ggerganov/llama.cpp

Pull Requests Commits

HIP: Use mmq on MFMA devices for MUL_MAT_ID in cases where a lot of splits would be generated (#18202)

IMbackK committed 3 hours ago

Verified 4ffc47cb

model : Plamo3 support (#17304)

mmnga committed 6 hours ago

Verified 9c675c71

Revert "ggml-cuda: use CMAKE_CUDA_ARCHITECTURES if set when GGML_NATIVE=ON (#18413)" (#18426)

am17an committed 10 hours ago

Verified 07a0c4ba

rpc: fix segfault on invalid endpoint format (#18387)

o7si committed 12 hours ago

Verified 60f17f56

llama-fit-params: fix step size for last device (#18415)

JohannesGaessler committed 13 hours ago

Verified f8d561eb

github: update issue templates [no ci] (#18410)

JohannesGaessler committed 13 hours ago

Verified e59efe6a

mtmd: clarify that we no longer accept AI-generated PRs (#18406)

ngxson committed 14 hours ago

Verified cffa5c46

cmake: Added more x86_64 CPU backends when building with `GGML_CPU_ALL_VARIANTS=On` (#18186)

bberberov committed 15 hours ago

Verified 94de74e7

ggml-cuda: use CMAKE_CUDA_ARCHITECTURES if set when GGML_NATIVE=ON (#18413)

QDelta committed 21 hours ago

Verified 4fd59e84

opencl: allow resizing transpose buffers (#18384)

lhez committed 23 hours ago

Verified 08566977

llama-fit-params: fix overflow check (#18354)

JohannesGaessler committed 1 day ago

Verified a4bf3588

llama: fix magic number of 999 for GPU layers (#18266)

JohannesGaessler committed 1 day ago

Verified 026d2ad4

ggml-cuda: Use same regex for GGML_NATIVE=OFF (#18407)

am17an committed 1 day ago

Verified 06705fdc

llama_fit_params: return enum for fail vs. error (#18374)

JohannesGaessler committed 1 day ago

Verified a52dc60b

llama-fit-params: fix Gemma 3 calculation (#18372)

JohannesGaessler committed 1 day ago

Verified 9045c9af

vulkan: preprocess mul_mat_id experts and discard workgroups more quickly (#18352)

jeffbolznv committed 2 days ago

Verified c9ced491

vulkan: optimize decodeFuncB in coopmat2 mul_mat_id shader (#18349)

jeffbolznv committed 2 days ago

Verified 7ac89021

vulkan: Use BK=32 for coopmat2 mul_mat_id (#18332)

jeffbolznv committed 2 days ago

Verified 9bf20d8a

vulkan: small dequantization improvements (#18380)

netrunnereve committed 2 days ago

Verified cb999704

vulkan: Support UPSCALE w/antialias (#18327)

jeffbolznv committed 2 days ago

Verified b96b82fc

vulkan: handle rope with large number of rows (#18306)

jeffbolznv committed 2 days ago

Verified 10dc500b

server : fix crash when seq_rm fails for hybrid/recurrent models (#18391)

o7si committed 2 days ago

Verified 4893cc07

docs: added note for pre SYCL Intel hardware (#18016)

alosslessdev committed 2 days ago

Verified af3be131

CANN: implement the SSM_CONV operator (#17737)

0Marble committed 2 days ago

Verified b07cda68

ggml-cuda: fix regex for arch list (#18371)

am17an committed 3 days ago

Verified 85c40c9b

cuda: optimize cumsum cub path (#18362)

am17an committed 3 days ago

Verified 83b3b1c2

ggml-cuda: fix blackwell native builds (#18361)

am17an committed 3 days ago

Verified b0fb0f0a

CANN: Add support for CONV_TRANSPOSE_1D when kernel size > 255 (#17934)

Intellouis committed 3 days ago

Verified e68c19b0

ggml : optimize cuda cumsum fallback kernel (#18343)

Aadeshveer committed 3 days ago

Verified c54bba86

server: (router) add stop-timeout option (#18350)

ngxson committed 4 days ago

Verified f5acfb2f

Older