Update app.py

app.py

@@ -1,42 +1,88 @@
+# app.py - Optimized for Hugging Face Spaces (Unsloth = 2-4x faster)
+
 import torch
-from fastapi import FastAPI
-from pydantic import BaseModel
-from transformers import AutoTokenizer, AutoModelForCausalLM
+import gradio as gr
+from unsloth import FastLanguageModel
 
-MODEL_ID = "saadkhi/SQL_Chat_finetuned_model"
+# ────────────────────────────────────────────────────────────────
+BASE_MODEL = "unsloth/Phi-3-mini-4k-instruct-bnb-4bit"
+LORA_PATH  = "saadkhi/SQL_Chat_finetuned_model"
 
-app = FastAPI()
+MAX_NEW_TOKENS = 180
+TEMPERATURE    = 0.0   # Greedy = fastest & deterministic
+# ────────────────────────────────────────────────────────────────
 
-# ---- LOAD ONCE ONLY ----
-tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
+print("Loading base model with Unsloth (4-bit)...")
+model, tokenizer = FastLanguageModel.from_pretrained(
+    model_name = BASE_MODEL,
+    max_seq_length = 2048,
+    dtype = None,               # Auto: bfloat16 on GPU
+    load_in_4bit = True,        # Already quantized base
+)
 
-model = AutoModelForCausalLM.from_pretrained(
-    MODEL_ID,
-    dtype=torch.float16,     # use dtype, not torch_dtype
-    device_map="auto",
-    low_cpu_mem_usage=True
+print("Applying your LoRA adapter...")
+model = FastLanguageModel.get_peft_model(
+    model,
+    r = 64,                     # Match your original rank
+    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
+                      "gate_proj", "up_proj", "down_proj"],
+    lora_alpha = 128,
+    lora_dropout = 0,
+    bias = "none",
+    use_gradient_checkpointing = "unsloth",
 )
 
-model.eval()
+# Enable 2x faster inference kernels
+FastLanguageModel.for_inference(model)
 
+print("Model ready! (very fast now)")
 
-class QueryRequest(BaseModel):
-    prompt: str
-    max_new_tokens: int = 256
+# ────────────────────────────────────────────────────────────────
+def generate_sql(prompt: str):
+    messages = [{"role": "user", "content": prompt}]
+    
+    inputs = tokenizer.apply_chat_template(
+        messages,
+        tokenize=True,
+        add_generation_prompt=True,
+        return_tensors="pt"
+    ).to("cuda" if torch.cuda.is_available() else "cpu")
 
+    outputs = model.generate(
+        input_ids = inputs,
+        max_new_tokens = MAX_NEW_TOKENS,
+        temperature = TEMPERATURE,
+        do_sample = (TEMPERATURE > 0.01),
+        use_cache = True,
+        pad_token_id = tokenizer.eos_token_id,
+    )
 
-@app.post("/generate")
-def generate(req: QueryRequest):
-    inputs = tokenizer(req.prompt, return_tensors="pt").to(model.device)
+    response = tokenizer.decode(outputs[0], skip_special_tokens=True)
+    
+    # Extract only assistant response
+    if "<|assistant|>" in response:
+        response = response.split("<|assistant|>", 1)[-1].strip()
+    response = response.split("<|end|>")[0].strip()
 
-    with torch.no_grad():
-        outputs = model.generate(
-            **inputs,
-            max_new_tokens=req.max_new_tokens,
-            do_sample=True,
-            temperature=0.7,
-            top_p=0.9
-        )
+    return response
+
+# ────────────────────────────────────────────────────────────────
+demo = gr.Interface(
+    fn = generate_sql,
+    inputs = gr.Textbox(
+        label = "Ask SQL question",
+        placeholder = "Delete duplicate rows from users table based on email",
+        lines = 3
+    ),
+    outputs = gr.Textbox(label="Generated SQL"),
+    title = "SQL Chatbot - Ultra Fast (Unsloth)",
+    description = "Phi-3-mini 4-bit + your LoRA",
+    examples = [
+        ["Find duplicate emails in users table"],
+        ["Top 5 highest paid employees"],
+        ["Count orders per customer last month"]
+    ]
+)
 
-    text = tokenizer.decode(outputs[0], skip_special_tokens=True)
-    return {"response": text}
+if __name__ == "__main__":
+    demo.launch()