perf_infer_gpu_one.md 22.7 KB
Newer Older
1
2
3
4
5
6
7
8
9
<!--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
10
11
12
13

鈿狅笍 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.

14
15
-->

16
# GPU inference
17

18
GPUs are the standard choice of hardware for machine learning, unlike CPUs, because they are optimized for memory bandwidth and parallelism. To keep up with the larger sizes of modern models or to run these large models on existing and older hardware, there are several optimizations you can use to speed up GPU inference. In this guide, you'll learn how to use FlashAttention-2 (a more memory-efficient attention mechanism), BetterTransformer (a PyTorch native fastpath execution), and bitsandbytes to quantize your model to a lower precision. Finally, learn how to use 馃 Optimum to accelerate inference with ONNX Runtime on Nvidia and AMD GPUs.
19
20
21

<Tip>

22
The majority of the optimizations described here also apply to multi-GPU setups!
23
24
25

</Tip>

26
## FlashAttention-2
27

28
<Tip>
29

30
FlashAttention-2 is experimental and may change considerably in future versions.
31

32
</Tip>
33

34
[FlashAttention-2](https://huggingface.co/papers/2205.14135) is a faster and more efficient implementation of the standard attention mechanism that can significantly speedup inference by:
35

36
37
1. additionally parallelizing the attention computation over sequence length
2. partitioning the work between GPU threads to reduce communication and shared memory reads/writes between them
38

39
40
41
FlashAttention-2 is currently supported for the following architectures:
* [Bark](https://huggingface.co/docs/transformers/model_doc/bark#transformers.BarkModel)
* [Bart](https://huggingface.co/docs/transformers/model_doc/bart#transformers.BartModel)
Saurabh Dash's avatar
Saurabh Dash committed
42
* [Cohere](https://huggingface.co/docs/transformers/model_doc/cohere#transformers.CohereModel)
43
* [DistilBert](https://huggingface.co/docs/transformers/model_doc/distilbert#transformers.DistilBertModel)
44
* [Gemma](https://huggingface.co/docs/transformers/model_doc/gemma#transformers.GemmaModel)
45
* [GPT2](https://huggingface.co/docs/transformers/model_doc/gpt2)
46
47
48
* [GPTBigCode](https://huggingface.co/docs/transformers/model_doc/gpt_bigcode#transformers.GPTBigCodeModel)
* [GPTNeo](https://huggingface.co/docs/transformers/model_doc/gpt_neo#transformers.GPTNeoModel)
* [GPTNeoX](https://huggingface.co/docs/transformers/model_doc/gpt_neox#transformers.GPTNeoXModel)
49
* [GPT-J](https://huggingface.co/docs/transformers/model_doc/gptj#transformers.GPTJModel)
amyeroberts's avatar
amyeroberts committed
50
* [Idefics2](https://huggingface.co/docs/transformers/model_doc/idefics2#transformers.Idefics2Model)
51
* [Falcon](https://huggingface.co/docs/transformers/model_doc/falcon#transformers.FalconModel)
tomeras91's avatar
tomeras91 committed
52
* [Jamba](https://huggingface.co/docs/transformers/model_doc/jamba#transformers.JambaModel)
53
54
* [Llama](https://huggingface.co/docs/transformers/model_doc/llama#transformers.LlamaModel)
* [Llava](https://huggingface.co/docs/transformers/model_doc/llava)
NielsRogge's avatar
NielsRogge committed
55
* [Llava-NeXT](https://huggingface.co/docs/transformers/model_doc/llava_next)
56
* [VipLlava](https://huggingface.co/docs/transformers/model_doc/vipllava)
57
* [M2M100](https://huggingface.co/docs/transformers/model_doc/m2m_100)
58
59
* [MBart](https://huggingface.co/docs/transformers/model_doc/mbart#transformers.MBartModel)
* [Mistral](https://huggingface.co/docs/transformers/model_doc/mistral#transformers.MistralModel)
60
* [Mixtral](https://huggingface.co/docs/transformers/model_doc/mixtral#transformers.MixtralModel)
61
62
* [Musicgen](https://huggingface.co/docs/transformers/model_doc/musicgen#transformers.MusicgenModel)
* [MusicGen Melody](https://huggingface.co/docs/transformers/model_doc/musicgen_melody#transformers.MusicgenMelodyModel)
63
* [NLLB](https://huggingface.co/docs/transformers/model_doc/nllb)
Shane A's avatar
Shane A committed
64
* [OLMo](https://huggingface.co/docs/transformers/model_doc/olmo#transformers.OlmoModel)
65
66
* [OPT](https://huggingface.co/docs/transformers/model_doc/opt#transformers.OPTModel)
* [Phi](https://huggingface.co/docs/transformers/model_doc/phi#transformers.PhiModel)
Jonathan Tow's avatar
Jonathan Tow committed
67
* [StableLm](https://huggingface.co/docs/transformers/model_doc/stablelm#transformers.StableLmModel)
RaymondLi0's avatar
RaymondLi0 committed
68
* [Starcoder2](https://huggingface.co/docs/transformers/model_doc/starcoder2#transformers.Starcoder2Model)
Junyang Lin's avatar
Junyang Lin committed
69
* [Qwen2](https://huggingface.co/docs/transformers/model_doc/qwen2#transformers.Qwen2Model)
Bo Zheng's avatar
Bo Zheng committed
70
* [Qwen2MoE](https://huggingface.co/docs/transformers/model_doc/qwen2_moe#transformers.Qwen2MoeModel)
71
72
73
* [Whisper](https://huggingface.co/docs/transformers/model_doc/whisper#transformers.WhisperModel)

