hf-model-inference（高频模型推理）

HuggingFace Model Inference Service

Overview

This skill provides procedural guidance for setting up HuggingFace model inference services. It covers model downloading, caching strategies, Flask API creation, and service deployment patterns.

Workflow

Phase 1: Environment Setup

1. Verify package manager availability

- Check for uv, pip, or conda before installing dependencies - Prefer uv for faster dependency resolution when available

1. Install required packages

- Core: transformers, torch (or tensorflow) - API: flask for REST endpoints - Set appropriate timeouts for large package installations (300+ seconds)

1. Create model cache directory

- Establish a dedicated directory for model storage (e.g., /app/model_cache/model_name) - Create parent directories as needed before downloading

Phase 2: Model Download

1. Download the model separately from API startup

- Use a dedicated download script or inline download before starting the service - This prevents timeout issues during API initialization

1. Specify cache directory explicitly from transformers import pipeline model = pipeline("task-type", model="model-name", cache_dir="/path/to/cache")
2. Verification step (commonly missed)

- After download, verify model files exist in the target directory - List directory contents to confirm successful download

Phase 3: API Creation

1. Flask application structure from flask import Flask, request, jsonify from transformers import pipeline app = Flask(__name__) model = None # Load at startup @app.route('/predict', methods=['POST']) def predict(): # Handle inference pass
2. Input validation requirements

- Check for required fields in request JSON - Validate field types (string, number, etc.) - Handle empty or whitespace-only inputs - Return descriptive error messages with appropriate HTTP status codes

1. Error response format

- Use consistent JSON structure: {"error": "message"} - Return 400 for client errors, 500 for server errors

Phase 4: Service Deployment

1. Host and port configuration

- Bind to 0.0.0.0 for external accessibility - Use specified port (commonly 5000) - Example: app.run(host='0.0.0.0', port=5000)

1. Background execution

- Start Flask in background mode for testing - Allow startup time (2-3 seconds) before sending test requests

Verification Strategies

Model Download Verification

• List cache directory contents after download
• Confirm expected model files exist (config.json, model weights, tokenizer files)

API Functionality Testing

Test these scenarios in order:

1. Positive case: Valid input that should succeed curl -X POST http://localhost:5000/predict \ -H "Content-Type: application/json" \ -d '{"text": "valid input text"}'
2. Negative case: Different valid input to verify varied responses curl -X POST http://localhost:5000/predict \ -H "Content-Type: application/json" \ -d '{"text": "different input text"}'
3. Error case: Missing required field curl -X POST http://localhost:5000/predict \ -H "Content-Type: application/json" \ -d '{}'

Extended Edge Cases (Optional)

• Empty string input
• Very long text input
• Non-JSON content type
• Malformed JSON
• Wrong field type (number instead of string)

Common Pitfalls

Installation Issues

• Insufficient timeout: Large packages like torch require extended timeouts (5+ minutes)
• Missing system dependencies: Some models require additional system packages

Model Loading Issues

• Cold start timeout: Loading models at first request causes timeouts; load at startup instead
• Memory constraints: Large models may exceed available RAM; check model requirements

API Issues

• Development server warning: Flask development server is not suitable for production; acceptable for testing but note the limitation
• No graceful shutdown: Consider signal handling for clean termination
• No health check endpoint: Adding /health endpoint aids debugging

Process Management

• Background process verification: After starting in background, verify the process is running
• Port conflicts: Check if the specified port is already in use before starting

Task Planning Template

When approaching HuggingFace inference tasks, structure work as follows:

1. Environment verification (package manager, system requirements)
2. Dependency installation with appropriate timeouts
3. Cache directory creation
4. Model download with explicit cache path
5. Model download verification
6. API script creation with validation
7. Service startup in background
8. Functional testing (positive, negative, error cases)
9. Edge case testing (if time permits)