Pruning Bert on Task MNLI

Workable Pruning Process

Here we show an effective transformer pruning process that NNI team has tried, and users can use NNI to discover better processes.

The entire pruning process can be divided into the following steps:

  1. Finetune the pre-trained model on the downstream task. From our experience, the final performance of pruning on the finetuned model is better than pruning directly on the pre-trained model. At the same time, the finetuned model obtained in this step will also be used as the teacher model for the following distillation training.

  2. Pruning the attention layer at first. Here we apply block-sparse on attention layer weight, and directly prune the head (condense the weight) if the head was fully masked. If the head was partially masked, we will not prune it and recover its weight.

  3. Retrain the head-pruned model with distillation. Recover the model precision before pruning FFN layer.

  4. Pruning the FFN layer. Here we apply the output channels pruning on the 1st FFN layer, and the 2nd FFN layer input channels will be pruned due to the pruning of 1st layer output channels.

  5. Retrain the final pruned model with distillation.

During the process of pruning transformer, we gained some of the following experiences:

  • We using Movement Pruner in step 2 and Taylor FO Weight Pruner in step 4. Movement Pruner has good performance on attention layers, and Taylor FO Weight Pruner method has good performance on FFN layers. These two pruners are all some kinds of gradient-based pruning algorithms, we also try weight-based pruning algorithms like L1 Norm Pruner, but it doesn’t seem to work well in this scenario.

  • Distillation is a good way to recover model precision. In terms of results, usually 1~2% improvement in accuracy can be achieved when we prune bert on mnli task.

  • It is necessary to gradually increase the sparsity rather than reaching a very high sparsity all at once.


The complete pruning process will take about 8 hours on one A100.


This section is mainly to get a finetuned model on the downstream task. If you are familiar with how to finetune Bert on GLUE dataset, you can skip this section.


Please set dev_mode to False to run this tutorial. Here dev_mode is True by default is for generating documents.

dev_mode = True

Some basic setting.

from pathlib import Path
from typing import Callable, Dict

pretrained_model_name_or_path = 'bert-base-uncased'
task_name = 'mnli'
experiment_id = 'pruning_bert_mnli'

# heads_num and layers_num should align with pretrained_model_name_or_path
heads_num = 12
layers_num = 12

# used to save the experiment log
log_dir = Path(f'./pruning_log/{pretrained_model_name_or_path}/{task_name}/{experiment_id}')
log_dir.mkdir(parents=True, exist_ok=True)

# used to save the finetuned model and share between different experiemnts with same pretrained_model_name_or_path and task_name
model_dir = Path(f'./models/{pretrained_model_name_or_path}/{task_name}')
model_dir.mkdir(parents=True, exist_ok=True)

# used to save GLUE data
data_dir = Path(f'./data')
data_dir.mkdir(parents=True, exist_ok=True)

# set seed
from transformers import set_seed

import torch
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

Create dataloaders.

from import DataLoader

from datasets import load_dataset
from transformers import BertTokenizerFast, DataCollatorWithPadding

task_to_keys = {
    'cola': ('sentence', None),
    'mnli': ('premise', 'hypothesis'),
    'mrpc': ('sentence1', 'sentence2'),
    'qnli': ('question', 'sentence'),
    'qqp': ('question1', 'question2'),
    'rte': ('sentence1', 'sentence2'),
    'sst2': ('sentence', None),
    'stsb': ('sentence1', 'sentence2'),
    'wnli': ('sentence1', 'sentence2'),

def prepare_dataloaders(cache_dir=data_dir, train_batch_size=32, eval_batch_size=32):
    tokenizer = BertTokenizerFast.from_pretrained(pretrained_model_name_or_path)
    sentence1_key, sentence2_key = task_to_keys[task_name]
    data_collator = DataCollatorWithPadding(tokenizer)

