Solar: PyTorch Model Analysis Toolkit

Solar is a toolkit for analyzing PyTorch model graphs, converting them to einsum representations, and performing hardware-aware SOL performance predictions.

Features

• 5-Stage Analysis Pipeline: Seamless conversion from PyTorch models to performance predictions
• Graph Extraction: Extract structured computation graphs from PyTorch models (torchview-based)
• Einsum Conversion: Convert PyTorch operations to einsum notation with automatic rank renaming
• Graph Visualization: Generate PDF visualizations of einsum graphs
• Hardware-Independent Analysis: Compute MACs, FLOPs, and memory footprints
• Performance Prediction: Architecture-aware roofline modeling (H100, A6000, etc.)
• Timeloop / Orojenesis Export: Convert to Timeloop workload format for architectural exploration
• Benchmark Support: Native support for kernelbench benchmark suites
• Human-Readable YAML: All outputs use clean YAML without anchors/aliases

Installation

# Install Solar in development mode
cd solar
pip install -e .

Dependencies:

# Core dependencies are in requirements.txt
pip install -r requirements.txt

# For graph visualization (optional)
pip install graphviz matplotlib

The 5-Stage Pipeline

Solar processes models through five distinct stages:

Stage 1: PyTorch Graph Extraction
  └─> pytorch_graph.yaml

Stage 2: Einsum Conversion + Rank Renaming
  └─> einsum_graph.yaml
  └─> einsum_graph_renamed.yaml
  └─> einsum_graph.pdf (optional)

Stage 3: Hardware-Independent Analysis
  └─> analysis.yaml

Stage 4: Performance Prediction
  └─> perf_<arch>.yaml

Stage 5: Timeloop Export (optional)
  └─> timeloop_graph.yaml

Examples

Solar includes several example models demonstrating different attention patterns:

Available Examples

Example	Description	Based On
Attention/	Multi-head self-attention	Standard Transformer
BERT/	BERT-like encoder model	BERT architecture

Running an Example

# Run the complete pipeline for any example
cd solar/examples/Attention
bash run_solar.sh

# Outputs:
#   - output/graph/pytorch_graph.yaml           (Stage 1)
#   - output/einsum/einsum_graph.yaml           (Stage 2)
#   - output/einsum/einsum_graph_renamed.yaml   (Stage 2 - with BFS rank renaming)
#   - output/einsum/einsum_graph.pdf            (Stage 2 - visualization)
#   - output/analysis/analysis.yaml             (Stage 3)
#   - output/perf/perf_H100_PCIe.yaml           (Stage 4)
#   - output/timeloop/timeloop_graph.yaml       (Stage 5)

Benchmark Suite (Kernelbench)

Process benchmark models:

# Process and analyze kernelbench models
solar-toeinsum --level level1 --kernel-ids 1 2 3

# Use different architecture
solar-toeinsum --level level1 --kernel-ids 1 --arch-config B200

CLI Commands

Single Model Processing

# Stage 1: Extract PyTorch graph
solar-process-model --model-file model.py --output-dir output/graph

# Stage 2: Convert to einsum (with optional PDF visualization)
solar-toeinsum-model --graph-path output/graph/pytorch_graph.yaml \
                     --output-dir output/einsum --no-copy-graph \
                     --save-graph

# Stage 3: Analyze (hardware-independent)
solar-analyze-model --einsum-graph-path output/einsum/einsum_graph_renamed.yaml \
                    --output-dir output/analysis

# Stage 4: Predict performance
solar-predict-perf-model --analysis-path output/analysis/analysis.yaml \
                         --output-dir output/perf --arch-config H100_PCIe

# Stage 5: Convert to Timeloop format
solar-totimeloop --einsum-graph-path output/einsum/einsum_graph_renamed.yaml \
                 --output-dir output/timeloop

Output File Formats

All output files use human-readable YAML without anchors/aliases:

