Spaces:
Paused
Paused
Commit History
-
fix: Display generated X-ray images in visualization panel
Updated visualization detection to include generated X-ray images alongside grounding and segmentation outputs. Now all three visualization types are properly displayed in the UI. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>39784b3 -
feat: Filter tools based on mode for Socratic teaching
Problem: Socratic mode should guide learning through questions, not give direct answers. Tools like segmentation and report generation defeat the purpose by providing complete answers. Solution: Filter tools based on agent mode - Assistant mode: All 8 tools available - Socratic mode: Only 5 teaching-appropriate tools Socratic Tools (5): ✓ Grounding - "Show me where you see that" ✓ VQA - Ask targeted questions about specific features ✓ Classification - Show probabilities to guide thinking ✓ Generation - Create practice cases ✓ DICOM Processor - Utility only Excluded from Socratic (3): ✗ Segmentation - Draws answers automatically ✗ Report Generation - Gives complete diagnosis ✗ Web Browser - Not relevant to case analysis This ensures Socratic mode truly guides discovery vs giving answers. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>9b867b2 -
feat: Enable RoentGen-v2 synthetic X-ray generation
Added RoentGen-v2 (Stanford MIMI) for generating synthetic chest X-rays from text descriptions. Changes: - Updated xray_generation.py to use stanfordmimi/RoentGen-v2 from HF - Changed from fp32 to fp16 for memory efficiency - Added tool loading in app.py for GPU environments - Now publicly available (no need for manual weight download) Usage examples: - "Generate an X-ray showing pneumonia in the right lower lobe" - "Create a chest X-ray with left-sided pleural effusion" - "Generate a normal chest radiograph" Tool count: 7 → 8 (with RoentGen-v2 generation) Paper: https://arxiv.org/abs/2508.16783 Model: https://huggingface.co/stanfordmimi/RoentGen-v2 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>8dd90e9
-
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>37bdbfa -
fix: Load grounding model without dtype parameter to avoid errors
Changed from using dtype parameter in from_pretrained to manually converting dtype after loading. This avoids potential JSON serialization and compatibility issues with the dtype parameter. Pattern: load() → .to(dtype=bfloat16) → .eval() → .to(device) 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>e7a5afc -
fix: Update grounding tool for transformers 4.56.0 and fix DICOM segmentation
Two critical fixes: 1. Grounding Tool (MAIRA-2): - Updated to use 'dtype' instead of deprecated 'torch_dtype' - Removed device_map parameter causing compatibility issues - Use .eval().to(device) pattern matching transformers 4.56.0 - Fixes: "'Maira2ForConditionalGeneration' object has no attribute 'model'" 2. Segmentation Tool: - Don't invert processed DICOM files (temp/processed_dicom_*.png) - These files are already in correct orientation from DICOM processor - Only invert regular PNG/JPEG X-rays (which have inverted intensities) - Fixes poor segmentation accuracy for DICOM images Both tools now work correctly with transformers 4.56.0. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>99d91bc
-
fix: Don't invert processed DICOM files for segmentation
Problem: DICOM files processed by the DICOM tool are saved as PNG in temp/processed_dicom_*.png, but were being treated as regular PNG files and incorrectly inverted, causing poor segmentation accuracy. Solution: Detect processed DICOM files by checking if "processed_dicom" is in the filename, and skip inversion for these files while still converting to HU range. Image handling now: - Raw PNG/JPEG: Invert + HU conversion (lungs bright → dark) - Processed DICOM PNG: HU conversion only (already correct orientation) - Raw DICOM: Keep original HU values This fixes segmentation accuracy for DICOM images. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>8bb9754 -
fix: Persist uploaded image across interactions to keep visualization visible
Problem: Visualization panel was empty after initial analysis because the uploaded image path wasn't persisted across multiple chat interactions. Solution: - Added gr.State to track current_image across interactions - When no visualization is generated (e.g., grounding fails), show the uploaded image as a fallback reference - Image now stays visible throughout the conversation until replaced This ensures users can always see their uploaded X-ray as reference. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>8edb8da -
