Optimize PyTorch training and inference with AMP, torch.compile, and safe checkpointing.
npx skills add https://github.com/ulf1/trading-regime --skill 323-pytorch
Please help me install this Agent Skill. Skill: 323-pytorch Source: https://github.com/ulf1/trading-regime/tree/main/.agent/skills/323-pytorch Command: npx skills add https://github.com/ulf1/trading-regime --skill 323-pytorch
This Skill prevents slow, brittle, and error-prone PyTorch training/inference implementations by enforcing production-grade patterns for device handling, mixed precision, data loading, and safe checkpointing.
Use case example: You’re training a neural model for daily retraining and hit GPU OOMs and inconsistent performance—use this Skill’s checklist to fix device transfers, enable AMP correctly, stabilize the training loop, and validate that checkpoints load cleanly.
Use the 323-pytorch skill to optimize and debug your PyTorch training code for correctness, AMP compatibility, and performance while ensuring safe checkpointing.
High-intent search queries and answers about installing and using this skill.
Optimize PyTorch DataLoader by setting pin_memory=True for CUDA transfers, adjusting num_workers for parallel data loading, and enabling persistent_workers to avoid reinitializing worker processes. Profile data loading to identify pipeline bottlenecks before modifying parameters.
PyTorch checkpoint loading fails across devices when map_location is not specified in torch.load. Safe state_dict checkpointing requires setting map_location to correctly remap storage tensors to the target device, whether CPU or CUDA.
Apply torch.compile to your PyTorch model before the training loop to optimize execution graphs and reduce kernel overhead. Ensure your forward function remains pure with no hidden state side effects for the compilation to succeed.
Using zero_grad(set_to_none=True) in PyTorch improves memory performance by setting gradients to None instead of zeroing them, allowing more efficient memory allocation during backpropagation. This avoids unnecessary memory writes and reduces optimizer overhead.
Prevent GPU OOM errors during PyTorch retraining by minimizing host-device transfers, enabling automatic mixed precision, detaching hidden states to avoid graph growth, and ensuring explicit device placement for tensors and models before execution.
rishikanthcDetect wiring and integration bugs in PyTorch model implementations before training runs.
Vegetable-bird10086Convert PyTorch models to portable .pte files for ExecuTorch deployment.
michaeldubuParse and analyze CUDA, CPU, and memory benchmark outputs across CSV, JSON, and text formats.
zeronx798Implement distributed training strategies for PyTorch models across GPUs and nodes.
