ggerganov/llama.cpp

Pull Requests Commits

metal : avoid extra copy during set_tensor

ggerganov committed 11 hours ago

Verified e65c53e9

metal : disable host buffer to prevent races

ggerganov committed 12 hours ago

Verified afd95d20

metal : enable private buffers + add global device queue

ggerganov committed 12 hours ago

Verified 040844cf

metal : support both private and shared buffers

ggerganov committed 14 hours ago

Verified 7b59f0f7

metal : clean-up loose ends, ready for tests

ggerganov committed 19 hours ago

Verified 0926cb49

metal : remove broken implementation of GGML_OP_SET

ggerganov committed 19 hours ago

Verified f288225d

metal : restore .alloc_buffer for buffer_from_ptr_type

ggerganov committed 19 hours ago

Verified 7fc2b3d5

metal : create only metal buffers, no wrapping of host memory

ggerganov committed 20 hours ago

Verified 85aaf52b

metal : remove deprecated ggml_backend_metal_buffer_from_ptr

ggerganov committed 1 day ago

Verified d91ba85d

metal : fix batch size for MUL_MAT_ID

ggerganov committed 1 day ago

Verified bdff7729

cont : add comments, extend op offload, clean up

ggerganov committed 1 day ago

Verified c5637cf3

metal : make the backend async

ggerganov committed 1 day ago

Verified 97b96c1a

CUDA: fix GET_ROWS for large tensors (#15882)

JohannesGaessler committed 1 day ago

Verified 550cf726

contrib : add notes about merging PRs (#15881)

ggerganov committed 1 day ago

Verified c252ce67

requirements : update transformers/torch for Embedding Gemma (#15828)

danbev committed 1 day ago

Verified 70cd37db

model-conversion : add extra debugging support for model conversion (#15877)

pwilkin committed 1 day ago

Verified acc1b008

json : support `enum` values within `allOf` (#15830)

aldehir committed 1 day ago

Verified 7057faf6

media : add llama1 icon (#15878)

06kellyjac committed 1 day ago

Verified fe1c92cd

vulkan: sort graph to allow more parallel execution (#15850)

jeffbolznv committed 1 day ago

Verified e68aa10d

CUDA: generate_cu_files.py - add missing mxfp4 (#15880)

am17an committed 1 day ago

Verified 0a16bf52

chat : Deepseek V3.1 reasoning and tool calling support (OpenAI Style) (#15533)

createthis committed 1 day ago

Verified 88021565

server : bring back timings_per_token (#15879)

ngxson committed 1 day ago

Verified 56920f56

cuda : fix supports_op condition for get_rows when number of blocks is too large (#15868)

ggerganov committed 1 day ago

Verified b0d52998

metal : refactor + optimize (#15857)

ggerganov committed 1 day ago

Verified f28d4f4a

ggml: allow casting between f32 and i32 (#15783)

ngxson committed 1 day ago

Verified 9fcb29f2

CUDA: non-contiguous src0 not supported for PAD (#15869)

CISC committed 1 day ago

Verified 5ef22d28

convert : force setting sliding_window from original config (#15867)

danbev committed 1 day ago

Verified 233d773d

batched-bench : fix llama_synchronize usage during prompt processing (#15835)

ggerganov committed 1 day ago

Verified a885dcff

context : fix n_outputs during reserve (#15858)

ggerganov committed 1 day ago

Verified 663027fd

model : avoid ggml_cont_3d for fused QKV weights (#15662)

ggerganov committed 1 day ago

Verified cf0e3ba1

Older