Skip to content

PyTorch NLP BERT-base

This tutorial aims to introduce how to compile and deploy the natural language processing model BERT for a masked language modeling task provided by Hugging Face. The model predicts the most probable word to fill the masked position of the given sentence.

The tutorial consists of two main steps:

  1. How to compile the PyTorch BERT-base model and save to local storage
  2. How to deploy the compiled model in the runtime-based inference environment

Prerequisite

Before we proceed, please make sure you have installed the following pip packages in your system:

Note

If you want to skip the details and quickly compile and deploy the models on RBLN NPU, you can directly jump to the summary section in this tutorial. The code summarized in this section includes all the necessary steps required to compile and deploy the model so it can be used as a quick starting point for your own project.

Step 1. How to compile

We will demonstrate how to compile the Hugging Face BERT-base model.

Prepare the model

To begin, we will import the BertForMaskedLM model from the transformers library.

1
2
3
4
5
6
7
import torch
from transformers import BertForMaskedLM
import rebel  # RBLN Compiler

# Instantiate HuggingFace PyTorch BERT-base model
bert_model = BertForMaskedLM.from_pretrained("bert-base-uncased", return_dict=False)
bert_model.eval()

Compile the model

Once the model torch.nn.Module is instantiated, we can simply compile it with rebel.compile_from_torch() method.

# Compile the model
MAX_SEQ_LEN = 128
input_info = [
    ("input_ids", [1, MAX_SEQ_LEN], "int64"),
    ("attention_mask", [1, MAX_SEQ_LEN], "int64"),
    ("token_type_ids", [1, MAX_SEQ_LEN], "int64"),
]
compiled_model = rebel.compile_from_torch(
    bert_model,
    input_info,
    # If the NPU is installed on your host machine, you can omit the `npu` argument.
    # The function will automatically detect and use the installed NPU.
    npu="RBLN-CA12",
)

If the NPU is installed on your host machine, you can omit the npu argument in the rebel.compile_from_torch() function. In this case, the function will automatically detect and use the installed NPU. However, if the NPU is not installed on your host machine, you need to specify the target NPU using the npu argument to avoid any errors.

Currently, there are two supported NPU names: RBLN-CA02, RBLN-CA12. If you are unsure about the name of your target NPU, you can check it by running the rbln-stat command in the shell on the host machine where the NPU is installed.

Save the compiled model

To save the compiled model to local storage, we can use the compiled_model.save() method as below.

# Save the compiled model to local storage
compiled_model.save("bert_base.rbln")

Step 2. How to deploy

In this section, we will learn how to load the compiled model, run inference on RBLN NPU, and check results inferred from the model.

Prepare the input

First, we need to preprocess the input text sequence for the masked language modeling task. We will use BertTokenizer from the transformers library to tokenize the input sequence.

1
2
3
4
5
6
7
8
9
import torch
from transformers import BertTokenizer, pipeline
import rebel  # RBLN Runtime

# Prepare the input
tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
text = "the color of rose is [MASK]." 
MAX_SEQ_LEN = 128
inputs = tokenizer(text, return_tensors="pt", padding="max_length", max_length=MAX_SEQ_LEN)

Run inference

The RBLN Runtime module rebel.Runtime() is used to load the compiled model by passing the path of the saved model as an input argument. The tensor_type argument in rebel.Runtime() specifies the type of tensor to be used for input and output data. It can be set to either "pt" for PyTorch tensors or "np" for NumPy arrays.

We can use the run() method of the instantiated runtime module rebel.Runtime() for running inference. Additionally, the forward() method and the __call__ magic method can also be used to run inference, maintaining compatibility with PyTorch's interface.

1
2
3
4
5
# Load the compiled model
module = rebel.Runtime("bert_base.rbln", tensor_type="pt")

# Run inference
out = module(**inputs)

You can see fundamental information of the runtime module, such as input/output shapes and the compiled model size, by using the print(module) function.

Check results

To decode the final logits to text, we will use fill-mask pipeline from the transformers library. It generates the most probable word to fill in [MASK] of the given sentence.

1
2
3
# Check results
unmasker = pipeline("fill-mask", model="bert-base-uncased", framework="pt")
print(unmasker.postprocess({"input_ids": inputs.input_ids, "logits": out}))

The results will look like:

1
2
3
4
5
6
7
[
    {'score': 0.23562490940093994, 'token': 2317, 'token_str': 'white', 'sequence': 'the color of rose is white.'},
    {'score': 0.10957575589418411, 'token': 2417, 'token_str': 'red', 'sequence': 'the color of rose is red.'},
    {'score': 0.08016733080148697, 'token': 2304, 'token_str': 'black', 'sequence': 'the color of rose is black.'},
    {'score': 0.07074742764234543, 'token': 3756, 'token_str': 'yellow', 'sequence': 'the color of rose is yellow.'},
    {'score': 0.05175992473959923, 'token': 2630, 'token_str': 'blue', 'sequence': 'the color of rose is blue.'},
]

Summary

The complete code for model compilation is:

import torch
from transformers import BertForMaskedLM
import rebel  # RBLN Compiler

# Instantiate HuggingFace PyTorch BERT-base model
bert_model = BertForMaskedLM.from_pretrained("bert-base-uncased", return_dict=False)
bert_model.eval()

# Compile the model
MAX_SEQ_LEN = 128
input_info = [
    ("input_ids", [1, MAX_SEQ_LEN], "int64"),
    ("attention_mask", [1, MAX_SEQ_LEN], "int64"),
    ("token_type_ids", [1, MAX_SEQ_LEN], "int64"),
]
compiled_model = rebel.compile_from_torch(
    bert_model,
    input_info,
    # If the NPU is installed on your host machine, you can omit the `npu` argument.
    # The function will automatically detect and use the installed NPU.
    npu="RBLN-CA12",
)

# Save the compiled model to local storage
compiled_model.save("bert_base.rbln")

The completed code for inference of the compiled model is as follows:

import torch
from transformers import BertTokenizer, pipeline
import rebel  # RBLN Runtime

# Prepare the input
MAX_SEQ_LEN = 128
tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
text = "the color of rose is [MASK]." 
inputs = tokenizer(text, return_tensors="pt", padding="max_length", max_length=MAX_SEQ_LEN)

# Load the compiled model
module = rebel.Runtime("bert_base.rbln", tensor_type="pt")

# Run inference
out = module(**inputs)

# Check results
unmasker = pipeline("fill-mask", model="bert-base-uncased", framework="pt")
print(unmasker.postprocess({"input_ids": inputs.input_ids, "logits": out}))