Upload 3 files

Dockerfile

@@ -9,7 +9,9 @@
 FROM python:3.11-slim
 
 # Whisper needs ffmpeg to read audio files, this installs it at the
-# system level since it can't come from pip.
+# system level since it can't come from pip. build-essential is needed
+# because some packages in requirements.txt compile C extensions during
+# install, and the slim base image has no compiler by default.
 RUN apt-get update && apt-get install -y --no-install-recommends \
     build-essential \
     ffmpeg \
@@ -26,6 +28,19 @@ WORKDIR /home/user/app
 COPY --chown=user requirements.txt .
 RUN pip install --no-cache-dir --upgrade -r requirements.txt
 
+# YarnGPT downloads its WavTokenizer checkpoint itself on first import,
+# using a bare requests.get() call with no error checking. If that
+# download hiccups at runtime (inside the running container, on the
+# app's first request), it silently saves a corrupted or partial file
+# and the app fails later with a confusing, unrelated error.
+#
+# This downloads the exact same two files ahead of time, at build time,
+# using huggingface_hub's properly tested download function instead. If
+# this fails, the build fails clearly, right here, instead of the app
+# failing mysteriously after it's already live. YarnGPT's own downloader
+# checks whether the file already exists before downloading, so it will
+# find these already in place and skip straight past its own fragile
+# download step.
 RUN python - <<'PY'
 from huggingface_hub import hf_hub_download
 import os
@@ -49,7 +64,6 @@ hf_hub_download(
 print("WavTokenizer files downloaded.")
 PY
 
-
 COPY --chown=user . .
 
 ENV ENVIRONMENT=production

tts.py

@@ -10,8 +10,15 @@ Nigerian Pidgin uses MMS-TTS (facebook/mms-tts-pcm) instead, since YarnGPT
 has no Pidgin support. MMS-TTS is the only free option found with a
 dedicated Pidgin checkpoint.
 
-Note: importing this file triggers a one-time download of YarnGPT's audio
-tokenizer files. This is expected and only happens once.
+Note on the WavTokenizer checkpoint: YarnGPT downloads this itself on
+first import, using a bare requests.get() with no error checking, which
+can silently save a corrupted file if the download hiccups. The
+Dockerfile pre-downloads the same two files at build time using
+huggingface_hub's properly tested download function instead, so by the
+time this file is imported, YarnGPT finds them already in place and
+skips its own fragile download step. Locally (not in Docker), the first
+import still triggers YarnGPT's own download as normal, this only
+matters for the deployed container.
 
 Performance note: YarnGPT's own generate_speech() function reloads its
 full model from disk on every call, which is too slow for a live app.
@@ -27,21 +34,6 @@ import scipy.io.wavfile
 from transformers import VitsModel, AutoTokenizer
 from yarngpt.core import load_model_and_tokenizer, SPEAKER_MAPPING, AVAILABLE_SPEAKERS
 
-import yarngpt
-import yarngpt.core
-import inspect
-
-import audiotokenizer
-import inspect
-print("AudioTokenizer file:", audiotokenizer.__file__)
-source = inspect.getsource(audiotokenizer)
-print(source[:8000])
-
-print("YarnGPT:", yarngpt.__file__)
-print("MODEL_PATH:", yarngpt.core.MODEL_PATH)
-print("CONFIG_PATH:", yarngpt.core.CONFIG_PATH)
-print(inspect.getsource(yarngpt.core.load_model_and_tokenizer))
-
 # Default speaker used for each YarnGPT-supported language.
 YARNGPT_SPEAKERS = {
     "english": "idera",
@@ -71,26 +63,13 @@ def _get_pidgin_model():
         _pidgin_model = VitsModel.from_pretrained("facebook/mms-tts-pcm")
     return _pidgin_model, _pidgin_tokenizer
 
+
 def preload_models():
+    """Loads both TTS backends into memory ahead of time. Call this once
+    when the Flask app starts, so the first real request isn't slow."""
     _get_yarngpt_model()
     _get_pidgin_model()
 
-    
-# ===== DEBUG START =====
-import audiotokenizer
-
-print(audiotokenizer.__file__)
-
-from pathlib import Path
-
-model = Path("/home/user/.yarngpt/models/wavtokenizer_large_speech_320_24k.ckpt")
-print("Exists:", model.exists())
-
-if model.exists():
-    print("Size:", model.stat().st_size)
-    with open(model, "rb") as f:
-        print("First 100 bytes:", f.read(100))
-# ===== DEBUG END =====
 
 def _generate_yarngpt_speech(text, speaker, language, temperature=0.1, repetition_penalty=1.1, max_length=4000):
     """