sync : llama.cpp#1327

Merged

ggerganov merged 41 commits intomasterfrom

sync-llama.cpp-25-08-13

Aug 14, 2025

Commits on Aug 13, 2025

CUDA: use mma FA kernel for gqa > 4 on RTX 4000 (llama/15035)

JohannesGaessler
authored and
ggerganov
committed
opencl: fix adreno compiler detection logic (llama/15029)

lhez
authored and
ggerganov
committed
vulkan: Use coopmat2 for conv2d (llama/14982)

jeffbolznv
authored and
ggerganov
committed
vulkan: fix build when using glslang that does not support coopmat2 (llama/15062)

jeffbolznv
authored and
ggerganov
committed
cmake: Add GGML_BACKEND_DIR option (llama/15074)

ckastner
authored and
ggerganov
committed
sycl: fix mul_mat selection (llama/15092)

Rbiessy
authored and
ggerganov
committed
llama : add gpt-oss (llama/15091)

committed
CANN: add support for ACL Graph (llama/15065)

noemotiovon
authored and
ggerganov
committed
ggml : fix fallback to CPU for ununsupported ops (llama/15118)

slaren
authored and
ggerganov
committed
opencl: add `swiglu_oai` and `add_id` (llama/15121)

lhez
authored and
ggerganov
committed
fix profiling crash (llama/15072)

rmatif
authored and
ggerganov
committed
CUDA: GEMM for FP32/FP16/BF16 and ne11 <= 16 (llama/15131)

JohannesGaessler
authored and
ggerganov
committed
ggml: Skip backend library linking code when GGML_BACKEND_DL=ON (llama/15094)

ckastner
authored and
ggerganov
committed
HIP: add cmake option to enable compiler output of kernel resource usage metrics (llama/15103)

IMbackK
authored and
ggerganov
committed
vulkan: Add env var to disable host visible vidmem (llama/15109)

jeffbolznv
authored and
ggerganov
committed
vulkan: support fattn sinks (llama/15126)

jeffbolznv
authored and
ggerganov
committed
opencl: support sink in `soft_max` (attn sinks) (llama/15152)

lhez
authored and
ggerganov
committed
CUDA: attention sinks for mma FlashAttention (llama/15157)

JohannesGaessler
authored and
ggerganov
committed
ggml : fix field name when new ggml_backend (llama/14944)

aisk
authored and
ggerganov
committed
gguf-py : add Numpy MXFP4 de/quantization support (llama/15111)

compilade
authored and
ggerganov
committed
CUDA: add attention sinks for tile and wmma (llama/15178)

am17an
authored and
ggerganov
committed
cuda: refactored ssm_scan and use CUB (llama/13291)

Your-Cheese
authored and
ggerganov
committed
kleidiai: fix unsigned overflow bug (llama/15150)

chaxu01
authored and
ggerganov
committed
CANN: Add broadcast for softmax and FA (llama/15208)

hipudding
authored and
ggerganov
committed
musa: fix failures in test-backend-ops for mul_mat_id op (llama/15236)

yeahdongcn
authored and
ggerganov
committed
CANN: GGML_OP_CPY optimization (llama/15070)

noemotiovon
authored and
ggerganov
committed
CUDA cmake: add `-lineinfo` for easier debug (llama/15260)

am17an
authored and
ggerganov
committed
opencl: allow mixed f16/f32 `add` (llama/15140)

rmatif
authored and
ggerganov
committed
sycl: Fix and disable more configurations of mul_mat (llama/15151)

Rbiessy
authored and
ggerganov
committed
HIP: disable sync warp shuffel operators from clr amd_warp_sync_functions.h (llama/15273)

IMbackK
authored and
ggerganov
committed
ggml-rpc: chunk send()/recv() to avoid EINVAL for very large tensors over RPC (macOS & others) (llama/15188)

authored and
ggerganov
committed
CUDA: Optimize `reduce_rows_f32` kernel, leading up to 25x perf improvement on kernel-level and 10% perf increase for Gemma3n (llama/15132)

ORippler
authored and
ggerganov
committed
ggml : repack block_iq4_nlx8 (llama/14904)
ggerganov
committed
ggml : update `ggml_rope_multi` (llama/12665)

foldl
and
ggerganov
committed
sync : llama.cpp
ggerganov
committed

Commits on Aug 14, 2025