sync : ggml#3912

Merged

ggerganov merged 30 commits into

masterfrom

sync-ggml-26-06-26

Jun 26, 2026

Commits on Jun 26, 2026

mtmd, arg: fix utf8 handling on windows (llama/24779)

ngxson
authored and
ggerganov
committed
ggml-webgpu: add adapter toggles for F16 on Vulkan + NVIDIA

yomaytk
authored and
ggerganov
committed
ggml : optimize AMX (llama/24806)

angt
authored and
ggerganov
committed
fix(hexagon): use padded stride for ssm-conv weights (llama/24470)

BiReRa
authored and
ggerganov
committed
support bf16 on bin_bcast OP and unary OPs (llama/24838)

arthw
authored and
ggerganov
committed
opencl: q8_0 gemv precision improvement (llama/24923)

shawngu-quic
authored and
ggerganov
committed
ggml-webgpu: improve MTP inference by using mat-vec path for small batches (llama/24811)

yomaytk
authored and
ggerganov
committed
vulkan: link ggml-cpu when GGML_VULKAN_CHECK_RESULTS / RUN_TESTS are enabled (llama/24444)

Detensable
authored and
ggerganov
committed
vulkan: make mul_mm ALIGNED a spec constant (llama/24689)

jeffbolznv
authored and
ggerganov
committed
vulkan: support CONV_3D (llama/24612)

jeffbolznv
authored and
ggerganov
committed
vulkan: Support GET_ROWS_BACK (llama/24883)

jeffbolznv
authored and
ggerganov
committed
vulkan: support all backend tests for SQR/SQRT/SIN/COS/CLAMP/LEAKY_RELU/NORM (llama/24582)

jeffbolznv
authored and
ggerganov
committed
vulkan: Apply bias before softmax in FA, to avoid overflow (llama/24909)

jeffbolznv
authored and
ggerganov
committed
vulkan: fail the build when a shader fails to compile (llama/24450)

liminfei-amd
authored and
ggerganov
committed
vulkan: allow reducing the graph submission batches to avoid timeouts (llama/24872)

wbruna
authored and
ggerganov
committed
hexagon: MUL_MAT and MUL_MAT_ID rework : 32x32 tiled weight repack, kernel-params, cached graphs (llama/24954)

max-krasnyansky
authored and
ggerganov
committed
opencl: support non-contig rows in norm (llama/24965)

lhez
authored and
ggerganov
committed
sycl : fix the failed UT cases of conv_3d (llama/24900)

arthw
authored and
ggerganov
committed
sycl : support --split-mode tensor (llama/24152)

Spruill-1
authored and
ggerganov
committed
ggml : address integer overflows in binary ops CUDA implementation (llama/24706)

authored and
ggerganov
committed
CUDA: Various fixes to `cpy.cu` (llama/25000)

ORippler
authored and
ggerganov
committed
opencl: flush profiling batch at shutdown for incomplete batches (llama/25016)

shaofeiqi
authored and
ggerganov
committed
CUDA: batch out_prod broadcast (dps2>1) path with cublasSgemmBatched (llama/24426)

leonardHONG
authored and
ggerganov
committed
sycl : clamp softmax input to avoid underflow (llama/24941)

Jassieluo
authored and
ggerganov
committed
ggml-cpu: fix SVE leftover path in ggml_vec_dot_f32 (llama/24699)

tdakhran
and
ggerganov
committed
CUDA: add cublasSgemmBatched mapping for HIP/MUSA vendor headers (llama/25033)

leonardHONG
authored and
ggerganov
committed
vulkan: Workaround compiler bug in conv2d coopmat2 path (llama/24924)

jeffbolznv
authored and
ggerganov
committed
ggml : bump version to 0.15.3 (ggml/1550)
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#3912

sync : ggml#3912
ggerganov merged 30 commits into
masterfrom
sync-ggml-26-06-26

Commits on Jun 26, 2026

mtmd, arg: fix utf8 handling on windows (llama/24779)

ggml-webgpu: add adapter toggles for F16 on Vulkan + NVIDIA

ggml : optimize AMX (llama/24806)

fix(hexagon): use padded stride for ssm-conv weights (llama/24470)

support bf16 on bin_bcast OP and unary OPs (llama/24838)

opencl: q8_0 gemv precision improvement (llama/24923)

ggml-webgpu: improve MTP inference by using mat-vec path for small batches (llama/24811)

vulkan: link ggml-cpu when GGML_VULKAN_CHECK_RESULTS / RUN_TESTS are enabled (llama/24444)

vulkan: make mul_mm ALIGNED a spec constant (llama/24689)

vulkan: support CONV_3D (llama/24612)

vulkan: Support GET_ROWS_BACK (llama/24883)

vulkan: support all backend tests for SQR/SQRT/SIN/COS/CLAMP/LEAKY_RELU/NORM (llama/24582)

vulkan: Apply bias before softmax in FA, to avoid overflow (llama/24909)

vulkan: fail the build when a shader fails to compile (llama/24450)

vulkan: allow reducing the graph submission batches to avoid timeouts (llama/24872)

hexagon: MUL_MAT and MUL_MAT_ID rework : 32x32 tiled weight repack, kernel-params, cached graphs (llama/24954)

opencl: support non-contig rows in norm (llama/24965)

sycl : fix the failed UT cases of conv_3d (llama/24900)

sycl : support --split-mode tensor (llama/24152)

ggml : address integer overflows in binary ops CUDA implementation (llama/24706)

CUDA: Various fixes to `cpy.cu` (llama/25000)

opencl: flush profiling batch at shutdown for incomplete batches (llama/25016)

CUDA: batch out_prod broadcast (dps2>1) path with cublasSgemmBatched (llama/24426)

sycl : clamp softmax input to avoid underflow (llama/24941)

ggml-cpu: fix SVE leftover path in ggml_vec_dot_f32 (llama/24699)

CUDA: add cublasSgemmBatched mapping for HIP/MUSA vendor headers (llama/25033)

vulkan: Workaround compiler bug in conv2d coopmat2 path (llama/24924)

ggml : bump version to 0.15.3 (ggml/1550)

sync : ggml

talk-llama : sync llama.cpp