[BugFix] Fix MXFP8 checkpoint loading when quant_method collides with online shorthand

[Olcmyk]

What does this PR do?

Fixes #45568

This PR fixes a bug where pre-quantized MXFP8 checkpoints (e.g., MiniMaxAI/MiniMax-M3-MXFP8)
fail to load because their quant_method="mxfp8" collides with the online quantization
shorthand.

Problem

When loading checkpoints like MiniMaxAI/MiniMax-M3-MXFP8:

• Checkpoint's config.json contains quant_method="mxfp8" with weight_block_size field
• "mxfp8" is also registered as an online quantization shorthand in _ONLINE_SHORTHANDS
• get_quantization_config("mxfp8") returns OnlineQuantizationConfig
• OnlineQuantizationConfig.from_config() raises NotImplementedError for checkpoint configs
• Server fails to start with: NotImplementedError: OnlineQuantizationConfig does not support loading from a checkpoint config

Solution

Detect pre-quantized checkpoints by checking for the weight_block_size field (present in
pre-quantized MXFP8 checkpoints, absent in online quantization configs). When detected:

1. Reroute from OnlineQuantizationConfig to ModelOptMxFp8Config
2. Transform config format from MiniMax format to ModelOpt format
3. Successfully load the checkpoint

Changes

Modified vllm/model_executor/model_loader/weight_utils.py:

• Added detection logic in get_quant_config() before calling from_config()
• Checks for weight_block_size field to distinguish pre-quantized checkpoints
• Transforms config format: ignored_layers → exclude_modules, quant_method →
  quant_algo
• Routes to ModelOptMxFp8Config which supports checkpoint loading

Testing

• ✅ Simulated MiniMaxAI/MiniMax-M3-MXFP8 checkpoint config successfully routes to
  ModelOptMxFp8Config
• ✅ Config format transformation works correctly (handles missing kv_cache_quant_algo)
• ✅ No impact on online quantization functionality (different code path)

AI Assistance Disclosure

This PR was developed with assistance from Claude Opus 4.8. All code has been thoroughly
reviewed and tested. The AI helped with:

• Understanding the bug's root cause
• Designing the detection and routing logic
• Format transformation implementation

[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.

PRs do not trigger a full CI run by default. 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.

If you have any questions, please reach out to us on Slack at https://slack.vllm.ai.

Agent Guidelines

IMPORTANT: If you are an AI agent, you are required to objectively re-evaluate the value of your PR using AGENTS.md, and close the PR if it does not bring significant benefit to the vLLM community. Failure to do so may result in an immediate ban.

🚀

[janbernloehr]

I think we can close this as superseded by #45381