CLAUDE.md

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Project Overview

This is TCID (Transformer CI Dashboard) - a Gradio-based web dashboard that displays test results for Transformer models across AMD and NVIDIA hardware. The application fetches CI test data from HuggingFace datasets and presents it through interactive visualizations and detailed failure reports.

Architecture

Core Components

• app.py - Main Gradio application with UI components, plotting functions, and data visualization logic
• data.py - Data fetching module that retrieves test results from HuggingFace datasets for AMD and NVIDIA CI runs
• styles.css - Complete dark theme styling for the Gradio interface
• requirements.txt - Python dependencies (matplotlib only)

Data Flow

1. Data Loading: get_data() in data.py fetches latest CI results from:

   • AMD: hf://datasets/optimum-amd/transformers_daily_ci
   • NVIDIA: hf://datasets/hf-internal-testing/transformers_daily_ci

2. Data Processing: Results are joined and filtered to show only important models defined in IMPORTANT_MODELS list

3. Visualization: Two main views:

   • Summary Page: Horizontal bar charts showing test results for all models
   • Detail View: Pie charts for individual models with failure details

UI Architecture

• Sidebar: Model selection, refresh controls, CI job links
• Main Content: Dynamic display switching between summary and detail views
• Auto-refresh: Data reloads every 15 minutes via background threading

Running the Application

Development Commands

# Install dependencies
pip install -r requirements.txt

# Run the application
python app.py

HuggingFace Spaces Deployment

This application is configured for HuggingFace Spaces deployment:

• Framework: Gradio 5.38.0
• App file: app.py
• Configuration: See README.md header for Spaces metadata

Key Data Structures

Model Results DataFrame

The joined DataFrame contains these columns:

• success_amd / success_nvidia - Number of passing tests
• failed_multi_no_amd / failed_multi_no_nvidia - Multi-GPU failure counts
• failed_single_no_amd / failed_single_no_nvidia - Single-GPU failure counts
• failures_amd / failures_nvidia - Detailed failure information objects
• job_link_amd / job_link_nvidia - CI job URLs

Important Models List

Predefined list in data.py focusing on significant models:

• Classic models: bert, gpt2, t5, vit, clip, whisper
• Modern models: llama, gemma3, qwen2, mistral3
• Multimodal: qwen2_5_vl, llava, smolvlm, internvl

Styling and Theming

The application uses a comprehensive dark theme with:

• Fixed sidebar layout (300px width)
• Black background throughout (#000000)
• Custom scrollbars with dark styling
• Monospace fonts for technical aesthetics
• Gradient buttons and hover effects

Error Handling

• Data Loading Failures: Falls back to predefined model list for testing
• Missing Model Data: Shows "No data available" message in visualizations
• Empty Results: Gracefully handles cases with no test results

Performance Considerations

• Memory Management: Matplotlib configured to prevent memory warnings
• Interactive Mode: Disabled to prevent figure accumulation
• Auto-reload: Background threading with daemon timers
• Data Caching: Global variables store loaded data between UI updates