Repository navigation

All pull requests

Labels (24)Milestones (0)New pull request

Open

(187)

Closed

(153)

feat(moe): split the offload expert cache into slot pools
#689jason-fxz opened Oct 14, 2026Collaborator
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
fix(rocm): verify native cpu moe graph handshake on rocm 10
#667zihaomu opened Oct 13, 2026Contributor
perf(moe): skip inactive block-fp8 decode routes
#665taking-lying-flat opened Oct 12, 2026Contributor
feat(moe): support gsq q2_0 routed experts
#660wenlive opened Oct 11, 2026
fix(kernel): enable disk row stores on windows
#659wenlive opened Oct 11, 2026
feat(models): add bounded encoder swa replay behind --swa-encoder-replay
#657void-mckenzie opened Oct 10, 2026Contributor
2
fix(moe): refuse e8m0 scale code 0xFF when MXFP4 expert pieces are read
#656Avicennasis opened Oct 10, 2026
fix(scheduler): stop the swa prefill commit from freeing pages the request still reads
#651ischencheng opened Oct 8, 2026Contributor
fix(engine): apply shrinking cache resizes before growing ones in live rebuild (Closes #643)
#650void-mckenzie opened Oct 8, 2026Contributor
fix(rotary): preserve fp32 inputs in mrope torch fallbacks
#641taking-lying-flat opened Oct 8, 2026Contributor
feat(qwen3_5_moe): tensor-parallel weight sharding for multi-GPU serving
#637zhangnju opened Oct 8, 20264 of 5 tasks
feat(quant): opt-in fp8-block QAT for the lm_head
#636zhangnju opened Oct 8, 20262 of 3 tasks
perf(moe): avoid intra-op fanout while filling NVFP4 host banks
#635earlvanze opened Oct 8, 2026Contributor
feat(attention): hand-written GQA-fused extend kernel for pre-sm70 (1.8x)
#618wac81 opened Oct 6, 2026
perf(fla): fix the pre-sm70 GDN chunk-kernel launch configs (7.4x / 11.3x)
#617wac81 opened Oct 6, 2026
fix(server): name the lifecycle stage in every API's 503
#616YevheniiKotyrlo opened Oct 6, 2026Contributor
fix(engine): start a single rank without a rendezvous port
#615YevheniiKotyrlo opened Oct 6, 2026Contributor

Pull requests