v0.3.1

ed-donner/llm_engineeringv0.3.1Jul 15, 2026by andimarafioti

AI Summary

Standardizes the backend warmup process across Torch and GGML backends to resolve interface mismatches, ensuring compatibility with external projects like speech-to-speech.

Key Highlights

Added public `model.warmup(prefill_len=100)` API shared by Torch and GGML backends
Kept the former private Torch `_warmup()` entry point as a compatibility alias
Implemented CUDA graph capture for Torch while treating GGML warmup as a no-op
Updated internal, demo, and test callers to use the public API

New Features

Public backend warmup API

Full Release Notes

## Highlights

- Add a public `model.warmup(prefill_len=100)` API shared by the Torch and GGML backends.
- Keep the former private Torch `_warmup()` entry point as a compatibility alias.
- Capture CUDA graphs once on Torch while treating GGML warmup as a safe no-op until qwentts.cpp exposes a native preparation hook.
- Update internal, demo, example, and test callers to use the public API.
- Install the published PyPI package in the demo Space deployment.

This resolves the backend interface mismatch reported in [huggingface/speech-to-speech#354](https://github.com/huggingface/speech-to-speech/issues/354).

## Installation

```bash
pip install faster-qwen3-tts==0.3.1
```

For the optional GGML backend:

```bash
pip install "faster-qwen3-tts[ggml]==0.3.1"
```

## Changes

- [#128](https://github.com/andimarafioti/faster-qwen3-tts/pull/128) — Expose public backend warmup API
- [#122](https://github.com/andimarafioti/faster-qwen3-tts/pull/122) — Use the PyPI package in the demo Space

**Full changelog:** https://github.com/andimarafioti/faster-qwen3-tts/compare/v0.3.0...v0.3.1