Commit Graph

  • 080bbbe852 examples : Remove leading space from txt output (#3921) master Cappuccino 2026-07-11 23:53:59 +08:00
  • 7695a53312 ggml : use ggml_vqtbl1q_u8 for 32-bit compat (#0) Georgi Gerganov 2026-07-10 11:06:42 +03:00
  • 289ecb05c9 talk-llama : sync llama.cpp Georgi Gerganov 2026-07-10 10:41:19 +03:00
  • afa6bffd2b sync : ggml Georgi Gerganov 2026-07-10 10:36:00 +03:00
  • 4fc2a2d922 ggml : bump version to 0.16.0 (ggml/1559) Georgi Gerganov 2026-07-10 10:25:06 +03:00
  • 6b17d2c05f ggml : fix conv 2d dw (llama/25490) Georgi Gerganov 2026-07-09 17:56:32 +03:00
  • ade3292d03 common : adapt to q2_0 (ggml/0) Georgi Gerganov 2026-07-09 14:58:55 +03:00
  • a2c50280d8 Only index by compile times + always multiply/add (llama/25445) Oliver Simons 2026-07-09 13:23:57 +02:00
  • 9cd788e703 metal : add CONV_2D_DW (depthwise convolution) support (llama/21565) Sou-ly 2026-07-09 18:29:15 +09:00
  • cc7a521508 ggml-hip: enable -funsafe-math-optimizations (llama/24668) RapidMark 2026-07-09 01:02:26 -07:00
  • 1b0b078a98 cuda: align snake fusion matcher with the other backends (llama/25460) Pascal 2026-07-09 10:00:06 +02:00
  • def36a5831 hexagon: add VISION RoPE support (llama/25216) Aparna M P 2026-07-09 10:25:00 +05:30
  • 4b914096e8 ggml-webgpu: tune subgroup split (d_split) in flash_attn_vec (llama/25418) Masashi Yoshimura 2026-07-09 08:34:19 +09:00
  • 8e7b672fb4 opencl: Q6_K GEMM/GEMV fix for ne01 of weights that are not multiples of 128. (llama/25464) Hongqiang Wang 2026-07-08 15:52:21 -07:00
  • b47f39dba2 vulkan: disable FA mask_opt on GCN to improve performance (llama/24362) Ruben Ortlam 2026-07-08 19:01:25 +02:00
  • 5f052fbfaf opencl: ragged-tile MoE prefill FP16 GEMM optimization (skip padded expert tiles) (llama/25433) Hongqiang Wang 2026-07-08 09:44:55 -07:00
  • 06492a14a3 vulkan: for small AMD GPUs, reduce submission threshold based on CU count (llama/25240) Ruben Ortlam 2026-07-08 18:15:18 +02:00
  • 50e0adb702 hexagon: new vtcm layouts and improved pipelines for MUL_MAT, MUL_MAT_ID and FLASH_ATTN_EXT (llama/25425) Max Krasnyansky 2026-07-08 07:38:27 -07:00
  • b83d1671db cuda : add support for f16->f16 GGML_OP_SET_ROWS (llama/25367) fairydreaming 2026-07-08 13:24:20 +02:00
  • 81c8bb90df metal : add set_rows with src0 f16 (llama/25434) fairydreaming 2026-07-08 08:49:07 +02:00
  • 3d521d1bd8 ggml : fix A indexing in simd_gemm scalar tail-column path (llama/25390) tyronecai 2026-07-08 13:00:05 +08:00
  • 72e3f210f3 ggml : add support for CPU f16->f16 GGML_OP_SET_ROWS (llama/25344) fairydreaming 2026-07-08 05:46:28 +02:00
  • 75f1f0c063 opencl: fix potential crash in aos reconstruct (llama/25383) lhez 2026-07-07 20:34:29 -07:00
  • 66e178ec84 Add Q2_0 quantization: type definition and CPU backend (llama/24448) Pasha Khosravi 2026-07-07 12:05:47 -07:00
  • 2061e3b6bf CUDA: Fuse MMVQ post-scale for NVFP4 (llama/24481) Oliver Simons 2026-07-07 17:12:19 +02:00
  • 38d16b9a7e ggml-hip : add -fno-finite-math-only alongside -ffast-math (llama/25373) asf0 2026-07-07 05:27:50 -06:00
