workaround for v1.20 missing attribution

Signed-off-by: Xin He <[email protected]>

Author: xinhe3

neural_compressor/torch/algorithms/fp8_quant/_core/utils.py

@@ -91,7 +91,7 @@ def quantize_dynamic_op(config, mod_type):
 
 
 def set_runtime_scale_patching_mode(scaling_method_name):
-    if is_runtime_scale_patching():
+    if is_runtime_scale_patching() and hasattr(htexp, "_set_scale_attributes"):
         assert (
             scaling_method_name in runtime_scale_patching_supported_methods_list
         ), f"Scaling method \"{scaling_method_name}\" is not supported for runtime scale patching (graph recompile reduction). Cannot set scaling attributes."

test/3x/torch/algorithms/fp8_quant/unit_tests/test_runtime_scale_patching.py

@@ -4,6 +4,7 @@
 import shutil
 import copy
 import habana_frameworks.torch.core as htcore
+import habana_frameworks.torch.utils.experimental as htexp
 
 from ..tester import RUNTIME_SCALE_PATCHING_SUPPORTED_METHODS_LIST, SCALE_METHODS_KEY_ERROR, run_with_raised_exception
 from neural_compressor.torch.algorithms.fp8_quant._core.common import is_runtime_scale_patching
@@ -47,6 +48,7 @@ def temp_directory():
     shutil.rmtree(temp_dir)
 
 
+@pytest.mark.skipif(not hasattr(htexp, "_set_scale_attributes"), reason="scale attributes not supported")
 @pytest.mark.parametrize("scale_method", ScaleMethod)
 @pytest.mark.parametrize("scale_format", ["SCALAR", "CONST"])
 @pytest.mark.parametrize("dynamic_scale_patching", [True, False])