move get_gpu_family to is_gpu_available (#1479)

intel_extension_for_transformers/tools/utils.py

@@ -34,34 +34,10 @@
 def supported_gpus():
     return ['flex', 'max', 'arc']
 
-def get_gpu_family():
-    """Get gpu device family info.
-
-    Return 'flex'|'max'|'arc'| 'no_gpu'| assert
-
-    Note, this function need to import intel_extension_for_pytorch
-
-
-    Additional info (common gpu name):
-      'Intel(R) Data Center GPU Flex 170'
-      'Intel(R) Data Center GPU Max 1100'
-      'Intel(R) Arc(TM) A770 Graphics'
-    """
-
+def is_intel_gpu_available():
     import intel_extension_for_pytorch as ipex
-    if not (hasattr(torch, "xpu") and torch.xpu.is_available()):
-        return 'no_gpu'
+    return hasattr(torch, "xpu") and torch.xpu.is_available()
 
-    name = torch.xpu.get_device_name()
-    if torch.xpu.has_xmx() and torch.xpu.has_2d_block_array():
-        return "PVC"
-    elif torch.xpu.has_xmx() and not torch.xpu.has_2d_block_array():
-        return "ARC"
-    elif not torch.xpu.has_xmx() and not torch.xpu.has_2d_block_array():
-        return "MTL"
-    else:
-        warnings.warn("Unverified GPU device: {}".format(name))
-        return name
 
 _ipex_available = importlib.util.find_spec("intel_extension_for_pytorch") is not None
 _ipex_version = "N/A"

intel_extension_for_transformers/transformers/modeling/modeling_auto.py

@@ -67,7 +67,7 @@
     convert_to_quantized_model,
     replace_linear,
 )
-from ...tools.utils import get_gpu_family, is_ipex_available
+from ...tools.utils import is_intel_gpu_available, is_ipex_available
 from accelerate import init_empty_weights
 from huggingface_hub import hf_hub_download
 from neural_compressor.adaptor.torch_utils.model_wrapper import WeightOnlyLinear
@@ -84,7 +84,7 @@
 
 from typing import Union
 
-if is_ipex_available() and get_gpu_family() != "no_gpu":
+if is_ipex_available() and is_intel_gpu_available():
     # pylint: disable=E0401
     from intel_extension_for_pytorch.nn.utils._quantize_convert import (
         WeightOnlyQuantizedLinear,

setup.py

@@ -11,7 +11,7 @@
 
 result = subprocess.Popen("pip install -r requirements.txt", shell=True)
 result.wait()
-from intel_extension_for_transformers.tools.utils import get_gpu_family
+from intel_extension_for_transformers.tools.utils import is_intel_gpu_available
 
 def check_env_flag(name: str, default: bool = False) -> bool:
     if default:  # if a flag meant to be true if not set / mal-formatted
@@ -29,7 +29,7 @@ def check_env_flag(name: str, default: bool = False) -> bool:
 ipex_available = importlib.util.find_spec(
     "intel_extension_for_pytorch") is not None
 IS_INTEL_GPU = False
-if ipex_available and (get_gpu_family() != "no_gpu"):
+if ipex_available and is_intel_gpu_available():
     SKIP_RUNTIME = True
     RUNTIME_ONLY = False
     IS_INTEL_GPU = True

tests/CI/test_weight_only_gpu.py

@@ -24,14 +24,10 @@
 from intel_extension_for_transformers.transformers import GPTQConfig, RtnConfig
 from math import isclose
 from transformers import AutoTokenizer
-from intel_extension_for_transformers.tools.utils import get_gpu_family, is_ipex_available
+from intel_extension_for_transformers.tools.utils import is_intel_gpu_available, is_ipex_available
 from torch.utils.data import DataLoader
 
 
-if is_ipex_available():
-    gpu_name = get_gpu_family()
-
-
 MODEL_NAME ="hf-internal-testing/tiny-random-gptj"
 
 
@@ -73,7 +69,7 @@ def forward(self, x):
         return self.linear(x)
 
 
-@unittest.skipIf(not is_ipex_available() or gpu_name == "no_gpu",
+@unittest.skipIf(not is_ipex_available() or not is_intel_gpu_available(),
     "There is no Intel GPU in this machine, skip this test!")
 class TestArcWeightOnly(unittest.TestCase):