Upload folder using huggingface_hub

models/model_loader.py

@@ -10,9 +10,7 @@ def load_embed_model(model_path: str = "nvidia/llama-nemotron-embed-vl-1b-v2"):
     print(f"🔄 Loading embedding model on {device}...")
     
     config = AutoConfig.from_pretrained(model_path, trust_remote_code=True)
-    config._attn_implementation = "sdpa"
-    if hasattr(config, 'llm_config'):
-        config.llm_config._attn_implementation = "sdpa"
+    # ✅ FIX: Removed SDPA config override which causes issues in HF Spaces
     
     # ✅ FIX: Use manual device instead of device_map="auto"
     model = AutoModel.from_pretrained(
@@ -21,6 +19,7 @@ def load_embed_model(model_path: str = "nvidia/llama-nemotron-embed-vl-1b-v2"):
         torch_dtype=torch.bfloat16 if torch.cuda.is_available() else torch.float32,
         trust_remote_code=True,
         low_cpu_mem_usage=True,  # ✅ CPU optimization
+        attn_implementation="eager", # ✅ FIX: Force eager execution
     ).to(device).eval()
     
     print(f"✅ Embedding model loaded on {device}")