้กน็›ฎๆ–‡ไปถๅคน

ๆ–‡ไปถ
wehub-resource-sync e06fe8e8c6
Secret Leaks / trufflehog (push) Failing after 1s
Build documentation / build (push) Failing after 1s
Build documentation / build_other_lang (push) Failing after 0s
CodeQL Security Analysis / CodeQL Analysis (push) Failing after 0s
PR CI / pr-ci (push) Failing after 1s
Slow tests on important models (on Push - A10) / Get all modified files (push) Failing after 1s
Slow tests on important models (on Push - A10) / Model CI (push) Has been skipped
Self-hosted runner (benchmark) / Benchmark (aws-g5-4xlarge-cache) (push) Has been cancelled
New model PR merged notification / Notify new model (push) Has been cancelled
Update Transformers metadata / build_and_package (push) Has been cancelled
chore: import upstream snapshot with attribution
2026-07-13 11:57:37 +08:00

126 ่กŒ
8.6 KiB
Markdown

<!--Copyright 2020 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
specific language governing permissions and limitations under the License.
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.
-->
# ONNX๋กœ ๋‚ด๋ณด๋‚ด๊ธฐ [[export-to-onnx]]
๐Ÿค— Transformers ๋ชจ๋ธ์„ ์ œํ’ˆ ํ™˜๊ฒฝ์—์„œ ๋ฐฐํฌํ•˜๊ธฐ ์œ„ํ•ด์„œ๋Š” ๋ชจ๋ธ์„ ์ง๋ ฌํ™”๋œ ํ˜•์‹์œผ๋กœ ๋‚ด๋ณด๋‚ด๊ณ  ํŠน์ • ๋Ÿฐํƒ€์ž„๊ณผ ํ•˜๋“œ์›จ์–ด์—์„œ ๋กœ๋“œํ•˜๊ณ  ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ์œผ๋ฉด ์œ ์šฉํ•ฉ๋‹ˆ๋‹ค.
๐Ÿค— Optimum์€ Transformers์˜ ํ™•์žฅ์œผ๋กœ, PyTorch ๋˜๋Š” TensorFlow์—์„œ ๋ชจ๋ธ์„ ONNX์™€ TFLite์™€ ๊ฐ™์€ ์ง๋ ฌํ™”๋œ ํ˜•์‹์œผ๋กœ ๋‚ด๋ณด๋‚ผ ์ˆ˜ ์žˆ๋„๋ก ํ•˜๋Š” `exporters` ๋ชจ๋“ˆ์„ ํ†ตํ•ด ์ œ๊ณต๋ฉ๋‹ˆ๋‹ค. ๐Ÿค— Optimum์€ ๋˜ํ•œ ์„ฑ๋Šฅ ์ตœ์ ํ™” ๋„๊ตฌ ์„ธํŠธ๋ฅผ ์ œ๊ณตํ•˜์—ฌ ํŠน์ • ํ•˜๋“œ์›จ์–ด์—์„œ ๋ชจ๋ธ์„ ํ›ˆ๋ จํ•˜๊ณ  ์‹คํ–‰ํ•  ๋•Œ ์ตœ๋Œ€ ํšจ์œจ์„ฑ์„ ๋‹ฌ์„ฑํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค.
์ด ์•ˆ๋‚ด์„œ๋Š” ๐Ÿค— Optimum์„ ์‚ฌ์šฉํ•˜์—ฌ ๐Ÿค— Transformers ๋ชจ๋ธ์„ ONNX๋กœ ๋‚ด๋ณด๋‚ด๋Š” ๋ฐฉ๋ฒ•์„ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค. TFLite๋กœ ๋ชจ๋ธ์„ ๋‚ด๋ณด๋‚ด๋Š” ์•ˆ๋‚ด์„œ๋Š” [TFLite๋กœ ๋‚ด๋ณด๋‚ด๊ธฐ ํŽ˜์ด์ง€](tflite)๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.
