metadata
license: apache-2.0
pipeline_tag: text-generation
library_name: transformers
AMD-Llama-135m-code-GGUF
Introduction
AMD-Llama-135m is a language model trained on AMD MI250 GPUs. Based on LLaMA2 model architecture, this model can be smoothly loaded as LlamaForCausalLM with huggingface transformers. Furthermore, we use the same tokenizer as LLaMA2, enabling it to be a draft model of speculative decoding for LLaMA2 and CodeLlama.
Quickstart
AMD-Llama-135m-code-GGUF can be loaded and used via Llama.cpp, here is a program with GUI.
pip install PyQt5 llama-cpp-python pymupdf
import sys
import os
from PyQt5.QtWidgets import (QApplication, QWidget, QLabel, QPushButton,
QLineEdit, QTextEdit, QVBoxLayout, QHBoxLayout,
QFileDialog, QProgressBar, QMessageBox, QMenu)
from PyQt5.QtCore import Qt, QThread, pyqtSignal
from llama_cpp import Llama
import fitz # For PDF processing
class Worker(QThread):
finished = pyqtSignal(str)
progress = pyqtSignal(int, int)
def __init__(self, model, messages, max_tokens):
super().__init__()
self.model = model
self.messages = messages
self.max_tokens = max_tokens
def run(self):
try:
response = self.model.create_chat_completion(
messages=self.messages,
max_tokens=self.max_tokens,
temperature=0.7,
stream=True
)
total_tokens = 0
full_response = ""
for chunk in response:
if "choices" in chunk:
content = chunk["choices"][0]["delta"].get("content", "")
full_response += content
total_tokens += 1
self.progress.emit(total_tokens, self.max_tokens)
self.finished.emit(full_response)
except Exception as e:
self.finished.emit(f"Error generating response: {str(e)}")
class ChatbotGUI(QWidget):
def __init__(self):
super().__init__()
self.setWindowTitle("Chatbot GUI")
self.resize(800, 600)
self.model = None
self.messages = [
{"role": "system", "content": "You are a helpful AI assistant."}
]
self.thread_count = 12
self.pdf_content = ""
self.initUI()
def initUI(self):
# Model loading section
model_label = QLabel("Model: No model loaded")
load_button = QPushButton("Load GGUF Model")
load_button.clicked.connect(self.load_model)
model_layout = QHBoxLayout()
model_layout.addWidget(model_label)
model_layout.addWidget(load_button)
# PDF upload section
pdf_label = QLabel("PDF: No PDF loaded")
upload_pdf_button = QPushButton("Upload PDF")
upload_pdf_button.clicked.connect(self.upload_pdf)
pdf_layout = QHBoxLayout()
pdf_layout.addWidget(pdf_label)
pdf_layout.addWidget(upload_pdf_button)
# Thread count section
thread_label = QLabel(f"Thread Count: {self.thread_count}")
self.thread_input = QLineEdit()
self.thread_input.setPlaceholderText("Enter new thread count")
update_thread_button = QPushButton("Update Threads")
update_thread_button.clicked.connect(self.update_thread_count)
thread_layout = QHBoxLayout()
thread_layout.addWidget(thread_label)
thread_layout.addWidget(self.thread_input)
thread_layout.addWidget(update_thread_button)
# Chat display
self.chat_display = QTextEdit()
self.chat_display.setReadOnly(True)
self.chat_display.setContextMenuPolicy(Qt.CustomContextMenu)
self.chat_display.customContextMenuRequested.connect(self.show_context_menu)
# User input
self.user_input = QLineEdit()
self.user_input.returnPressed.connect(self.send_message)
send_button = QPushButton("Send")
send_button.clicked.connect(self.send_message)
input_layout = QHBoxLayout()
input_layout.addWidget(self.user_input)
input_layout.addWidget(send_button)
# Progress bar
self.progress_bar = QProgressBar()
self.progress_bar.hide()
# Clear conversation button
clear_button = QPushButton("Clear Conversation")
clear_button.clicked.connect(self.clear_conversation)
# Main layout
main_layout = QVBoxLayout()
main_layout.addLayout(model_layout)
main_layout.addLayout(pdf_layout) # PDF before threads
main_layout.addLayout(thread_layout)
main_layout.addWidget(self.chat_display)
main_layout.addWidget(self.progress_bar)
main_layout.addLayout(input_layout)
main_layout.addWidget(clear_button)
self.setLayout(main_layout)
def load_model(self):
model_path, _ = QFileDialog.getOpenFileName(self, "Load GGUF Model", "", "GGUF Files (*.gguf)")
if model_path:
try:
self.model = Llama(model_path=model_path, n_ctx=2048, n_gpu_layers=-1, n_threads=self.thread_count)
model_name = os.path.basename(model_path)
self.layout().itemAt(0).itemAt(0).widget().setText(f"Model: {model_name}")
QMessageBox.information(self, "Success", "Model loaded successfully!")
