IPEX Duplicate importer V2 (#11310)

* Add gguf support.
* Avoid error when import ipex-llm for multiple times.
* Add check to avoid duplicate replace and revert.
* Add calling from check to avoid raising exceptions in the submodule.
* Add BIGDL_CHECK_DUPLICATE_IMPORT for controlling duplicate checker. Default is true.
This commit is contained in:
Qiyuan Gong 2024-06-19 16:29:19 +08:00 committed by GitHub
parent 271d82a4fc
commit 1eb884a249
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
4 changed files with 74 additions and 5 deletions

View file

@ -26,13 +26,15 @@ from .llm_patching import llm_patch, llm_unpatch
import sys import sys
import types import types
# Default is false, set to true to auto importing Intel Extension for PyTorch. # Default is True, set to False to disable auto importing Intel Extension for PyTorch.
USE_NPU = os.getenv("BIGDL_USE_NPU", 'False').lower() in ('true', '1', 't') USE_NPU = os.getenv("BIGDL_USE_NPU", 'False').lower() in ('true', '1', 't')
BIGDL_IMPORT_IPEX = os.getenv("BIGDL_IMPORT_IPEX", 'True').lower() in ('true', '1', 't') BIGDL_IMPORT_IPEX = os.getenv("BIGDL_IMPORT_IPEX", 'True').lower() in ('true', '1', 't')
BIGDL_IMPORT_IPEX = not USE_NPU and BIGDL_IMPORT_IPEX BIGDL_IMPORT_IPEX = not USE_NPU and BIGDL_IMPORT_IPEX
if BIGDL_IMPORT_IPEX: if BIGDL_IMPORT_IPEX:
# Import Intel Extension for PyTorch as ipex if XPU version is installed # Import Intel Extension for PyTorch as ipex if XPU version is installed
from .utils.ipex_importer import ipex_importer from .utils.ipex_importer import ipex_importer
# Avoid duplicate import
if ipex_importer.get_ipex_version() is None:
ipex_importer.import_ipex() ipex_importer.import_ipex()
# Default is true, set to true to auto patching bigdl-llm to ipex_llm. # Default is true, set to true to auto patching bigdl-llm to ipex_llm.

View file

@ -773,6 +773,9 @@ def ggml_convert_low_bit(model, qtype, optimize_model=True,
f"{list(gguf_mixed_qtype.keys())[index]} " f"{list(gguf_mixed_qtype.keys())[index]} "
f"format......") f"format......")
modules_to_not_convert = [] if modules_to_not_convert is None else modules_to_not_convert modules_to_not_convert = [] if modules_to_not_convert is None else modules_to_not_convert
# Disable ipex duplicate import checker
from ipex_llm.utils.ipex_importer import revert_import
revert_import()
# using ipex_llm optimizer before changing to bigdl linear # using ipex_llm optimizer before changing to bigdl linear
_enable_ipex = get_enable_ipex() _enable_ipex = get_enable_ipex()

View file

@ -30,6 +30,9 @@ qtype_map = {
def load_gguf_model(fpath: str, dtype: torch.dtype = torch.float, low_bit: str = "sym_int4"): def load_gguf_model(fpath: str, dtype: torch.dtype = torch.float, low_bit: str = "sym_int4"):
from .gguf import GGUFFileLoader from .gguf import GGUFFileLoader
# Disable ipex duplicate import checker
from ipex_llm.utils.ipex_importer import revert_import
revert_import()
loader = GGUFFileLoader(fpath) loader = GGUFFileLoader(fpath)
model_family = loader.config["general.architecture"] model_family = loader.config["general.architecture"]

View file

@ -18,15 +18,73 @@ from importlib.metadata import distribution, PackageNotFoundError
import logging import logging
import builtins import builtins
import sys import sys
from ipex_llm.utils.common import log4Error import os
import inspect import inspect
from ipex_llm.utils.common import log4Error
# Save the original __import__ function
original_import = builtins.__import__ # Default is True, set to False to disable IPEX duplicate checker
BIGDL_CHECK_DUPLICATE_IMPORT = os.getenv("BIGDL_CHECK_DUPLICATE_IMPORT",
'True').lower() in ('true', '1', 't')
RAW_IMPORT = None
IS_IMPORT_REPLACED = False
ipex_duplicate_import_error = "intel_extension_for_pytorch has already been automatically " + \ ipex_duplicate_import_error = "intel_extension_for_pytorch has already been automatically " + \
"imported. Please avoid importing it again!" "imported. Please avoid importing it again!"
def replace_import():
global RAW_IMPORT, IS_IMPORT_REPLACED
# Avoid multiple replacement
if not IS_IMPORT_REPLACED and RAW_IMPORT is None:
# Save the original __import__ function
RAW_IMPORT = builtins.__import__
builtins.__import__ = custom_ipex_import
IS_IMPORT_REPLACED = True
def revert_import():
if not BIGDL_CHECK_DUPLICATE_IMPORT:
return
global RAW_IMPORT, IS_IMPORT_REPLACED
# Only revert once
if RAW_IMPORT is not None and IS_IMPORT_REPLACED:
builtins.__import__ = RAW_IMPORT
IS_IMPORT_REPLACED = False
def get_calling_package():
"""
Return calling package name, e.g., ipex_llm.transformers
"""
# Get the current stack frame
frame = inspect.currentframe()
# Get the caller's frame
caller_frame = frame.f_back.f_back
# Get the caller's module
module = inspect.getmodule(caller_frame)
if module:
# Return the module's package name
return module.__package__
return None
def custom_ipex_import(name, globals=None, locals=None, fromlist=(), level=0):
"""
Custom import function to avoid importing ipex again
"""
if fromlist is not None or '.' in name:
return RAW_IMPORT(name, globals, locals, fromlist, level)
# Avoid errors in submodule import
calling = get_calling_package()
if calling is not None:
return RAW_IMPORT(name, globals, locals, fromlist, level)
# Only check ipex for main thread
if name == "ipex" or name == "intel_extension_for_pytorch":
log4Error.invalidInputError(False,
ipex_duplicate_import_error)
return RAW_IMPORT(name, globals, locals, fromlist, level)
class IPEXImporter: class IPEXImporter:
""" """
Auto import Intel Extension for PyTorch as ipex, Auto import Intel Extension for PyTorch as ipex,
@ -71,6 +129,9 @@ class IPEXImporter:
ipex_duplicate_import_error) ipex_duplicate_import_error)
self.directly_import_ipex() self.directly_import_ipex()
self.ipex_version = ipex.__version__ self.ipex_version = ipex.__version__
# Replace builtin import to avoid duplicate ipex import
if BIGDL_CHECK_DUPLICATE_IMPORT:
replace_import()
logging.info("intel_extension_for_pytorch auto imported") logging.info("intel_extension_for_pytorch auto imported")
def directly_import_ipex(self): def directly_import_ipex(self):