Dockerfile

# Pocket TTS gets its own image: `pocket-tts` pins numpy>=2 and pulls a server stack
# (fastapi/uvicorn/typer) plus sentencepiece, and it is the only backend here that needs no
# torchaudio at all — its audio I/O is stdlib `wave` + soundfile.
FROM nvidia/cuda:12.9.0-runtime-ubuntu24.04

# Avoid interactive prompts during package installation
ENV DEBIAN_FRONTEND=noninteractive

# System deps: libsndfile1 for soundfile I/O (writing generated wavs + decoding the dataset's
# prompt_audio column); ffmpeg for librosa/soundfile fallback decoders; git for pip VCS installs.
RUN apt-get update && apt-get install -y --no-install-recommends \
    python3 \
    python3-pip \
    python3-dev \
    git \
    ffmpeg \
    libsndfile1 \
    && rm -rf /var/lib/apt/lists/*

# Set Python alias (Ubuntu 24.04 ships Python 3.12; pocket-tts requires >=3.10,<3.15)
RUN ln -sf /usr/bin/python3 /usr/bin/python

# Allow pip to install packages system-wide in the container (PEP 668)
ENV PIP_BREAK_SYSTEM_PACKAGES=1

WORKDIR /app

# Upgrade pip so it prefers prebuilt manylinux wheels. --ignore-installed is required on
# Ubuntu 24.04: the Debian-installed pip has no RECORD file and cannot be uninstalled.
RUN pip install --no-cache-dir --upgrade --ignore-installed pip setuptools wheel

# Install PyTorch first (cu128 wheels for CUDA 12.8+/12.9 compat) so `pip install pocket-tts`
# does not pull a CPU wheel for its `torch>=2.5.0` requirement. NOTE: no torchaudio — pocket_tts
# reads wavs with the stdlib `wave` module and other formats with soundfile.
RUN pip install --no-cache-dir \
    torch==2.8.0 \
    --index-url https://download.pytorch.org/whl/cu128

# pocket-tts, version-pinned for reproducibility: the effective sampling temperature is a
# per-version detail (2.1.0 hard-defaults temp=0.7 in load_model's signature; upstream main moves
# it to a per-config `default_temperature` of 0.3 for English), and the package type-checks its own
# arguments at runtime with beartype, so a signature change is a hard failure rather than a
# silent one. The `[audio]` extra adds soundfile, needed for non-wav voice prompts.
RUN pip install --no-cache-dir "pocket-tts[audio]==2.1.0"

RUN pip install --no-cache-dir "datasets<4.0" tqdm soundfile librosa

# NOTE: the weights are deliberately NOT baked into this image, unlike the other backends.
# kyutai/pocket-tts is a GATED repo (auto-approve), and this image is published as a PUBLIC Space
# for HF Jobs to pull — baking gated weights into it would redistribute them outside the gate.
# They are fetched at runtime instead, so the job needs HF_TOKEN from an account that has accepted
# the terms at https://huggingface.co/kyutai/pocket-tts (submit_jobs.sh passes --secrets HF_TOKEN).
# Downloaded per run: the gated `languages/<lang>/model.safetensors`, plus the tokenizer and any
# predefined-voice states, which live in the ungated kyutai/pocket-tts-without-voice-cloning repo.

# Copy the full repository
COPY . /app

# Default entrypoint
ENTRYPOINT ["bash"]

# Keep-alive CMD so the Space runtime stays healthy. HF Jobs and `docker run`
# override this with their own command (e.g. run_eval.sh).
EXPOSE 7860
CMD ["-c", "python3 -m http.server 7860"]