Commits

Breadcrumbs

History for

onmain

User selector

All users

All time

Commit history

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

Commits on Jun 12, 2026

• [trainer] fix: use FullyAsyncLLMServerClient for async trainer (#6716)

  Show description for e2cde0a
  wuxibin89

  authored
  e2cde0a

  Copy full SHA for e2cde0a

• [trainer] feat: add unify trainer abstraction for sync and async training (#6710)

  Show description for 45632b5
  wuxibin89

  authored
  45632b5

  Copy full SHA for 45632b5

• [reward] fix: add optional per-sample compute_score timeout to NaiveRewardManager (#6673)

  Show description for 9f18abe
  zqinglin

  authored
  9f18abe

  Copy full SHA for 9f18abe

• [hardware] feat: add ROCm/HIP platform backend (PlatformROCm) (#6702)

  Show description for ab8da13
  

  xiaohong42

  and

  cursoragent

  authored
  ab8da13

  Copy full SHA for ab8da13

• [rollout] fix: clone LoRA weights out of the reused IPC buffer before add_lora (#6688)

  Show description for 173945f
  

  HaozheZhang6

  and

  claude

  authored
  173945f

  Copy full SHA for 173945f

• [fsdp] fix: handle missing chunked top-k config (#6704)

  Show description for 4a0ff06
  Luosuu

  authored
  4a0ff06

  Copy full SHA for 4a0ff06

• [fsdp, trainer] fix: detach model_output and loss metrics to stop per-micro-batch graph retention in actor update (#6699)

  Show description for dbbf085
  YoungZSh

  authored
  dbbf085

  Copy full SHA for dbbf085

Commits on Jun 10, 2026

• [megatron] fix: propagate disabled MTP config to bridge providers (#6626)

  Show description for 41a5244
  Mengyuyang

  authored
  41a5244

  Copy full SHA for 41a5244

• [rollout] fix: propagate ignore_eos to vLLM sampling params (#6630)

  Show description for 15560ea
  Mengyuyang

  authored
  15560ea

  Copy full SHA for 15560ea

• [trainer, megatron] fix: fix MindSpeed repatch config bug (#6674)

  Show description for a1fad97
  chengminhua

  authored
  a1fad97

  Copy full SHA for a1fad97