Fix float8 noise generation and fix gpu container download cache miss

app.py

@@ -106,11 +106,10 @@ def preload_model_files():
         print(f"Warning: Could not pre-download Unicalli_Pro: {e}")
         local_dir = None
     
-    # 2. T5 text encoder (includes model-00001-of-00002.safetensors, model-00002-of-00002.safetensors)
+    # 2. T5 text encoder
     try:
         snapshot_download(
-            "XLabs-AI/xflux_text_encoders",
-            allow_patterns=["*.safetensors", "*.json", "*.txt", "*.safetensors.index.json"],
+            "xlabs-ai/xflux_text_encoders",
             token=hf_token
         )
         print("✓ T5 text encoder cached")
@@ -121,7 +120,6 @@ def preload_model_files():
     try:
         snapshot_download(
             "openai/clip-vit-large-patch14",
-            allow_patterns=["*.safetensors", "*.json", "*.txt", "*.bin"],
             token=hf_token
         )
         print("✓ CLIP text encoder cached")

src/flux/xflux_pipeline.py

@@ -323,6 +323,9 @@ class XFluxPipeline:
         else:
             # Use model's dtype for efficient inference (fp16/bf16)
             inference_dtype = next(self.model.parameters()).dtype
+            # PyTorch's torch.randn does not support Float8_e4m3fn
+            if getattr(torch, "float8_e4m3fn", None) and inference_dtype == torch.float8_e4m3fn:
+                inference_dtype = torch.bfloat16
         
         print(f"Using {inference_dtype} for inference")