You can request to add FlashAttention-2 support for another model by opening a GitHub Issue or Pull Request.
74

Steven Liu's avatar
Steven Liu committed
75
Before you begin, make sure you have FlashAttention-2 installed.
76

Steven Liu's avatar
Steven Liu committed
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
<hfoptions id="install">
<hfoption id="NVIDIA">

```bash
pip install flash-attn --no-build-isolation
```

We strongly suggest referring to the detailed [installation instructions](https://github.com/Dao-AILab/flash-attention?tab=readme-ov-file#installation-and-features) to learn more about supported hardware and data types!

</hfoption>
<hfoption id="AMD">

FlashAttention-2 is also supported on AMD GPUs and current support is limited to **Instinct MI210** and **Instinct MI250**. We strongly suggest using this [Dockerfile](https://github.com/huggingface/optimum-amd/tree/main/docker/transformers-pytorch-amd-gpu-flash/Dockerfile) to use FlashAttention-2 on AMD GPUs.

</hfoption>
</hfoptions>
93

94
To enable FlashAttention-2, pass the argument `attn_implementation="flash_attention_2"` to [`~AutoModelForCausalLM.from_pretrained`]:
95
96
97
98
99
100
101
102
103

```python
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, LlamaForCausalLM

model_id = "tiiuae/falcon-7b"
tokenizer = AutoTokenizer.from_pretrained(model_id)

model = AutoModelForCausalLM.from_pretrained(
amyeroberts's avatar
amyeroberts committed
104
105
    model_id,
    torch_dtype=torch.bfloat16,
106
    attn_implementation="flash_attention_2",
107
108
109
)
```

110
<Tip>
111

112
FlashAttention-2 can only be used when the model's dtype is `fp16` or `bf16`. Make sure to cast your model to the appropriate dtype and load them on a supported device before using FlashAttention-2.
113

Steven Liu's avatar
Steven Liu committed
114
115
116
<br>

You can also set `use_flash_attention_2=True` to enable FlashAttention-2 but it is deprecated in favor of `attn_implementation="flash_attention_2"`.
amyeroberts's avatar
amyeroberts committed
117

118
</Tip>
119

120
FlashAttention-2 can be combined with other optimization techniques like quantization to further speedup inference. For example, you can combine FlashAttention-2 with 8-bit or 4-bit quantization:
121

122
```py
123
124
125
126
127
128
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, LlamaForCausalLM

model_id = "tiiuae/falcon-7b"
tokenizer = AutoTokenizer.from_pretrained(model_id)

129
# load in 8bit
130
model = AutoModelForCausalLM.from_pretrained(
amyeroberts's avatar
amyeroberts committed
131
    model_id,
132
    load_in_8bit=True,
133
    attn_implementation="flash_attention_2",
134
135
)

136
# load in 4bit
137
model = AutoModelForCausalLM.from_pretrained(
amyeroberts's avatar
amyeroberts committed
138
    model_id,
139
    load_in_4bit=True,
140
    attn_implementation="flash_attention_2",
141
142
143
)
```

