skills.rest
sovr610
sovr610COMMUNITY
Explore all skills by this author →
WebsiteGitHub

Compiler & Kernel Fusion (torch.compile) Integration

Wrap models with a safe torch.compile harness for training and inference.

Updated Feb 28, 2026
Search Tags:#fsdp#ddp#torch.compile#inductor#torchdynamo#shape-stabilization
One-click install
npx skills add https://github.com/sovr610/refffiy --skill compiler-kernel-fusion-torch-compile-integration
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: Compiler & Kernel Fusion (torch.compile) Integration
Source: https://github.com/sovr610/refffiy/tree/main/brain-ai-dev/skills/torch-compile-integration
Command: npx skills add https://github.com/sovr610/refffiy --skill compiler-kernel-fusion-torch-compile-integration

SYSTEM DOCUMENTATION & REQUIREMENTS

💡 This Skill requires torch, yaml, and includes scripts (resource) and references (resource) and assets (resource) components.

What problem does it solve?

Torch-based compilation accelerates training/inference by reducing Python overhead and fusion improvements through TorchDynamo/AOT Autograd/TorchInductor, while maintaining safe fallbacks.

Core Features & Use Cases

  • Safe wrapper for safe compilation: maybe_compile ensures non-breaking fallback on failure.
  • Selective compilation via allowlist/blocklist and full-model compilation with various backends/modes.
  • Shape management and bucketing to reduce recompilations, including dynamic shape hints and bucketing strategies for variable-length inputs.
  • DDP/FSDP friendly workflows and smoketest health checks to verify correctness.

Quick Start

Wrap your model with maybe_compile(cfg, logger, sample_batch) to obtain a compiled model ready for training.

Frequently Asked Questions about Compiler & Kernel Fusion (torch.compile) Integration

High-intent search queries and answers about installing and using this skill.

FAQPage Schema
How do I speed up PyTorch training with torch.compile without breaking my model?▼

Use a safe compile wrapper like maybe_compile to wrap your model with TorchDynamo and TorchInductor, ensuring non-breaking fallbacks on compilation failure while reducing Python overhead.

Does torch.compile work with DDP and FSDP multi-GPU training?▼

Yes, torch.compile integrates with DDP and FSDP multi-GPU workflows through a safe compile harness, providing smoketest health checks to verify correctness across distributed training setups.

How do I stop torch.compile from recompiling when input shapes change?▼

Manage dynamic shapes by applying shape bucketing and dynamic shape hints through a CompileConfig, which groups varying input shapes to reduce recompilations during training.

Can I selectively compile parts of my model instead of the full graph?▼

Yes, you can selectively apply compilation by configuring an allowlist or blocklist within your CompileConfig, enabling modular selection of model components for TorchInductor optimization.

What do I need to set up before applying a torch.compile wrapper to my model?▼

You need PyTorch installed and must define a CompileConfig, optionally providing a sample batch and logger to run smoketest health checks before obtaining the compiled model.

Related Skills

Explore More Software Engineering Skills→
jinzhezenggroupjinzhezenggroup
124

deepmd-train-se-e2-a

Train DeePMD-kit models with SE_E2_A descriptor and PyTorch backend.

Official
Advanced
GonglitianGonglitian
2

gpu-train-monitor

Monitors GPU health and summarizes training progress from live system data.

Community
Advanced
Tianyi-Billy-MaTianyi-Billy-Ma

pytorch-lightning

Automate PyTorch training workflows with LightningModule, Trainer, and LightningDataModule.

Community
Advanced
aayushsoamaayushsoam
1

simpo-training

Train preference-aligned language models without a reference model.

Community
Intermediate