Auto-enable router dp_aware routing when DP attention is on

[Shi-Dong]

Summary

When DP attention is enabled (--sglang-data-parallel-size > 1, which Miles already forces alongside --sglang-enable-dp-attention), the SGLang engine shards its KV cache across DP ranks — each rank keeps an independent KV pool and its own radix prefix tree. The sgl-router, however, defaults to dp_aware = False, so it dispatches at engine granularity and lets SGLang scatter requests across DP ranks by load. That fragments the radix prefix cache across ranks and can sharply reduce the KV-cache hit rate (a shared prefix that used to be cached once is recomputed on every rank it lands on).

This PR couples the two: in the sgl-router path, when sglang_dp_size > 1, Miles now auto-enables router_args.dp_aware so the router routes at DP-rank granularity (preserving prefix-cache locality).

Behavior

• Triggers only on the sgl-router path (start_router's non-use_miles_router branch). The miles-router path (RadixTreeMiddleware) and externally-managed routers (--sglang-router-ip set) are untouched.
• Honors an explicit --router-dp-aware (already True → left as-is).
• Provides an opt-out via MILES_DISABLE_AUTO_DP_AWARE=1. Because --router-dp-aware is a store_true flag (no native "explicit False"), the env var is the supported way to keep DP attention on while routing per-engine — and it warns when used, since it reintroduces the prefix-cache fragmentation.
• Routing policy is left alone (router already defaults to cache_aware; dp_aware is orthogonal, controlling worker granularity rather than the selection policy).

Test plan

• Unit test (tests/fast/router/test_router_dp_aware.py) covering the four branches: auto-enable when dp_size > 1, no-op when dp_size == 1, respect an explicit dp_aware=True, and opt-out keeps it disabled + warns.
• Helper logic exercised directly across all four cases locally.
• Smoke: a short DP-attention rollout run (dp_size >= 2) confirming the router log shows dp_aware active and the KV-cache prefix hit-rate metric recovers vs. a baseline without the change.

[gemini-code-assist]

Code Review

This pull request introduces automatic DP-aware routing when DP attention is enabled (sglang_dp_size > 1) to preserve prefix-cache locality, along with an opt-out mechanism via the MILES_DISABLE_AUTO_DP_AWARE environment variable. It also adds unit tests to verify this behavior. Feedback was provided regarding a potential TypeError if sglang_dp_size is explicitly set to None, and an AttributeError if router_args lacks the dp_aware attribute, suggesting safer attribute retrieval.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

If args.sglang_dp_size is explicitly set to None (which is common for optional CLI arguments), getattr(args, "sglang_dp_size", 1) will return None instead of the default 1. This will result in a TypeError: '<=' not supported between instances of 'NoneType' and 'int' when evaluating None <= 1.

Additionally, to be more defensive and avoid potential AttributeErrors if router_args does not have the dp_aware attribute, we should use getattr(router_args, "dp_aware", False).

    dp_size = getattr(args, "sglang_dp_size", 1) or 1
    dp_aware = getattr(router_args, "dp_aware", False)
    if dp_size <= 1 or dp_aware:
        return

[guapisolo]

I suggest put this implicit argument setting in arguments.py

[lawrence-harmonic]

This may break PD disaggregation if DP aware does not work correctly with PD disaggregation yet.

[lawrence-harmonic]

DP aware support for PD router is in: smg-project/smg#1522

This needs to be added to https://github.com/radixark/sgl-router-for-miles I think.