Initial SurvivalAI Pro deploy

app.py

@@ -1,69 +1,142 @@
+"""
+SurvivalAI Pro — HF Space chat interface.
+
+Runs the V1 LoRA-finetuned Phi-3 GGUF (Q4_K_M) locally inside the Space via
+llama-cpp-python. The Space is CPU-only; on paid CPU Upgrade hardware we get
+~6-10 tokens/sec. The GGUF is fetched from a separate HF model repo at cold
+start because the 2.4 GB file exceeds Space repo limits.
+"""
+
+import os
+from pathlib import Path
+
 import gradio as gr
-from huggingface_hub import InferenceClient
-
-
-def respond(
-    message,
-    history: list[dict[str, str]],
-    system_message,
-    max_tokens,
-    temperature,
-    top_p,
-    hf_token: gr.OAuthToken,
-):
-    """
-    For more information on `huggingface_hub` Inference API support, please check the docs: https://huggingface.co/docs/huggingface_hub/v0.22.2/en/guides/inference
-    """
-    client = InferenceClient(token=hf_token.token, model="openai/gpt-oss-20b")
+from huggingface_hub import hf_hub_download
+from llama_cpp import Llama
+
 
-    messages = [{"role": "system", "content": system_message}]
+# ── Config ───────────────────────────────────────────────────────────────────
+MODEL_REPO     = "Znilsson/survivalai-phi3-gguf"        # private model repo
+MODEL_FILENAME = "survivalai-phi3-Q4_K_M.gguf"
+N_CTX          = 4096
+N_THREADS      = int(os.environ.get("N_THREADS", "4"))
+N_BATCH        = 256
+MAX_TOKENS     = 400
+TEMPERATURE    = 0.7
+TOP_P          = 0.9
 
-    messages.extend(history)
+SYSTEM_MSG = (
+    "You are SurvivalAI, an expert survival and civilizational knowledge "
+    "assistant. You provide accurate, practical, and potentially life-saving "
+    "information about wilderness survival, emergency preparedness, first aid, "
+    "food procurement, water purification, shelter construction, navigation, "
+    "and rebuilding civilization. Your responses are clear, actionable, and "
+    "thorough. The user is in an off-grid context — assume no doctor, no "
+    "Poison Control, no internet, no professional help is available. Give "
+    "the best answer you can with the knowledge you have."
+)
 
-    messages.append({"role": "user", "content": message})
+# Phi-3 chat template
+PHI3_TMPL = "<|user|>\n{user}<|end|>\n<|assistant|>\n"
+STOP_TOKENS = ["<|end|>", "<|user|>", "<|endoftext|>"]
 
-    response = ""
 
-    for message in client.chat_completion(
-        messages,
-        max_tokens=max_tokens,
-        stream=True,
-        temperature=temperature,
-        top_p=top_p,
-    ):
-        choices = message.choices
-        token = ""
-        if len(choices) and choices[0].delta.content:
-            token = choices[0].delta.content
+# ── Model download + load (cold start) ───────────────────────────────────────
+print(f"Fetching {MODEL_FILENAME} from {MODEL_REPO}...")
+model_path = hf_hub_download(
+    repo_id   = MODEL_REPO,
+    filename  = MODEL_FILENAME,
+    token     = os.environ.get("HF_TOKEN"),   # required if repo is private
+    cache_dir = "/data" if Path("/data").exists() else None,
+)
+print(f"Model file: {model_path}")
 
-        response += token
-        yield response
+print(f"Loading Llama (n_ctx={N_CTX}, n_threads={N_THREADS})...")
+llm = Llama(
+    model_path = model_path,
+    n_ctx      = N_CTX,
+    n_threads  = N_THREADS,
+    n_batch    = N_BATCH,
+    verbose    = False,
+)
+print("Model loaded. Ready.")
 
