All pull requests

Pull requests

• 修复 P2pNcclConnector 三个问题：tensor key 对不上、对端重启挂 listener、producer 断言崩溃

  #1250·kezboard233 opened Oct 23, 2026·

  ·
  2

• [Bugfix][Core] Reserve the null KV block during startup admission

  #1247·Leonccaa opened Oct 21, 2026Contributor·3 of 4 tasks·

  ·

• [Perf][Distributed] Add topology-aware all-reduce selection and extend to any p2p

  #1245·trademark789 opened Oct 21, 2026·4 tasks done·

  ·
  1

• [Perf][KV Offload] Restore Mamba states where EAGLE GPU hits can reuse them

  #1244·Leonccaa opened Oct 21, 2026Contributor·3 of 4 tasks·

  ·
  1

• [Model][SM70] DeepSeek-V4.1-Flash-1017 on 8x V100: EXL3 experts, graph attention, DSpark speculative decoding, 1M context (stacks on #1031)

  #1243·rivetphilbot opened Oct 21, 2026Contributor·

  ·

• Bump the minor-update group across 1 directory with 171 updates

  dependencies
  #1241·dependabot[bot] opened Oct 20, 2026Bot·

  ·

• [Bugfix][Core] Align Mamba replay boundaries without an EAGLE block drop

  #1240·Peuqui opened Oct 20, 2026Contributor·

  ·
  1

• [Bugfix][Qwen4Exp] Reduce the PLE offload config to DP1 before it is built

  #1237·Peuqui opened Oct 19, 2026Contributor·

  ·
  1

• [Perf][Qwen4Exp] Run the native page4 QSA prefill routes for 12/24-head ranks, several KV heads and Turing

  #1236·Peuqui opened Oct 19, 2026Contributor·

  ·

• [Perf][Qwen4Exp] Bound the page prefetch of the mapped PLE row gather

  #1235·Peuqui opened Oct 19, 2026Contributor·

  ·

• [Bugfix][Qwen4Exp] Decide the PLE cascade merge when the layer is built

  #1234·Peuqui opened Oct 19, 2026Contributor·

  ·

• [Bugfix][KV Offload] Finish reused-block stores before MRv2 zeroes them

  #1232·Leonccaa opened Oct 18, 2026Contributor·

  ·

• [Bugfix][KV Cache] Zero packed DCP pages from the page base

  #1230·Leonccaa opened Oct 18, 2026Contributor·

  ·

• [Perf][Qwen4Exp] Admit the cuBLAS QSA indexer and exact top-k on Turing

  #1229·Peuqui opened Oct 17, 2026Contributor·

  ·

• [Perf][GDN] Run the FlashQLA TileLang prefill kernel on Turing

  #1228·Peuqui opened Oct 17, 2026Contributor·

  ·

• [Perf][Attention] Pack GQA heads for FA2 multi-token decode verification

  #1227·Peuqui opened Oct 17, 2026Contributor·

  ·

• [Perf][FLA] Let the delta_h autotuner lift the ptxas register cap

  #1226·Peuqui opened Oct 17, 2026Contributor·

  ·

• [Bugfix][MRv2] Bound prompt logprobs memory with small chunks and request limits

  #1225·Leonccaa opened Oct 17, 2026Contributor·3 of 4 tasks·

  ·

• [Perf][SM70][DeepSeek-V4] Keep the cuBLAS decode indexer above 131,072 tokens of context

  #1222·Peuqui opened Oct 16, 2026Contributor·

  ·

• [Spec Decode][MRV2] DSpark speculator and compressed-index context buckets for Model Runner V2

  #1221·Peuqui opened Oct 16, 2026Contributor·

  ·

• [Bugfix][Distributed] Verify P2P before provisioning compact collectives without custom all-reduce

  #1219·Peuqui opened Oct 15, 2026Contributor·

  ·
  2

• [Spec Decode][MRV2] Reduced draft vocabulary for MTP (backport of vllm#59740)

  #1218·Peuqui opened Oct 15, 2026Contributor·

  ·

• [Perf][Qwen4Exp] Overlap the PLE gather of the next prefill chunk

  #1217·Peuqui opened Oct 15, 2026Contributor·

  ·

• [Kernel] Two-rank FP16 all-reduce through pinned host memory for GPUs without P2P

  #1216·Peuqui opened Oct 15, 2026Contributor·

  ·

• [Test] Repair stale fakes in PLE offload and SM70 decode-graph tests

  #1214·Peuqui opened Oct 15, 2026Contributor·

  ·