F12/F33: SGLang base_gpu_id is not 0 in RLix mode — doc/code mismatch (draft, needs direction)

[TianyeGGBond]

Draft / discussion vehicle. Issues are disabled on this repo, so this is filed as a draft PR. It contains only the safe documentation fix (Option A). Please pick a direction before it is un-drafted.

The mismatch

The F12 design docstring promises that in RLix / shared-PG mode SGLang runs with base_gpu_id=0 (each rollout worker gets its own CUDA_VISIBLE_DEVICES, so post-CVD it only sees cuda:0..tp-1). The rollout code does not implement this — base_gpu_id is the physical GPU id.

Docstring (miles/ray/placement_provider.py, F33 note, before this PR):

SGLang base_gpu_id MUST be 0 in RLix mode. CVD is set per worker … SGLang itself reads base_gpu_id=0.

Actual code (miles/ray/rollout.py):

• The SGLang actor's env_vars uses NOSET_VISIBLE_DEVICES_ENV_VARS_LIST (all "1") — Ray does not set CUDA_VISIBLE_DEVICES for the rollout actor. Unlike the train path in actor_group.py, there is no per-worker CVD.
• base_gpu_id = int(reordered_gpu_ids[gpu_index]) is the physical id (e.g. 2 for inference pool [2, 3]).
• sglang_engine._to_local_gpu_id(base) returns the id unchanged when CVD is unset, so base_gpu_id stays physical, not 0.

Net: for infer_device_mapping=[2, 3] SGLang launches with base_gpu_id=2, contradicting "MUST be 0".

Options

Option A — fix the docs (this draft). One-hunk change to the placement_provider.py F33 note describing actual behavior (base_gpu_id = physical id; rollout actors do not get per-worker CVD). Zero behavior change.

Option B — fix the code to honor base_gpu_id=0. Give each rollout actor a per-worker CUDA_VISIBLE_DEVICES (mirroring the train path) and pass base_gpu_id=0. This is a real GPU-scheduling change and needs validation on multi-GPU hardware (colocate vs non-colocate, tp>1, multi-node); it should not land unverified. Not included here.

Ask

Which direction do you want? If A, this draft can be reviewed as-is. If B, I'll open a separate PR once it can be validated on GPUs.