Update modeling_text_encoder.py

modeling_text_encoder.py

@@ -5,13 +5,12 @@ from transformers import (
     CLIPTextModelWithProjection,
     CLIPTokenizer,
     T5EncoderModel,
-    T5TokenizerFast,
-    BitsAndBytesConfig  # Import for 8-bit quantization
+    T5TokenizerFast
 )
 from typing import Union, List, Optional
 
 class SD3TextEncoderWithMask(nn.Module):
-    def __init__(self, model_path, torch_dtype):
+    def __init__(self, model_path, torch_dtype=torch.float16):
         super().__init__()
 
         # Tokenizers for CLIP and T5
@@ -27,9 +26,6 @@ class SD3TextEncoderWithMask(nn.Module):
         self.torch_dtype = torch_dtype
         self.tokenizer_max_length = self.tokenizer.model_max_length
 
-        # Quantization config for T5 model
-        self.quantization_config = BitsAndBytesConfig(load_in_8bit=True)  # Quantize T5 to 8-bit
-
         # Freeze parameters to avoid unnecessary gradient computation
         self._freeze()
 
@@ -51,12 +47,11 @@ class SD3TextEncoderWithMask(nn.Module):
             ).to("cuda")
 
         if self.text_encoder_3 is None:
-            # Load the T5 model with 8-bit quantization
+            # Load the FP8 T5 model (adjust path if needed for specific FP8 model)
             self.text_encoder_3 = T5EncoderModel.from_pretrained(
-                os.path.join(self.model_path, 'text_encoder_3'),
-                torch_dtype=self.torch_dtype,
-                quantization_config=self.quantization_config  # Apply quantization
-            )  # Do NOT use .to("cuda") for 8-bit quantized models, as they handle device placement automatically
+                os.path.join(self.model_path, 'text_encoder_3_fp8'),  # Use the FP8 version
+                torch_dtype=torch.float8  # FP8 precision
+            ).to("cuda")
 
     def _get_t5_prompt_embeds(
         self,
@@ -65,7 +60,7 @@ class SD3TextEncoderWithMask(nn.Module):
         device: Optional[torch.device] = None,
         max_sequence_length: int = 128,
     ):
-        """ Get embeddings from the T5 model. """
+        """ Get embeddings from T5 model. """
         self._load_models_if_needed()  # Lazy loading
         prompt = [prompt] if isinstance(prompt, str) else prompt
         text_inputs = self.tokenizer_3(
@@ -80,7 +75,7 @@ class SD3TextEncoderWithMask(nn.Module):
         text_input_ids = text_inputs.input_ids.to(device)
         prompt_attention_mask = text_inputs.attention_mask.to(device)
 
-        # Use the T5 model to generate embeddings (quantized)
+        # Use the T5 model to generate embeddings in FP8
         prompt_embeds = self.text_encoder_3(text_input_ids, attention_mask=prompt_attention_mask)[0]
         prompt_embeds = prompt_embeds.to(dtype=self.text_encoder_3.dtype)  # Ensure correct dtype
 
@@ -116,7 +111,7 @@ class SD3TextEncoderWithMask(nn.Module):
             return_tensors="pt",
         )
 
-        text_input_ids = text_inputs.input_ids.to(device)
+        text_input_ids = text_inputs.input.ids.to(device)
         prompt_embeds = text_encoder(text_input_ids, output_hidden_states=True)[0]
 
         # Duplicate embeddings for each image generation
@@ -138,7 +133,7 @@ class SD3TextEncoderWithMask(nn.Module):
         pooled_prompt_2_embed = self._get_clip_prompt_embeds(prompt, num_images_per_prompt=num_images_per_prompt, device=device, clip_model_index=1)
         pooled_prompt_embeds = torch.cat([pooled_prompt_embed, pooled_prompt_2_embed], dim=-1)
 
-        # Get T5 embeddings
+        # Get T5 embeddings in FP8
         prompt_embeds, prompt_attention_mask = self._get_t5_prompt_embeds(prompt, num_images_per_prompt=num_images_per_prompt, device=device)
 
         return prompt_embeds, prompt_attention_mask, pooled_prompt_embeds