b, 0x2b081919, 0x0808082b, 0x2b082b2b, 0x0808082b, 0x2b191908, 0x0808082b,
+    0x2b2b082b, 0x0808082b, 0x08080819, 0x08081908, 0x08081908, 0x08081908, 0x0808192b, 0x08081908,
+    0x08082b19, 0x08081908, 0x08190808, 0x08081908, 0x0819082b, 0x08081908, 0x08191919, 0x08081908,
+    0x08192b08, 0x08081908, 0x082b0819, 0x08081908, 0x082b1908, 0x08081908, 0x082b192b, 0x08081908,
+    0x082b2b19, 0x08081908, 0x19080808, 0x08081908, 0x1908082b, 0x08081908, 0x19081919, 0x08081908,
+    0x19082b08, 0x08081908, 0x19082b2b, 0x08081908, 0x19190819, 0x08081908, 0x19191908, 0x08081908,
+    0x1919192b, 0x08081908, 0x19192b19, 0x08081908, 0x192b0808, 0x08081908, 0x192b082b, 0x08081908,
+    0x192b1919, 0x08081908, 0x2b080819, 0x08081908, 0x2b081908, 0x08081908, 0x2b08192b, 0x08081908,
+    0x2b082b19, 0x08081908, 0x2b190808, 0x08081908, 0x2b191919, 0x08081908, 0x2b192b08, 0x08081908,
+    0x2b2b0819, 0x08081908, 0x2b2b1908, 0x08081908, 0x08080808, 0x08081919, 0x0808082b, 0x08081919,
+    0x08081919, 0x08081919, 0x08082b08, 0x08081919, 0x08082b2b, 0x08081919, 0x08190819, 0x08081919,
+    0x08191908, 0x08081919, 0x0819192b, 0x08081919, 0x08192b19, 0x08081919, 0x082b0808, 0x08081919,
+    0x082b1919, 0x08081919, 0x082b2b08, 0x08081919, 0x19080819, 0x08081919, 0x19081908, 0x08081919,
+    0x1908192b, 0x08081919, 0x19082b19, 0x08081919, 0x19190808, 0x08081919, 0x1919082b, 0x08081919,
+    0x19191919, 0x08081919, 0x19192b08, 0x08081919, 0x192b0819, 0x08081919, 0x192b1908, 0x08081919,
+    0x2b080808, 0x08081919, 0x2b08082b, 0x08081919, 0x2b081919, 0x08081919, 0x2b082b08, 0x08081919,
+    0x2b190819, 0x08081919, 0x2b191908, 0x08081919, 0x2b2b0808, 0x08081919, 0x08080819, 0x0808192b,
+    0x08081908, 0x0808192b, 0x0808192b, 0x0808192b, 0x08082b19, 0x0808192b, 0x08190808, 0x0808192b,
+    0x08191919, 0x0808192b, 0x19080808, 0x0808192b, 0x19081919, 0x0808192b, 0x19082b08, 0x0808192b,
+    0x19190819, 0x0808192b, 0x19191908, 0x0808192b, 0x192b0808, 0x0808192b, 0x2b080819, 0x0808192b,
+    0x2b081908, 0x0808192b, 0x2b190808, 0x0808192b, 0x08080808, 0x08082b08, 0x0808082b, 0x08082b08,
+    0x08081919, 0x08082b08, 0x08082b08, 0x08082b08, 0x08190819, 0x08082b08, 0x08191908, 0x08082b08,
+    0x0819192b, 0x08082b08, 0x08192b19, 0x08082b08, 0x082b0808, 0x08082b08, 0x082b1919, 0x08082b08,
+    0x082b2b2b, 0x08082b08, 0x19080819, 0x08082b08, 0x19081908, 0x08082b08, 0x1908192b, 0x08082b08,
+    0x19082b19, 0x08082b08, 0x19190808, 0x08082b08, 0x1919082b, 0x08082b08, 0x19191919, 0x08082b08,
+    0x19192b08, 0x08082b08, 0x192b0819, 0x08082b08, 0x192b1908, 0x08082b08, 0x2b080808, 0x08082b08,
+    0x2b081919, 0x08082b08, 0x2b191908, 0x08082b08, 0x2b2b2b2b, 0x08082b08, 0x08080819, 0x08082b19,
+    0x08081908, 0x08082b19, 0x08190808, 0x08082b19, 0x0819082b, 0x08082b19, 0x08191919, 0x08082b19,
+    0x08192b08, 0x08082b19, 0x082b0819, 0x08082b19, 0x19080808, 0x08082b19, 0x19081919, 0x08082b19,
+    0x19082b08, 0x08082b19, 0x19190819, 0x08082b19, 0x19191908, 0x08082b19, 0x192b0808, 0x08082b19,
+    0x2b080819, 0x08082b19, 0x2b190808, 0x08082b19, 0x08080808, 0x08082b2b, 0x08190819, 0x08082b2b,
+    0x08191908, 0x08082b2b, 0x082b082b, 0x08082b2b, 0x082b2b08, 0x08082b2b, 0x082b2b2b, 0x08082b2b,
+    0x19190808, 0x08082b2b, 0x2b192b19, 0x08082b2b, 0x08080819, 0x08190808, 0x08081908, 0x08190808,
+    0x0808192b, 0x08190808, 0x08082b19, 0x08190808, 0x08190808, 0x08190808, 0x0819082b, 0x08190808,
+    0x08191919, 0x08190808, 0x08192b08, 0x08190808, 0x082b0819, 0x08190808, 0x082b1908, 0x08190808,
+    0x082b192b, 0x08190808, 0x19080808, 0x08190808, 0x1908082b, 0x08190808, 0x19081919, 0x08190808,
+    0x19082b08, 0x08190808, 0x19190819, 0x08190808, 0x19191908, 0x08190808, 0x1919192b, 0x08190808,
+    0x19192b19, 0x08190808, 0x192b0808, 0x08190808, 0x192b082b, 0x08190808, 0x192b1919, 0x08190808,
+    0x192b2b08, 0x08190808, 0x2b080819, 0x08190808, 0x2b081908, 0x08190808, 0x2b08192b, 0x08190808,
+    0x2b190808, 0x08190808, 0x2b191919, 0x08190808, 0x2b192b08, 0x08190808, 0x2b2b0819, 0x08190808,
+    0x2b2b1908, 0x08190808, 0x08080808, 0x08190819, 0x0808082b, 0x08190819, 0x08081919, 0x08190819,
+    0x08082b08, 0x08190819, 0x08082b2b, 0x08190819, 0x08190819, 0x08190819, 0x08191908, 0x08190819,
+    0x0819192b, 0x08190819, 0x08192b19, 0x08190819, 0x082b0808, 0x08190819, 0x082b082b, 0x08190819,
+    0x082b1919, 0x08190819, 0x082b2b08, 0x08190819, 0x19080819, 0x08190819, 0x19081908, 0x08190819,
+    0x1908192b, 0x08190819, 0x19082b19, 0x08190819, 0x19190808, 0x08190819, 0x1919082b, 0x08190819,
+    0x19191919, 0x08190819, 0x19192b08, 0x08190819, 0x192b0819, 0x08190819, 0x192b1908, 0x08190819,
+    0x2b080808, 0x08190819, 0x2b08082b, 0x08190819, 0x2b081919, 0x08190819, 0x2b082b08, 0x08190819,
+    0x2b190819, 0x08190819, 0x2b191908, 0x08190819, 0x08080819, 0x0819082b, 0x08081908, 0x0819082b,
+    0x08082b19, 0x0819082b, 0x08190808, 0x0819082b, 0x08191919, 0x0819082b, 0x082b0819, 0x0819082b,
+    0x082b1908, 0x0819082b, 0x19080808, 0x0819082b, 0x19081919, 0x0819082b, 0x19190819, 0x0819082b,
+    0x19191908, 0x0819082b, 0x2b080819, 0x0819082b, 0x2b081908, 0x0819082b, 0x2b190808, 0x0819082b,
+    0x08080808, 0x08191908, 0x0808082b, 0x08191908, 0x08081919, 0x08191908, 0x08082b08, 0x08191908,
+    0x08190819, 0x08191908, 0x08191908, 0x08191908, 0x0819192b, 0x08191908, 0x08192b19, 0x08191908,
+    0x082b0808, 0x08191908, 0x082b1919, 0x08191908, 0x082b2b08, 0x08191908, 0x19080819, 0x08191908,
+    0x19081908, 0x08191908, 0x1908192b, 0x08191908, 0x19082b19, 0x08191908, 0x19190808, 0x08191908,
+    0x1919082b, 0x08191908, 0x19191919, 0x08191908, 0x19192b08, 0x08191908, 0x192b0819, 0x08191908,
+    0x192b1908, 0x08191908, 0x2b080808, 0x08191908, 0x2b08082b, 0x08191908, 0x2b081919, 0x08191908,
+    0x2b082b08, 0x08191908, 0x2b190819, 0x08191908, 0x2b191908, 0x08191908, 0x2b2b0808, 0x08191908,
+    0x08080819, 0x08191919, 0x08081908, 0x08191919, 0x0808192b, 0x08191919, 0x08082b19, 0x08191919,
+    0x08190808, 0x08191919, 0x0819082b, 0x08191919, 0x08191919, 0x08191919, 0x08192b08, 0x08191919,
+    0x082b0819, 0x08191919, 0x082b1908, 0x08191919, 0x19080808, 0x08191919, 0x1908082b, 0x08191919,
+    0x19081919, 0x08191919, 0x19082b08, 0x08191919, 0x19190819, 0x08191919, 0x19191908, 0x08191919,
+    0x192b0808, 0x08191919, 0x2b080819, 0x08191919, 0x2b081908, 0x08191919, 0x2b190808, 0x08191919,
+    0x08080808, 0x0819192b, 0x08081919, 0x0819192b, 0x08082b08, 0x0819192b, 0x08190819, 0x0819192b,
+    0x08191908, 0x0819192b, 0x082b0808, 0x0819192b, 0x19080819, 0x0819192b, 0x19081908, 0x0819192b,
+    0x19190808, 0x0819192b, 0x2b080808, 0x0819192b, 0x2b2b2b2b, 0x0819192b, 0x08080819, 0x08192b08,
+    0x08081908, 0x08192b08, 0x0808192b, 0x08192b08, 0x08082b19, 0x08192b08, 0x08190808, 0x08192b08,
+    0x08191919, 0x08192b08, 0x08192b08, 0x08192b08, 0x082b0819, 0x08192b08, 0x19080808, 0x08192b08,
+    0x1908082b, 0x08192b08, 0x19081919, 0x08192b08, 0x19082b08, 0x08192b08, 0x19190819, 0x08192b08,
+    0x19191908, 0x08192b08, 0x192b0808, 0x08192b08, 0x2b080819, 0x08192b08, 0x2b081908, 0x08192b08,
+    0x08080808, 0x08192b19, 0x0808082b, 0x08192b19, 0x08081919, 0x08192b19, 0x08082b08, 0x08192b19,
+    0x08190819, 0x08192b19, 0x08191908, 0x08192b19, 0x082b0808, 0x08192b19, 0x19080819, 0x08192b19,
+    0x19081908, 0x08192b19, 0x19190808, 0x08192b19, 0x192b2b19, 0x08192b19, 0x2b2b082b, 0x08192b19,
+    0x08081908, 0x08192b2b, 0x08190808, 0x08192b2b, 0x19080808, 0x08192b2b, 0x1919192b, 0x08192b2b,
+    0x08080808, 0x082b0808, 0x0808082b, 0x082b0808, 0x08081919, 0x082b0808, 0x08082b08, 0x082b0808,
+    0x08190819, 0x082b0808, 0x08191908, 0x082b0808, 0x0819192b, 0x082b0808, 0x08192b19, 0x082b0808,
+    0x082b0808, 0x082b0808, 0x082b1919, 0x082b0808, 0x082b2b2b, 0x082b0808, 0x19080819, 0x082b0808,
+    0x19081908, 0x082b0808, 0x19190808, 0x082b0808, 0x1919082b, 0x082b0808, 0x19191919, 0x082b0808,
+    0x192b1908, 0x082b0808, 0x2b080808, 0x082b0808, 0x2b082b2b, 0x082b0808, 0x2b191908, 0x082b0808,
