Skip to content
Merged
4 changes: 2 additions & 2 deletions docs/source/ko/_toctree.yml
Original file line number Diff line number Diff line change
Expand Up @@ -119,8 +119,8 @@
title: (๋ฒˆ์—ญ์ค‘) Training on TPU with TensorFlow
- local: in_translation
title: (๋ฒˆ์—ญ์ค‘) Training on Specialized Hardware
- local: in_translation
title: (๋ฒˆ์—ญ์ค‘) Inference on CPU
- local: perf_infer_cpu
title: CPU๋กœ ์ถ”๋ก ํ•˜๊ธฐ
- local: in_translation
title: (๋ฒˆ์—ญ์ค‘) Inference on one GPU
- local: in_translation
Expand Down
73 changes: 73 additions & 0 deletions docs/source/ko/perf_infer_cpu.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,73 @@
<!--Copyright 2022 The HuggingFace Team. 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

โš ๏ธ Note that this file is in Markdown but contain specific syntax for our doc-builder (similar to MDX) that may not be
rendered properly in your Markdown viewer.

-->

# CPU์—์„œ ํšจ์œจ์ ์ธ ์ถ”๋ก ํ•˜๊ธฐ [[efficient-inference-on-cpu]]

์ด ๊ฐ€์ด๋“œ๋Š” CPU์—์„œ ๋Œ€๊ทœ๋ชจ ๋ชจ๋ธ์„ ํšจ์œจ์ ์œผ๋กœ ์ถ”๋ก ํ•˜๋Š” ๋ฐฉ๋ฒ•์— ์ค‘์ ์„ ๋‘๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.

## ๋” ๋น ๋ฅธ ์ถ”๋ก ์„ ์œ„ํ•œ `BetterTransformer` [[bettertransformer-for-faster-inference]]

์šฐ๋ฆฌ๋Š” ์ตœ๊ทผ CPU์—์„œ ํ…์ŠคํŠธ, ์ด๋ฏธ์ง€ ๋ฐ ์˜ค๋””์˜ค ๋ชจ๋ธ์˜ ๋น ๋ฅธ ์ถ”๋ก ์„ ์œ„ํ•ด `BetterTransformer`๋ฅผ ํ†ตํ•ฉํ–ˆ์Šต๋‹ˆ๋‹ค. ์ด ํ†ตํ•ฉ์— ๋Œ€ํ•œ ๋” ์ž์„ธํ•œ ๋‚ด์šฉ์€ [์ด ๋ฌธ์„œ](https://huggingface.co/docs/optimum/bettertransformer/overview)๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.

## PyTorch JIT ๋ชจ๋“œ (TorchScript) [[pytorch-jitmode-torchscript]]
TorchScript๋Š” PyTorch ์ฝ”๋“œ์—์„œ ์ง๋ ฌํ™”์™€ ์ตœ์ ํ™”๊ฐ€ ๊ฐ€๋Šฅํ•œ ๋ชจ๋ธ์„ ์ƒ์„ฑํ• ๋•Œ ์“ฐ์ž…๋‹ˆ๋‹ค. TorchScript๋กœ ๋งŒ๋“ค์–ด์ง„ ํ”„๋กœ๊ทธ๋žจ์€ ๊ธฐ์กด Python ํ”„๋กœ์„ธ์Šค์—์„œ ์ €์žฅํ•œ ๋’ค, ์ข…์†์„ฑ์ด ์—†๋Š” ์ƒˆ๋กœ์šด ํ”„๋กœ์„ธ์Šค๋กœ ๊ฐ€์ ธ์˜ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. PyTorch์˜ ๊ธฐ๋ณธ ์„ค์ •์ธ `eager` ๋ชจ๋“œ์™€ ๋น„๊ตํ–ˆ์„๋•Œ, `jit` ๋ชจ๋“œ๋Š” ์—ฐ์‚ฐ์ž ๊ฒฐํ•ฉ๊ณผ ๊ฐ™์€ ์ตœ์ ํ™” ๋ฐฉ๋ฒ•๋ก ์„ ํ†ตํ•ด ๋ชจ๋ธ ์ถ”๋ก ์—์„œ ๋Œ€๋ถ€๋ถ„ ๋” ๋‚˜์€ ์„ฑ๋Šฅ์„ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค.

