Torch compile¶
Compile model-owned execution methods while preserving checkpoint keys and reversible eager fallbacks.
Use¶
print(model.available_optimization_passes())
result = model.apply_optimization_plan("compile", mode="inference")
print(result.manifest())
model.restore_optimization_plan(mode="inference")
Support¶
| Property | Value |
|---|---|
| Availability | Registered public pass: compile |
| Fidelity | Exact intent; verify numerical and audio equivalence for the concrete graph |
| Runtime | CPU or CUDA; float32, float16, or bfloat16 |
| Registry name | compile |
| Pass ID | torch.compile |
| Pass version | 1 |
| Restore | model.restore_optimization_plan(mode="inference") |
Unsupported explicit configurations must fail before mutation. A pass that
does not match a model reports not-applicable; it is not an acceleration.
Paper and GitHub¶
- Paper: PyTorch 2: Faster Machine Learning Through Dynamic Python Bytecode Transformation and Graph Compilation
- Upstream GitHub: PyTorch
- VoiceHub source: VoiceHub implementation
Verify¶
Compare the eager and optimized paths with the same checkpoint, input, seed, warm-up, device, and dtype. Record latency, memory, output quality, the exact source revision, and the optimization manifest.
See the related workflow and optimization API.