## ONNX๋กœ ๋‚ด๋ณด๋‚ด๊ธฐ [[export-to-onnx]]
[ONNX (Open Neural Network eXchange)](http://onnx.ai)๋Š” PyTorch์™€ TensorFlow๋ฅผ ํฌํ•จํ•œ ๋‹ค์–‘ํ•œ ํ”„๋ ˆ์ž„์›Œํฌ์—์„œ ์‹ฌ์ธต ํ•™์Šต ๋ชจ๋ธ์„ ๋‚˜ํƒ€๋‚ด๋Š” ๋ฐ ์‚ฌ์šฉ๋˜๋Š” ๊ณตํ†ต ์—ฐ์‚ฐ์ž ์„ธํŠธ์™€ ๊ณตํ†ต ํŒŒ์ผ ํ˜•์‹์„ ์ •์˜ํ•˜๋Š” ์˜คํ”ˆ ํ‘œ์ค€์ž…๋‹ˆ๋‹ค. ๋ชจ๋ธ์ด ONNX ํ˜•์‹์œผ๋กœ ๋‚ด๋ณด๋‚ด์ง€๋ฉด ์ด๋Ÿฌํ•œ ์—ฐ์‚ฐ์ž๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ์‹ ๊ฒฝ๋ง์„ ํ†ตํ•ด ๋ฐ์ดํ„ฐ๊ฐ€ ํ๋ฅด๋Š” ํ๋ฆ„์„ ๋‚˜ํƒ€๋‚ด๋Š” ๊ณ„์‚ฐ ๊ทธ๋ž˜ํ”„(์ผ๋ฐ˜์ ์œผ๋กœ _์ค‘๊ฐ„ ํ‘œํ˜„_์ด๋ผ๊ณ  ํ•จ)๊ฐ€ ๊ตฌ์„ฑ๋ฉ๋‹ˆ๋‹ค.
ํ‘œ์ค€ํ™”๋œ ์—ฐ์‚ฐ์ž์™€ ๋ฐ์ดํ„ฐ ์œ ํ˜•์„ ๊ฐ€์ง„ ๊ทธ๋ž˜ํ”„๋ฅผ ๋…ธ์ถœํ•จ์œผ๋กœ์จ, ONNX๋Š” ํ”„๋ ˆ์ž„์›Œํฌ ๊ฐ„์— ์‰ฝ๊ฒŒ ์ „ํ™˜ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, PyTorch์—์„œ ํ›ˆ๋ จ๋œ ๋ชจ๋ธ์„ ONNX ํ˜•์‹์œผ๋กœ ๋‚ด๋ณด๋‚ด๊ณ  TensorFlow์—์„œ ๊ฐ€์ ธ์˜ฌ ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค(๊ทธ ๋ฐ˜๋Œ€๋„ ๊ฐ€๋Šฅํ•ฉ๋‹ˆ๋‹ค).
ONNX ํ˜•์‹์œผ๋กœ ๋‚ด๋ณด๋‚ธ ๋ชจ๋ธ์€ ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค:
- [๊ทธ๋ž˜ํ”„ ์ตœ์ ํ™”](https://huggingface.co/docs/optimum/onnxruntime/usage_guides/optimization) ๋ฐ [์–‘์žํ™”](https://huggingface.co/docs/optimum/onnxruntime/usage_guides/quantization)์™€ ๊ฐ™์€ ๊ธฐ๋ฒ•์„ ์‚ฌ์šฉํ•˜์—ฌ ์ถ”๋ก ์„ ์œ„ํ•ด ์ตœ์ ํ™”๋ฉ๋‹ˆ๋‹ค.
- ONNX Runtime์„ ํ†ตํ•ด ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. [`ORTModelForXXX` ํด๋ž˜์Šค๋“ค](https://huggingface.co/docs/optimum/onnxruntime/package_reference/modeling_ort)์„ ํ†ตํ•ด ๋™์ผํ•œ `AutoModel` API๋ฅผ ๋”ฐ๋ฆ…๋‹ˆ๋‹ค. ์ด API๋Š” ๐Ÿค— Transformers์—์„œ ์‚ฌ์šฉํ•˜๋Š” ๊ฒƒ๊ณผ ๋™์ผํ•ฉ๋‹ˆ๋‹ค.
- [์ตœ์ ํ™”๋œ ์ถ”๋ก  ํŒŒ์ดํ”„๋ผ์ธ](https://huggingface.co/docs/optimum/main/en/onnxruntime/usage_guides/pipelines)์„ ์‚ฌ์šฉํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์ด๋Š” ๐Ÿค— Transformers์˜ [`pipeline`] ํ•จ์ˆ˜์™€ ๋™์ผํ•œ API๋ฅผ ๊ฐ€์ง€๊ณ  ์žˆ์Šต๋‹ˆ๋‹ค.