  • 7c04837122 fix unsupported UT cases of CONT & CPY (llama/25231) Neo Zhang 2026-07-07 17:20:52 +08:00
  • bcbc5bc554 support op col2im_1d (llama/25264) Neo Zhang 2026-07-07 16:07:46 +08:00
  • a0b8896847 support OP cross_entropy_loss, cross_entropy_loss_back (llama/25236) Neo Zhang 2026-07-07 15:48:50 +08:00
  • e1d865d834 sycl : set K_QUANTS_PER_ITERATION to 1 on DMMV path (llama/25063) Todd Malsbary 2026-07-07 00:43:41 -07:00
  • 120fc3729d fix unsupport ACC UT cases for noncontiguous (llama/25124) Neo Zhang 2026-07-07 15:40:38 +08:00
  • bfd86f4304 sycl : enhance argsort to support all UT cases (llama/25125) Neo Zhang 2026-07-07 15:39:29 +08:00
  • f353a27988 sycl : use sycl func to fix AOT double type issue (llama/25081) Neo Zhang 2026-07-07 15:38:33 +08:00
  • 2c6e470ced sycl : rename the env vars from "disable" to "enable" (llama/25042) Neo Zhang 2026-07-07 15:33:51 +08:00
  • 753e101595 ggml : make ggml_time_init idempotent (llama/24422) An Long 2026-07-07 16:29:17 +09:00
  • bd5ce1ed48 vulkan : check src0 type in GGML_OP_SET_ROWS to avoid failures due to unimplemented f16 support (llama/25351) fairydreaming 2026-07-07 06:56:02 +02:00
  • 3a1387caf8 opencl: general flash attention decode performance optimizations (llama/25366) Hongqiang Wang 2026-07-06 19:57:52 -07:00
  • 91ca310157 metal: add col2im_1d op (f32/f16/bf16) (llama/25176) Pascal 2026-07-06 20:47:36 +02:00
  • 3f05208e5c CUDA: remove -sm row, refactor cuBLAS (llama/24216) Johannes Gäßler 2026-07-06 20:04:53 +02:00
  • 82ac928a5d CUDA: extend K-type validation to V-types for flash attention (llama/24403) Alexey Kopytko 2026-07-06 23:26:50 +09:00
  • 7d8e64fe37 ggml-cpu: use UE4M3 LUT in ARM NVFP4 dot product (llama/25331) ragz4125 2026-07-06 16:36:40 +05:30
  • ce2c003f17 ggml-cpu: Enable tiled matmul on AIX (llama/25199) shalinib-ibm 2026-07-06 15:48:17 +05:30
  • 2ec3883df0 vulkan: fix 32-bit integer overflow in CEIL_DIV (llama/25245) hokanosekai 2026-07-06 10:35:57 +02:00
  • 9fe449cbd0 ggml-hip: enable -ffast-math for HIP builds (llama/23862) a-huk 2026-07-06 09:02:26 +02:00
  • ec6e0f2c29 ggml-cuda: optimize conv_transpose_1d indexing (llama/25310) adavyas 2026-07-05 20:49:06 -07:00
  • 38c6260238 abort if we see a multi buffer (llama/25276) Eve 2026-07-05 18:38:47 +00:00
  • f6f1252b05 ggml : fix tensor-parallel + -ncmoe crash on MoE models (llama/25028) liminfei-amd 2026-07-06 01:56:11 +08:00
  • 1222c01f0e ggml: Update VMM Pool allocation ggml-cuda.cu - Turing P2P access fix (fixes #24489) (llama/24491) Vexxie 2026-07-05 18:10:09 +01:00
  • 93cff36f15 cuda : concat implementation for quantized types (llama/25303) fairydreaming 2026-07-05 17:26:24 +02:00
  • 0a58b4b546 ggml : fix broken CPU concat implementation for quantized types (llama/25247) fairydreaming 2026-07-04 13:37:37 +02:00
  • b28289008c cuda: enable topk-moe fusion for 288 experts (llama/25267) Piotr Wilkin (ilintar) 2026-07-03 15:36:55 +02:00
