The open-source AI voice studio. Clone, dictate, create.
fix: install chatterbox-tts with --no-deps to avoid numpy pin conflict
James Pine9beb9d7fec4b
diff --git a/Makefile b/Makefileindex 3ff4080c..38918613 100644--- a/Makefile+++ b/Makefile@@ -48,6 +48,7 @@ setup-python: $(VENV)/bin/activate ## Set up Python virtual environment and depe@echo -e "$(BLUE)Installing Python dependencies...$(NC)"$(PIP) install --upgrade pip$(PIP) install -r $(BACKEND_DIR)/requirements.txt+ $(PIP) install --no-deps chatterbox-tts@if [ "$$(uname -m)" = "arm64" ] && [ "$$(uname)" = "Darwin" ]; then \echo -e "$(BLUE)Detected Apple Silicon - installing MLX dependencies...$(NC)"; \$(PIP) install -r $(BACKEND_DIR)/requirements-mlx.txt; \diff --git a/backend/backends/chatterbox_backend.py b/backend/backends/chatterbox_backend.pyindex 84eb8365..9078f686 100644--- a/backend/backends/chatterbox_backend.py+++ b/backend/backends/chatterbox_backend.py@@ -18,7 +18,6 @@ import numpy as npfrom . import TTSBackendfrom ..utils.audio import normalize_audio, load_audiofrom ..utils.progress import get_progress_manager-from ..utils.hf_progress import HFProgressTracker, create_hf_progress_callbackfrom ..utils.tasks import get_task_managerlogger = logging.getLogger(__name__)@@ -110,63 +109,59 @@ class ChatterboxTTSBackend:is_cached = self._is_model_cached()- try:- progress_callback = create_hf_progress_callback(model_name, progress_manager)- tracker = HFProgressTracker(progress_callback, filter_non_downloads=is_cached)+ if not is_cached:+ task_manager.start_download(model_name)+ progress_manager.update_progress(+ model_name=model_name,+ current=0,+ total=0,+ filename="Downloading Chatterbox model...",+ status="downloading",+ )- if not is_cached:- task_manager.start_download(model_name)- progress_manager.update_progress(- model_name=model_name,- current=0,- total=0,- filename="Downloading Chatterbox model...",- status="downloading",- )+ try:+ device = self._get_device()+ self._device = device- with tracker.patch_download():- device = self._get_device()- self._device = device+ logger.info(f"Loading Chatterbox Multilingual TTS on {device}...")- logger.info(f"Loading Chatterbox Multilingual TTS on {device}...")+ import torch+ from chatterbox.mtl_tts import ChatterboxMultilingualTTS++ # Monkey-patch torch.load for CPU loading. The model's .pt files+ # were saved on CUDA; from_pretrained() doesn't pass map_location+ # so loading on CPU fails without this.+ if device == "cpu":+ _orig_torch_load = torch.load++ def _patched_load(*args, **kwargs):+ kwargs.setdefault("map_location", "cpu")+ return _orig_torch_load(*args, **kwargs)++ with ChatterboxTTSBackend._load_lock:+ torch.load = _patched_load+ try:+ self.model = ChatterboxMultilingualTTS.from_pretrained(+ device=device,+ )+ finally:+ torch.load = _orig_torch_load+ else:+ self.model = ChatterboxMultilingualTTS.from_pretrained(+ device=device,+ )- import torch- from chatterbox.mtl_tts import ChatterboxMultilingualTTS-- # Monkey-patch torch.load for CPU loading. The model's .pt files- # were saved on CUDA; from_pretrained() doesn't pass map_location- # so loading on CPU fails without this.- if device == "cpu":- _orig_torch_load = torch.load-- def _patched_load(*args, **kwargs):- kwargs.setdefault("map_location", "cpu")- return _orig_torch_load(*args, **kwargs)-- with ChatterboxTTSBackend._load_lock:- torch.load = _patched_load- try:- self.model = ChatterboxMultilingualTTS.from_pretrained(- device=device,- )- finally:- torch.load = _orig_torch_load- else:- self.model = ChatterboxMultilingualTTS.from_pretrained(- device=device,- )-- # Fix: transformers >= 4.36 defaults LlamaModel to sdpa attention- # which doesn't support output_attentions=True (needed by- # Chatterbox's AlignmentStreamAnalyzer). Force eager attention.- t3_tfmr = self.model.t3.tfmr- if hasattr(t3_tfmr, "config") and hasattr(- t3_tfmr.config, "_attn_implementation"- ):- t3_tfmr.config._attn_implementation = "eager"- for layer in getattr(t3_tfmr, "layers", []):- if hasattr(layer, "self_attn"):- layer.self_attn._attn_implementation = "eager"+ # Fix: transformers >= 4.36 defaults LlamaModel to sdpa attention+ # which doesn't support output_attentions=True (needed by+ # Chatterbox's AlignmentStreamAnalyzer). Force eager attention.+ t3_tfmr = self.model.t3.tfmr+ if hasattr(t3_tfmr, "config") and hasattr(+ t3_tfmr.config, "_attn_implementation"+ ):+ t3_tfmr.config._attn_implementation = "eager"+ for layer in getattr(t3_tfmr, "layers", []):+ if hasattr(layer, "self_attn"):+ layer.self_attn._attn_implementation = "eager"if not is_cached:progress_manager.mark_complete(model_name)@@ -179,13 +174,15 @@ class ChatterboxTTSBackend:"chatterbox-tts package not found. ""Install with: pip install chatterbox-tts")- progress_manager.mark_error(model_name, str(e))- task_manager.error_download(model_name, str(e))+ if not is_cached:+ progress_manager.mark_error(model_name, str(e))+ task_manager.error_download(model_name, str(e))raiseexcept Exception as e:logger.error(f"Failed to load Chatterbox: {e}")- progress_manager.mark_error(model_name, str(e))- task_manager.error_download(model_name, str(e))+ if not is_cached:+ progress_manager.mark_error(model_name, str(e))+ task_manager.error_download(model_name, str(e))raisedef unload_model(self) -> None:diff --git a/backend/requirements.txt b/backend/requirements.txtindex 53e9e6d3..10574711 100644--- a/backend/requirements.txt+++ b/backend/requirements.txt@@ -21,8 +21,17 @@ qwen-tts>=0.0.5linacodec @ git+https://github.com/ysharma3501/LinaCodec.gitZipvoice @ git+https://github.com/ysharma3501/LuxTTS.git-# Chatterbox TTS (multilingual voice cloning, includes Hebrew)-chatterbox-tts>=0.1.0+# Chatterbox TTS sub-dependencies (chatterbox-tts itself is installed+# --no-deps in the setup script because it pins numpy<1.26 / torch==2.6+# which are incompatible with Python 3.12+)+conformer>=0.3.2+diffusers>=0.29.0+omegaconf+pykakasi+resemble-perth>=1.0.1+s3tokenizer+spacy-pkuseg+pyloudnorm# Audio processinglibrosa>=0.10.0diff --git a/justfile b/justfileindex b0e1981f..ad172fa8 100644--- a/justfile+++ b/justfile@@ -38,6 +38,8 @@ setup-python:echo "Installing Python dependencies..."{{ pip }} install --upgrade pip -q{{ pip }} install -r {{ backend_dir }}/requirements.txt+ # Chatterbox pins numpy<1.26 / torch==2.6 which break on Python 3.12++ {{ pip }} install --no-deps chatterbox-tts# Apple Silicon: install MLX backendif [ "$(uname -m)" = "arm64" ] && [ "$(uname)" = "Darwin" ]; thenecho "Detected Apple Silicon — installing MLX dependencies..."