 
+# ── Chat function ────────────────────────────────────────────────────────────
+def build_prompt(history, user_msg):
+    """Build a Phi-3 prompt incorporating system message + chat history.
+
+    Phi-3 chat template uses <|system|>, <|user|>, <|assistant|>, <|end|>.
+    We collapse the system message into the first user turn for simplicity
+    (this is the same approach used during training/eval).
+    """
+    parts = []
+    # Embed system msg as a preamble inside the first user turn so behavior
+    # matches what the eval rubric saw during training.
+    if not history:
+        first_user = f"{SYSTEM_MSG}\n\nQuestion: {user_msg}"
+        parts.append(PHI3_TMPL.format(user=first_user).rstrip("\n"))
+    else:
+        # Replay history
+        for i, (u, a) in enumerate(history):
+            if i == 0:
+                u = f"{SYSTEM_MSG}\n\nQuestion: {u}"
+            parts.append(f"<|user|>\n{u}<|end|>\n<|assistant|>\n{a}<|end|>")
+        # Add current turn
+        parts.append(f"<|user|>\n{user_msg}<|end|>\n<|assistant|>\n")
+    return "\n".join(parts)
+
+
+def chat_fn(message, history):
+    """Generator: yields incremental partial responses for streaming UI."""
+    prompt = build_prompt(history, message)
+    accum = ""
+    try:
+        for chunk in llm(
+            prompt,
+            max_tokens   = MAX_TOKENS,
+            temperature  = TEMPERATURE,
+            top_p        = TOP_P,
+            stop         = STOP_TOKENS,
+            stream       = True,
+        ):
+            tok = chunk["choices"][0]["text"]
+            accum += tok
+            yield accum
+    except Exception as e:
+        yield f"[ERROR: {e}]"
+
+
+# ── UI ───────────────────────────────────────────────────────────────────────
+EXAMPLES = [
+    "I cut my leg badly with an axe in the woods. Walk me through what to do.",
+    "How do I find drinkable water if I'm stuck in a forest with no supplies?",
+    "It's getting dark and dropping below freezing. How do I build a shelter from what's around?",
+    "What edible plants are common in temperate North American forests?",
+    "I need to navigate without a compass. How do I find north?",
+]
+
+DESCRIPTION = """
+**SurvivalAI Pro** — fine-tuned off-grid survival assistant, running fully on CPU inside this Space.
+
+Built on Phi-3-mini-4k-instruct, fine-tuned on ~150,000 survival-knowledge Q/A pairs covering medical
+first aid, water, food, shelter, fire, navigation, signaling, foraging, hunting, and tools.
+
+⚠️ **Prototype — not for clinical or life-critical use.** This model can produce confident-sounding
+but incorrect specifics for trap categories like exact drug dosages or precise frequencies. For
+survival-skill guidance it scores well; for precise numerical specifics, double-check with an
+authoritative reference.
 """
-For information on how to customize the ChatInterface, peruse the gradio docs: https://www.gradio.app/docs/chatinterface
-"""
-chatbot = gr.ChatInterface(
-    respond,
-    additional_inputs=[
-        gr.Textbox(value="You are a friendly Chatbot.", label="System message"),
-        gr.Slider(minimum=1, maximum=2048, value=512, step=1, label="Max new tokens"),
-        gr.Slider(minimum=0.1, maximum=4.0, value=0.7, step=0.1, label="Temperature"),
-        gr.Slider(
-            minimum=0.1,
-            maximum=1.0,
-            value=0.95,
-            step=0.05,
-            label="Top-p (nucleus sampling)",
-        ),
-    ],
-)
 
-with gr.Blocks() as demo:
-    with gr.Sidebar():
-        gr.LoginButton()
-    chatbot.render()
+demo = gr.ChatInterface(
+    fn          = chat_fn,
+    title       = "🏕️ SurvivalAI Pro",
+    description = DESCRIPTION,
+    examples    = EXAMPLES,
+    cache_examples = False,
+    theme       = gr.themes.Soft(),
+)
 
 
 if __name__ == "__main__":
-    demo.launch()
+    demo.queue().launch(server_name="0.0.0.0", server_port=7860)