Huawei Cloud Ascend Small Model Migration

Overview

This skill guides the migration workflow for small vision models to Ascend NPU, covering structure analysis → migration verification → performance optimization.

Architecture: Model Analysis → Environment Setup → NPU Inference → Performance Profiling → Bottleneck Analysis → Optimization Recommendations

Related Skills:

• huawei-cloud-msmodelslim-model-analysis - Model structure analysis for migration path determination
• huawei-cloud-msot-msopprof-operator-profiler - Operator performance data collection
• huawei-cloud-ascend-profiler-db-explorer - Profiling database analysis for bottleneck identification
• huawei-cloud-ascendc-operator-performance-optim - Optional: AscendC operator optimization for bottleneck operators

Architecture Components

This skill involves the following cloud services and components:

• Ascend NPU: Target hardware for model deployment (Ascend 910B series)
• torch_npu: PyTorch adapter for Ascend NPU
• MSProf: Ascend profiling tool for performance analysis
• Ultralytics: YOLO model framework support
• Docker: Container environment for consistent deployment

Use Cases

Typical Problem Scenarios:

• Migrating vision models from GPU to Ascend NPU
• Deploying YOLO/ResNet/UNet models on Ascend hardware
• Optimizing small model performance on NPU
• Verifying model accuracy after migration
• Identifying performance bottlenecks in computer vision models

Typical User Phrases:

• "Migrate YOLOv8 to Ascend NPU
• "How to run ResNet on Ascend?
• "Optimize UNet inference on NPU
• "Verify model accuracy after migration
• "Analyze performance bottlenecks in my vision model
• "YOLOModelMigrationAscendNPU
• "AscendModel？
• "ModelMigrationNPU？

Scope

Supported:

• Encoder-only architectures (ResNet, VGG, EfficientNet)
• Detection models (YOLO, Faster-RCNN, SSD)
• Segmentation models (UNet, DeepLab)
• Other non-Decoder-only LLM models

Not supported:

• Decoder-only LLM (Qwen, LLaMA, DeepSeek) - requires adapter-based quantization approach
• Understanding VLM text backbone - requires adapter-based quantization approach

Workflow

┌─────────────────────────────────────────────────────────────┐
│  Step 1: Model Structure Analysis                           │
│  → Determine msmodelslim compatibility                      │
│  → Output structure analysis + migration path suggestion    │
└─────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────┐
│  Step 2: Environment Preparation + Migration Verification  │
│  → Configure torch_npu environment                         │
│  → Run inference test                                      │
│  → Verify accuracy                                         │
└─────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────┐
│  Step 3: Performance Data Collection                       │
│  → Collect operator performance data                       │
│  → Output performance data location                        │
└─────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────┐
│  Step 4: Performance Analysis                              │
│  → Analyze profiling data for bottlenecks                  │
│  → Output complete operator time distribution              │
│  → Identify bottleneck and well-performing operators       │
└─────────────────────────────────────────────────────────────┘
                              ↓
┌─────────────────────────────────────────────────────────────┐
│  Step 5: Optimization Suggestions                          │
│  → Provide optimization solutions for bottleneck operators │
│  → Optional operator optimization                          │
└─────────────────────────────────────────────────────────────┘

Step 1: Model Structure Analysis

1.1 Analysis Process

Read the model configuration to analyze:

• Model implementation source (transformers or local directory)
• Architecture type (Decoder-only / Encoder-only / Encoder-Decoder)
• Layer-by-layer loading requirements
• MoE fused weight risks

1.2 Migration Path Determination

1.3 Output Analysis Report

## Model Structure Analysis Result

### Basic Information
- Model Name: xxx
- Architecture Type: Encoder-only / Decoder-only / Encoder-Decoder
- Parameter Count: xxx
- Source: transformers / local directory

### Support Status
- msmodelslim Support: Yes/No
- Recommended Migration Path: torch_npu direct migration / msmodelslim adaptation

