[main][refractor] Refractor forward metadata retrieval across DP nodes to reduce redundant padding.

[yx0716]

Before refactoring cross-DP decoding metadata aggregation, clean up the token‐padding logic .

What this PR does：

1. First checks whether any DP instance is in the prefill phase.

2. If in the decode phase and torchair_graph_enabled is true, pads each DP instance’s token count up to the global maximum.

3. If in the prefill phase, or in decode phase with graph mode disabled, returns each DP instance’s original token count without padding.

This reordering removes the previous two‐step padding/unpadding flow and ensures padding only occurs when strictly necessary.

• vLLM version: v0.10.0
• vLLM main: vllm-project/vllm@bd3db7f

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 76.34%. Comparing base (807f089) to head (b269550).
⚠️ Report is 613 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #2062   +/-   ##
=======================================
  Coverage   76.34%   76.34%           
=======================================
  Files         110      110           
  Lines       12473    12473           
=======================================
  Hits         9522     9522           
  Misses       2951     2951           

Flag	Coverage Δ
unittests	76.34% <ø> (ø)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[ApsarasX]

I think the third scenario should be considered, in the decode phase and torchair_graph_enabled is true, but with_prefill is true

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.

[yx0716]

I think the third scenario should be considered, in the decode phase and torchair_graph_enabled is true, but with_prefill is true

• In a single DP instance, TorchAir’s graph mode and chunk-prefill mode are not enabled at the same time, thus prefill and decode will not exist simultaneously.
• Across multiple DP instances, if even one instance enters the prefill stage, all instances automatically revert to non-graph mode.

Given these two scenarios, it seems unlikely that a third case would arise at this point.

[wangxiyuan]

@ApsarasX I'm working on torchair model runner refacor, let's do more work there #2204

[ApsarasX]

@ApsarasX I'm working on torchair model runner refacor, let's do more work there #2204

OK