Stick trace tensor back in to prevent bad fusion.

monorimet · monorimet · commit d76426d0d411 · 2024-06-13T05:44:20.000-05:00
diff --git a/src/diffusers/models/attention_processor.py b/src/diffusers/models/attention_processor.py
@@ -1152,7 +1152,7 @@ def __call__(
             hidden_states[:, residual.shape[1] :],
         )
         hidden_states_cl = hidden_states.clone()
-        #trace_tensor("attn_out", hidden_states_cl[0,0,0])
+        trace_tensor("attn_out", hidden_states_cl[0,0,0])
         # linear proj
         hidden_states = attn.to_out[0](hidden_states_cl)
         # dropout

Original file line number	Diff line number	Diff line change
`@@ -1152,7 +1152,7 @@ def __call__(`
`1152`	`1152`	`hidden_states[:, residual.shape[1] :],`
`1153`	`1153`	`)`
`1154`	`1154`	`hidden_states_cl = hidden_states.clone()`
`1155`		`- #trace_tensor("attn_out", hidden_states_cl[0,0,0])`
	`1155`	`+ trace_tensor("attn_out", hidden_states_cl[0,0,0])`
`1156`	`1156`	`# linear proj`
`1157`	`1157`	`hidden_states = attn.to_out[0](hidden_states_cl)`
`1158`	`1158`	`# dropout`