All pull requests

Pull requests

• perf(train): replicate frozen parameters under FSDP

  #1324·Madhav-Kanda opened Oct 27, 2026Contributor·3 of 4 tasks·

  ·
  5

• feat(train): adapt flex_attention for NPU

  dco-failed
  #1323·sunny-infra opened Oct 27, 2026Contributor·4 tasks·

  ·
  3

• Added AIPerf based eval support

  documentation
  #1322·shanjiaz opened Oct 26, 2026Collaborator·4 tasks done·

  ·
  7

• perf(models): count masked accuracy positions without masked_select

  #1315·orestis-z opened Oct 23, 2026Collaborator·2 of 4 tasks·

  ·
  2

• perf(train): fuse Muon's momentum and weight-update passes into two kernels

  needs-rebase
  #1314·orestis-z opened Oct 23, 2026Collaborator·7/7·2 of 4 tasks·

  ·
  3

• perf(train): form Muon's Newton-Schulz Gram products block-triangularly

  #1313·orestis-z opened Oct 23, 2026Collaborator·6/7·2 of 4 tasks·

  ·
  2

• [Crucible] dspark training step: -34.5 ms/step

  quality-failed
  #1312·orestis-z opened Oct 23, 2026Collaborator·

  ·
  3

• Async dataloader (+ refactor of data loading)

  needs-rebase
  #1310·fynnsu opened Oct 22, 2026Collaborator·3 of 4 tasks·

  ·
  12

• [Crucible] dspark training step: -23.19 ms/step

  #1309·orestis-z opened Oct 22, 2026Collaborator·4 tasks·

  ·
  2

• fix(evaluate): detect per-series counter discontinuities

  documentation
  #1304·Liuchenbing-2026 opened Oct 22, 2026Contributor·3 of 4 tasks·

  ·
  1

  4

• perf(train): keep frozen linear weights in the autocast dtype

  #1301·orestis-z opened Oct 21, 2026Collaborator·5/7·2 of 4 tasks·

  ·
  2

  

• perf(train): render console logs off the training thread

  #1297·orestis-z opened Oct 21, 2026Collaborator·4/7·2 of 4 tasks·

  ·
  2

• perf(losses): compute the fused KL and RKL forwards in a single pass

  #1296·orestis-z opened Oct 21, 2026Collaborator·3/7·2 of 4 tasks·

  ·
  2

  

• [WIP] docs(research): DFlash adaptive verification with measured cost curves

  documentation
  #1293·tomasruizt opened Oct 21, 2026Contributor·3 of 4 tasks·

  ·
  5

• fix(provenance): record a failed git diff instead of a clean tree

  #1289·YeonwooSung opened Oct 21, 2026·2 of 4 tasks·

  ·
  1

  2

• fix(hs_connectors): refuse non-finite tensors in the FP8 hidden-state writer

  #1288·YeonwooSung opened Oct 21, 2026·2 of 4 tasks·

  ·
  1

  4

• fix(models): keep P-EAGLE anchors inside the sequence and the document

  #1287·YeonwooSung opened Oct 21, 2026·2 of 4 tasks·

  ·
  1

  2

• fix(train): persist global_step on --save-best checkpoints

  #1286·YeonwooSung opened Oct 21, 2026·2 of 4 tasks·

  ·
  1

  2

• fix(cli): refuse to append regenerate-responses output without --resume

  #1285·YeonwooSung opened Oct 21, 2026·2 of 4 tasks·

  ·
  1

  2

• fix(cli): reject an incomplete prepare-data directory

  #1284·YeonwooSung opened Oct 21, 2026·2 of 4 tasks·

  ·
  1

  2

• fix(cli): stop regenerate-responses on the completion cap, not the prompt

  #1283·YeonwooSung opened Oct 21, 2026·2 of 4 tasks·

  ·
  1

  2

• fix(train): ignore non-finite validation loss when picking checkpoint_best

  #1282·YeonwooSung opened Oct 21, 2026·2 of 4 tasks·

  ·
  1

  2

• fix(train): keep run.yaml and speculators.patch under --save-best

  #1281·YeonwooSung opened Oct 21, 2026·2 of 4 tasks·

  ·
  1

  2

• Training support for H-Spec draft model

  dco-failedneeds-rebase
  #1267·weifanjiang opened Oct 21, 2026Contributor·2 of 4 tasks·

  ·
  4

• fix(convert): read the Eagle3 checkpoint dtype from either config key

  #1261·wguo7 opened Oct 20, 2026·4 tasks·

  ·
  3