provide more models, secure memory usage

README.md

@@ -1,6 +1,6 @@
 ---
-title: Qwen2.5 7B Instruct Llama.cpp
-emoji: 🌍
+title: Multi-GGUF LLM Inference
+emoji: 🧠
 colorFrom: pink
 colorTo: purple
 sdk: streamlit
@@ -8,7 +8,28 @@ sdk_version: 1.44.1
 app_file: app.py
 pinned: false
 license: apache-2.0
-short_description: Run Qwen2.5-7B on Llama.cpp
+short_description: Run multiple GGUF models (Qwen2.5, Gemma-3, Phi-4) via llama.cpp
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+This Streamlit app lets you run **chat-based inference** on different GGUF models with `llama.cpp` and `llama-cpp-python`. 
+
+### 🔄 Supported Models:
+- `Qwen/Qwen2.5-7B-Instruct-GGUF` → `qwen2.5-7b-instruct-q2_k.gguf`
+- `unsloth/gemma-3-4b-it-GGUF` → `gemma-3-4b-it-Q5_K_M.gguf`
+- `unsloth/Phi-4-mini-instruct-GGUF` → `Phi-4-mini-instruct-Q5_K_M.gguf`
+
+### ⚙️ Features:
+- Model selection in sidebar
+- Custom system prompt and generation parameters
+- Chat-style UI with streaming responses
+
+### 🧠 Memory-Safe Design (for HuggingFace Spaces):
+- Only **one model is loaded at a time** (no persistent memory bloat)
+- Uses **manual unloading and `gc.collect()`** to free memory when switching
+- Reduces `n_ctx` context length to stay under 16 GB RAM limit
+- Automatically downloads models only when needed
+- Trims history to the **last 8 user-assistant turns** to avoid context overflow
+
+Perfect for deploying multi-GGUF chat models on **free-tier HuggingFace Spaces**!
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app.py

@@ -1,19 +1,66 @@
 import streamlit as st
 from llama_cpp import Llama
 from huggingface_hub import hf_hub_download
+import os
+import gc
 
-hf_hub_download(
-    repo_id="Qwen/Qwen2.5-7B-Instruct-GGUF",
-    filename="qwen2.5-7b-instruct-q2_k.gguf",
-    local_dir="./models",
-)
-
-# Load the model (on first run)
-@st.cache_resource
-def load_model():
-    return Llama(
-        model_path="models/qwen2.5-7b-instruct-q2_k.gguf",
-        n_ctx=2048,
+# Available models
+MODELS = {
+    "Qwen2.5-7B-Instruct (Q2_K)": {
+        "repo_id": "Qwen/Qwen2.5-7B-Instruct-GGUF",
+        "filename": "qwen2.5-7b-instruct-q2_k.gguf",
+        "description": "Qwen2.5-7B Instruct (Q2_K)"
+    },
+    "Gemma-3-4B-IT (Q5_K_M)": {
+        "repo_id": "unsloth/gemma-3-4b-it-GGUF",
+        "filename": "gemma-3-4b-it-Q5_K_M.gguf",
+        "description": "Gemma 3 4B IT (Q5_K_M)"
+    },
+    "Phi-4-mini-Instruct (Q5_K_M)": {
+        "repo_id": "unsloth/Phi-4-mini-instruct-GGUF",
+        "filename": "Phi-4-mini-instruct-Q5_K_M.gguf",
+        "description": "Phi-4 Mini Instruct (Q5_K_M)"
+    },
+}
+
+with st.sidebar:
+    st.header("⚙️ Settings")
+    selected_model_name = st.selectbox("Select Model", list(MODELS.keys()))
+    system_prompt = st.text_area("System Prompt", value="You are a helpful assistant.", height=80)
+    max_tokens = st.slider("Max tokens", 64, 2048, 512, step=32)
+    temperature = st.slider("Temperature", 0.1, 2.0, 0.7)
+    top_k = st.slider("Top-K", 1, 100, 40)
+    top_p = st.slider("Top-P", 0.1, 1.0, 0.95)
+    repeat_penalty = st.slider("Repetition Penalty", 1.0, 2.0, 1.1)
+
+# Model info
+selected_model = MODELS[selected_model_name]
+model_path = os.path.join("models", selected_model["filename"])
+
+# Initialize model cache state
+if "model_name" not in st.session_state:
+    st.session_state.model_name = None
+if "llm" not in st.session_state:
+    st.session_state.llm = None
+
+# Download model if needed
+if not os.path.exists(model_path):
+    hf_hub_download(
+        repo_id=selected_model["repo_id"],
+        filename=selected_model["filename"],
+        local_dir="./models",
+        local_dir_use_symlinks=False,
+    )
+
+# Load model only if it changed
+if st.session_state.model_name != selected_model_name:
+    if st.session_state.llm is not None:
+        # Clean up old model to free memory
+        del st.session_state.llm
+        gc.collect()
+    st.session_state.llm = Llama(
+        model_path=model_path,
+        n_ctx=1024,  # Reduced for RAM safety
         n_threads=2,
         n_threads_batch=2,
         n_batch=4,
@@ -22,43 +69,34 @@ def load_model():
         use_mmap=True,
         verbose=False,
     )
+    st.session_state.model_name = selected_model_name
 
-llm = load_model()
+llm = st.session_state.llm
 
-# Session state for chat history
+# Chat history state
 if "chat_history" not in st.session_state:
     st.session_state.chat_history = []
 
-st.title("🧠 Qwen2.5-7B-Instruct (Streamlit + GGUF)")
-st.caption("Powered by `llama.cpp` and `llama-cpp-python` | 2-bit Q2_K inference")
-
-with st.sidebar:
-    st.header("⚙️ Settings")
-    system_prompt = st.text_area("System Prompt", value="You are a helpful assistant.", height=80)
-    max_tokens = st.slider("Max tokens", 64, 2048, 512, step=32)
-    temperature = st.slider("Temperature", 0.1, 2.0, 0.7)
-    top_k = st.slider("Top-K", 1, 100, 40)
-    top_p = st.slider("Top-P", 0.1, 1.0, 0.95)
-    repeat_penalty = st.slider("Repetition Penalty", 1.0, 2.0, 1.1)
+st.title(f"🧠 {selected_model['description']} (Streamlit + GGUF)")
+st.caption(f"Powered by `llama.cpp` | Model: {selected_model['filename']}")
 
-# Input box
 user_input = st.chat_input("Ask something...")
 
 if user_input:
-    # Add user message to chat
     st.session_state.chat_history.append({"role": "user", "content": user_input})
 
-    # Display user message
     with st.chat_message("user"):
         st.markdown(user_input)
 
-    # Construct the prompt
-    messages = [{"role": "system", "content": system_prompt}] + st.session_state.chat_history
+    # Trim conversation history to max 8 turns (user+assistant)
+    MAX_TURNS = 8
+    trimmed_history = st.session_state.chat_history[-MAX_TURNS * 2:]
+    messages = [{"role": "system", "content": system_prompt}] + trimmed_history
 
-    # Stream response
     with st.chat_message("assistant"):
         full_response = ""
         response_area = st.empty()
+
         stream = llm.create_chat_completion(
             messages=messages,
             max_tokens=max_tokens,