change convert and use GLMModel instead of GLMForCasualLM

Author: zRzRzRzRzRzRzR

examples/cogview4-control/train_control_cogview4.py

@@ -660,7 +660,7 @@ def prepare_train_dataset(dataset, accelerator):
         [
             transforms.Resize((args.resolution, args.resolution), interpolation=transforms.InterpolationMode.BILINEAR),
             transforms.ToTensor(),
-            transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[0.5, 0.5, 0.5]),
+            transforms.Lambda(lambda x: x * 2 - 1)
         ]
     )
 

scripts/convert_cogview4_to_diffusers_megatron.py

@@ -25,7 +25,7 @@
 
 import torch
 from tqdm import tqdm
-from transformers import GlmForCausalLM, PreTrainedTokenizerFast
+from transformers import GlmModel, PreTrainedTokenizerFast
 
 from diffusers import AutoencoderKL, CogView4Pipeline, CogView4Transformer2DModel, FlowMatchEulerDiscreteScheduler
 from diffusers.loaders.single_file_utils import convert_ldm_vae_checkpoint
@@ -326,7 +326,7 @@ def main(args):
     # Load the text encoder and tokenizer
     text_encoder_id = "THUDM/glm-4-9b-hf"
     tokenizer = PreTrainedTokenizerFast.from_pretrained(text_encoder_id)
-    text_encoder = GlmForCausalLM.from_pretrained(
+    text_encoder = GlmModel.from_pretrained(
         text_encoder_id,
         cache_dir=args.text_encoder_cache_dir,
         torch_dtype=torch.bfloat16 if args.dtype == "bf16" else torch.float32,

src/diffusers/pipelines/cogview4/pipeline_cogview4.py

@@ -215,7 +215,7 @@ def _get_glm_embeds(
             )
             text_input_ids = torch.cat([pad_ids, text_input_ids], dim=1)
         prompt_embeds = self.text_encoder(
-            text_input_ids.to(self.text_encoder.model.device), output_hidden_states=True
+            text_input_ids.to(self.text_encoder.device), output_hidden_states=True
         ).hidden_states[-2]
 
         prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)

src/diffusers/pipelines/cogview4/pipeline_cogview4_control.py

@@ -219,7 +219,7 @@ def _get_glm_embeds(
             )
             text_input_ids = torch.cat([pad_ids, text_input_ids], dim=1)
         prompt_embeds = self.text_encoder(
-            text_input_ids.to(self.text_encoder.model.device), output_hidden_states=True
+            text_input_ids.to(self.text_encoder.device), output_hidden_states=True
         ).hidden_states[-2]
 
         prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)