init

configs/models/qwen2_5_1_5b_radio_sd3_dynamic_puffin.py

@@ -6,7 +6,7 @@ from src.models.radiov3.hf_model import RADIOModel
 from diffusers import AutoencoderKL, FlowMatchEulerDiscreteScheduler
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
-llm_name_or_path = 'configs/qwen2.5'
+llm_name_or_path = 'Qwen/Qwen2.5-1.5B-Instruct'
 sd3_model_name_or_path = "configs/sd3"
 radiov3_model_name_or_path = "configs/radiov3"
 
@@ -69,17 +69,14 @@ model = dict(type=Qwen2p5RadioStableDiffusion3HFDynamic,
              freeze_visual_encoder=True,
              freeze_llm=True,
              llm=dict(
-                 type=AutoModelForCausalLM.from_config,
+                 type=AutoModelForCausalLM.from_pretrained,
                  pretrained_model_name_or_path=llm_name_or_path,
                  torch_dtype=torch.bfloat16,
-                 #local_files_only=True,
-                 #attn_implementation='flash_attention_2',
+                 attn_implementation='flash_attention_2',
              ),
              tokenizer=dict(
-                 type=AutoTokenizer.from_config,
-                 pretrained_model_name_or_path=llm_name_or_path,
-                 #local_files_only=True,
-                 ),
+                 type=AutoTokenizer.from_pretrained,
+                 pretrained_model_name_or_path=llm_name_or_path),
              prompt_template=prompt_template,
              pretrained_pth=None,
              use_activation_checkpointing=False,