[Config] add "qwen" as a native eagle3 target supported model

[lec77]

Purpose

Add "qwen" to the eagle3_target_supported model list in config.py, so that Qwen models can run with Eagle3 by setting speculative_config.
Previously, "qwen" was only enabled for Eagle3 when draft_model_config was provided with SpeculatorsConfig.

The architectures of Qwen3-Eagle3 drafter models listed in EAGLE project are all LlamaForCausalLMEagle3:

• AngelSlim/Qwen3-8B_eagle3
• Tengyunw/qwen3_8b_eagle3
• AngelSlim/Qwen3-1.7B_eagle3

This needs to be added to the model registry for compatibility.

Test Plan

• Run Eagle3 inference with a "qwen" model without Speculatorsconfig (Qwen/Qwen3-8B + AngelSlim/Qwen3-8B_eagle3).
• Confirm that the model loads successfully and speculative decoding works as intended.

Test Result

total_num_output_tokens: 1000
num_drafts: 352
num_draft_tokens: 1056
num_accepted_tokens: 680
mean acceptance length: 2.93
--------------------------------------------------
acceptance at token 0: 1.00
acceptance at token 1: 0.62
acceptance at token 2: 0.31

[gemini-code-assist]

Code Review

This pull request adds 'qwen' to the list of natively supported target models for Eagle3 speculative decoding. The change simplifies the configuration logic by removing a conditional check that previously enabled Qwen support only when a SpeculatorsConfig was used. The updated code is more direct and improves maintainability. Based on the provided description and test results, this change appears correct and aligns with the goal of broader model support for Eagle3. I have no major concerns with this change.

[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.

🚀

[hmellor]

Could you please share the command used for testing?

[lec77]

Sure!
I ran the tests using the following command:

pytest test_spec_decode.py -v -k qwen3_eagle3

The acceptance_rate and related metrics were obtained by calling spec_llm.get_metrics() and computing the statistics:

try:
    metrics = spec_llm.get_metrics()
except AssertionError:
    print("Metrics are not supported in the V0 engine.")
    return

total_num_output_tokens = sum(
    len(output.outputs[0].token_ids) for output in spec_outputs
)
num_drafts = 0
num_draft_tokens = 0
num_accepted_tokens = 0
acceptance_counts = [0] * 3
for metric in metrics:
    if metric.name == "vllm:spec_decode_num_drafts":
        assert isinstance(metric, Counter)
        num_drafts += metric.value
    elif metric.name == "vllm:spec_decode_num_draft_tokens":
        assert isinstance(metric, Counter)
        num_draft_tokens += metric.value
    elif metric.name == "vllm:spec_decode_num_accepted_tokens":
        assert isinstance(metric, Counter)
        num_accepted_tokens += metric.value
    elif metric.name == "vllm:spec_decode_num_accepted_tokens_per_pos":
        assert isinstance(metric, Vector)
        for pos in range(len(metric.values)):
            acceptance_counts[pos] += metric.values[pos]

print("-" * 50)
print(f"total_num_output_tokens: {total_num_output_tokens}")
print(f"num_drafts: {num_drafts}")
print(f"num_draft_tokens: {num_draft_tokens}")
print(f"num_accepted_tokens: {num_accepted_tokens}")
acceptance_length = 1 + (num_accepted_tokens / num_drafts) if num_drafts > 0 else 1
print(f"mean acceptance length: {acceptance_length:.2f}")
print("-" * 50)

# print acceptance at each token position
for i in range(len(acceptance_counts)):
    acceptance_rate = acceptance_counts[i] / num_drafts if num_drafts > 0 else 0
    print(f"acceptance at token {i}: {acceptance_rate:.2f}")

[hmellor]

Thanks for adding the test, LGTM as long as the relevant tests pass. (Please merge from main to resolve some of the unrelated errors)

[hmellor]

Can you make sure pre-commit passes?

[lec77]

Sorry, I am not familiar with the process. I will fix it.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @lec77.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[DarkLight1337]

Sorry there is merge conflict again, can you resolve it?

[DarkLight1337]

Looks like https://buildkite.com/vllm/ci/builds/26496/steps/canvas?sid=0198934d-abd0-4623-ac97-7ab5f3be8cfc is caused by this PR, can you fix it?

[lec77]

Thanks for pointing this out. I’ll submit a fix in a follow-up PR.

Looks like https://buildkite.com/vllm/ci/builds/26496/steps/canvas?sid=0198934d-abd0-4623-ac97-7ab5f3be8cfc is caused by this PR, can you fix it?

[DarkLight1337]

See #22611