All pull requests

Pull requests

• Fix speculators build error found in downstream source build

  quality-failedready
  #1193·dhuangnm opened Oct 2, 2026Collaborator·4 tasks·

  ·
  6

• [Mooncake] Return transfer manifest through kv_transfer_params

  ready
  #1192·fynnsu opened Oct 1, 2026Collaborator·3 of 4 tasks·

  ·
  7

• Enable gradient accumulation

  documentationneeds-rebaseready
  #1178·shanjiaz opened Sep 30, 2026Collaborator·4 tasks done·

  ·
  6

• [feat]: add DeepSeek V4 aliases

  #1177·pavelgein opened Sep 30, 2026Contributor·2 of 4 tasks·

  ·
  3

• [Train] Add --draft-num-key-value-heads; warn on MLA verifiers

  documentation
  #1176·Frozen7771 opened Sep 30, 2026Contributor·4 tasks done·

  ·
  1

  2

• [Refactor][Data] Reject raw presets in prepare-data [1/n]

  documentationready
  #1170·WindChimeRan opened Sep 26, 2026Contributor·4 tasks done·

  ·
  4

• Add e2e, regression, and multi-GPU tests to Buildkite nightly pipeline

  ready
  #1161·rashmigottipati opened Sep 24, 2026Collaborator·4 tasks·

  ·
  4

• fix(models): treat the Qwen3.5 MoE family as Gemma-style for the veri…

  #1145·avshalomman opened Sep 18, 2026·4 tasks·

  ·
  2

• fix(eagle3): fail conversion on unmatched checkpoint keys instead of saving random weights

  #1131·YuEfSaEDU opened Sep 16, 2026·

  ·
  1

  2

• Loss weight schedule

  documentationready
  #1130·fynnsu opened Sep 16, 2026Collaborator·3 of 4 tasks·

  ·
  5

• feat(benchmark): add 'comment' subcommand for GitHub PR summaries

  #1125·orestis-z opened Sep 15, 2026Collaborator·4/4·4 tasks done·

  ·
  7

  

• feat(benchmark): add HTML and Markdown report generation

  needs-rebase
  #1124·orestis-z opened Sep 15, 2026Collaborator·3/4·4 tasks done·

  ·
  8

  

• feat(benchmark): add torch.profiler support and GPU kernel breakdown

  #1123·orestis-z opened Sep 15, 2026Collaborator·2/4·4 tasks done·

  ·
  6

  

• fix: clean up stale docstrings, dead exports and script refs.

  documentation
  #1113·Adityarya11 opened Sep 12, 2026·3 of 4 tasks·

  ·
  1

  16

• feat(benchmark): add detailed profiling, visualization, and statistical tests

  needs-rebase
  #1110·orestis-z opened Sep 9, 2026Collaborator·5 tasks done·

  ·
  5

• refactor(metrics): remove legacy training accuracy and acceptance reports

  documentationneeds-rebase
  #1102·WindChimeRan opened Sep 9, 2026Contributor·4 tasks done·

  ·
  3

• add chunk loss for dspark

  dco-failed
  #1099·walterchenchn opened Sep 8, 2026Contributor·3 of 4 tasks·

  ·
  6

• perf(eagle3): skip unused teacher projections while preserving metrics

  #1094·julyanghar opened Sep 8, 2026·

  ·
  2

• test(train): add unit tests, typing, and docstrings for train utilities

  #1093·YuEfSaEDU opened Sep 7, 2026·

  ·
  1

  4

• feat(metrics): add shared reference-prefix metrics across drafters

  documentationneeds-rebase
  #1088·WindChimeRan opened Sep 6, 2026Contributor·4 tasks done·

  ·
  22

• fix(eval): support spline smoothing for small sweeps

  #1082·edgeai1 opened Sep 4, 2026Contributor·7 tasks done·

  ·
  4

• fix(training): Fix regression on DFlash/DFlash2/DSpark: --max-anchors is now per packed document, not per packed sequence

  documentation
  #1071·WindChimeRan opened Sep 2, 2026Contributor·4 tasks done·

  ·
  7

• Add XPress: a causal refiner over the DFlash block drafter

  documentation
  #1061·ZKBig opened Aug 30, 2026·4 tasks done·

  ·
  65

• [DFlash2] Add configurable MoE draft FFN

  needs-rebase
  #1059·qianlihuang opened Aug 30, 2026Contributor·2 of 4 tasks·

  ·
  3

• [Train] Add warmup-stable-decay learning rate scheduler

  #1058·qianlihuang opened Aug 30, 2026Contributor·3 of 4 tasks·

  ·
  7