๐Ÿค— Optimum์€ ๊ตฌ์„ฑ ๊ฐ์ฒด๋ฅผ ํ™œ์šฉํ•˜์—ฌ ONNX ๋‚ด๋ณด๋‚ด๊ธฐ๋ฅผ ์ง€์›ํ•ฉ๋‹ˆ๋‹ค. ์ด๋Ÿฌํ•œ ๊ตฌ์„ฑ ๊ฐ์ฒด๋Š” ์—ฌ๋Ÿฌ ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜์— ๋Œ€ํ•ด ๋ฏธ๋ฆฌ ์ค€๋น„๋˜์–ด ์žˆ์œผ๋ฉฐ ๋‹ค๋ฅธ ์•„ํ‚คํ…์ฒ˜์— ์‰ฝ๊ฒŒ ํ™•์žฅํ•  ์ˆ˜ ์žˆ๋„๋ก ์„ค๊ณ„๋˜์—ˆ์Šต๋‹ˆ๋‹ค.
๋ฏธ๋ฆฌ ์ค€๋น„๋œ ๊ตฌ์„ฑ ๋ชฉ๋ก์€ [๐Ÿค— Optimum ๋ฌธ์„œ](https://huggingface.co/docs/optimum/exporters/onnx/overview)๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”.
๐Ÿค— Transformers ๋ชจ๋ธ์„ ONNX๋กœ ๋‚ด๋ณด๋‚ด๋Š” ๋‘ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์ด ์žˆ์Šต๋‹ˆ๋‹ค. ์—ฌ๊ธฐ์—์„œ ๋‘ ๊ฐ€์ง€ ๋ฐฉ๋ฒ•์„ ๋ชจ๋‘ ๋ณด์—ฌ์ค๋‹ˆ๋‹ค:
- ๐Ÿค— Optimum์„ ์‚ฌ์šฉํ•˜์—ฌ CLI๋กœ ๋‚ด๋ณด๋‚ด๊ธฐ
- `optimum.onnxruntime`์„ ์‚ฌ์šฉํ•˜์—ฌ ๐Ÿค— Optimum์œผ๋กœ ONNX๋กœ ๋‚ด๋ณด๋‚ด๊ธฐ
### CLI๋ฅผ ์‚ฌ์šฉํ•˜์—ฌ ๐Ÿค— Transformers ๋ชจ๋ธ์„ ONNX๋กœ ๋‚ด๋ณด๋‚ด๊ธฐ [[exporting-a-transformers-model-to-onnx-with-cli]]
๐Ÿค— Transformers ๋ชจ๋ธ์„ ONNX๋กœ ๋‚ด๋ณด๋‚ด๋ ค๋ฉด ๋จผ์ € ์ถ”๊ฐ€ ์ข…์†์„ฑ์„ ์„ค์น˜ํ•˜์„ธ์š”:
```bash
pip install optimum-onnx
```
์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๋ชจ๋“  ์ธ์ˆ˜๋ฅผ ํ™•์ธํ•˜๋ ค๋ฉด [๐Ÿค— Optimum ๋ฌธ์„œ](https://huggingface.co/docs/optimum/exporters/onnx/usage_guides/export_a_model#exporting-a-model-to-onnx-using-the-cli)๋ฅผ ์ฐธ์กฐํ•˜๊ฑฐ๋‚˜ ๋ช…๋ น์ค„์—์„œ ๋„์›€๋ง์„ ๋ณด์„ธ์š”.