144
### Expected speedups
145

146
You can benefit from considerable speedups for inference, especially for inputs with long sequences. However, since FlashAttention-2 does not support computing attention scores with padding tokens, you must manually pad/unpad the attention scores for batched inference when the sequence contains padding tokens. This leads to a significant slowdown for batched generations with padding tokens.
147

148
To overcome this, you should use FlashAttention-2 without padding tokens in the sequence during training (by packing a dataset or [concatenating sequences](https://github.com/huggingface/transformers/blob/main/examples/pytorch/language-modeling/run_clm.py#L516) until reaching the maximum sequence length).
149

150
For a single forward pass on [tiiuae/falcon-7b](https://hf.co/tiiuae/falcon-7b) with a sequence length of 4096 and various batch sizes without padding tokens, the expected speedup is:
151

152
153
154
<div style="text-align: center">
<img src="https://huggingface.co/datasets/ybelkada/documentation-images/resolve/main/falcon-7b-inference-large-seqlen.png">
</div>
155

156
For a single forward pass on [meta-llama/Llama-7b-hf](https://hf.co/meta-llama/Llama-7b-hf) with a sequence length of 4096 and various batch sizes without padding tokens, the expected speedup is:
157

158
159
160
<div style="text-align: center">
<img src="https://huggingface.co/datasets/ybelkada/documentation-images/resolve/main/llama-7b-inference-large-seqlen.png">
</div>
161

162
For sequences with padding tokens (generating with padding tokens), you need to unpad/pad the input sequences to correctly compute the attention scores. With a relatively small sequence length, a single forward pass creates overhead leading to a small speedup (in the example below, 30% of the input is filled with padding tokens):
163

164
165
166
<div style="text-align: center">
<img src="https://huggingface.co/datasets/ybelkada/documentation-images/resolve/main/llama-2-small-seqlen-padding.png">
</div>
167

168
But for larger sequence lengths, you can expect even more speedup benefits:
169
170
171

<Tip>

172
FlashAttention is more memory efficient, meaning you can train on much larger sequence lengths without running into out-of-memory issues. You can potentially reduce memory usage up to 20x for larger sequence lengths. Take a look at the [flash-attention](https://github.com/Dao-AILab/flash-attention) repository for more details.
173

174
</Tip>
Younes Belkada's avatar
Younes Belkada committed
175

176
177
178
<div style="text-align: center">
<img src="https://huggingface.co/datasets/ybelkada/documentation-images/resolve/main/llama-2-large-seqlen-padding.png">
</div>
179

Steven Liu's avatar
Steven Liu committed
180
## PyTorch scaled dot product attention
181

Steven Liu's avatar
Steven Liu committed
182
PyTorch's [`torch.nn.functional.scaled_dot_product_attention`](https://pytorch.org/docs/master/generated/torch.nn.functional.scaled_dot_product_attention.html) (SDPA) can also call FlashAttention and memory-efficient attention kernels under the hood. SDPA support is currently being added natively in Transformers and is used by default for `torch>=2.1.1` when an implementation is available.
183

