ggerganov/llama.cpp

Pull Requests Commits

cont : adjust nsg

ggerganov committed 3 days ago

Verified cd3a4883

qwen4 : enable sparse attention

ggerganov committed 3 days ago

Verified 7916e96d

metal : add top-k radix implementation

ggerganov committed 3 days ago

Verified d33a6e32

tests : add perf cases for sparse flash attention prefill

ggerganov committed 3 days ago

Verified e9e6801c

metal : single-pass flash attention sparse index compaction

ggerganov committed 3 days ago

Verified 08d28dd8

cont : use sparse vec FA for prefill

ggerganov committed 3 days ago

Verified 552309b1

metal : fix sparse flash attention row addressing

ggerganov committed 3 days ago

Verified c31b682b

metal : support n_kv_max sparse mask hint in flash attention vec kernel

ggerganov committed 3 days ago

Verified 39a0be84

am17an committed 3 days ago

Verified 06a3b04b

CUDA + ggml: add sparse flash attention

am17an committed 3 days ago

Verified c2d23183

kv-cache : optimize restoring non-contiguous cells (#27991)

itsnotoger committed 3 days ago

Verified 2d8d612e

opencl: tune the quant paths for Intel Xe-LP GPUs to improve its TG and PP performance (#26438)

wanghqc committed 3 days ago

Verified 010be968

ui: copy the displayed text of grouped agentic responses (#27832)

ServeurpersoCom committed 3 days ago

Verified 774ee0e2

webgpu : avoid crash when offset is not multiple of 4 in WebGPU ggml_backend_tensor_get() implementation (#28045)

fairydreaming committed 3 days ago

Verified 8e53fcef

ROCm: add radix TOP_K for long rows (#27466)

jadenmach2 committed 3 days ago

Verified f8dbcd61

metal : add fa-vec tunings for M1 (#28078)

nikwen committed 4 days ago

Verified 5d4a3be2

CUDA: extend MOE fusion to specdec, earlier MOE glu fusion and topk-router fusion were restricted to 1 token (#27621)

ynankani committed 4 days ago

Verified 41ef91f7

sycl : Enhance to get the free memory of Intel GPU (#27968)

arthw committed 4 days ago

Verified a32af33d

ci : add check for unzip (#28082)

CISC committed 4 days ago

Verified 580e88d8

spec : fuse the DFlash encoder into the KV cache injection (#27310)

wjinxu committed 4 days ago

Verified 662a0b01

vulkan: tune mat-vec rows for batched inference on Strix Halo (#27909)

SimonTeixidor committed 4 days ago

Verified 2cdae802

ggml : add MUL_MAT to the list of ops that may need additional memory (for WebGPU) (#28071)

fairydreaming committed 4 days ago

Verified 557614e0

vulkan: top_k radix select for k >= 1024 for Qwen 3.8 Flash Next (#28032)

0cc4m committed 4 days ago

Verified daef7b68

hexagon: fix CPY fence bug (#28033)

yshsharke committed 4 days ago

Verified 9723942a

metal : add remaining Q4_1/Q5_0/Q5_1 fa-vec tunings for M2 (#28017)

ring2003 committed 4 days ago

Verified bd55e6aa

rpc: avoid serializing buffers from other servers (#26500)

hmirin committed 4 days ago

Verified a7cc83bb

ggml : fix ggml_backend_buft_get_alloc_size() guard (#28038)

ggerganov committed 4 days ago

Verified 6d1479c1

kv-cells: stop the sequence scan once all sequences are seen (#28011)

ServeurpersoCom committed 4 days ago

Verified 62acc89c

ggml: add SWIGLU_CLAMP (#27930)

am17an committed 4 days ago

Verified 0190529e

llama: improve TENSOR_READ_LAZY handling (#27837)

ngxson committed 4 days ago

Verified 25781383

Older