[megatron] feat: a bunch of optimzation on vram, sequence packing

[ISEEKYAN]

What does this PR do?

add a bunch of optimizations for megatron training, including:

1. aggressive_empty_cache to avoid OOM on hybrid engine. Before this sometimes the cache could use as much as 30GB so bring OOMs.
2. better sequence packing pre/post-process. Before this there are a few times of d2h sync when pre/post-process the sequence packing.
3. make override_ddp_config compatible to mbridge.

The optimized implementations have replaced the old ones, no options needed to enable them.

Add concise overview of what this PR aims to achieve or accomplish. Reference related GitHub issues and PRs that help with the review.

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.

[gemini-code-assist]

Code Review

This pull request introduces optimizations for Megatron training, focusing on VRAM usage and sequence packing. Key changes include a new aggressive_empty_cache utility, optimized sequence packing pre/post-processing, and refactoring of override_ddp_config. A critical issue was identified in verl/utils/memory_utils.py regarding hardcoded GPU device indices, which has been flagged for immediate attention.

[gemini-code-assist]

Hardcoding device index 0 for memory statistics can lead to incorrect behavior in multi-GPU environments where the current device is not GPU 0. The memory functions should operate on the current device. The device.memory_reserved() and device.memory_allocated() functions default to the current device when no device index is provided. Please remove the hardcoded 0.

Suggested change

	before_reserved = device.memory_reserved(0)
	before_allocated = device.memory_allocated(0)
	before_reserved = device.memory_reserved()
	before_allocated = device.memory_allocated()

[gemini-code-assist]

Similar to the previous point, the device index 0 is hardcoded here. This should be removed to ensure the function queries memory statistics for the current device.

Suggested change

	after_reserved = device.memory_reserved(0)
	after_allocated = device.memory_allocated(0)
	after_reserved = device.memory_reserved()
	after_allocated = device.memory_allocated()

[gemini-code-assist]

This function hardcodes the device index 0 for all memory statistics. This will report incorrect information on multi-GPU systems where the current process is not on device 0.

Most torch.cuda memory functions default to the current device if no index is provided. For get_device_properties, you need to explicitly pass the current device index, which you can get via device.current_device().

    device = get_torch_device()
    device_id = device.current_device()

    return {
        "total_memory_gb": device.get_device_properties(device_id).total_memory / 1024**3,
        "reserved_memory_gb": device.memory_reserved() / 1024**3,
        "allocated_memory_gb": device.memory_allocated() / 1024**3,
        "cached_memory_gb": (device.memory_reserved() - device.memory_allocated()) / 1024**3,
        "max_memory_allocated_gb": device.max_memory_allocated() / 1024**3,
        "max_memory_reserved_gb": device.max_memory_reserved() / 1024**3,
    }