### Migration Suggestions
[Specific recommendations]

Step 2: Environment Preparation + Migration Verification

2.1 Default Verification Environment

• Server: ascend-server-01
• Container: skill-the
• Image: quay.io/ascend/vllm-ascend:v0.18.0
• NPU: 8× Ascend 910B3

2.2 Environment Configuration

# Enter container
docker exec -it skill-the bash

# Install dependencies
pip install torch_npu
pip install ultralytics  # For YOLO series
# Or other model-specific libraries

# OpenCV dependencies (if needed)
apt install libgl1 libglib2.0-0

2.3 Migration Verification Script

import torch
import torch_npu

# Check NPU availability
print(f"NPU available: {torch.npu.is_available()}")
print(f"NPU count: {torch.npu.device_count()}")

# Load model
model = ...  # Model loading code
model = model.to('npu:0')

# Inference test
with torch.no_grad():
    output = model(input_tensor)

print(f"Inference success: {output is not None}")

2.4 Output Migration Verification Report

## Migration Verification Result

### Environment Information
- Server: ascend-server-01
- Container: skill-the
- torch_npu Version: xxx
- NPU Status: Normal

### Inference Test
- Model Loading: Success/Failure
- NPU Inference: Success/Failure
- Accuracy Verification: Pass/Fail

### Performance Metrics
- Average Inference Time: xxx ms
- FPS: xxx

Step 3: Performance Data Collection

3.1 Performance Collection Process

Collect operator performance data using the profiling skill:

• On-board collection (device mode)
• Or simulation collection (simulator mode)

3.2 Output Performance Data Location

## Performance Collection Result

### Data Location
- Server: ascend-server-01
- Path: /home/xxx/PROF_xxx/
- Database: msprof_xxx.db
- Collection Time: xxx

### Collection Configuration
- Mode: device / simulator
- NPU: npu:0
- Collection Duration: xxx s

Step 4: Performance Analysis

4.1 Analysis Process

Query and analyze:

1. Top N operator time consumption
2. Group statistics by operator type
3. AI_CPU / AI_CORE / AI_VECTOR_CORE distribution

4.2 SQL Query Example

-- Top 20 operators by time
SELECT op_name, op_type, total_time, call_times
FROM op_summary
ORDER BY total_time DESC
LIMIT 20;

-- Group by type
SELECT op_type, SUM(total_time) as type_time, COUNT(*) as count
FROM op_summary
GROUP BY op_type
ORDER BY type_time DESC;

4.3 Output Performance Analysis Report

## Performance Analysis Result

### Operator Time Distribution (TOP 20)
| Rank | Operator Name | Type | Time(ms) | Percentage | Call Count |
|------|--------------|------|----------|------------|------------|
| 1 | xxx | AI_CPU | xxx | xx% | xxx |
| ... | ... | ... | ... | ... | ... |

### Statistics by Type
| Type | Total Time | Percentage | Operator Count |
|------|------------|------------|----------------|
| AI_CPU | xxx | xx% | xxx |
| AI_CORE | xxx | xx% | xxx |
| AI_VECTOR_CORE | xxx | xx% | xxx |

### Bottleneck Operators (>5% usage)
| Operator | Type | Percentage | Issue |
|----------|------|------------|-------|
| xxx | AI_CPU | xx% | [Specific issue] |

### Well-performing Operators
| Operator | Type | Description |
|----------|------|-------------|
| Conv2D | AI_CORE | High Cube utilization, normal |

Step 5: Optimization Suggestions

5.1 Common Bottlenecks and Solutions

5.2 Output Optimization Suggestions Report

## Optimization Suggestions

### Priority Ranking
| Priority | Operator | Issue | Solution | Expected Gain |
|----------|----------|-------|----------|---------------|
| P0 | xxx | xxx | xxx | xx% |
| P1 | xxx | xxx | xxx | xx% |

### Next Steps
1. [Specific optimization steps]
2. Operator optimization may be performed for bottleneck operators

Complete Report Template

After completing each migration task, output a complete report:

