Diffusion sampling¶
Reduce model evaluations through explicit schedule, guidance, cache, or solver policies.
Use¶
result = model.apply_optimization_plan(
"diffusion-sampling",
mode="inference",
)
print(result.manifest())
Support¶
| Property | Value |
|---|---|
| Availability | Registered public pass: diffusion-sampling |
| Fidelity | Approximate; generated audio may change |
| Runtime | CPU, CUDA, or MPS; inference only |
| Registry name | diffusion-sampling |
| Pass ID | voicehub.diffusion-sampling |
| 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: DPM-Solver: A Fast ODE Solver for Diffusion Probabilistic Model Sampling; Flow Matching for Generative Modeling
- Upstream GitHub: DPM-Solver
- 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.