fix: Update tokenizers version to be compatible with transformers 4.56.0
Transformers 4.56.0 requires tokenizers>=0.22.0, but we had <0.22. Updated tokenizers constraint from >=0.21,<0.22 to >=0.22,<0.24. This resolves the dependency conflict that was preventing the build. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>be1ba0f -
fix: Upgrade transformers to 4.56.0 to fix NV-Reason-CXR dtype error
Root cause: We were using transformers 4.51.3, but NVIDIA's official demo uses 4.56.0. The older version has a bug where the 'dtype' parameter causes JSON serialization errors. Solution: - Upgrade transformers from 4.51.3 to 4.56.0 (matches NVIDIA demo) - Use NVIDIA's EXACT loading code with dtype=torch.bfloat16 - This is the same version and code that works in NVIDIA's official Space This should finally resolve the persistent NV-Reason-CXR loading error. Reference: https://huggingface.co/spaces/nvidia/nv-reason-cxr 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>6a7c30f -
fix: Load NV-Reason-CXR without dtype parameter to avoid JSON error
The 'dtype' parameter in from_pretrained() causes JSON serialization errors in our environment. Instead, load the model with default settings, then manually convert to bfloat16 after loading. This approach achieves the same result as NVIDIA's demo (bfloat16 model) but avoids the JSON serialization TypeError. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>7a6a9a6 -
fix: Use 'dtype' instead of deprecated 'torch_dtype' for NV-Reason-CXR
Based on NVIDIA's official Gradio demo code, the correct parameter is 'dtype' (not 'torch_dtype' which is deprecated and causes the dict error). Changes: - Use 'dtype=torch.bfloat16' (matches NVIDIA demo exactly) - Use 'pretrained_model_name_or_path' parameter name - Add 'use_fast=True' to processor loading - Switch from float16 to bfloat16 (as per NVIDIA's demo) This should resolve the persistent 'dict' object has no attribute 'to_dict' error. Reference: https://huggingface.co/spaces/nvidia/nv-reason-cxr/blob/main/app.py 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>f49ba8b -
feat: Display uploaded X-ray image as reference before visualization
When users upload an X-ray image, immediately display it in the visualization panel as a reference. Once segmentation or grounding is complete, the visualization replaces the original image. Flow: 1. User uploads X-ray → Shows original image on right 2. User requests segmentation → Original image visible as reference 3. Segmentation completes → Overlay replaces original image This provides visual context while the agent processes the request. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>b2960ee -
fix: Handle empty assistant responses when visualizations are generated
When segmentation/grounding tools run, Gemini sometimes returns empty responses, causing blank chat bubbles in the UI. Solution: Provide helpful default messages when: - Segmentation visualization is created - Grounding visualization is created - Any other empty response occurs This ensures users always see a message explaining what happened. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>d2991ca -
fix: Invert PNG/JPEG X-rays before HU conversion for accurate segmentation
Problem: PNG/JPEG X-rays have inverted intensities compared to DICOM - DICOM: lungs = dark (low HU), bones = bright (high HU) - PNG/JPEG: lungs = bright (high pixels), bones = dark (low pixels) Solution: Invert PNG/JPEG images (255 - pixel_value) before converting to Hounsfield Units. This ensures consistent input for TorchXRayVision. Result: PNG/JPEG segmentation accuracy now matches DICOM accuracy. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>85e170b -
refactor: Remove FastAPI integration, keep Gradio only
Removed all FastAPI endpoints and dependencies: - Removed FastAPI, uvicorn, pydantic imports - Removed MedGemma API model loading - Removed /api/medgemma/* endpoints - Simplified to pure Gradio interface - Back to demo.launch() instead of uvicorn This simplifies deployment and focuses on the interactive Gradio UI. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>0b51074 -
debug: Add verbose logging to NV-Reason-CXR loading
Added detailed logging and traceback to help diagnose the persistent dict.to_dict() error on Hugging Face Spaces deployment. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>e88cf5f -
fix: Preserve FastAPI routes when mounting Gradio
Previously, the FastAPI app variable was being overwritten when mounting Gradio, causing all API routes (/docs, /api/medgemma/*) to return 404. Now using combined_app variable to preserve both FastAPI routes and Gradio UI. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>3e5432f -
