pytorch-knowledge-patch

What problem does it solve?

This Skill helps developers and code-generating agents align PyTorch code with breaking changes, new defaults, and runtime improvements introduced between PyTorch 2.6 and 2.11, preventing runtime failures, unsafe model loads, and degraded performance.

Core Features & Use Cases

• Migration guidance for the new weights_only default in model loading and safe alternatives for loading full nn.Module objects.
• Distributed training and sharding updates including the FSDP2 fully_shard workflow, context parallelism, symmetric memory, and differentiable collectives.
• Compilation and export patterns covering torch.compile stances, mega cache artifacts, hierarchical compilation, control flow operators, and torch.export with Dim.AUTO.
• Attention and performance notes for varlen_attn, FlexAttention with FA4, and platform compatibility including CUDA and Python versions.
• Use case: audit a repository to find unsafe torch.load usages, recommend state_dict migrations, and advise on compile and distributed training best practices.

Quick Start

Ask the skill to scan your PyTorch codebase and summarize necessary API and compatibility changes to upgrade to PyTorch 2.11.