Steven Liu's avatar
Steven Liu committed
184
For now, Transformers supports SDPA inference and training for the following architectures:
185
* [Bart](https://huggingface.co/docs/transformers/model_doc/bart#transformers.BartModel)
Saurabh Dash's avatar
Saurabh Dash committed
186
* [Cohere](https://huggingface.co/docs/transformers/model_doc/cohere#transformers.CohereModel)
187
188
* [GPTBigCode](https://huggingface.co/docs/transformers/model_doc/gpt_bigcode#transformers.GPTBigCodeModel)
* [Falcon](https://huggingface.co/docs/transformers/model_doc/falcon#transformers.FalconModel)
189
* [Gemma](https://huggingface.co/docs/transformers/model_doc/gemma#transformers.GemmaModel)
tomeras91's avatar
tomeras91 committed
190
* [Jamba](https://huggingface.co/docs/transformers/model_doc/jamba#transformers.JambaModel)
191
* [Llama](https://huggingface.co/docs/transformers/model_doc/llama#transformers.LlamaModel)
Shane A's avatar
Shane A committed
192
* [OLMo](https://huggingface.co/docs/transformers/model_doc/olmo#transformers.OlmoModel)
JB (Don)'s avatar
JB (Don) committed
193
* [Phi](https://huggingface.co/docs/transformers/model_doc/phi#transformers.PhiModel)
194
195
* [Idefics](https://huggingface.co/docs/transformers/model_doc/idefics#transformers.IdeficsModel)
* [Whisper](https://huggingface.co/docs/transformers/model_doc/whisper#transformers.WhisperModel)
196
197
* [Mistral](https://huggingface.co/docs/transformers/model_doc/mistral#transformers.MistralModel)
* [Mixtral](https://huggingface.co/docs/transformers/model_doc/mixtral#transformers.MixtralModel)
198
* [StableLm](https://huggingface.co/docs/transformers/model_doc/stablelm#transformers.StableLmModel)
RaymondLi0's avatar
RaymondLi0 committed
199
* [Starcoder2](https://huggingface.co/docs/transformers/model_doc/starcoder2#transformers.Starcoder2Model)
Junyang Lin's avatar
Junyang Lin committed
200
* [Qwen2](https://huggingface.co/docs/transformers/model_doc/qwen2#transformers.Qwen2Model)
Bo Zheng's avatar
Bo Zheng committed
201
* [Qwen2MoE](https://huggingface.co/docs/transformers/model_doc/qwen2_moe#transformers.Qwen2MoeModel)
202
203
* [Musicgen](https://huggingface.co/docs/transformers/model_doc/musicgen#transformers.MusicgenModel)
* [MusicGen Melody](https://huggingface.co/docs/transformers/model_doc/musicgen_melody#transformers.MusicgenMelodyModel)
204

Steven Liu's avatar
Steven Liu committed
205
206
<Tip>

207
FlashAttention can only be used for models with the `fp16` or `bf16` torch type, so make sure to cast your model to the appropriate type first. The memory-efficient attention backend is able to handle `fp32` models.
208

Steven Liu's avatar
Steven Liu committed
209
210
211
</Tip>

By default, SDPA selects the most performant kernel available but you can check whether a backend is available in a given setting (hardware, problem size) with [`torch.backends.cuda.sdp_kernel`](https://pytorch.org/docs/master/backends.html#torch.backends.cuda.sdp_kernel) as a context manager:
212
213
214
215
216
217

```diff
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("facebook/opt-350m")
218
model = AutoModelForCausalLM.from_pretrained("facebook/opt-350m", torch_dtype=torch.float16).to("cuda")
219
220
221
222
223
224
225
226
227
228
229
230
# convert the model to BetterTransformer
model.to_bettertransformer()

input_text = "Hello my dog is cute and"
inputs = tokenizer(input_text, return_tensors="pt").to("cuda")

+ with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=False):
    outputs = model.generate(**inputs)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))
```

Steven Liu's avatar
Steven Liu committed
231
If you see a bug with the traceback below, try using the nightly version of PyTorch which may have broader coverage for FlashAttention:
232
233

```bash
234
RuntimeError: No available kernel. Aborting execution.
235

236
# install PyTorch nightly
237
238
239
pip3 install -U --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/cu118
```

240
241
242
243
## BetterTransformer

<Tip warning={true}>

Steven Liu's avatar
Steven Liu committed
244
Some BetterTransformer features are being upstreamed to Transformers with default support for native `torch.nn.scaled_dot_product_attention`. BetterTransformer still has a wider coverage than the Transformers SDPA integration, but you can expect more and more architectures to natively support SDPA in Transformers.
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275

</Tip>

<Tip>

Check out our benchmarks with BetterTransformer and scaled dot product attention in the [Out of the box acceleration and memory savings of 馃 decoder models with PyTorch 2.0](https://pytorch.org/blog/out-of-the-box-acceleration/) and learn more about the fastpath execution in the [BetterTransformer](https://medium.com/pytorch/bettertransformer-out-of-the-box-performance-for-huggingface-transformers-3fbe27d50ab2) blog post.

</Tip>

BetterTransformer accelerates inference with its fastpath (native PyTorch specialized implementation of Transformer functions) execution. The two optimizations in the fastpath execution are:

1. fusion, which combines multiple sequential operations into a single "kernel" to reduce the number of computation steps
2. skipping the inherent sparsity of padding tokens to avoid unnecessary computation with nested tensors

BetterTransformer also converts all attention operations to use the more memory-efficient [scaled dot product attention (SDPA)](https://pytorch.org/docs/master/generated/torch.nn.functional.scaled_dot_product_attention), and it calls optimized kernels like [FlashAttention](https://huggingface.co/papers/2205.14135) under the hood.

