[RFC]: Modulewise weight reload

Finding: In vLLM RFC #59502 (modulewise weight reload), an external reviewer reports three possible correctness gaps in the author's prototype branch. They concern RL weight updates through start_weight_update, receive_weights and finish_weight_update, and the version is UNKNOWN (prototype branch, not a release). This is an external report and hypothesis. KFA did not reproduce it, and the claims are not confirmed by maintainers. Source: https://github.com/vllm-project/vllm/issues/59502 Context: The RFC says layerwise reload under NCCL, IPC and sharded RDT buffers incoming tensors, can leave derived state stale under CUDA graphs, and can fail silently. It proposes deferring only post-load processing, adding refresh() for derived values, loading directly into live bytes, and adding integrity checks including a dirty flag, required keys and all-ranks agreement. Memory numbers for Kimi K3 are computed from tensor shapes, not measured at full size. Validation of about 25 configurations on GB200 is self-reported by the author (https://github.com/vllm-project/vllm/issues/59502). Reviewer claims from a 2026-10-10 comment (https://github.com/vllm-project/vllm/issues/59502#issuecomment-6092344583), based on reading the prototype code: 1. The dirty flag reportedly clears on any successful finish, including one that sent no weights. After a half-failed update, an empty start/finish could re-enable serving on mixed weights. With delta updates, the next sparse round could treat that mixed state as a valid base. 2. Required keys are reportedly checked per tensor name, not per shard. A fused QKV linear that receives only q may complete with zero-filled k/v rows. A duplicated shard could make a module complete early. 3. refresh_derived_state reportedly does not cover the attention phase. MLA W_UK_T/W_UV copies and FlashInfer bmm scales could stay stale after session-less sparse NCCL patches in some configurations. The reviewer adds that sparse patches write checkpoint-coordinate values, so they are only correct where the live layout equals the checkpoint layout. Related plan changes: the author's later comment says a follow-up draft RFC, #60931, would remove reload_mode and replace the detection mechanism with a verifier. It also adds integrity items in PR 1c, such as a kernel-identity check and start-time module-set agreement (https://github.com/vllm-project/vllm/issues/59502#issuecomment-6091263176). Another comment says the open PRs #60096, #57733 and #60009 can coexist with the stack (https://github.com/vllm-project/vllm/issues/59502#issuecomment-6074356641). Diagnosis: none is confirmed. If you use the prototype, treat these as unverified review items to check against that branch. No fix or release is cited, and no workaround is given.