init

configs/models/qwen2_5_1_5b_radio_sd3_dynamic_puffin.py

@@ -2,7 +2,6 @@ import torch
 from src.models.puffin.model import Qwen2p5RadioStableDiffusion3HFDynamic
 from src.models.stable_diffusion3.transformer_sd3_dynamic import SD3Transformer2DModel
 from src.models.radiov3.hf_model import RADIOModel
-from transformers import AutoConfig
 from diffusers import AutoencoderKL, FlowMatchEulerDiscreteScheduler
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
@@ -41,45 +40,51 @@ model = dict(type=Qwen2p5RadioStableDiffusion3HFDynamic,
                  ),
              transformer=dict(
                  type=SD3Transformer2DModel.from_pretrained,
-                 #pretrained_model_name_or_path=sd3_model_name_or_path,
-                 #subfolder="transformer",
-                 #torch_dtype=torch.bfloat16
+                 pretrained_model_name_or_path=sd3_model_name_or_path,
+                 subfolder="transformer",
+                 torch_dtype=torch.bfloat16,
+                 local_files_only=True,
                  ),
              test_scheduler=dict(
                  type=FlowMatchEulerDiscreteScheduler.from_pretrained,
-                 #pretrained_model_name_or_path=sd3_model_name_or_path,
-                 #subfolder="scheduler"
+                 pretrained_model_name_or_path=sd3_model_name_or_path,
+                 subfolder="scheduler",
+                 local_files_only=True,
                  ),
              train_scheduler=dict(
                  type=FlowMatchEulerDiscreteScheduler.from_pretrained,
-                 #pretrained_model_name_or_path=sd3_model_name_or_path,
-                 #subfolder="scheduler"
+                 pretrained_model_name_or_path=sd3_model_name_or_path,
+                 subfolder="scheduler",
+                 local_files_only=True,
                  ),
              vae=dict(
                  type=AutoencoderKL.from_pretrained,
-                 #pretrained_model_name_or_path=sd3_model_name_or_path,
-                 #subfolder="vae",
-                 #torch_dtype=torch.bfloat16
+                 pretrained_model_name_or_path=sd3_model_name_or_path,
+                 subfolder="vae",
+                 torch_dtype=torch.bfloat16,
+                 local_files_only=True,
                  ),
              freeze_visual_encoder=True,
              freeze_llm=True,
              llm=dict(
                  type=AutoModelForCausalLM.from_pretrained,
-                 #pretrained_model_name_or_path=llm_name_or_path,
-                 #torch_dtype=torch.bfloat16,
+                 pretrained_model_name_or_path=llm_name_or_path,
+                 torch_dtype=torch.bfloat16,
+                 local_files_only=True,
                  #attn_implementation='flash_attention_2',
              ),
              tokenizer=dict(
                  type=AutoTokenizer.from_pretrained,
-                 #pretrained_model_name_or_path=llm_name_or_path
+                 pretrained_model_name_or_path=llm_name_or_path,
+                 local_files_only=True,
                  ),
              prompt_template=prompt_template,
              pretrained_pth=None,
              use_activation_checkpointing=False,
              visual_encoder=dict(
                  type=RADIOModel.from_pretrained,
-                 config=AutoConfig.from_pretrained("nvidia/C-RADIOv3-H"),
-                 #pretrained_model_name_or_path="nvidia/C-RADIOv3-H",
-                 #torch_dtype=torch.bfloat16,
+                 pretrained_model_name_or_path="nvidia/C-RADIOv3-H",
+                 torch_dtype=torch.bfloat16,
+                 local_files_only=True,
                  ),
              )