ggml-org/llama.cpp

Pull Requests Commits

ggerganov committed 24 days ago

Verified e7fbfc9b

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

aendk committed 24 days ago

Verified d46bd7ef

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

aendk committed 24 days ago

Verified 07093368

Simplifies synchronizations to adhere to `saaasg` pattern.

aendk committed 24 days ago

Verified 1528c841

Corrects initialization of ggml_backend_sync_mode in

aendk committed 24 days ago

Verified ff28ae93

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

aendk committed 24 days ago

Verified 01d89f9b

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

aendk committed 24 days ago

Verified e74b070e

aendk committed 24 days 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 24 days ago

Verified d776354d

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

aendk committed 24 days ago

Verified 79a77277

Adds macro guards to allow compilation in non-CUDA builds

aendk committed 24 days ago

Verified 44e481bb

Exchanges synchronous copy with async copy function.

aendk committed 24 days ago

Verified 91c6026b

Adds function to relax sync requirements between input copies on

aendk committed 24 days ago

Verified 2ad0d391

Adds CPU-to-CUDA copy capability to

aendk committed 24 days ago

Verified dd9f1faf

metal : fix event synchronization in cpy_tensor_async (#19402)

ggerganov committed 24 days ago

Verified a554bdd7

fix output pattern

JohannesGaessler committed 25 days ago

02ee504f

re-use buffers + ggml contexts

JohannesGaessler committed 26 days ago

94c66557

unconditional peer access

JohannesGaessler committed 27 days ago

6ab02d09

add support for 4/8 GPUs

JohannesGaessler committed 27 days ago

2b705538

partial Vulkan fix

JohannesGaessler committed 28 days ago

b7fd1066

support for GPT-OSS, Qwen 3 MoE

JohannesGaessler committed 28 days ago

a630b27d

ggml: backend-agnostic tensor parallelism

JohannesGaessler committed 29 days ago

39b96f8f

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

ggerganov committed 29 days ago

Verified 22cae832

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

jeffbolznv committed 29 days ago

Verified 449ec2ab

benches : update models + numbers (#19359)

ggerganov committed 30 days ago

Verified 3795cc1e

docker : fix vulkan build (#19352)

CISC committed 30 days ago

Verified b828e18c

vendor : update BoringSSL to 0.20260204.0 (#19333)

angt committed 30 days ago

Verified a4ea7a18

metal : add diag (#19330)

ggerganov committed 30 days ago

Verified 7a4f97d1

vulkan: fix GPU deduplication logic. (#19222)

okuvshynov committed 30 days ago

Verified a498c75a

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

jeffbolznv committed 30 days ago

Verified 3409ab84

Older