Switch build trigger from push to workflow_dispatch

This reverts commit eb6a39d292.

.gitea/workflows/build.yml

@@ -1,10 +1,7 @@
 name: Build ROCm Image
 
 on:
-  push:
+  workflow_dispatch:
-    branches:
-      - main
-      - dev
 
 jobs:
   build:
@@ -23,21 +20,14 @@ jobs:
           username: ${{ secrets.REGISTRY_USERNAME }}
           password: ${{ secrets.REGISTRY_TOKEN }}
 
-      - name: Set tags
-        id: tags
-        run: |
-          if [ "${{ gitea.ref_name }}" = "main" ]; then
-            echo "tags=git.sdgarren.com/scott/rocm-chatterbox-whisper:latest,git.sdgarren.com/scott/rocm-chatterbox-whisper:${{ gitea.sha }}" >> $GITHUB_OUTPUT
-          else
-            echo "tags=git.sdgarren.com/scott/rocm-chatterbox-whisper:${{ gitea.ref_name }},git.sdgarren.com/scott/rocm-chatterbox-whisper:${{ gitea.sha }}" >> $GITHUB_OUTPUT
-          fi
-
       - name: Build and Push
         uses: docker/build-push-action@v6
         with:
           context: .
           file: Dockerfile.rocm
           push: true
-          tags: ${{ steps.tags.outputs.tags }}
+          tags: |
+            git.sdgarren.com/scott/rocm-chatterbox-whisper:latest
+            git.sdgarren.com/scott/rocm-chatterbox-whisper:${{ gitea.sha }}
           cache-from: type=registry,ref=git.sdgarren.com/scott/rocm-chatterbox-whisper:latest
           cache-to: type=inline

Dockerfile.rocm

@@ -1,4 +1,4 @@
-FROM rocm/dev-ubuntu-22.04:7.2
+FROM rocm/dev-ubuntu-22.04:6.1
 
 ENV DEBIAN_FRONTEND=noninteractive \
     PYTHONDONTWRITEBYTECODE=1 \
@@ -7,37 +7,27 @@ ENV DEBIAN_FRONTEND=noninteractive \
     PIP_NO_CACHE_DIR=1
 
 RUN apt-get update && apt-get install -y --no-install-recommends \
-    software-properties-common \
+    python3 \
-    && add-apt-repository ppa:deadsnakes/ppa \
+    python3-pip \
-    && apt-get update && apt-get install -y --no-install-recommends \
+    python3-dev \
-    python3.11 \
-    python3.11-dev \
-    python3.11-distutils \
     git \
     ffmpeg \
     libsndfile1 \
     && rm -rf /var/lib/apt/lists/*
 
-# Install pip for Python 3.11
-RUN curl -sS https://bootstrap.pypa.io/get-pip.py | python3.11
-
-# Make python3.11 the default python3
-RUN update-alternatives --install /usr/bin/python3 python3 /usr/bin/python3.11 1 \
-    && update-alternatives --install /usr/bin/python python /usr/bin/python3.11 1
-
 WORKDIR /app
 
 # Step 1: Install ROCm-compatible PyTorch stack first.
 # This must happen before anything else to prevent pip from pulling CPU wheels.
 COPY requirements-rocm-init.txt .
-RUN python3 -m pip install -r requirements-rocm-init.txt
+RUN pip3 install -r requirements-rocm-init.txt
 
 # Step 2: Install remaining dependencies (pinned to avoid overwriting torch).
 COPY requirements-rocm.txt .
-RUN python3 -m pip install -r requirements-rocm.txt
+RUN pip3 install -r requirements-rocm.txt
 
 # Step 3: Install chatterbox with --no-deps so pip cannot replace ROCm torch.
-RUN python3 -m pip install --no-deps chatterbox-tts
+RUN pip3 install --no-deps chatterbox-tts
 
 # Application source
 COPY engine.py config.py wyoming_handler.py wyoming_voices.py main.py ./

config.yaml

@@ -13,11 +13,11 @@ tts_engine:
 generation_defaults:
   # Turbo model: uses temperature only (exaggeration/cfg_weight ignored)
   # Standard model: uses exaggeration and cfg_weight (temperature ignored)
-  temperature: 0.9
+  temperature: 0.8
   exaggeration: 0.5
   cfg_weight: 0.5
   # seed: 0 = random each call, >0 = reproducible output
-  seed: 1960
+  seed: 0
 
 wyoming:
   host: "0.0.0.0"

engine.py

@@ -8,6 +8,12 @@ chatterbox_model = None
 _sample_rate = 24000
 _is_turbo = False
 
