[DCP][Bugfix][CI] Fix accuracy issue of DCP when using FLASH_ATTN_MLA

[FENP]

Purpose

#25049 add MTP support for DCP with FA3, but this only works when cp_kv_cache_interleave_size=1.

For FA3 backend, we should check cp_kv_cache_interleave_size and set supports_dcp_with_varlen accordingly.

Test Plan

vllm serve deepseek-ai/DeepSeek-V2-Lite-Chat/ --gpu-memory-utilization 0.9 --tensor-parallel-size 4 --decode-context-parallel-size 4 --cp-kv-cache-interleave-size 64

python ./tests/evals/gsm8k/gsm8k_eval.py

Test Result

• Main

Running GSM8K evaluation: 1319 questions, 5-shot
Evaluating: 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1319/1319 [00:40<00:00, 32.62it/s]

Results:
Accuracy: 0.028
Invalid responses: 0.007
Total latency: 40.447 s
Questions per second: 32.611
Total output tokens: 189017
Output tokens per second: 4673.216

• This PR

Running GSM8K evaluation: 1319 questions, 5-shot
Evaluating: 100%|████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████████| 1319/1319 [00:45<00:00, 28.81it/s]

Results:
Accuracy: 0.639
Invalid responses: 0.002
Total latency: 45.794 s
Questions per second: 28.803
Total output tokens: 161057
Output tokens per second: 3516.982

cc @LucasWilkinson @minosfuture @pisceskkk

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.
• (Optional) Release notes update. If your change is user facing, please update the release notes draft in the Google Doc.

[chatgpt-codex-connector]

Codex usage limits have been reached for code reviews. Please check with the admins of this repo to increase the limits by adding credits.

[gemini-code-assist]

Code Review

This pull request addresses an accuracy issue in Decoded Context Parallelism (DCP) when used with FlashAttention MLA and a cp_kv_cache_interleave_size greater than 1. The fix correctly disables supports_dcp_with_varlen in this configuration, forcing requests into the prefill path and resolving the bug. The code change is correct and directly targets the issue. My main feedback is to enhance the test coverage to include the specific configuration that was manually tested and shown to be fixed, ensuring this scenario is covered by CI to prevent future regressions.

[gemini-code-assist]

The PR description shows a manual test with dcp_size == tp_size (i.e., dcp_multiplier=1) and cp_kv_cache_interleave_size=64 to demonstrate the fix. However, this specific test case seems to be missing from the automated tests after the changes. To prevent future regressions, it would be beneficial to include this configuration in the test suite. You can achieve this by adding 1 to the dcp_multipliers list for the test case with cp_kv_cache_interleave_size=64.

Suggested change

	dcp_multipliers=[0.5],
	dcp_multipliers=[0.5, 1],

[LucasWilkinson]

LGTM