Open
Description
One of the key design goals of slime is to stay compatible with the latest releases of sglang. However, to support certain reinforcement learning (RL) features, we occasionally introduce minor modifications to sglang itself.
This issue serves as a centralized place to track relevant pull requests—both ongoing and merged—that affect or are affected by slime. It aims to facilitate collaboration and maintain alignment between the slime and sglang communities.
Ultimately, our goal is to eliminate the need for maintaining a separate sglang.patch file by upstreaming necessary changes directly into sglang.
We’ll keep this list updated as development progresses.
🔄 Open / Pending PRs
- [RL] support update_weights_from_tensor for mtp sgl-project/sglang#7415
- [WIP][RL] fix fp8 update weight sgl-project/sglang#7421
✅ Merged PRs
- [router] support http2 in router sgl-project/sglang#6487
- [router] Add /list_workers endpoint to router sgl-project/sglang#6366
- [RL] allow weight updation with dp attention enabled sgl-project/sglang#6311
- [RL] Remove the w13 weight_scale and input_scale for UnquantizedEPMoE… sgl-project/sglang#6308
- Add endpoint /abort sgl-project/sglang#5966
or Fix request abortion sgl-project/sglang#6184 - [fix][RL] Remove the incorrect barrier in init_weights_update_group sgl-project/sglang#5914
- [fix][RL] Fix DeepSeekV3ForCausalLM.post_load_weights for multiple update weight sgl-project/sglang#6265
- [RL] support update_weights_from_distributed with different group and multiple weights sgl-project/sglang#7292
- [RL] add --skip-warmup sgl-project/sglang#7416
- [router] add --log-level to sgl-router sgl-project/sglang#6512
- [RL] support abort all and fix abort on waiting queue sgl-project/sglang#6855
or Support updating weights at once by stopping all requests sgl-project/sglang#6698 - [RL] Add --nccl-port to prevent port conflict sgl-project/sglang#7418
- [RL] add pause and continue generation for async rl training sgl-project/sglang#7419
Metadata
Metadata
Assignees
Labels
No labels