```bash
optimum-cli export onnx --help
```
์˜ˆ๋ฅผ ๋“ค์–ด, ๐Ÿค— Hub์—์„œ `distilbert/distilbert-base-uncased-distilled-squad`์™€ ๊ฐ™์€ ๋ชจ๋ธ์˜ ์ฒดํฌํฌ์ธํŠธ๋ฅผ ๋‚ด๋ณด๋‚ด๋ ค๋ฉด ๋‹ค์Œ ๋ช…๋ น์„ ์‹คํ–‰ํ•˜์„ธ์š”:
```bash
optimum-cli export onnx --model distilbert/distilbert-base-uncased-distilled-squad distilbert_base_uncased_squad_onnx/
```
์œ„์™€ ๊ฐ™์ด ์ง„ํ–‰ ์ƒํ™ฉ์„ ๋‚˜ํƒ€๋‚ด๋Š” ๋กœ๊ทธ๊ฐ€ ํ‘œ์‹œ๋˜๊ณ  ๊ฒฐ๊ณผ์ธ `model.onnx`๊ฐ€ ์ €์žฅ๋œ ์œ„์น˜๊ฐ€ ํ‘œ์‹œ๋ฉ๋‹ˆ๋‹ค.
```bash
Validating ONNX model distilbert_base_uncased_squad_onnx/model.onnx...
-[โœ“] ONNX model output names match reference model (start_logits, end_logits)
- Validating ONNX Model output "start_logits":
-[โœ“] (2, 16) matches (2, 16)
-[โœ“] all values close (atol: 0.0001)
- Validating ONNX Model output "end_logits":
-[โœ“] (2, 16) matches (2, 16)
-[โœ“] all values close (atol: 0.0001)
The ONNX export succeeded and the exported model was saved at: distilbert_base_uncased_squad_onnx
```
์œ„์˜ ์˜ˆ์ œ๋Š” ๐Ÿค— Hub์—์„œ ์ฒดํฌํฌ์ธํŠธ๋ฅผ ๋‚ด๋ณด๋‚ด๋Š” ๊ฒƒ์„ ์„ค๋ช…ํ•ฉ๋‹ˆ๋‹ค. ๋กœ์ปฌ ๋ชจ๋ธ์„ ๋‚ด๋ณด๋‚ผ ๋•Œ์—๋Š” ๋ชจ๋ธ์˜ ๊ฐ€์ค‘์น˜์™€ ํ† ํฌ๋‚˜์ด์ € ํŒŒ์ผ์„ ๋™์ผํ•œ ๋””๋ ‰ํ† ๋ฆฌ(`local_path`)์— ์ €์žฅํ–ˆ๋Š”์ง€ ํ™•์ธํ•˜์„ธ์š”. CLI๋ฅผ ์‚ฌ์šฉํ•  ๋•Œ์—๋Š” ๐Ÿค— Hub์˜ ์ฒดํฌํฌ์ธํŠธ ์ด๋ฆ„ ๋Œ€์‹  `model` ์ธ์ˆ˜์— `local_path`๋ฅผ ์ „๋‹ฌํ•˜๊ณ  `--task` ์ธ์ˆ˜๋ฅผ ์ œ๊ณตํ•˜์„ธ์š”. ์ง€์›๋˜๋Š” ์ž‘์—…์˜ ๋ชฉ๋ก์€ [๐Ÿค— Optimum ๋ฌธ์„œ](https://huggingface.co/docs/optimum/exporters/task_manager)๋ฅผ ์ฐธ์กฐํ•˜์„ธ์š”. `task` ์ธ์ˆ˜๊ฐ€ ์ œ๊ณต๋˜์ง€ ์•Š์œผ๋ฉด ์ž‘์—…์— ํŠนํ™”๋œ ํ—ค๋“œ ์—†์ด ๋ชจ๋ธ ์•„ํ‚คํ…์ฒ˜๋กœ ๊ธฐ๋ณธ ์„ค์ •๋ฉ๋‹ˆ๋‹ค.
