adapt to minimax_m2

[Feng-xiaosuo]

What this PR does / why we need it?

This PR fixes Minimax model loading in vLLM Ascend backend by:

Adding model type check for "minimax" and "minimax_m2" to replace "mlp" prefix with "block_sparse_moe"
Implementing special handling for Minimax expert layer naming conventions
Adding Minimax configuration to packed_modules_model_mapping for proper qkv_proj and experts module handling
Without these changes, Minimax models fail to load on Ascend devices due to incompatible layer naming and module packing.

Does this PR introduce any user-facing change?

Yes. Users can now successfully load and run Minimax models on Ascend hardware with vLLM. This enables inference capabilities for this model family on Ascend devices.

How was this patch tested?

Local Testing:
Verified model loading for minimax-xxx and minimax_m2-xxx model variants on Atlas 800I A2 hardware
Tested inference with sample prompts using vLLM's OpenAI-compatible API server

Benchmark Validation:
Compared throughput and latency metrics against GPU baseline
Verified memory usage stays within expected limits for different batch sizes
Tested multi-card inference scenarios with tensor parallelism

• vLLM version: v0.13.0
• vLLM main: vllm-project/vllm@8be6432

[gemini-code-assist]

Code Review

This pull request adds support for the minimax_m2 model, primarily by updating quantization configurations. The changes involve modifying layer prefixes for MoE layers and adding a packed module mapping for minimax_m2. My review identifies a critical issue in the prefix modification logic for MoE layers that could lead to incorrect quantization for models with multiple experts. I've provided a suggestion to make the implementation more robust.

[gemini-code-assist]

This line appears to have an off-by-one error in the slice index. By slicing up to exp_idx + 1, the expert index is stripped from the prefix. For example, ...experts.1.w1 becomes ...experts. The downstream logic then incorrectly assumes expert 0 due to the hardcoded packed_modules_model_mapping.

This will cause incorrect quantization behavior for any expert with an index other than 0. This is a critical bug for models with multiple experts.

A more robust solution is to slice up to exp_idx + 2 to preserve the expert index in the prefix (e.g., ...experts.1). This would mean the quantization configuration should be defined per-expert (e.g., ...experts.1.weight), which is a cleaner design. This change would also make the experts entry in packed_modules_model_mapping for minimax_m2 unnecessary.

Suggested change

	parts = parts[:exp_idx + 1]
	parts = parts[:exp_idx + 2]

[github-actions]

👋 Hi! Thank you for contributing to the vLLM Ascend project. The following points will speed up your PR merge:‌‌

• A PR should do only one thing, smaller PRs enable faster reviews.
• Every PR should include unit tests and end-to-end tests ‌to ensure it works and is not broken by other future PRs.
• Write the commit message by fulfilling the PR description to help reviewer and future developers understand.

If CI fails, you can run linting and testing checks locally according Contributing and Testing.

[github-actions]

This pull request has conflicts, please resolve those before we can evaluate the pull request.

[realliujiaxu]

Thanks for your contribution! Could you explain why replacing "mlp" prefix with "block_sparse_moe"?

[Feng-xiaosuo]

Thanks for your contribution! Could you explain why replacing "mlp" prefix with "block_sparse_moe"?

In the vLLM Ascend backend, we use a unified quantization loading mechanism that expects specific module naming patterns. For most models (like Llama, Qwen), the expert layers use "mlp" as the prefix (e.g., model.layers.0.mlp.gate_proj).

However, Minimax models have a different architecture:

They use a block-sparse mixture-of-experts (MoE) implementation
Their weight files contain layers named with "block_sparse_moe" prefix (e.g., model.layers.0.block_sparse_moe.gate)
The standard "mlp" prefix doesn't exist in their actual weight files
Without this replacement:

During quantization loading, the system tries to find weights with "mlp" prefix
These weights don't exist in Minimax model files
The quant_config.py module fails to map weight names to modules
Results in KeyError when loading quantized models
This replacement ensures the weight loading path correctly matches Minimax's actual layer naming convention. It's not a generic transformation but a model-specific adaptation required for compatibility with the Ascend backend's quantization infrastructure.