All pull requests

Pull requests

• [Attention] DEC107: add SM107 FP8 and MXFP4-K decode (default off)

  #166·leopck opened Oct 8, 2026Collaborator·

  ·

• [KV Cache] VLLM_LOCALITY_SPLIT: place shared cache blocks by domain (default off)

  #165·leopck opened Oct 8, 2026Collaborator·

  ·

• [Locality] BACKFILL: add gated domain streams and fork/join (default off)

  #164·leopck opened Oct 8, 2026Collaborator·

  ·

• [Attention] PRE107: retain reference prefill routing (default off)

  #163·leopck opened Oct 8, 2026Collaborator·

  ·

• [Model] VLLM_L2X: prefetch non-MoE weights (default off)

  #162·leopck opened Oct 8, 2026Collaborator·

  ·

• [MoE] W4A8: add packed experts and gain correction (default off)

  #161·leopck opened Oct 8, 2026Collaborator·

  ·

• [MoE] PDL: add routing early release and standalone finalize scheduling (default off)

  #160·leopck opened Oct 8, 2026Collaborator·

  ·

• [Attention] VLLM_FLASHINFER_F1_PREFILL: opt into SM107 gen routing (default off)

  #159·leopck opened Oct 8, 2026Collaborator·

  ·

• [Attention] FI_DECODE_SPLITKV: select a bounded decode split policy (default off)

  #158·leopck opened Oct 8, 2026Collaborator·

  ·

• [GDN] PREFILL: add exact state transport and scheduling variants (default off)

  #157·leopck opened Oct 8, 2026Collaborator·

  ·

• [GDN] UCACHE: preserve source-faithful deferred-state port (default off)

  #156·leopck opened Oct 8, 2026Collaborator·

  ·

• [GEMM] VLLM_LOWM_BF16_GEMM_MAX_M: cap low-M dispatch (default off)

  #155·leopck opened Oct 8, 2026Collaborator·

  ·

• [GDN] Block extent: preserve unconditional source startup guard

  #154·leopck opened Oct 8, 2026Collaborator·

  ·

• [GDN] ZERO_DRAFT_AS_SPEC: preserve accepted state and counters (default on)

  #153·leopck opened Oct 8, 2026Collaborator·

  ·

• [KV] CoW: copy the widest aliased block view (default on)

  #152·leopck opened Oct 8, 2026Collaborator·

  ·

• [FlashInfer][sm_107] K=64 MXFP8 tactics for the autotuned CuTe-DSL mm_mxfp8 (VLLM_FLASHINFER_MXFP8_K64)

  #149·leopck opened Oct 6, 2026Collaborator·

  ·
  1

• [FlashInfer] VLLM_FLASHINFER_AUTOTUNE_FILE: seed the autotune cache from a pinned file

  #148·leopck opened Oct 6, 2026Collaborator·

  ·
  1

• [lcd] GB300: PDL launch of _nqf_quant_rows_kernel under LCD_PDL_TRITON (sm_107 unchanged)

  #147·leopck opened Oct 6, 2026Collaborator·

  ·
  1

• [U02] Suffix-only token staging for new requests (VLLM_SUFFIX_TOKEN_STAGING)

  #146·leopck opened Oct 6, 2026Collaborator·

  ·
  1

• [F05] M-RoPE text-only fast path (VLLM_MROPE_TEXT_ONLY_FAST_PATH, default ON)

  #145·leopck opened Oct 6, 2026Collaborator·

  ·
  2

• [gb-fix] GB300 s8g fixes: FMHA_GEN_RULE=gb300, VLLM_GDN_FI_VSPLIT_V1, MoE routing module cache, eager F46 conv load

  #144·leopck opened Oct 6, 2026Collaborator·

  ·
  1

• [glue] glue tier 1: NQRR / FNQ_F0 / FNQ_XB / EWS CUDA QKV prologue + PDL, GLUE_GSC_QO plumbing, AOT salts (GLUE_*)

  #143·leopck opened Oct 6, 2026Collaborator·

  ·
  1

• [rx-cascade] FRONT: exact front grouping of shared-prefix decode requests (PREFIX_SPREAD*)

  #142·leopck opened Oct 6, 2026Collaborator·

  ·
  1

• [lowc2] host-path family: FEH memo, KV-copy table cache, sampler dirty-state skip, sampler pre-capture

  #141·leopck opened Oct 6, 2026Collaborator·

  ·
  1

• [gdnpf] GDN spec||prefill overlap (FAST, sideearly) + step-plan hoists + VSDIRECT fix/guard + FI persistent KV counter

  #140·leopck opened Oct 6, 2026Collaborator·

  ·
  1