Fix get value from qconfig (#3318)

intel · Oct 21, 2024 · 1c51f3f · 1c51f3f
1 parent bb76079
commit 1c51f3f
Showing 1 changed file with 8 additions and 7 deletions.
diff --git a/intel_extension_for_pytorch/nn/modules/weight_only_quantization.py b/intel_extension_for_pytorch/nn/modules/weight_only_quantization.py
@@ -287,14 +287,15 @@ def from_int4_weight(
         lowp_mode = 2
         act_quant_mode = 1
         cache_weight_for_large_batch = False
-        if qconfig is not None:
-            if hasattr(qconfig, "lowp_mode"):
-                lowp_mode = qconfig.lowp_mode
-            if hasattr(qconfig, "act_quant_mode"):
-                act_quant_mode = qconfig.act_quant_mode
-            if hasattr(qconfig, "cache_weight_for_large_batch"):
+        if qconfig is not None and hasattr(qconfig, "global_qconfig"):
+            if hasattr(qconfig.global_qconfig, "lowp_mode"):
+                lowp_mode = qconfig.global_qconfig.lowp_mode
+            if hasattr(qconfig.global_qconfig, "act_quant_mode"):
+                act_quant_mode = qconfig.global_qconfig.act_quant_mode
+            if hasattr(qconfig.global_qconfig, "cache_weight_for_large_batch"):
                 cache_weight_for_large_batch = (
-                    qconfig.cache_weight_for_large_batch and lowp_mode == 2
+                    qconfig.global_qconfig.cache_weight_for_large_batch
+                    and lowp_mode in (2, 3)
                 )
 
         w_dtype = qweight.dtype