Agent Skills
› aiming-lab/AutoResearchClaw
› mixed-precision
mixed-precision
GitHub指导在PyTorch中使用torch.cuda.amp进行FP16/BF16混合精度训练，以加速模型训练并降低显存占用。包含自动缩放、梯度缩放及NaN处理等最佳实践，适用于GPU性能优化场景。
Trigger Scenarios
需要优化GPU训练速度
显存不足需要降低内存占用
配置AMP混合精度训练
Install
npx skills add aiming-lab/AutoResearchClaw --skill mixed-precision -g -y
SKILL.md
Frontmatter
{
"name": "mixed-precision",
"metadata": {
"author": "researchclaw",
"version": "1.0",
"category": "tooling",
"priority": "5",
"references": "Micikevicius et al., Mixed Precision Training, ICLR 2018",
"code-template": "scaler = torch.cuda.amp.GradScaler()\nfor batch in dataloader:\n optimizer.zero_grad()\n with torch.cuda.amp.autocast():\n output = model(batch)\n loss = criterion(output, target)\n scaler.scale(loss).backward()\n scaler.step(optimizer)\n scaler.update()",
"trigger-keywords": "training,gpu,memory,speed,precision,fp16,bf16",
"applicable-stages": "10,12"
},
"description": "Use FP16\/BF16 mixed precision to accelerate training and reduce memory. Use when optimizing GPU performance."
}
Mixed Precision Training Best Practice
Use torch.cuda.amp for automatic mixed precision:
- Wrap forward pass in torch.cuda.amp.autocast()
- Use GradScaler for loss scaling
- BF16 preferred over FP16 on Ampere+ GPUs (RTX 3xxx, A100, RTX 4xxx)
- Watch for NaN gradients — reduce learning rate if needed
- Do NOT use amp with custom CUDA kernels unless tested
Version History
- e2e23c9 Current 2026-07-25 07:48
