fix: handle non-contiguous tensors in IPC weight refit

[jlcanta]

What does this PR do ?

Make pack_tensor and packed_broadcast_producer accept non-contiguous tensors yielded by the params iterator, instead of crashing with RuntimeError: view size is not compatible with input tensor's size and stride.

Issues

Fixes #2417

Usage

No user-facing API change. Internal helpers (stream_weights_via_ipc_zmq_impl.pack_tensor and packed_broadcast_producer) become robust to non-contiguous inputs:

# before: RuntimeError
tensor.data.view(-1).view(dtype=torch.uint8)

# after: works whether the tensor is contiguous or not
tensor.data.reshape(-1).view(dtype=torch.uint8)

Before your PR is "Ready for review"

Pre checks:

• Make sure you read and followed Contributor guidelines
• Did you write any new necessary tests? → tests/unit/models/policy/test_utils.py::TestStreamWeightsViaIPC::test_stream_weights_via_ipc_zmq_impl_non_contiguous
• Did you run the unit tests and functional tests locally? → ran ruff check and ruff format --check on the touched files; the unit test requires CUDA which I don't have locally
• Did you add or update any necessary documentation? → no docs change needed (internal helper, behaviour is now strictly more permissive)

Additional Information

Root cause. stream_weights_via_ipc_zmq_impl consumes tensors from megatron_bridge.export_hf_weights. For the Qwen3ForCausalLM converter, QKVMapping.megatron_to_hf → split_qkv_weights finishes with q.reshape(-1, hidden_size) (and same for k, v). Tensor.reshape is allowed to return a view with strides that are compatible with reshape but not with view. The next step in NeMo-RL was tensor.data.view(-1), which raises.

Fix.

1. nemo_rl/models/policy/utils.py — pack_tensor now uses reshape(-1) instead of view(-1). reshape returns the same view in the contiguous case (zero copy) and falls back to a copy when not — and since the next operation is .copy_() into a buffer, the extra copy in the rare non-contiguous case has no extra cost.
2. nemo_rl/utils/packed_tensor.py — packed_broadcast_producer adds .contiguous() before .view(torch.uint8).view(-1) (defense in depth — same root cause).
3. tests/unit/models/policy/test_utils.py — refactored the existing IPC roundtrip test into a reusable helper _run_stream_weights_roundtrip, and added test_stream_weights_via_ipc_zmq_impl_non_contiguous that exercises the helper with three kinds of non-contiguous inputs (transpose, strided slicing, permute). Without the fix, these new inputs trigger the original RuntimeError.

Tested manually.

Model	converter_type	Result before fix	Result after fix
Qwen/Qwen3-1.7B	Qwen2ForCausalLM	refit works	refit works
Qwen/Qwen3.5-2B	Qwen3ForCausalLM	crashes in pack_tensor	refit works

Compatibility. Strictly additive — when the input tensor is already contiguous (the common path) the behaviour is identical to before.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[youngeunkwon0405]

/ok to test 8384171

[youngeunkwon0405]

LGTM

@terrykong, I think I had a similar issue when I tested the qwen3.5 model, and the fix was pretty much similar to this one at that time.

[youngeunkwon0405]

/okay to test e1d8e95

[youngeunkwon0405]

@terrykong, I believe, to run QWEN3.5 we need this fix. It is almost one line change, could you please help reviewing?