Before you start, make sure you have 馃 Optimum [installed](https://huggingface.co/docs/optimum/installation).

Then you can enable BetterTransformer with the [`PreTrainedModel.to_bettertransformer`] method:

```python
model = model.to_bettertransformer()
```

You can return the original Transformers model with the [`~PreTrainedModel.reverse_bettertransformer`] method. You should use this before saving your model to use the canonical Transformers modeling:

```py
model = model.reverse_bettertransformer()
model.save_pretrained("saved_model")
```

276
## bitsandbytes
277

278
bitsandbytes is a quantization library that includes support for 4-bit and 8-bit quantization. Quantization reduces your model size compared to its native full precision version, making it easier to fit large models onto GPUs with limited memory.
279

Stas Bekman's avatar
Stas Bekman committed
280
Make sure you have bitsandbytes and 馃 Accelerate installed:
281

282
283
284
```bash
# these versions support 8-bit and 4-bit
pip install bitsandbytes>=0.39.0 accelerate>=0.20.0
285

286
287
288
# install Transformers
pip install transformers
```
289

290
### 4-bit
291

292
To load a model in 4-bit for inference, use the `load_in_4bit` parameter. The `device_map` parameter is optional, but we recommend setting it to `"auto"` to allow 馃 Accelerate to automatically and efficiently allocate the model given the available resources in the environment.
293
294
295
296
297

```py
from transformers import AutoModelForCausalLM

model_name = "bigscience/bloom-2b5"
298
model_4bit = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", load_in_4bit=True)
299
300
```

301
To load a model in 4-bit for inference with multiple GPUs, you can control how much GPU RAM you want to allocate to each GPU. For example, to distribute 600MB of memory to the first GPU and 1GB of memory to the second GPU:
302
303
304
305

```py
max_memory_mapping = {0: "600MB", 1: "1GB"}
model_name = "bigscience/bloom-3b"
306
model_4bit = AutoModelForCausalLM.from_pretrained(
307
308
309
310
    model_name, device_map="auto", load_in_4bit=True, max_memory=max_memory_mapping
)
```

311
### 8-bit
312

313
<Tip>
314

315
If you're curious and interested in learning more about the concepts underlying 8-bit quantization, read the [Gentle Introduction to 8-bit Matrix Multiplication for transformers at scale using Hugging Face Transformers, Accelerate and bitsandbytes](https://huggingface.co/blog/hf-bitsandbytes-integration) blog post.
316
317
318

</Tip>

319
To load a model in 8-bit for inference, use the `load_in_8bit` parameter. The `device_map` parameter is optional, but we recommend setting it to `"auto"` to allow 馃 Accelerate to automatically and efficiently allocate the model given the available resources in the environment:
320

321
```py
322
323
from transformers import AutoModelForCausalLM

324
325
326
327
model_name = "bigscience/bloom-2b5"
model_8bit = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", load_in_8bit=True)
```

328
If you're loading a model in 8-bit for text generation, you should use the [`~transformers.GenerationMixin.generate`] method instead of the [`Pipeline`] function which is not optimized for 8-bit models and will be slower. Some sampling strategies, like nucleus sampling, are also not supported by the [`Pipeline`] for 8-bit models. You should also place all inputs on the same device as the model:
329
330
331
332
333
334
335
336

```py
from transformers import AutoModelForCausalLM, AutoTokenizer

model_name = "bigscience/bloom-2b5"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model_8bit = AutoModelForCausalLM.from_pretrained(model_name, device_map="auto", load_in_8bit=True)

337
prompt = "Hello, my llama is cute"
338
339
340
341
342
inputs = tokenizer(prompt, return_tensors="pt").to("cuda")
generated_ids = model.generate(**inputs)
outputs = tokenizer.batch_decode(generated_ids, skip_special_tokens=True)
```

343
To load a model in 4-bit for inference with multiple GPUs, you can control how much GPU RAM you want to allocate to each GPU. For example, to distribute 1GB of memory to the first GPU and 2GB of memory to the second GPU:
344
345
346
347
348
349
350
351
352

```py
max_memory_mapping = {0: "1GB", 1: "2GB"}
model_name = "bigscience/bloom-3b"
model_8bit = AutoModelForCausalLM.from_pretrained(
    model_name, device_map="auto", load_in_8bit=True, max_memory=max_memory_mapping
)
```