  • 28ef70c68c Remove redundant CUDA copies after gated_delta_net. (llama/23940) Gaurav Garg 2026-07-03 14:36:29 +05:30
  • 759197a754 opencl: allow loading precompiled binary kernels from library (llama/23042) lhez 2026-07-01 10:29:22 -07:00
  • cae089f4a1 hexagon: flash attention rework (optimizations, accuracy improvements, etc) (llama/25085) Max Krasnyansky 2026-07-01 06:59:19 -07:00
  • 388a1fc962 CUDA: consistent use of __restrict__ + PDL for FA (llama/25185) Johannes Gäßler 2026-07-01 10:55:14 +02:00
  • dd1c9a552b ggml-cpu: add AVX2 optimization for nvfp4 dot product and use UE4M3 LUT (llama/23961) ragz4125 2026-07-01 13:01:20 +05:30
  • 9804f9736c opencl: initial q1_0 support (llama/25160) lhez 2026-06-30 21:43:20 -07:00
  • 52302312e6 cuda : prevent integer truncation and overflow errors when using KQ mask strides in flash_attn_mask_to_KV_max kernel (llama/24945) fairydreaming 2026-06-30 20:47:05 +02:00
  • 63285ff81f CUDA: fix get_rows_back for tables with more than 65535 rows (grid-y clamp + stride) (llama/25103) Matt Jallo 2026-06-30 05:16:24 -07:00
  • 2b35abc294 CUDA: fix Gemma E4B MTP FlashAttention (llama/25148) Johannes Gäßler 2026-06-30 14:06:54 +02:00
  • 899c7ed4ac vulkan: roll bk loop in matmul for asahi linux (llama/24663) Kevin Liu 2026-06-30 06:27:38 -04:00
  • 2ea7a752af HIP: use hipBLAS for dense prefill on gfx900, keep MMQ for MoE (llama/24588) zduford 2026-06-30 05:51:38 -04:00
  • b313fd3ce8 ggml-webgpu: add support for NVFP4 (llama/25143) Masashi Yoshimura 2026-06-30 17:20:04 +09:00
  • 7999638e04 Revert "sched : reintroduce less synchronizations during split compute (#20793)" (llama/25138) Oliver Simons 2026-06-30 02:41:45 +02:00
  • ba36235565 vulkan: use flops instead of weight tensor size for submission heuristic (llama/25005) Ruben Ortlam 2026-06-29 15:24:44 +02:00
  • 5f51daf516 opencl: flash attention improvement (llama/25069) Hongqiang Wang 2026-06-27 15:36:06 -07:00
  • 868367609f Added a cudaMemcpy2DAsync fast path to ggml_cuda_cpy (llama/25057) Gaurav Garg 2026-06-27 17:46:21 +05:30
  • a4ac1c0369 sycl : fix failed ut cases of norm (llama/25044) Neo Zhang 2026-06-27 17:13:43 +08:00
  • d54eb4fe2e vulkan: fix step operator for 0 input (llama/25036) Ruben Ortlam 2026-06-27 10:57:31 +02:00
  • e4234013e9 sched : reintroduce less synchronizations during split compute (llama/20793) Andreas Kieslinger 2026-06-26 16:18:30 +02:00
  • 3e97b98452 openvino: Update to OV 2026.2.1, self-contained release packages, operator improvements (llama/24974) Ravi Panchumarthy 2026-06-26 05:07:19 -07:00
  • c497d75e00 vulkan: opt mul_mat_vecq for mi50 (llama/22933) nullname 2026-06-26 19:49:24 +08:00
  • 1c8ea5fb92 vulkan: add INTEL_XE1 arch enum and enable coopmat1 on Intel Xe-LPG Plus (llama/24404) Jiang, Fish 2026-06-26 11:26:22 +00:00
  • 6fc7c33b4c whisper : expose internal VAD speech segments (#3916) Lin Xiaodong 2026-07-01 19:10:16 +08:00
  • 167d225f3a ci, examples : add package-lock.json to examples addon node (#3919) Daniel Bevenius 2026-07-01 09:04:21 +02:00
