Update app.py

app.py

@@ -67,7 +67,7 @@ llm = Llama(
     
     # MoE Cache & RAM Bandwidth Optimizations:
     # 1. Quantized KV Cache (Q4_0) saves RAM and frees CPU cache space for routing
-    type_k=getattr(llama_cpp, "GGML_TYPE_Q4_0", 2),
+    type_k=getattr(llama_cpp, "GGML_TYPE_Q8_0", 2),
     type_v=getattr(llama_cpp, "GGML_TYPE_Q4_0", 2),
     
     # 2. Smaller ubatch prevents L2/L3 cache eviction during active expert evaluation
@@ -76,10 +76,10 @@ llm = Llama(
     
     use_mlock=False,
     use_mmap=True,
-    flash_attn=False,
+    flash_attn=True,
 )
 
-MODEL_NAME = "Qwen3.6-35B-A3B"
+MODEL_NAME = "unsloth/Qwen3.6-35B-A3B-GGUF:UD-IQ3_XXS"
 
 # ---------------------------------------------------------------------------
 # ⚡ 3. FASTAPI INITIALIZATION & OPENAI PYDANTIC SCHEMAS