[vllm, rollout] fix: handle non-contiguous weights in bucketed transfer

[Mecoli1219]

What does this PR do?

Fixes vLLM rollout weight updates when the actor exports a valid but non-contiguous tensor. The bucketed weight sender previously packed weights with weight.view(-1).view(torch.uint8), which crashes for strided tensors with:

RuntimeError: view size is not compatible with input tensor's size and stride

This PR adds a small byte-packing helper that first compacts the tensor with .contiguous() before reinterpreting it as torch.uint8. This keeps the common contiguous path cheap while allowing exporters such as Megatron-Bridge to return strided tensor views.

Checklist Before Starting

• Search for similar PRs. Paste at least one query link here:
  • https://github.com/verl-project/verl/pulls?q=is%3Apr+is%3Aopen+bucketed_weight_transfer
  • https://github.com/verl-project/verl/pulls?q=is%3Apr+is%3Aopen+non-contiguous+tensor+update_weights
  • https://github.com/verl-project/verl/pulls?q=is%3Apr+is%3Aopen+view+size+is+not+compatible+input+tensor+stride
• 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

• Added a unit test for it.
• Tested with the latest megatron-bridge v0.5.0, and verified it is working.

API and Usage Example

No API or user-facing usage changes.

Design & Code Changes

• Added a helper who calls weight.contiguous().view(-1).view(torch.uint8) so valid strided/non-contiguous tensors can be serialized into the contiguous bucket buffer.

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.
• [ x 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.

[CLAassistant]

All committers have signed the CLA.

[gemini-code-assist]

Code Review

This pull request introduces a helper function _weight_to_bytes to handle non-contiguous (strided) tensors when converting weights to bytes, along with a corresponding unit test. The reviewer suggests optimizing this process by directly copying the weight into a reshaped view of the destination buffer, which avoids the temporary GPU memory allocation and extra copy overhead caused by calling .contiguous().

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

Instead of making a contiguous copy of the non-contiguous weight tensor via _weight_to_bytes(weight) (which calls weight.contiguous()), we can avoid any temporary GPU memory allocations and extra copy overhead by viewing the destination buffer slice with the target dtype and shape, and copying the weight directly.

PyTorch's copy_ natively supports copying from non-contiguous/strided source tensors to contiguous destination tensors.

Why this is better:

1. Zero Temporary Allocations: weight.contiguous() allocates a temporary tensor of the same size as weight on the GPU. For large models, this can cause GPU memory spikes and potential Out-Of-Memory (OOM) errors.
2. Fewer Copies: It reduces the operation from two copies (weight -> temporary contiguous tensor -> self.buffer) to a single direct copy (weight -> self.buffer).

Suggested change

	self.buffer[offset : offset + weight.nbytes].copy_(_weight_to_bytes(weight), non_blocking=True)
	self.buffer[offset : offset + weight.nbytes].view(dtype=weight.dtype).view(weight.shape).copy_(weight, non_blocking=True)

[wuxibin89]

RuntimeError: view size is not compatible with input tensor's size and stride

@Mecoli1219 In which case is parameter non-contiguous? I never meet this error before.

[Mecoli1219]

I met this error when using megatron as the trainer via megatron-bridge v0.5.0. I am suspecting the gate tensors here is the one that causing the non-continuous tensor, but need to verify it though.