fix: defer rag_pipeline import to inside @spaces.GPU worker — eliminates CUDA init error on zero-a10g

app.py

@@ -2,32 +2,31 @@
 
 import os
 import sys
-from threading import Thread
 
 sys.path.insert(0, os.path.join(os.path.dirname(__file__), "src"))
 
 import gradio as gr
-import spaces
-from rag_pipeline import AethronPipeline
+
+try:
+    import spaces
+    import spaces.zero
+    from spaces.zero import torch as spaces_torch
+    spaces_torch.patch()
+except ImportError:
+    spaces = None
 
 print("=" * 55)
 print("AETHRON PORTFOLIO AGENT — CPU Mode")
 print("=" * 55)
 
-# Dummy GPU function required by HF Spaces runtime detection on zero-a10g.
-# Never called — all inference runs on CPU.
-@spaces.GPU(duration=120)
-def _gpu_placeholder() -> str:
-    return "CPU-only mode"
-
-
-# Lazy-load pipeline on first query (model on CPU)
+# Lazy-loaded inside @spaces.GPU context to avoid CUDA init in main process
 _pipeline = None
 
 
 def get_pipeline():
     global _pipeline
     if _pipeline is None:
+        from rag_pipeline import AethronPipeline
         print("Loading pipeline on CPU (first query only)...")
         _pipeline = AethronPipeline(
             build_index=not os.path.exists("data/index/faiss.index")
@@ -35,28 +34,45 @@ def get_pipeline():
     return _pipeline
 
 
-def respond(message, history):
-    if not message or not message.strip():
+if spaces is not None:
+    @spaces.GPU(duration=120)
+    def respond(message, history):
+        if not message or not message.strip():
+            return history
+        pipeline = get_pipeline()
+        result = pipeline.query(message.strip())
+        response = result["answer"]
+        if result["sources"]:
+            source_names = []
+            for s in result["sources"][:3]:
+                name = s["section"].replace("_", " ").title()
+                if "Summary" in name:
+                    name = name.replace("Summary", "(Summary)")
+                source_names.append(name)
+            response += "\n\n**Sources:** " + " | ".join(source_names)
+        history.append({"role": "user", "content": message})
+        history.append({"role": "assistant", "content": response})
+        return history
+else:
+    from rag_pipeline import AethronPipeline
+
+    def respond(message, history):
+        if not message or not message.strip():
+            return history
+        pipeline = get_pipeline()
+        result = pipeline.query(message.strip())
+        response = result["answer"]
+        if result["sources"]:
+            source_names = []
+            for s in result["sources"][:3]:
+                name = s["section"].replace("_", " ").title()
+                if "Summary" in name:
+                    name = name.replace("Summary", "(Summary)")
+                source_names.append(name)
+            response += "\n\n**Sources:** " + " | ".join(source_names)
+        history.append({"role": "user", "content": message})
+        history.append({"role": "assistant", "content": response})
         return history
-
-    pipeline = get_pipeline()
-    result = pipeline.query(message.strip())
-    response = result["answer"]
-    if result["sources"]:
-        source_names = []
-        for s in result["sources"][:3]:
-            name = s["section"].replace("_", " ").title()
-            if "Summary" in name:
-                name = name.replace("Summary", "(Summary)")
-            source_names.append(name)
-        response += "\n\n**Sources:** " + " | ".join(source_names)
-    history.append({"role": "user", "content": message})
-    history.append({"role": "assistant", "content": response})
-    return history
-
-
-def clear_chat():
-    return []
 
 
 CUSTOM_CSS = """
@@ -117,7 +133,7 @@ with gr.Blocks(css=CUSTOM_CSS, title="Aethron | Chat with Arash's Portfolio") as
 
     msg_input.submit(respond, [msg_input, chatbot], [chatbot])
     send_btn.click(respond, [msg_input, chatbot], [chatbot])
-    clear_btn.click(clear_chat, None, [chatbot], queue=False)
+    clear_btn.click(lambda: [], None, [chatbot], queue=False)
 
     gr.Markdown("""
     ---