+    0x2b2b2b2b, 0x082b0808, 0x08080819, 0x082b0819, 0x08081908, 0x082b0819, 0x08190808, 0x082b0819,
+    0x0819082b, 0x082b0819, 0x08191919, 0x082b0819, 0x082b0819, 0x082b0819, 0x19080808, 0x082b0819,
+    0x1908082b, 0x082b0819, 0x19081919, 0x082b0819, 0x19190819, 0x082b0819, 0x19191908, 0x082b0819,
+    0x192b0808, 0x082b0819, 0x2b080819, 0x082b0819, 0x2b081908, 0x082b0819, 0x2b190808, 0x082b0819,
+    0x08080808, 0x082b082b, 0x08082b2b, 0x082b082b, 0x082b082b, 0x082b082b, 0x082b2b08, 0x082b082b,
+    0x082b2b2b, 0x082b082b, 0x19081908, 0x082b082b, 0x19190808, 0x082b082b, 0x2b082b08, 0x082b082b,
+    0x2b082b2b, 0x082b082b, 0x2b2b2b08, 0x082b082b, 0x08080819, 0x082b1908, 0x08081908, 0x082b1908,
+    0x0808192b, 0x082b1908, 0x08082b19, 0x082b1908, 0x08190808, 0x082b1908, 0x08191919, 0x082b1908,
+    0x08192b08, 0x082b1908, 0x082b0819, 0x082b1908, 0x082b1908, 0x082b1908, 0x19080808, 0x082b1908,
+    0x1908082b, 0x082b1908, 0x19081919, 0x082b1908, 0x19082b08, 0x082b1908, 0x19190819, 0x082b1908,
+    0x19191908, 0x082b1908, 0x192b0808, 0x082b1908, 0x2b080819, 0x082b1908, 0x2b081908, 0x082b1908,
+    0x2b190808, 0x082b1908, 0x08080808, 0x082b1919, 0x08081919, 0x082b1919, 0x08082b08, 0x082b1919,
+    0x08190819, 0x082b1919, 0x08191908, 0x082b1919, 0x082b0808, 0x082b1919, 0x19080819, 0x082b1919,
+    0x19081908, 0x082b1919, 0x19190808, 0x082b1919, 0x192b192b, 0x082b1919, 0x2b080808, 0x082b1919,
+    0x08080819, 0x082b192b, 0x08081908, 0x082b192b, 0x08190808, 0x082b192b, 0x19080808, 0x082b192b,
+    0x19192b19, 0x082b192b, 0x08080808, 0x082b2b08, 0x08081919, 0x082b2b08, 0x08190819, 0x082b2b08,
+    0x08191908, 0x082b2b08, 0x19080819, 0x082b2b08, 0x19081908, 0x082b2b08, 0x19190808, 0x082b2b08,
+    0x2b082b2b, 0x082b2b08, 0x2b2b2b2b, 0x082b2b08, 0x08080819, 0x082b2b19, 0x08081908, 0x082b2b19,
+    0x08190808, 0x082b2b19, 0x2b191919, 0x082b2b19, 0x08082b2b, 0x082b2b2b, 0x082b082b, 0x082b2b2b,
+    0x192b1908, 0x082b2b2b, 0x2b082b08, 0x082b2b2b, 0x2b082b2b, 0x082b2b2b, 0x08080819, 0x19080808,
+    0x08081908, 0x19080808, 0x0808192b, 0x19080808, 0x08082b19, 0x19080808, 0x08190808, 0x19080808,
+    0x0819082b, 0x19080808, 0x08191919, 0x19080808, 0x08192b08, 0x19080808, 0x08192b2b, 0x19080808,
+    0x082b0819, 0x19080808, 0x082b1908, 0x19080808, 0x082b192b, 0x19080808, 0x19080808, 0x19080808,
+    0x1908082b, 0x19080808, 0x19081919, 0x19080808, 0x19082b08, 0x19080808, 0x19082b2b, 0x19080808,
+    0x19190819, 0x19080808, 0x19191908, 0x19080808, 0x1919192b, 0x19080808, 0x19192b19, 0x19080808,
+    0x192b0808, 0x19080808, 0x192b082b, 0x19080808, 0x192b1919, 0x19080808, 0x2b080819, 0x19080808,
+    0x2b081908, 0x19080808, 0x2b190808, 0x19080808, 0x2b191919, 0x19080808, 0x2b192b08, 0x19080808,
+    0x2b2b0819, 0x19080808, 0x2b2b1908, 0x19080808, 0x08080808, 0x19080819, 0x0808082b, 0x19080819,
+    0x08081919, 0x19080819, 0x08082b08, 0x19080819, 0x08190819, 0x19080819, 0x08191908, 0x19080819,
+    0x0819192b, 0x19080819, 0x08192b19, 0x19080819, 0x082b0808, 0x19080819, 0x082b082b, 0x19080819,
+    0x082b1919, 0x19080819, 0x19080819, 0x19080819, 0x19081908, 0x19080819, 0x1908192b, 0x19080819,
+    0x19082b19, 0x19080819, 0x19190808, 0x19080819, 0x1919082b, 0x19080819, 0x19191919, 0x19080819,
+    0x19192b08, 0x19080819, 0x192b0819, 0x19080819, 0x192b1908, 0x19080819, 0x2b080808, 0x19080819,
+    0x2b08082b, 0x19080819, 0x2b081919, 0x19080819, 0x2b082b08, 0x19080819, 0x2b190819, 0x19080819,
+    0x2b191908, 0x19080819, 0x2b2b0808, 0x19080819, 0x08080819, 0x1908082b, 0x08081908, 0x1908082b,
+    0x08190808, 0x1908082b, 0x0819082b, 0x1908082b, 0x08191919, 0x1908082b, 0x08192b08, 0x1908082b,
+    0x082b1908, 0x1908082b, 0x19080808, 0x1908082b, 0x19081919, 0x1908082b, 0x19082b08, 0x1908082b,
+    0x19190819, 0x1908082b, 0x19191908, 0x1908082b, 0x192b0808, 0x1908082b, 0x2b080819, 0x1908082b,
+    0x2b081908, 0x1908082b, 0x08080808, 0x19081908, 0x0808082b, 0x19081908, 0x08081919, 0x19081908,
+    0x08082b08, 0x19081908, 0x08082b2b, 0x19081908, 0x08190819, 0x19081908, 0x08191908, 0x19081908,
+    0x0819192b, 0x19081908, 0x08192b19, 0x19081908, 0x082b0808, 0x19081908, 0x082b082b, 0x19081908,
+    0x082b1919, 0x19081908, 0x082b2b08, 0x19081908, 0x19080819, 0x19081908, 0x19081908, 0x19081908,
+    0x1908192b, 0x19081908, 0x19082b19, 0x19081908, 0x19190808, 0x19081908, 0x1919082b, 0x19081908,
+    0x19191919, 0x19081908, 0x19192b08, 0x19081908, 0x192b0819, 0x19081908, 0x192b1908, 0x19081908,
+    0x2b080808, 0x19081908, 0x2b08082b, 0x19081908, 0x2b081919, 0x19081908, 0x2b082b08, 0x19081908,
+    0x2b190819, 0x19081908, 0x2b191908, 0x19081908, 0x2b2b0808, 0x19081908, 0x08080819, 0x19081919,
+    0x08081908, 0x19081919, 0x0808192b, 0x19081919, 0x08082b19, 0x19081919, 0x08190808, 0x19081919,
+    0x0819082b, 0x19081919, 0x08191919, 0x19081919, 0x08192b08, 0x19081919, 0x082b0819, 0x19081919,
+    0x082b1908, 0x19081919, 0x19080808, 0x19081919, 0x1908082b, 0x19081919, 0x19081919, 0x19081919,
+    0x19082b08, 0x19081919, 0x19190819, 0x19081919, 0x19191908, 0x19081919, 0x192b0808, 0x19081919,
+    0x192b2b2b, 0x19081919, 0x2b080819, 0x19081919, 0x2b081908, 0x19081919, 0x2b190808, 0x19081919,
+    0x08080808, 0x1908192b, 0x0808082b, 0x1908192b, 0x08081919, 0x1908192b, 0x08082b08, 0x1908192b,
+    0x08190819, 0x1908192b, 0x08191908, 0x1908192b, 0x082b0808, 0x1908192b, 0x19080819, 0x1908192b,
+    0x19081908, 0x1908192b, 0x19190808, 0x1908192b, 0x2b080808, 0x1908192b, 0x2b2b1919, 0x1908192b,
+    0x08080819, 0x19082b08, 0x08081908, 0x19082b08, 0x08082b19, 0x19082b08, 0x08190808, 0x19082b08,
+    0x0819082b, 0x19082b08, 0x08191919, 0x19082b08, 0x08192b08, 0x19082b08, 0x082b0819, 0x19082b08,
+    0x082b1908, 0x19082b08, 0x19080808, 0x19082b08, 0x1908082b, 0x19082b08, 0x19081919, 0x19082b08,
+    0x19082b08, 0x19082b08, 0x19190819, 0x19082b08, 0x19191908, 0x19082b08, 0x192b0808, 0x19082b08,
+    0x2b081908, 0x19082b08, 0x2b190808, 0x19082b08, 0x08080808, 0x19082b19, 0x0808082b, 0x19082b19,
+    0x08081919, 0x19082b19, 0x08082b08, 0x19082b19, 0x08190819, 0x19082b19, 0x08191908, 0x19082b19,
+    0x082b0808, 0x19082b19, 0x19080819, 0x19082b19, 0x19081908, 0x19082b19, 0x19190808, 0x19082b19,
+    0x2b080808, 0x19082b19, 0x2b19192b, 0x19082b19, 0x08080819, 0x19082b2b, 0x08081908, 0x19082b2b,
+    0x08190808, 0x19082b2b, 0x19080808, 0x19082b2b, 0x08080808, 0x19190808, 0x0808082b, 0x19190808,
+    0x08081919, 0x19190808, 0x08082b08, 0x19190808, 0x08190819, 0x19190808, 0x08191908, 0x19190808,
+    0x0819192b, 0x19190808, 0x08192b19, 0x19190808, 0x082b0808, 0x19190808, 0x082b082b, 0x19190808,
+    0x082b1919, 0x19190808, 0x082b2b08, 0x19190808, 0x19080819, 0x19190808, 0x19081908, 0x19190808,
+    0x1908192b, 0x19190808, 0x19082b19, 0x19190808, 0x19190808, 0x19190808, 0x1919082b, 0x19190808,
+    0x19191919, 0x19190808, 0x19192b08, 0x19190808, 0x192b0819, 0x19190808, 0x192b1908, 0x19190808,
+    0x2b080808, 0x19190808, 0x2b08082b, 0x19190808, 0x2b081919, 0x19190808, 0x2b082b08, 0x19190808,
+    0x2b190819, 0x19190808, 0x2b191908, 0x19190808, 0x08080819, 0x19190819, 0x08081908, 0x19190819,
+    0x0808192b, 0x19190819, 0x08082b19, 0x19190819, 0x08190808, 0x19190819, 0x0819082b, 0x19190819,
+    0x08191919, 0x19190819, 0x08192b08, 0x19190819, 0x082b0819, 0x19190819, 0x082b1908, 0x19190819,
+    0x19080808, 0x19190819, 0x1908082b, 0x19190819, 0x19081919, 0x19190819, 0x19082b08, 0x19190819,
+    0x19190819, 0x19190819, 0x19191908, 0x19190819, 0x192b0808, 0x19190819, 0x2b080819, 0x19190819,
+    0x2b081908, 0x19190819, 0x2b190808, 0x19190819, 0x08080808, 0x1919082b, 0x08081919, 0x1919082b,
+    0x08082b08, 0x1919082b, 0x08190819, 0x1919082b, 0x08191908, 0x1919082b, 0x082b0808, 0x1919082b,
+    0x19080819, 0x1919082b, 0x19081908, 0x1919082b, 0x19190808, 0x1919082b, 0x192b2b19, 0x1919082b,
+    0x2b080808, 0x1919082b, 0x08080819, 0x19191908, 0x08081908, 0x19191908, 0x0808192b, 0x19191908,
+    0x08082b19, 0x19191908, 0x08190808, 0x19191908, 0x0819082b, 0x19191908, 0x08191919, 0x19191908,
