[None][fix] Cherry-pick PR #14276: Handle unset attention_dp_relax in ADP routers

[lancelly]

Summary

Cherry-pick of upstream main PR #14276 (3de344d384) onto the feat/deepseek_v4 development branch.

Why: feat/deepseek_v4 does not yet have PR #14276 (merged on main 2026-05-19). Without it, the feat/deepseek_v4 line crashes at every Phase-transition point where the request batch contains multiple SchedulingParams objects with attention_dp_relax = None:

File ".../_torch/pyexecutor/scheduler/adp_router.py", line 517, in route_requests
    sorted_requests = sorted(new_requests, key=get_relax_value)
TypeError: '<' not supported between instances of 'NoneType' and 'NoneType'

This was reproduced on an agg DSV4-Pro RWLT validation run (2 nodes × GB300, TP8/TEP8/MTP3) at the Phase 0 → Phase 1 transition (concurrency 1 → 4) — all 8 ranks crashed simultaneously on both DefaultADPRouter.route_requests (line 232) and KVCacheAwareADPRouter.route_requests (line 517), since both routers share the same get_relax_value helper that dereferences scheduling_params.attention_dp_relax which can be None.

The upstream fix in PR #14276 changes the root cause — the default value of SchedulingParams.attention_dp_relax:

-    attention_dp_relax: Optional[bool] = None
+    attention_dp_relax: bool = True

so the router's sorted(...) no longer sees None keys. Treating unset = "relaxed" (True) matches the semantics: an unset request has no specific rank preference and can be routed anywhere.

[lancelly]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #49805 [ run ] triggered by Bot. Commit: c3539a5 Link to invocation

[tensorrt-cicd]

PR_Github #49805 [ run ] completed with state SUCCESS. Commit: c3539a5
/LLM/main/L0_MergeRequest_PR pipeline #39392 completed with status: 'SUCCESS'

CI Report

Link to invocation