+# Cache: voice file path → prepared conditionals object.
+# prepare_conditionals loads audio, runs s3tokenizer + voice encoder, and
+# builds mel embeddings — expensive work that only depends on the reference
+# audio, not the text. Cache it so multi-chunk requests pay the cost once.
+_cond_cache: dict = {}
+
 
 def _test_cuda() -> bool:
     try:
@@ -51,6 +57,26 @@ def load_model() -> bool:
         return False
 
 
+def prepare_voice(audio_prompt_path: str) -> None:
+    """
+    Pre-compute and cache the voice conditionals for a reference audio file.
+    Calling this once avoids repeating the s3tokenizer + voice encoder work
+    on every synthesis chunk that uses the same voice.
+    """
+    if chatterbox_model is None:
+        return
+    if audio_prompt_path in _cond_cache:
+        return
+    if not _is_turbo:
+        return  # only turbo exposes prepare_conditionals
+
+    logger.info(f"Preparing voice conditionals for '{audio_prompt_path}'")
+    with torch.inference_mode():
+        chatterbox_model.prepare_conditionals(audio_prompt_path)
+    _cond_cache[audio_prompt_path] = chatterbox_model.conds
+    logger.info("Voice conditionals cached")
+
+
 def get_sample_rate() -> int:
     return _sample_rate
 
@@ -71,8 +97,16 @@ def synthesize(
         if torch.cuda.is_available():
             torch.cuda.manual_seed_all(seed)
 
+    # Restore cached conditionals so generate() skips prepare_conditionals.
+    if audio_prompt_path and _is_turbo:
+        if audio_prompt_path not in _cond_cache:
+            prepare_voice(audio_prompt_path)
+        chatterbox_model.conds = _cond_cache[audio_prompt_path]
+
     kwargs: dict = {}
-    if audio_prompt_path:
+    # Don't pass audio_prompt_path — conds are already set above.
+    # For non-turbo models there's no cache, pass path as normal.
+    if audio_prompt_path and not _is_turbo:
         kwargs["audio_prompt_path"] = audio_prompt_path
 
     if _is_turbo:

main.py

@@ -19,8 +19,18 @@ logger = logging.getLogger(__name__)
 
 
 def _warmup(voices: dict) -> None:
-    """Run one synthesis to populate MIOpen's in-memory kernel cache."""
+    """Pre-compute voice conditionals and populate MIOpen's kernel cache."""
     from wyoming_voices import resolve_voice
+
+    # Pre-compute conditionals for all discovered voices so the first real
+    # request doesn't pay the s3tokenizer + voice encoder cost.
+    for name, path in voices.items():
+        try:
+            engine.prepare_voice(path)
+        except Exception:
+            logger.warning(f"Failed to prepare voice '{name}' (non-fatal)", exc_info=True)
+
+    # Synthesis warmup to populate MIOpen's in-memory kernel cache.
     audio_prompt = resolve_voice(None, voices) if voices else None
     logger.info("Running warmup synthesis...")
     try:

requirements-rocm-init.txt

@@ -1,5 +1,5 @@
---index-url https://download.pytorch.org/whl/rocm7.2
+--index-url https://download.pytorch.org/whl/rocm6.1
-torch==2.11.0
+torch==2.5.1
-torchaudio==2.11.0
+torchaudio==2.5.1
-torchvision==0.26.0
+torchvision==0.20.1
-pytorch_triton_rocm==3.3.0
+pytorch_triton_rocm==3.1.0

requirements-rocm.txt

@@ -5,9 +5,9 @@ librosa==0.11.0
 pyloudnorm
 
 # ML dependencies (pinned to match chatterbox without overwriting ROCm torch)
-transformers>=4.50.0
+transformers==4.46.3
-diffusers>=0.32.0
+diffusers==0.29.0
-safetensors>=0.4.5
+safetensors>=0.4.1
 huggingface-hub
 omegaconf