Deploy a Hugging Face Pruned Model on CPU

Author: Josh Fromm

This tutorial demonstrates how to take any pruned model, in this case PruneBert from Hugging Face, and use TVM to leverage the model’s sparsity support to produce real speedups. Although the primary purpose of this tutorial is to realize speedups on already pruned models, it may also be useful to estimate how fast a model would be if it were pruned. To this end, we also provide a function that takes an unpruned model and replaces its weights with random and pruned weights at a specified sparsity. This may be a useful feature when trying to decide if a model is worth pruning or not.

Before we get into the code, it’s useful to discuss sparsity and pruning and dig into the two different types of sparsity: structured and unstructured.

Pruning is a technique primarily used to reduce the parameter size of a model by replacing weight values with 0s. Although many methods exist for choosing which weights should be set to 0, the most straight forward is by picking the weights with the smallest value. Typically, weights are pruned to a desired sparsity percentage. For example, a 95% sparse model would have only 5% of its weights non-zero. Pruning to very high sparsities often requires finetuning or full retraining as it tends to be a lossy approximation. Although parameter size benefits are quite easy to obtain from a pruned model through simple compression, leveraging sparsity to yield runtime speedups is more complicated.

In structured sparsity weights are pruned with the goal of clustering pruned weights together. In other words, they are pruned using both their value and location. The benefit of bunching up pruned weights is that it allows an algorithm such as matrix multiplication to skip entire blocks. It turns out that some degree of block sparsity is very important to realizing significant speedups on most hardware available today. This is because when loading memory in most CPUs or GPUs, it doesn’t save any work to skip reading a single value at a time, instead an entire chunk or tile is read in and executed using something like vectorized instructions.

Unstructured sparse weights are those that are pruned only on the value of the original weights. They may appear to be scattered randomly throughout a tensor rather than in chunks like we’d see in block sparse weights. At low sparsities, unstructured pruning techniques are difficult to accelerate. However, at high sparsities many blocks of all 0 values will naturally appear, making it possible to accelerate.

This tutorial interacts with both structured and unstructured sparsity. Hugging Face’s PruneBert model is unstructured but 95% sparse, allowing us to apply TVM’s block sparse optimizations to it, even if not optimally. When generating random sparse weights for an unpruned model, we do so with structured sparsity. A fun exercise is comparing the real speed of PruneBert with the block sparse speed using fake weights to see the benefit of structured sparsity.

Load Required Modules

Other than TVM, scipy, the latest transformers, and tensorflow 2.2+ are required.

import os
import tvm
import time
import itertools
import numpy as np
import tensorflow as tf
from tvm import relay
from tvm.contrib import graph_runtime
from tvm.relay import data_dep_optimization as ddo
from tensorflow.python.framework.convert_to_constants import (
import scipy.sparse as sp

Configure Settings

Let’s start by defining some parameters that define the type of model and sparsity to run.

# The name of the transformer model to download and run.
name = "huggingface/prunebert-base-uncased-6-finepruned-w-distil-squad"
# The number of batches in an input.
batch_size = 1
# The length of each input sequence.
seq_len = 128
# TVM platform identifier. Although cuda is also supported, it requires
# tuning that is outside the scope of this tutorial. Note that best
# cpu performance can be achieved by setting -mcpu appropriately for
# your specific machine.
target = "llvm"
# Which device to run on. Should be one of tvm.cpu() or tvm.gpu().
ctx = tvm.cpu()
# If true, then a sparse variant of the network will be run and
# benchmarked.
measure_sparse = True
# The block size of structured sparsity to convert weight tensors
# into. Changing this parameter may yield speedups for some platforms.
bs_r = 1
# For models besides PruneBert (which is 95% sparse), this parameter
# determines how sparse the generated weights should be. The higher
# the sparsity, the faster the result.
sparsity = 0.85

Download and Convert Transformers Model

Now we’ll grab a model from the transformers module, download it, convert it into a TensorFlow graphdef in preperation for converting that graphdef into a relay graph that we can optimize and deploy.

