ggerganov/llama.cpp

Pull Requests Commits

vulkan: make the radix top-k tie selection deterministic

ServeurpersoCom committed 21 minutes ago

03b8cc2f

mtmd: refactoring: remove clip.h, consolidate code into mtmd (#31865)

ngxson committed 17 hours ago

Verified 7d7f0287

cpu: enable SpacemiT IME2 MXFP4 dense repack (#31776)

gitGurugu committed 17 hours ago

Verified b65bf692

llamafile: impl bf16 and q8_0 vxe (#31866)

taronaeo committed 17 hours ago

Verified a3c7cf08

ggml : update CUDA/Metal radix Top-K to make tied selections deterministic (#31084)

sanmai committed 17 hours ago

Verified 57d04820

ggml-cuda : add d_state=64 fast path to SSM_SCAN kernel (#31818)

sargentmajord committed 18 hours ago

Verified 99820049

moe-cache : ensure CPU part does not wait for the cached experts (#31888)

sanmai committed 18 hours ago

Verified cf34d311

webgpu : add COUNT_EQUAL support (#31578)

LeON-Nie-code committed 22 hours ago

Verified a691b234

ci: move runners to Ubuntu 26.04 (#31529)

taronaeo committed 22 hours ago

Verified 2277e752

ggml : switch to F32 src1 in the CPU backend (#31854)

ggerganov committed 22 hours ago

Verified 12ab5979

llama, mtmd: follow the reference precision of ggml_prec_set_acc calls (#31851)

ServeurpersoCom committed 1 day ago

Verified d6be0655

speculative : do not apply the MoE cache to the draft model (#31497)

boshjerns committed 1 day ago

Verified a2b30183

opencl: improve q8_0 and MoE small-n verification for spec (#31831)

shaofeiqi committed 1 day ago

Verified c19043f6

common : sanitize invalid UTF-8 in PEG tool call arguments (#31861)

apuravmanhas committed 1 day ago

Verified 372dc5ca

server : downgrade per-request proxy logs to trace (#31859)

ggerganov committed 1 day ago

Verified f477f795

hexagon: support for automatic loading of libhtp without legacy env vars (#31781)

aparmp-quic committed 1 day ago

Verified 76a67730

sycl: exact rounding in CPY to Q4_1, Q5_1 and Q8_0 (#31465)

Torchit1 committed 1 day ago

Verified 12e9084d

sycl: refuse device buffers that would exceed VRAM (#31285)

KyleHagy committed 1 day ago

Verified 818ac6b0

hexagon: n_hvx and n_threads cleanup (#31857)

max-krasnyansky committed 1 day ago

Verified e9a2a44f

mtmd: move log-mel computation to cgraph (#31808)

ngxson committed 1 day ago

Verified a2654ec4

mtmd: support MageVL-Cowork model (#31819)

ngxson committed 1 day ago

Verified 9f0e5098

llama_moe_cache: run missed experts on CPU, add tuning script (#30966)

am17an committed 1 day ago

Verified e95426f1

sycl: add f16 support for SWIGLU_OAI (#30857)

mctylr-gh committed 1 day ago

Verified f7591a6a

metal : fix GROUP_NORM for ne03 > 1 and non-contiguous inputs (#31852)

ggerganov committed 1 day ago

Verified 80d2c0c2

server: visit parent slots before child slots (#30779)

harkanwalbedi committed 1 day ago

Verified 09566845

webgpu : Implement group norm kernel (#31490)

nikhilJain17 committed 2 days ago

Verified cf659f24

vulkan: Add 'readonly' to decodeBuf declarations. (#31833)

jeffbolznv committed 2 days ago

Verified 19381526

CUDA: fix shared memory bank conflicts in the lightning indexer (#31803)

aalramadan committed 2 days ago

Verified 5d7d4aec

llama : add missing link to pthread (#31785)

angt committed 2 days ago

Verified c0ab6e5c

opencl: add arange (#31829)

shaofeiqi committed 2 days ago

Verified 31424b7c

Older