Fix multiple tool loading issues

- Remove all quantization code from NV-Reason-CXR to avoid dict.to_dict() error
- Fix Classification and Report Generation tools by removing unsupported temp_dir parameter
- Add tensorflow>=2.15.0 to requirements for VQA tool
- NV-Reason-CXR will now run at bfloat16 precision (~6GB memory)

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

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

app.py

@@ -73,7 +73,6 @@ if device == "cuda":
         from medrax.tools.classification import TorchXRayVisionClassifierTool
         classification_tool = TorchXRayVisionClassifierTool(
             device=device,
-            temp_dir="temp",
             load_in_4bit=True
         )
         tools.append(classification_tool)
@@ -85,7 +84,6 @@ if device == "cuda":
         from medrax.tools.report_generation import ChestXRayReportGeneratorTool
         report_tool = ChestXRayReportGeneratorTool(
             device=device,
-            temp_dir="temp",
             load_in_4bit=True
         )
         tools.append(report_tool)

medrax/tools/nv_reason_cxr.py

@@ -66,38 +66,18 @@ class NVReasonCXRTool(BaseTool):
         super().__init__()
         self.device = device
 
-        # Setup quantization config
-        quantization_config = None
-        if load_in_4bit and device == "cuda":
-            try:
-                quantization_config = BitsAndBytesConfig(
-                    load_in_4bit=True,
-                    bnb_4bit_compute_dtype=torch.bfloat16,
-                    bnb_4bit_use_double_quant=True,
-                    bnb_4bit_quant_type="nf4",
-                )
-            except Exception as e:
-                print(f"Warning: Could not setup 4-bit quantization: {e}")
-                quantization_config = None
-
-        # Load model
+        # Load model without quantization to avoid compatibility issues
         try:
             print(f"Loading NV-Reason-CXR model from {model_path}...")
 
-            # Load without quantization config if it's causing issues
-            model_kwargs = {
-                "device_map": self.device,
-                "cache_dir": cache_dir,
-                "torch_dtype": torch.bfloat16,
-                "trust_remote_code": True,
-            }
-
-            if quantization_config is not None:
-                model_kwargs["quantization_config"] = quantization_config
-
+            # Note: Skipping quantization due to dict.to_dict() error
+            # Model will run at bfloat16 precision (~6GB memory)
             self.model = AutoModelForImageTextToText.from_pretrained(
                 model_path,
-                **model_kwargs
+                device_map=self.device,
+                cache_dir=cache_dir,
+                torch_dtype=torch.bfloat16,
+                trust_remote_code=True,
             ).eval()
 
             self.processor = AutoProcessor.from_pretrained(

requirements.txt

@@ -47,6 +47,7 @@ openai>=0.27.0
 backoff>=1.10.0
 torch>=2.2.0
 torchvision>=0.10.0
+tensorflow>=2.15.0
 scikit-image>=0.18.0
 opencv-python>=4.8.0
 matplotlib>=3.8.0