Commits

Breadcrumbs

History for

onmain

User selector

All users

All time

Commit history

Commits on May 27, 2025

• run ppo

  shwgao

  committed
  eb222e3View commit details

  Copy full SHA for eb222e3

  View code at this pointBrowse repository at this point

Commits on May 23, 2025

• [NPU] feat: Support FSDP worker and vLLM Ascend (#332)

  Show description for 0528ba1
  

  3 peopleauthored
  0528ba1View commit details

  Copy full SHA for 0528ba1

  View code at this pointBrowse repository at this point

• fix: add `loss_agg_mode` to critics (#1340)

  Show description for 9ddc725
  tongyx361

  authored
  9ddc725View commit details

  Copy full SHA for 9ddc725

  View code at this pointBrowse repository at this point

• Activation Offloading (#1220)

  Show description for aaaaaab
  imh966

  authored
  aaaaaabView commit details

  Copy full SHA for aaaaaab

  View code at this pointBrowse repository at this point

• [megatron] feat: save hf model config in megatron checkpoint manager (#1562)

  Show description for 54a5e6e
  0x404

  authored
  54a5e6eView commit details

  Copy full SHA for 54a5e6e

  View code at this pointBrowse repository at this point

• Add explicit position_ids to model.generate in hf rollout (#1637)

  Show description for 2c179da
  Geaming2002

  authored
  2c179daView commit details

  Copy full SHA for 2c179da

  View code at this pointBrowse repository at this point

Commits on May 22, 2025

• Mert/passatk advantage (#1621)

  mertunsall

  authored
  01abd3cView commit details

  Copy full SHA for 01abd3c

  View code at this pointBrowse repository at this point

• [megatron] optimization: avoid padding to logits (#1629)

  Show description for 04acd09
  ISEEKYAN

  authored
  04acd09View commit details

  Copy full SHA for 04acd09

  View code at this pointBrowse repository at this point

• [BugFix] Fix sglang and vllm engine args (#1634)

  Show description for c803b1f
  ETOgaosion

  authored
  c803b1fView commit details

  Copy full SHA for c803b1f

  View code at this pointBrowse repository at this point

• [PRIME]bug fix: reward scoring hangs after progress reaches 100% (#1466)

  Show description for 6cbb051
  CajZella

  authored
  6cbb051View commit details

  Copy full SHA for 6cbb051

  View code at this pointBrowse repository at this point

• [Megatron][BREAKING] Allow override of transformer config to enable custom megatron features like variable PP layers distribution, with CI tests (#1555)

  Show description for 1cfa2be
  ETOgaosion

  authored
  1cfa2beView commit details

  Copy full SHA for 1cfa2be

  View code at this pointBrowse repository at this point

• [vllm] feat: add `engine_kwargs` in vllm_rollout_spmd to set params in vllm engine (#1442)

  Show description for c07013e
  zhangyongxin121

  authored
  c07013eView commit details

  Copy full SHA for c07013e

  View code at this pointBrowse repository at this point

Commits on May 21, 2025

• Expose `engine_kwargs` from SGLang to Verl configuration (#1616)

  Show description for 72683a7
  hebiao064

  authored
  72683a7View commit details

  Copy full SHA for 72683a7

  View code at this pointBrowse repository at this point

• [minor] fix: use init_empty_weights instead of torch.device("meta") (#1587)

  0x404

  authored
  d475654View commit details

  Copy full SHA for d475654

  View code at this pointBrowse repository at this point

• [recipe] feat: support running dapo using main_ppo (#1612)

  Show description for 2dc3e0e
  vermouth1992

  authored
  2dc3e0eView commit details

  Copy full SHA for 2dc3e0e

  View code at this pointBrowse repository at this point

• Simplify FSDP SGLang pre and post process (#1609)

  Show description for 8970cb0
  hebiao064

  authored
  8970cb0View commit details

  Copy full SHA for 8970cb0

  View code at this pointBrowse repository at this point

• [constant scheduler] fix: model won't be updated on first training step (#1463)

  0x404

  authored
  80af51bView commit details

  Copy full SHA for 80af51b

  View code at this pointBrowse repository at this point

• [megatron] support megatron expert parallel (#1467)

  Show description for add17f0
  

  ISEEKYAN

  and

  ETOgaosion

  authored
  add17f0View commit details

  Copy full SHA for add17f0

  View code at this pointBrowse repository at this point

Commits on May 20, 2025

• Remove unnecessary broadcast (#1597)

