[V1] Fix local chunked attention always disabled

[sarckk]

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

#21188 and #19351 made similar and conflicting changes around self.use_irope in Attention layer, causing self.use_irope to always be False in V1:

self.use_irope = extra_impl_args.pop("use_irope", False)
...
self.use_irope = extra_impl_args.get("use_irope", False)

We should not pop use_irope in V0 as attention backends still expect use_irope as an arg (example)

Test Plan

ruler niah_multikey_2

VLLM_USE_V1=1 lm_eval --model vllm --tasks niah_multikey_2 --model_args pretrained=meta-llama/Llama-4-Scout-17B-16E-Instruct,tensor_parallel_size=4,max_model_len=256000  --metadata='{"max_seq_lengths":[4096,8192,16384,32768]}' --batch_size auto  > /tmp/test_irope_fix.log 2>&1 &

Test Result

baseline:

|---------------|------:|------|-----:|-----:|---|----:|---|------|
|niah_multikey_2|      1|none  |     0| 16384|↑  |0.980|±  |   N/A|
|               |       |none  |     0| 32768|↑  |0.000|±  |   N/A|
|               |       |none  |     0|  4096|↑  |1.000|±  |   N/A|
|               |       |none  |     0|  8192|↑  |0.996|±  |   N/A|

This PR:

|     Tasks     |Version|Filter|n-shot|Metric|   |Value|   |Stderr|
|---------------|------:|------|-----:|-----:|---|----:|---|------|
|niah_multikey_2|      1|none  |     0| 16384|↑  |0.980|±  |   N/A|
|               |       |none  |     0| 32768|↑  |0.944|±  |   N/A|
|               |       |none  |     0|  4096|↑  |1.000|±  |   N/A|
|               |       |none  |     0|  8192|↑  |0.996|±  |   N/A|

cc: @luccafong @minosfuture @houseroad @yeqcharlotte

[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.

🚀

[gemini-code-assist]

Code Review

This PR correctly fixes a bug where self.use_irope was always being set to False due to being overwritten. While the fix is logically sound for the stated problem, I've identified a potential critical issue with the underlying logic of using pop vs get for V1 and V0 code paths. This could be a latent bug related to how quantization parameters are handled. My review includes a suggestion to make the behavior consistent and likely more correct.

[facebook-github-bot]

@vladmihailescu has imported this pull request. If you are a Meta employee, you can view this in D78782520.

[vladmihailescu]

Importing this diff internally for an A/B perf test for Llama4 Maverick on H100

[yeqcharlotte]

Good catch @sarckk! Are we missing some unit test coverage for local attention? I would expect some test failure when we disable things.

[LucasWilkinson]

Good catch!! I appreciate you fixing this! Thanks for the contribution

[sarckk]

Good catch @sarckk! Are we missing some unit test coverage for local attention? I would expect some test failure when we disable things.

yes, I will add some unit test to catch this in the future (EDIT: added in #21478)