Commits

Breadcrumbs

History for

onmain

User selector

All users

All time

Commit history

Commits on Jun 29, 2026

• [fully_async, doc] fix: ignore temperature config for teacher prompt_logprobs and warn when non-default value is set (#6867)

  Show description for c165b3d
  

  kenkenpa2126

  and

  claude

  authored
  c165b3d

  Copy full SHA for c165b3d

• [model] fix: tokenizer.hf_processor: GLM-4V arm uses wrong processor class name (#6814) (#6873)

  Show description for 7ba7a5d
  yushuosun

  authored
  7ba7a5d

  Copy full SHA for 7ba7a5d

• [fully_async] fix: correct the use of partial_rollout (#6882)

  Show description for 5371b99
  zpltys

  authored
  5371b99

  Copy full SHA for 5371b99

• [trainer, rollout] feat: log rollout moe load-balance metrics (#6853)

  Show description for 3d66a3d
  Luosuu

  authored
  3d66a3d

  Copy full SHA for 3d66a3d

Commits on Jun 27, 2026

• [worker] fix: Update external_lib config annotation in HFModelConfig to allow a list of string (#6862)

  Show description for 1ff76cc
  

  wyettzeng

  and

  huaye.zeng

  authored
  1ff76cc

  Copy full SHA for 1ff76cc

Commits on Jun 26, 2026

• [fsdp] fix: re-wrap fused log_probs/entropy when `use_remove_padding=False` (#6846)

  Show description for 2390a3f
  Mecoli1219

  authored
  2390a3f

  Copy full SHA for 2390a3f

• fix typo: clip_cvo_ratio to clip_cov_ratio (#6855)

  Show description for 6fef6a7
  GZYZG

  authored
  6fef6a7

  Copy full SHA for 6fef6a7

• [rollout, algo] fix: compute rollout_is_seq_fraction from raw weights (#6806)

  Show description for bdfedd4
  

  EazyReal

  and

  claude

  authored
  bdfedd4

  Copy full SHA for bdfedd4

Commits on Jun 25, 2026

• [ckpt] fix: write HF config before bridge save to fix shard inference (#6842)

  Show description for b6db2c1
  wizkdc

  authored
  b6db2c1

  Copy full SHA for b6db2c1

Commits on Jun 24, 2026

• [BREAKING][trainer, cfg] chore: enable V1 trainer by default (#6823)

  Show description for bc6fb96
  Begunner

  authored
  bc6fb96

  Copy full SHA for bc6fb96

Commits on Jun 23, 2026

• [reward] feat: colocated reward model for v1 sync/colocate_async trainer (#6818)

  Show description for f22e715
  Begunner

  authored
  f22e715

  Copy full SHA for f22e715

• [hardware] refactor: per-model NPU patches with fault isolation (#6777)

  Show description for c5a9ed5
  tardis-key

  authored
  c5a9ed5

  Copy full SHA for c5a9ed5

• [rollout, reward] feat: add full determinism support for vLLM rollout and reward model (#6572)

  Show description for f7daadb
  KaisennHu

  authored
  f7daadb

  Copy full SHA for f7daadb

• [megatron] feat: align optimizer states and DDP grad bucket with model precision (#6526)

  Show description for 64ea454
  kolehma8

  authored
  64ea454

  Copy full SHA for 64ea454

• [ci] chore: solve dapo error and add three baselines for npu nightly ci (#6711)

  Show description for 9d5188a
  daikang6

  authored
  9d5188a

  Copy full SHA for 9d5188a

Commits on Jun 22, 2026

• [rollout] feat: add Continuous Token for Agentic Rollout (#6779)

  Show description for 603944b
  gxlvera

  authored
  603944b

  Copy full SHA for 603944b

• [fully_async, trainer] fix: align aggregated metrics logging with current step (#6796)

  Show description for cb59290
  huaiyizhao

  authored
  cb59290

  Copy full SHA for cb59290

Commits on Jun 18, 2026

• [fsdp, model] feat: per-unit LoRA summon, FSDP1/2 compatibility, and strip-modules support (#6512)