    # used to preprocess the raw data
    def preprocess_function(examples):
        # Tokenize the texts
        args = (
            (examples[sentence1_key],) if sentence2_key is None else (examples[sentence1_key], examples[sentence2_key])
        result = tokenizer(*args, padding=False, max_length=128, truncation=True)

        if 'label' in examples:
            # In all cases, rename the column to labels because the model will expect that.
            result['labels'] = examples['label']
        return result

    raw_datasets = load_dataset('glue', task_name, cache_dir=cache_dir)
    for key in list(raw_datasets.keys()):
        if 'test' in key:

    processed_datasets =, batched=True,

    train_dataset = processed_datasets['train']
    if task_name == 'mnli':
        validation_datasets = {
            'validation_matched': processed_datasets['validation_matched'],
            'validation_mismatched': processed_datasets['validation_mismatched']
        validation_datasets = {
            'validation': processed_datasets['validation']

    train_dataloader = DataLoader(train_dataset, shuffle=True, collate_fn=data_collator, batch_size=train_batch_size)
    validation_dataloaders = {
        val_name: DataLoader(val_dataset, collate_fn=data_collator, batch_size=eval_batch_size) \
            for val_name, val_dataset in validation_datasets.items()

    return train_dataloader, validation_dataloaders

train_dataloader, validation_dataloaders = prepare_dataloaders()

Training function & evaluation function.

import functools
import time

import torch.nn.functional as F
from datasets import load_metric
from transformers.modeling_outputs import SequenceClassifierOutput

def training(model: torch.nn.Module,
             optimizer: torch.optim.Optimizer,
             criterion: Callable[[torch.Tensor, torch.Tensor], torch.Tensor],
             lr_scheduler: torch.optim.lr_scheduler._LRScheduler = None,
             max_steps: int = None,
             max_epochs: int = None,
             train_dataloader: DataLoader = None,
             distillation: bool = False,
             teacher_model: torch.nn.Module = None,
             distil_func: Callable = None,
             log_path: str = Path(log_dir) / 'training.log',
             save_best_model: bool = False,
             save_path: str = None,
             evaluation_func: Callable = None,
             eval_per_steps: int = 1000,

    assert train_dataloader is not None

    if teacher_model is not None:
    current_step = 0
    best_result = 0

    total_epochs = max_steps // len(train_dataloader) + 1 if max_steps else max_epochs if max_epochs else 3
    total_steps = max_steps if max_steps else total_epochs * len(train_dataloader)

    print(f'Training {total_epochs} epochs, {total_steps} steps...')

    for current_epoch in range(total_epochs):
        for batch in train_dataloader:
            if current_step >= total_steps:
            outputs = model(**batch)
            loss = outputs.loss

            if distillation:
                assert teacher_model is not None
                with torch.no_grad():
                    teacher_outputs = teacher_model(**batch)
                distil_loss = distil_func(outputs, teacher_outputs)
                loss = 0.1 * loss + 0.9 * distil_loss

            loss = criterion(loss, None)

            # per step schedule
            if lr_scheduler:

            current_step += 1

            if current_step % eval_per_steps == 0 or current_step % len(train_dataloader) == 0:
                result = evaluation_func(model) if evaluation_func else None
                with (log_path).open('a+') as f:
                    msg = '[{}] Epoch {}, Step {}: {}\n'.format(time.asctime(time.localtime(time.time())), current_epoch, current_step, result)
                # if it's the best model, save it.
                if save_best_model and (result is None or best_result < result['default']):
                    assert save_path is not None
          , save_path)
                    best_result = None if result is None else result['default']

def distil_loss_func(stu_outputs: SequenceClassifierOutput, tea_outputs: SequenceClassifierOutput, encoder_layer_idxs=[]):
    encoder_hidden_state_loss = []
    for i, idx in enumerate(encoder_layer_idxs[:-1]):
        encoder_hidden_state_loss.append(F.mse_loss(stu_outputs.hidden_states[i], tea_outputs.hidden_states[idx]))
    logits_loss = F.kl_div(F.log_softmax(stu_outputs.logits / 2, dim=-1), F.softmax(tea_outputs.logits / 2, dim=-1), reduction='batchmean') * (2 ** 2)

    distil_loss = 0
    for loss in encoder_hidden_state_loss:
        distil_loss += loss
    distil_loss += logits_loss
    return distil_loss

def evaluation(model: torch.nn.Module, validation_dataloaders: Dict[str, DataLoader] = None, device=None):
    assert validation_dataloaders is not None
    training =

    is_regression = task_name == 'stsb'
    metric = load_metric('glue', task_name)

