Fix 3.2 11B inference, by updating padded_collate_tiled_images_and_mask args (#1431)

Jack-Khuu · web-flow · commit b1b32f1f13ca · 2024-12-19T10:16:45.000-08:00
diff --git a/torchchat/generate.py b/torchchat/generate.py
@@ -949,7 +949,7 @@ def _gen_model_input(
 
             if image_found:
                 batch = padded_collate_tiled_images_and_mask(
-                    [data], pad_direction="left", pad_max_images=1
+                    [data], pad_direction="left", pad_max_images=1, pad_max_tiles=transform.max_num_tiles
                 )
                 encoded = batch.pop("tokens").to(device).view(-1)
                 seq_len = encoded.size(0)

Original file line number	Diff line number	Diff line change
`@@ -949,7 +949,7 @@ def _gen_model_input(`
`949`	`949`
`950`	`950`	`if image_found:`
`951`	`951`	`batch = padded_collate_tiled_images_and_mask(`
`952`		`- [data], pad_direction="left", pad_max_images=1`
	`952`	`+ [data], pad_direction="left", pad_max_images=1, pad_max_tiles=transform.max_num_tiles`
`953`	`953`	`)`
`954`	`954`	`encoded = batch.pop("tokens").to(device).view(-1)`
`955`	`955`	`seq_len = encoded.size(0)`