Update MODEL_PATH to Q8_K_P, add threading + reasoning flags

entrypoint.sh

@@ -1,23 +1,39 @@
 #!/bin/bash
 set -e
 
-# Use pre-downloaded local model
-MODEL_PATH="/data/model/Gemma-4-E4B-Uncensored-HauhauCS-Aggressive-Q4_K_M.gguf"
+MODEL_PATH="/data/model/Gemma-4-E4B-Uncensored-HauhauCS-Aggressive-Q8_K_P.gguf"
 
-# Start llama-server on internal port 8080
+# Detect CPU cores for max threading
+NPROC=$(nproc)
+echo "Using $NPROC threads"
+
+# Context window = 131072 tokens.
+# --n-predict 25000 sets the default/maximum generated tokens per request.
+#
+# Reasoning ("thinking") is ENABLED BY DEFAULT:
+#   --jinja                      use the model's embedded chat template
+#   --reasoning-format deepseek  extract <think>...</think> into a SEPARATE
+#                                `reasoning_content` field in the OpenAI
+#                                response/stream (instead of inline tags)
+#   --reasoning-budget -1        unrestricted thinking length
 /app/llama-server \
     --model "$MODEL_PATH" \
     --port 8080 \
     --host 127.0.0.1 \
     --ctx-size 131072 \
+    --n-predict 25000 \
     --parallel 1 \
-    --jinja &
+    --threads "$NPROC" \
+    --threads-batch "$NPROC" \
+    --batch-size 512 \
+    --jinja \
+    --reasoning-format deepseek \
+    --reasoning-budget -1 &
 
 LLAMA_PID=$!
 
-# Wait for llama-server to be ready
 echo "Waiting for llama-server to start..."
-for i in {1..300}; do
+for i in {1..600}; do
     if curl -s http://127.0.0.1:8080/health > /dev/null 2>&1; then
         echo "llama-server is ready!"
         break
@@ -25,5 +41,4 @@ for i in {1..300}; do
     sleep 1
 done
 
-# Start the proxy on port 8000 (HF Spaces expects this)
 exec uvicorn proxy:app --host 0.0.0.0 --port 8000 --proxy-headers