[megatron] fix: add missing FP8 padding for router replay

[eternally-z]

What does this PR do?

The router replay path lacks FP8 padding logic. Consequently, enabling router replay during FP8 training leads to incorrect training results. This PR adds the missing FP8 padding support.

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here: ...
• Format the PR title as [{modules}] {type}: {description} (This will be checked by the CI)
  • {modules} include fsdp, megatron, veomni, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data, cfg, reward, fully_async, one_step_off
  • If this PR involves multiple modules, separate them with , like [megatron, fsdp, doc]
  • {type} is in feat, fix, refactor, chore, test
  • If this PR breaks any API (CLI arguments, config, function signature, etc.), add [BREAKING] to the beginning of the title.
  • Example: [BREAKING][fsdp, megatron] feat: dynamic batching

Test

For changes that can not be tested by CI (e.g., algorithm implementation, new model support), validate by experiment(s) and show results like training curve plots, evaluation results, etc.

API and Usage Example

Demonstrate how the API changes if any, and provide usage example(s) if possible.

# Add code snippet or script demonstrating how to use this

Design & Code Changes

Demonstrate the high-level design if this PR is complex, and list the specific changes.

Checklist Before Submitting

Important

Please check all the following items before requesting a review, otherwise the reviewer might deprioritize this PR for review.

• Read the Contribute Guide.
• Apply pre-commit checks: pre-commit install && pre-commit run --all-files --show-diff-on-failure --color=always
• Add / Update the documentation.
• Add unit or end-to-end test(s) to the CI workflow to cover all the code. If not feasible, explain why: ...
• Once your PR is ready for CI, send a message in the ci-request channel in the verl Slack workspace. (If not accessible, please try the Feishu group (飞书群).)
• If your PR is related to the recipe submodule, please also update the reference to the submodule commit via git submodule update --remote or cd recipe && git pull origin main.

[gemini-code-assist]

Code Review

This pull request integrates FP8 padding support into the router replay utilities by passing a use_fp8_padding flag to the preprocessing functions. Feedback identifies a potential AttributeError when accessing tf_config.fp8 directly and suggests using getattr for safety. Additionally, it is recommended to set pre_process=False in merge_router_topk_indices to avoid unnecessary memory allocation and computation for tensors that are not utilized.

[gemini-code-assist]

Accessing tf_config.fp8 directly can lead to an AttributeError if the attribute is missing in certain Megatron versions or custom configurations. It is safer to use getattr(tf_config, 'fp8', None).

Additionally, in merge_router_topk_indices, the calls to preprocess_thd_engine and preprocess_packed_seqs only require the packed_seq_params for the subsequent post-processing step. Setting pre_process=False avoids redundant memory allocation and computation for the processed tensor which is currently discarded.

        use_fp8_padding = getattr(tf_config, 'fp8', None) in ["e4m3", "hybrid"]

        if input_ids.is_nested:
            batch_size = input_ids.shape[0]
            _, packed_seq_params, _ = preprocess_thd_engine(
                input_ids, pre_process=False, use_fp8_padding=use_fp8_padding
            )
            layers_topk_idx = postprocess_thd_engine(
                layers_topk_idx, packed_seq_params, input_ids, batch_size, post_process=True
            )
        else:
            batch_size, seq_len = attention_mask.shape[:2]
            _, packed_seq_params = preprocess_packed_seqs(
                input_ids, attention_mask, pre_process=False, use_fp8_padding=use_fp8_padding
            )

[gemini-code-assist]

Use getattr(tf_config, 'fp8', None) to safely check for FP8 configuration and avoid potential AttributeError if the attribute is not present in the configuration object.

Suggested change

	fp8 = tf_config.fp8
	use_fp8_padding = fp8 in ["e4m3", "hybrid"]
	use_fp8_padding = getattr(tf_config, 'fp8', None) in ["e4m3", "hybrid"]