sync : ggml#3478

Merged

ggerganov merged 23 commits intomasterfrom

sync-ggml-25-10-14

Oct 15, 2025

Commits on Oct 14, 2025

CUDA: faster tile FA, add oob checks, more HSs (llama/16492)

JohannesGaessler
authored and
ggerganov
committed
ggml: Correct SVE implementation in ggml_vec_dot_f16_unroll (llama/16518)

authored and
ggerganov
committed
ggml : Fix FP16 ELU positive branch (llama/16519)

authored and
ggerganov
committed
fix UT fault cases: count-equal, argsort, pad OPs (llama/16521)

authored and
ggerganov
committed
metal : add opt_step_adamw and op_sum (llama/16529)

cern1710
authored and
ggerganov
committed
CANN: Update several operators to support FP16 data format (llama/16251)

authored and
ggerganov
committed
ggml : fix scalar path for computing norm (llama/16558)
ggerganov
committed
metal: add support for opt_step_sgd (llama/16539)

cern1710
authored and
ggerganov
committed
CANN: fix CPU memory leak in CANN backend (llama/16549)

noemotiovon
authored and
ggerganov
committed
ggml : fix build broken with -march=armv9-a on MacOS (llama/16520)

authored and
ggerganov
committed
CUDA: fix numerical issues in tile FA kernel (llama/16540)

JohannesGaessler
authored and
ggerganov
committed
opencl: fix build targeting CL 2 (llama/16554)

lhez
authored and
ggerganov
committed
metal : FA support F32 K and V and head size = 32 (llama/16531)
ggerganov
committed
cuda : remove legacy copy-op pointer indirection code (llama/16485)

anavp-nvidia
authored and
ggerganov
committed
CUDA: add fp kernel for larger batch size MoE (llama/16512)

am17an
authored and
ggerganov
committed
CUDA: use fastdiv + ggml_cuda_mad for mmvf (llama/16557)

am17an
authored and
ggerganov
committed
CUDA: enable FA for FP32 KV cache (llama/16546)

JohannesGaessler
authored and
ggerganov
committed
vulkan: Improve build time for MSVC (llama/16545)

jeffbolznv
authored and
ggerganov
committed
vulkan: Support FA with K/V in F32 (llama/16543)

jeffbolznv
authored and
ggerganov
committed
CUDA + openCL: fix bug in accessing rms_norm->src while doing fusion (llama/16577)

am17an
authored and
ggerganov
committed
vulkan: Add ACC_TYPE_VEC2 implementation (llama/16203)

authored and
ggerganov
committed
sync : ggml
ggerganov
committed
talk-llama : sync llama.cpp
ggerganov
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

sync : ggml#3478

sync : ggml#3478
ggerganov merged 23 commits intomasterfrom
sync-ggml-25-10-14

Commits on Oct 14, 2025

CUDA: faster tile FA, add oob checks, more HSs (llama/16492)

ggml: Correct SVE implementation in ggml_vec_dot_f16_unroll (llama/16518)

ggml : Fix FP16 ELU positive branch (llama/16519)

fix UT fault cases: count-equal, argsort, pad OPs (llama/16521)

metal : add opt_step_adamw and op_sum (llama/16529)

CANN: Update several operators to support FP16 data format (llama/16251)

ggml : fix scalar path for computing norm (llama/16558)

metal: add support for opt_step_sgd (llama/16539)

CANN: fix CPU memory leak in CANN backend (llama/16549)

ggml : fix build broken with -march=armv9-a on MacOS (llama/16520)

CUDA: fix numerical issues in tile FA kernel (llama/16540)

opencl: fix build targeting CL 2 (llama/16554)

metal : FA support F32 K and V and head size = 32 (llama/16531)

cuda : remove legacy copy-op pointer indirection code (llama/16485)

CUDA: add fp kernel for larger batch size MoE (llama/16512)

CUDA: use fastdiv + ggml_cuda_mad for mmvf (llama/16557)

CUDA: enable FA for FP32 KV cache (llama/16546)

vulkan: Improve build time for MSVC (llama/16545)

vulkan: Support FA with K/V in F32 (llama/16543)

CUDA + openCL: fix bug in accessing rms_norm->src while doing fusion (llama/16577)

vulkan: Add ACC_TYPE_VEC2 implementation (llama/16203)

sync : ggml

talk-llama : sync llama.cpp

sync : ggml#3478ggerganov merged 23 commits intomasterggml-org/whisper.cpp:masterfrom sync-ggml-25-10-14ggml-org/whisper.cpp:sync-ggml-25-10-14Copy head branch name to clipboard

Commits

Commits on Oct 14, 2025

sync : ggml#3478
ggerganov merged 23 commits intomasterfrom
sync-ggml-25-10-14