Skip to content

The agent engineering platform.

huggingface: Add IPEX models support (#29179)

Ella Charlaix6f95db81b71a
diff --git a/libs/partners/huggingface/langchain_huggingface/llms/huggingface_pipeline.py b/libs/partners/huggingface/langchain_huggingface/llms/huggingface_pipeline.py
index 3e743a6428..daecda0d87 100644
--- a/libs/partners/huggingface/langchain_huggingface/llms/huggingface_pipeline.py
+++ b/libs/partners/huggingface/langchain_huggingface/llms/huggingface_pipeline.py
@@ -9,6 +9,14 @@ from langchain_core.language_models.llms import BaseLLM
from langchain_core.outputs import Generation, GenerationChunk, LLMResult
from pydantic import ConfigDict, model_validator
+from ..utils.import_utils import (
+ IMPORT_ERROR,
+ is_ipex_available,
+ is_openvino_available,
+ is_optimum_intel_available,
+ is_optimum_intel_version,
+)
+
DEFAULT_MODEL_ID = "gpt2"
DEFAULT_TASK = "text-generation"
VALID_TASKS = (
@@ -18,6 +26,8 @@ VALID_TASKS = (
"translation",
)
DEFAULT_BATCH_SIZE = 4
+_MIN_OPTIMUM_VERSION = "1.21"
+
logger = logging.getLogger(__name__)
@@ -126,70 +136,70 @@ class HuggingFacePipeline(BaseLLM):
_model_kwargs["device_map"] = device_map
tokenizer = AutoTokenizer.from_pretrained(model_id, **_model_kwargs)
- try:
- if task == "text-generation":
- if backend == "openvino":
- try:
- from optimum.intel.openvino import ( # type: ignore[import]
- OVModelForCausalLM,
- )
-
- except ImportError:
- raise ValueError(
- "Could not import optimum-intel python package. "
- "Please install it with: "
- "pip install 'optimum[openvino,nncf]' "
- )
- try:
- # use local model
- model = OVModelForCausalLM.from_pretrained(
- model_id, **_model_kwargs
- )
-
- except Exception:
- # use remote model
- model = OVModelForCausalLM.from_pretrained(
- model_id, export=True, **_model_kwargs
- )
- else:
- model = AutoModelForCausalLM.from_pretrained(
- model_id, **_model_kwargs
- )
- elif task in ("text2text-generation", "summarization", "translation"):
- if backend == "openvino":
- try:
- from optimum.intel.openvino import OVModelForSeq2SeqLM
-
- except ImportError:
- raise ValueError(
- "Could not import optimum-intel python package. "
- "Please install it with: "
- "pip install 'optimum[openvino,nncf]' "
- )
- try:
- # use local model
- model = OVModelForSeq2SeqLM.from_pretrained(
- model_id, **_model_kwargs
- )
-
- except Exception:
- # use remote model
- model = OVModelForSeq2SeqLM.from_pretrained(
- model_id, export=True, **_model_kwargs
- )
- else:
- model = AutoModelForSeq2SeqLM.from_pretrained(
- model_id, **_model_kwargs
- )
- else:
+ if backend in {"openvino", "ipex"}:
+ if task not in VALID_TASKS:
raise ValueError(
f"Got invalid task {task}, "
f"currently only {VALID_TASKS} are supported"
)
- except ImportError as e:
- raise ValueError(
- f"Could not load the {task} model due to missing dependencies."
- ) from e
+
+ err_msg = f'Backend: {backend} {IMPORT_ERROR.format(f"optimum[{backend}]")}'
+ 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):
+ raise ImportError(
+ f"Backend: {backend} requires optimum-intel>="
+ f"{min_optimum_version}. You can install it with pip: "
+ "`pip install --upgrade --upgrade-strategy eager "
+ f"`optimum[{backend}]`."
+ )
+
+ 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]
+ )
+
+ model_cls = IPEXModelForCausalLM
+ else:
+ from optimum.intel import (
+ IPEXModelForSeq2SeqLM, # type: ignore[import]
+ )
+
+ model_cls = IPEXModelForSeq2SeqLM
+
+ else:
+ model_cls = (
+ AutoModelForCausalLM
+ if task == "text-generation"
+ else AutoModelForSeq2SeqLM
+ )
+
+ model = model_cls.from_pretrained(model_id, **_model_kwargs)
if tokenizer.pad_token is None:
tokenizer.pad_token_id = model.config.eos_token_id
diff --git a/libs/partners/huggingface/langchain_huggingface/utils/import_utils.py b/libs/partners/huggingface/langchain_huggingface/utils/import_utils.py
new file mode 100644
index 0000000000..fe215c9965
--- /dev/null
+++ b/libs/partners/huggingface/langchain_huggingface/utils/import_utils.py
@@ -0,0 +1,116 @@
+import importlib.metadata
+import importlib.util
+import operator as op
+from typing import Union
+
+from packaging import version
+
+STR_OPERATION_TO_FUNC = {
+ ">": op.gt,
+ ">=": op.ge,
+ "==": op.eq,
+ "!=": op.ne,
+ "<=": op.le,
+ "<": op.lt,
+}
+
+
+_optimum_available = importlib.util.find_spec("optimum") is not None
+_optimum_version = "N/A"
+if _optimum_available:
+ try:
+ _optimum_version = importlib.metadata.version("optimum")
+ except importlib.metadata.PackageNotFoundError:
+ _optimum_available = False
+
+
+_optimum_intel_available = (
+ _optimum_available and importlib.util.find_spec("optimum.intel") is not None
+)
+_optimum_intel_version = "N/A"
+if _optimum_intel_available:
+ try:
+ _optimum_intel_version = importlib.metadata.version("optimum-intel")
+ 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
+
+
+# This function was copied from: https://github.com/huggingface/accelerate/blob/874c4967d94badd24f893064cc3bef45f57cadf7/src/accelerate/utils/versions.py#L319
+def compare_versions(
+ library_or_version: Union[str, version.Version],
+ operation: str,
+ requirement_version: str,
+) -> bool:
+ """
+ Compare a library version to some requirement using a given operation.
+
+ Arguments:
+ library_or_version (`str` or `packaging.version.Version`):
+ A library name or a version to check.
+ operation (`str`):
+ A string representation of an operator, such as `">"` or `"<="`.
+ requirement_version (`str`):
+ The version to compare the library version against
+ """
+ if operation not in STR_OPERATION_TO_FUNC.keys():
+ raise ValueError(
+ f"`operation` must be one of {list(STR_OPERATION_TO_FUNC.keys())}"
+ f", received {operation}"
+ )
+ if isinstance(library_or_version, str):
+ library_or_version = version.parse(
+ importlib.metadata.version(library_or_version)
+ )
+ return STR_OPERATION_TO_FUNC[operation](
+ library_or_version, version.parse(requirement_version)
+ )
+
+
+def is_optimum_available() -> bool:
+ return _optimum_available
+
+
+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
+
+
+def is_optimum_version(operation: str, reference_version: str) -> bool:
+ """
+ Compare the current Optimum version to a given reference with an operation.
+ """
+ if not _optimum_version:
+ return False
+ return compare_versions(
+ version.parse(_optimum_version), operation, reference_version
+ )
+
+
+def is_optimum_intel_version(operation: str, reference_version: str) -> bool:
+ """
+ Compare the current Optimum Intel version to a given reference with an operation.
+ """
+ if not _optimum_intel_version:
+ return False
+ return compare_versions(
+ version.parse(_optimum_intel_version), operation, reference_version
+ )
+
+
+IMPORT_ERROR = """
+requires the {0} library but it was not found in your environment.
+You can install it with pip: `pip install {0}`.
+Please note that you may need to restart your runtime after installation.
+"""