ggml-org/llama.cpp

Pull Requests Commits

ggerganov committed 1 day ago

Verified e7fbfc9b

Apply suggestion from @ggerganov (src->buffer to buf_src) v2

aendk committed 1 day ago

Verified d46bd7ef

Apply suggestion from @ggerganov (src->buffer to buf_src)

aendk committed 1 day ago

Verified 07093368

Simplifies synchronizations to adhere to `saaasg` pattern.

aendk committed 1 day ago

Verified 1528c841

Corrects initialization of ggml_backend_sync_mode in

aendk committed 1 day ago

Verified ff28ae93

Reintroduces stricter check for CPU->CUDA backend async copy via

aendk committed 1 day ago

Verified 01d89f9b

Makes opt-in to relax use of explicit syncs more general. Backends like

aendk committed 1 day ago

Verified e74b070e

aendk committed 1 day ago

Verified b7376c3e

Relax requirement of checks in async CUDA copies from backend and buffer type to just buffer type, to avoid linking issues

aendk committed 1 day ago

Verified d776354d

Reworked backend detection in ggml-backend.cpp to avoid linking

aendk committed 1 day ago

Verified 79a77277

Adds macro guards to allow compilation in non-CUDA builds

aendk committed 1 day ago

Verified 44e481bb

Exchanges synchronous copy with async copy function.

aendk committed 1 day ago

Verified 91c6026b

Adds function to relax sync requirements between input copies on

aendk committed 1 day ago

Verified 2ad0d391

Adds CPU-to-CUDA copy capability to

aendk committed 1 day ago

Verified dd9f1faf

metal : fix event synchronization in cpy_tensor_async (#19402)

ggerganov committed 1 day ago

Verified a554bdd7

fix output pattern

JohannesGaessler committed 3 days ago

02ee504f

re-use buffers + ggml contexts

JohannesGaessler committed 4 days ago

94c66557

unconditional peer access

JohannesGaessler committed 4 days ago

6ab02d09

add support for 4/8 GPUs

JohannesGaessler committed 5 days ago

2b705538

partial Vulkan fix

JohannesGaessler committed 5 days ago

b7fd1066

support for GPT-OSS, Qwen 3 MoE

JohannesGaessler committed 6 days ago

a630b27d

ggml: backend-agnostic tensor parallelism

JohannesGaessler committed 7 days ago

39b96f8f

metal : adaptive CPU/GPU interleave based on number of nodes (#19369)

ggerganov committed 7 days ago

Verified 22cae832

vulkan: Preprocess FA mask to detect all-neg-inf and all-zero. (#19281)

jeffbolznv committed 7 days ago

Verified 449ec2ab

benches : update models + numbers (#19359)

ggerganov committed 7 days ago

Verified 3795cc1e

docker : fix vulkan build (#19352)

CISC committed 7 days ago

Verified b828e18c

vendor : update BoringSSL to 0.20260204.0 (#19333)

angt committed 7 days ago

Verified a4ea7a18

metal : add diag (#19330)

ggerganov committed 7 days ago

Verified 7a4f97d1

vulkan: fix GPU deduplication logic. (#19222)

okuvshynov committed 7 days ago

Verified a498c75a

vulkan: Set k_load_shmem to false when K is too large (#19301)

jeffbolznv committed 7 days ago

Verified 3409ab84

Older