Huawei Cloud msModelSlim Model Analysis

Overview

• huawei-cloud-msmodelslim-model-adapt

  - Adapter creation based on analysis results

Architecture Components

• msModelSlim: Huawei Cloud's model quantization framework
• Transformers Library: Hugging Face Transformers for model loading
• ModelScope: Model download and management platform
• config.json: Model configuration file for analysis

┌─────────────────────────────────────────────────────────────┐
│           msModelSlim Model Analysis Skill                  │
├─────────────────────────────────────────────────────────────┤
│  ┌──────────────┐    ┌──────────────┐    ┌──────────────┐ │
│  │  Model       │───▶│  Source      │───▶│  Structure   │ │
│  │  Input       │    │  Detection   │    │  Analysis    │ │
│  │  (config)    │    │              │    │              │ │
│  └──────────────┘    └──────────────┘    └──────────────┘ │
│         │                   │                   │          │
│         ▼                   ▼                   ▼          │
│  ┌──────────────┐    ┌──────────────┐    ┌──────────────┐ │
│  │  Type        │    │  MoE         │    │  Risk        │ │
│  │  Classification│   │  Assessment │    │  Assessment │ │
│  └──────────────┘    └──────────────┘    └──────────────┘ │
└─────────────────────────────────────────────────────────────┘

Use Cases

• Assessing model adaptation feasibility before creating msModelSlim adapters
• Analyzing model structure and type classification
• Evaluating MoE compatibility for quantization
• Determining if a model can be quantized with msModelSlim
• Identifying potential risks before adapter development

• "Analyze my model for msModelSlim compatibility"
• "Check if this model can be quantized"
• "Evaluate MoE fused weights risk"
• "Assess model adaptation feasibility"
• "Analyze model structure for quantization"
• "AnalysisModelmsModelSlim"
• "ModelQuantization"
• "CheckMoE"

Scope

• Decoder-only LLM
• VLM text backbone analysis (LLM/text path only)

• Non-transformers implementations
• Multimodal generation models (image/video/audio generation)

Required Input

• Model path or model repository identifier

• config.json

• Optional:

  modeling_*.py

  ,

  model.safetensors.index.json

  in the model directory
• If files are missing locally:
  • Download non-weight files using:

    modelscope download --model <org>/<model> --local_dir ./models/<name> --exclude '*.safetensors'

  • Read

    config.json

    and

    modeling_*.py

    from the download directory as input for analysis.

Hard Requirement: Parse Implementation Source First

1. Read

   config.json

   :
   • Get

     model_type

   • Get

     auto_map

     (if present)
2. Try parsing from transformers:
   • Check if

     transformers

     library supports the

     model_type

     .
   • Check if path exists:

     transformers/models/<model_type>/modeling_<model_type>.py

     .
   • If exists, record as

     transformers

     implementation.
3. If not parsed, try model-local implementation:
   • Check if files pointed by

     auto_map

     exist in the model directory.
   • Check if

     modeling_*.py

     files exist in the model directory.
   • If exists, record as

     model-local

     implementation.
4. If neither path available:
   • Stop analysis.
   • Request user to provide readable model implementation code.

Minimum Workflow

1. Parse implementation source (complete hard requirement above).
2. Determine model type, structural differences, and connections:
   • Type: Pure LLM / Multimodal understanding / Multimodal generation
   • Compare with common Qwen2-like LLMs, record special structural designs (e.g., MoE, non-standard attention, SSM/hybrid blocks, additional heads or parallel branches)
   • Check special structure connections (location, dependencies, serial/parallel/residual connections, impact on backbone traversal)
3. Identify structural features:
   • Decoder layer class, attention/MLP module naming, forward signature
4. Determine features affecting adaptation:
   • Layer traversal path and order
   • Whether layer-by-layer loading is needed
   • MoE fused expert weight risk
   • Quantized model dequantization script risk
   • MTP structure implementation availability and weight handling risk
5. Output structured analysis results (refer to template below).
6. Provide next steps:
   • Proceed to adapter creation workflow
   • Or block and explain what user needs to provide

Model Type, Structural Differences, and Connection Determination

• Pure LLM: Text token input only, backbone is decoder-only language model.
• Multimodal understanding: Contains vision/audio encoders, but generation path centers on text backbone; only text portion can be analyzed and adapted.
• Multimodal generation: Core goal is image/video/audio generation; current workflow does not support, should block and explain reason directly.
• Structural differences only need to record "existence + impact direction", no deep implementation details required.
• Connection relationships should record at minimum: which stage special structure is located in backbone, which modules it connects to, connection type (serial/parallel/residual), and impact on traversal/forward alignment.

