AgentSkillsCN

ml-optimization

优化 TensorFlow.js 的 WebGPU 后端，提升 WebLLM 的离线推理性能，并为重型机器学习模型实现内存安全机制。

SKILL.md
--- frontmatter
name: ml-optimization
description: Optimizes TensorFlow.js WebGPU backend, WebLLM offline inference, and implements memory safety for heavy ML models
license: MIT
compatibility: opencode
metadata:
  audience: developers
  workflow: development

What I do

I optimize the ML pipeline for performance and memory efficiency. I manage TF.js WebGPU backend configuration, WebLLM offline inference, and ensure proper memory cleanup to prevent leaks.

When to use me

Use this when:

  • •You're working with TF.js or WebLLM models
  • •You notice memory leaks or performance regressions
  • •You're updating model weights or configurations

Key Concepts

  • •TF.js WebGPU: GPU acceleration for browser ML
  • •WebLLM: Browser-based LLM for offline inference
  • •SmolLM2: Efficient LLM model variant
  • •Memory Safety: Proper tensor cleanup
  • •Manual Chunks: Code splitting for model loading

Source Files

  • •services/vision.ts: ML model integration
  • •vite.config.ts: Build configuration for chunks
  • •plans/02_edge_ml_implementation.md: ML optimization plan

Code Patterns

  • •Use tf.tidy() for automatic tensor cleanup
  • •Lazy load heavy models on user interaction
  • •Expose unload() method for component unmount
  • •Manual chunks in Vite for model code splitting

Operational Constraints

  • •All TF.js operations MUST use tf.tidy() or dispose()
  • •Heavy models must expose unload() method
  • •Clean up on component unmount
  • •Monitor memory usage in production