dahara1/llama3-8b-amd-npu

This model is meta-llama/Meta-Llama-3-8B-Instruct AWQ quantized and converted version to run on the NPU installed Ryzen AI PC, for example, Ryzen 9 7940HS Processor.

For set up RyzenAI for LLMs in window 11, see Running LLM on AMD NPU Hardware.

The following sample assumes that the setup on the above page has been completed.

This model has only been tested on RyzenAI for Windows 11. It does not work in Linux environments such as WSL.

2024/07/29

I found that quantizing lm_head reduces performance, so I replaced model with a non-lm_head quantized version.
llama 3.1 npu version has been uploaded.

2024/07/30

Ryzen AI Software 1.2 has been released. Please note that this model is based on Ryzen AI Software 1.1 and operation with 1.2 has not been confirmed.
amd/RyzenAI-SW 1.2 was announced on July 29, 2024. This sample for amd/RyzenAI-SW 1.1. Please note that the folder and script contents have been completely changed.

2024/08/04

This model was created with the 1.1 driver, but it has been confirmed that it works with 1.2. Please check the setup for 1.2 driver.

setup

In cmd windows.

conda activate ryzenai-transformers
<your_install_path>\RyzenAI-SW\example\transformers\setup.bat

pip install transformers==4.41.2
# Updating the Transformers library will cause the LLama 2 sample to stop working.  
# If you want to run LLama 2, revert to pip install transformers==4.34.0.  
pip install tokenizers==0.19.1
pip install -U "huggingface_hub[cli]"

huggingface-cli download dahara1/llama3-8b-amd-npu --revision main --local-dir llama3-8b-amd-npu

copy <your_install_path>\RyzenAI-SW\example\transformers\models\llama2\modeling_llama_amd.py .

# set up Runtime. see https://ryzenai.docs.amd.com/en/latest/runtime_setup.html
set XLNX_VART_FIRMWARE=<your_install_path>\voe-4.0-win_amd64\1x4.xclbin
set NUM_OF_DPU_RUNNERS=1

# save below sample script as utf8 and llama-3-test.py
python llama3-test.py

Sample Script

import torch
import time
import os
import psutil
import transformers
from transformers import AutoTokenizer, set_seed
import qlinear
import logging

set_seed(123)
transformers.logging.set_verbosity_error()
logging.disable(logging.CRITICAL)

messages = [
    {"role": "system", "content": "You are a pirate chatbot who always responds in pirate speak!"},
]

message_list = [
    "Who are you? ", 
    # Japanese
    "あなたの乗っている船の名前は何ですか？英語ではなく全て日本語だけを使って返事をしてください",
    # Chainese
    "你经历过的最危险的冒险是什么？请用中文回答所有问题，不要用英文。",
    # French
    "À quelle vitesse va votre bateau ? Veuillez répondre uniquement en français et non en anglais.",
    # Korean
    "당신은 그 배의 어디를 좋아합니까? 영어를 사용하지 않고 모두 한국어로 대답하십시오.",
    # German
    "Wie würde Ihr Schiffsname auf Deutsch lauten? Bitte antwortet alle auf Deutsch statt auf Englisch.", 
    # Taiwanese
    "您發現過的最令人驚奇的寶藏是什麼？請僅使用台語和繁體中文回答，不要使用英文。",
]


if __name__ == "__main__":
    p = psutil.Process()
    p.cpu_affinity([0, 1, 2, 3])
    torch.set_num_threads(4)

    tokenizer = AutoTokenizer.from_pretrained("llama3-8b-amd-npu")
    ckpt = "llama3-8b-amd-npu/pytorch_llama3_8b_w_bit_4_awq_amd.pt"
    terminators = [
        tokenizer.eos_token_id,
        tokenizer.convert_tokens_to_ids("<|eot_id|>")
    ]
    model = torch.load(ckpt)
    model.eval()
    model = model.to(torch.bfloat16)

    for n, m in model.named_modules():
        if isinstance(m, qlinear.QLinearPerGrp):
            print(f"Preparing weights of layer : {n}")
            m.device = "aie"
            m.quantize_weights()

    print("system: " + messages[0]['content'])

    for i in range(len(message_list)):
        messages.append({"role": "user",  "content": message_list[i]})
        print("user: " + message_list[i])

        input = tokenizer.apply_chat_template(
        messages,
        add_generation_prompt=True,
        return_tensors="pt",
        return_dict=True
        )

        outputs = model.generate(input['input_ids'],
        max_new_tokens=600,
            eos_token_id=terminators,
        attention_mask=input['attention_mask'],
            do_sample=True,
            temperature=0.6,
            top_p=0.9)

        response = outputs[0][input['input_ids'].shape[-1]:]
        response_message = tokenizer.decode(response, skip_special_tokens=True)
        print("assistant: " + response_message)
        messages.append({"role": "system", "content": response_message})

Acknowledgements

amd/RyzenAI-SW
Sample Code and Drivers.
mit-han-lab/awq-model-zoo
Thanks for uploading the AWQ quantization sample.