fix: hicache with eagle didn't manage the draft model's kv cache.

[cicirori]

Motivation

When both speculative decoding and HiCache are enabled, the draft model KV cache is not managed by the HiRadix cache.

When the original draft model KV cache is evicted, but the target model hits HiCache, the draft model ends up drafting based on a random KV cache.

Modifications

Add dedicated L2 HiCache support for the draft model KV cache.

Accuracy Tests

• Before the fix
  When HiCache was triggered, the acceptance length dropped for the same request.

• After the fix
  The acceptance length stays almost the same.

see #16964

Benchmarking and Profiling

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

Review Process

1. Ping Merge Oncalls to start the PR flow. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • /tag-run-ci-label, /rerun-failed-ci, /tag-and-rerun-ci
4. After green CI and required approvals, ask Merge Oncalls to merge.

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[cicirori]

/tag-run-ci-label

[cicirori]

/tag-and-rerun-ci

[xiezhq-hermann]

@ispobock should the draft model just re-compute when the cache is already evicted for the draft model?

[hnyls2002]

/rerun-stage stage-b-test-large-1-gpu

[cicirori]

/rerun-stage stage-b-test-large-1-gpu

[github-actions]

✅ Triggered stage-b-test-large-1-gpu to run independently (skipping dependencies).

[github-actions]

🔗 View workflow run