[rollout] fix: apply copy_to_local before init hf config

[ZornWang]

Change-Id: Ic0ddfdfa13a38a56571b9c59125e9ebeea5c7802

What does this PR do?

• Fixed a bug where the original HDFS path was passed due to not using copy_to_local when initializing the hf config.

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, sglang, vllm, rollout, trainer, ci, training_utils, recipe, hardware, deployment, ray, worker, single_controller, misc, perf, model, algo, env, tool, ckpt, doc, data
  • 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).)

[gemini-code-assist]

Code Review

The pull request correctly fixes a bug where remote HDFS paths were used instead of local paths for initializing Hugging Face configurations and tokenizers. This is achieved by consistently using the copy_to_local utility. However, the implementation introduces redundant calls to copy_to_local, which can lead to performance degradation during worker startup, especially in a distributed environment. My review includes a suggestion to optimize this by ensuring each unique path is copied only once.

[gemini-code-assist]

While the fix is correct in using copy_to_local for all relevant paths, it introduces potential performance issues by calling copy_to_local multiple times for what might be the same path. If self.path, self.tokenizer_path, and self.hf_config_path point to the same remote resource, copy_to_local will be executed three times.

Although copy_to_local has caching mechanisms, it still involves overhead like hashing, file system checks, and locking, which can be costly, especially during the initialization of many workers in a distributed environment.

To avoid this redundancy and improve startup performance, you can collect all unique paths and copy each one only once.

        # Efficiently copy remote paths to local, avoiding redundant copies.
        remote_paths = {self.path, self.tokenizer_path, self.hf_config_path}
        local_path_map = {p: copy_to_local(p, use_shm=self.use_shm) for p in remote_paths}

        self.local_path = local_path_map[self.path]

        # constuct tokenizer
        local_tokenizer_path = local_path_map[self.tokenizer_path]
        self.tokenizer = hf_tokenizer(local_tokenizer_path, trust_remote_code=self.trust_remote_code)
        self.processor = hf_processor(local_tokenizer_path, trust_remote_code=self.trust_remote_code)

        local_hf_config_path = local_path_map[self.hf_config_path]
        self.generation_config = get_generation_config(local_hf_config_path, trust_remote_code=self.trust_remote_code)

        # constuct hf_config
        attn_implementation = self.override_config.get("attn_implementation", "flash_attention_2")
        self.hf_config = AutoConfig.from_pretrained(
            local_hf_config_path, trust_remote_code=self.trust_remote_code, attn_implementation=attn_implementation
        )

[vermouth1992]

Could you assign a variable to self? e.g., self. local_tokenizer_path

[ZornWang]

Could you assign a variable to self? e.g., self. local_tokenizer_path

OK, I will do so.

[ZornWang]

Could you assign a variable to self? e.g., self. local_tokenizer_path

Done, plz check if I'm doing right

[vermouth1992]

Have to add to _mutable_fields as well

[ZornWang]

Have to add to _mutable_fields as well

@vermouth1992 Done