sync : ggml#3470

Merged

ggerganov merged 36 commits intomasterfrom

sync-ggml-25-11-12

Oct 12, 2025

Commits on Oct 12, 2025

opencl: support ne3 in get_rows (llama/15866)

lhez
authored and
ggerganov
committed
ggml webgpu: support for rope,div,sub,glu,scale,cont operators (llama/16187)

reeselevine
and
ggerganov
committed
opencl: support pad_ext (llama/15888)

lhez
authored and
ggerganov
committed
vulkan: make ggml_vk_default_dispatcher support older vulkan headers (llama/16345)

netrunnereve
authored and
ggerganov
committed
HIP: Disable ROCWMMA fattn on CDNA when compiled against ROCWMMA 2.0.0 (llama/16221)

IMbackK
authored and
ggerganov
committed
musa: update compile flags (llama/16265)

yeahdongcn
authored and
ggerganov
committed
model : Apertus model implementation (llama/15852)

committed
ggml webgpu: add support for soft_max, optimize rms_norm (llama/16357)

reeselevine
and
ggerganov
committed
vulkan: in flash attention, bounds check against nem1 (don't rely on GGML_KQ_MASK_PAD) (llama/16316)

jeffbolznv
authored and
ggerganov
committed
vulkan: Fix FA coopmat1 invalid array indexing (llama/16365)

jeffbolznv
authored and
ggerganov
committed
vulkan: Replace uses of maxMemoryAllocationSize and VK_WHOLE_SIZE (llama/16354)

jeffbolznv
authored and
ggerganov
committed
ggml : fix graph reallocation with multiple chunks (llama/16396)

Acly
authored and
ggerganov
committed
metal : fix loop bound in ggml_mem_ranges (llama/16412)
ggerganov
committed
vulkan : incremental shader builds (llama/16341)

authored and
ggerganov
committed
rpc : add support for multiple devices (llama/16276)

rgerganov
authored and
ggerganov
committed
rpc : check src buffer when copying tensor (llama/16421)

rgerganov
authored and
ggerganov
committed
vulkan: use a more appropriate amount of threads when generating shaders (llama/16418)

netrunnereve
authored and
ggerganov
committed
ggml webgpu: actually add softmax, fix rms_norm offset (llama/16400)

reeselevine
authored and
ggerganov
committed
ggml-cpu : fix leftover handling in ggml_vec_scale_f32 for SVE (llama/16443)

danbev
authored and
ggerganov
committed
ggml : fix unaligned access in AMX code (llama/16315)
ggerganov
committed
metal : various optimizations + refactoring (llama/16446)
ggerganov
committed
tests : add -INF blocks to the KQ mask in the FA tests (llama/16380)

ggerganov
and
jeffbolznv
committed
metal : add support for non-padded FA KV (llama/16148)
ggerganov
committed
ggml webgpu: profiling, CI updates, reworking of command submission (llama/16452)

reeselevine
authored and
ggerganov
committed
metal : mark FA blocks (llama/16372)
ggerganov
committed
Disable CUDA host buffers on integrated GPUs (llama/16308)

ai-fonsi
authored and
ggerganov
committed
refactor soft_max, add soft_max_back (llama/16472)

authored and
ggerganov
committed
kleidiai: kernel interface refactoring (llama/16460)

chaxu01
authored and
ggerganov
committed
CANN: Improve ACL graph matching (llama/16166)

noemotiovon
authored and
ggerganov
committed
cpu : optimize the ggml NORM operation (llama/15953)

committed
cmake : Dont define XOPENSOURCE on AIX (llama/16481)

mehendarkarprajwal
authored and
ggerganov
committed
cuda : avoid initializing unused devices (llama/16510)

slaren
authored and
ggerganov
committed
metal : fix mul-mm condition + fix mul-mv permuted kernels (llama/16494)
ggerganov
committed
sync : ggml
ggerganov
committed
talk-llama : sync llama.cpp
ggerganov
committed
bench : update [no ci]
ggerganov
committed