Update app.py

app.py

@@ -1,59 +1,61 @@
-# app.py
-# Stable CPU-only Hugging Face Space
-# Phi-3-mini + LoRA (NO bitsandbytes, NO SSR issues)
-
 import warnings
-warnings.filterwarnings("ignore", category=FutureWarning)
+warnings.filterwarnings("ignore")
 
 import torch
 import gradio as gr
 from transformers import AutoTokenizer, AutoModelForCausalLM
 from peft import PeftModel
 
-# ─────────────────────────────────────────────
-# Config
-# ─────────────────────────────────────────────
+# ─────────────────────────────
 BASE_MODEL = "unsloth/Phi-3-mini-4k-instruct"
 LORA_PATH  = "saadkhi/SQL_Chat_finetuned_model"
 
 MAX_NEW_TOKENS = 180
-TEMPERATURE    = 0.0
-DO_SAMPLE      = False
-
-# ─────────────────────────────────────────────
-# Load model & tokenizer (CPU SAFE)
-# ─────────────────────────────────────────────
-print("Loading base model on CPU...")
-
-model = AutoModelForCausalLM.from_pretrained(
-    BASE_MODEL,
-    device_map="cpu",
-    torch_dtype=torch.float32,
-    trust_remote_code=True,
-    low_cpu_mem_usage=True,
-)
 
-print("Loading LoRA adapter...")
-model = PeftModel.from_pretrained(model, LORA_PATH)
+model = None
+tokenizer = None
+
+# ─────────────────────────────
+# Lazy load (VERY IMPORTANT)
+# ─────────────────────────────
+def load_model():
+    global model, tokenizer
+
+    if model is not None:
+        return
+
+    print("🔄 Loading model (first request only)...")
+
+    base = AutoModelForCausalLM.from_pretrained(
+        BASE_MODEL,
+        device_map="cpu",
+        torch_dtype=torch.float16,   # lighter
+        low_cpu_mem_usage=True,
+        trust_remote_code=True,
+    )
+
+    base = PeftModel.from_pretrained(base, LORA_PATH)
+
+    print("Merging LoRA...")
+    model_loaded = base.merge_and_unload()
+
+    tokenizer_loaded = AutoTokenizer.from_pretrained(BASE_MODEL)
+
+    model_loaded.eval()
 
-print("Merging LoRA weights...")
-model = model.merge_and_unload()
+    model = model_loaded
+    tokenizer = tokenizer_loaded
 
-tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
+    print("✅ Model ready")
 
-model.eval()
-print("Model & tokenizer loaded successfully")
+# ─────────────────────────────
+def generate_sql(question):
+    if not question.strip():
+        return "Enter a question"
 
-# ─────────────────────────────────────────────
-# Inference
-# ─────────────────────────────────────────────
-def generate_sql(question: str) -> str:
-    if not question or not question.strip():
-        return "Please enter a SQL-related question."
+    load_model()
 
-    messages = [
-        {"role": "user", "content": question.strip()}
-    ]
+    messages = [{"role": "user", "content": question}]
 
     input_ids = tokenizer.apply_chat_template(
         messages,
@@ -66,60 +68,29 @@ def generate_sql(question: str) -> str:
         output_ids = model.generate(
             input_ids=input_ids,
             max_new_tokens=MAX_NEW_TOKENS,
-            temperature=TEMPERATURE,
-            do_sample=DO_SAMPLE,
+            temperature=0.0,
+            do_sample=False,
             pad_token_id=tokenizer.eos_token_id,
-            use_cache=True,
         )
 
-    response = tokenizer.decode(
-        output_ids[0],
-        skip_special_tokens=True
-    )
+    response = tokenizer.decode(output_ids[0], skip_special_tokens=True)
 
-    # Clean Phi-3 chat artifacts
-    for token in ["<|assistant|>", "<|user|>", "<|end|>"]:
-        if token in response:
-            response = response.split(token)[-1]
+    for t in ["<|assistant|>", "<|user|>", "<|end|>"]:
+        if t in response:
+            response = response.split(t)[-1]
 
-    return response.strip() or "(empty response)"
+    return response.strip()
 
-# ─────────────────────────────────────────────
-# Gradio UI
-# ─────────────────────────────────────────────
+# ─────────────────────────────
 demo = gr.Interface(
     fn=generate_sql,
-    inputs=gr.Textbox(
-        label="SQL Question",
-        placeholder="Find duplicate emails in users table",
-        lines=3,
-    ),
-    outputs=gr.Textbox(
-        label="Generated SQL",
-        lines=8,
-    ),
-    title="SQL Chat – Phi-3-mini (CPU)",
-    description=(
-        "CPU-only Hugging Face Space.\n"
-        "First response may take 60–180 seconds. "
-        "Subsequent requests are faster."
-    ),
-    examples=[
-        ["Find duplicate emails in users table"],
-        ["Top 5 highest paid employees"],
-        ["Count orders per customer last month"],
-        ["Delete duplicate rows based on email"],
-    ],
-    cache_examples=False,
+    inputs=gr.Textbox(lines=3, label="SQL Question"),
+    outputs=gr.Textbox(lines=8, label="SQL"),
+    title="SQL Chat Phi-3 CPU",
+    description="First request loads model (60-120s)",
 )
 
-# ─────────────────────────────────────────────
-# Launch
-# ─────────────────────────────────────────────
+demo.queue(concurrency_count=1, max_size=5)
+
 if __name__ == "__main__":
-    print("Launching Gradio interface...")
-    demo.launch(
-        server_name="0.0.0.0",
-        ssr_mode=False,   # important: avoids asyncio FD bug
-        show_error=True,
-    )
+    demo.launch(server_name="0.0.0.0", show_error=True)