```bash
optimum-cli export onnx --model local_path --task question-answering distilbert_base_uncased_squad_onnx/
```
๊ทธ ๊ฒฐ๊ณผ๋กœ ์ƒ์„ฑ๋œ `model.onnx` ํŒŒ์ผ์€ ONNX ํ‘œ์ค€์„ ์ง€์›ํ•˜๋Š” ๋งŽ์€ [๊ฐ€์†๊ธฐ](https://onnx.ai/supported-tools.html#deployModel) ์ค‘ ํ•˜๋‚˜์—์„œ ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, [ONNX Runtime](https://onnxruntime.ai/)์„ ์‚ฌ์šฉํ•˜์—ฌ ๋ชจ๋ธ์„ ๋กœ๋“œํ•˜๊ณ  ์‹คํ–‰ํ•  ์ˆ˜ ์žˆ์Šต๋‹ˆ๋‹ค:
```python
>>> from transformers import AutoTokenizer
>>> from optimum.onnxruntime import ORTModelForQuestionAnswering
>>> tokenizer = AutoTokenizer.from_pretrained("distilbert_base_uncased_squad_onnx")
>>> model = ORTModelForQuestionAnswering.from_pretrained("distilbert_base_uncased_squad_onnx")
>>> inputs = tokenizer("What am I using?", "Using DistilBERT with ONNX Runtime!", return_tensors="pt")
>>> outputs = model(**inputs)
```
Hub์˜ TensorFlow ์ฒดํฌํฌ์ธํŠธ์— ๋Œ€ํ•ด์„œ๋„ ๋™์ผํ•œ ํ”„๋กœ์„ธ์Šค๊ฐ€ ์ ์šฉ๋ฉ๋‹ˆ๋‹ค. ์˜ˆ๋ฅผ ๋“ค์–ด, [Keras organization](https://huggingface.co/keras-io)์—์„œ ์ˆœ์ˆ˜ํ•œ TensorFlow ์ฒดํฌํฌ์ธํŠธ๋ฅผ ๋‚ด๋ณด๋‚ด๋Š” ๋ฐฉ๋ฒ•์€ ๋‹ค์Œ๊ณผ ๊ฐ™์Šต๋‹ˆ๋‹ค:
```bash
optimum-cli export onnx --model keras-io/transformers-qa distilbert_base_cased_squad_onnx/
```
### `optimum.onnxruntime`์„ ์‚ฌ์šฉํ•˜์—ฌ ๐Ÿค— Transformers ๋ชจ๋ธ์„ ONNX๋กœ ๋‚ด๋ณด๋‚ด๊ธฐ [[exporting-a-transformers-model-to-onnx-with-optimumonnxruntime]]
CLI ๋Œ€์‹ ์— `optimum.onnxruntime`์„ ์‚ฌ์šฉํ•˜์—ฌ ํ”„๋กœ๊ทธ๋ž˜๋ฐ ๋ฐฉ์‹์œผ๋กœ ๐Ÿค— Transformers ๋ชจ๋ธ์„ ONNX๋กœ ๋‚ด๋ณด๋‚ผ ์ˆ˜๋„ ์žˆ์Šต๋‹ˆ๋‹ค. ๋‹ค์Œ๊ณผ ๊ฐ™์ด ์ง„ํ–‰ํ•˜์„ธ์š”:
```python
>>> from optimum.onnxruntime import ORTModelForSequenceClassification
>>> from transformers import AutoTokenizer
>>> model_checkpoint = "distilbert_base_uncased_squad"
>>> save_directory = "onnx/"
>>> # Load a model from transformers and export it to ONNX
>>> ort_model = ORTModelForSequenceClassification.from_pretrained(model_checkpoint, export=True)
>>> tokenizer = AutoTokenizer.from_pretrained(model_checkpoint)
>>> # Save the onnx model and tokenizer
>>> ort_model.save_pretrained(save_directory)
>>> tokenizer.save_pretrained(save_directory)
```
### ์ง€์›๋˜์ง€ ์•Š๋Š” ์•„ํ‚คํ…์ฒ˜์˜ ๋ชจ๋ธ ๋‚ด๋ณด๋‚ด๊ธฐ [[exporting-a-model-for-an-unsupported-architecture]]
ํ˜„์žฌ ๋‚ด๋ณด๋‚ผ ์ˆ˜ ์—†๋Š” ๋ชจ๋ธ์„ ์ง€์›ํ•˜๊ธฐ ์œ„ํ•ด ๊ธฐ์—ฌํ•˜๋ ค๋ฉด, ๋จผ์ € [`optimum.exporters.onnx`](https://huggingface.co/docs/optimum/exporters/onnx/overview)์—์„œ ์ง€์›๋˜๋Š”์ง€ ํ™•์ธํ•œ ํ›„ ์ง€์›๋˜์ง€ ์•Š๋Š” ๊ฒฝ์šฐ์—๋Š” [๐Ÿค— Optimum์— ๊ธฐ์—ฌ](https://huggingface.co/docs/optimum/exporters/onnx/usage_guides/contribute)ํ•˜์„ธ์š”.