• pytorch_graph.yaml: Structured graph with layers, shapes, weights, connections
• einsum_graph.yaml: Einsum equations + shapes for each layer
• einsum_graph_renamed.yaml: Einsum graph with consistent dimension labels (BFS-based)
• einsum_graph.pdf: Visual representation of the computation graph
• analysis.yaml: Hardware-independent metrics (MACs, FLOPs, bytes)
• perf_.yaml: Architecture-specific performance predictions
• timeloop_graph.yaml: Timeloop workload format for architectural exploration

Testing

# Run all tests
bash run_tests.sh

# Quick smoke tests
bash run_tests.sh quick

# Run specific test categories
bash run_tests.sh graph      # Graph processing tests
bash run_tests.sh einsum     # Einsum analyzer tests
bash run_tests.sh unit       # All unit tests
bash run_tests.sh integration # Integration tests

# Test examples
bash run_tests.sh examples   # Run all example scripts

# Test benchmark compatibility
bash run_tests.sh kernelbench

# Verbose output
bash run_tests.sh all -v

See TESTING_GUIDE.md for detailed testing documentation.

Python API

Graph Processing

from solar.graph import PyTorchProcessor

processor = PyTorchProcessor()
processor.process_model_file("model.py", output_dir="outputs/my_model")

Einsum Conversion

from solar.einsum import PyTorchToEinsum

converter = PyTorchToEinsum()
einsum_graph = converter.convert(
    "outputs/my_model/pytorch_graph.yaml",
    "outputs/my_model",
    copy_graph=False  # Don't duplicate input graph
)
# Produces both einsum_graph.yaml and einsum_graph_renamed.yaml

Graph Visualization

from solar.einsum import EinsumGraphVisualizer

visualizer = EinsumGraphVisualizer()
visualizer.save_graph_pdf(
    "outputs/my_model/einsum_graph_renamed.yaml",
    "outputs/my_model/einsum_graph.pdf"
)

Analysis

from solar.analysis import EinsumGraphAnalyzer

analyzer = EinsumGraphAnalyzer()
analysis = analyzer.analyze_graph(
    "outputs/my_model/einsum_graph_renamed.yaml",
    "outputs/my_model"
)

Performance Prediction

from solar.perf import EinsumGraphPerfModel

perf_model = EinsumGraphPerfModel()
perf = perf_model.predict(
    "outputs/my_model/analysis.yaml",
    "outputs/my_model",
    arch_config="H100_PCIe"
)

Timeloop Export

from solar.einsum import EinsumToTimeloop

converter = EinsumToTimeloop()
result = converter.convert(
    "outputs/my_model/einsum_graph_renamed.yaml",
    "outputs/my_model/timeloop_graph.yaml"
)

Architecture

solar/
├── solar/
│   ├── common/        # Shared types, constants, utilities (NoAliasDumper)
│   ├── graph/         # Stage 1: PyTorch graph extraction
│   ├── einsum/        # Stage 2: Einsum conversion, visualization, Timeloop export
│   ├── analysis/      # Stage 3: Hardware-independent analysis
│   ├── perf/          # Stage 4: Performance prediction
│   └── cli/           # Command-line interfaces
├── tests/             # Comprehensive test suite
├── examples/          # Example models (Attention, BERT, sparse attention variants)
└── configs/           # Architecture configs (H100_PCIe.yaml, A6000.yaml)

Key Components

• NoAliasDumper: Custom YAML dumper for human-readable output (no &id001 references)
• EinsumRankRenamer: BFS-based dimension label renaming for consistent einsum equations
• EinsumGraphVisualizer: PDF visualization of computation graphs
• EinsumToTimeloop: Export to Timeloop workload format
• Node Registry: Extensible registry for operation handlers
• LLM Agent (optional): Dynamic handler generation for unknown operations
• Benchmark Processors: Specialized handling for kernelbench file structures

Active Contributors

• hqjennynv
• sdamani-nvidia
• askiad
• LemonAndRabbit

Contributing

• Follow Google's Python Style Guide
• Add tests for new features
• Update documentation for API changes
• Run bash run_tests.sh before submitting PRs

Documentation

• TESTING_GUIDE.md: Comprehensive testing documentation
• REFACTORING_SUMMARY.md: Design decisions and refactoring history
• MIGRATION_COMPLETE.md: Migration guide from legacy JSON format

License

MIT License