ggerganov/llama.cpp

Pull Requests Commits

metal : remove the in-kernel dequantization from the FA kernels

ggerganov committed 20 days ago

Verified 1134adc9

test : make the FA V-is-view-of-K case a test case parameter

ggerganov committed 20 days ago

Verified 242dbc8c

metal : dequantize quantized KV to F16 before flash attention (#27390)

ggerganov committed 20 days ago

Verified 70aff252

Revert "tensor-split meta backend fixes (#26502)" (#27433)

ggerganov committed 20 days ago

Verified f20395da

ggml: fix backend split scheduler race condition (#26040)

0cc4m committed 20 days ago

Verified 84979813

convert: fix get block count error for Nemotron 3 Ultra (#27101)

rockchenn committed 20 days ago

Verified a3b1effc

ggml-cuda: provide static workspace for cuBLAS handles (#26574)

heislera763 committed 20 days ago

Verified d9b6be07

graph : create V as a view of K in the k_iswa build_attn (#27392)

ggerganov committed 20 days ago

Verified 929d47a3

spec : avoid binding reference to null pointer (#27404)

ggerganov committed 20 days ago

Verified f466cfa3

vulkan : add source groups for shaders (#26666)

mtavenrath committed 20 days ago

Verified 2cfdb5fc

opencl: make the MoE expert scatter deterministic (#26464)

wanghqc committed 20 days ago

Verified 9ee9fc04

tensor-split meta backend fixes (#26502)

max-krasnyansky committed 21 days ago

Verified d59d455f

hexagon: fix FA HMX queue ordering and pack the rescale D matrices (#27042)

njsyw1997 committed 21 days ago

Verified 990e3bfe

opencl: port fused ssm_scan kernel (Mamba-2, d_state in {128, 256}) to GPU (#26439)

wanghqc committed 21 days ago

Verified b062ba73

ggml-cpu: gate __fp16 on __ARM_FP16_FORMAT_IEEE (#26860)

ServeurpersoCom committed 21 days ago

Verified cd644c39

server: refactor sleep handling, allow access /metrics during sleep (#27376)

ngxson committed 21 days ago

Verified 947fd9bb

server : make models endpoints private when authentication is enabled (#26347)

s0mecode committed 21 days ago

Verified ee0ea03a

vulkan : dequant q8_0 KV once in coopmat1 (#25494)

Nathanw1014 committed 21 days ago

Verified dc72703f

vulkan: add null checks in ggml_vk_queue_command_pools_cleanup (#27353)

jetsontan committed 21 days ago

Verified b95502ba

Revert "common: share thread pools when `n_threads` differ (#27138)" (#27337)

nikwen committed 21 days ago

Verified 3e734467

model : GraniteSWAForCausalLM / GraniteMoeSWAForCausalLM (#25505)

gabe-l-hart committed 21 days ago

Verified 7221e24f

sycl: report zero devices instead of aborting when the host has none (#27291)

Titaniumtown committed 21 days ago

Verified 6cc504a2

ci : add release attestation url (#27389)

CISC committed 21 days ago

Verified 01ac3ad7

models : remove duplicate metadata load (#27378)

CISC committed 21 days ago

Verified 2e92ecd0

ci : re-enable release dependency for sycl (#27385)

CISC committed 21 days ago

Verified 645ca283

ggml: add ggml_rope_set_offset (+ metal support) (#27120)

ngxson committed 21 days ago

Verified fe8156f7

ui: read persisted settings before the API key probe (#27365)

ServeurpersoCom committed 21 days ago

Verified 77acca43

mtmd: add mtmd_bitmap_set_mergeable (#27348)

ngxson committed 21 days ago

Verified 95c409c1

metal : dequantize q8_0 using packed types (#27370)

ggerganov committed 21 days ago

Verified 8ef78e64

server: add dedup-cache-models preset option (#27346)

ngxson committed 21 days ago

Verified ee4c505a

Older