Update app.py

app.py

@@ -75,13 +75,13 @@ Rules:
 def load_model():
     return Llama(
         model_path=hf_hub_download(
-            repo_id="Qwen/Qwen2.5-1.5B-Instruct-GGUF",
-            filename="qwen2.5-1.5b-instruct-q4_k_m.gguf",
+            repo_id="Valtry/llama3.2-3b-q4-gguf",
+            filename="llama3.2-3b-q4.gguf",
             token=HF_TOKEN,
             cache_dir="/data"
         ),
         n_ctx=512,        # ✅ reduced from 2048 — fits prompt + 80 token reply
-        n_threads=2,      # ✅ matches HF free tier vCPU count (was 4)
+        n_threads=4,      # ✅ matches HF free tier vCPU count (was 4)
         n_batch=128,      # ✅ smaller batch = faster for single requests (was 512)
         use_mmap=True,
         use_mlock=False,  # ✅ set False — mlock can fail on HF shared infra (was True)
@@ -132,7 +132,7 @@ async def chat(req: ChatRequest):
         partial(
             model,
             prompt,
-            max_tokens=100,          # ✅ reduced from 2048 — 30 words ≈ 60-80 tokens
+            max_tokens=80,          # ✅ reduced from 2048 — 30 words ≈ 60-80 tokens
             temperature=req.temperature,
             top_p=0.9,
             top_k=40,               # ✅ added — limits candidates, speeds up sampling