fix: Match NVIDIA's official NV-Reason-CXR loading approach
- Updated NV-Reason-CXR to use torch.float16 (not bfloat16) - Removed device_map parameter causing dict.to_dict() error - Simplified to NVIDIA's exact pattern: .eval().to(device) - Removed unnecessary parameters (trust_remote_code, low_cpu_mem_usage, use_fast) - Added missing Optional import for FastAPI endpoints This resolves the persistent 'dict' object has no attribute 'to_dict' error by following NVIDIA's official documentation exactly. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>c1c4dbf -
fix: Disable NV-Reason-CXR quantization in app.py to resolve dict.to_dict() error
The nv_reason_cxr.py tool was fixed, but app.py was still passing load_in_4bit=True which triggered the BitsAndBytesConfig error. Changed: - load_in_4bit=True → load_in_4bit=False This matches the fix in nv_reason_cxr.py and grounding.py. Model will run in bfloat16 precision (~7GB VRAM). Now NV-Reason-CXR should load successfully! Co-Authored-By: Claude <noreply@anthropic.com>223042e -
feat: Add FastAPI to expose MedGemma via REST API
Added FastAPI alongside Gradio for direct model access: New Features: - FastAPI mounted on same port (7860) as Gradio - Direct MedGemma model access without agent overhead - Two API endpoints: 1. POST /api/medgemma/analyze - JSON with base64 image 2. POST /api/medgemma/upload - Multipart file upload - Auto-generated docs at /docs - Separate MedGemma instance for API (independent from agent) Architecture: ┌─────────────────────────────────────┐ │ Port 7860 (Unified Server) │ ├─────────────────────────────────────┤ │ / → Gradio UI │ │ /api/* → FastAPI endpoints │ │ /docs → Interactive API docs │ └─────────────────────────────────────┘ Example API Usage: ```bash # Analyze with base64 curl -X POST "http://localhost:7860/api/medgemma/analyze" \ -H "Content-Type: application/json" \ -d '{"query": "What do you see?", "image_base64": "..."}' # Upload file curl -X POST "http://localhost:7860/api/medgemma/upload" \ -F "file=@xray.jpg…3a26781 -
fix: Remove quantization parameters from NV-Reason-CXR to fix dict.to_dict() error
Same BitsAndBytesConfig issue as grounding tool - passing quantization_config=None or load_in_4bit=False still triggers the 'dict' object has no attribute 'to_dict' error. Solution: Remove ALL quantization-related parameters entirely - Removed: quantization_config=None - Removed: load_in_4bit=False - Removed: load_in_8bit=False - Added: low_cpu_mem_usage=True (for efficiency) Model will now run in bfloat16 (~7GB VRAM) without quantization. This matches the fix applied to grounding.py that resolved the issue. Co-Authored-By: Claude <noreply@anthropic.com>90107a8
-
refactor: Integrate message box into chat interface
Made the message input feel like part of the chat box: Changes: - Reduced chatbot height from 600px to 520px - Removed labels from both chatbot and message box - Set container=False on message box to reduce gaps - Message box now feels integrated at bottom of chat Before: ┌─────────────────┐ │ Chatbot │ │ (separate) │ └─────────────────┘ ┌─────────────────┐ │ Message Box │ └─────────────────┘ After: ┌─────────────────┐ │ Chatbot │ │ │ ├─────────────────┤ │ Message Box │ ← Feels unified └─────────────────┘ This creates a WhatsApp-like unified chat interface. Co-Authored-By: Claude <noreply@anthropic.com>30b53f2 -
feat: Side-by-side layout with chat and visualization
Changed from stacked to side-by-side layout for better UX: Before (stacked): - Chat messages on top - Visualization below - Had to scroll to see both After (side-by-side): - Chat on left (2/3 width) - Visualization on right (1/3 width) - See both simultaneously - Fixed height of 600px for both Layout structure: ┌─────────────────────────┬──────────────┐ │ Chat Messages │ Visualization│ │ (scale=2) │ (scale=1) │ │ │ │ │ Message Input │ │ └─────────────────────────┴──────────────┘ Benefits: - Better use of screen space - View segmentation/grounding while chatting - More professional clinical interface - Similar to PACS viewer layouts Co-Authored-By: Claude <noreply@anthropic.com>227178d -
fix: Display proper WhatsApp-style chat conversation flow
