Error Analyzer - Ascend NPU

Quick Start

Input Format

Expect user to provide error information in this format:

## Error Information
[Error message / log / stack trace]

## Environment
[OS, version, library versions, NPU info, etc.]

## Context
[What were you trying to do?]

## Repository Path (optional)
[Path to repository for source code reference - if available]

Workflow

1. Parse - Extract error details using the parse_error.py script or manually
2. Match - Compare against known error patterns in references
3. Analyze - Determine root cause using debugging checklist
4. Research - Optionally search repository for source code context
5. Respond - Generate customer-friendly response using templates

Using the Parser Script

Run the parsing script to extract structured information:

# From text
python scripts/parse_error.py "Error: Module not found"

# From file
python scripts/parse_error.py --file error.log

# Interactive mode
python scripts/parse_error.py --interactive

# Output formats
python scripts/parse_error.py --output json   # JSON
python scripts/parse_error.py --output markdown  # Markdown
python scripts/parse_error.py --output summary   # Short summary

Multi-Repository Support

This skill can analyze errors against multiple repositories by accepting a repository path parameter:

Step 1: Identify Error Context

Determine which repository the error relates to:

• MindSpeed-RL: Reinforcement learning on Ascend NPU
• MultimodalSDK: Multimodal LLM preprocessing
• Vision SDK: Image/video processing on Ascend
• AgentSDK: Agent framework integration

Step 2: Configure Repository Path

When user provides a repository path, search for:

1. Error message in source code (grep for error strings)
2. Related configuration or usage patterns
3. Recent changes that might cause the issue

Step 3: Cross-Reference

Use the repository path to:

• Find exact line numbers in stack traces
• Identify version-specific behaviors
• Check for known issues in the codebase

Error Pattern Matching

When analyzing errors:

1. Extract the key error type and message
2. Match against patterns in error-patterns.md
3. For Ascend-specific errors, check sdk-knowledge.md
4. Look for version mismatches, missing dependencies, NPU issues
5. Check for known issues in the error domain

Ascend NPU Specific Errors

This skill specializes in Ascend NPU hardware scenarios:

CANN Errors

• ascend error: CANN initialization failures
• RuntimeError: CANN: NPU runtime errors
• NPU error: NPU device errors

MultimodalSDK Errors

• mm.: MultimodalSDK API errors
• AdapterError: Preprocessor adapter failures
• TensorError: Tensor handling errors

Memory Errors on NPU

• NPU out of memory: NPU memory exhaustion
• ACL error: Ascend ACL errors

Vision SDK Errors

• mxvision: Vision SDK errors
• Image decode error: Image processing failures

Response Generation

Always follow these templates when responding to users:

• Known Issue: Use Template 1 from response-templates.md
• Need Info: Use Template 2 - ask for missing details
• Version Issue: Use Template 3 - explain compatibility
• Config Error: Use Template 4 - provide correct settings
• Permission: Use Template 5 - explain required access
• NPU Specific: Use Ascend-specific solutions from sdk-knowledge.md

Debugging Checklist

For complex errors, follow the systematic approach in debugging-checklist.md:

1. Information Gathering - extract error, env, context
2. Initial Analysis - classify, check versions, analyze logs
3. Root Cause Determination - form and test hypotheses
4. Resolution - develop and verify solution
5. Communication - prepare clear response

Output Format

For each error analysis, always include:

## Issue Analysis

**Root Cause**: [Brief explanation]

**Solution**: [Step-by-step resolution]

**Prevention**: [Tips to avoid this issue]

**NPU Context**: [If applicable, Ascend-specific considerations]

Scripts

parse_error.py

Extracts structured information from error logs.

python scripts/parse_error.py < error.log

Outputs JSON with fields: error_type, error_message, category, environment, traceback, npu_specific.

analyze_error.py (optional advanced script)

For deeper analysis with repository context:

python scripts/analyze_error.py \
  --error-log error.log \
  --repo-path /path/to/repo \
  --output analysis.md

References

• error-patterns.md - Common error patterns (general + Ascend)
• sdk-knowledge.md - Ascend NPU/CANN specific knowledge
• response-templates.md - Response templates
• debugging-checklist.md - Debugging approach
• examples.md - Analysis examples
• solutions.md - Common solutions and workarounds

Example Usage

Example 1: NPU Memory Error

Input:

Error: RuntimeError: NPU out of memory. Tried to allocate 2.0 GB on device 0.
Environment: Ubuntu 22.04, CANN 8.5.0, Python 3.9
Context: Running MultimodalSDK preprocessing

Analysis:

1. Pattern match: NPU OOM error
2. Root cause: Insufficient NPU memory for batch
3. Solution: Reduce batch size, enable memory optimization

Example 2: CANN Import Error

Input:

Error: ImportError: cannot import name 'acl' from 'ascend'
Environment: CentOS 7.9, CANN 8.0.0
Context: Initializing Ascend NPU

Analysis:

1. Pattern match: CANN not properly installed
2. Root cause: CANN environment variables not set
3. Solution: Source CANN set_env.sh

Example 3: Vision SDK Configuration Error

Input:

Error: KeyError: 'device_id'
Environment: Python 3.9, Vision SDK 3.0
Context: Loading pipeline configuration

Analysis:

1. Pattern match: Configuration key missing
2. Root cause: Missing required configuration parameter
3. Solution: Add device_id to config

Best Practices

1. Always validate input completeness - Ask for missing environment info if needed
2. Be specific in solutions - Provide exact commands, file paths, line numbers
3. Explain the why - Don't just give fixes, explain why they work
4. Acknowledge uncertainty - If the cause is unclear, say so and suggest diagnostic steps
5. Keep responses actionable - Every suggestion should have a clear next step
6. Consider NPU specifics - For Ascend errors, always check CANN version and NPU status

Limitations

• This skill analyzes based on provided information and known patterns
• Complex issues may require additional debugging
• Some errors may need developer investigation
• Always recommend creating an issue for persistent problems

Integration Notes

This skill is designed to work independently and can be used:

• In CI/CD pipelines for automated error triage
• In support workflows for first-line response
• In development workflows for self-service debugging
• As a standalone tool for error analysis
• With configurable repository paths for source code reference