    result = {}
    default_result = 0
    for val_name, validation_dataloader in validation_dataloaders.items():
        for batch in validation_dataloader:
            outputs = model(**batch)
            predictions = outputs.logits.argmax(dim=-1) if not is_regression else outputs.logits.squeeze()
        result[val_name] = metric.compute()
        default_result += result[val_name].get('f1', result[val_name].get('accuracy', 0))
    result['default'] = default_result / len(result)

    return result

evaluation_func = functools.partial(evaluation, validation_dataloaders=validation_dataloaders, device=device)

def fake_criterion(loss, _):
    return loss

Prepare pre-trained model and finetuning on downstream task.

from torch.optim import Adam
from torch.optim.lr_scheduler import LambdaLR
from transformers import BertForSequenceClassification

def create_pretrained_model():
    is_regression = task_name == 'stsb'
    num_labels = 1 if is_regression else (3 if task_name == 'mnli' else 2)
    model = BertForSequenceClassification.from_pretrained(pretrained_model_name_or_path, num_labels=num_labels)
    model.bert.config.output_hidden_states = True
    return model

def create_finetuned_model():
    finetuned_model = create_pretrained_model()
    finetuned_model_state_path = Path(model_dir) / 'finetuned_model_state.pth'

    if finetuned_model_state_path.exists():
        finetuned_model.load_state_dict(torch.load(finetuned_model_state_path, map_location='cpu'))
    elif dev_mode:
        steps_per_epoch = len(train_dataloader)
        training_epochs = 3
        optimizer = Adam(finetuned_model.parameters(), lr=3e-5, eps=1e-8)

        def lr_lambda(current_step: int):
            return max(0.0, float(training_epochs * steps_per_epoch - current_step) / float(training_epochs * steps_per_epoch))

        lr_scheduler = LambdaLR(optimizer, lr_lambda)
        training(finetuned_model, optimizer, fake_criterion, lr_scheduler=lr_scheduler,
                 max_epochs=training_epochs, train_dataloader=train_dataloader, log_path=log_dir / 'finetuning_on_downstream.log',
                 save_best_model=True, save_path=finetuned_model_state_path, evaluation_func=evaluation_func, device=device)
    return finetuned_model

finetuned_model = create_finetuned_model()


According to experience, it is easier to achieve good results by pruning the attention part and the FFN part in stages. Of course, pruning together can also achieve the similar effect, but more parameter adjustment attempts are required. So in this section, we do pruning in stages.

First, we prune the attention layer with MovementPruner.

steps_per_epoch = len(train_dataloader)

# Set training steps/epochs for pruning.

if not dev_mode:
    total_epochs = 4
    total_steps = total_epochs * steps_per_epoch
    warmup_steps = 1 * steps_per_epoch
    cooldown_steps = 1 * steps_per_epoch
    total_epochs = 1
    total_steps = 3
    warmup_steps = 1
    cooldown_steps = 1

# Initialize evaluator used by MovementPruner.

import nni
from nni.algorithms.compression.v2.pytorch import TorchEvaluator

movement_training = functools.partial(training, train_dataloader=train_dataloader,
                                      log_path=log_dir / 'movement_pruning.log',
                                      evaluation_func=evaluation_func, device=device)
traced_optimizer = nni.trace(Adam)(finetuned_model.parameters(), lr=3e-5, eps=1e-8)

def lr_lambda(current_step: int):
    if current_step < warmup_steps:
        return float(current_step) / warmup_steps
    return max(0.0, float(total_steps - current_step) / float(total_steps - warmup_steps))

traced_scheduler = nni.trace(LambdaLR)(traced_optimizer, lr_lambda)
evaluator = TorchEvaluator(movement_training, traced_optimizer, fake_criterion, traced_scheduler)

# Apply block-soft-movement pruning on attention layers.
# Note that block sparse is introduced by `sparse_granularity='auto'`, and only support `bert`, `bart`, `t5` right now.

from nni.compression.pytorch.pruning import MovementPruner

config_list = [{
    'op_types': ['Linear'],
    'op_partial_names': ['bert.encoder.layer.{}.attention'.format(i) for i in range(layers_num)],
    'sparsity': 0.1

pruner = MovementPruner(model=finetuned_model,
                        cool_down_beginning_step=total_steps - cooldown_steps,
_, attention_masks = pruner.compress()
pruner.show_pruned_weights(), Path(log_dir) / 'attention_masks.pth')

Load a new finetuned model to do speedup, you can think of this as using the finetuned state to initialize the pruned model weights. Note that nni speedup don’t support replacing attention module, so here we manully replace the attention module.

