Repository navigation

[ckpt, trainer, rollout, vllm, doc] feat: decoupled_async trainer with rollout-pulled weights over one-sided reads - #8526

Draft

Luosuu wants to merge 26 commits into

verl-project:mainfrom

Luosuu:feat/transport-weight-registry

Draft

[ckpt, trainer, rollout, vllm, doc] feat: decoupled_async trainer with rollout-pulled weights over one-sided reads#8526
Luosuu wants to merge 26 commits into
verl-project:mainfrom
Luosuu:feat/transport-weight-registry

Commits on Nov 3, 2026
- [ckpt] feat: add weight version registry for publish/pull weight sync
  
  Luosuu
  and
  claude
  committed
  259c1d9
- [ckpt] feat: add transport_pull checkpoint engine (publish/pull over torch.distributed._transport)
  
  Luosuu
  and
  claude
  committed
  41998ac
- [ckpt] feat: nixl_native transport and cross-GPU publish/pull test
  
  Luosuu
  and
  claude
  committed
  3baaf95
- [ckpt, test] run transport_pull correctness test on both transports
  
  Luosuu
  and
  claude
  committed
  ad2b70c
- [ckpt] fix: rename transport_pull option to transport, avoid clash with registry backend arg
  
  Luosuu
  and
  claude
  committed
  f6176e0
- [ckpt, test] transport_pull GPU test: use sdpa, always shut down ray
  
  Luosuu
  and
  claude
  committed
  916195e
- [ckpt, test] transport_pull GPU test: let UCX choose transports
  
  Luosuu
  and
  claude
  committed
  38ebe17
- [ckpt] feat: transport_pull casts fp32 weights to rollout_dtype before publishing
  
  Luosuu
  and
  claude
  committed
  4d6fdbc
- [ckpt] transport_pull: default to pinned-host slots and nixl_native
  
  Luosuu
  and
  claude
  committed
  ba67436
- [trainer, ckpt] feat: decoupled weight sync for v1 separate_async
  
  Luosuu
  and
  claude
  committed
  2e2775e
- [doc] feat: design doc for decoupled weight sync
  
  Luosuu
  and
  claude
  committed
  4566765
- [doc] decoupled weight sync: point follow-ups at the prompt dispatcher design
  
  Luosuu
  and
  claude
  committed
  22df432
- [rollout, vllm] fix: send host-memory weights larger than the bucket over CUDA IPC
  
  Luosuu
  and
  claude
  committed
  28899ec
- [trainer, rollout] decoupled weight sync: agent lifecycle, freed IPC staging, vLLM agent test
  
  Luosuu
  and
  claude
  committed
  d62c60a
- [test] vLLM agent test: normalize chat-template token ids
  
  Luosuu
  and
  claude
  committed
  aa93c33
- [test] vLLM agent test: keep the replica busy across the update
  
  Luosuu
  and
  claude
  committed
  3be4a63
- [doc] decoupled weight sync: add Qwen3-8B results
  
  Luosuu
  and
  claude
  committed
  7126898
- [trainer] feat: decoupled_async trainer mode with rollout-side prompt dispatcher
  
  Luosuu
  and
  claude
  committed
  25efc9a
- [doc] decoupled weight sync: describe the decoupled_async mode and prompt dispatcher
  
  Luosuu
  and
  claude
  committed
  923edaf
- [trainer] test: prompt dispatcher and decoupled_async mode; fix missing on_sample_end
  
  Luosuu
  and
  claude
  committed
  fadb268
- [ckpt] test: registry Ray test reuses an existing Ray session
  
  Luosuu
  and
  claude
  committed
  5d5dedc
- [trainer] decoupled_async: count rollout GPUs in throughput, as separate_async does
  
  Luosuu
  and
  claude
  committed
  90c9b79
- [doc] decoupled_async: add 0.5B dispatcher results and the path to rollout outside the trainer's Ray job
  
  Luosuu
  and
  claude
  committed
  29a958a
- [ckpt] fix: update agent always loads the first published version
  
  Luosuu
  and
  claude
  committed
  9ad282e
- [doc] decoupled_async: lead with what is decoupled, add 8B bypass results and the vLLM deployment path
  
  Luosuu
  and
  claude
  committed
  39ffacf
- [ckpt, trainer] fix: address review findings on decoupled_async
  
  Luosuu
  and
  claude
  committed
  788c1ca

Commits

Commits on Nov 3, 2026