def load_keras_model(module, name, seq_len, batch_size, report_runtime=True):
    model = module.from_pretrained(name)
    dummy_input = tf.keras.Input(shape=[seq_len], batch_size=batch_size, dtype="int32")
    dummy_out = model(dummy_input)  # Propagate shapes through the keras model.
    if report_runtime:
        np_input = np.random.uniform(size=[batch_size, seq_len], low=0, high=seq_len).astype(
        start = time.time()
        repeats = 50
        for i in range(repeats):
            np_out = model(np_input)
        end = time.time()
        print("Keras Runtime: %f ms." % (1000 * ((end - start) / repeats)))
    return model

def convert_to_graphdef(model, batch_size, seq_len):
    model_func = tf.function(lambda x: model(x))
    input_dict = model._saved_model_inputs_spec
    input_spec = input_dict[list(input_dict.keys())[0]]
    model_func = model_func.get_concrete_function(
        tf.TensorSpec([batch_size, seq_len], input_spec.dtype)
    frozen_func = convert_variables_to_constants_v2(model_func)
    return frozen_func.graph.as_graph_def()

def download_model(name, batch_size, seq_len):
    import transformers

    module = getattr(transformers, "TFBertForSequenceClassification")
    model = load_keras_model(module, name=name, batch_size=batch_size, seq_len=seq_len)
    return convert_to_graphdef(model, batch_size, seq_len)

Convert to Relay Graph

We now have all the tooling to get a transformers model in the right format for relay conversion. Let’s import it! In the following function we save the imported graph in relay’s json format so that we dont have to reimport from tensorflow each time this script is run.

def import_graphdef(
    abs_path = os.path.dirname(os.path.abspath(__file__))
    shape_dict = {"input_1": (batch_size, seq_len)}
    relay_file = ("%s_%d_%d_%s" % (name, batch_size, seq_len, relay_file)).replace("/", "_")
    relay_params = ("%s_%d_%d_%s" % (name, batch_size, seq_len, relay_params)).replace("/", "_")
    if os.path.exists(os.path.join(abs_path, relay_file)) and os.path.exists(
        os.path.join(abs_path, relay_params)
        with open(os.path.join(abs_path, relay_file), "r") as fi:
            mod =
        with open(os.path.join(abs_path, relay_params), "rb") as fi:
            params = relay.load_param_dict(
        graph_def = download_model(name, batch_size, seq_len)

        mod, params = relay.frontend.from_tensorflow(graph_def, shape=shape_dict)

        if save_relay:
            with open(os.path.join(abs_path, relay_file), "w") as fo:
            with open(os.path.join(abs_path, relay_params), "wb") as fo:

    return mod, params, shape_dict

Run the Dense Graph

Let’s run the default version of the imported model. Note that even if the weights are sparse, we won’t see any speedup because we are using regular dense matrix multiplications on these dense (but mostly zero) tensors instead of sparse aware kernels.

def run_relay_graph(mod, params, shape_dict, target, ctx):
    with relay.build_config(opt_level=3):
        lib =, target=target, params=params)
    input_shape = shape_dict["input_1"]
    dummy_data = np.random.uniform(size=input_shape, low=0, high=input_shape[1]).astype("int32")

    m = graph_runtime.GraphModule(lib["default"](ctx))
    m.set_input(0, dummy_data)
    tvm_output = m.get_output(0)

    ftimer = m.module.time_evaluator("run", ctx, repeat=5, number=5)
    prof_res = np.array(ftimer().results) * 1000
        "%-20s %-19s (%s)"
        % ("Runtime:", "%.2f ms" % np.mean(prof_res), "%.2f ms" % np.std(prof_res))
    return tvm_output

def run_dense(mod, params, shape_dict, target, ctx):
    print("Dense Model Benchmark:")
    return run_relay_graph(mod, params, shape_dict, target, ctx)

Run the Sparse Graph

Next we’ll convert the graph into a sparse representation and generate fake sparse weights if needed. Then we’ll use the same benchmarking script as dense to see how much faster we go! We apply a few relay passes to the graph to get it leveraging sparsity. First we use simplify_fc_transpose to use transposes on the weights of dense layers into the parameters. This makes it easier to convert to matrix multiplies to sparse versions. Next we apply bsr_dense.convert to identify all weight matrices that can be sparse, and automatically replace them.

The bsr_dense.convert call below is doing the heavy lifting of identifying which weights in the model can be made sparse by checking if they are at least sparsity_threshold percent sparse. If so, it converts those weights into Block Compressed Row Format (BSR). BSR is essentially a representation that indexes into the nonzero chunks of the tensor, making it easy for an algorithm to load those non-zero chunks and ignore the rest of the tensor. Once the sparse weights are in BSR format, relay.transform.DenseToSparse is applied to actually replace relay.dense operations with relay.sparse_dense calls that can be run faster.

