Update app.py

app.py

@@ -34,8 +34,8 @@ app.add_middleware(
 # --- Configuration ---
 # Map filenames to "Hannah" names
 MODEL_MAP: Dict[str, str] = {
-    "qwen2.5-0.5b-instruct-q2_k.gguf": "Hannah-1.0 Light",
-    "qwen2.5-0.5b-instruct-q4_k_m.gguf": "Hannah-1.0 Heavy",
+    "qwen2.5-0.5b-instruct-q2_k.gguf": "Hannah-1.1 Light",
+    "qwen2.5-0.5b-instruct-q4_k_m.gguf": "Hannah-1.1 Heavy",
 }
 
 current_model: Optional[Llama] = None
@@ -116,21 +116,22 @@ def get_model(model_name: str) -> Llama:
         del current_model
 
     # --- PERFORMANCE TUNING (HF Free CPU) ---
-    # 4096 ctx can be too memory heavy on small Spaces; start safer, then tune up later.
+    # Increased context for Hannah 1.1 with better memory management
+    # 4096 ctx provides more context awareness; fallback to 2048 if needed
     threads = int(os.getenv("N_THREADS", "2"))
-    n_ctx = int(os.getenv("N_CTX", "2048"))
-    n_batch = int(os.getenv("N_BATCH", "256"))
+    n_ctx = int(os.getenv("N_CTX", "4096"))  # Increased from 2048
+    n_batch = int(os.getenv("N_BATCH", "512"))  # Increased from 256
 
     try:
         current_model = _try_load_model(
             model_path, n_ctx=n_ctx, n_threads=threads, n_batch=n_batch
         )
     except Exception as e:
-        # Retry with very conservative settings in case this is memory pressure.
+        # Retry with conservative settings in case of memory pressure
         print(f"Model load failed with N_CTX={n_ctx}, N_BATCH={n_batch}: {e}")
         try:
             current_model = _try_load_model(
-                model_path, n_ctx=1024, n_threads=threads, n_batch=64
+                model_path, n_ctx=2048, n_threads=threads, n_batch=256
             )
         except Exception as e2:
             print(f"Model load retry failed: {e2}")
@@ -151,7 +152,7 @@ def get_model(model_name: str) -> Llama:
 
 @app.get("/")
 async def root():
-    return {"status": "ok", "name": "Hannah-1.0"}
+    return {"status": "ok", "name": "Hannah-1.1"}
 
 
 @app.get("/api/models")
@@ -420,7 +421,7 @@ async def chat(request: Request):
 
         stream = llm(
             prompt,
-            max_tokens=2048,
+            max_tokens=4096,  # Increased from 2048 for better responses
             stop=["<|im_end|>", "User:", "System:"],
             stream=True,
         )