Fixed chat history to show back-and-forth conversation instead of combining all responses into one bubble. Before: - Only assistant responses shown - All responses merged into single bubble - No user messages visible in chat After: - Both user and assistant messages shown - Each message in separate bubble - Proper conversation flow like WhatsApp/messaging apps Changes: - Manually add user message to chat_history (not auto-added by Gradio) - Extract text from multimodal messages properly - Add [Image uploaded] indicator when files attached - Append both user and assistant messages in sequence Now shows: User: What do you see? Assistant: I see several findings... User: Show me where Assistant: Here's the location... Co-Authored-By: Claude <noreply@anthropic.com>4167f4b -
refactor: Simplify segmentation to single overlay panel
Changed from 3-panel debug view to clean single overlay: - Removed: Original image panel (left) - Removed: Pure mask panel (middle) - Kept: Overlay panel with contours and legend Benefits: - Cleaner, more professional appearance - Focuses on the actual result (overlay) - Larger visualization size - Better for production use The 3-panel view was useful for debugging "no masks detected" but now that segmentation works, we only need the overlay. Co-Authored-By: Claude <noreply@anthropic.com>0e36bfb -
fix: Use correct Hounsfield Unit range for torchxrayvision segmentation
CRITICAL FIX: The torchxrayvision PSPNet model expects images in Hounsfield Unit range [-1024, 1024], NOT normalized [0, 1] range! The warning message revealed this: "Input image has the range [0.22,0.91] which doesn't seem to be in the [-1024,1024] range" This caused probabilities of 0.004 instead of proper detection. Changes: - For 8-bit images (0-255): Scale to [-1024, 600] (lung window in HU) - For DICOM/16-bit: Keep original HU range - Removed the incorrect preprocess_medical_image() normalization Formula: (pixel / 255.0) * 1624 - 1024 Maps: 0 → -1024 (air), 255 → 600 (bone/tissue) This should dramatically increase segmentation probabilities! Co-Authored-By: Claude <noreply@anthropic.com>f6108da -
debug: Add comprehensive logging for segmentation pipeline
Added extensive debug output to diagnose "No Masks Detected" issue: - Image loading stats (shape, dtype, range) - Preprocessing output stats - Transform effects - Final tensor stats (mean, std, min, max) - Raw model predictions before sigmoid - Probabilities after sigmoid (both mean and max per organ) - Pixel counts after thresholding This will show us exactly where the pipeline is failing: - If probabilities are all < 0.3, we need to lower threshold further - If probabilities are reasonable but masks are empty, it's a thresholding issue - If model output is all zeros/very low, it's a model loading or input format problem Co-Authored-By: Claude <noreply@anthropic.com>940e3ae -
fix: Add 3-panel diagnostic visualization for segmentation
Created side-by-side visualization to diagnose mask detection issues: - Left panel: Original X-ray - Middle panel: Pure mask visualization (shows if model detects anything) - Right panel: Overlay with contours If middle panel shows "No Masks Detected" in red, the model isn't finding organs at the 0.3 threshold. If middle panel has colors but right panel doesn't show overlay, it's a rendering issue. This diagnostic view will definitively show what's failing. Co-Authored-By: Claude <noreply@anthropic.com>6d677d9 -
fix: Lower segmentation threshold and add probability debugging
Critical fix: - Lowered segmentation threshold from 0.5 to 0.3 for non-DICOM images - The model is less confident on JPG/PNG vs DICOM, causing all masks to be filtered out - Added probability debugging output to show actual confidence scores for each organ This fixes the issue where segmentation visualization only showed the original image without any overlays. The 0.5 threshold was too strict for compressed image formats. Debug output now shows: === Segmentation Probabilities === Left Lung: 0.427 Right Lung: 0.438 ... ================================== This helps diagnose when masks aren't being detected. Co-Authored-By: Claude <noreply@anthropic.com>7f44ee5
-
fix: Enhanced segmentation overlay with contours and better alignment
Improvements: - Added contour line drawing for clear organ boundaries - Implemented alignment quality check - falls back to simple resize if >90% of mask is lost - Switched to tab10 colormap for more distinct colors - Used numpy masked arrays for cleaner overlay rendering - Added matplotlib Patch objects for proper legend rendering - Improved debug logging to track alignment issues This should make segmentation overlays much more visible, especially when the alignment algorithm is losing mask data during coordinate transformation. Co-Authored-By: Claude <noreply@anthropic.com>89eb957 -
fix: Improve segmentation overlay and add DICOM support