+    0x08192b08, 0x19191908, 0x082b0819, 0x19191908, 0x082b1908, 0x19191908, 0x19080808, 0x19191908,
+    0x1908082b, 0x19191908, 0x19081919, 0x19191908, 0x19082b08, 0x19191908, 0x19190819, 0x19191908,
+    0x19191908, 0x19191908, 0x192b0808, 0x19191908, 0x2b080819, 0x19191908, 0x2b081908, 0x19191908,
+    0x2b190808, 0x19191908, 0x08080808, 0x19191919, 0x0808082b, 0x19191919, 0x08081919, 0x19191919,
+    0x08082b08, 0x19191919, 0x08190819, 0x19191919, 0x08191908, 0x19191919, 0x082b0808, 0x19191919,
+    0x19080819, 0x19191919, 0x19081908, 0x19191919, 0x19190808, 0x19191919, 0x2b080808, 0x19191919,
+    0x08080819, 0x1919192b, 0x08081908, 0x1919192b, 0x08190808, 0x1919192b, 0x082b192b, 0x1919192b,
+    0x19080808, 0x1919192b, 0x08080808, 0x19192b08, 0x0808082b, 0x19192b08, 0x08081919, 0x19192b08,
+    0x08082b08, 0x19192b08, 0x08190819, 0x19192b08, 0x08191908, 0x19192b08, 0x082b0808, 0x19192b08,
+    0x19080819, 0x19192b08, 0x19081908, 0x19192b08, 0x19190808, 0x19192b08, 0x19192b2b, 0x19192b08,
+    0x2b080808, 0x19192b08, 0x08080819, 0x19192b19, 0x08081908, 0x19192b19, 0x08190808, 0x19192b19,
+    0x19080808, 0x19192b19, 0x08080808, 0x19192b2b, 0x08192b19, 0x19192b2b, 0x2b081919, 0x19192b2b,
+    0x2b2b2b08, 0x19192b2b, 0x08080819, 0x192b0808, 0x08081908, 0x192b0808, 0x0808192b, 0x192b0808,
+    0x08190808, 0x192b0808, 0x0819082b, 0x192b0808, 0x08191919, 0x192b0808, 0x08192b08, 0x192b0808,
+    0x082b0819, 0x192b0808, 0x082b1908, 0x192b0808, 0x19080808, 0x192b0808, 0x19081919, 0x192b0808,
+    0x19082b08, 0x192b0808, 0x19190819, 0x192b0808, 0x19191908, 0x192b0808, 0x192b0808, 0x192b0808,
+    0x2b081908, 0x192b0808, 0x2b190808, 0x192b0808, 0x08080808, 0x192b0819, 0x0808082b, 0x192b0819,
+    0x08081919, 0x192b0819, 0x08082b08, 0x192b0819, 0x08190819, 0x192b0819, 0x08191908, 0x192b0819,
+    0x082b0808, 0x192b0819, 0x19080819, 0x192b0819, 0x19081908, 0x192b0819, 0x19190808, 0x192b0819,
+    0x2b080808, 0x192b0819, 0x2b192b19, 0x192b0819, 0x08081908, 0x192b082b, 0x08190808, 0x192b082b,
+    0x19080808, 0x192b082b, 0x1919192b, 0x192b082b, 0x2b2b0819, 0x192b082b, 0x08080808, 0x192b1908,
+    0x08081919, 0x192b1908, 0x08082b08, 0x192b1908, 0x08190819, 0x192b1908, 0x08191908, 0x192b1908,
+    0x082b0808, 0x192b1908, 0x19080819, 0x192b1908, 0x19081908, 0x192b1908, 0x19190808, 0x192b1908,
+    0x2b080808, 0x192b1908, 0x08080819, 0x192b1919, 0x08081908, 0x192b1919, 0x08190808, 0x192b1919,
+    0x19080808, 0x192b1919, 0x19082b2b, 0x192b1919, 0x192b2b08, 0x192b1919, 0x2b19082b, 0x192b1919,
+    0x08080808, 0x192b192b, 0x2b191908, 0x192b192b, 0x08080819, 0x192b2b08, 0x08081908, 0x192b2b08,
+    0x08190808, 0x192b2b08, 0x192b1919, 0x192b2b08, 0x2b192b08, 0x192b2b08, 0x08080808, 0x192b2b19,
+    0x082b2b2b, 0x192b2b19, 0x1908082b, 0x192b2b2b, 0x2b2b0819, 0x192b2b2b, 0x08080808, 0x2b080808,
+    0x0808082b, 0x2b080808, 0x08081919, 0x2b080808, 0x08082b08, 0x2b080808, 0x08190819, 0x2b080808,
+    0x08191908, 0x2b080808, 0x08192b19, 0x2b080808, 0x082b0808, 0x2b080808, 0x082b1919, 0x2b080808,
+    0x19080819, 0x2b080808, 0x19081908, 0x2b080808, 0x19190808, 0x2b080808, 0x1919082b, 0x2b080808,
+    0x19191919, 0x2b080808, 0x19192b08, 0x2b080808, 0x192b0819, 0x2b080808, 0x2b080808, 0x2b080808,
+    0x2b081919, 0x2b080808, 0x2b190819, 0x2b080808, 0x2b191908, 0x2b080808, 0x08080819, 0x2b080819,
+    0x08081908, 0x2b080819, 0x08082b19, 0x2b080819, 0x08190808, 0x2b080819, 0x0819082b, 0x2b080819,
+    0x08191919, 0x2b080819, 0x08192b08, 0x2b080819, 0x082b0819, 0x2b080819, 0x082b1908, 0x2b080819,
+    0x19080808, 0x2b080819, 0x1908082b, 0x2b080819, 0x19081919, 0x2b080819, 0x19082b08, 0x2b080819,
+    0x19190819, 0x2b080819, 0x19191908, 0x2b080819, 0x2b080819, 0x2b080819, 0x2b081908, 0x2b080819,
+    0x2b190808, 0x2b080819, 0x2b2b2b19, 0x2b080819, 0x08080808, 0x2b08082b, 0x08081919, 0x2b08082b,
+    0x08082b2b, 0x2b08082b, 0x08190819, 0x2b08082b, 0x08191908, 0x2b08082b, 0x19080819, 0x2b08082b,
+    0x19081908, 0x2b08082b, 0x19190808, 0x2b08082b, 0x08080819, 0x2b081908, 0x08081908, 0x2b081908,
+    0x0808192b, 0x2b081908, 0x08082b19, 0x2b081908, 0x08190808, 0x2b081908, 0x0819082b, 0x2b081908,
+    0x08191919, 0x2b081908, 0x08192b08, 0x2b081908, 0x082b0819, 0x2b081908, 0x19080808, 0x2b081908,
+    0x1908082b, 0x2b081908, 0x19081919, 0x2b081908, 0x19082b08, 0x2b081908, 0x19190819, 0x2b081908,
+    0x19191908, 0x2b081908, 0x192b0808, 0x2b081908, 0x2b080819, 0x2b081908, 0x2b081908, 0x2b081908,
+    0x2b190808, 0x2b081908, 0x08080808, 0x2b081919, 0x0808082b, 0x2b081919, 0x08081919, 0x2b081919,
+    0x08082b08, 0x2b081919, 0x08190819, 0x2b081919, 0x08191908, 0x2b081919, 0x082b0808, 0x2b081919,
+    0x19080819, 0x2b081919, 0x19081908, 0x2b081919, 0x19190808, 0x2b081919, 0x2b080808, 0x2b081919,
+    0x2b082b2b, 0x2b081919, 0x08080819, 0x2b08192b, 0x08081908, 0x2b08192b, 0x08190808, 0x2b08192b,
+    0x082b2b19, 0x2b08192b, 0x19080808, 0x2b08192b, 0x08080808, 0x2b082b08, 0x08081919, 0x2b082b08,
+    0x08190819, 0x2b082b08, 0x08191908, 0x2b082b08, 0x19080819, 0x2b082b08, 0x19081908, 0x2b082b08,
+    0x19190808, 0x2b082b08, 0x2b2b082b, 0x2b082b08, 0x08080819, 0x2b082b19, 0x08081908, 0x2b082b19,
+    0x19080808, 0x2b082b19, 0x192b1919, 0x2b082b19, 0x082b082b, 0x2b082b2b, 0x19192b08, 0x2b082b2b,
+    0x19192b2b, 0x2b082b2b, 0x2b08082b, 0x2b082b2b, 0x2b2b082b, 0x2b082b2b, 0x08080819, 0x2b190808,
+    0x08081908, 0x2b190808, 0x08082b19, 0x2b190808, 0x08190808, 0x2b190808, 0x0819082b, 0x2b190808,
+    0x08191919, 0x2b190808, 0x08192b08, 0x2b190808, 0x082b1908, 0x2b190808, 0x19080808, 0x2b190808,
+    0x1908082b, 0x2b190808, 0x19081919, 0x2b190808, 0x19082b08, 0x2b190808, 0x19190819, 0x2b190808,
+    0x19191908, 0x2b190808, 0x192b0808, 0x2b190808, 0x2b080819, 0x2b190808, 0x2b081908, 0x2b190808,
+    0x2b190808, 0x2b190808, 0x08080808, 0x2b190819, 0x08081919, 0x2b190819, 0x08190819, 0x2b190819,
+    0x08191908, 0x2b190819, 0x19080819, 0x2b190819, 0x19081908, 0x2b190819, 0x19190808, 0x2b190819,
+    0x19192b2b, 0x2b190819, 0x08080819, 0x2b19082b, 0x08081908, 0x2b19082b, 0x08190808, 0x2b19082b,
+    0x19080808, 0x2b19082b, 0x2b2b192b, 0x2b19082b, 0x08080808, 0x2b191908, 0x0808082b, 0x2b191908,
+    0x08081919, 0x2b191908, 0x08082b08, 0x2b191908, 0x08190819, 0x2b191908, 0x08191908, 0x2b191908,
+    0x082b0808, 0x2b191908, 0x19080819, 0x2b191908, 0x19081908, 0x2b191908, 0x19190808, 0x2b191908,
+    0x2b080808, 0x2b191908, 0x2b19192b, 0x2b191908, 0x08080819, 0x2b191919, 0x08081908, 0x2b191919,
+    0x08190808, 0x2b191919, 0x19080808, 0x2b191919, 0x2b192b08, 0x2b191919, 0x2b2b0819, 0x2b191919,
+    0x08080808, 0x2b19192b, 0x1908192b, 0x2b19192b, 0x192b1908, 0x2b19192b, 0x08080819, 0x2b192b08,
+    0x08081908, 0x2b192b08, 0x08190808, 0x2b192b08, 0x082b192b, 0x2b192b08, 0x19080808, 0x2b192b08,
+    0x2b2b2b19, 0x2b192b08, 0x08080808, 0x2b192b19, 0x19082b19, 0x2b192b19, 0x1919082b, 0x2b192b19,
+    0x2b190808, 0x2b192b2b, 0x08080808, 0x2b2b0808, 0x08081919, 0x2b2b0808, 0x08082b2b, 0x2b2b0808,
+    0x08191908, 0x2b2b0808, 0x082b082b, 0x2b2b0808, 0x082b2b2b, 0x2b2b0808, 0x19080819, 0x2b2b0808,
+    0x19081908, 0x2b2b0808, 0x19190808, 0x2b2b0808, 0x2b2b082b, 0x2b2b0808, 0x2b2b2b2b, 0x2b2b0808,
+    0x19080808, 0x2b2b0819, 0x192b1919, 0x2b2b0819, 0x0808082b, 0x2b2b082b, 0x08082b2b, 0x2b2b082b,
+    0x082b082b, 0x2b2b082b, 0x082b2b08, 0x2b2b082b, 0x082b2b2b, 0x2b2b082b, 0x2b08082b, 0x2b2b082b,
+    0x2b082b08, 0x2b2b082b, 0x2b082b2b, 0x2b2b082b, 0x2b2b2b08, 0x2b2b082b, 0x08080819, 0x2b2b1908,
+    0x08081908, 0x2b2b1908, 0x08190808, 0x2b2b1908, 0x19080808, 0x2b2b1908, 0x2b082b19, 0x2b2b1908,
+    0x2b2b1908, 0x2b2b1908, 0x08080808, 0x2b2b1919, 0x08192b19, 0x2b2b1919, 0x19190819, 0x2b2b192b,
+    0x08082b2b, 0x2b2b2b08, 0x082b2b08, 0x2b2b2b08, 0x2b2b082b, 0x2b2b2b08, 0x19191908, 0x2b2b2b19,
+    0x2b08192b, 0x2b2b2b19, 0x08082b08, 0x2b2b2b2b, 0x08082b2b, 0x2b2b2b2b, 0x082b0808, 0x2b2b2b2b,
+    0x082b082b, 0x2b2b2b2b, 0x082b2b08, 0x2b2b2b2b, 0x2b082b08, 0x2b2b2b2b, 0x2b2b2b2b, 0x2b2b2b2b
+};

