Guard weight_dequant.args[1] access in _quantize_fused_conv_bias pass

ethansfng · web-flow · commit d62addbcdadc · 2026-05-18T19:43:49.000Z
Differential Revision: D105503056 Pull Request resolved: pytorch#19623
diff --git a/backends/transforms/quantize_fused_convbn_bias_pass.py b/backends/transforms/quantize_fused_convbn_bias_pass.py
@@ -238,6 +238,8 @@ def _quantize_fused_conv_bias(
         )
 
         quant_min = -(2**31) + 1 if use_symmetric_quantization else -(2**31)
+        if len(weight_dequant.args) < 2:
+            continue
         if isinstance(weight_dequant.args[1], torch.fx.node.Node):
             weight_scale = get_weight_scale_tensor(weight_dequant.args[1])
             bias_scale = input_dequant.args[1] * weight_scale

Original file line number	Diff line number	Diff line change
`@@ -238,6 +238,8 @@ def _quantize_fused_conv_bias(`
`238`	`238`	`)`
`239`	`239`
`240`	`240`	`quant_min = -(231) + 1 if use_symmetric_quantization else -(231)`
	`241`	`+ if len(weight_dequant.args) < 2:`
	`242`	`+ continue`
`241`	`243`	`if isinstance(weight_dequant.args[1], torch.fx.node.Node):`
`242`	`244`	`weight_scale = get_weight_scale_tensor(weight_dequant.args[1])`
`243`	`245`	`bias_scale = input_dequant.args[1] * weight_scale`