Spaces:
Sleeping
Sleeping
Update app.py
Browse files
app.py
CHANGED
|
@@ -52,7 +52,7 @@ from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
|
|
| 52 |
from peft import PeftModel
|
| 53 |
import torch
|
| 54 |
|
| 55 |
-
#
|
| 56 |
quant_config = BitsAndBytesConfig(
|
| 57 |
load_in_4bit=True,
|
| 58 |
bnb_4bit_quant_type="nf4",
|
|
@@ -60,17 +60,17 @@ quant_config = BitsAndBytesConfig(
|
|
| 60 |
bnb_4bit_use_double_quant=True,
|
| 61 |
)
|
| 62 |
|
| 63 |
-
# Load base
|
| 64 |
base_model_name = "unsloth/Phi-3-mini-4k-instruct-bnb-4bit"
|
| 65 |
lora_model_name = "saadkhi/SQL_Chat_finetuned_model"
|
| 66 |
|
| 67 |
-
print("Loading model (20–
|
| 68 |
base_model = AutoModelForCausalLM.from_pretrained(
|
| 69 |
base_model_name,
|
| 70 |
quantization_config=quant_config,
|
| 71 |
device_map="auto",
|
| 72 |
trust_remote_code=True,
|
| 73 |
-
|
| 74 |
)
|
| 75 |
|
| 76 |
model = PeftModel.from_pretrained(base_model, lora_model_name)
|
|
@@ -80,7 +80,7 @@ model.eval()
|
|
| 80 |
print("Model ready!")
|
| 81 |
|
| 82 |
def chat(message, history):
|
| 83 |
-
#
|
| 84 |
messages = []
|
| 85 |
for user, assistant in history:
|
| 86 |
messages.append({"role": "user", "content": user})
|
|
@@ -88,7 +88,6 @@ def chat(message, history):
|
|
| 88 |
messages.append({"role": "assistant", "content": assistant})
|
| 89 |
messages.append({"role": "user", "content": message})
|
| 90 |
|
| 91 |
-
# Tokenize with chat template
|
| 92 |
inputs = tokenizer.apply_chat_template(
|
| 93 |
messages,
|
| 94 |
tokenize=True,
|
|
@@ -96,7 +95,7 @@ def chat(message, history):
|
|
| 96 |
return_tensors="pt"
|
| 97 |
).to(model.device)
|
| 98 |
|
| 99 |
-
#
|
| 100 |
outputs = model.generate(
|
| 101 |
inputs,
|
| 102 |
max_new_tokens=256,
|
|
@@ -104,20 +103,19 @@ def chat(message, history):
|
|
| 104 |
do_sample=True,
|
| 105 |
top_p=0.9,
|
| 106 |
repetition_penalty=1.1,
|
| 107 |
-
use_cache=True, # KV
|
| 108 |
eos_token_id=tokenizer.eos_token_id,
|
| 109 |
)
|
| 110 |
|
| 111 |
-
# Decode only the new response
|
| 112 |
response = tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True)
|
| 113 |
|
| 114 |
history.append((message, response))
|
| 115 |
return history, ""
|
| 116 |
|
| 117 |
-
#
|
| 118 |
with gr.Blocks(title="SQL Chatbot", theme=gr.themes.Soft()) as demo:
|
| 119 |
gr.Markdown("# SQL Chat Assistant")
|
| 120 |
-
gr.Markdown("Fine-tuned Phi-3 Mini for SQL
|
| 121 |
|
| 122 |
chatbot = gr.Chatbot(height=500)
|
| 123 |
msg = gr.Textbox(label="Your Question", placeholder="e.g., delete duplicate rows from users table based on email", lines=2)
|
|
|
|
| 52 |
from peft import PeftModel
|
| 53 |
import torch
|
| 54 |
|
| 55 |
+
# Best 4-bit config for speed + low memory
|
| 56 |
quant_config = BitsAndBytesConfig(
|
| 57 |
load_in_4bit=True,
|
| 58 |
bnb_4bit_quant_type="nf4",
|
|
|
|
| 60 |
bnb_4bit_use_double_quant=True,
|
| 61 |
)
|
| 62 |
|
| 63 |
+
# Load base + your LoRA once
|
| 64 |
base_model_name = "unsloth/Phi-3-mini-4k-instruct-bnb-4bit"
|
| 65 |
lora_model_name = "saadkhi/SQL_Chat_finetuned_model"
|
| 66 |
|
| 67 |
+
print("Loading model (20–40s first time)...")
|
| 68 |
base_model = AutoModelForCausalLM.from_pretrained(
|
| 69 |
base_model_name,
|
| 70 |
quantization_config=quant_config,
|
| 71 |
device_map="auto",
|
| 72 |
trust_remote_code=True,
|
| 73 |
+
# Removed flash_attention_2 — avoids install issues
|
| 74 |
)
|
| 75 |
|
| 76 |
model = PeftModel.from_pretrained(base_model, lora_model_name)
|
|
|
|
| 80 |
print("Model ready!")
|
| 81 |
|
| 82 |
def chat(message, history):
|
| 83 |
+
# Full conversation history
|
| 84 |
messages = []
|
| 85 |
for user, assistant in history:
|
| 86 |
messages.append({"role": "user", "content": user})
|
|
|
|
| 88 |
messages.append({"role": "assistant", "content": assistant})
|
| 89 |
messages.append({"role": "user", "content": message})
|
| 90 |
|
|
|
|
| 91 |
inputs = tokenizer.apply_chat_template(
|
| 92 |
messages,
|
| 93 |
tokenize=True,
|
|
|
|
| 95 |
return_tensors="pt"
|
| 96 |
).to(model.device)
|
| 97 |
|
| 98 |
+
# Optimized generation
|
| 99 |
outputs = model.generate(
|
| 100 |
inputs,
|
| 101 |
max_new_tokens=256,
|
|
|
|
| 103 |
do_sample=True,
|
| 104 |
top_p=0.9,
|
| 105 |
repetition_penalty=1.1,
|
| 106 |
+
use_cache=True, # KV cache = faster sequential tokens
|
| 107 |
eos_token_id=tokenizer.eos_token_id,
|
| 108 |
)
|
| 109 |
|
|
|
|
| 110 |
response = tokenizer.decode(outputs[0][inputs.shape[-1]:], skip_special_tokens=True)
|
| 111 |
|
| 112 |
history.append((message, response))
|
| 113 |
return history, ""
|
| 114 |
|
| 115 |
+
# UI
|
| 116 |
with gr.Blocks(title="SQL Chatbot", theme=gr.themes.Soft()) as demo:
|
| 117 |
gr.Markdown("# SQL Chat Assistant")
|
| 118 |
+
gr.Markdown("Fine-tuned Phi-3 Mini for SQL. Fast responses (3–8s on GPU).")
|
| 119 |
|
| 120 |
chatbot = gr.Chatbot(height=500)
|
| 121 |
msg = gr.Textbox(label="Your Question", placeholder="e.g., delete duplicate rows from users table based on email", lines=2)
|