ggml/src/ggml-opencl/kernels/iq2_xs_grid.cl

@@ -0,0 +1,135 @@
+// IQ2_XS codebook, stored as pairs of four packed magnitudes.
+constant uint iq2_xs_grid[1024] = {
+    0x08080808, 0x08080808, 0x0808082b, 0x08080808, 0x08081919, 0x08080808, 0x08082b08, 0x08080808,
+    0x08082b2b, 0x08080808, 0x08190819, 0x08080808, 0x08191908, 0x08080808, 0x0819192b, 0x08080808,
+    0x08192b19, 0x08080808, 0x082b0808, 0x08080808, 0x082b082b, 0x08080808, 0x082b1919, 0x08080808,
+    0x082b2b08, 0x08080808, 0x19080819, 0x08080808, 0x19081908, 0x08080808, 0x1908192b, 0x08080808,
+    0x19082b19, 0x08080808, 0x19190808, 0x08080808, 0x1919082b, 0x08080808, 0x19191919, 0x08080808,
+    0x19192b08, 0x08080808, 0x192b0819, 0x08080808, 0x192b1908, 0x08080808, 0x2b080808, 0x08080808,
+    0x2b08082b, 0x08080808, 0x2b081919, 0x08080808, 0x2b082b08, 0x08080808, 0x2b190819, 0x08080808,
+    0x2b191908, 0x08080808, 0x2b192b19, 0x08080808, 0x2b2b0808, 0x08080808, 0x08080819, 0x08080819,
+    0x08081908, 0x08080819, 0x0808192b, 0x08080819, 0x08082b19, 0x08080819, 0x08190808, 0x08080819,
+    0x0819082b, 0x08080819, 0x08191919, 0x08080819, 0x08192b08, 0x08080819, 0x08192b2b, 0x08080819,
+    0x082b0819, 0x08080819, 0x082b1908, 0x08080819, 0x19080808, 0x08080819, 0x1908082b, 0x08080819,
+    0x19081919, 0x08080819, 0x19082b08, 0x08080819, 0x19190819, 0x08080819, 0x19191908, 0x08080819,
+    0x192b0808, 0x08080819, 0x192b2b08, 0x08080819, 0x2b080819, 0x08080819, 0x2b081908, 0x08080819,
+    0x2b190808, 0x08080819, 0x08080808, 0x0808082b, 0x0808082b, 0x0808082b, 0x08081919, 0x0808082b,
+    0x08082b08, 0x0808082b, 0x08190819, 0x0808082b, 0x08191908, 0x0808082b, 0x082b0808, 0x0808082b,
+    0x19080819, 0x0808082b, 0x19081908, 0x0808082b, 0x19190808, 0x0808082b, 0x19191919, 0x0808082b,
+    0x2b080808, 0x0808082b, 0x2b082b2b, 0x0808082b, 0x08080819, 0x08081908, 0x08081908, 0x08081908,
+    0x0808192b, 0x08081908, 0x08082b19, 0x08081908, 0x08190808, 0x08081908, 0x0819082b, 0x08081908,
+    0x08191919, 0x08081908, 0x08192b08, 0x08081908, 0x082b0819, 0x08081908, 0x082b1908, 0x08081908,
+    0x19080808, 0x08081908, 0x1908082b, 0x08081908, 0x19081919, 0x08081908, 0x19082b08, 0x08081908,
+    0x19190819, 0x08081908, 0x19191908, 0x08081908, 0x1919192b, 0x08081908, 0x192b0808, 0x08081908,
+    0x2b080819, 0x08081908, 0x2b081908, 0x08081908, 0x2b190808, 0x08081908, 0x08080808, 0x08081919,
+    0x0808082b, 0x08081919, 0x08081919, 0x08081919, 0x08082b08, 0x08081919, 0x08190819, 0x08081919,
+    0x08191908, 0x08081919, 0x082b0808, 0x08081919, 0x19080819, 0x08081919, 0x19081908, 0x08081919,
+    0x19190808, 0x08081919, 0x192b0819, 0x08081919, 0x2b080808, 0x08081919, 0x08080819, 0x0808192b,
+    0x08081908, 0x0808192b, 0x08190808, 0x0808192b, 0x082b192b, 0x0808192b, 0x19080808, 0x0808192b,
+    0x1908082b, 0x0808192b, 0x2b081908, 0x0808192b, 0x08080808, 0x08082b08, 0x0808082b, 0x08082b08,
+    0x08081919, 0x08082b08, 0x08082b08, 0x08082b08, 0x08082b2b, 0x08082b08, 0x08190819, 0x08082b08,
+    0x08191908, 0x08082b08, 0x082b0808, 0x08082b08, 0x082b1919, 0x08082b08, 0x19080819, 0x08082b08,
+    0x19081908, 0x08082b08, 0x19190808, 0x08082b08, 0x19192b08, 0x08082b08, 0x2b080808, 0x08082b08,
+    0x2b2b0808, 0x08082b08, 0x2b2b2b2b, 0x08082b08, 0x08080819, 0x08082b19, 0x08081908, 0x08082b19,
+    0x08190808, 0x08082b19, 0x19080808, 0x08082b19, 0x2b080819, 0x08082b19, 0x2b082b19, 0x08082b19,
+    0x08080808, 0x08082b2b, 0x082b0808, 0x08082b2b, 0x082b2b08, 0x08082b2b, 0x2b19192b, 0x08082b2b,
+    0x2b2b0808, 0x08082b2b, 0x08080819, 0x08190808, 0x08081908, 0x08190808, 0x0808192b, 0x08190808,
+    0x08082b19, 0x08190808, 0x08190808, 0x08190808, 0x0819082b, 0x08190808, 0x08191919, 0x08190808,
+    0x08192b08, 0x08190808, 0x082b0819, 0x08190808, 0x082b1908, 0x08190808, 0x19080808, 0x08190808,
+    0x1908082b, 0x08190808, 0x19081919, 0x08190808, 0x19082b08, 0x08190808, 0x19190819, 0x08190808,
+    0x19191908, 0x08190808, 0x192b0808, 0x08190808, 0x192b2b2b, 0x08190808, 0x2b080819, 0x08190808,
+    0x2b081908, 0x08190808, 0x2b190808, 0x08190808, 0x08080808, 0x08190819, 0x0808082b, 0x08190819,
+    0x08081919, 0x08190819, 0x08082b08, 0x08190819, 0x08190819, 0x08190819, 0x08191908, 0x08190819,
+    0x082b0808, 0x08190819, 0x19080819, 0x08190819, 0x19081908, 0x08190819, 0x19190808, 0x08190819,
+    0x2b080808, 0x08190819, 0x2b191908, 0x08190819, 0x2b19192b, 0x08190819, 0x08080819, 0x0819082b,
+    0x08081908, 0x0819082b, 0x0808192b, 0x0819082b, 0x08190808, 0x0819082b, 0x19080808, 0x0819082b,
+    0x192b0808, 0x0819082b, 0x08080808, 0x08191908, 0x0808082b, 0x08191908, 0x08081919, 0x08191908,
+    0x08082b08, 0x08191908, 0x08190819, 0x08191908, 0x08191908, 0x08191908, 0x082b0808, 0x08191908,
+    0x19080819, 0x08191908, 0x19081908, 0x08191908, 0x19082b19, 0x08191908, 0x19190808, 0x08191908,
+    0x192b1908, 0x08191908, 0x2b080808, 0x08191908, 0x08080819, 0x08191919, 0x08081908, 0x08191919,
+    0x08190808, 0x08191919, 0x19080808, 0x08191919, 0x08080808, 0x0819192b, 0x08191908, 0x0819192b,
+    0x19082b19, 0x0819192b, 0x08080819, 0x08192b08, 0x08081908, 0x08192b08, 0x08190808, 0x08192b08,
+    0x0819082b, 0x08192b08, 0x19080808, 0x08192b08, 0x19191908, 0x08192b08, 0x2b08192b, 0x08192b08,
+    0x08080808, 0x08192b19, 0x08081919, 0x08192b19, 0x192b192b, 0x08192b19, 0x19190819, 0x08192b2b,
+    0x2b2b2b19, 0x08192b2b, 0x08080808, 0x082b0808, 0x0808082b, 0x082b0808, 0x08081919, 0x082b0808,
+    0x08082b08, 0x082b0808, 0x08082b2b, 0x082b0808, 0x08190819, 0x082b0808, 0x08191908, 0x082b0808,
+    0x082b0808, 0x082b0808, 0x19080819, 0x082b0808, 0x19081908, 0x082b0808, 0x19190808, 0x082b0808,
+    0x2b080808, 0x082b0808, 0x2b2b0808, 0x082b0808, 0x08080819, 0x082b0819, 0x08081908, 0x082b0819,
+    0x08190808, 0x082b0819, 0x19080808, 0x082b0819, 0x19082b08, 0x082b0819, 0x192b1919, 0x082b0819,
+    0x08080808, 0x082b082b, 0x082b082b, 0x082b082b, 0x2b080808, 0x082b082b, 0x2b2b2b08, 0x082b082b,
+    0x08080819, 0x082b1908, 0x08081908, 0x082b1908, 0x08190808, 0x082b1908, 0x082b2b19, 0x082b1908,
+    0x19080808, 0x082b1908, 0x08080808, 0x082b1919, 0x19080819, 0x082b1919, 0x1919082b, 0x082b1919,
+    0x2b192b19, 0x082b1919, 0x08080819, 0x082b192b, 0x08192b2b, 0x082b192b, 0x2b2b192b, 0x082b192b,
+    0x08080808, 0x082b2b08, 0x08082b08, 0x082b2b08, 0x08082b2b, 0x082b2b08, 0x082b0808, 0x082b2b08,
+    0x19191919, 0x082b2b08, 0x2b082b08, 0x082b2b08, 0x2b2b082b, 0x082b2b08, 0x192b2b08, 0x082b2b19,
+    0x2b190808, 0x082b2b19, 0x08082b08, 0x082b2b2b, 0x082b0808, 0x082b2b2b, 0x2b08082b, 0x082b2b2b,
+    0x2b082b08, 0x082b2b2b, 0x2b082b2b, 0x082b2b2b, 0x08080819, 0x19080808, 0x08081908, 0x19080808,
+    0x0808192b, 0x19080808, 0x08082b19, 0x19080808, 0x08190808, 0x19080808, 0x0819082b, 0x19080808,
+    0x08191919, 0x19080808, 0x08192b08, 0x19080808, 0x082b0819, 0x19080808, 0x082b1908, 0x19080808,
+    0x19080808, 0x19080808, 0x1908082b, 0x19080808, 0x19081919, 0x19080808, 0x19082b08, 0x19080808,
+    0x19082b2b, 0x19080808, 0x19190819, 0x19080808, 0x19191908, 0x19080808, 0x192b0808, 0x19080808,
+    0x192b1919, 0x19080808, 0x2b080819, 0x19080808, 0x2b081908, 0x19080808, 0x2b190808, 0x19080808,
+    0x08080808, 0x19080819, 0x0808082b, 0x19080819, 0x08081919, 0x19080819, 0x08082b08, 0x19080819,
+    0x08190819, 0x19080819, 0x08191908, 0x19080819, 0x082b0808, 0x19080819, 0x19080819, 0x19080819,
+    0x19081908, 0x19080819, 0x19190808, 0x19080819, 0x2b080808, 0x19080819, 0x2b081919, 0x19080819,
+    0x2b2b082b, 0x19080819, 0x08080819, 0x1908082b, 0x08081908, 0x1908082b, 0x08190808, 0x1908082b,
+    0x0819082b, 0x1908082b, 0x082b2b19, 0x1908082b, 0x19080808, 0x1908082b, 0x08080808, 0x19081908,
+    0x0808082b, 0x19081908, 0x08081919, 0x19081908, 0x08082b08, 0x19081908, 0x08190819, 0x19081908,
+    0x08191908, 0x19081908, 0x08192b19, 0x19081908, 0x082b0808, 0x19081908, 0x19080819, 0x19081908,
+    0x19081908, 0x19081908, 0x19190808, 0x19081908, 0x2b080808, 0x19081908, 0x2b191908, 0x19081908,
+    0x08080819, 0x19081919, 0x08081908, 0x19081919, 0x08190808, 0x19081919, 0x082b1908, 0x19081919,
+    0x19080808, 0x19081919, 0x2b192b2b, 0x19081919, 0x08080808, 0x1908192b, 0x08082b2b, 0x1908192b,
+    0x19081908, 0x1908192b, 0x19190808, 0x1908192b, 0x08080819, 0x19082b08, 0x08081908, 0x19082b08,
+    0x08190808, 0x19082b08, 0x19080808, 0x19082b08, 0x19081919, 0x19082b08, 0x19191908, 0x19082b08,
+    0x192b082b, 0x19082b08, 0x08080808, 0x19082b19, 0x08190819, 0x19082b19, 0x19081908, 0x19082b19,
+    0x19190808, 0x19082b19, 0x192b2b19, 0x19082b19, 0x08081908, 0x19082b2b, 0x08080808, 0x19190808,
+    0x0808082b, 0x19190808, 0x08081919, 0x19190808, 0x08082b08, 0x19190808, 0x08190819, 0x19190808,
+    0x08191908, 0x19190808, 0x082b0808, 0x19190808, 0x082b2b08, 0x19190808, 0x19080819, 0x19190808,
+    0x19081908, 0x19190808, 0x19190808, 0x19190808, 0x2b080808, 0x19190808, 0x08080819, 0x19190819,
+    0x08081908, 0x19190819, 0x08190808, 0x19190819, 0x08191919, 0x19190819, 0x19080808, 0x19190819,
+    0x1908082b, 0x19190819, 0x08080808, 0x1919082b, 0x19081908, 0x1919082b, 0x2b2b2b2b, 0x1919082b,
+    0x08080819, 0x19191908, 0x08081908, 0x19191908, 0x08190808, 0x19191908, 0x082b0819, 0x19191908,
+    0x19080808, 0x19191908, 0x192b0808, 0x19191908, 0x2b080819, 0x19191908, 0x2b2b0819, 0x19191908,
+    0x08080808, 0x19191919, 0x08082b08, 0x19191919, 0x2b080808, 0x19191919, 0x2b082b08, 0x19191919,
+    0x082b0819, 0x1919192b, 0x192b2b08, 0x1919192b, 0x2b2b0819, 0x1919192b, 0x08080808, 0x19192b08,
+    0x08191908, 0x19192b08, 0x19080819, 0x19192b08, 0x19190808, 0x19192b08, 0x2b192b19, 0x19192b08,
+    0x08192b2b, 0x19192b19, 0x19080808, 0x19192b19, 0x1908082b, 0x19192b19, 0x2b081919, 0x19192b2b,
+    0x08080819, 0x192b0808, 0x08081908, 0x192b0808, 0x08190808, 0x192b0808, 0x19080808, 0x192b0808,
+    0x19191908, 0x192b0808, 0x192b082b, 0x192b0808, 0x2b08192b, 0x192b0808, 0x2b2b2b19, 0x192b0808,
+    0x08080808, 0x192b0819, 0x082b1908, 0x192b082b, 0x19082b2b, 0x192b082b, 0x2b19082b, 0x192b082b,
+    0x08080808, 0x192b1908, 0x0819192b, 0x192b1908, 0x08190808, 0x192b1919, 0x19080808, 0x192b1919,
+    0x19081919, 0x192b1919, 0x2b2b1908, 0x192b1919, 0x08080819, 0x192b2b08, 0x192b2b2b, 0x192b2b08,
+    0x082b1919, 0x192b2b19, 0x0808192b, 0x192b2b2b, 0x19191908, 0x192b2b2b, 0x192b082b, 0x192b2b2b,
+    0x08080808, 0x2b080808, 0x0808082b, 0x2b080808, 0x08081919, 0x2b080808, 0x08082b08, 0x2b080808,
+    0x08190819, 0x2b080808, 0x08191908, 0x2b080808, 0x082b0808, 0x2b080808, 0x082b2b2b, 0x2b080808,
+    0x19080819, 0x2b080808, 0x19081908, 0x2b080808, 0x19190808, 0x2b080808, 0x2b080808, 0x2b080808,
+    0x2b08082b, 0x2b080808, 0x2b2b2b08, 0x2b080808, 0x2b2b2b2b, 0x2b080808, 0x08080819, 0x2b080819,
+    0x08081908, 0x2b080819, 0x0808192b, 0x2b080819, 0x08190808, 0x2b080819, 0x19080808, 0x2b080819,
+    0x19190819, 0x2b080819, 0x19192b19, 0x2b080819, 0x08080808, 0x2b08082b, 0x082b0808, 0x2b08082b,
+    0x2b080808, 0x2b08082b, 0x2b08082b, 0x2b08082b, 0x2b2b0808, 0x2b08082b, 0x2b2b2b08, 0x2b08082b,
+    0x08080819, 0x2b081908, 0x08081908, 0x2b081908, 0x08190808, 0x2b081908, 0x0819082b, 0x2b081908,
+    0x08191919, 0x2b081908, 0x19080808, 0x2b081908, 0x192b0808, 0x2b081908, 0x2b082b19, 0x2b081908,
+    0x08080808, 0x2b081919, 0x19081908, 0x2b081919, 0x2b2b1919, 0x2b081919, 0x08192b08, 0x2b08192b,
+    0x192b2b2b, 0x2b08192b, 0x08080808, 0x2b082b08, 0x08082b08, 0x2b082b08, 0x082b1919, 0x2b082b08,
+    0x19192b2b, 0x2b082b08, 0x2b080808, 0x2b082b08, 0x2b08082b, 0x2b082b08, 0x2b2b2b08, 0x2b082b08,
+    0x0808192b, 0x2b082b19, 0x082b082b, 0x2b082b2b, 0x2b080808, 0x2b082b2b, 0x2b082b08, 0x2b082b2b,
+    0x2b19192b, 0x2b082b2b, 0x2b2b2b08, 0x2b082b2b, 0x08080819, 0x2b190808, 0x08081908, 0x2b190808,
+    0x08190808, 0x2b190808, 0x19080808, 0x2b190808, 0x1919192b, 0x2b190808, 0x2b081908, 0x2b190808,
+    0x08080808, 0x2b190819, 0x082b082b, 0x2b190819, 0x192b1908, 0x2b190819, 0x1919192b, 0x2b19082b,
+    0x2b082b19, 0x2b19082b, 0x08080808, 0x2b191908, 0x08081919, 0x2b191908, 0x19081908, 0x2b191908,
+    0x19190808, 0x2b191908, 0x19192b08, 0x2b191908, 0x082b2b19, 0x2b191919, 0x2b190808, 0x2b191919,
+    0x2b19082b, 0x2b191919, 0x19080819, 0x2b19192b, 0x19190819, 0x2b192b08, 0x2b2b192b, 0x2b192b08,
+    0x19082b19, 0x2b192b19, 0x08191919, 0x2b192b2b, 0x192b0808, 0x2b192b2b, 0x08080808, 0x2b2b0808,
+    0x0808082b, 0x2b2b0808, 0x08082b08, 0x2b2b0808, 0x08082b2b, 0x2b2b0808, 0x082b0808, 0x2b2b0808,
+    0x082b2b2b, 0x2b2b0808, 0x2b2b0808, 0x2b2b0808, 0x19190819, 0x2b2b0819, 0x19192b19, 0x2b2b0819,
+    0x2b2b192b, 0x2b2b0819, 0x08080808, 0x2b2b082b, 0x0808082b, 0x2b2b082b, 0x08082b08, 0x2b2b082b,
+    0x082b2b2b, 0x2b2b082b, 0x2b080808, 0x2b2b082b, 0x2b2b0808, 0x2b2b082b, 0x19080808, 0x2b2b1908,
+    0x2b191919, 0x2b2b1908, 0x192b1919, 0x2b2b192b, 0x2b192b08, 0x2b2b192b, 0x08082b2b, 0x2b2b2b08,
+    0x082b0808, 0x2b2b2b08, 0x082b082b, 0x2b2b2b08, 0x082b2b08, 0x2b2b2b08, 0x2b2b0808, 0x2b2b2b08,
+    0x2b2b2b08, 0x2b2b2b08, 0x08081908, 0x2b2b2b19, 0x2b081908, 0x2b2b2b19, 0x2b08192b, 0x2b2b2b19,
+    0x082b2b08, 0x2b2b2b2b, 0x082b2b2b, 0x2b2b2b2b, 0x2b190819, 0x2b2b2b2b, 0x2b2b2b2b, 0x2b2b2b2b
+};
+
+inline uint iq2_xs_signs(uint code) {
+    return code | ((popcount(code) & 1u) << 7);
+}

