attention_backend is hard-coded to "fa3", making sglang.attention_backend config option ineffective

[zzs97str]

Hi there,

I noticed that in the current codebase, the attention_backend is hard-coded to "fa3" in verl/workers/rollout/sglang_rollout/sglang_rollout.py line 463 464

attention_backend = "fa3" # hard-coded

However, the configuration already exposes a user-configurable option via:
actor_rollout_ref.rollout.engine_kwargs.sglang.attention_backend

Unfortunately, due to the hard-coded value, the configuration setting is currently ignored, rendering the option non-functional.

[vermouth1992]

You can propose a PR to modify it back and set the default to fa3. I remember to hardcode it for CI issues