ggerganov/llama.cpp

Pull Requests Commits

vendor: apply deep nested json patch from upstream

ngxson committed 24 days ago

f0440d9e

CUDA: fix round issue, under MSVC the CPU and GPU agree (#30229)

KafKafrnZ committed 24 days ago

Verified 79e2e74e

graph : reorder get_rows for embeddings (#30160)

ggerganov committed 24 days ago

Verified 8e2d31e0

ui: Models Manager Follow-up Improvements (#30228)

allozaur committed 24 days ago

Verified baef3ed9

llama-bench: respect -fitc if bigger than required benchmark size (#28331)

MartinEmrich committed 24 days ago

Verified f39148a9

CUDA: Remove redundant CUDA copies after SSM_SCAN (#29807)

anavp-nvidia committed 24 days ago

Verified 50e3e3e4

opencl: fix kernel compilation for a6x GPUs (#30176)

jimmyzhouj committed 24 days ago

Verified 64df9183

model : use exact GELU for ModernBERT encoders (#30108)

boshjerns committed 24 days ago

Verified 6184e92c

chat : refactor API (#30210)

aldehir committed 25 days ago

Verified 8b543610

llama: keep the backend sampling graph static across ubatches (#30223)

ServeurpersoCom committed 25 days ago

Verified a518119d

ggml: fix OOB write in ggml_acc with negative offset (#30135)

Jasmine-tim committed 25 days ago

Verified 8ae38670

meta : handle host views (#30217)

ggerganov committed 25 days ago

Verified e60eff95

webgpu: use 2D workgroup dispatch for all the ops which use 1D dispatch (e.g., rms_norm) (#30219)

yomaytk committed 25 days ago

Verified ba6439a6

vulkan: fix rms_norm workgroup count overflow (#30145)

0cc4m committed 25 days ago

Verified 5e4878e9

convert : support compressed-tensor mixed-precision NVFP4 checkpoint (#28636)

ynankani committed 25 days ago

Verified 609290be

musa: drop mp_21 from the default architectures (#30203)

yeahdongcn committed 25 days ago

Verified 86a28353

ci: fix Models Backend webgpu by supporting GGML_OP_DUP (#30216)

ServeurpersoCom committed 25 days ago

Verified e94acad8

ci : set run-name for publish release (#30214)

CISC committed 25 days ago

Verified 5e1d7404

server: use port 9931 by default (#30159)

ngxson committed 25 days ago

Verified b42b7e6d

ui : add the models manager (#29583)

allozaur committed 25 days ago

Verified b013e56a

ci: fix HIP quality check by ignoring the 320/256 FA kernel spill (#30212)

ServeurpersoCom committed 25 days ago

Verified 2c564f43

metal: add the 128/96 flash attention kernels, accept every compiled pair (#30209)

ServeurpersoCom committed 25 days ago

Verified 1f8fa523

CUDA: pass src1 precision to host MMQ config helpers (#30168)

ynankani committed 25 days ago

Verified 8a1a9b51

hexagon: fix IM2COL patch-embed DMA ring overflow (#30189)

marktodorovich committed 25 days ago

Verified d4d82d67

sycl : Q5_K reorder-layout MMVQ and fused GLU (#29375)

NickM-27 committed 25 days ago

Verified 3d65c90d

Musa FWHT fix (#30167)

bri-prism committed 25 days ago

Verified de7fa0a3

CUDA: improve top-k algorithm selection (#28713)

praneshgo committed 25 days ago

Verified 71ad0590

model : fix DFlash output head sharing (#30111)

hthadicherla committed 25 days ago

Verified a11f57ba

CUDA: fix MMQ out-of-bounds reads (#29953)

JohannesGaessler committed 26 days ago

Verified fc9ce6b9

CUDA : looped PAD kernel for more than 65535 rows or slices (#30147)

pskrunner14 committed 26 days ago

Verified c35b6674

Older