If the head is entire masked, physically prune it and create config_list for FFN pruning.

attention_pruned_model = create_finetuned_model().to(device)
attention_masks = torch.load(Path(log_dir) / 'attention_masks.pth')

ffn_config_list = []
layer_remained_idxs = []
module_list = []
for i in range(0, layers_num):
    prefix = f'bert.encoder.layer.{i}.'
    value_mask: torch.Tensor = attention_masks[prefix + 'attention.self.value']['weight']
    head_mask = (value_mask.reshape(heads_num, -1).sum(-1) == 0.)
    head_idxs = torch.arange(len(head_mask))[head_mask].long().tolist()
    print(f'layer {i} prune {len(head_idxs)} head: {head_idxs}')
    if len(head_idxs) != heads_num:
        # The final ffn weight remaining ratio is the half of the attention weight remaining ratio.
        # This is just an empirical configuration, you can use any other method to determine this sparsity.
        sparsity = 1 - (1 - len(head_idxs) / heads_num) * 0.5
        # here we use a simple sparsity schedule, we will prune ffn in 12 iterations, each iteration prune `sparsity_per_iter`.
        sparsity_per_iter = 1 - (1 - sparsity) ** (1 / 12)
            'op_names': [f'bert.encoder.layer.{len(layer_remained_idxs)}.intermediate.dense'],
            'sparsity': sparsity_per_iter

attention_pruned_model.bert.encoder.layer = torch.nn.ModuleList(module_list)
distil_func = functools.partial(distil_loss_func, encoder_layer_idxs=layer_remained_idxs)

Retrain the attention pruned model with distillation.

if not dev_mode:
    total_epochs = 5
    total_steps = None
    distillation = True
    total_epochs = 1
    total_steps = 1
    distillation = False

teacher_model = create_finetuned_model()
optimizer = Adam(attention_pruned_model.parameters(), lr=3e-5, eps=1e-8)

def lr_lambda(current_step: int):
    return max(0.0, float(total_epochs * steps_per_epoch - current_step) / float(total_epochs * steps_per_epoch))

lr_scheduler = LambdaLR(optimizer, lr_lambda)
at_model_save_path = log_dir / 'attention_pruned_model_state.pth'
training(attention_pruned_model, optimizer, fake_criterion, lr_scheduler=lr_scheduler, max_epochs=total_epochs,
         max_steps=total_steps, train_dataloader=train_dataloader, distillation=distillation, teacher_model=teacher_model,
         distil_func=distil_func, log_path=log_dir / 'retraining.log', save_best_model=True, save_path=at_model_save_path,
         evaluation_func=evaluation_func, device=device)

if not dev_mode:

Iterative pruning FFN with TaylorFOWeightPruner in 12 iterations. Finetuning 3000 steps after each pruning iteration, then finetuning 2 epochs after pruning finished.

NNI will support per-step-pruning-schedule in the future, then can use an pruner to replace the following code.

if not dev_mode:
    total_epochs = 7
    total_steps = None
    taylor_pruner_steps = 1000
    steps_per_iteration = 3000
    total_pruning_steps = 36000
    distillation = True
    total_epochs = 1
    total_steps = 6
    taylor_pruner_steps = 2
    steps_per_iteration = 2
    total_pruning_steps = 4
    distillation = False

from nni.compression.pytorch.pruning import TaylorFOWeightPruner
from nni.compression.pytorch.speedup import ModelSpeedup

distil_training = functools.partial(training, train_dataloader=train_dataloader, distillation=distillation,
                                    teacher_model=teacher_model, distil_func=distil_func, device=device)
traced_optimizer = nni.trace(Adam)(attention_pruned_model.parameters(), lr=3e-5, eps=1e-8)
evaluator = TorchEvaluator(distil_training, traced_optimizer, fake_criterion)