  Show description for 8a69493
  

  3 peopleauthored
  8a69493

  Copy full SHA for 8a69493

• [docker, misc] chore: upgrade vllm, nccl and numpy > 2.0.0, to resolve packages dependency error (#6551)

  Show description for 14574ec
  

  ETOgaosion

  and

  xiefan46

  authored
  14574ec

  Copy full SHA for 14574ec

• [megatron] feat: Support Megatron chunk entropy (#6446)

  Show description for 33f97ca
  ZLiao097

  authored
  33f97ca

  Copy full SHA for 33f97ca

• [worker] feat: add per-step optimizer param overrides (#6765)

  Show description for 09de985
  Luosuu

  authored
  09de985

  Copy full SHA for 09de985

Commits on Jun 17, 2026

• [data] fix: offload process_vision_info to a thread executor to unblock the agent-loop event loop (#6789)

  Show description for b55a2b3
  huaiyizhao

  authored
  b55a2b3

  Copy full SHA for b55a2b3

• [trainer] feat: A runnable separate async trainer (#6790)

  Show description for 1d76c23
  Begunner

  authored
  1d76c23

  Copy full SHA for 1d76c23

Commits on Jun 16, 2026

• [trainer] feat: support staleness control strategy (#6778)

  Show description for 86e8123
  wuxibin89

  authored
  86e8123

  Copy full SHA for 86e8123

• [model] fix: NPU patches for the Qwen3-MoE compatible with different transformers (#6708)

  Show description for 5a6bf08
  mikequan0425

  authored
  5a6bf08

  Copy full SHA for 5a6bf08

• [rollout, sglang] feat: support sglang ROCm backend via aiter defaults and ray init env (#6664)

  Show description for 7c37cc5
  

  xiaohong42

  and

  Cursor

  authored
  7c37cc5

  Copy full SHA for 7c37cc5

• [rollout] fix: skip redundant clone in get_named_tensor_buckets to avoid OOM during SGLang weight sync (#6738)

  Show description for 996ab64
  

  HaozheZhang6

  and

  claude

  authored
  996ab64

  Copy full SHA for 996ab64

• [trainer] feat: add off_policy metrics (#6736)

  Show description for 96b74ad
  wuxibin89

  authored
  96b74ad

  Copy full SHA for 96b74ad

Commits on Jun 15, 2026

• [worker] feat: add tinker training worker primitives (#6717)

  Show description for ed89419
  Luosuu

  authored
  ed89419

  Copy full SHA for ed89419

• [BREAKING][ckpt] refactor: rename misleading `trainer` arg to `actor_wg` (#6724)

  Show description for e71ae41
  

  PeterSH6

  and

  claude

  authored
  e71ae41

  Copy full SHA for e71ae41

• [rollout] fix: Adjust cuda graph capture sizes config logic for vLLM >= 0.11.1 compatibility (#6728)

  Show description for db33bc8
  

  ConanZH429

  and

  Zhou Hang

  authored
  db33bc8

  Copy full SHA for db33bc8

Commits on Jun 14, 2026

• [model] feat: support Gemma4 multimodal models in RL (GRPO) (#6715)

  Show description for 919b593
  dbuos

  authored
  919b593

  Copy full SHA for 919b593

• [rollout, vllm] fix: preserve MTP drafter weights during hybrid sleep (#6661)

  Show description for 0c2d339
  sunnweiwei

  authored
  0c2d339

  Copy full SHA for 0c2d339

• [rollout, trainer] fix: handle empty token sequences in _pad_token_ids and metric timing (#6675)

  Show description for ccebe7c
  GJWu-zyx

  authored
  ccebe7c

  Copy full SHA for ccebe7c

Commits on Jun 13, 2026

• [reward] fix: only require max_resp_len when DAPO overlong penalty is enabled (#6709)

  Show description for 334d9f8
  

  discobot

  and

  claude

  authored
  334d9f8

  Copy full SHA for 334d9f8