fix: Downgrade transformers to 4.51.3 to fix grounding tool

Problem: MAIRA-2 grounding tool requires transformers <4.52, but we
upgraded to 4.56.0 for NV-Reason-CXR, breaking grounding.

Solution: Prioritize grounding tool (more useful for visualization)
- Downgrade transformers from 4.56.0 → 4.51.3
- Revert tokenizers from >=0.22 → >=0.21,<0.22
- Revert grounding tool to use torch_dtype with transformers 4.51.3
- Disable NV-Reason-CXR tool (incompatible with transformers 4.51.3)

Now 7 tools will work:
✓ Grounding (MAIRA-2) - Now working
✓ Segmentation
✓ Classification
✓ VQA (CheXagent)
✓ Report Generation
✓ DICOM Processor
✓ Web Browser
⊗ NV-Reason-CXR - Disabled (transformers conflict)

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

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

app.py

@@ -38,18 +38,20 @@ tools = []
 if device == "cuda":
     # Load GPU-based tools
 
-    # NV-Reason-CXR - Re-enabled for L40S (48GB VRAM)
-    # Quantization disabled due to dict.to_dict() error - runs in bfloat16 (~7GB)
-    try:
-        from medrax.tools import NVReasonCXRTool
-        nv_reason_tool = NVReasonCXRTool(
-            device=device,
-            load_in_4bit=False  # Disabled - causes BitsAndBytesConfig error
-        )
-        tools.append(nv_reason_tool)
-        print("✓ Loaded NV-Reason-CXR tool")
-    except Exception as e:
-        print(f"✗ Failed to load NV-Reason-CXR tool: {e}")
+    # NV-Reason-CXR - Disabled due to transformers version conflict
+    # Requires transformers 4.56.0, but MAIRA-2 (grounding) requires <4.52
+    # Prioritizing grounding tool for visualization
+    # try:
+    #     from medrax.tools import NVReasonCXRTool
+    #     nv_reason_tool = NVReasonCXRTool(
+    #         device=device,
+    #         load_in_4bit=False
+    #     )
+    #     tools.append(nv_reason_tool)
+    #     print("✓ Loaded NV-Reason-CXR tool")
+    # except Exception as e:
+    #     print(f"✗ Failed to load NV-Reason-CXR tool: {e}")
+    print("⊗ NV-Reason-CXR tool disabled (transformers version conflict)")
 
     # MAIRA-2 Grounding - Re-enabled for L40S (48GB VRAM)
     try:

medrax/tools/grounding.py

@@ -67,13 +67,15 @@ class XRayPhraseGroundingTool(BaseTool):
         super().__init__()
         self.device = torch.device(device) if device else "cuda"
 
-        # Load model - convert to bfloat16 after loading to avoid dtype parameter issues
-        # Load with default dtype, then manually convert to bfloat16
+        # Load model without quantization - works with transformers 4.51.3
         self.model = AutoModelForCausalLM.from_pretrained(
             model_path,
+            device_map=self.device,
             cache_dir=cache_dir,
             trust_remote_code=True,
-        ).to(dtype=torch.bfloat16).eval().to(self.device)
+            torch_dtype=torch.bfloat16,
+        )
+        self.model = self.model.eval()
 
         self.processor = AutoProcessor.from_pretrained(
             model_path,

requirements.txt

@@ -21,9 +21,9 @@ Pillow>=8.0.0
 PyPDF2>=3.0.0
 pdfplumber>=0.10.0
 torchxrayvision>=0.0.37
-transformers==4.56.0
+transformers==4.51.3
 datasets>=2.15.0
-tokenizers>=0.22,<0.24
+tokenizers>=0.21,<0.22
 sentencepiece>=0.1.95
 shortuuid>=1.0.0
 tqdm>=4.64.0