fix: Add fallback for model loading with better error handling

**Model Loading Improvements:**
- Explicitly set use_safetensors=True for primary loading attempt
- Add try-except fallback to default loading if safetensors fails
- Keep torch_dtype=torch.float32 (dtype not supported in transformers 4.30)
- Better error messages for debugging

**Error Handling:**
- Primary: Try loading with use_safetensors=True
- Fallback: Try loading without use_safetensors if primary fails
- Print warning message when fallback is used
- Prevents complete failure when safetensors has issues

This should fix the model loading error on HF Spaces while
maintaining compatibility with both safetensors and legacy formats.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

Co-Authored-By: Claude <noreply@anthropic.com>

app.py

@@ -82,13 +82,26 @@ def load_model(model_name):
                 tokenizer.pad_token = tokenizer.eos_token
 
             # Load model with safetensors support
-            model = AutoModelForCausalLM.from_pretrained(
-                model_name,
-                token=HF_TOKEN,
-                dtype=torch.float32,
-                low_cpu_mem_usage=True,
-                trust_remote_code=True
-            )
+            try:
+                model = AutoModelForCausalLM.from_pretrained(
+                    model_name,
+                    token=HF_TOKEN,
+                    torch_dtype=torch.float32,
+                    low_cpu_mem_usage=True,
+                    trust_remote_code=True,
+                    use_safetensors=True
+                )
+            except Exception as e:
+                # Fallback to default loading if safetensors fails
+                print(f"⚠️ Safetensors loading failed, trying default method: {e}")
+                model = AutoModelForCausalLM.from_pretrained(
+                    model_name,
+                    token=HF_TOKEN,
+                    torch_dtype=torch.float32,
+                    low_cpu_mem_usage=True,
+                    trust_remote_code=True
+                )
+
             model.to(device)
             model.eval()