Fix F.sdpa attn drop prob

rwightman · rwightman · commit 935950cc11de · 2024-02-10T20:14:47.000-08:00
diff --git a/timm/models/nextvit.py b/timm/models/nextvit.py
@@ -263,7 +263,7 @@ def forward(self, x):
         if self.fused_attn:
             x = F.scaled_dot_product_attention(
                 q, k, v,
-                dropout_p=self.attn_drop if self.training else 0.,
+                dropout_p=self.attn_drop.p if self.training else 0.,
             )
         else:
             q = q * self.scale

Original file line number	Diff line number	Diff line change
`@@ -263,7 +263,7 @@ def forward(self, x):`
`263`	`263`	`if self.fused_attn:`
`264`	`264`	`x = F.scaled_dot_product_attention(`
`265`	`265`	`q, k, v,`
`266`		`- dropout_p=self.attn_drop if self.training else 0.,`
	`266`	`+ dropout_p=self.attn_drop.p if self.training else 0.,`
`267`	`267`	`)`
`268`	`268`	`else:`
`269`	`269`	`q = q * self.scale`