change to phi-3

app.py

@@ -1,46 +1,56 @@
 import gradio as gr
-from transformers import AutoModelForCausalLM, AutoTokenizer
 import torch
+from transformers import AutoModelForCausalLM, AutoTokenizer
+
+# ── Model & tokenizer ───────────────────────────────────────────────────────────
+model_name = "microsoft/Phi-3-mini-4k-instruct"  # 3.8 B‑param open instruct model :contentReference[oaicite:0]{index=0}
+
+# seed for reproducibility
+torch.random.manual_seed(0)
+
+# load tokenizer (no gating, open access) :contentReference[oaicite:1]{index=1}
+tokenizer = AutoTokenizer.from_pretrained(model_name, trust_remote_code=True)
 
-model_name = "TheBloke/openinstruct‑mistral‑7B‑AWQ"
-tokenizer = AutoTokenizer.from_pretrained(model_name)
+# load as causal‑LM, quantize to int8 to halve memory & speed up inference :contentReference[oaicite:2]{index=2}
 model = AutoModelForCausalLM.from_pretrained(model_name)
+# compile once for ~30% further speed‑up :contentReference[oaicite:3]{index=3}
 model = torch.compile(model)
 
+# ── Prompt template ──────────────────────────────────────────────────────────────
 prompt_prefix = """
-You are an energy-saving expert tasked to help households reduce their monthly electricity bills.
+You are an energy‑saving expert tasked to help households reduce their monthly electricity bills.
 Given the user's appliance usage information (device name, wattage, hours used per day, days used per week):
 1. Flag the highest energy consumers.
 2. Recommend practical, empathetic, achievable actions.
 3. Suggest appliance swaps (e.g. LED, inverter AC) and habit changes.
 Format with bullet points.
-Here is the input of the user:
+Here is the summary:
 """
 
+# ── Generation function ──────────────────────────────────────────────────────────
 def generate_recommendation(appliance_info: str) -> str:
-    # Build the full prompt
     prompt = prompt_prefix + appliance_info + "\n\nRecommendations:"
-    # Tokenize and move inputs to the model device
     inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
-    # Generate with no grad, limited tokens
     with torch.no_grad():
         outputs = model.generate(
             **inputs,
-            max_new_tokens=120,
-            use_cache=True
+            max_new_tokens=150,
+            use_cache=True,
+            do_sample=False,
+            temperature=0.0
         )
-    # Decode and return only the recommendations section
     text = tokenizer.decode(outputs[0], skip_special_tokens=True)
+    # return only the recommendations section
     return text.split("Recommendations:")[-1].strip()
 
-# Define the Gradio interface
+# ── Gradio interface ────────────────────────────────────────────────────────────
 iface = gr.Interface(
     fn=generate_recommendation,
-    inputs=gr.Textbox(lines=10, placeholder="Enter appliance usage details..."),
+    inputs=gr.Textbox(lines=8, placeholder="e.g. Refrigerator: 150 W, 8 h/day, 7 days/week\n..."),
     outputs="text",
-    title="Energy-Saving Recommendation Generator",
-    description="Provide appliance usage details to receive energy-saving tips."
+    title="Energy‑Saving Tips (Phi‑3‑Mini‑4K‑Instruct)",
+    description="Paste your per‑appliance summary to get targeted energy‑saving recommendations."
 )
 
 if __name__ == "__main__":
-    iface.launch()
+    iface.launch()