  Show description for 0d3360a
  hebiao064

  authored
  0d3360aView commit details

  Copy full SHA for 0d3360a

  View code at this pointBrowse repository at this point

• [BugFix] Megatron: fix missing grad_norm and lr calculation, and fix fsdp grad_norm storage (#1601)

  Show description for 1d4e23a
  ETOgaosion

  authored
  1d4e23aView commit details

  Copy full SHA for 1d4e23a

  View code at this pointBrowse repository at this point

• Support multi-turn rollout with Qwen chat template (#1593)

  Show description for 15b1b15
  jybsuper

  authored
  15b1b15View commit details

  Copy full SHA for 15b1b15

  View code at this pointBrowse repository at this point

• [rollout] perf: replace AsyncOpenAI to aiohttp client in ChatCompletionScheduler (#1588)

  Show description for 3eaaf24
  wuxibin89

  authored
  3eaaf24View commit details

  Copy full SHA for 3eaaf24

  View code at this pointBrowse repository at this point

• [feat] support logging to ClearML (#1582)

  Show description for 457ccd9
  pavelgein

  authored
  457ccd9View commit details

  Copy full SHA for 457ccd9

  View code at this pointBrowse repository at this point

• [BugFix] Megatron: fix checkpoint manager, some states only rank 0 need to save to avoid errors (#1586)

  Show description for 88527e6
  ETOgaosion

  authored
  88527e6View commit details

  Copy full SHA for 88527e6

  View code at this pointBrowse repository at this point

• Bump to sglang 0.4.6.post4 & unified generate sequences ability between sgl and sgl async (#1577)

  Show description for 8160ec6
  

  SwordFaith

  and

  ocss884

  authored
  8160ec6View commit details

  Copy full SHA for 8160ec6

  View code at this pointBrowse repository at this point

Commits on May 19, 2025

• [doc] single_controller: Adding doc strings and doc pages for public methods in `single_controller` (#1396)

  Show description for 8788e55
  hongpeng-guo

  authored
  8788e55View commit details

  Copy full SHA for 8788e55

  View code at this pointBrowse repository at this point

• [misc] use lazy import on megatron utils components (#1551)

  Show description for 5b24e01
  hongpeng-guo

  authored
  5b24e01View commit details

  Copy full SHA for 5b24e01

  View code at this pointBrowse repository at this point

• [Megatron] Qwen3moe-part3: fix mcore qwen3 moe config, no need for patching now, offer option to freeze moe router (#1540)

  Show description for 8176d3b
  ETOgaosion

  authored
  8176d3bView commit details

  Copy full SHA for 8176d3b

  View code at this pointBrowse repository at this point

Commits on May 18, 2025

• [misc] feat: support return full prompt with chat template in RLHFDataset (#1567)

  PeterSH6

  authored
  8653b1bView commit details

  Copy full SHA for 8653b1b

  View code at this pointBrowse repository at this point

• disable fused kernels by default (#1568)

  mingruimingrui

  authored
  b9a6890View commit details

  Copy full SHA for b9a6890

  View code at this pointBrowse repository at this point

Commits on May 17, 2025

• [BUG] Fix silent bug of using dtype from previous loop scope in build_memory_reference_from_module() (#1553)

  Show description for 40dcabe
  ccchow

  authored
  40dcabeView commit details

  Copy full SHA for 40dcabe

  View code at this pointBrowse repository at this point

Commits on May 16, 2025

• Feat/memory optimized loss (#1212)

  Show description for eb077f6
  

  mingruimingrui

  and

  ETOgaosion

  authored
  eb077f6View commit details

  Copy full SHA for eb077f6

  View code at this pointBrowse repository at this point

• [megatron] Qwen3moe-part 2: Allow Infer and train tp to be different with CI tests, Fix vllm resharding process (#1444)

  Show description for b529564
  ETOgaosion

  authored
  b529564View commit details

  Copy full SHA for b529564

  View code at this pointBrowse repository at this point

• [Refactor] Add middle truncation (#1488)

  Show description for 12bb857
  

  GaokaiZhang

  and

  Wangmerlyn

  authored
  12bb857View commit details

  Copy full SHA for 12bb857

  View code at this pointBrowse repository at this point

• [megatron] fix head_dim in GQA model when load from hf ckpt (#1513)

  Show description for 2c991f6
  ShareLer

  authored
  2c991f6View commit details

  Copy full SHA for 2c991f6

  View code at this pointBrowse repository at this point