124 lines
4.8 KiB
Python
124 lines
4.8 KiB
Python
import os
|
|
import torch
|
|
import logging
|
|
|
|
import folder_paths
|
|
|
|
from ..src.voxcpm.core import VoxCPM
|
|
from ..src.voxcpm.model.voxcpm import LoRAConfig
|
|
from .model_info import AVAILABLE_VOXCPM_MODELS, MODEL_CONFIGS
|
|
from .downloader import DownloadManager, DownloadCancelledError, get_download_manager
|
|
|
|
logger = logging.getLogger(__name__)
|
|
|
|
LOADED_MODELS_CACHE = {}
|
|
|
|
# ── Default LoRA Configuration Constants ─────────────────────────────
|
|
# These values are optimized for single-speaker voice cloning:
|
|
# - r=32: Good balance between capacity and overfitting risk for voice cloning.
|
|
# Higher rank (64+) is recommended for style/language adaptation.
|
|
# - alpha=16: Scaling factor (scaling = alpha / r = 0.5).
|
|
# Usually set to r or 2*r. Adjust to control LoRA influence strength.
|
|
# - enable_lm=True: Essential for voice quality (language model backbone).
|
|
# - enable_dit=True: Essential for voice quality (diffusion transformer).
|
|
# - enable_proj=False: Projection layers rarely benefit from LoRA.
|
|
DEFAULT_LORA_RANK = 32
|
|
DEFAULT_LORA_ALPHA = 16
|
|
DEFAULT_LORA_ENABLE_LM = True
|
|
DEFAULT_LORA_ENABLE_DIT = True
|
|
DEFAULT_LORA_ENABLE_PROJ = False
|
|
|
|
|
|
class VoxCPMModelHandler(torch.nn.Module):
|
|
"""
|
|
A lightweight handler for a VoxCPM model. It acts as a container
|
|
that ComfyUI's ModelPatcher can manage, while the actual heavy model
|
|
is loaded on demand.
|
|
"""
|
|
def __init__(self, model_name: str):
|
|
super().__init__()
|
|
self.model_name = model_name
|
|
self.model = None # This will hold the actual loaded VoxCPM instance
|
|
|
|
# Get model size from config
|
|
model_config = MODEL_CONFIGS.get(model_name, {})
|
|
size_gb = model_config.get("size_gb", 2.5)
|
|
# Allocate slightly more than model size for offloading calculations
|
|
self.size = int((size_gb + 0.5) * (1024**3))
|
|
|
|
|
|
class VoxCPMLoader:
|
|
@staticmethod
|
|
def load_model(model_name: str, client_id: str = None):
|
|
"""
|
|
Loads a VoxCPM model, downloading it if necessary. Caches the loaded model instance.
|
|
|
|
Args:
|
|
model_name: Name of the model to load.
|
|
client_id: Optional ComfyUI client ID for targeted progress events.
|
|
|
|
Raises:
|
|
DownloadCancelledError: If the user cancels the download.
|
|
RuntimeError: If the model path cannot be determined.
|
|
ValueError: If the model name is not found.
|
|
"""
|
|
if model_name in LOADED_MODELS_CACHE:
|
|
logger.info(f"Using cached VoxCPM model instance: {model_name}")
|
|
return LOADED_MODELS_CACHE[model_name]
|
|
|
|
model_info = AVAILABLE_VOXCPM_MODELS.get(model_name)
|
|
if not model_info:
|
|
# Fall back to MODEL_CONFIGS if not in AVAILABLE_VOXCPM_MODELS
|
|
model_info = MODEL_CONFIGS.get(model_name)
|
|
if not model_info:
|
|
raise ValueError(f"Model '{model_name}' not found. Available models: {list(MODEL_CONFIGS.keys())}")
|
|
|
|
voxcpm_path = None
|
|
|
|
if model_info.get("type") == "local":
|
|
voxcpm_path = model_info["path"]
|
|
logger.info(f"Loading local model from: {voxcpm_path}")
|
|
|
|
else:
|
|
# Official model from HuggingFace
|
|
base_tts_path = os.path.join(folder_paths.get_folder_paths("tts")[0])
|
|
voxcpm_models_dir = os.path.join(base_tts_path, "VoxCPM")
|
|
os.makedirs(voxcpm_models_dir, exist_ok=True)
|
|
|
|
voxcpm_path = os.path.join(voxcpm_models_dir, model_name)
|
|
|
|
has_bin = os.path.exists(os.path.join(voxcpm_path, "pytorch_model.bin"))
|
|
has_safe = os.path.exists(os.path.join(voxcpm_path, "model.safetensors"))
|
|
|
|
if not (has_bin or has_safe):
|
|
logger.info(f"Downloading official VoxCPM model '{model_name}' from {model_info['repo_id']}...")
|
|
download_manager = get_download_manager()
|
|
download_manager.download_model(
|
|
model_name=model_name,
|
|
repo_id=model_info["repo_id"],
|
|
local_dir=voxcpm_path,
|
|
client_id=client_id,
|
|
)
|
|
|
|
if not voxcpm_path:
|
|
raise RuntimeError(f"Could not determine path for model '{model_name}'")
|
|
|
|
logger.info("Instantiating VoxCPM model...")
|
|
|
|
default_lora_config = LoRAConfig(
|
|
enable_lm=DEFAULT_LORA_ENABLE_LM,
|
|
enable_dit=DEFAULT_LORA_ENABLE_DIT,
|
|
enable_proj=DEFAULT_LORA_ENABLE_PROJ,
|
|
r=DEFAULT_LORA_RANK,
|
|
alpha=DEFAULT_LORA_ALPHA
|
|
)
|
|
|
|
model_instance = VoxCPM(
|
|
voxcpm_model_path=voxcpm_path,
|
|
enable_denoiser=False,
|
|
optimize=False,
|
|
lora_config=default_lora_config
|
|
)
|
|
|
|
LOADED_MODELS_CACHE[model_name] = model_instance
|
|
return model_instance |