  • 5cad7abb1e whisper : map token timestamps to original time when VAD is enabled (#3910) Lin Xiaodong 2026-07-01 14:23:50 +08:00
  • 909307c8bd whisper : make voice_length() utf-8 aware for CJK (#3915) Lin Xiaodong 2026-07-01 14:21:35 +08:00
  • 0874de3e8e ci : set GGML_NATIVE=OFF for sanitize builds (#3920) Daniel Bevenius 2026-06-30 17:47:22 +02:00
  • 65dc3fece6 examples : Improved inference performance of Android example project (#3913) Yongmin Yoo 유용민 2026-06-30 19:12:04 +09:00
  • 0ae02cdb2c talk-llama : sync llama.cpp Georgi Gerganov 2026-06-26 15:06:05 +03:00
  • 0279b5386a sync : ggml Georgi Gerganov 2026-06-26 15:05:10 +03:00
  • 974db58c5f ggml : bump version to 0.15.3 (ggml/1550) Georgi Gerganov 2026-06-26 14:37:43 +03:00
  • df03e266f1 vulkan: Workaround compiler bug in conv2d coopmat2 path (llama/24924) Jeff Bolz 2026-06-26 04:53:32 -05:00
  • b564d5364a CUDA: add cublasSgemmBatched mapping for HIP/MUSA vendor headers (llama/25033) leonardHONG 2026-06-26 17:42:56 +08:00
  • 5e7ef7e4b0 ggml-cpu: fix SVE leftover path in ggml_vec_dot_f32 (llama/24699) Tarek Dakhran 2026-06-26 09:41:56 +02:00
  • 02be871218 sycl : clamp softmax input to avoid underflow (llama/24941) Jassieluo 2026-06-26 15:02:42 +08:00
  • 96996579f0 CUDA: batch out_prod broadcast (dps2>1) path with cublasSgemmBatched (llama/24426) leonardHONG 2026-06-26 13:51:25 +08:00
  • 7ce3616ae8 opencl: flush profiling batch at shutdown for incomplete batches (llama/25016) shaofeiqi 2026-06-25 18:48:24 -07:00
  • 9cad795918 CUDA: Various fixes to cpy.cu (llama/25000) Oliver Simons 2026-06-25 17:29:23 +02:00
  • 82686b518f ggml : address integer overflows in binary ops CUDA implementation (llama/24706) fairydreaming 2026-06-25 10:06:44 +02:00
  • ffd5b0b5c7 sycl : support --split-mode tensor (llama/24152) David Spruill 2026-06-25 01:35:21 -04:00
  • c4ba5e2e13 sycl : fix the failed UT cases of conv_3d (llama/24900) Neo Zhang 2026-06-25 13:27:58 +08:00
  • fe37d2e489 opencl: support non-contig rows in norm (llama/24965) lhez 2026-06-24 19:21:25 -07:00
  • e0a0f1256f hexagon: MUL_MAT and MUL_MAT_ID rework : 32x32 tiled weight repack, kernel-params, cached graphs (llama/24954) Max Krasnyansky 2026-06-24 12:14:25 -07:00
  • a201c20505 vulkan: allow reducing the graph submission batches to avoid timeouts (llama/24872) Wagner Bruna 2026-06-24 11:29:24 -03:00
  • f081d84a4e vulkan: fail the build when a shader fails to compile (llama/24450) liminfei-amd 2026-06-24 17:42:03 +08:00
  • 792da0ecdb vulkan: Apply bias before softmax in FA, to avoid overflow (llama/24909) Jeff Bolz 2026-06-23 22:34:00 -05:00
  • 3fc2c57a63 vulkan: support all backend tests for SQR/SQRT/SIN/COS/CLAMP/LEAKY_RELU/NORM (llama/24582) Jeff Bolz 2026-06-23 09:48:24 -05:00
  • 1fdbfb19eb vulkan: Support GET_ROWS_BACK (llama/24883) Jeff Bolz 2026-06-23 08:39:37 -05:00
  • c25eb26c32 vulkan: support CONV_3D (llama/24612) Jeff Bolz 2026-06-23 08:39:20 -05:00