Update app1.py

app1.py

@@ -33,16 +33,23 @@ pipe.to("cuda")
 # Load FP8 Text Encoder
 # ------------------------
 log("Loading FP8 text encoder...")
+from safetensors.torch import load_file
+import torch
+from transformers import AutoConfig
+from qwen.modeling_qwen import QwenForCausalLM  # ensure Qwen3 model class is imported
+
+# Load config
 config = AutoConfig.from_pretrained("jiangchengchengNLP/qwen3-4b-fp8-scaled")
-fp8_encoder = AutoModel.from_pretrained(
-    "jiangchengchengNLP/qwen3-4b-fp8-scaled",
-    config=config,
-    device_map="auto",
-    torch_dtype=torch.bfloat16,
-)
-pipe.text_encoder = fp8_encoder
-log("FP8 text encoder loaded successfully.")
-log(f"Tokenizer remains: {pipe.tokenizer.__class__.__name__}")
+
+# Initialize model
+fp8_encoder = QwenForCausalLM(config)
+
+# Load weights manually
+weights = load_file("qwen3_4b_fp8_scaled.safetensors")
+fp8_encoder.load_state_dict(weights, strict=False)
+
+fp8_encoder.to("cuda")
+
 
 # ------------------------
 # PIPELINE DEBUG INFO