fix(vllm): default attention backend to TRITON_ATTN on ROCm

[amd-fuweiy]

What

create_async_llm (roll/third_party/vllm/__init__.py) defaults attention_config.backend to FLASH_ATTN when the caller does not pass one. This PR keeps that default on every non-ROCm platform and uses TRITON_ATTN on ROCm.

Why

On ROCm the FLASH_ATTN backend depends on the CUDA-only vllm_flash_attn package. vLLM's ROCm path returns None from get_flash_attn_version() on purpose (vllm/v1/attention/backends/fa_utils.py, "ROCm doesn't use vllm_flash_attn"). FlashAttentionImpl.forward then asserts that value is not None, so every rollout engine fails during profile_run, before the first training step:

AssertionError: FlashAttention version not detected.

TRITON_ATTN is the Triton FlashAttention implementation vLLM ships as its ROCm default.

Behavior

• CUDA and other platforms: unchanged, still FLASH_ATTN.
• ROCm: TRITON_ATTN.
• An explicit attention_config from the caller still takes precedence, because the existing "attention_config" not in kwargs guard is untouched.

Without this change, ROCm users have to set actor_infer.strategy_args.strategy_config.attention_config.backend: TRITON_ATTN in every pipeline config.

Testing

• 8x AMD Instinct MI308X (gfx942), base image rocm/primus:v26.7 (torch 2.12.0+rocm10.0.0), vLLM v0.31.0.
• ROLL RLVR pipeline with vLLM rollout and an FSDP2 trainer in colocate mode, Qwen2.5-0.5B. Smoke run of 3 steps at learning rate 0: engines initialize, the trainer-to-vLLM weight sync is bit-exact, and the checkpoint save/resume works.
• The same pipeline without this change fails at engine init with the assertion above.
• Not tested on CUDA in this PR; the CUDA branch of the expression is the existing default.

[CLAassistant]

All committers have signed the CLA.