All pull requests

Pull requests

• feat: add emergency response log tool

  #2149·Ray961123 opened Sep 25, 2026Contributor·

  ·
  2

• [Fix] Handle 0-D scalar buffers in ZCC EMA reshard load

  #2148·wangbingguang2026 opened Sep 24, 2026Contributor·

  ·
  4

• Pr 2086 review fix

  #2147·swgu98 opened Sep 24, 2026Member·

  ·
  3

• test: port #2080 coverage_test migration/rewrite to release/0.4

  #2146·liuhao2638 opened Sep 24, 2026Contributor·

  ·
  3

• Revert "fix fleet install"

  #2145·zhengshengning opened Sep 24, 2026Collaborator·

  ·
  3

• dev paddle test

  #2144·swgu98 opened Sep 24, 2026Member·

  ·
  8

• Fix the incorrect usage of fused_ffn in AOA

  #2143·xingmingyyj opened Sep 24, 2026Collaborator·

  ·
  8

• [Fsdp] support fp8 and fused linear

  #2142·zhengshengning opened Sep 24, 2026Collaborator·

  ·
  2

• [UAC] Accuracy-compatible loss and deferred token normalization in trainer

  #2140·zrr1999 opened Sep 24, 2026Member·8/9·

  ·
  6

• [UAC] Accuracy-compatible MoE: router, dense MLP, grouped experts and MoE layer

  #2139·zrr1999 opened Sep 24, 2026Member·7/9·

  ·
  10

• [UAC] Accuracy-compatible MLA/DSA attention, absorbed core and RoPE inv_freq

  #2138·zrr1999 opened Sep 24, 2026Member·6/9·

  ·
  12

• [UAC] Accuracy-compatible TP linear, vocab embedding, MTP projection and residual order

  #2137·zrr1999 opened Sep 24, 2026Member·5/9·

  ·
  11

• [SFT] GLM-5.2 training contract, pretokenized datasets and glm5_2 template

  #2136·zrr1999 opened Sep 24, 2026Member·4/9·

  ·
  9

• [Trainer] Fused-expert FlexCheckpoint at sharding=1 and HF export cadence

  #2135·zrr1999 opened Sep 24, 2026Member·3/9·

  ·
  7

• [GLM-5.2] Sequence-parallel DSA mask alignment, SP RoPE slicing and dense MTP masks

  #2134·zrr1999 opened Sep 24, 2026Member·2/9·

  ·
  9

• [GLM-5.2] DSA indexer skip/share: config, AOA mapping and runtime

  #2133·zrr1999 opened Sep 24, 2026Member·1/9·

  ·
  22

• chore: refresh connection cache (retry 3x)

  #2132·Ray961123 opened Sep 24, 2026Contributor·

  ·
  2

• [Optim] Support DeepSelect for TopK optimization in paddlefleet_ops

  #2131·Enigmatisms opened Sep 23, 2026Collaborator·

  ·
  6

• [Feature] Add DSpark read-only embedding and projection utilities

  #2130·LetThoughtsFlowOut opened Sep 23, 2026·

  ·
  4

• [Rcompute] Add block-scoped recompute spanning several layers (#2062)

  #2128·Difers opened Sep 23, 2026Collaborator·

  ·
  2

• Add modular AOA statements on the MTP block and the LM head

  #2127·wangbingguang2026 opened Sep 23, 2026Contributor·

  ·
  5

• [alignment] Kimi-K2 Megatron-aligned CE / MoE grad order / dense wgrad

  #2126·liym27 opened Sep 23, 2026Collaborator·

  ·
  13

• 修复无法触发coverage approval

  #2125·Ray961123 opened Sep 23, 2026Contributor·

  ·
  5

• private

  #2119·swgu98 opened Sep 22, 2026Member·

  ·
  1

• refined recompute: add MLA o_proj point and config support

  #2118·LCStayingdullCircuit opened Sep 22, 2026Contributor·

  ·
  8