current_step = 0
best_result = 0
init_lr = 3e-5

dummy_input = torch.rand(8, 128, 768).to(device)

for current_epoch in range(total_epochs):
    for batch in train_dataloader:
        if total_steps and current_step >= total_steps:
        # pruning with TaylorFOWeightPruner & reinitialize optimizer
        if current_step % steps_per_iteration == 0 and current_step < total_pruning_steps:
            check_point = attention_pruned_model.state_dict()
            pruner = TaylorFOWeightPruner(attention_pruned_model, ffn_config_list, evaluator, taylor_pruner_steps)
            _, ffn_masks = pruner.compress()
            renamed_ffn_masks = {}
            # rename the masks keys, because we only speedup the bert.encoder
            for model_name, targets_mask in ffn_masks.items():
                renamed_ffn_masks[model_name.split('bert.encoder.')[1]] = targets_mask
            ModelSpeedup(attention_pruned_model.bert.encoder, dummy_input, renamed_ffn_masks).speedup_model()
            optimizer = Adam(attention_pruned_model.parameters(), lr=init_lr)
        # manually schedule lr
        for params_group in optimizer.param_groups:
            params_group['lr'] = (1 - current_step / (total_epochs * steps_per_epoch)) * init_lr

        outputs = attention_pruned_model(**batch)
        loss = outputs.loss

        # distillation
        if distillation:
            assert teacher_model is not None
            with torch.no_grad():
                teacher_outputs = teacher_model(**batch)
            distil_loss = distil_func(outputs, teacher_outputs)
            loss = 0.1 * loss + 0.9 * distil_loss


        current_step += 1

        if current_step % 1000 == 0 or current_step % len(train_dataloader) == 0:
            result = evaluation_func(attention_pruned_model)
            with (log_dir / 'ffn_pruning.log').open('a+') as f:
                msg = '[{}] Epoch {}, Step {}: {}\n'.format(time.asctime(time.localtime(time.time())),
                                                            current_epoch, current_step, result)
            if current_step >= total_pruning_steps and best_result < result['default']:
      , log_dir / 'best_model.pth')
                best_result = result['default']


The speedup is test on the entire validation dataset with batch size 128 on A100. We test under two pytorch version and found the latency varying widely.

Setting 1: pytorch 1.12.1

Setting 2: pytorch 1.10.0

Prune Bert-base-uncased on MNLI

Attention Pruning Method

FFN Pruning Method

Total Sparsity


Acc. Drop

Speedup (S1)

Speedup (S2)

85.1M (-0.0%)

84.85 / 85.28

+0.0 / +0.0

25.60s (x1.00)

8.10s (x1.00)

Movement Pruner (soft, sparsity=0.1, regular_scale=1)

Taylor FO Weight Pruner

54.1M (-36.43%)

85.38 / 85.41

+0.53 / +0.13

17.93s (x1.43)

7.22s (x1.12)

Movement Pruner (soft, sparsity=0.1, regular_scale=5)

Taylor FO Weight Pruner

37.1M (-56.40%)

84.73 / 85.12

-0.12 / -0.16

12.83s (x2.00)

5.61s (x1.44)

Movement Pruner (soft, sparsity=0.1, regular_scale=10)

Taylor FO Weight Pruner

24.1M (-71.68%)

84.14 / 84.78

-0.71 / -0.50

8.93s (x2.87)

4.55s (x1.78)

Movement Pruner (soft, sparsity=0.1, regular_scale=20)

Taylor FO Weight Pruner

14.3M (-83.20%)

83.26 / 82.96

-1.59 / -2.32

5.98s (x4.28)

3.56s (x2.28)

Movement Pruner (soft, sparsity=0.1, regular_scale=30)

Taylor FO Weight Pruner

9.9M (-88.37%)

82.22 / 82.19

-2.63 / -3.09

4.36s (x5.88)

3.12s (x2.60)

Movement Pruner (soft, sparsity=0.1, regular_scale=40)

Taylor FO Weight Pruner

8.8M (-89.66%)

81.64 / 82.39

-3.21 / -2.89

3.88s (x6.60)

2.81s (x2.88)

Total running time of the script: ( 0 minutes 20.822 seconds)

Gallery generated by Sphinx-Gallery