Skip to content

The agent engineering platform.

fix(huggingface): remove retired IPEX backend (#39649)

wanhakim3539ca08d92e
diff --git a/libs/partners/huggingface/langchain_huggingface/embeddings/huggingface.py b/libs/partners/huggingface/langchain_huggingface/embeddings/huggingface.py
index 8c55348cee..2ba2bc9f20 100644
--- a/libs/partners/huggingface/langchain_huggingface/embeddings/huggingface.py
+++ b/libs/partners/huggingface/langchain_huggingface/embeddings/huggingface.py
@@ -1,19 +1,11 @@
from __future__ import annotations
+import warnings
from typing import Any
from langchain_core.embeddings import Embeddings
from pydantic import BaseModel, ConfigDict, Field
-from langchain_huggingface.utils.import_utils import (
- IMPORT_ERROR,
- is_ipex_available,
- is_optimum_intel_available,
- is_optimum_intel_version,
-)
-
-_MIN_OPTIMUM_VERSION = "1.22"
-
class HuggingFaceEmbeddings(BaseModel, Embeddings):
"""HuggingFace sentence_transformers embedding models.
@@ -74,25 +66,15 @@ class HuggingFaceEmbeddings(BaseModel, Embeddings):
raise ImportError(msg) from exc
if self.model_kwargs.get("backend", "torch") == "ipex":
- if not is_optimum_intel_available() or not is_ipex_available():
- msg = f"Backend: ipex {IMPORT_ERROR.format('optimum[ipex]')}"
- raise ImportError(msg)
-
- if is_optimum_intel_version("<", _MIN_OPTIMUM_VERSION):
- msg = (
- f"Backend: ipex requires optimum-intel>="
- f"{_MIN_OPTIMUM_VERSION}. You can install it with pip: "
- "`pip install --upgrade --upgrade-strategy eager "
- "`optimum[ipex]`."
- )
- raise ImportError(msg)
-
- from optimum.intel import IPEXSentenceTransformer # type: ignore[import]
-
- model_cls = IPEXSentenceTransformer
+ msg = (
+ "`backend='ipex'` is no longer supported; the default `torch` "
+ "backend will be used instead. Intel GPU acceleration is "
+ "natively supported in PyTorch 2.5 and later."
+ )
+ warnings.warn(msg, UserWarning, stacklevel=2)
+ self.model_kwargs["backend"] = "torch"
- else:
- model_cls = sentence_transformers.SentenceTransformer
+ model_cls = sentence_transformers.SentenceTransformer
self._client = model_cls(
self.model_name, cache_folder=self.cache_folder, **self.model_kwargs
diff --git a/libs/partners/huggingface/langchain_huggingface/llms/huggingface_pipeline.py b/libs/partners/huggingface/langchain_huggingface/llms/huggingface_pipeline.py
index c710997613..a354191c1f 100644
--- a/libs/partners/huggingface/langchain_huggingface/llms/huggingface_pipeline.py
+++ b/libs/partners/huggingface/langchain_huggingface/llms/huggingface_pipeline.py
@@ -2,6 +2,7 @@ from __future__ import annotations # type: ignore[import-not-found]
import importlib.util
import logging
+import warnings
from collections.abc import Iterator, Mapping
from typing import Any
@@ -14,7 +15,6 @@ from typing_extensions import Self
from langchain_huggingface._version import __version__
from langchain_huggingface.utils.import_utils import (
IMPORT_ERROR,
- is_ipex_available,
is_openvino_available,
is_optimum_intel_available,
is_optimum_intel_version,
@@ -158,7 +158,16 @@ class HuggingFacePipeline(BaseLLM):
_model_kwargs["device_map"] = device_map
tokenizer = AutoTokenizer.from_pretrained(model_id, **_model_kwargs)
- if backend in {"openvino", "ipex"}:
+ if backend == "ipex":
+ msg = (
+ "`backend='ipex'` is no longer supported; the default "
+ "`transformers` backend will be used instead. Intel GPU acceleration "
+ "is natively supported in PyTorch 2.5 and later."
+ )
+ warnings.warn(msg, UserWarning, stacklevel=2)
+ backend = "default"
+
+ if backend == "openvino":
if task not in VALID_TASKS:
msg = (
f"Got invalid task {task}, "
@@ -170,51 +179,26 @@ class HuggingFacePipeline(BaseLLM):
if not is_optimum_intel_available():
raise ImportError(err_msg)
- # TODO: upgrade _MIN_OPTIMUM_VERSION to 1.22 after release
- min_optimum_version = (
- "1.22"
- if backend == "ipex" and task != "text-generation"
- else _MIN_OPTIMUM_VERSION
- )
- if is_optimum_intel_version("<", min_optimum_version):
+ if is_optimum_intel_version("<", _MIN_OPTIMUM_VERSION):
msg = (
f"Backend: {backend} requires optimum-intel>="
- f"{min_optimum_version}. You can install it with pip: "
+ f"{_MIN_OPTIMUM_VERSION}. You can install it with pip: "
"`pip install --upgrade --upgrade-strategy eager "
f"`optimum[{backend}]`."
)
raise ImportError(msg)
- if backend == "openvino":
- if not is_openvino_available():
- raise ImportError(err_msg)
-
- from optimum.intel import ( # type: ignore[import]
- OVModelForCausalLM,
- OVModelForSeq2SeqLM,
- )
-
- model_cls = (
- OVModelForCausalLM
- if task == "text-generation"
- else OVModelForSeq2SeqLM
- )
- else:
- if not is_ipex_available():
- raise ImportError(err_msg)
-
- if task == "text-generation":
- from optimum.intel import (
- IPEXModelForCausalLM, # type: ignore[import]
- )
+ if not is_openvino_available():
+ raise ImportError(err_msg)
- model_cls = IPEXModelForCausalLM
- else:
- from optimum.intel import (
- IPEXModelForSeq2SeqLM, # type: ignore[import]
- )
+ from optimum.intel import ( # type: ignore[import]
+ OVModelForCausalLM,
+ OVModelForSeq2SeqLM,
+ )
- model_cls = IPEXModelForSeq2SeqLM
+ model_cls = (
+ OVModelForCausalLM if task == "text-generation" else OVModelForSeq2SeqLM
+ )
else:
model_cls = (
diff --git a/libs/partners/huggingface/langchain_huggingface/utils/import_utils.py b/libs/partners/huggingface/langchain_huggingface/utils/import_utils.py
index f217517a98..bee78a26c7 100644
--- a/libs/partners/huggingface/langchain_huggingface/utils/import_utils.py
+++ b/libs/partners/huggingface/langchain_huggingface/utils/import_utils.py
@@ -35,9 +35,6 @@ if _optimum_intel_available:
except importlib.metadata.PackageNotFoundError:
_optimum_intel_available = False
-
-_ipex_available = importlib.util.find_spec("intel_extension_for_pytorch") is not None
-
_openvino_available = importlib.util.find_spec("openvino") is not None
@@ -81,10 +78,6 @@ def is_optimum_intel_available() -> bool:
return _optimum_intel_available
-def is_ipex_available() -> bool:
- return _ipex_available
-
-
def is_openvino_available() -> bool:
return _openvino_available