ggml/src/ggml-opencl/kernels/iq2_xxs_grid.cl

@@ -0,0 +1,71 @@
+// IQ2_XXS codebook, stored as pairs of four packed magnitudes.
+constant uint iq2_xxs_grid[512] = {
+    0x08080808, 0x08080808, 0x0808082b, 0x08080808, 0x08081919, 0x08080808, 0x08082b08, 0x08080808,
+    0x08082b2b, 0x08080808, 0x08190819, 0x08080808, 0x08191908, 0x08080808, 0x082b0808, 0x08080808,
+    0x082b082b, 0x08080808, 0x082b2b08, 0x08080808, 0x082b2b2b, 0x08080808, 0x19080819, 0x08080808,
+    0x19081908, 0x08080808, 0x19190808, 0x08080808, 0x19192b08, 0x08080808, 0x192b0819, 0x08080808,
+    0x192b1908, 0x08080808, 0x2b080808, 0x08080808, 0x2b08082b, 0x08080808, 0x2b082b2b, 0x08080808,
+    0x2b2b082b, 0x08080808, 0x08080819, 0x08080819, 0x08081908, 0x08080819, 0x08190808, 0x08080819,
+    0x08191919, 0x08080819, 0x19080808, 0x08080819, 0x2b081908, 0x08080819, 0x2b192b08, 0x08080819,
+    0x08080808, 0x0808082b, 0x0808082b, 0x0808082b, 0x082b082b, 0x0808082b, 0x2b08082b, 0x0808082b,
+    0x08080819, 0x08081908, 0x08081908, 0x08081908, 0x08190808, 0x08081908, 0x082b0819, 0x08081908,
+    0x082b1908, 0x08081908, 0x19080808, 0x08081908, 0x1908082b, 0x08081908, 0x19082b08, 0x08081908,
+    0x192b0808, 0x08081908, 0x2b080819, 0x08081908, 0x2b081908, 0x08081908, 0x2b190808, 0x08081908,
+    0x2b2b1908, 0x08081908, 0x08080808, 0x08081919, 0x0808082b, 0x08081919, 0x08082b08, 0x08081919,
+    0x082b0808, 0x08081919, 0x1908192b, 0x08081919, 0x192b2b19, 0x08081919, 0x2b080808, 0x08081919,
+    0x2b190819, 0x08081919, 0x08082b19, 0x0808192b, 0x08190808, 0x0808192b, 0x19080808, 0x0808192b,
+    0x2b081908, 0x0808192b, 0x2b2b1908, 0x0808192b, 0x08080808, 0x08082b08, 0x08081919, 0x08082b08,
+    0x08082b08, 0x08082b08, 0x08191908, 0x08082b08, 0x082b2b08, 0x08082b08, 0x19080819, 0x08082b08,
+    0x19081908, 0x08082b08, 0x19190808, 0x08082b08, 0x1919082b, 0x08082b08, 0x2b082b08, 0x08082b08,
+    0x08081908, 0x08082b19, 0x19080808, 0x08082b19, 0x0808082b, 0x08082b2b, 0x08191908, 0x08082b2b,
+    0x08080819, 0x08190808, 0x08081908, 0x08190808, 0x08190808, 0x08190808, 0x082b0819, 0x08190808,
+    0x19080808, 0x08190808, 0x192b0808, 0x08190808, 0x2b081908, 0x08190808, 0x2b190808, 0x08190808,
+    0x2b191919, 0x08190808, 0x08080808, 0x08190819, 0x08082b08, 0x08190819, 0x082b0808, 0x08190819,
+    0x19190808, 0x08190819, 0x19192b2b, 0x08190819, 0x2b080808, 0x08190819, 0x082b1908, 0x0819082b,
+    0x19081919, 0x0819082b, 0x08080808, 0x08191908, 0x08082b08, 0x08191908, 0x082b0808, 0x08191908,
+    0x082b1919, 0x08191908, 0x19082b19, 0x08191908, 0x2b080808, 0x08191908, 0x08192b08, 0x08191919,
+    0x192b082b, 0x08191919, 0x08080808, 0x0819192b, 0x0819192b, 0x0819192b, 0x08080819, 0x08192b08,
+    0x08081908, 0x08192b08, 0x08190808, 0x08192b08, 0x19080808, 0x08192b08, 0x2b080819, 0x08192b08,
+    0x08080808, 0x08192b19, 0x08081919, 0x08192b19, 0x2b2b0808, 0x08192b19, 0x19190819, 0x08192b2b,
+    0x08080808, 0x082b0808, 0x0808082b, 0x082b0808, 0x08082b2b, 0x082b0808, 0x19081908, 0x082b0808,
+    0x192b0819, 0x082b0808, 0x2b080808, 0x082b0808, 0x2b08082b, 0x082b0808, 0x082b2b19, 0x082b0819,
+    0x19082b08, 0x082b0819, 0x08080808, 0x082b082b, 0x0808082b, 0x082b082b, 0x08080819, 0x082b1908,
+    0x08081908, 0x082b1908, 0x08190808, 0x082b1908, 0x19080808, 0x082b1908, 0x1919192b, 0x082b1908,
+    0x08080808, 0x082b1919, 0x19080819, 0x082b1919, 0x192b1908, 0x082b1919, 0x2b190808, 0x082b192b,
+    0x08082b08, 0x082b2b08, 0x082b0808, 0x082b2b08, 0x2b191908, 0x082b2b08, 0x19081908, 0x082b2b2b,
+    0x08080819, 0x19080808, 0x08081908, 0x19080808, 0x08190808, 0x19080808, 0x08192b08, 0x19080808,
+    0x082b0819, 0x19080808, 0x082b1908, 0x19080808, 0x19080808, 0x19080808, 0x19082b08, 0x19080808,
+    0x1919192b, 0x19080808, 0x192b0808, 0x19080808, 0x2b080819, 0x19080808, 0x2b081908, 0x19080808,
+    0x2b190808, 0x19080808, 0x08080808, 0x19080819, 0x082b0808, 0x19080819, 0x192b0819, 0x19080819,
+    0x2b080808, 0x19080819, 0x2b081919, 0x19080819, 0x08080819, 0x1908082b, 0x08190808, 0x1908082b,
+    0x19082b08, 0x1908082b, 0x1919192b, 0x1908082b, 0x192b2b08, 0x1908082b, 0x08080808, 0x19081908,
+    0x08082b08, 0x19081908, 0x082b0808, 0x19081908, 0x2b080808, 0x19081908, 0x2b192b19, 0x19081908,
+    0x0819082b, 0x19081919, 0x082b1908, 0x19081919, 0x08080808, 0x1908192b, 0x08080819, 0x19082b08,
+    0x08081908, 0x19082b08, 0x08190808, 0x19082b08, 0x19080808, 0x19082b08, 0x19081919, 0x19082b08,
+    0x08080808, 0x19082b19, 0x19192b08, 0x19082b19, 0x192b0819, 0x19082b19, 0x2b08082b, 0x19082b19,
+    0x19081919, 0x19082b2b, 0x2b190808, 0x19082b2b, 0x08080808, 0x19190808, 0x08082b08, 0x19190808,
+    0x08190819, 0x19190808, 0x08192b19, 0x19190808, 0x082b0808, 0x19190808, 0x2b080808, 0x19190808,
+    0x2b082b08, 0x19190808, 0x08081908, 0x19190819, 0x1908082b, 0x19190819, 0x2b2b1908, 0x19190819,
+    0x2b190819, 0x1919082b, 0x2b190808, 0x19191908, 0x2b19082b, 0x19191908, 0x08082b2b, 0x19191919,
+    0x08080819, 0x1919192b, 0x19191908, 0x1919192b, 0x08080808, 0x19192b08, 0x08190819, 0x19192b08,
+    0x08192b19, 0x19192b08, 0x192b1908, 0x19192b08, 0x19080808, 0x19192b19, 0x08082b08, 0x19192b2b,
+    0x08081908, 0x192b0808, 0x08190808, 0x192b0808, 0x19080808, 0x192b0808, 0x192b2b08, 0x192b0808,
+    0x08080808, 0x192b0819, 0x19191919, 0x192b0819, 0x08192b08, 0x192b082b, 0x192b0808, 0x192b082b,
+    0x08080808, 0x192b1908, 0x08081919, 0x192b1908, 0x08190808, 0x192b1919, 0x0819082b, 0x192b1919,
+    0x2b081908, 0x192b1919, 0x1908082b, 0x192b2b08, 0x08080808, 0x2b080808, 0x0808082b, 0x2b080808,
+    0x08082b2b, 0x2b080808, 0x19080819, 0x2b080808, 0x2b08082b, 0x2b080808, 0x08081908, 0x2b080819,
+    0x08192b08, 0x2b080819, 0x19080808, 0x2b080819, 0x08190819, 0x2b08082b, 0x08080819, 0x2b081908,
+    0x08081908, 0x2b081908, 0x08190808, 0x2b081908, 0x08191919, 0x2b081908, 0x19080808, 0x2b081908,
+    0x192b0808, 0x2b081908, 0x08080808, 0x2b081919, 0x1908192b, 0x2b081919, 0x2b191908, 0x2b081919,
+    0x08082b19, 0x2b08192b, 0x19080808, 0x2b08192b, 0x192b0808, 0x2b08192b, 0x0808082b, 0x2b082b08,
+    0x08081908, 0x2b082b19, 0x08190819, 0x2b082b2b, 0x08081908, 0x2b190808, 0x08190808, 0x2b190808,
+    0x082b1908, 0x2b190808, 0x19080808, 0x2b190808, 0x2b2b0819, 0x2b190808, 0x0819192b, 0x2b190819,
+    0x2b080808, 0x2b190819, 0x19081919, 0x2b19082b, 0x08080808, 0x2b191908, 0x082b082b, 0x2b191908,
+    0x19081908, 0x2b191908, 0x19190819, 0x2b191919, 0x2b080819, 0x2b192b08, 0x082b0808, 0x2b192b19,
+    0x0808082b, 0x2b2b0808, 0x19190808, 0x2b2b0808, 0x2b081919, 0x2b2b0808, 0x08082b19, 0x2b2b0819,
+    0x08080808, 0x2b2b082b, 0x08192b08, 0x2b2b1908, 0x19190808, 0x2b2b2b08, 0x08081908, 0x2b2b2b19
+};
+
+inline uint iq2_xxs_signs(uint code) {
+    return code | ((popcount(code) & 1u) << 7);
+}

ggml/src/ggml-opencl/kernels/mul_mm_iq2_s_f32_l4_lm.cl

@@ -0,0 +1,191 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_required_subgroup_size
+#define INTEL_GPU 1
+#endif
+
+#define QK_K 256
+
+struct block_iq2_s {
+    half d;
+    uchar qs[QK_K/4];
+    uchar qh[QK_K/32];
+    uchar scales[QK_K/32];
+};
+
+#define LOAD_VEC_A 4
+#define LOAD_VEC_B 4
+
+#define BM 64
+#define BN 64
+#define BK 32
+#ifdef INTEL_GPU
+#define TM 8
+#define TN 8
+#else
+#define TM 4
+#define TN 8
+#endif
+
+kernel void kernel_mul_mm_iq2_s_f32_l4_lm(
+    global uchar * src0_qs,
+    global uchar * src0_sg,
+    global uchar * src0_qh,
+    global uchar * src0_sc,
+    global half * src0_d,
+    ulong offset0,
+    global float4 * src1,
+    ulong offset1,
+    global float  * dst,
+    ulong offsetd,
+
+    int ne00,
+    int ne01,
+    int ne02,
+    int ne11,
+    int ne12,
+
+    int stride_a,
+    int stride_b,
+    int stride_d,
+
+    int batch_stride_a,
+    int batch_stride_b,
+    int batch_stride_d,
+
+    int r2,
+    int r3
+) {
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst  = (global float *)((global char*)dst  + offsetd);
+
+    local float buf_a[BM * BK];
+    local float buf_b[BN * BK];
+
+    const int batch_idx = get_global_id(2);
+
+    const int i13 = batch_idx / ne12;
+    const int i12 = batch_idx % ne12;
+
+    const int i03 = i13 / r3;
+    const int i02 = i12 / r2;
+
+    const int batch_idx_a = i03 * ne02 + i02;
+
+    const int ir = get_group_id(0);
+    const int ic = get_group_id(1);
+
+    const int tid = get_local_id(0);
+    const int th_r  = tid % (BM / TM);
+    const int th_c  = tid / (BM / TM);
+
+    const int loadr_a = get_local_id(0) % (BK / LOAD_VEC_A);
+    const int loadc_a = get_local_id(0) / (BK / LOAD_VEC_A);
+    const int loadr_b = get_local_id(0) % (BK / LOAD_VEC_B);
+    const int loadc_b = get_local_id(0) / (BK / LOAD_VEC_B);
+
+    const int loadstride_a = get_local_size(0) * LOAD_VEC_A / BK;
+    const int loadstride_b = get_local_size(0) * LOAD_VEC_B / BK;
+
+    // pos_a counts elements, not blocks - one iq2_s super block holds QK_K of them
+    int pos_a = batch_idx_a * batch_stride_a + ir * BM * stride_a;
+    int pos_b = (batch_idx   * batch_stride_b + ic * BN * stride_b) / LOAD_VEC_B;
+
+    float sums[TM * TN];
+    float cache_a[TM];
+    float cache_b[TN];
+
+    for (int i = 0; i < TM * TN; i++) {
+        sums[i] = 0.0f;
+    }
+
+    for (int block = 0; block < ne00; block += BK) {
+        for (int l = 0; l < BM; l += loadstride_a) {
+            if (ir*BM + loadc_a + l < ne01) {
+                int idx = pos_a + (loadc_a + l) * stride_a + loadr_a * LOAD_VEC_A;
+                int ib  = idx / QK_K;
+                ib += offset0/sizeof(struct block_iq2_s);
+                int e   = idx % QK_K;
+
+                int ib32 = e >> 5;
+                int rem  = e & 31;
+                int lg    = rem >> 3;
+                int j0   = rem & 7;
+
+                uchar scb = src0_sc[8*ib + ib32];
+                float db  = (float)src0_d[ib] * (0.5f + (float)((lg < 2) ? (scb & 0xf) : (scb >> 4))) * 0.25f;
+
+                uint  gi = (uint)src0_qs[32*ib + 4*ib32 + lg] | (((uint)src0_qh[8*ib + ib32] << (8-2*lg)) & 0x300);
+                uchar sg = src0_sg[32*ib + 4*ib32 + lg];
+                uint  g  = iq2_s_grid[2*gi + (j0 >> 2)];
+
+                float4 v1;
+                v1.s0 = db * (float)((g >>  0) & 0xFF) * ((sg & (1 << (j0+0))) ? -1.f : 1.f);
+                v1.s1 = db * (float)((g >>  8) & 0xFF) * ((sg & (1 << (j0+1))) ? -1.f : 1.f);
+                v1.s2 = db * (float)((g >> 16) & 0xFF) * ((sg & (1 << (j0+2))) ? -1.f : 1.f);
+                v1.s3 = db * (float)((g >> 24) & 0xFF) * ((sg & (1 << (j0+3))) ? -1.f : 1.f);
+
+                buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = v1.s0;
+                buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = v1.s1;
+                buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = v1.s2;
+                buf_a[(loadr_a * LOAD_VEC_A + 3) * BM + loadc_a + l] = v1.s3;
+            } else {
+                buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * LOAD_VEC_A + 3) * BM + loadc_a + l] = 0.0f;
+            }
+        }
+
+        for (int l = 0; l < BN; l += loadstride_b) {
+            if (ic*BN + loadc_b + l < ne11) {
+                int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b;
+                buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0;
+                buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1;
+                buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = src1[idx].s2;
+                buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = src1[idx].s3;
+            } else {
+                buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = 0.0f;
+            }
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        pos_a += BK;
+        pos_b += BK / LOAD_VEC_B;
+
+        for (int i = 0; i < BK; i++) {
+            for (int j = 0; j < TM; j++) {
+                cache_a[j] = buf_a[(i) * BM + th_r * TM + j];
+            }
+
+            for (int j = 0; j < TN; j++) {
+                cache_b[j] = buf_b[(i) * BN + th_c * TN + j];
+            }
+
+            for (int cc = 0; cc < TN; cc++) {
+                for (int cr = 0; cr < TM; cr++) {
+                    const int sums_idx = cc*TM + cr;
+                    sums[sums_idx] = mad(cache_a[cr], cache_b[cc], sums[sums_idx]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const int dr = ir * BM + th_r * TM;
+    const int dc = ic * BN + th_c * TN;
+
+    const ulong offsets = batch_idx * batch_stride_d;
+
+    for (int cc = 0; cc < TN; cc++) {
+        for (int cr = 0; cr < TM; cr++) {
+            if (dr + cr < ne01 && dc + cc < ne11) {
+                dst[offsets + (dc + cc) * stride_d + dr + cr] = sums[cc * TM + cr];
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mm_iq2_xs_f32_l4_lm.cl

