GPT-SoVITS/tools/portable_runtime.py

98 lines
3.5 KiB
Python

"""Runtime initialization for the bundled GPT-SoVITS environment."""
from __future__ import annotations
import os
import sys
from importlib import import_module
from importlib.machinery import ModuleSpec
from importlib.metadata import version
from pathlib import Path
from threading import RLock
from types import ModuleType
PROJECT_ROOT = Path(__file__).resolve().parents[1]
RUNTIME_DIR = PROJECT_ROOT / "runtime"
SITE_PACKAGES = RUNTIME_DIR / "Lib" / "site-packages"
FFMPEG_EXE = RUNTIME_DIR / "ffmpeg.exe"
FFPROBE_EXE = RUNTIME_DIR / "ffprobe.exe"
_DLL_HANDLES = []
_DLL_READY = False
_ACTIVE = False
_FLASH_IMPORT_READY = False
_FLASH_IMPORT_LOCK = RLock()
def _prepare_dlls() -> None:
global _DLL_READY
if _DLL_READY or os.name != "nt":
return
directories = [
RUNTIME_DIR,
SITE_PACKAGES / "torch" / "lib",
PROJECT_ROOT / "GPT_SoVITS" / "Accel" / "vendor" / "cuda" / "bin",
*sorted((SITE_PACKAGES / "nvidia").glob("*/bin")),
]
for directory in directories:
if directory.is_dir():
_DLL_HANDLES.append(os.add_dll_directory(str(directory)))
# cuDNN dynamically loads its component DLLs through the process PATH.
os.environ["PATH"] = str(directory) + os.pathsep + os.environ.get("PATH", "")
_DLL_READY = True
def _lazy_apply_rotary_emb(*args, **kwargs):
return import_module("flash_attn.layers.rotary").apply_rotary_emb(*args, **kwargs)
def prepare_transformers_flash_attention() -> None:
global _FLASH_IMPORT_READY
with _FLASH_IMPORT_LOCK:
if _FLASH_IMPORT_READY:
return
_prepare_dlls()
import_utils = import_module("transformers.utils.import_utils")
if import_utils.is_flash_attn_2_available():
from tools.acceleration import _flash_attention_module
if _flash_attention_module() is None:
import_utils.is_flash_attn_2_available = lambda: False
import_module("transformers.utils").is_flash_attn_2_available = import_utils.is_flash_attn_2_available
if os.name != "nt" or version("transformers") != "4.51.3":
_FLASH_IMPORT_READY = True
return
# Transformers 4.51.3 imports Triton rotary even for BERT/HuBERT, which never use it.
rotary_name = "flash_attn.layers.rotary"
previous_rotary = sys.modules.get(rotary_name)
if previous_rotary is None:
rotary = ModuleType(rotary_name)
rotary.__spec__ = ModuleSpec(rotary_name, loader=None)
rotary.apply_rotary_emb = _lazy_apply_rotary_emb
sys.modules[rotary_name] = rotary
try:
import_module("transformers.modeling_flash_attention_utils")
finally:
if previous_rotary is None:
sys.modules.pop(rotary_name, None)
_FLASH_IMPORT_READY = True
def activate(*, change_cwd: bool = False) -> None:
global _ACTIVE
if change_cwd:
os.chdir(PROJECT_ROOT)
if not _ACTIVE:
sys.path.insert(0, str(PROJECT_ROOT))
sys.path.insert(0, str(PROJECT_ROOT / "GPT_SoVITS"))
os.environ["PATH"] = str(RUNTIME_DIR) + os.pathsep + os.environ.get("PATH", "")
os.environ.setdefault("G2PW", "0")
os.environ["NLTK_DATA"] = str(RUNTIME_DIR / "nltk_data")
try:
import _distutils_hack
_distutils_hack.add_shim()
except ImportError:
pass
_prepare_dlls()
_ACTIVE = True
prepare_transformers_flash_attention()