Skip to content

cli.md

Latest commit

 

History

History
71 lines (55 loc) · 1.56 KB

cli.md

File metadata and controls

71 lines (55 loc) · 1.56 KB

CLI Reference

Use src.optimizer.workflow for headless or scripted runs.

Commands

1. Profile

python -m src.optimizer.workflow profile --project <project_name>

Captures operator I/O tensors and baseline benchmark data.

2. Generate kernels

# All operators
python -m src.optimizer.workflow generate \
  --project <project_name> \
  --target-device cuda

# Selected operators only
python -m src.optimizer.workflow generate \
  --project <project_name> \
  --ops torch_nn_functional_conv2d,torch_nn_functional_linear \
  --target-device cuda

3. Optimize + benchmark (end-to-end)

python -m src.optimizer.workflow generate \
  --project <project_name> \
  --target-device cuda \
  --optimize \
  --benchmark \
  --iterations 5

4. Optimize existing kernels

python -m src.optimizer.workflow optimize \
  --project <project_name> \
  --ops torch_nn_functional_conv2d \
  --target-device cuda \
  --iterations 5 \
  --benchmark

5. Benchmark only

python -m src.optimizer.workflow benchmark --project <project_name>

Project artifact layout

kernels/projects/<project_name>/
├── state.json                          # job state (progress, pause/cancel)
├── io/
│   ├── summary.json
│   └── individual_ops/                 # captured tensor I/O per operator
├── kernels/
│   └── generated/individual_op_kernels/<op>/
├── trees/<op>/                         # MCTS nodes and kernel source per attempt
├── benchmarks/op_benchmarks.json
└── logs/