@@ -0,0 +1,189 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_required_subgroup_size
+#define INTEL_GPU 1
+#endif
+
+#define QK_K 256
+
+struct block_iq2_xs {
+    half d;
+    ushort qs[QK_K/8];
+    uchar scales[QK_K/32];
+};
+
+#define LOAD_VEC_A 4
+#define LOAD_VEC_B 4
+
+#define BM 64
+#define BN 64
+#define BK 32
+#ifdef INTEL_GPU
+#define TM 8
+#define TN 8
+#else
+#define TM 4
+#define TN 8
+#endif
+
+kernel void kernel_mul_mm_iq2_xs_f32_l4_lm(
+    global ushort * src0_qs,
+    global uchar * src0_sc,
+    global half * src0_d,
+    ulong offset0,
+    global float4 * src1,
+    ulong offset1,
+    global float  * dst,
+    ulong offsetd,
+
+    int ne00,
+    int ne01,
+    int ne02,
+    int ne11,
+    int ne12,
+
+    int stride_a,
+    int stride_b,
+    int stride_d,
+
+    int batch_stride_a,
+    int batch_stride_b,
+    int batch_stride_d,
+
+    int r2,
+    int r3
+) {
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst  = (global float *)((global char*)dst  + offsetd);
+
+    local float buf_a[BM * BK];
+    local float buf_b[BN * BK];
+
+    const int batch_idx = get_global_id(2);
+
+    const int i13 = batch_idx / ne12;
+    const int i12 = batch_idx % ne12;
+
+    const int i03 = i13 / r3;
+    const int i02 = i12 / r2;
+
+    const int batch_idx_a = i03 * ne02 + i02;
+
+    const int ir = get_group_id(0);
+    const int ic = get_group_id(1);
+
+    const int tid = get_local_id(0);
+    const int th_r  = tid % (BM / TM);
+    const int th_c  = tid / (BM / TM);
+
+    const int loadr_a = get_local_id(0) % (BK / LOAD_VEC_A);
+    const int loadc_a = get_local_id(0) / (BK / LOAD_VEC_A);
+    const int loadr_b = get_local_id(0) % (BK / LOAD_VEC_B);
+    const int loadc_b = get_local_id(0) / (BK / LOAD_VEC_B);
+
+    const int loadstride_a = get_local_size(0) * LOAD_VEC_A / BK;
+    const int loadstride_b = get_local_size(0) * LOAD_VEC_B / BK;
+
+    // pos_a counts elements, not blocks - one iq2_xs super block holds QK_K of them
+    int pos_a = batch_idx_a * batch_stride_a + ir * BM * stride_a;
+    int pos_b = (batch_idx   * batch_stride_b + ic * BN * stride_b) / LOAD_VEC_B;
+
+    float sums[TM * TN];
+    float cache_a[TM];
+    float cache_b[TN];
+
+    for (int i = 0; i < TM * TN; i++) {
+        sums[i] = 0.0f;
+    }
+
+    for (int block = 0; block < ne00; block += BK) {
+        for (int l = 0; l < BM; l += loadstride_a) {
+            if (ir*BM + loadc_a + l < ne01) {
+                int idx = pos_a + (loadc_a + l) * stride_a + loadr_a * LOAD_VEC_A;
+                int ib  = idx / QK_K;
+                ib += offset0/sizeof(struct block_iq2_xs);
+                int e   = idx % QK_K;
+
+                int ib32 = e >> 5;
+                int rem  = e & 31;
+                int lg    = rem >> 3;
+                int j0   = rem & 7;
+
+                uchar scb = src0_sc[8*ib + ib32];
+                float db  = (float)src0_d[ib] * (0.5f + (float)((lg < 2) ? (scb & 0xf) : (scb >> 4))) * 0.25f;
+
+                uint q = src0_qs[32*ib + 4*ib32 + lg];
+                uint gi = q & 511u;
+                uint sg = iq2_xs_signs(q >> 9);
+                uint  g  = iq2_xs_grid[2*gi + (j0 >> 2)];
+
+                float4 v1;
+                v1.s0 = db * (float)((g >>  0) & 0xFF) * ((sg & (1 << (j0+0))) ? -1.f : 1.f);
+                v1.s1 = db * (float)((g >>  8) & 0xFF) * ((sg & (1 << (j0+1))) ? -1.f : 1.f);
+                v1.s2 = db * (float)((g >> 16) & 0xFF) * ((sg & (1 << (j0+2))) ? -1.f : 1.f);
+                v1.s3 = db * (float)((g >> 24) & 0xFF) * ((sg & (1 << (j0+3))) ? -1.f : 1.f);
+
+                buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = v1.s0;
+                buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = v1.s1;
+                buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = v1.s2;
+                buf_a[(loadr_a * LOAD_VEC_A + 3) * BM + loadc_a + l] = v1.s3;
+            } else {
+                buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * LOAD_VEC_A + 3) * BM + loadc_a + l] = 0.0f;
+            }
+        }
+
+        for (int l = 0; l < BN; l += loadstride_b) {
+            if (ic*BN + loadc_b + l < ne11) {
+                int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b;
+                buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0;
+                buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1;
+                buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = src1[idx].s2;
+                buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = src1[idx].s3;
+            } else {
+                buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = 0.0f;
+            }
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        pos_a += BK;
+        pos_b += BK / LOAD_VEC_B;
+
+        for (int i = 0; i < BK; i++) {
+            for (int j = 0; j < TM; j++) {
+                cache_a[j] = buf_a[(i) * BM + th_r * TM + j];
+            }
+
+            for (int j = 0; j < TN; j++) {
+                cache_b[j] = buf_b[(i) * BN + th_c * TN + j];
+            }
+
+            for (int cc = 0; cc < TN; cc++) {
+                for (int cr = 0; cr < TM; cr++) {
+                    const int sums_idx = cc*TM + cr;
+                    sums[sums_idx] = mad(cache_a[cr], cache_b[cc], sums[sums_idx]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const int dr = ir * BM + th_r * TM;
+    const int dc = ic * BN + th_c * TN;
+
+    const ulong offsets = batch_idx * batch_stride_d;
+
+    for (int cc = 0; cc < TN; cc++) {
+        for (int cr = 0; cr < TM; cr++) {
+            if (dr + cr < ne01 && dc + cc < ne11) {
+                dst[offsets + (dc + cc) * stride_d + dr + cr] = sums[cc * TM + cr];
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mm_iq2_xxs_f32_l4_lm.cl

@@ -0,0 +1,187 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_required_subgroup_size
+#define INTEL_GPU 1
+#endif
+
+#define QK_K 256
+
+struct block_iq2_xxs {
+    half d;
+    ushort qs[QK_K/8];
+};
+
+#define LOAD_VEC_A 4
+#define LOAD_VEC_B 4
+
+#define BM 64
+#define BN 64
+#define BK 32
+#ifdef INTEL_GPU
+#define TM 8
+#define TN 8
+#else
+#define TM 4
+#define TN 8
+#endif
+
+kernel void kernel_mul_mm_iq2_xxs_f32_l4_lm(
+    global uchar * src0_qs,
+    global uint * src0_sas,
+    global half * src0_d,
+    ulong offset0,
+    global float4 * src1,
+    ulong offset1,
+    global float  * dst,
+    ulong offsetd,
+
+    int ne00,
+    int ne01,
+    int ne02,
+    int ne11,
+    int ne12,
+
+    int stride_a,
+    int stride_b,
+    int stride_d,
+
+    int batch_stride_a,
+    int batch_stride_b,
+    int batch_stride_d,
+
+    int r2,
+    int r3
+) {
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst  = (global float *)((global char*)dst  + offsetd);
+
+    local float buf_a[BM * BK];
+    local float buf_b[BN * BK];
+
+    const int batch_idx = get_global_id(2);
+
+    const int i13 = batch_idx / ne12;
+    const int i12 = batch_idx % ne12;
+
+    const int i03 = i13 / r3;
+    const int i02 = i12 / r2;
+
+    const int batch_idx_a = i03 * ne02 + i02;
+
+    const int ir = get_group_id(0);
+    const int ic = get_group_id(1);
+
+    const int tid = get_local_id(0);
+    const int th_r  = tid % (BM / TM);
+    const int th_c  = tid / (BM / TM);
+
+    const int loadr_a = get_local_id(0) % (BK / LOAD_VEC_A);
+    const int loadc_a = get_local_id(0) / (BK / LOAD_VEC_A);
+    const int loadr_b = get_local_id(0) % (BK / LOAD_VEC_B);
+    const int loadc_b = get_local_id(0) / (BK / LOAD_VEC_B);
+
+    const int loadstride_a = get_local_size(0) * LOAD_VEC_A / BK;
+    const int loadstride_b = get_local_size(0) * LOAD_VEC_B / BK;
+
+    // pos_a counts elements, not blocks - one iq2_xxs super block holds QK_K of them
+    int pos_a = batch_idx_a * batch_stride_a + ir * BM * stride_a;
+    int pos_b = (batch_idx   * batch_stride_b + ic * BN * stride_b) / LOAD_VEC_B;
+
+    float sums[TM * TN];
+    float cache_a[TM];
+    float cache_b[TN];
+
+    for (int i = 0; i < TM * TN; i++) {
+        sums[i] = 0.0f;
+    }
+
+    for (int block = 0; block < ne00; block += BK) {
+        for (int l = 0; l < BM; l += loadstride_a) {
+            if (ir*BM + loadc_a + l < ne01) {
+                int idx = pos_a + (loadc_a + l) * stride_a + loadr_a * LOAD_VEC_A;
+                int ib  = idx / QK_K;
+                ib += offset0/sizeof(struct block_iq2_xxs);
+                int e   = idx % QK_K;
+
+                int ib32 = e >> 5;
+                int rem  = e & 31;
+                int lg    = rem >> 3;
+                int j0   = rem & 7;
+
+                uint aux = src0_sas[8*ib + ib32];
+                float db = (float)src0_d[ib] * (0.5f + (float)(aux >> 28)) * 0.25f;
+
+                uint gi = src0_qs[32*ib + 4*ib32 + lg];
+                uint sg = iq2_xxs_signs((aux >> (7*lg)) & 127u);
+                uint  g  = iq2_xxs_grid[2*gi + (j0 >> 2)];
+
+                float4 v1;
+                v1.s0 = db * (float)((g >>  0) & 0xFF) * ((sg & (1 << (j0+0))) ? -1.f : 1.f);
+                v1.s1 = db * (float)((g >>  8) & 0xFF) * ((sg & (1 << (j0+1))) ? -1.f : 1.f);
+                v1.s2 = db * (float)((g >> 16) & 0xFF) * ((sg & (1 << (j0+2))) ? -1.f : 1.f);
+                v1.s3 = db * (float)((g >> 24) & 0xFF) * ((sg & (1 << (j0+3))) ? -1.f : 1.f);
+
+                buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = v1.s0;
+                buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = v1.s1;
+                buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = v1.s2;
+                buf_a[(loadr_a * LOAD_VEC_A + 3) * BM + loadc_a + l] = v1.s3;
+            } else {
+                buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * LOAD_VEC_A + 3) * BM + loadc_a + l] = 0.0f;
+            }
+        }
+
+        for (int l = 0; l < BN; l += loadstride_b) {
+            if (ic*BN + loadc_b + l < ne11) {
+                int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b;
+                buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0;
+                buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1;
+                buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = src1[idx].s2;
+                buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = src1[idx].s3;
+            } else {
+                buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = 0.0f;
+            }
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        pos_a += BK;
+        pos_b += BK / LOAD_VEC_B;
+
+        for (int i = 0; i < BK; i++) {
+            for (int j = 0; j < TM; j++) {
+                cache_a[j] = buf_a[(i) * BM + th_r * TM + j];
+            }
+
+            for (int j = 0; j < TN; j++) {
+                cache_b[j] = buf_b[(i) * BN + th_c * TN + j];
+            }
+
+            for (int cc = 0; cc < TN; cc++) {
+                for (int cr = 0; cr < TM; cr++) {
+                    const int sums_idx = cc*TM + cr;
+                    sums[sums_idx] = mad(cache_a[cr], cache_b[cc], sums[sums_idx]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const int dr = ir * BM + th_r * TM;
+    const int dc = ic * BN + th_c * TN;
+
+    const ulong offsets = batch_idx * batch_stride_d;
+
+    for (int cc = 0; cc < TN; cc++) {
+        for (int cr = 0; cr < TM; cr++) {
+            if (dr + cr < ne01 && dc + cc < ne11) {
+                dst[offsets + (dc + cc) * stride_d + dr + cr] = sums[cc * TM + cr];
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mm_q2_0_f32_l4_lm.cl

