Async rl - #273
Draft
meichangsu1 wants to merge 19 commits into
Draft
Conversation
- Rename low-level component methods to match new API (e.g., `submit_sample` → `sample_to_data_plane`, `submit_forward_only` → `forward_only`) - Clarify that token tensors and log-probabilities remain server-side in async RL flow - Add `TWINKLE_TEMPLATE_MODEL_ID` environment variable for DPO template configuration - Remove `_extract_ref_outputs` helper as it's no longer needed with updated API
…trated-async-rl # Conflicts: # src/twinkle/metric/__init__.py # src/twinkle/model/__init__.py # src/twinkle/sampler/vllm_sampler/vllm_sampler.py # src/twinkle/server/model/twinkle_handlers.py # src/twinkle/server/sampler/tinker_handlers.py
meichangsu1
marked this pull request as draft
August 24, 2026 06:46
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
PR type
PR information
本 PR 新增多租户异步 RL 能力,同时支持 YAML 管理的 CLI 模式和客户端编排的 C/S 模式。
主要改动如下:
新增 YAML 管理的异步多 LoRA GRPO 流程:
新增客户端编排的异步 GRPO:
DataRefAPI,用于传递 Rollout 与训练数据。新增 Dynamic Micro-batching:
micro_batch_scale()保持不同 Loss 的归约语义。新增异步 RL 运行组件:
新增 C/S 服务端能力:
改进 Ray GPU 资源管理:
新增 Reward 和 Metrics:
新增 Cookbook 示例:
cookbook/rl/async_rl:YAML 管理的多 LoRA 异步 GRPO。cookbook/client/async_rl:基于 Qwen3.5-4B 的客户端编排异步 GRPO。Experiment results
功能测试结果: