fix: avoid redundant device-to-device copy in NCCL weight-transfer receiver

[dumko2001]

Summary

• receive_state_dict (src/prime_rl/inference/vllm/worker/nccl.py) does concatenated[...].view(shape).clone() before yielding each tensor, and every consumer then copies that tensor straight into its destination parameter. So each weight is copied twice on-device on every sync.
• Both consumers copy eagerly: load_weights_kernel calls param.copy_(tensor), and vLLM's load_weights uses default_weight_loader (param.data.copy_(...)). concatenated stays alive for the whole dtype group, so a plain view is valid until the consumer has copied from it and the clone is pure overhead.
• Fix: yield concatenated[offset : offset + numel].view(shape) without .clone().

Verification

Microbenchmark of the split-and-consume pattern on an RTX 3090 (torch 2.8, CUDA 12.8), a 4 GB bf16 group in 2000 tensors:

clone + copy:  54.8 ms   ( 73 GB/s)
view + copy:   27.6 ms   (145 GB/s)

About 2x on the copy, scaling with checkpoint size (~0.7 s/sync at 100 GB). Peak memory is unchanged — the allocator reuses the transient clone — so this is bandwidth, not memory.

---

Note

Low Risk
Single-path performance change with unchanged copy semantics; views remain valid for the lifetime of the dtype-group buffer.

Overview
receive_state_dict in nccl.py no longer calls .clone() on each slice of the per-dtype concatenated broadcast buffer. It yields view(shape) slices only, with comments noting that load_weights_kernel and vLLM’s load_weights already copy_ into destination parameters before the next tensor is consumed, while concatenated stays alive for the whole dtype group.

The per-tensor try/finally that deleted each yielded tensor was removed as unnecessary with views.

^{Reviewed by Cursor Bugbot for commit 28db8ac. Bugbot is set up for automated code reviews on this repo. Configure here.}