a2c1675546
* Add a new CPU example of Yuan2-2B-hf * Add a new CPU generate.py of Yuan2-2B-hf example * Add a new GPU example of Yuan2-2B-hf * Add Yuan2 to README table * In CPU example:1.Use English as default prompt; 2.Provide modified files in yuan2-2B-instruct * In GPU example:1.Use English as default prompt;2.Provide modified files * GPU example:update README * update Yuan2-2B-hf in README table * Add CPU example for Yuan2-2B in Pytorch-Models * Add GPU example for Yuan2-2B in Pytorch-Models * Add license in generate.py; Modify README * In GPU Add license in generate.py; Modify README * In CPU yuan2 modify README * In GPU yuan2 modify README * In CPU yuan2 modify README * In GPU example, updated the readme for Windows GPU supports * In GPU torch example, updated the readme for Windows GPU supports * GPU hf example README modified * GPU example README modified
78 lines
No EOL
3.5 KiB
Python
78 lines
No EOL
3.5 KiB
Python
#
|
|
# Copyright 2016 The BigDL Authors.
|
|
#
|
|
# 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.
|
|
#
|
|
|
|
import torch, transformers
|
|
import sys, os, time
|
|
import intel_extension_for_pytorch as ipex
|
|
import argparse
|
|
from transformers import LlamaTokenizer
|
|
from bigdl.llm.transformers import AutoModelForCausalLM
|
|
|
|
# Refer to https://huggingface.co/IEITYuan/Yuan2-2B-hf#Usage
|
|
YUAN2_PROMPT_FORMAT = """
|
|
{prompt}
|
|
"""
|
|
|
|
if __name__ == '__main__':
|
|
parser = argparse.ArgumentParser(description='Generate text using Yuan2-2B model')
|
|
parser.add_argument('--repo-id-or-model-path', type=str, default="IEITYuan/Yuan2-2B-hf",
|
|
help='The huggingface repo id for the Yuan2 to be downloaded'
|
|
', or the path to the huggingface checkpoint folder')
|
|
parser.add_argument('--prompt', type=str, default="What is AI?",
|
|
help='Prompt for the model')
|
|
parser.add_argument('--n-predict', type=int, default=100,
|
|
help='Number of tokens to generate')
|
|
|
|
args = parser.parse_args()
|
|
model_path = args.repo_id_or_model_path
|
|
|
|
# Load tokenizer
|
|
print("Creating tokenizer...")
|
|
tokenizer = LlamaTokenizer.from_pretrained(model_path, add_eos_token=False, add_bos_token=False, eos_token='<eod>')
|
|
tokenizer.add_tokens(['<sep>', '<pad>', '<mask>', '<predict>', '<FIM_SUFFIX>', '<FIM_PREFIX>', '<FIM_MIDDLE>','<commit_before>',
|
|
'<commit_msg>','<commit_after>','<jupyter_start>','<jupyter_text>','<jupyter_code>','<jupyter_output>','<empty_output>'], special_tokens=True)
|
|
|
|
# Load model in 4 bit,
|
|
# which convert the relevant layers in the model into INT4 format
|
|
# When running LLMs on Intel iGPUs for Windows users, we recommend setting `cpu_embedding=True` in the from_pretrained function.
|
|
# This will allow the memory-intensive embedding layer to utilize the CPU instead of iGPU.
|
|
print("Creating model...")
|
|
model = AutoModelForCausalLM.from_pretrained(model_path, trust_remote_code=True, load_in_4bit=True).eval()
|
|
# Convert the model to xpu
|
|
model = model.to('xpu')
|
|
|
|
prompt = YUAN2_PROMPT_FORMAT.format(prompt=args.prompt)
|
|
inputs = tokenizer(prompt, return_tensors="pt")["input_ids"]
|
|
# Convert the inputs to xpu
|
|
inputs = inputs.to('xpu')
|
|
|
|
# Default warmup since the first generate() is slow
|
|
outputs = model.generate(inputs, do_sample=True, top_k=5, max_length=args.n_predict)
|
|
print('Finish warmup')
|
|
|
|
# Measure the inference time
|
|
start_time = time.time()
|
|
# if your selected model is capable of utilizing previous key/value attentions
|
|
# to enhance decoding speed, but has `"use_cache": false` in its model config,
|
|
# it is important to set `use_cache=True` explicitly in the `generate` function
|
|
# to obtain optimal performance with BigDL-LLM INT4 optimizations
|
|
outputs = model.generate(inputs, do_sample=True, top_k=5, max_length=args.n_predict)
|
|
end_time = time.time()
|
|
|
|
output_str = tokenizer.decode(outputs[0])
|
|
print(f'Inference time: {end_time - start_time} seconds')
|
|
print('-'*20, 'Output', '-'*20)
|
|
print(output_str) |