# [Model Name] Ascend Migration Report

## 1. Model Structure Analysis
[Step 1 output]

## 2. Migration Verification
[Step 2 output]

## 3. Performance Collection
[Step 3 output]

## 4. Performance Analysis
[Step 4 output]

## 5. Optimization Suggestions
[Step 5 output]

## Summary
- Migration Status: Success/Failure
- Inference Performance: xxx ms / xxx FPS
- Main Bottlenecks: xxx
- Optimization Direction: xxx

Default Environment

• Server: ascend-server-01:22 (root/Hhuawei@smb)
• Container: skill-the
• NPU: 8× Ascend 910B3 (64G HBM each)
• CANN: cann-version-placeholder.220

Prerequisites

System Requirements

• Python 3.8+
• torch_npu >= 2.0.0
• msprof >= 7.0.0
• ultralytics >= 8.0.0 (for YOLO models)

Environment Check

Prerequisite check: Python3 + torch_npu + msprof required

python3 --version  # Python3 >= 3.8
python3 -c "import torch_npu; print('OK')"  # NPU PyTorch support
python3 -c "import msprof; print('OK')"  # Profiling library

If not installed: pip3 install --user torch_npu msprof ultralytics

Additional System Dependencies

For computer vision models:

apt install libgl1 libglib2.0-0  # OpenCV dependencies

Enhanced Features

Performance Baseline Comparison Module

This skill includes a performance baseline comparison mechanism that compares current model performance against industry-standard baselines:

Features:

• Pre-defined Baselines: Baseline data for common models (YOLOv8, ResNet50, UNet, EfficientNet) on Ascend NPU
• Delta Analysis: Generates performance gap analysis and optimization potential assessment
• Performance Ranking: Compares against similar models in the benchmark database
• Trend Analysis: Tracks performance improvements across migration iterations

Baseline Database:

Delta Analysis Output:

## Performance Baseline Comparison
- Target Model: YOLOv8s
- Baseline Reference: YOLOv8s @ Ascend 910B

### Performance Gap
| Metric | Current | Baseline | Gap |
|--------|---------|----------|-----|
| Latency | 5.2 ms | 4.8 ms | +8.3% |
| Throughput | 192 FPS | 208 FPS | -7.7% |
| Accuracy | 60.2% | 60.6% | -0.4% |

### Optimization Potential
- Priority P0: Reduce latency by optimizing Conv operators
- Priority P1: Improve memory access pattern
- Expected Gain: ~10-15% performance improvement

Resource Estimation & Planning Tool

This skill provides pre-migration resource estimation capabilities:

Features:

• Memory Requirements Prediction: Estimates NPU memory usage based on model size and batch configuration
• Inference Time Estimation: Predicts latency and throughput before deployment
• Batch Size Recommendation: Suggests optimal batch size based on target latency constraints
• Multi-card Scaling Guidance: Provides scaling recommendations for multi-device deployment
• Cost-Benefit Analysis: Evaluates optimization investment vs. expected performance gain

Resource Estimation Output:

## Resource Estimation Report
- Model: YOLOv8s
- Input Resolution: 640x640

### Memory Requirements
| Component | Size |
|-----------|------|
| Model Weights | 21 MB |
| Activation (BS=16) | 480 MB |
| Total Estimated | 501 MB |

### Performance Prediction
| Batch Size | Estimated Latency | Estimated Throughput |
|------------|-------------------|---------------------|
| 8 | 3.2 ms | 250 FPS |
| 16 | 4.8 ms | 208 FPS |
| 32 | 8.5 ms | 188 FPS |

### Recommended Configuration
- Optimal Batch Size: 16
- Target Latency: 4.8 ms
- Expected Throughput: 208 FPS
- Memory Utilization: ~78% of 64GB HBM

Reference Documents

Prerequisites (Duplicate - See Above)

• torch_npu >= 2.0.0 installed
• msprof >= 7.0.0 installed
• Ascend NPU environment configured
• Model code to be migrated

Core Commands

# Analyze model migration feasibility
python3 scripts/analyze_model.py --model /path/to/model

# Verify NPU inference
python3 scripts/verify_npu.py --model /path/to/model --input test.jpg

Parameter Confirmation