def random_bsr_matrix(M, N, BS_R, BS_C, density, dtype="float32"):
    Y = np.zeros((M, N), dtype=dtype)
    assert M % BS_R == 0
    assert N % BS_C == 0
    nnz = int(density * M * N)
    num_blocks = int(nnz / (BS_R * BS_C)) + 1
    candidate_blocks = np.asarray(list(itertools.product(range(0, M, BS_R), range(0, N, BS_C))))
    assert candidate_blocks.shape[0] == M // BS_R * N // BS_C
    chosen_blocks = candidate_blocks[
        np.random.choice(candidate_blocks.shape[0], size=num_blocks, replace=False)
    for i in range(len(chosen_blocks)):
        r, c = chosen_blocks[i]
        Y[r : r + BS_R, c : c + BS_C] = np.random.uniform(-0.1, 0.1, (BS_R, BS_C))
    s = sp.bsr_matrix(Y, blocksize=(BS_R, BS_C))
    assert == (num_blocks, BS_R, BS_C)
    assert >= nnz
    assert s.indices.shape == (num_blocks,)
    assert s.indptr.shape == (M // BS_R + 1,)
    return s.todense()

def random_sparse_bert_params(func, params, density, BS_R, BS_C):
    def deepcopy(param_dic):
        ret = {}
        for k, v in param_dic.items():
            ret[k] = tvm.nd.array(v.asnumpy())
        return ret

    new_params = deepcopy(params)
    dense_weight_names = relay.analysis.sparse_dense._search_dense_op_weight(func)
    for item in dense_weight_names:
        name = str(item)
        shape = new_params[name].shape
        if shape[0] % BS_R == 0 and shape[1] % BS_C == 0:
            new_w = random_bsr_matrix(shape[0], shape[1], BS_R, BS_C, density)
            new_params[name] = tvm.nd.array(new_w)
    return new_params

def run_sparse(mod, params, shape_dict, target, ctx, bs_r, sparsity, gen_weights):
    mod, params = ddo.simplify_fc_transpose.convert(mod["main"], params)
    if gen_weights:
        params = random_sparse_bert_params(mod, params, BS_R=bs_r, BS_C=1, density=1 - sparsity)
    mod, params = ddo.bsr_dense.convert(mod, params, (bs_r, 1), sparsity_threshold=0.8)
    print("Block Sparse Model with {blocksize}x1 blocks:".format(blocksize=bs_r))
    return run_relay_graph(mod, params, shape_dict, target, ctx)

Run All the Code!

And that’s it! Now we’ll simply call all the needed function to benchmark the model according to the set parameters. Note that to run this code you’ll need to uncomment the last line first.

def benchmark():
    mod, params, shape_dict = import_graphdef(name, batch_size, seq_len)
    run_dense(mod, params, shape_dict, target, ctx)
    if measure_sparse:
        gen_weights = "prune" not in name
        run_sparse(mod, params, shape_dict, target, ctx, bs_r, sparsity, gen_weights)

# benchmark()

Sample Output

For reference, below is the output of the script when run on an AMD CPU and shows about a 2.5X speedup from using sparsity.

# Dense Model Benchmark:
# Cannot find config for target=llvm, workload=('dense_nopack.x86', ('TENSOR', (1, 768), 'float32'), ('TENSOR', (2, 768), 'float32'), None, 'float32'). A fallback configuration is used, which may bring great performance regression.
# Cannot find config for target=llvm, workload=('dense_nopack.x86', ('TENSOR', (1, 768), 'float32'), ('TENSOR', (768, 768), 'float32'), None, 'float32'). A fallback configuration is used, which may bring great performance regression.
# Cannot find config for target=llvm, workload=('dense_nopack.x86', ('TENSOR', (128, 3072), 'float32'), ('TENSOR', (768, 3072), 'float32'), None, 'float32'). A fallback configuration is used, which may bring great performance regression.
# Cannot find config for target=llvm, workload=('dense_nopack.x86', ('TENSOR', (128, 768), 'float32'), ('TENSOR', (3072, 768), 'float32'), None, 'float32'). A fallback configuration is used, which may bring great performance regression.
# Cannot find config for target=llvm, workload=('dense_nopack.x86', ('TENSOR', (128, 768), 'float32'), ('TENSOR', (768, 768), 'float32'), None, 'float32'). A fallback configuration is used, which may bring great performance regression.
# Cannot find config for target=llvm, workload=('batch_matmul.x86', ('TENSOR', (12, 128, 128), 'float32'), ('TENSOR', (12, 64, 128), 'float32')). A fallback configuration is used, which may bring great performance regression.
# Cannot find config for target=llvm, workload=('batch_matmul.x86', ('TENSOR', (12, 128, 64), 'float32'), ('TENSOR', (12, 128, 64), 'float32')). A fallback configuration is used, which may bring great performance regression.
# Runtime:             165.26 ms           (12.83 ms)
# Block Sparse Model with 1x1 blocks:
# Runtime:             67.75 ms            (8.83 ms)

Gallery generated by Sphinx-Gallery