alexlimh
/

jasper_en_vision_language_v1

sentence-transformers

Model card Files Files and versions Community

alexlimh commited on Jan 16

Commit

b10a706

·

verified ·

1 Parent(s): 2466c58

Update modeling_jasper_vl.py

Files changed (1) hide show

modeling_jasper_vl.py +1 -1

modeling_jasper_vl.py CHANGED Viewed

@@ -1168,7 +1168,7 @@ class JasperVL(PreTrainedModel):
             vit_embeds = self.vision_model(pixel_values=pixel_values, return_dict=True)["last_hidden_state"]
             vit_embeds = self.adaptive_avg_pool2d(vit_embeds)
             selected = (input_ids.reshape(B * N) == self.config.img_token_id)
-            vit_embeds = vit_embeds.to(dtype=inputs_embeds)
             inputs_embeds[selected] = vit_embeds.reshape(-1, C)
             inputs_embeds = inputs_embeds.reshape(B, N, C)
             last_hidden_state = self.model(

             vit_embeds = self.vision_model(pixel_values=pixel_values, return_dict=True)["last_hidden_state"]
             vit_embeds = self.adaptive_avg_pool2d(vit_embeds)
             selected = (input_ids.reshape(B * N) == self.config.img_token_id)
+            vit_embeds = vit_embeds.to(dtype=inputs_embeds.dtype)
             inputs_embeds[selected] = vit_embeds.reshape(-1, C)
             inputs_embeds = inputs_embeds.reshape(B, N, C)
             last_hidden_state = self.model(