Update app.py

app.py

@@ -12,6 +12,15 @@ import re
 import psutil
 import subprocess
 import traceback
+import torch  # For checking CUDA availability
+
+# Set VLLM logging level to DEBUG for detailed output
+os.environ["VLLM_LOGGING_LEVEL"] = "DEBUG"
+
+# If no GPU is available, force CPU usage by hiding CUDA devices
+if not torch.cuda.is_available():
+    print("No GPU detected. Forcing CPU mode by setting CUDA_VISIBLE_DEVICES to an empty string.")
+    os.environ["CUDA_VISIBLE_DEVICES"] = ""
 
 # Persistent directory setup
 persistent_dir = "/data/hf_cache"
@@ -26,6 +35,7 @@ vllm_cache_dir = os.path.join(persistent_dir, "vllm_cache")
 for directory in [model_cache_dir, tool_cache_dir, file_cache_dir, report_dir, vllm_cache_dir]:
     os.makedirs(directory, exist_ok=True)
 
+# Update environment variables to use HF_HOME
 os.environ["HF_HOME"] = model_cache_dir
 os.environ["TRANSFORMERS_CACHE"] = model_cache_dir
 os.environ["VLLM_CACHE_DIR"] = vllm_cache_dir
@@ -38,7 +48,7 @@ sys.path.insert(0, src_path)
 
 from txagent.txagent import TxAgent
 
-# Medical Keywords used in the PDF processing
+# Medical keywords for processing PDF files
 MEDICAL_KEYWORDS = {'diagnosis', 'assessment', 'plan', 'results', 'medications',
                     'allergies', 'summary', 'impression', 'findings', 'recommendations'}
 
@@ -53,11 +63,9 @@ def extract_priority_pages(file_path: str, max_pages: int = 20) -> str:
     try:
         text_chunks = []
         with pdfplumber.open(file_path) as pdf:
-            # Always include the first three pages
             for i, page in enumerate(pdf.pages[:3]):
                 text = page.extract_text() or ""
                 text_chunks.append(f"=== Page {i+1} ===\n{text.strip()}")
-            # Include further pages if they contain any medical keywords
             for i, page in enumerate(pdf.pages[3:max_pages], start=4):
                 page_text = page.extract_text() or ""
                 if any(re.search(rf'\b{kw}\b', page_text.lower()) for kw in MEDICAL_KEYWORDS):
@@ -138,6 +146,8 @@ def init_agent():
             seed=100,
             additional_default_tools=[],
         )
+        # This call attempts to load the models. If device inference fails,
+        # it will now produce DEBUG-level logs.
         agent.init_model()
         log_system_usage("After Load")
         print("✅ Agent Ready")
@@ -166,7 +176,10 @@ def create_ui(agent):
                 file_hash_value = ""
                 if files:
                     with ThreadPoolExecutor(max_workers=4) as executor:
-                        futures = [executor.submit(convert_file_to_json, f.name, f.name.split(".")[-1].lower()) for f in files]
+                        futures = [
+                            executor.submit(convert_file_to_json, f.name, f.name.split(".")[-1].lower())
+                            for f in files
+                        ]
                         results = []
                         for future in as_completed(futures):
                             try:
@@ -194,8 +207,6 @@ Medical Records:
                 print(prompt)
 
                 full_response = ""
-                finish_detected = False
-
                 for chunk in agent.run_gradio_chat(
                     message=prompt,
                     history=[],
@@ -215,23 +226,19 @@ Medical Records:
                             chunk_content = "".join([c.content for c in chunk if hasattr(c, "content") and c.content])
                             print("DEBUG: Received list chunk:", chunk_content)
                             full_response += chunk_content
-                            if '"name": "Finish"' in chunk_content:
-                                finish_detected = True
                         else:
                             print("DEBUG: Received unknown type chunk", type(chunk))
                     except Exception as e:
                         print("❌ Error processing chunk:", str(e))
                         traceback.print_exc()
 
-                    # Yield intermediate full (raw) response for debugging
+                    # Yield intermediate raw response for debugging purposes
                     history[-1] = {"role": "assistant", "content": full_response}
                     yield history, None
 
-                # Final processing
                 if not full_response:
                     full_response = "⚠️ No clear oversights identified or model output was invalid."
 
-                # Save full report
                 report_path = None
                 if file_hash_value:
                     report_path = os.path.join(report_dir, f"{file_hash_value}_report.txt")