Repository navigation

All pull requests

Labels (25)Milestones (0)New pull request

Open

(184)

Closed

(265)

perf(ggml): step the flat moe grid tile width up when the mean fills the tile
#861void-mckenzie opened Nov 4, 2026Contributor
fix(gguf): log why mmproj auto-discovery skips a file
#860void-mckenzie opened Nov 4, 2026Contributor
feat(spec): add DFlash 2 speculative decoding
#857jason-fxz opened Nov 3, 2026Collaborator
feat(kvcache): add opt-in fp8 kv cache for r9700
#853zihaomu opened Nov 3, 2026Contributor
fix(server): reject empty completions prompt and keep error code
#850techsaurabh-394 opened Nov 2, 2026
fix(scheduler): clip max_tokens on a copy of the shared SamplingParams
#849techsaurabh-394 opened Nov 2, 2026
Feat/completions n
#843A5bhinav opened Oct 28, 2026Contributor
[PoC, DO NOT MERGE] feat(spec): add speculative decoding with MTP drafts
#832jason-fxz opened Oct 27, 2026Collaborator
feat(windows): expand Qwen and Gemma checkpoint support for OOB Huggingface uncensored models
#831trcwebdesign opened Oct 27, 2026
feat(models): add nvidia/Qwen3.8-Flash-Next-FP4-Local checkpoint support
#820chrisqianz opened Oct 24, 2026Contributor
fix(gguf): rename the ggml shim's type_info lookup, ambiguous under MSVC
#809YevheniiKotyrlo opened Oct 23, 2026Contributor
perf(ple): add an opt-in bounded disk row cache
#792wenlive opened Oct 22, 2026Contributor
fix(scheduler): preserve shared sampling parameters when clipping
#788wenlive opened Oct 22, 2026Contributor
perf(attention): fuse short bf16 decode output gates on gfx1201
#781zihaomu opened Oct 21, 2026Contributor
perf(kernel): add opt-in hip fp8 decode kernels for r9700
#779zihaomu opened Oct 21, 2026Contributor
test(models): cover the Ornith-1.5-122B NVFP4 export
#777chrisqianz opened Oct 21, 2026Contributor
1
feat(mimo): add text-only MiMo v2.6 Flash support
#761jcmh0782 opened Oct 20, 2026
fix(server): opt-in backend recovery after runtime CUDA OOM
#688oscarlius opened Oct 14, 2026
fix(rocm): route moe_align around Triton 3.8 miscompile
#687thedandano opened Oct 14, 2026
perf(qwen4): add opt-in layer-major long prefill
#686oscarlius opened Oct 14, 2026
fix(moe): admit a wide offloaded decode step's experts 128 at a time
#680YevheniiKotyrlo opened Oct 13, 2026Contributor
feat(moe): load a short prefill chunk's experts on demand through the slot cache
#676YevheniiKotyrlo opened Oct 13, 2026Contributor
2
perf(moe): promote recurring hybrid misses into empty cache slots
#674taking-lying-flat opened Oct 13, 2026Contributor
feat(kvcache): --kv-placement host keeps the QSA K/V in pinned host memory
#672YevheniiKotyrlo opened Oct 13, 2026Contributor
test(moe): grant one bf16 ulp only at rounding ties in the ds_fp4 transcription check
#668void-mckenzie opened Oct 13, 2026Contributor

Pull requests