ggerganov/llama.cpp

Pull Requests Commits

mimo2 : expose t_layer_inp for the DFlash drafter

ggerganov committed 27 days ago

Verified 52cfe0f4

convert : handle dict/list-wrapped mask_embedding.pt in DFlash drafter

ggerganov committed 27 days ago

1ea03b57

convert : add partial rotary, value scale and mask embd to DFlash drafter

ggerganov committed 27 days ago

50aa599b

vulkan: fuse qwen4exp's SCALE -> SIGMOID -> SCALE -> hc_post chain (#29520)

fxgsell committed 27 days ago

Verified 03a667aa

HIP: fix template skip for DKQ > 256 mfma kernels (#29559)

IMbackK committed 27 days ago

Verified c2a9e160

metal: support left and circular padding in GGML_OP_PAD (#29561)

ServeurpersoCom committed 27 days ago

Verified 4364bf72

context : do not re-reserve the scheduler when toggling causal_attn (#28751)

sihanyu03 committed 27 days ago

Verified ed7ac35e

Enables Windows ARM64 build with MSVC cl.exe (#28362)

sarahwu185 committed 27 days ago

Verified 0c6a6a7c

tests : fix ggml init (#29554)

ggerganov committed 27 days ago

Verified 81ef10ea

vulkan: fix wrong results when a mul_mat reads a slice of a larger cache (#28956)

ServeurpersoCom committed 27 days ago

Verified 52624716

server : allow RANK pooling batch splitting for causal LLM rerankers (ie. Qwen3 and Qwen3-VL) (#28876)

timothywang21 committed 28 days ago

Verified 4da63377

common : avoid side effects around params parsing (#29537)

ggerganov committed 28 days ago

Verified a97cce86

common : make string_split<T> throw on invalid values (#29518)

angt committed 28 days ago

Verified 136887b6

convert : export YaRN scaling parameters for PLaMo-3 (#29528)

tokinasin committed 28 days ago

Verified 9adc7f42

ci : bump ty to 0.0.84 (#29529)

CISC committed 28 days ago

Verified 6fd50a40

jinja : add support for dict builtin (#29477)

CISC committed 28 days ago

Verified 33c923db

opencl: refine bin kernel loading condition (#29503)

lhez committed 28 days ago

Verified c9064dde

sycl: FWHT kernels for block widths above 512 (#29243)

bri-prism committed 28 days ago

Verified c8296709

CUDA: tune fp16 tile FlashAttention configs for head sizes 40-112 (#26289)

animeshsri14 committed 28 days ago

Verified 36d7b083

HIP: Enable fattn-mma kernel on cdna for dkq > 256 for large batch sizes (#28907)

IMbackK committed 28 days ago

Verified 2ebd9ae6

vulkan: fix argsort kernel selection for Adreno (#29469)

0cc4m committed 28 days ago

Verified cea74625

common : throw instead of abort on grammar without llguidance (#29516)

angt committed 28 days ago

Verified da6c28eb

RPC: use RDMA completion channel to not spin (#29440)

am17an committed 28 days ago

Verified d7fb90e8

ci : enable GGML_SCHED_DEBUG_REALLOC=1 for ctest workflows (#29514)

ggerganov committed 28 days ago

Verified 7fb2b082

llama-bench : fix OOB access of hf_file (#29515)

angt committed 28 days ago

Verified 187664b5

hrm : fix layer placement of `z_l_init` weight (#29512)

ggerganov committed 28 days ago

Verified 85ca3b52

hexagon: support tiled Q4_0 and Q8_0 GET_ROWS (#29511)

kurquhar committed 28 days ago

Verified 7ac59a6e

hexagon: support for backend sampler (#29502)

max-krasnyansky committed 28 days ago

Verified 2b129ccf

cuda: support Nemotron 3 Puzzle state size 96 for ssm scan (#28717)

anavp-nvidia committed 29 days ago

Verified 95887577

musa: build the docker images from the PH1 MUSA SDK image (#29481)

yeahdongcn committed 29 days ago

Verified 694ec235

Older