353
<Tip>
354

355
Feel free to try running a 11 billion parameter [T5 model](https://colab.research.google.com/drive/1YORPWx4okIHXnjW7MSAidXN29mPVNT7F?usp=sharing) or the 3 billion parameter [BLOOM model](https://colab.research.google.com/drive/1qOjXfQIAULfKvZqwCen8-MoWKGdSatZ4?usp=sharing) for inference on Google Colab's free tier GPUs!
356

357
</Tip>
358

359
## 馃 Optimum
360

361
362
<Tip>

363
Learn more details about using ORT with 馃 Optimum in the [Accelerated inference on NVIDIA GPUs](https://huggingface.co/docs/optimum/onnxruntime/usage_guides/gpu#accelerated-inference-on-nvidia-gpus) and [Accelerated inference on AMD GPUs](https://huggingface.co/docs/optimum/onnxruntime/usage_guides/amdgpu#accelerated-inference-on-amd-gpus) guides. This section only provides a brief and simple example.
364
365
366

</Tip>

367
ONNX Runtime (ORT) is a model accelerator that supports accelerated inference on Nvidia GPUs, and AMD GPUs that use [ROCm](https://www.amd.com/en/products/software/rocm.html) stack. ORT uses optimization techniques like fusing common operations into a single node and constant folding to reduce the number of computations performed and speedup inference. ORT also places the most computationally intensive operations on the GPU and the rest on the CPU to intelligently distribute the workload between the two devices.
368

369
ORT is supported by 馃 Optimum which can be used in 馃 Transformers. You'll need to use an [`~optimum.onnxruntime.ORTModel`] for the task you're solving, and specify the `provider` parameter which can be set to either [`CUDAExecutionProvider`](https://huggingface.co/docs/optimum/onnxruntime/usage_guides/gpu#cudaexecutionprovider), [`ROCMExecutionProvider`](https://huggingface.co/docs/optimum/onnxruntime/usage_guides/amdgpu) or [`TensorrtExecutionProvider`](https://huggingface.co/docs/optimum/onnxruntime/usage_guides/gpu#tensorrtexecutionprovider). If you want to load a model that was not yet exported to ONNX, you can set `export=True` to convert your model on-the-fly to the ONNX format:
370
371
372
373
374

```py
from optimum.onnxruntime import ORTModelForSequenceClassification

ort_model = ORTModelForSequenceClassification.from_pretrained(
375
  "distilbert/distilbert-base-uncased-finetuned-sst-2-english",
376
377
378
379
  export=True,
  provider="CUDAExecutionProvider",
)
```
380

381
382
383
384
385
386
Now you're free to use the model for inference:

```py
from optimum.pipelines import pipeline
from transformers import AutoTokenizer

387
tokenizer = AutoTokenizer.from_pretrained("distilbert/distilbert-base-uncased-finetuned-sst-2-english")
388
389
390
391
392
393
394
395

pipeline = pipeline(task="text-classification", model=ort_model, tokenizer=tokenizer, device="cuda:0")
result = pipeline("Both the music and visual were astounding, not to mention the actors performance.")
```

## Combine optimizations

It is often possible to combine several of the optimization techniques described above to get the best inference performance possible for your model. For example, you can load a model in 4-bit, and then enable BetterTransformer with FlashAttention:
396
397
398
399
400

```py
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig

401
# load model in 4-bit
402
403
404
405
406
407
408
409
quantization_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_compute_dtype=torch.float16
)

tokenizer = AutoTokenizer.from_pretrained("facebook/opt-350m")
model = AutoModelForCausalLM.from_pretrained("facebook/opt-350m", quantization_config=quantization_config)

410
411
412
# enable BetterTransformer
model = model.to_bettertransformer()

413
414
415
input_text = "Hello my dog is cute and"
inputs = tokenizer(input_text, return_tensors="pt").to("cuda")

416
# enable FlashAttention
417
418
419
420
with torch.backends.cuda.sdp_kernel(enable_flash=True, enable_math=False, enable_mem_efficient=False):
    outputs = model.generate(**inputs)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))
421
```