TorchScript์— ๋Œ€ํ•œ ์นœ์ ˆํ•œ ์†Œ๊ฐœ๋Š” [PyTorch TorchScript ํŠœํ† ๋ฆฌ์–ผ](https://pytorch.org/tutorials/beginner/Intro_to_TorchScript_tutorial.html#tracing-modules)์„ ์ฐธ์กฐํ•˜์„ธ์š”.

### JIT ๋ชจ๋“œ์™€ ํ•จ๊ป˜ํ•˜๋Š” IPEX ๊ทธ๋ž˜ํ”„ ์ตœ์ ํ™” [[ipex-graph-optimization-with-jitmode]]
Intelยฎ Extension for PyTorch(IPEX)๋Š” Transformers ๊ณ„์—ด ๋ชจ๋ธ์˜ jit ๋ชจ๋“œ์—์„œ ์ถ”๊ฐ€์ ์ธ ์ตœ์ ํ™”๋ฅผ ์ œ๊ณตํ•ฉ๋‹ˆ๋‹ค. jit ๋ชจ๋“œ์™€ ๋”๋ถˆ์–ด Intelยฎ Extension for PyTorch(IPEX)๋ฅผ ํ™œ์šฉํ•˜์‹œ๊ธธ ๊ฐ•๋ ฅํžˆ ๊ถŒ์žฅ๋“œ๋ฆฝ๋‹ˆ๋‹ค. Transformers ๋ชจ๋ธ์—์„œ ์ž์ฃผ ์‚ฌ์šฉ๋˜๋Š” ์ผ๋ถ€ ์—ฐ์‚ฐ์ž ํŒจํ„ด์€ ์ด๋ฏธ jit ๋ชจ๋“œ ์—ฐ์‚ฐ์ž ๊ฒฐํ•ฉ(operator fusion)์˜ ํ˜•ํƒœ๋กœ Intelยฎ Extension for PyTorch(IPEX)์—์„œ ์ง€์›๋˜๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค. Multi-head-attention, Concat Linear, Linear+Add, Linear+Gelu, Add+LayerNorm ๊ฒฐํ•ฉ ํŒจํ„ด ๋“ฑ์ด ์ด์šฉ ๊ฐ€๋Šฅํ•˜๋ฉฐ ํ™œ์šฉํ–ˆ์„ ๋•Œ ์„ฑ๋Šฅ์ด ์šฐ์ˆ˜ํ•ฉ๋‹ˆ๋‹ค. ์—ฐ์‚ฐ์ž ๊ฒฐํ•ฉ์˜ ์ด์ ์€ ์‚ฌ์šฉ์ž์—๊ฒŒ ๊ณ ์Šค๋ž€ํžˆ ์ „๋‹ฌ๋ฉ๋‹ˆ๋‹ค. ๋ถ„์„์— ๋”ฐ๋ฅด๋ฉด, ์งˆ์˜ ์‘๋‹ต, ํ…์ŠคํŠธ ๋ถ„๋ฅ˜ ๋ฐ ํ† ํฐ ๋ถ„๋ฅ˜์™€ ๊ฐ™์€ ๊ฐ€์žฅ ์ธ๊ธฐ ์žˆ๋Š” NLP ํƒœ์Šคํฌ ์ค‘ ์•ฝ 70%๊ฐ€ ์ด๋Ÿฌํ•œ ๊ฒฐํ•ฉ ํŒจํ„ด์„ ์‚ฌ์šฉํ•˜์—ฌ Float32 ์ •๋ฐ€๋„์™€ BFloat16 ํ˜ผํ•ฉ ์ •๋ฐ€๋„ ๋ชจ๋‘์—์„œ ์„ฑ๋Šฅ์ƒ์˜ ์ด์ ์„ ์–ป์„ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

[IPEX ๊ทธ๋ž˜ํ”„ ์ตœ์ ํ™”](https://intel.github.io/intel-extension-for-pytorch/cpu/latest/tutorials/features/graph_optimization.html)์— ๋Œ€ํ•œ ์ž์„ธํ•œ ์ •๋ณด๋ฅผ ํ™•์ธํ•˜์„ธ์š”.

#### IPEX ์„ค์น˜: [[ipex-installation]]

IPEX ๋ฐฐํฌ ์ฃผ๊ธฐ๋Š” PyTorch๋ฅผ ๋”ฐ๋ผ์„œ ์ด๋ฃจ์–ด์ง‘๋‹ˆ๋‹ค. ์ž์„ธํ•œ ์ •๋ณด๋Š” [IPEX ์„ค์น˜ ๋ฐฉ๋ฒ•](https://intel.github.io/intel-extension-for-pytorch/)์„ ํ™•์ธํ•˜์„ธ์š”.

