refactor: mirror slime structure in update_weight + vllm_rollout

[aoshen02]

Summary

• update_weight_from_distributed.py: Add update_weight_metrics + pop_metrics(), _on_chunk() hook, restructure _send_weights() dual-pass to call _on_chunk before each broadcast, align _iter_non_expert_chunks buffer accounting to convert-first-then-measure (matches slime), remove _send_hf_chunk/_update_weights_vllm_packed wrappers, remove early-exit guard in _ep_gather_and_convert
• vllm_rollout.py: Add docstrings and inline comments matching slime@44d29ee (generate_rollout, generate_rollout_async, eval_rollout_single_dataset, generate_and_rm_group, generate_and_rm)
• vllm_utils/: No changes needed — vllm_config.py already identical, arguments.py and vllm_engine.py are genuine sglang→vLLM divergences

Reference: slime commit 44d29ee59dfa1489c369937779cbbdb1798ce701

Test plan

• ruff check + ruff format pass
• tests/unit/backends/megatron_utils/update_weight/test_update_weight_from_distributed.py — 22/22 pass
• tests/unit/rollout/test_vllm_rollout.py — 26/26 pass
• Integration test on h200 (colocate + non-colocate weight sync)

🤖 Generated with Claude Code

[gemini-code-assist]

Code Review

This pull request refactors the weight update and rollout generation logic, introducing subclass hooks, weight update metrics, and improved documentation/comments across the codebase. A critical issue was identified in update_weight_from_distributed.py where removing the safety check on pbar before calling pbar.update(1) can lead to an AttributeError when pbar is None.

[gemini-code-assist]

Unconditionally calling pbar.update(1) will raise an AttributeError if pbar is None. Since pbar is typed as tqdm | None = None and can be None (for example, on non-PP-source ranks or when not provided), we should restore the if pbar is not None: guard.

Suggested change

	pbar.update(1)
	if pbar is not None:
	pbar.update(1)