torch.compile¶
torch.compile provides compiler-driven graph fusion and specialization for stable operator sequences. In Sol-Engine, it is used where compiler fusion is more practical than maintaining custom kernels for every shape.
Sol-Engine placement¶
| Pipeline | Scope | Role |
|---|---|---|
| SANA-Video | compiled denoising path | part of the 2.77x full optimization stack |
| LTX-2.3 | selected graph regions | candidate when stable under target shapes |
Practical notes¶
- cold compile time should be excluded from warmup-excluded latency numbers.
- compile mode and cache directory can change observed performance.
- compiler fusion should be disabled for regions that deadlock or regress.
Validation¶
Run at least one warmup before quoting latency and compare generated video outputs under the same prompt and seed.