@@ -0,0 +1,171 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : enable
+
+#ifdef cl_intel_required_subgroup_size
+#define INTEL_GPU 1
+#endif
+
+#define QK2_0 64
+
+struct block_q2_0 {
+    half d;
+    uchar qs[16];
+};
+
+#define LOAD_VEC_A 4
+#define LOAD_VEC_B 4
+
+#define BM 64
+#define BN 64
+#define BK 32
+#ifdef INTEL_GPU
+#define TM 8
+#define TN 8
+#else
+#define TM 4
+#define TN 8
+#endif
+
+kernel void kernel_mul_mm_q2_0_f32_l4_lm(
+    global uchar * src0_qs,
+    global half * src0_d,
+    ulong offset0,
+    global float4 * src1,
+    ulong offset1,
+    global float  * dst,
+    ulong offsetd,
+
+    int ne00,
+    int ne01,
+    int ne02,
+    int ne11,
+    int ne12,
+
+    int stride_a,
+    int stride_b,
+    int stride_d,
+
+    int batch_stride_a,
+    int batch_stride_b,
+    int batch_stride_d,
+
+    int r2,
+    int r3
+) {
+    src1 = (global float4*)((global char*)src1 + offset1);
+    dst  = (global float *)((global char*)dst  + offsetd);
+
+    local float buf_a[BM * BK];
+    local float buf_b[BN * BK];
+
+    const int batch_idx = get_global_id(2);
+
+    const int i13 = batch_idx / ne12;
+    const int i12 = batch_idx % ne12;
+
+    const int i03 = i13 / r3;
+    const int i02 = i12 / r2;
+
+    const int batch_idx_a = i03 * ne02 + i02;
+
+    const int ir = get_group_id(0);
+    const int ic = get_group_id(1);
+
+    const int tid = get_local_id(0);
+    const int th_r  = tid % (BM / TM);
+    const int th_c  = tid / (BM / TM);
+
+    const int loadr_a = get_local_id(0) % (BK / LOAD_VEC_A);
+    const int loadc_a = get_local_id(0) / (BK / LOAD_VEC_A);
+    const int loadr_b = get_local_id(0) % (BK / LOAD_VEC_B);
+    const int loadc_b = get_local_id(0) / (BK / LOAD_VEC_B);
+
+    const int loadstride_a = get_local_size(0) * LOAD_VEC_A / BK;
+    const int loadstride_b = get_local_size(0) * LOAD_VEC_B / BK;
+
+    int pos_a = batch_idx_a * batch_stride_a + ir * BM * stride_a;
+    int pos_b = (batch_idx   * batch_stride_b + ic * BN * stride_b) / LOAD_VEC_B;
+
+    float sums[TM * TN];
+    float cache_a[TM];
+    float cache_b[TN];
+
+    for (int i = 0; i < TM * TN; i++) {
+        sums[i] = 0.0f;
+    }
+
+    for (int block = 0; block < ne00; block += BK) {
+        for (int l = 0; l < BM; l += loadstride_a) {
+            if (ir*BM + loadc_a + l < ne01) {
+                int idx = pos_a + (loadc_a + l) * stride_a + loadr_a * LOAD_VEC_A;
+                int ib  = idx / QK2_0;
+                ib += offset0/sizeof(struct block_q2_0);
+                int e   = idx % QK2_0;
+
+                const uint q = src0_qs[16*ib + e/4];
+                const int4 v = convert_int4(((uint4)(q) >> (uint4)(0, 2, 4, 6)) & (uint4)(3)) - (int4)(1);
+                const float4 v1 = (float)src0_d[ib] * convert_float4(v);
+
+                buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = v1.s0;
+                buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = v1.s1;
+                buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = v1.s2;
+                buf_a[(loadr_a * LOAD_VEC_A + 3) * BM + loadc_a + l] = v1.s3;
+            } else {
+                buf_a[(loadr_a * LOAD_VEC_A + 0) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * LOAD_VEC_A + 1) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * LOAD_VEC_A + 2) * BM + loadc_a + l] = 0.0f;
+                buf_a[(loadr_a * LOAD_VEC_A + 3) * BM + loadc_a + l] = 0.0f;
+            }
+        }
+
+        for (int l = 0; l < BN; l += loadstride_b) {
+            if (ic*BN + loadc_b + l < ne11) {
+                int idx = pos_b + (loadc_b + l) * stride_b / LOAD_VEC_B + loadr_b;
+                buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = src1[idx].s0;
+                buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = src1[idx].s1;
+                buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = src1[idx].s2;
+                buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = src1[idx].s3;
+            } else {
+                buf_b[(loadr_b * LOAD_VEC_B + 0) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 1) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 2) * BN + loadc_b + l] = 0.0f;
+                buf_b[(loadr_b * LOAD_VEC_B + 3) * BN + loadc_b + l] = 0.0f;
+            }
+        }
+
+        barrier(CLK_LOCAL_MEM_FENCE);
+
+        pos_a += BK;
+        pos_b += BK / LOAD_VEC_B;
+
+        for (int i = 0; i < BK; i++) {
+            for (int j = 0; j < TM; j++) {
+                cache_a[j] = buf_a[(i) * BM + th_r * TM + j];
+            }
+
+            for (int j = 0; j < TN; j++) {
+                cache_b[j] = buf_b[(i) * BN + th_c * TN + j];
+            }
+
+            for (int cc = 0; cc < TN; cc++) {
+                for (int cr = 0; cr < TM; cr++) {
+                    const int sums_idx = cc*TM + cr;
+                    sums[sums_idx] = mad(cache_a[cr], cache_b[cc], sums[sums_idx]);
+                }
+            }
+        }
+        barrier(CLK_LOCAL_MEM_FENCE);
+    }
+
+    const int dr = ir * BM + th_r * TM;
+    const int dc = ic * BN + th_c * TN;
+
+    const ulong offsets = batch_idx * batch_stride_d;
+
+    for (int cc = 0; cc < TN; cc++) {
+        for (int cr = 0; cr < TM; cr++) {
+            if (dr + cr < ne01 && dc + cc < ne11) {
+                dst[offsets + (dc + cc) * stride_d + dr + cr] = sums[cc * TM + cr];
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mv_iq2_s_f32.cl

@@ -0,0 +1,147 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : 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 QK_K 256
+
+// qs holds 32 grid low bytes then 32 sign bytes
+typedef struct {
+    half  d;
+    uchar qs[QK_K/4];
+    uchar qh[QK_K/32];
+    uchar scales[QK_K/32];
+} block_iq2_s;
+
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 16
+#elif defined (ADRENO_GPU)
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+#undef  BLOCK_STRIDE
+// 8 threads cover one super block, one 32 element sub block each
+#define BLOCK_STRIDE (N_SIMDWIDTH/8)
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_iq2_s_f32(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne12,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int ix = get_sub_group_local_id()/8;  // super block index
+    int it = get_sub_group_local_id()%8;  // sub block inside the super block
+
+    int nb = ne00/QK_K;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+    ulong offset_src1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+    global block_iq2_s * x = (global block_iq2_s *) (src0 + offset_src0);
+    global float         * y = (global float         *) (src1 + offset_src1);
+
+    float yl[32];
+    float sumf[N_DST] = {0.f};
+    float all_sum;
+
+    global float * y4 = y + ix * QK_K + 32 * it;
+
+    for (int ib = ix; ib < nb; ib += BLOCK_STRIDE) {
+        for (int i = 0; i < 32; ++i) {
+            yl[i] = y4[i];
+        }
+
+        global char * xrow = (global char *)(x + ib);
+
+        // only ne01 output rows exist; reading past them can pull in an inf scale
+        for (int row = 0; row < N_DST && first_row + row < ne01; row++) {
+            global block_iq2_s * xb = (global block_iq2_s *)(xrow + row*nb01);
+
+            uchar scb = xb->scales[it];
+            float dlo = (float)xb->d * (0.5f + (float)(scb & 0xf)) * 0.25f;
+            float dhi = (float)xb->d * (0.5f + (float)(scb >>  4)) * 0.25f;
+
+            uchar qhb = xb->qh[it];
+            global uchar * qsb = xb->qs + 4*it;
+            global uchar * sgb = xb->qs + 32 + 4*it;
+
+            float alo = 0.f, ahi = 0.f;
+            for (int l = 0; l < 4; ++l) {
+                uint  gi = (uint)qsb[l] | (((uint)qhb << (8-2*l)) & 0x300);
+                uchar sg = sgb[l];
+                uint  lo = iq2_s_grid[2*gi+0];
+                uint  hi = iq2_s_grid[2*gi+1];
+                float a  = 0.f;
+                for (int j = 0; j < 4; ++j) {
+                    a += yl[8*l+j+0] * (float)((lo >> (8*j)) & 0xFF) * ((sg & (1 << j))     ? -1.f : 1.f);
+                    a += yl[8*l+j+4] * (float)((hi >> (8*j)) & 0xFF) * ((sg & (1 << (j+4))) ? -1.f : 1.f);
+                }
+                if (l < 2) alo += a; else ahi += a;
+            }
+
+            sumf[row] += dlo * alo + dhi * ahi;
+        }
+
+        y4 += BLOCK_STRIDE * QK_K;
+    }
+
+    global float * dst_f32 = (global float *) dst + im*ne0*ne1 + r1*ne0;
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = sub_group_reduce_add(sumf[row]);
+        if (first_row + row < ne01) {
+            if (get_sub_group_local_id() == 0) {
+                dst_f32[first_row + row] = all_sum;
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mv_iq2_s_f32_flat.cl

@@ -0,0 +1,148 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : 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 QK_K 256
+
+// qs holds 32 grid low bytes then 32 sign bytes
+typedef struct {
+    half  d;
+    uchar qs[QK_K/4];
+    uchar qh[QK_K/32];
+    uchar scales[QK_K/32];
+} block_iq2_s;
+
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 16
+#elif defined (ADRENO_GPU)
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+#undef  BLOCK_STRIDE
+// 8 threads cover one super block, one 32 element sub block each
+#define BLOCK_STRIDE (N_SIMDWIDTH/8)
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_iq2_s_f32_flat(
+        global uchar * src0_qs,
+        global uchar * src0_sg,
+        global uchar * src0_qh,
+        global uchar * src0_sc,
+        global half * src0_d,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne12,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int ix = get_sub_group_local_id()/8;  // super block index
+    int it = get_sub_group_local_id()%8;  // sub block inside the super block
+
+    int nb = ne00/QK_K;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = (offset0 + first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03)/sizeof(block_iq2_s);
+    ulong row_stride = nb01/sizeof(block_iq2_s);
+    ulong offset_src1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+    global float         * y = (global float         *) (src1 + offset_src1);
+
+    float yl[32];
+    float sumf[N_DST] = {0.f};
+    float all_sum;
+
+    global float * y4 = y + ix * QK_K + 32 * it;
+
+    for (int ib = ix; ib < nb; ib += BLOCK_STRIDE) {
+        for (int i = 0; i < 32; ++i) {
+            yl[i] = y4[i];
+        }
+
+        // only ne01 output rows exist; reading past them can pull in an inf scale
+        for (int row = 0; row < N_DST && first_row + row < ne01; row++) {
+            const ulong b = offset_src0 + ib + row*row_stride;
+
+            uchar scb = src0_sc[8*b + it];
+            float dlo = (float)src0_d[b] * (0.5f + (float)(scb & 0xf)) * 0.25f;
+            float dhi = (float)src0_d[b] * (0.5f + (float)(scb >>  4)) * 0.25f;
+
+            uchar qhb = src0_qh[8*b + it];
+            global uchar * qsb = src0_qs + 32*b + 4*it;
+            global uchar * sgb = src0_sg + 32*b + 4*it;
+
+            float alo = 0.f, ahi = 0.f;
+            for (int l = 0; l < 4; ++l) {
+                uint  gi = (uint)qsb[l] | (((uint)qhb << (8-2*l)) & 0x300);
+                uchar sg = sgb[l];
+                uint  lo = iq2_s_grid[2*gi+0];
+                uint  hi = iq2_s_grid[2*gi+1];
+                float a  = 0.f;
+                for (int j = 0; j < 4; ++j) {
+                    a += yl[8*l+j+0] * (float)((lo >> (8*j)) & 0xFF) * ((sg & (1 << j))     ? -1.f : 1.f);
+                    a += yl[8*l+j+4] * (float)((hi >> (8*j)) & 0xFF) * ((sg & (1 << (j+4))) ? -1.f : 1.f);
+                }
+                if (l < 2) alo += a; else ahi += a;
+            }
+
+            sumf[row] += dlo * alo + dhi * ahi;
+        }
+
+        y4 += BLOCK_STRIDE * QK_K;
+    }
+
+    global float * dst_f32 = (global float *) dst + im*ne0*ne1 + r1*ne0;
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = sub_group_reduce_add(sumf[row]);
+        if (first_row + row < ne01) {
+            if (get_sub_group_local_id() == 0) {
+                dst_f32[first_row + row] = all_sum;
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mv_iq2_xs_f32.cl

@@ -0,0 +1,143 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : 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 QK_K 256
+
+typedef struct {
+    half   d;
+    ushort qs[QK_K/8];
+    uchar  scales[QK_K/32];
+} block_iq2_xs;
+
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 16
+#elif defined (ADRENO_GPU)
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+#undef  BLOCK_STRIDE
+// 8 threads cover one super block, one 32 element sub block each
+#define BLOCK_STRIDE (N_SIMDWIDTH/8)
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_iq2_xs_f32(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne12,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int ix = get_sub_group_local_id()/8;  // super block index
+    int it = get_sub_group_local_id()%8;  // sub block inside the super block
+
+    int nb = ne00/QK_K;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+    ulong offset_src1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+    global block_iq2_xs * x = (global block_iq2_xs *) (src0 + offset_src0);
+    global float         * y = (global float         *) (src1 + offset_src1);
+
+    float yl[32];
+    float sumf[N_DST] = {0.f};
+    float all_sum;
+
+    global float * y4 = y + ix * QK_K + 32 * it;
+
+    for (int ib = ix; ib < nb; ib += BLOCK_STRIDE) {
+        for (int i = 0; i < 32; ++i) {
+            yl[i] = y4[i];
+        }
+
+        global char * xrow = (global char *)(x + ib);
+
+        // only ne01 output rows exist; reading past them can pull in an inf scale
+        for (int row = 0; row < N_DST && first_row + row < ne01; row++) {
+            global block_iq2_xs * xb = (global block_iq2_xs *)(xrow + row*nb01);
+
+            uchar scb = xb->scales[it];
+            float dlo = (float)xb->d * (0.5f + (float)(scb & 0xf)) * 0.25f;
+            float dhi = (float)xb->d * (0.5f + (float)(scb >>  4)) * 0.25f;
+
+            global ushort * q16 = xb->qs + 4*it;
+
+            float alo = 0.f, ahi = 0.f;
+            for (int l = 0; l < 4; ++l) {
+                ushort q  = q16[l];
+                uchar  sg = iq2_xs_signs(q >> 9);
+                uint   lo = iq2_xs_grid[2*(q & 511)+0];
+                uint   hi = iq2_xs_grid[2*(q & 511)+1];
+                float  a  = 0.f;
+                for (int j = 0; j < 4; ++j) {
+                    a += yl[8*l+j+0] * (float)((lo >> (8*j)) & 0xFF) * ((sg & (1 << j))     ? -1.f : 1.f);
+                    a += yl[8*l+j+4] * (float)((hi >> (8*j)) & 0xFF) * ((sg & (1 << (j+4))) ? -1.f : 1.f);
+                }
+                if (l < 2) alo += a; else ahi += a;
+            }
+
+            sumf[row] += dlo * alo + dhi * ahi;
+        }
+
+        y4 += BLOCK_STRIDE * QK_K;
+    }
+
+    global float * dst_f32 = (global float *) dst + im*ne0*ne1 + r1*ne0;
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = sub_group_reduce_add(sumf[row]);
+        if (first_row + row < ne01) {
+            if (get_sub_group_local_id() == 0) {
+                dst_f32[first_row + row] = all_sum;
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mv_iq2_xs_f32_flat.cl