except Exception as e:
error_message = f"Error loading model: {str(e)}"
QMessageBox.critical(self, "Error", error_message)
def update_thread_count(self):
try:
new_thread_count = int(self.thread_input.text())
if new_thread_count > 0:
self.thread_count = new_thread_count
self.layout().itemAt(2).itemAt(0).widget().setText(f"Thread Count: {self.thread_count}") # Updated index
self.thread_input.clear()
if self.model:
self.model.set_thread_count(self.thread_count)
QMessageBox.information(self, "Success", f"Thread count updated to {self.thread_count}")
else:
raise ValueError("Thread count must be a positive integer")
except ValueError as e:
QMessageBox.warning(self, "Invalid Input", str(e))
def upload_pdf(self):
pdf_path, _ = QFileDialog.getOpenFileName(self, "Upload PDF", "", "PDF Files (*.pdf)")
if pdf_path:
try:
doc = fitz.open(pdf_path)
self.pdf_content = ""
for page in doc:
self.pdf_content += page.get_text()
self.layout().itemAt(1).itemAt(0).widget().setText(f"PDF: {os.path.basename(pdf_path)}") # Updated index
QMessageBox.information(self, "Success", "PDF loaded successfully!")
except Exception as e:
QMessageBox.critical(self, "Error", f"Error loading PDF: {str(e)}")
def send_message(self):
user_message = self.user_input.text()
if user_message and self.model:
self.messages.append({"role": "user", "content": user_message})
self.update_chat_display(f"You: {user_message}")
self.user_input.clear()
max_tokens = 1000
self.progress_bar.show()
self.progress_bar.setRange(0, max_tokens)
self.progress_bar.setValue(0)
# Add PDF content if available
if self.pdf_content:
self.messages.append({"role": "user", "content": self.pdf_content})
self.worker = Worker(self.model, self.messages, max_tokens)
self.worker.finished.connect(self.on_response_finished)
self.worker.progress.connect(self.on_response_progress)
self.worker.start()
def on_response_finished(self, assistant_message):
self.progress_bar.hide()
self.messages.append({"role": "assistant", "content": assistant_message})
self.update_chat_display(f"Assistant: {assistant_message}")
# Python Code Download
if assistant_message.startswith("```python") and assistant_message.endswith("```"):
self.offer_code_download(assistant_message)
def on_response_progress(self, current_tokens, total_tokens):
self.progress_bar.setValue(current_tokens)
def offer_code_download(self, code):
reply = QMessageBox.question(self, "Download Code",
"The assistant generated Python code. Do you want to download it?",
QMessageBox.Yes | QMessageBox.No)
if reply == QMessageBox.Yes:
file_path, _ = QFileDialog.getSaveFileName(self, "Save Python Code", "code.py", "Python Files (*.py)")
if file_path:
try:
with open(file_path, "w") as f:
f.write(code.strip("```python").strip("```"))
QMessageBox.information(self, "Success", "Code saved successfully!")
except Exception as e:
QMessageBox.critical(self, "Error", f"Error saving code: {str(e)}")
def update_chat_display(self, message):
self.chat_display.append(message + "\n")
self.chat_display.verticalScrollBar().setValue(self.chat_display.verticalScrollBar().maximum())
def clear_conversation(self):
self.messages = [
{"role": "system", "content": "You are a helpful AI assistant."}
]
self.chat_display.clear()
self.pdf_content = "" # Clear PDF content
self.layout().itemAt(1).itemAt(0).widget().setText("PDF: No PDF loaded") # Updated index
def show_context_menu(self, point):
menu = QMenu(self)
copy_action = menu.addAction("Copy")
copy_action.triggered.connect(self.copy_text)
menu.exec_(self.chat_display.mapToGlobal(point))
def copy_text(self):
cursor = self.chat_display.textCursor()
if cursor.hasSelection():
text = cursor.selectedText()
QApplication.clipboard().setText(text)
if __name__ == "__main__":
app = QApplication(sys.argv)
gui = ChatbotGUI()
gui.show()
sys.exit(app.exec_())
Training and finetuning cost
It takes 6 days to pretrain AMD-Llama-135m on 4 MI250 nodes each of which has 4 MI250 GPUs (8 virtual GPU cards, 64G memory for each). It takes 4 days to finetune AMD-Llama-135m-code on 4 MI250 GPUs. It takes 11T disk space to store raw and processed SlimPajama, project gutenberg and Starcoder datasets.
License
Copyright (c) 2018-2024 Advanced Micro Devices, Inc. All Rights Reserved.
Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at
http://www.apache.org/licenses/LICENSE-2.0
Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.