All pull requests

Pull requests

• [model] fix: preserve gradients for empty local PLE shards

  fix
  #1381·YLlllllllllll opened Oct 28, 2026Contributor·

  ·
  1

• [lora, ops, dist] feat: add SwiGLU-OAI expert LoRA for M3.1-Flash

  #1380·Li-Jicheng opened Oct 28, 2026·7 of 8 tasks·

  ·
  14

• [BREAKING][model, ops] refactor: use module-level ops config

  #1379·Zx55 opened Oct 28, 2026Contributor·

  ·
  14

• [BREAKING][trainer, data] fix: preserve DPO pair semantics

  fix
  #1378·YLlllllllllll opened Oct 28, 2026Contributor·

  ·
  3

• [data] fix: preserve dynamic batching checkpoint position

  ckptfix
  #1377·YLlllllllllll opened Oct 28, 2026Contributor·

  ·
  2

• [data] fix: support Ascend tail-segment cumprod

  ascendfix
  #1375·YLlllllllllll opened Oct 28, 2026Contributor·

  ·
  2

• [ckpt, optim, model] fix: preserve GDN FP32 and bound cold-resume memory

  ckptfix
  #1374·YLlllllllllll opened Oct 28, 2026Contributor·

  ·
  2

• Read compute units from an NPU instead of falling back to CPU threads

  ascend
  #1373·li-lizhe opened Oct 27, 2026Contributor·

  ·
  3

• [BREAKING][dist, ops, ckpt, perf] feat: add DeepGEMM FP8 expert all-gather

  ckpt
  #1372·FoolPlayer opened Oct 27, 2026Collaborator·

  ·
  12

• [omni, model, ckpt, dist, ops, config, docs] feat: support DeepSeek-V4.1-Flash in SeedOmni

  ckptdoc
  #1370·Coach257 opened Oct 26, 2026Collaborator·5 tasks done·

  ·
  44

• [ops] fix: prevent GDN backward NaNs on Ascend950 and migrate to fla_npu v26.10.0

  fix
  #1369·Feng0w0 opened Oct 26, 2026Collaborator·3 of 5 tasks·

  ·
  8

• [dist, model, ops] feat: add opt-in tensorwise float8 training

  #1364·Levius-Fubuki opened Oct 26, 2026Contributor·

  ·
  1

• [ops] feat: run the DeepSeek-V4 TileLang kernels and FP8 QAT on AMD gfx942/gfx950

  #1363·xiaohong42 opened Oct 25, 2026·5 tasks done·

  ·
  18

• [ops, perf] fix: unroll sparse MLA split-store loop

  fix
  #1362·dajiaohuang opened Oct 25, 2026·

  ·
  1

• [dist, perf] refactor: coalesce EP host metadata transfers

  #1359·Levius-Fubuki opened Oct 25, 2026Contributor·4 tasks done·

  ·
  5

• [omni, model, config, docs, agent, ci] feat: support Qwen3 in SeedOmni

  cidoc
  #1357·Coach257 opened Oct 23, 2026Collaborator·5 tasks done·

  ·
  1

  26

• [model, optim] feat: freeze Qwen4-Exp PLE embedding tables

  #1354·zyang6 opened Oct 23, 2026·5 of 7 tasks·

  ·
  1

• [parallel, model] feat: support Qwen3-VL USP with bidirectional ViT ring

  #1352·yeyeyeping opened Oct 23, 2026Contributor·

  ·
  4

• [ops, perf] feat: reuse AscendC GDN forward intermediates

  #1346·YLlllllllllll opened Oct 23, 2026Contributor·

  ·
  22

• [ci] test: cover low-precision gradient accumulation

  ci
  #1341·amenomori opened Oct 22, 2026·

  ·
  1

• [lora] feat: support disabling dense adapters for reference inference

  #1337·Levius-Fubuki opened Oct 21, 2026Contributor·7 tasks done·

  ·
  10

• [ci] fix: verify workspace NPU runtime before Ascend tests

  ascendcifix
  #1334·YLlllllllllll opened Oct 20, 2026Contributor·

  ·
  7

• [parallel] feat: opt in Qwen4-Exp to combined Ulysses and CP meshes

  #1332·YLlllllllllll opened Oct 20, 2026Contributor·

  ·
  25

• [model] test: preserve Qwen4-Exp dummy vision attention contract

  ci
  #1331·YLlllllllllll opened Oct 20, 2026Contributor·

  ·
  13

• [ops, perf] refactor: tighten merged MoE same-NK launch bounds

  #1323·LeoCristallo opened Oct 20, 2026Contributor·4 tasks done·

  ·
  6