fix dflash converter issues

[shanjiaz]

Purpose

Fixing two separate issues:

1. For z-lab style dflash models, older models have block_size on top level but newer models have them nested under dflash_config.

Tests

Tested both locally.

Checklist

I have filled in:

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan/results, such as providing test command and pasting the results.
• (Optional) The necessary documentation update.
• I (a human) have written or reviewed the code in this pr to the best of my ability.

---

^{Need help on this PR? Tag /codesmith with what you need. Autofix is disabled.}

[coderabbitai]

Important

Review skipped

Auto incremental reviews are disabled on this repository.

Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Pro

Run ID: 87cd4676-889a-4aa3-b38d-b16f5a985547

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

• 🔍 Trigger review

📝 Walkthrough

Walkthrough

The PR updates DFlash conversion to resolve block_size from either checkpoint config layout and changes SpeculatorModel.from_pretrained to accept and forward an explicit verifier argument during external checkpoint conversion.

Changes

Checkpoint conversion and verifier forwarding

Layer / File(s)	Summary
DFlash block-size resolution src/speculators/convert/dflash/converter.py	_build_config reads block_size from either the top-level checkpoint config or dflash_config, raises if neither is present, and uses the resolved value for speculative_tokens and DFlashSpeculatorConfig.block_size.
Explicit verifier forwarding src/speculators/model.py	SpeculatorModel.from_pretrained adds a verifier parameter and passes it to maybe_convert_external_checkpoint instead of reading it from kwargs.

Possibly related PRs

• vllm-project/speculators#617: Refines the same DFlash _build_config path that resolves block_size from multiple checkpoint config locations.

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 50.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title is concise and accurately reflects the main dflash converter fixes.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Description check	✅ Passed	The description matches the changes, covering dflash block_size layout differences and the verifier argument update.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Commit unit tests in branch fix-converter-issues

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[mergify]

The quality checks have failed. Please run make style and make quality under
the root directory to address the lint failures. You will need to install the
dev optional install to get the required linting packages:
https://github.com/vllm-project/speculators/blob/main/CONTRIBUTING.md

[rahul-tuli]

By making verifier a named param and consuming it (no longer in **kwargs), it's never forwarded to the constructor anymore. So EagleSpeculator gets verifier=None, falls back to the config's name_or_path, and tries to load that instead of using the passed-in object. I believe that would break test_eagle_speculator_from_pretrained_different_layers and ..._local_marshalling_different_layers at runtime (they assert model.verifier == mock_verifier)

[rahul-tuli]

LGTM; The failing tests seem to be unrelated to this PR 🚢