Solar Roofline

Overview

AKG Agents' roofline support is provided through an installed solar Python package.

Runtime constraints:

• AKG does not depend on a local SOLAR source tree path
• AKG does not require modifying or patching the SOLAR repository
• AKG maintains the missing bridge pieces internally:
  • SOLBench workload wrapper generation
  • roofline architecture configs for Ascend / A100 / V100

In other words, AKG now only requires:

python -c "import solar"

to succeed.

Installation

Recommended:

bash download.sh --with_solar

or:

SOLAR_DIR=/path/to/SOLAR bash download.sh --with_solar
SOLAR_REF=<tag|branch|commit> bash download.sh --with_solar

The script will:

1. Clone the official Solar repo into thirdparty/SOLAR by default
2. Optionally override the directory with SOLAR_DIR and check out a specific SOLAR_REF
3. Prefer Solar's own install.sh
4. Install Solar dependencies, including patched torchview
5. Install Solar into the current Python environment in editable mode
6. Verify import solar and the core APIs used by AKG

Runtime behavior

When roofline is enabled during profiling, AKG calls Solar's Python APIs directly:

• solar.graph.PyTorchProcessor
• solar.einsum.PyTorchToEinsum
• solar.analysis.EinsumGraphAnalyzer
• solar.perf.EinsumGraphPerfModel

Outputs include:

• roofline_profile_result.json
• profile return fields:
  • roofline_time
  • roofline_speedup
  • roofline

Where:

roofline_speedup = roofline_time / gen_time

1.0x means the generated kernel is on the roofline bound; values below 1.0x mean it is still below the theoretical roofline.

Graceful degradation

If Solar is not installed, or Solar analysis fails:

• the main correctness / profiling flow still continues
• roofline degrades to missing data only
• AKG profiling does not fail solely because roofline is unavailable