Skip to content

README.md

Latest commit

 

History

History
62 lines (44 loc) · 1.88 KB

README.md

File metadata and controls

62 lines (44 loc) · 1.88 KB

Kernel Forge Source (src)

This directory contains the backend generation, optimization, and benchmarking logic.

Structure

src/
├── generator/                  # Kernel generation + correctness validation
│   ├── main.py                 # Generator entrypoint (operator-level)
│   ├── generator.py            # LLM interaction and code extraction
│   └── prompts/                # Prompt templates/system prompts
├── optimizer/
│   ├── workflow.py             # Canonical orchestration CLI (profile/generate/optimize/benchmark)
│   ├── pipeline.py             # Optimization pipeline (MCTS-driven)
│   ├── benchmarking/           # Baseline + optimized benchmarking pipeline
│   ├── core/                   # MCTS core types/logic
│   └── backends/               # CUDA and Triton backend abstractions (Metal is a skeleton)
└── progress.py                 # Job progress + pause/cancel helpers via state.json

Canonical entrypoints

Orchestration (preferred)

python -m src.optimizer.workflow <action> [flags]

Actions:

  • profile
  • generate
  • optimize
  • benchmark

Generator direct

src.generator.main is called by workflow for generation.

Optimizer direct

src.optimizer.pipeline is called by workflow for optimization.

Current execution model

  • profile: captures operator data and baseline benchmark artifacts.
  • generate: runs per-operator generation, verifies success markers, then optionally optimize+benchmark.
  • optimize: runs optimization by operator, optionally benchmark.
  • benchmark: rebuilds op_benchmarks.json from current project outputs.

Project artifact location

kernels/projects/<project_name>/...

Key files:

  • state.json
  • io/summary.json
  • kernels/generated/individual_op_kernels/<op>/
  • trees/<op>/
  • benchmarks/op_benchmarks.json