Segmentation improvements: - Improved matplotlib overlay rendering with better color opacity - Added debug logging to track mask detection and alignment - Changed to subplot-based rendering for better overlay composition - Increased overlay opacity from 30% to 40% for better visibility - Added mask count tracking to verify segmentation success DICOM file support: - Added DICOM file extensions to Gradio file upload (.dcm, .dicom) - Updated placeholder text to indicate DICOM support - Added DICOM file detection in chat function - DICOM files are passed to agent for processing with DICOM tool - Added error traceback printing for better debugging The segmentation overlay should now properly show colored masks on top of the X-ray image, and DICOM files can be uploaded without errors. Co-Authored-By: Claude <noreply@anthropic.com>3a18164 -
fix: Enable Gemini vision support for image analysis
- Added PIL, base64, and BytesIO imports for image processing - Updated chat function to encode images as base64 for Gemini - Images are now passed as multimodal content to support Gemini 2.0 Flash vision - Resize images larger than 4096x4096 to meet Gemini limits - Include image path in message text for tool access - This fixes the issue where tools couldn't access uploaded images Now both Assistant and Socratic modes can properly analyze X-ray images and invoke tools like grounding and segmentation. Co-Authored-By: Claude <noreply@anthropic.com>9dc78d7 -
feat: Add Socratic AI Tutor mode with toggle
- Added SOCRATIC_TUTOR prompt to system_prompts.txt with comprehensive teaching guidelines - Modified Agent class to accept 'mode' parameter (assistant/socratic) - Added UI toggle in app.py to switch between Assistant and Tutor modes - Implemented dynamic agent creation based on selected mode - Created test script to validate both modes The Socratic mode guides learners through medical reasoning with questions instead of providing direct answers, while Assistant mode maintains the original direct response behavior. Co-Authored-By: Claude <noreply@anthropic.com>11b0dac
-
Display segmentation visualizations in UI
Fixed chat() function to also check for segmentation images alongside grounding visualizations. Changes: - Added glob pattern for temp/segmentation_*.png files - Updated UI label to "Visualization (Grounding/Segmentation)" - Now displays most recent visualization from either tool This allows segmentation overlays to appear in the UI just like grounding bounding boxes. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>7d28f28 -
Fix NV-Reason-CXR quantization config error
Explicitly disable quantization parameters to prevent 'dict' object has no attribute 'to_dict' error from embedded model config. Added explicit parameters: - quantization_config=None - load_in_4bit=False - load_in_8bit=False Model will run at bfloat16 precision (~6-7GB VRAM). 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>f1877c8 -
Add ChestXRaySegmentationTool for anatomical structure segmentation
Enables pixel-perfect segmentation of 14 anatomical structures: - Left/Right Clavicle, Scapula, Lung, Hilus Pulmonis - Heart, Aorta, Diaphragm, Mediastinum, Weasand, Spine Features: - Automatic segmentation of all organs - Detailed metrics (area, centroid, bbox, size, intensity) - Segmentation visualization overlays - Uses TorchXRayVision PSPNet model This completes MedRAX2 with 8 tools total: 1. NV-Reason-CXR - Clinical reasoning 2. MAIRA-2 Grounding - Phrase grounding with bboxes 3. CheXagent VQA - Interactive Q&A 4. Classification - Pathology detection 5. Report Generation - Radiology reports 6. Segmentation - Anatomical structure segmentation (NEW) 7. DICOM Processor - File handling 8. Web Browsing - Medical information 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>f372684 -
Re-enable NV-Reason-CXR for L40S GPU (48GB VRAM)
With L40S, we now have enough VRAM for all 7 tools: - NV-Reason-CXR: ~7GB (with 4-bit quantization) - MAIRA-2 Grounding: ~22GB - CheXagent VQA: ~2-3GB - Classification + Report Gen: ~2GB - Total: ~33-34GB out of 48GB (14GB free) This completes the full MedRAX2 medical AI assistant with: 1. NV-Reason-CXR - Clinical reasoning 2. MAIRA-2 Grounding - Visual phrase grounding 3. CheXagent VQA - Interactive Q&A 4. Classification - Pathology detection 5. Report Generation - Radiology reports 6. DICOM Processor - File handling 7. Web Browsing - Medical information 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>18ae80c -
Fix Gradio messages format based on official documentation
