Перейти к содержанию

Optimization overview

VoiceHub exposes one optimization lifecycle across TTS, ASR, and VAD. A pass can target speed, memory, or both, but compatibility and output behavior depend on the loaded model, execution mode, hardware, dtype, and input shape.

Use the same explicit lifecycle for every registered model:

def run_optimization(model, pass_name):
    print(model.available_optimization_passes())
    result = model.apply_optimization_plan(pass_name, mode="inference")
    print(model.optimization_manifest(mode="inference"))
    model.restore_optimization_plan(mode="inference")
    return result

The list reports global discovery, not compatibility with one runtime; validation happens before mutation, a failure rolls back earlier reversible passes, and the manifest records what was actually applied.

Technique Public pass Evidence boundary
Compilation compile The concrete graph, mode, device, dtype, and fixed-seed evidence decide support.
Attention backends flash-attention-4 Requires a compatible attention surface and optional CUDA backend.
General kernels custom-kernels Uses registered kernels only after runtime validation.
Codec kernels codec-kernels Applies only to discovered codec operations and records each selected backend.
Diffusion caching diffusion-cache Approximate reuse may change generated audio.
Diffusion sampling diffusion-sampling Step, guidance, or solver changes may change generated audio.

Compilation

Compilation can reduce Python overhead and fuse operations after an initial warm-up. VoiceHub checks the model-owned compile targets and preserves the original runtime for deterministic restoration. Models with failed real-checkpoint equivalence stay eager when automatic selection is requested and reject an explicit requirement.

Attention backends

An attention backend is useful only when the loaded architecture exposes the required attention protocol. VoiceHub does not select one from a provider name. The pass validates the concrete call, optional dependency, CUDA device, dtype, and execution mode before changing the runtime.

Kernels

General and codec-specific kernel passes resolve registered implementations by capability. They do not build an extension implicitly, and they retain canonical state-dict keys. Explicit unsupported backends fail instead of silently using a different implementation.

Diffusion caching

Diffusion caching reuses intermediate work across related sampling steps. It is architecture- and schedule-sensitive and can trade quality for latency or memory. Keep it disabled when the model has no retained checkpoint evidence or when exact output behavior is required.

Diffusion sampling

Sampling optimization can reduce steps, alter guidance, or select a different solver. These are semantic changes rather than guaranteed-equivalent kernels. Record the complete configuration and compare audio quality against the eager baseline before retaining a policy.

Boundaries

VoiceHub currently has no registry-wide public quantization pass. Quantized checkpoint formats and provider-local loaders therefore remain model-specific and are not advertised as universal optimization support.

Parallelism is a training or serving topology, not a reversible model pass. Continuous batching belongs to a serving scheduler, not the model-mutation lifecycle. Both stay outside this registry until they have a model-independent public contract and complete coverage.

Next steps