i

modules_forge/forge_clip.py

@@ -1,3 +1,5 @@
+import torch
+
 from modules.sd_hijack_clip import FrozenCLIPEmbedderWithCustomWords
 from ldm_patched.modules import model_management
 
@@ -52,6 +54,8 @@ class CLIP_SD_XL_G(FrozenCLIPEmbedderWithCustomWords):
         else:
             z = outputs.hidden_states[self.wrapped.layer_idx]
 
-        z.pooled = outputs.pooler_output
-
+        pooled_output = outputs.pooler_output
+        text_projection = self.wrapped.text_projection
+        pooled_output = pooled_output.float().to(text_projection.device) @ text_projection.float()
+        z.pooled = pooled_output
         return z

modules_forge/forge_loader.py

@@ -174,6 +174,7 @@ def load_model_for_a1111(timer, checkpoint_info=None, state_dict=None):
             elif typename == 'FrozenOpenCLIPEmbedder2':  # SDXL Clip G
                 embedder.tokenizer = forge_object.clip.tokenizer.clip_g.tokenizer
                 embedder.transformer = forge_object.clip.cond_stage_model.clip_g.transformer
+                embedder.text_projection = forge_object.clip.cond_stage_model.clip_g.text_projection
                 model_embeddings = embedder.transformer.text_model.embeddings
                 model_embeddings.token_embedding = sd_hijack.EmbeddingsWithFixes(
                     model_embeddings.token_embedding, sd_hijack.model_hijack, textual_inversion_key='clip_g')