MoE Layout Determination

• Non-fused MoE: Experts expanded by module/list (commonly each expert has its own

  gate/up/down

  linear layers).
• Fused MoE: Multiple expert weights packaged as tensor parameters, no longer independent linear layers.
• If any of

  gate/up/down

  stored in

  [..., num_experts, ...]

  or

  [num_experts, ...]

  form, treat as "fused".
• Three-dimensional expert weights (e.g., gate/up/down each fused into 3D parameters) uniformly classified as

  MoE fused

  , with "may need unpack" marked in report.

Required Output: Analysis Report

# Analysis Report

## Model Identification
- Model Path/Repository: {model_path}
- `model_type`: {model_type}
- `architectures`: {architectures}

## Implementation Source Analysis
- Result: `transformers` | `model-local` | `unsupported`
- Basis:
  - Resolved file path: {path}
  - Related configuration fields (`model_type`, `auto_map`): {details}

## Model Features and Specifications
- Hidden size: {hidden_size}
- Number of layers: {num_layers}
- Attention heads / KV heads: {num_heads} / {num_kv_heads}
- Analyze only VLM text portion: Yes/No

## Model Type, Structural Differences and Connections
- Model type: Pure LLM | Multimodal understanding | Multimodal generation
- Special structures vs common Qwen2: {special_structures}
- Special structure connections: {special_structure_connections}
- Impact on adaptation workflow: {structure_impact}

## Layer-by-Layer Loading Assessment
- Need layer-by-layer loading: Yes/No
- Reason: {reason}
- Constraints (memory/runtime environment): {constraints}

## MoE Assessment
- Contains MoE: Yes/No
- Layout type: No MoE | Non-fused MoE | Fused MoE
- Suspected fused keys/modules: {keys}
- Expert weight form: Independent linear layers | Packaged tensors
- Needs unpack: Yes/No

## Adaptation Impact Points
- Decoder traversal path: {traversal_path}
- Attention module naming: {attn_module}
- MLP module naming: {mlp_module}
- `visit/forward` strict alignment points: {alignment_points}

## Quantization and MTP Risk Assessment
- Model already quantized: Yes/No
- Quantization determination basis: {quant_evidence}
- Dequantization script provided: Yes/No
- Dequantization script status: {dequant_status}
- MTP structure exists: Yes/No
- MTP implementation code accessibility: Accessible/Not accessible
- MTP risk description: {mtp_risk}

## Risks and Next Steps
- Risk level: Low | Medium | High
- Blockers: {blockers}
- Recommended next steps:
  - Proceed to adapter creation workflow
  - Or request user to provide implementation code

Risk Identification and User Communication Requirements (Mandatory)

• If identified as "model already quantized", must mark "missing dequantization script" as blocker, explicitly requiring user to actively provide dequantization script before continuing adaptation.
• If MTP structure identified but implementation code inaccessible, must explicitly inform:
  • Agent may not be able to fully implement MTP structure adaptation;
  • To continue, user needs to copy MTP-related weights themselves (map according to user-side implementation).
• When at least one of above two risk types hits,

  risk level

  must not be lower than "Medium".

Pass/Fail Criteria

• Pass: Implementation source is

  transformers

  or

  model-local

  , model type is pure LLM or multimodal understanding, and report is complete; if quantization/MTP risks hit, clear user action requirements given in report.
• Fail: Source not parsed, unsupported implementation type, determined as multimodal generation model, or hits "quantized model without dequantization script" blocking condition.

Enhanced Features

Automated Compatibility Checker

• Migration Blocker Detection: Identifies unsupported operators, custom layers, and framework-specific features
• Early Warning System: Provides early warning for known issues with suggested workarounds
• Compatibility Score: Generates compatibility score with detailed breakdown
• Operator Coverage Analysis: Reports operator coverage rate for Ascend NPU support

## Compatibility Check Result
- Overall Score: XX/100
- Passed: X/XX checks
- Warning: X items require attention
- Blockers: X items preventing migration

### Detailed Results
| Check Item        | Status     | Details                           |
|-------------------|------------|-----------------------------------|
| Operator coverage | ✓ Pass     | 95% of operators supported         |
| Custom layers     | ⚠️ Warning | 2 custom ops need AscendC impl    |
| Weight format     | ✓ Pass     | Standard Hugging Face format      |

Reference Documents

• Analysis Checklist - Analysis verification checklist
• Acceptance Criteria - Functional acceptance criteria
• Verification Method - Verification approach
• Troubleshooting - Common issues and solutions

Prerequisites

• transformers >= 4.40.0 installed
• Model code available for analysis
• Basic understanding of model structure

Analysis Workflow

1. Parse model configuration (

   config.json

   )
2. Determine implementation source (transformers or model-local)
3. Analyze model architecture and structural features
4. Assess MoE layout and fused weight risks
5. Generate structured analysis report
6. Provide adaptation recommendations

Parameter Reference