Update app.py

app.py

@@ -1,12 +1,25 @@
 # app.py — ClimaMind on Hugging Face Spaces (Gradio)
+# Modes:
+#   PROVIDER=hf_model (default) -> calls HF Inference API for K2 (recommended for demo)
+#   PROVIDER=local              -> loads model with transformers (requires GPU Space)
+#   PROVIDER=stub               -> offline canned answers
+#
+# Space Secrets / Variables to set:
+#   HF_TOKEN        -> your Hugging Face token (Read + Inference permissions)
+#   MODEL_ID        -> default: MBZUAI-IFM/K2-Think-SFT (fallback: LLM360/K2-Think)
+#   PROVIDER        -> "hf_model" | "local" | "stub"
+#   HF_HUB_DISABLE_TELEMETRY=1 (optional)
+
 import os, time, json, random
 import requests
 import gradio as gr
 
+# -------- Config --------
 PROVIDER = os.getenv("PROVIDER", "hf_model").strip()
 MODEL_ID = os.getenv("MODEL_ID", "MBZUAI-IFM/K2-Think-SFT").strip()
 HF_TOKEN = os.getenv("HF_TOKEN", "").strip()
 
+# -------- Data fetch (Open-Meteo + OpenAQ) --------
 def _get(url, params=None, headers=None, timeout=12, retries=2, backoff=1.6):
     for i in range(retries + 1):
         try:
@@ -63,6 +76,7 @@ def fetch_factors(lat, lon):
     }
     return {"factors": factors, "raw": wx}
 
+# -------- Indices --------
 def drying_index(temp_c, rh, wind_kmh, cloud_frac=None):
     base = (temp_c or 0) * 1.2 + (wind_kmh or 0) * 0.8 - (rh or 0) * 0.9
     if cloud_frac is not None:
@@ -73,6 +87,7 @@ def heat_stress_index(temp_c, rh, wind_kmh):
     hs = (temp_c or 0) * 1.1 + (rh or 0) * 0.3 - (wind_kmh or 0) * 0.2
     return max(0, min(100, round(hs)))
 
+# -------- Prompt --------
 PROMPT = """You are ClimaMind, a climate reasoning assistant. Use ONLY the observations provided and return STRICT JSON.
 
 Location: {loc} (lat={lat}, lon={lon}), local time: {t_local}
@@ -86,12 +101,13 @@ Steps:
  3) Give a concise recommendation with time window(s) and a confidence.
  4) Output a short WHY-TRACE (3 bullets).
 Return JSON ONLY:
-{{
+{
  "answer": "...",
  "why_trace": ["...", "...", "..."],
  "risk_badge": "Low"|"Moderate"|"High"
-}}"""
+}"""
 
+# -------- Reasoning providers --------
 def call_stub(_prompt:str)->str:
     return json.dumps({
         "answer": "Based on 32°C, 50% RH and 12 km/h wind, cotton dries in ~2–3h (faster after 2pm).",
@@ -104,6 +120,7 @@ def call_stub(_prompt:str)->str:
     })
 
 def call_hf_model(prompt:str)->str:
+    # Hugging Face Inference API (serverless).
     from huggingface_hub import InferenceClient
     client = InferenceClient(model=MODEL_ID, token=(HF_TOKEN or None))
     out = client.text_generation(
@@ -133,7 +150,7 @@ def _ensure_local_loaded():
     model = AutoModelForCausalLM.from_pretrained(
         MODEL_ID,
         trust_remote_code=True,
-        device_map="auto",
+        device_map="auto",           # allows CPU offload if needed
         quantization_config=bnb_cfg,
         low_cpu_mem_usage=True,
     )
@@ -176,6 +193,7 @@ def reason_answer(loc, coords, factors, query):
     else:
         raw = call_stub(prompt)
 
+    # Extract largest JSON block
     start, end = raw.find("{"), raw.rfind("}")
     if start == -1 or end == -1:
         return {
@@ -192,6 +210,7 @@ def reason_answer(loc, coords, factors, query):
             "risk_badge": "Low"
         }
 
+# -------- Gradio UI --------
 def app(city, question):
     geo = geocode_city(city)
     data = fetch_factors(geo["lat"], geo["lon"])
@@ -227,9 +246,12 @@ demo = gr.Interface(
     outputs=gr.Markdown(label="ClimaMind"),
     title="ClimaMind — K2-Think + Live Climate Data",
     description="Provider = hf_model (Inference API) | local (GPU Space) | stub (offline). Configure env in Space settings.",
-    allow_flagging="never"
+    allow_flagging="never",
+    concurrency_limit=2,   # <— replaces deprecated queue(concurrency_count=...)
 )
-demo.queue(concurrency_count=2, max_size=8)
+
+# Optional queue for request buffering (no deprecated args)
+demo.queue(max_size=8)
 
 if __name__ == "__main__":
     demo.launch()