better solution for checking g_idx support

neural_compressor/torch/algorithms/weight_only/modules.py

@@ -26,10 +26,9 @@
 from torch.nn import functional as F
 
 from neural_compressor.torch.utils import (
-    Version,
     accelerator,
     can_pack_with_numba,
-    get_hpex_version,
+    is_hpex_support_g_idx,
     logger,
 )
 
@@ -731,7 +730,7 @@ def __init__(
             )
         else:
             self.g_idx = None
-        self.support_g_idx = True if get_hpex_version() >= Version("1.23.0") else False
+        self.support_g_idx = is_hpex_support_g_idx()
 
         self.half_indim = self.in_features // 2
 

neural_compressor/torch/utils/environ.py

@@ -17,6 +17,7 @@
 import importlib
 import os
 import sys
+from functools import lru_cache
 
 import torch
 from packaging.version import Version
@@ -79,19 +80,20 @@ def is_hpu_available():
     return get_accelerator().name() == "hpu"
 
 
-def get_hpex_version():
-    """Return ipex version if ipex exists."""
+@lru_cache(None)
+def is_hpex_support_g_idx():
+    """Check if HPEX supports group_index in the schema of hpu::convert_from_int4."""
     if is_hpex_available():
         try:
             import habana_frameworks.torch
+            import torch
 
-            hpex_version = habana_frameworks.torch.__version__
-        except ValueError as e:  # pragma: no cover
-            assert False, "Got an unknown version of habana_frameworks.torch: {}".format(e)
-        version = Version(hpex_version)
-        return version
+            schema = torch._C._get_schema("hpu::convert_from_int4", "")
+            return "group_index" in str(schema)
+        except:  # pragma: no cover
+            return False
     else:
-        return None
+        return False
 
 
 ## check optimum