skills.rest
TorchedHat
TorchedHatOFFICIAL
Explore all skills by this author →
GitHub

vllm-compile

Diagnoses and explains JetBrains Odyssey Odyssey Odyssey Odyssey Odyssey Odyssey Odyssey Odyssey Odyssey Odyssey Odyssey issues.

6|8|Updated May 7, 2026
Search Tags:#pytorch#vllm#torch.compile#inductor#dynamo#cuda graphs#compiler debugging
One-click install
npx skills add https://github.com/TorchedHat/ai-marketplace --skill vllm-compile
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: vllm-compile
Source: https://github.com/TorchedHat/ai-marketplace/tree/main/vllm-compile/skills/vllm-compile
Command: npx skills add https://github.com/TorchedHat/ai-marketplace --skill vllm-compile

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

This skill helps PyTorch and vLLM developers understand, configure, debug, and optimize vLLM's custom torch.compile pipeline, including guard dropping, graph breaks, fusion failures, graph splitting, compilation cache issues, and CUDA graph capture problems.

Core Features & Use Cases

  • Compilation Pipeline Guidance: Explains the flow from the support_torch_compile decorator through Dynamo capture, Inductor optimization, piecewise graph execution, and CUDA graph replay.
  • Fusion and Graph Optimization: Provides implementation patterns for vLLM-specific Inductor passes, including RoPE and KV-cache fusion, collective fusion, activation quantization, and normalization optimizations.
  • Debugging and Configuration: Helps investigate graph breaks, missing split points, guard violations, compilation out-of-memory errors, cache invalidation, and CUDA graph failures using relevant configuration options, logs, and generated artifacts.
  • Use Case: When a vLLM model fails during compilation or shows unexpected performance, use this skill to identify the affected pipeline stage, inspect the relevant cache or logs, adjust compilation settings, and isolate the issue by disabling individual components.

Quick Start

Use the vLLM compilation skill to diagnose a torch.compile failure in my model and recommend the relevant logs, configuration changes, and next debugging steps.

Frequently Asked Questions about vllm-compile

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

FAQPage Schema
How do I debug torch.compile graph breaks and guard dropping in vLLM?▼

Debug vLLM torch.compile graph breaks and guard dropping by inspecting Dynamo graph capture logs, identifying guard violations, and adjusting vLLM configuration settings to isolate the failing compilation pipeline stage.

Why does vLLM CUDA graph capture fail during torch.compile compilation?▼

vLLM CUDA graph capture fails during torch.compile due to unsupported operations or guard violations within the compiled graph. Investigate the generated compilation artifacts and debug logs to disable individual components and isolate the capture failure.

How do I optimize Inductor fusion passes for vLLM inference workloads?▼

Optimize Inductor fusion passes for vLLM inference workloads by applying vLLM-specific implementation patterns for RoPE and KV-cache fusion, collective fusion, activation quantization, and normalization optimizations within the Inductor pipeline.

What is the best way to troubleshoot vLLM compilation cache invalidation and out-of-memory errors?▼

The best way to troubleshoot vLLM compilation cache invalidation and out-of-memory errors is to examine the generated compilation artifacts and debug logs, disable individual components sequentially, and modify the relevant vLLM configuration settings.

Do I need to understand PyTorch Dynamo and Inductor behavior to configure vLLM compilation?▼

Yes, configuring vLLM compilation requires understanding PyTorch Dynamo and Inductor behavior, as the pipeline flows from the support_torch_compile decorator through Dynamo capture, Inductor optimization, piecewise graph execution, and CUDA graph replay.

When should I disable piecewise attention graph splitting in vLLM torch.compile?▼

Disable piecewise attention graph splitting in vLLM torch.compile when you encounter missing split points, fusion failures, or unexpected performance drops, using debug logging to isolate issues by turning off individual compilation components.

Related Skills

Explore More Software Engineering Skills→
TorchedHatTorchedHat
6

compile-bisect

Bisect PyTorch compiler stages to isolate failing backend, subsystem, and operation.

Official
Intermediate
TorchedHatTorchedHat
6

compile-overview

Explain torch.compile pipeline stages and intermediate representations for debugging.

Official
Intermediate
TorchedHatTorchedHat
6

compile-trace-dynamo

Diagnose PyTorch Dynamo compilation failures via FX graph inspection.

Official
Intermediate
SnakeOSnakeO
9

refactor:pytorch

Refactor PyTorch code with modular nn.Module, torch.compile, AMP, and optimized DataLoader.

Community
Advanced