[None][perf] reduce rank-0 GIL contention in disaggregated generation

[lancelly]

What

Reduce per-iteration GIL contention on the disaggregated generation rank-0
process. At high concurrency rank 0 runs the executor loop, the request ingress,
and the attention-DP request broadcast on a single GIL, so host work on rank 0
paces the iteration. Two redundant, hot GIL costs on the executor-loop thread are
removed:

1. safe_broadcast (_torch/distributed/communicator.py) — the root rank
   rebuilt its return value by pickle.loads-ing its own just-serialized bytes,
   a full deep copy of the broadcast payload (which carries the new requests'
   token ids) on every iteration. Return the original object instead. Also drops a
   now-dead size check that compared a length against itself.

2. DefaultADPRouter._balance_requests_across_ranks
   (_torch/pyexecutor/scheduler/adp_router.py) — input_token_ids is a C++
   getter that copies the whole token list on every access, and it was read twice
   per request per iteration (the descending-token sort key and the heap loop).
   Read it once per request.

Both run on the rank-0 executor-loop thread on every iteration at scale.

Impact

Measured on DeepSeek-V4-Pro 7P1D disaggregated (7×CTX-DEP4 + 1×GEN-DEP8, GB300),
concurrency 2880, py-spy on the rank-0 generation worker:

• GEN host_step_time p50: 123.5 ms → ~105 ms (−15%)
• GEN device_step_time p50: 123 ms → ~99 ms (−19%) — under the overlap
  scheduler the device step was being dragged up to the host step, i.e. the
  generation step was host-bound; relieving the host work surfaces the true
  device time
• per-request output speed (SSE) p25: +~6%

No functional change.

Test

• Router change is order-equivalent to the original (same descending-token sort).
• safe_broadcast root return is value-identical: non-root ranks still
  reconstruct the same object from the broadcast bytes; the root already holds it.

[lancelly]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #52976 [ run ] triggered by Bot. Commit: 0279247 Link to invocation

[tensorrt-cicd]

PR_Github #52976 [ run ] completed with state SUCCESS. Commit: 0279247
/LLM/main/L0_MergeRequest_PR pipeline #42210 completed with status: 'FAILURE'

CI Report

⚠️ Action Required:

• Please check the failed tests and fix your PR
• If you cannot view the failures, ask the CI triggerer to share details
• Once fixed, request an NVIDIA team member to trigger CI again

CI Agent Failure Analysis

Link to invocation

[lancelly]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #53053 [ run ] triggered by Bot. Commit: 0279247 Link to invocation

[tensorrt-cicd]

PR_Github #53053 [ run ] completed with state SUCCESS. Commit: 0279247
/LLM/main/L0_MergeRequest_PR pipeline #42267 completed with status: 'SUCCESS'

CI Report

Link to invocation