Arm backend: Fix bug in decompose_linear_pass (pytorch#12160)

oscarandersson8218 · Tanish2101 · commit b7cc0951da55 · 2025-07-09T15:15:38.000Z
The output of the decomposition did not get any metadata which is
troublesome when you have two consecutive linear layers.

Signed-off-by: Oscar Andersson &lt;oscar.andersson@arm.com&gt;
diff --git a/backends/arm/_passes/decompose_linear_pass.py b/backends/arm/_passes/decompose_linear_pass.py
@@ -88,6 +88,7 @@ def call(self, graph_module):
                     op_target=exir_ops.edge.aten.view_copy.default,
                     args=(conv, list(output_shape)),
                     kwargs={},
+                    from_node=node,
                 )
 
             node.replace_all_uses_with(output)

Original file line number	Diff line number	Diff line change
`@@ -88,6 +88,7 @@ def call(self, graph_module):`
`88`	`88`	`op_target=exir_ops.edge.aten.view_copy.default,`
`89`	`89`	`args=(conv, list(output_shape)),`
`90`	`90`	`kwargs={},`
	`91`	`+ from_node=node,`
`91`	`92`	`)`
`92`	`93`
`93`	`94`	`node.replace_all_uses_with(output)`