Pull requests

• feat: support Qwen hybrid mixed FP8 and NVFP4 checkpoints

  auto-lgtm:needs-approvalreview:running
  #1541·hy-guo opened Oct 21, 2026·

  ·
  1

• fix(cache): preserve device metadata and validate TCP handoff buffers

  #1540·wangzhao-11a opened Oct 21, 2026·

  ·

• support multimodal mtp and mrope and fix per token fp8 kv cache

  auto-lgtm:needs-approvalreview:running
  #1539·HongminTan opened Oct 21, 2026Collaborator·

  ·
  2

• Develop/embedding cuda graph

  auto-lgtm:needs-approvalreview:running
  #1538·baifengyang1-cloud opened Oct 21, 2026·

  ·
  2

• docs: propose Qwen3.5 prefill/decode serving on Ascend

  auto-lgtm:needs-approvalreview:blocked
  #1537·wangzhao-11a opened Oct 21, 2026·

  ·
  3

• Fix flaky tests, retire redundant DeepEP H20 smoke cases, reorganize ARM/SM103 suites

  auto-lgtm:needs-approvalreview:running
  #1536·LLLLKKKK opened Oct 20, 2026Collaborator·

  ·
  2

• Feature/deepseek v4 topp pruning

  auto-lgtm:needs-approvalreview:running
  #1535·DCodiver opened Oct 20, 2026·

  ·
  16

• Develop/feat 2026 10 16 graph

  auto-lgtm:needs-approvalreview:delayed
  #1534·baifengyang1-cloud opened Oct 20, 2026·

  ·
  6

• feat: add per-level codebook constraints and improve pruned-vocab beam search

  auto-lgtm:needs-approvalreview:running
  #1533·parkerpang opened Oct 20, 2026Collaborator·

  ·
  8

• perf: optimize DSv4.1 decode attention, MoE packing, and MTP

  auto-lgtm:needs-approvalreview:running
  #1532·guluguluhhhh opened Oct 19, 2026Contributor·

  ·
  8

• feat(loader): introduce NewLoader and migrate supported models

  auto-lgtm:needs-approvalreview:running
  #1531·Oneydauh opened Oct 19, 2026Collaborator·

  ·
  1

• feat(kimi-k3): integrate K3 modeling, kernels and distributed serving

  auto-lgtm:needs-approvalreview:running
  #1530·lhcezx opened Oct 17, 2026Collaborator·

  ·
  93

• [WIP][smoke] DiffusionGemma bring-up: model, weights, proposer + dgm_chat_basic smoke

  auto-lgtm:needs-approvalreview:running
  #1529·LLLLKKKK opened Oct 16, 2026Collaborator·

  ·
  56

• Feature/maxpander

  auto-lgtm:needs-approvalreview:blocked
  #1528·Natureal opened Oct 16, 2026·

  ·
  42

• perf(crc): 将 CRC32C 校验隐藏进 KV cache 拷贝窗口

  auto-lgtm:authorizedreview:blocked
  #1527·alibaba-miji opened Oct 16, 2026Collaborator·

  ·
  17

• fix(flexlb): isolate cache sync from worker health and bound scheduling freshness

  auto-lgtm:needs-approvalreview:delayed
  #1525·wzy-99 opened Oct 16, 2026Collaborator·

  ·
  18

• Feat/k3 dev zhd

  auto-lgtm:needs-approvalreview:running
  #1524·HoniiTro19 opened Oct 16, 2026Collaborator·

  ·
  7

• fix(scr): reuse RTP ranks and C2 gang member lookup

  auto-lgtm:needs-approvalreview:blocked
  #1523·ABNER-1 opened Oct 16, 2026Collaborator·

  ·
  12

• fix: make FP16 INT8 communication scales opt-in

  auto-lgtm:needs-approvalreview:blocked
  #1522·hy-guo opened Oct 15, 2026·

  ·
  22

• feat(engine): support pipeline parallelism with PD and speculative de…

  auto-lgtm:needs-approvalreview:running
  #1521·a1iive opened Oct 15, 2026Collaborator·

  ·
  37

• fix: ignore unsupported thinking in DashSC requests

  auto-lgtm:needs-approvalreview:delayed
  #1520·hy-guo opened Oct 15, 2026·

  ·
  75

• feat(cache): add CUDA13 3D batch device-host copy

  auto-lgtm:needs-approvalreview:running
  #1517·happyhyaa opened Oct 14, 2026Collaborator·

  ·
  25

• fix(flexlb): preserve selected ViT address in batch dispatch

  auto-lgtm:needs-approvalreview:delayed
  #1516·ySingularity opened Oct 14, 2026Collaborator·

  ·
  2

• Feature/glm5 inference core

  auto-lgtm:needs-approvalreview:running
  #1515·Nancheng-11 opened Oct 14, 2026Collaborator·

  ·
  2

• Dev/btc 1.0 to main

  auto-lgtm:authorizedreview:blocked
  #1514·MMeecatfish opened Oct 14, 2026Collaborator·

  ·
  18