Issue: Gradio msg.submit automatically adds user message to history Solution: Only append assistant response, not both user and assistant Based on Gradio 5 official docs for type="messages": - msg.submit automatically handles adding user message - Response function should only append assistant's reply - Format: {"role": "assistant", "content": bot_message} Fixes: "Data incompatible with messages format" error 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>e7caa1e -
Fix Gradio messages format for updated HF Spaces
- HF Spaces now uses newer Gradio requiring messages format - Added try/except for type="messages" parameter with fallback - Changed chat history to dict format with role/content keys - Fixes: "Data incompatible with messages format" error 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>79ed555 -
Revert to old Gradio format for HF Spaces compatibility
- Removed type="messages" parameter (not supported in HF Gradio) - Use list of [user, bot] pairs instead of dict format - Fixes: TypeError: Chatbot.__init__() got an unexpected keyword argument 'type' 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>f901e76 -
Re-enable MAIRA-2 Grounding for L40S GPU
- L40S has 48GB VRAM, enough for both MAIRA-2 and VQA - MAIRA-2: ~22GB - VQA: ~2GB - Others: ~2GB - Free: ~22GB 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>a782449 -
Fix Gradio messages format for newer HF Spaces version
- Updated chatbot to use type="messages" - Changed chat history to use dict format with role/content - Fixes: "Data incompatible with messages format" error This resolves compatibility with updated Gradio on HF Spaces. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>d49cf95 -
Disable MAIRA-2 Grounding to prioritize VQA on L4 GPU
Root cause: MAIRA-2 uses ~22GB VRAM (almost all of L4's 24GB) - Quantization disabled in grounding.py to fix dict.to_dict() errors - load_in_4bit parameter is ignored in grounding tool implementation - No room left for VQA (~2GB needed) Prioritizing VQA over Grounding: - VQA: Interactive Q&A about X-rays (more useful) - Grounding: Visual bounding boxes (nice-to-have) New tool config on L4 (24GB): - CheXagent VQA (~2-3GB) - Classification (~1GB) - Report Generation (~1GB) - DICOM + Web Browsing (CPU) Total: ~4-5GB with 18-19GB free 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>cf6f864 -
Disable NV-Reason-CXR to prioritize VQA tool on L4 GPU
- Commented out NV-Reason-CXR (~7GB VRAM) - MAIRA-2 Grounding uses ~15GB, leaving ~7GB for other tools - Prioritizing CheXagent VQA for interactive Q&A capabilities - With L4 (24GB), can run: Grounding + VQA + Classification + Report Gen VRAM allocation with this config: - MAIRA-2: ~15GB - VQA: ~2GB - Others: ~1GB - Free: ~6GB 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>2ae6f9b -
Re-enable NV-Reason-CXR for L4 GPU upgrade
- Uncommented NV-Reason-CXR tool loading - Enabled 4-bit quantization to optimize VRAM usage - With L4 (24GB VRAM), can now run all 7 tools: * NV-Reason-CXR (~2-3GB with quantization) * MAIRA-2 Grounding (~14.65GB) * CheXagent VQA (~1-2GB) * Classification, Report Gen, DICOM, Web Browsing Expected VRAM usage: ~18-20GB out of 24GB available 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>78ca1de -
Revert to older Gradio chatbot format for compatibility
- Removed type="messages" parameter (not supported in HF Spaces Gradio version) - Reverted chat history to list of [user, bot] pairs format - Fixes TypeError: Chatbot.__init__() got an unexpected keyword argument 'type' 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>6a95523 -
Fix Gradio chatbot message format error
- Updated chatbot to use messages format with role/content dicts - Changed grounding tool to explicitly disable quantization - Fixed chat history format to match Gradio's expectations This resolves the "Data incompatible with messages format" error. Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>cd8c9a2 -
Disable NV-Reason-CXR to save storage space
Commenting out NV-Reason-CXR tool (~8GB) to stay within HF Space 50GB limit. Can be re-enabled when using Space with more storage. Remaining tools: - MAIRA-2 Grounding - CheXagent VQA - Classification - Report Generation - DICOM - Web Browsing 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>3d05726 -
Fix MAIRA-2 grounding tool - remove BitsAndBytesConfig
This was the actual source of the dict.to_dict() error. All quantization removed from grounding tool to ensure compatibility. 🤖 Generated with [Claude Code](https://claude.com/claude-code) Co-Authored-By: Claude <noreply@anthropic.com>39ab112