[Bugfix][Qwen][DCA] fixes bug in dual-chunk-flash-attn backend for qwen 1m models.

[sighingnow]

…en 1m models.

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.

Purpose

Fixes a bug in the DCA backend. The error was introduced during rebasing previous PR: #11844

[gemini-code-assist]

Code Review

This pull request addresses a bug in the dual-chunk-flash-attention backend for Qwen 1M models. The changes involve removing an erroneous block_table parameter from several function calls within vllm/attention/backends/dual_chunk_flash_attn.py. My review confirms that this parameter was being passed incorrectly to functions that expect contiguous tensors, and its removal is the correct fix. The changes are consistent and well-contained, resolving the bug introduced during a previous rebase. The code quality is good.

[noooop]

@sighingnow

Please help verify whether using VLLM_ALLOW_LONG_MAX_MODEL_LEN in examples/offline_inference/qwen_1m.py may potentially cause issues leading to nan.

Refer to #20904

…en 1m models.

I didn't even know what kind of machine could run it
And I'm not familiar with the dual-chunk-flash-attention backend either.

Nan can cause the detokener to output !!!!!!!!!!!!!!
very easy to verify.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[sighingnow]

@sighingnow

Please help verify whether using VLLM_ALLOW_LONG_MAX_MODEL_LEN in examples/offline_inference/qwen_1m.py may potentially cause issues leading to nan.

Refer to #20904

…en 1m models.

I didn't even know what kind of machine could run it And I'm not familiar with the dual-chunk-flash-attention backend either.

Nan can cause the detokener to output !!!!!!!!!!!!!! very easy to verify.

This issue should be orthogonal with this PR. I will take a look, but could we merge this PR first to make qwen-1m work?

[youkaichao]

cc @LucasWilkinson

[youkaichao]

LGTM, approving since @sighingnow maintains the dual-chunk attention thing and this change looks reasonable.

[sighingnow]

LGTM, approving since @sighingnow maintains the dual-chunk attention thing and this change looks reasonable.

Thanks! The failed test looks not caused by this PR.