skills.rest
tenstorrent
tenstorrentOFFICIAL
Explore all skills by this author →
WebsiteGitHub

graph-break-analysis

Diagnose graph breaks in TT-XLA compilation logs and generate fix reports.

74|32|Updated Sep 13, 2024
Search Tags:#debugging#jax#compilation#pytorch#mlir#tt-xla#graph-break
One-click install
npx skills add https://github.com/tenstorrent/tt-xla --skill graph-break-analysis
Or copy as Structured Prompt for Agent▼
Please help me install this Agent Skill.
Skill: graph-break-analysis
Source: https://github.com/tenstorrent/tt-xla/tree/main/.claude/skills/graph-break-analysis
Command: npx skills add https://github.com/tenstorrent/tt-xla --skill graph-break-analysis

SYSTEM DOCUMENTATION & REQUIREMENTS

What problem does it solve?

Graph breaks occur when a model, pipeline, or script is split into more graphs than necessary during TT-XLA compilation, making logs noisy and debugging harder. In TT-XLA, graph breaks often arise from tracing stages such as torchdynamo tracing, torch_xla tracing, or rare byproducts of torch.export, and are not simply different MLIR modules. This guide helps you distinguish true graph breaks from expected graph variants and provides a structured approach to diagnosing and fixing the root causes.

Core Features & Use Cases

  • Count and categorize MLIR module blocks per graph to quantify graph breaks.
  • Map each graph to its source code path (torchdynamo, torch_xla, export) to identify responsible stage.
  • Produce a prioritized action list with reproducible steps and optional patch scripts to fix the break.

Quick Start

Input the TT-XLA debug log and I will generate a graph-break analysis with root causes and fixes.

Frequently Asked Questions about graph-break-analysis

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

FAQPage Schema
What causes graph breaks during TT-XLA compilation in PyTorch or JAX workflows?▼

Graph breaks during TT-XLA compilation happen when torchdynamo, torch_xla, or torch.export tracing splits a model into more MLIR modules than necessary, creating noisy logs. They represent actual tracing interruptions rather than just distinct MLIR graph variants.

How do I identify and diagnose graph breaks in MLIR module logs?▼

To identify graph breaks, count and categorize MLIR module blocks per graph, then map each graph to its source code path like torchdynamo or torch_xla. This pinpoints the responsible tracing stage causing the fragmentation.

How can I distinguish true graph breaks from expected MLIR graph variants?▼

Distinguish true graph breaks from expected MLIR graph variants by analyzing TT-XLA debug logs to quantify graph counts and categorize module blocks. Expected variants are normal architectural separations, whereas true breaks stem from tracing interruptions.

Can I generate reproduction scripts to validate TT-XLA graph break fixes?▼

Yes, graph break analysis generates reproduction scripts alongside prioritized action lists and optional patch scripts. These outputs allow you to execute and validate proposed fixes for TT-XLA compilation tracing interruptions directly.

What is the best way to fix torchdynamo or torch_xla tracing interruptions?▼

The best way to fix tracing interruptions is to input the TT-XLA debug log for analysis, which produces a structured report of root causes and a prioritized action list with executable patch scripts to resolve the breaks.

Related Skills

Explore More Software Engineering Skills→
tenstorrenttenstorrent
295

validate-tt-mlir-against-tt-xla

Automate cross-repo validation of tt-mlir pull requests against tt-xla CI.

Official
Advanced
tenstorrenttenstorrent
74

code-reviewer

Automate structured code-review checks for the tt-xla project.

Official
Advanced
TorchedHatTorchedHat
6

vllm-compile

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

Official
Advanced
tenstorrenttenstorrent
341

tt-bug-report

Creates GitHub issues for bug reports with minimized reproducers across TT repos.

Official
Intermediate