ggml-org/llama.cpp

Pull Requests Commits

ggerganov committed 3 days ago

Verified 6c8a0457

ngram-map : take into account the input can become shorter

ggerganov committed 4 days ago

Verified 003c9035

ngram-map : fix uninitialized values

ggerganov committed 4 days ago

Verified 9f8401a5

common : rename speculative.draftless_type -> speculative.type

ggerganov committed 4 days ago

Verified bc338380

Merge branch 'master' into pr/18471

ggerganov committed 4 days ago

Verified 351e798b

[CUDA] Reduce CPU-side stalls due to the CUDA command buffer being full (#19042)

gaugarg-nv committed 4 days ago

Verified a83c73a1

common : clarify HTTPS build options in error message (#19103)

danbev committed 4 days ago

Verified fc3cdf32

ggml-cpu: Enable FP16 MMA kernels on PPC (#19060)

shalinib-ibm committed 4 days ago

Verified 7afdfc9b

opencl: add flattened q6_K mv (#19054)

lhez committed 4 days ago

Verified 94eeb596

CUDA: fix padding of GQA to power of 2 in FA (#19115)

JohannesGaessler committed 4 days ago

Verified b0311c16

CODEOWNERS: add common/ngram-map.* (#18471)

srogmann committed 4 days ago

dd23149d

minor: comments

srogmann committed 4 days ago

72f416e9

graph : fix nkvo offload with FA (#19105)

ggerganov committed 4 days ago

Verified 8f80d1b2

ci : use new 1vCPU runner for lightweight jobs (#19107)

CISC committed 4 days ago

Verified 142cbe2a

minor : cleanup + fix build

ggerganov committed 5 days ago

Verified 1f8d3666

common : better names

ggerganov committed 5 days ago

Verified a3300937

minor : cleanup

ggerganov committed 5 days ago

Verified f895bca7

model : add correct type for GLM 4.7 Flash (#19106)

ggerganov committed 5 days ago

Verified 56f3ebf3

common: print performance in spec decoding

srogmann committed 5 days ago

fd4d803c

doc: (draftless) speculative decoding

srogmann committed 5 days ago

288ab505

spec: remove --spec-config

srogmann committed 5 days ago

8ea068e5

CUDA: faster FA for GQA > 1 but not power of 2 (#19092)

JohannesGaessler committed 5 days ago

Verified 0c21677e

ggerganov committed 5 days ago

Verified 9ac88176

metal : fix recommendedMaxWorkingSetSize availability on legacy iOS/macOS (#19088)

ccbinn committed 5 days ago

Verified 0440bfd1

convert : yield Gemma3N custom_map tensors directly (#19091)

CISC committed 5 days ago

Verified 0bf56369

spec : refactor

ggerganov committed 5 days ago

Verified 924517dd

common: cleanup (use common_speculative_state_draft)

srogmann committed 5 days ago

af382c38

ggml-cpu: Use tiled FA for prompt-processing (#19012)

am17an committed 5 days ago

Verified bcb43163

kv-cache : support V-less cache (#19067)

ggerganov committed 5 days ago

Verified d9c6ce46

convert : fix Gemma3N, GraniteMoe and Ernie4.5Moe (#19084)

CISC committed 6 days ago

Verified 70d86082

Older