### JIT ๋ชจ๋“œ ์‚ฌ์šฉ๋ฒ• [[usage-of-jitmode]]
ํ‰๊ฐ€ ๋˜๋Š” ์˜ˆ์ธก์„ ์œ„ํ•ด Trainer์—์„œ JIT ๋ชจ๋“œ๋ฅผ ์‚ฌ์šฉํ•˜๋ ค๋ฉด Trainer์˜ ๋ช…๋ น ์ธ์ˆ˜์— `jit_mode_eval`์„ ์ถ”๊ฐ€ํ•ด์•ผ ํ•ฉ๋‹ˆ๋‹ค.

<Tip warning={true}>

PyTorch์˜ ๋ฒ„์ „์ด 1.14.0 ์ด์ƒ์ด๋ผ๋ฉด, jit ๋ชจ๋“œ๋Š” jit.trace์—์„œ dict ์ž…๋ ฅ์ด ์ง€์›๋˜๋ฏ€๋กœ, ๋ชจ๋“  ๋ชจ๋ธ์˜ ์˜ˆ์ธก๊ณผ ํ‰๊ฐ€๊ฐ€ ๊ฐœ์„ ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.

PyTorch์˜ ๋ฒ„์ „์ด 1.14.0 ๋ฏธ๋งŒ์ด๋ผ๋ฉด, ์งˆ์˜ ์‘๋‹ต ๋ชจ๋ธ๊ณผ ๊ฐ™์ด forward ๋งค๊ฐœ๋ณ€์ˆ˜์˜ ์ˆœ์„œ๊ฐ€ jit.trace์˜ ํŠœํ”Œ ์ž…๋ ฅ ์ˆœ์„œ์™€ ์ผ์น˜ํ•˜๋Š” ๋ชจ๋ธ์— ๋“์ด ๋  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ํ…์ŠคํŠธ ๋ถ„๋ฅ˜ ๋ชจ๋ธ๊ณผ ๊ฐ™์ด forward ๋งค๊ฐœ๋ณ€์ˆ˜ ์ˆœ์„œ๊ฐ€ jit.trace์˜ ํŠœํ”Œ ์ž…๋ ฅ ์ˆœ์„œ์™€ ๋‹ค๋ฅธ ๊ฒฝ์šฐ, jit.trace๊ฐ€ ์‹คํŒจํ•˜๋ฉฐ ์˜ˆ์™ธ๊ฐ€ ๋ฐœ์ƒํ•ฉ๋‹ˆ๋‹ค. ์ด๋•Œ ์˜ˆ์™ธ์ƒํ™ฉ์„ ์‚ฌ์šฉ์ž์—๊ฒŒ ์•Œ๋ฆฌ๊ธฐ ์œ„ํ•ด Logging์ด ์‚ฌ์šฉ๋ฉ๋‹ˆ๋‹ค.

</Tip>

[Transformers ์งˆ์˜ ์‘๋‹ต](https://github.com/huggingface/transformers/tree/main/examples/pytorch/question-answering)์˜ ์‚ฌ์šฉ ์‚ฌ๋ก€ ์˜ˆ์‹œ๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.


- CPU์—์„œ jit ๋ชจ๋“œ๋ฅผ ์‚ฌ์šฉํ•œ ์ถ”๋ก :
<pre>python run_qa.py \
--model_name_or_path csarron/bert-base-uncased-squad-v1 \
--dataset_name squad \
--do_eval \
--max_seq_length 384 \
--doc_stride 128 \
--output_dir /tmp/ \
--no_cuda \
<b>--jit_mode_eval </b></pre>

- CPU์—์„œ IPEX์™€ ํ•จ๊ป˜ jit ๋ชจ๋“œ๋ฅผ ์‚ฌ์šฉํ•œ ์ถ”๋ก :
<pre>python run_qa.py \
--model_name_or_path csarron/bert-base-uncased-squad-v1 \
--dataset_name squad \
--do_eval \
--max_seq_length 384 \
--doc_stride 128 \
--output_dir /tmp/ \
--no_cuda \
<b>--use_ipex \</b>
<b>--jit_mode_eval</b></pre>