All pull requests

Pull requests

• [8/8] docs(dflash): tuned Qwen3.8-27B DFlash2 recipe for one 8x H200 node

  #903·maocheng23 opened Sep 23, 2026Collaborator·

  ·

• [7/8] perf(dflash): fused Triton DFlash2 unary head with in-kernel top-k

  #902·maocheng23 opened Sep 23, 2026Collaborator·

  ·

• [6/8] perf(dflash): fuse DFlash2 grouped dynamic conv into Triton kernels

  #901·maocheng23 opened Sep 23, 2026Collaborator·

  ·

• [5/8] perf(dflash): opt out of capturing the teacher-only final hidden state

  #900·maocheng23 opened Sep 23, 2026Collaborator·

  ·

• [4/8] perf(dflash): one host sync per optimizer step, foreach/fused optimizer

  #899·maocheng23 opened Sep 23, 2026Collaborator·

  ·

• [3/8] perf(dflash): overlap durable ack with the next step on a Gloo group

  #898·maocheng23 opened Sep 23, 2026Collaborator·

  ·

• [2/8] perf(dflash): continuous online prompt ingest without per-chunk drain

  #897·maocheng23 opened Sep 23, 2026Collaborator·

  ·

• [1/8] perf(dflash): expose SGLang capture-server knobs and reuse HTTP sessions

  #896·maocheng23 opened Sep 23, 2026Collaborator·

  ·

• fix(eagle3): align loss mask during target preprocessing

  #895·navaneethkrishnansuresh opened Sep 21, 2026·

  ·
  1

  2

• Register glm-5.3 chat template

  #894·hiteshjain118 opened Sep 19, 2026·

  ·
  1

• Fix positional embedding for USP attention backend (eagle3 offline)

  #893·cih9088 opened Sep 19, 2026Contributor·

  ·

• [Intel GPU] Add XPU support and update basic documentation

  #892·polisettyvarma opened Sep 18, 2026·

  ·

• feat(eval): organize summary and per-position diagnostics

  #891·maocheng23 opened Sep 17, 2026Collaborator·

  ·

• fix(export): keep checkpoint precision in SGLang exports and config dtype

  #890·StevenWang-CY opened Sep 17, 2026Contributor·

  ·
  3

• feat(eval): replay checkpoints on frozen teacher features

  #889·maocheng23 opened Sep 17, 2026Collaborator·

  ·

• feat(data): add DeepSeek-V4.1 chat templates with the official encoder

  #887·bighead-liat opened Sep 17, 2026·

  ·
  4

• feat(capture): add an SGLang main (923e4a56) target for the spec-capture patch

  #886·bighead-liat opened Sep 17, 2026·

  ·
  8

• feat(moe): partition the routed experts across an expert-parallel group (moe.py variant)

  #884·fg11991 opened Sep 16, 2026·

  ·
  2

• feat(moe): partition the routed experts across an expert-parallel group

  #883·fg11991 opened Sep 16, 2026·

  ·
  2

• feat: expert parallelism for draft models with sharded experts

  #882·fg11991 opened Sep 16, 2026·

  ·

• offline capture: compat shims for sglang-main API drift (sglang 0.5.18 path unchanged)

  #876·bighead-liat opened Sep 12, 2026·

  ·
  3

• Fix target head weight load and refactor checkpoint loading

  #875·cih9088 opened Sep 11, 2026Contributor·

  ·

• feat(dflash): add experimental Draft-OPD replay support

  #874·maocheng23 opened Sep 10, 2026Collaborator·

  ·

• [refactor] unify benchmark entry points under specforge benchmark

  #873·FrankLeeeee opened Sep 10, 2026Collaborator·

  ·
  2

• [refactor] add hardware-specific installation instruction

  #861·FrankLeeeee opened Sep 8, 2026Collaborator·

  ·
  6