@@ -0,0 +1,142 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : 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 QK_K 256
+
+typedef struct {
+    half   d;
+    ushort qs[QK_K/8];
+    uchar  scales[QK_K/32];
+} block_iq2_xs;
+
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 16
+#elif defined (ADRENO_GPU)
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+#undef  BLOCK_STRIDE
+// 8 threads cover one super block, one 32 element sub block each
+#define BLOCK_STRIDE (N_SIMDWIDTH/8)
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_iq2_xs_f32_flat(
+        global ushort * src0_qs,
+        global uchar * src0_sc,
+        global half * src0_d,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne12,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int ix = get_sub_group_local_id()/8;  // super block index
+    int it = get_sub_group_local_id()%8;  // sub block inside the super block
+
+    int nb = ne00/QK_K;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = (offset0 + first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03)/sizeof(block_iq2_xs);
+    ulong row_stride = nb01/sizeof(block_iq2_xs);
+    ulong offset_src1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+    global float         * y = (global float         *) (src1 + offset_src1);
+
+    float yl[32];
+    float sumf[N_DST] = {0.f};
+    float all_sum;
+
+    global float * y4 = y + ix * QK_K + 32 * it;
+
+    for (int ib = ix; ib < nb; ib += BLOCK_STRIDE) {
+        for (int i = 0; i < 32; ++i) {
+            yl[i] = y4[i];
+        }
+
+        // only ne01 output rows exist; reading past them can pull in an inf scale
+        for (int row = 0; row < N_DST && first_row + row < ne01; row++) {
+            const ulong b = offset_src0 + ib + row*row_stride;
+
+            uchar scb = src0_sc[8*b + it];
+            float dlo = (float)src0_d[b] * (0.5f + (float)(scb & 0xf)) * 0.25f;
+            float dhi = (float)src0_d[b] * (0.5f + (float)(scb >>  4)) * 0.25f;
+
+            global ushort * q16 = src0_qs + 32*b + 4*it;
+
+            float alo = 0.f, ahi = 0.f;
+            for (int l = 0; l < 4; ++l) {
+                ushort q  = q16[l];
+                uchar  sg = iq2_xs_signs(q >> 9);
+                uint   lo = iq2_xs_grid[2*(q & 511)+0];
+                uint   hi = iq2_xs_grid[2*(q & 511)+1];
+                float  a  = 0.f;
+                for (int j = 0; j < 4; ++j) {
+                    a += yl[8*l+j+0] * (float)((lo >> (8*j)) & 0xFF) * ((sg & (1 << j))     ? -1.f : 1.f);
+                    a += yl[8*l+j+4] * (float)((hi >> (8*j)) & 0xFF) * ((sg & (1 << (j+4))) ? -1.f : 1.f);
+                }
+                if (l < 2) alo += a; else ahi += a;
+            }
+
+            sumf[row] += dlo * alo + dhi * ahi;
+        }
+
+        y4 += BLOCK_STRIDE * QK_K;
+    }
+
+    global float * dst_f32 = (global float *) dst + im*ne0*ne1 + r1*ne0;
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = sub_group_reduce_add(sumf[row]);
+        if (first_row + row < ne01) {
+            if (get_sub_group_local_id() == 0) {
+                dst_f32[first_row + row] = all_sum;
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mv_iq2_xxs_f32.cl

@@ -0,0 +1,140 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : 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 QK_K 256
+
+typedef struct {
+    half   d;
+    ushort qs[QK_K/8];
+} block_iq2_xxs;
+
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 16
+#elif defined (ADRENO_GPU)
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+#undef  BLOCK_STRIDE
+// 8 threads cover one super block, one 32 element sub block each
+#define BLOCK_STRIDE (N_SIMDWIDTH/8)
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_iq2_xxs_f32(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne12,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int ix = get_sub_group_local_id()/8;  // super block index
+    int it = get_sub_group_local_id()%8;  // sub block inside the super block
+
+    int nb = ne00/QK_K;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+    ulong offset_src1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+    global block_iq2_xxs * x = (global block_iq2_xxs *) (src0 + offset_src0);
+    global float         * y = (global float         *) (src1 + offset_src1);
+
+    float yl[32];
+    float sumf[N_DST] = {0.f};
+    float all_sum;
+
+    global float * y4 = y + ix * QK_K + 32 * it;
+
+    for (int ib = ix; ib < nb; ib += BLOCK_STRIDE) {
+        for (int i = 0; i < 32; ++i) {
+            yl[i] = y4[i];
+        }
+
+        global char * xrow = (global char *)(x + ib);
+
+        // only ne01 output rows exist; reading past them can pull in an inf scale
+        for (int row = 0; row < N_DST && first_row + row < ne01; row++) {
+            global block_iq2_xxs * xb = (global block_iq2_xxs *)(xrow + row*nb01);
+
+            global ushort * q16 = xb->qs + 4*it;
+            uint a0 = (uint)q16[0] | ((uint)q16[1] << 16);
+            uint a1 = (uint)q16[2] | ((uint)q16[3] << 16);
+
+            float db = (float)xb->d * (0.5f + (float)(a1 >> 28)) * 0.25f;
+
+            float acc = 0.f;
+            for (int l = 0; l < 4; ++l) {
+                uint  gi = (a0 >> (8*l)) & 0xFF;
+                uchar sg = iq2_xxs_signs((a1 >> (7*l)) & 127);
+                uint  lo = iq2_xxs_grid[2*gi+0];
+                uint  hi = iq2_xxs_grid[2*gi+1];
+                for (int j = 0; j < 4; ++j) {
+                    acc += yl[8*l+j+0] * (float)((lo >> (8*j)) & 0xFF) * ((sg & (1 << j))     ? -1.f : 1.f);
+                    acc += yl[8*l+j+4] * (float)((hi >> (8*j)) & 0xFF) * ((sg & (1 << (j+4))) ? -1.f : 1.f);
+                }
+            }
+
+            sumf[row] += db * acc;
+        }
+
+        y4 += BLOCK_STRIDE * QK_K;
+    }
+
+    global float * dst_f32 = (global float *) dst + im*ne0*ne1 + r1*ne0;
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = sub_group_reduce_add(sumf[row]);
+        if (first_row + row < ne01) {
+            if (get_sub_group_local_id() == 0) {
+                dst_f32[first_row + row] = all_sum;
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mv_iq2_xxs_f32_flat.cl

@@ -0,0 +1,137 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : 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 QK_K 256
+
+typedef struct {
+    half   d;
+    ushort qs[QK_K/8];
+} block_iq2_xxs;
+
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 16
+#elif defined (ADRENO_GPU)
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+#undef  BLOCK_STRIDE
+// 8 threads cover one super block, one 32 element sub block each
+#define BLOCK_STRIDE (N_SIMDWIDTH/8)
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_iq2_xxs_f32_flat(
+        global uchar * src0_qs,
+        global uint * src0_sas,
+        global half * src0_d,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne12,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int ix = get_sub_group_local_id()/8;  // super block index
+    int it = get_sub_group_local_id()%8;  // sub block inside the super block
+
+    int nb = ne00/QK_K;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = (offset0 + first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03)/sizeof(block_iq2_xxs);
+    ulong row_stride = nb01/sizeof(block_iq2_xxs);
+    ulong offset_src1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+    global float         * y = (global float         *) (src1 + offset_src1);
+
+    float yl[32];
+    float sumf[N_DST] = {0.f};
+    float all_sum;
+
+    global float * y4 = y + ix * QK_K + 32 * it;
+
+    for (int ib = ix; ib < nb; ib += BLOCK_STRIDE) {
+        for (int i = 0; i < 32; ++i) {
+            yl[i] = y4[i];
+        }
+
+        // only ne01 output rows exist; reading past them can pull in an inf scale
+        for (int row = 0; row < N_DST && first_row + row < ne01; row++) {
+            const ulong b = offset_src0 + ib + row*row_stride;
+            uint a0 = as_uint(vload4(8*b + it, src0_qs));
+            uint a1 = src0_sas[8*b + it];
+
+            float db = (float)src0_d[b] * (0.5f + (float)(a1 >> 28)) * 0.25f;
+
+            float acc = 0.f;
+            for (int l = 0; l < 4; ++l) {
+                uint  gi = (a0 >> (8*l)) & 0xFF;
+                uchar sg = iq2_xxs_signs((a1 >> (7*l)) & 127);
+                uint  lo = iq2_xxs_grid[2*gi+0];
+                uint  hi = iq2_xxs_grid[2*gi+1];
+                for (int j = 0; j < 4; ++j) {
+                    acc += yl[8*l+j+0] * (float)((lo >> (8*j)) & 0xFF) * ((sg & (1 << j))     ? -1.f : 1.f);
+                    acc += yl[8*l+j+4] * (float)((hi >> (8*j)) & 0xFF) * ((sg & (1 << (j+4))) ? -1.f : 1.f);
+                }
+            }
+
+            sumf[row] += db * acc;
+        }
+
+        y4 += BLOCK_STRIDE * QK_K;
+    }
+
+    global float * dst_f32 = (global float *) dst + im*ne0*ne1 + r1*ne0;
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = sub_group_reduce_add(sumf[row]);
+        if (first_row + row < ne01) {
+            if (get_sub_group_local_id() == 0) {
+                dst_f32[first_row + row] = all_sum;
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mv_q2_0_f32.cl

@@ -0,0 +1,134 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : 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 QK2_0 64
+
+// 00=-1, 01=0, 10=+1, 11=+2
+typedef struct {
+    half  d;
+    uchar qs[QK2_0/4];
+} block_q2_0;
+
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 4 // number of rows each SIMD group works on
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // SIMD group size
+#elif defined (ADRENO_GPU)
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+#undef  BLOCK_STRIDE
+// 4 threads cover one block
+#define BLOCK_STRIDE (N_SIMDWIDTH/4)
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_q2_0_f32(
+        global char * src0,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne12,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src0 = src0 + offset0;
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int ix = get_sub_group_local_id()/4;  // block index
+    int it = get_sub_group_local_id()%4;  // part of the block
+
+    int nb = ne00/QK2_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03;
+    ulong offset_src1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+    global block_q2_0 * x = (global block_q2_0 *) (src0 + offset_src0);
+    global float   * y = (global float   *) (src1 + offset_src1);
+
+    float yl[16];
+    float sumf[N_DST] = {0.f};
+    float all_sum;
+
+    global float * y4 = y + ix * QK2_0 + 16 * it;
+
+    for (int ib = ix; ib < nb; ib += BLOCK_STRIDE) {
+        for (int i = 0; i < 16; ++i) {
+            yl[i] = y4[i];
+        }
+
+        global char * xrow = (global char *)(x + ib);
+
+        // only ne01 output rows exist; reading past them can pull in an inf scale
+        for (int row = 0; row < N_DST && first_row + row < ne01; row++) {
+            global block_q2_0 * xb = (global block_q2_0 *)(xrow + row*nb01);
+
+            global uchar * q = xb->qs + 4*it;
+
+            float acc = 0.f;
+            for (int i = 0; i < 4; ++i) {
+                uchar b = q[i];
+                acc += yl[4*i+0] * (float)(((b >> 0) & 3) - 1);
+                acc += yl[4*i+1] * (float)(((b >> 2) & 3) - 1);
+                acc += yl[4*i+2] * (float)(((b >> 4) & 3) - 1);
+                acc += yl[4*i+3] * (float)(((b >> 6) & 3) - 1);
+            }
+
+            sumf[row] += (float)xb->d * acc;
+        }
+
+        y4 += BLOCK_STRIDE * QK2_0;
+    }
+
+    global float * dst_f32 = (global float *) dst + im*ne0*ne1 + r1*ne0;
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = sub_group_reduce_add(sumf[row]);
+        if (first_row + row < ne01) {
+            if (get_sub_group_local_id() == 0) {
+                dst_f32[first_row + row] = all_sum;
+            }
+        }
+    }
+}

ggml/src/ggml-opencl/kernels/mul_mv_q2_0_f32_flat.cl

@@ -0,0 +1,132 @@
+#pragma OPENCL EXTENSION cl_khr_fp16 : 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 QK2_0 64
+
+// 00=-1, 01=0, 10=+1, 11=+2
+typedef struct {
+    half  d;
+    uchar qs[QK2_0/4];
+} block_q2_0;
+
+#undef N_DST
+#undef N_SIMDGROUP
+#undef N_SIMDWIDTH
+
+#ifdef INTEL_GPU
+#define N_DST 4 // number of rows each SIMD group works on
+#define N_SIMDGROUP 1 // number of SIMD groups in a thread group
+#define N_SIMDWIDTH 16 // SIMD group size
+#elif defined (ADRENO_GPU)
+#define N_DST 4
+#define N_SIMDGROUP 1
+#define N_SIMDWIDTH 64
+#endif
+
+#undef  BLOCK_STRIDE
+// 4 threads cover one block
+#define BLOCK_STRIDE (N_SIMDWIDTH/4)
+
+#ifdef INTEL_GPU
+REQD_SUBGROUP_SIZE_16
+#elif defined (ADRENO_GPU)
+REQD_SUBGROUP_SIZE_64
+#endif
+kernel void kernel_mul_mv_q2_0_f32_flat(
+        global uchar * src0_qs,
+        global half * src0_d,
+        ulong offset0,
+        global char * src1,
+        ulong offset1,
+        global char * dst,
+        ulong offsetd,
+        int ne00,
+        int ne01,
+        ulong nb01,
+        ulong nb02,
+        ulong nb03,
+        int ne12,
+        ulong nb11,
+        ulong nb12,
+        ulong nb13,
+        int ne0,
+        int ne1,
+        int r2,
+        int r3
+) {
+    src1 = src1 + offset1;
+    dst  = dst  + offsetd;
+
+    int ix = get_sub_group_local_id()/4;  // block index
+    int it = get_sub_group_local_id()%4;  // part of the block
+
+    int nb = ne00/QK2_0;
+
+    int r0 = get_group_id(0);
+    int r1 = get_group_id(1);
+    int im = get_group_id(2);
+    int first_row = (r0 * N_SIMDGROUP + get_sub_group_id()) * N_DST;
+
+    int i12 = im%ne12;
+    int i13 = im/ne12;
+
+    ulong offset_src0 = (offset0 + first_row*nb01 + (i12/r2)*nb02 + (i13/r3)*nb03)/sizeof(block_q2_0);
+    ulong row_stride = nb01/sizeof(block_q2_0);
+    ulong offset_src1 =        r1*nb11 + (i12   )*nb12 + (i13   )*nb13;
+
+    global float   * y = (global float   *) (src1 + offset_src1);
+
+    float yl[16];
+    float sumf[N_DST] = {0.f};
+    float all_sum;
+
+    global float * y4 = y + ix * QK2_0 + 16 * it;
+
+    for (int ib = ix; ib < nb; ib += BLOCK_STRIDE) {
+        for (int i = 0; i < 16; ++i) {
+            yl[i] = y4[i];
+        }
+
+        // only ne01 output rows exist; reading past them can pull in an inf scale
+        for (int row = 0; row < N_DST && first_row + row < ne01; row++) {
+            const ulong b = offset_src0 + ib + row*row_stride;
+
+            global uchar * q = src0_qs + 16*b + 4*it;
+
+            float acc = 0.f;
+            for (int i = 0; i < 4; ++i) {
+                uchar b = q[i];
+                acc += yl[4*i+0] * (float)(((b >> 0) & 3) - 1);
+                acc += yl[4*i+1] * (float)(((b >> 2) & 3) - 1);
+                acc += yl[4*i+2] * (float)(((b >> 4) & 3) - 1);
+                acc += yl[4*i+3] * (float)(((b >> 6) & 3) - 1);
+            }
+
+            sumf[row] += (float)src0_d[b] * acc;
+        }
+
+        y4 += BLOCK_STRIDE * QK2_0;
+    }
+
+    global float * dst_f32 = (global float *) dst + im*ne0*ne1 + r1*ne0;
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = sub_group_reduce_add(sumf[row]);
+        if (first_row + row < ne01) {
+            if (get_sub_group_local_id() == 0) {
+                dst_f32[first_row + row] = all_sum;
+            }
+        }
+    }
+}