tensorizer

Module, Model, and Tensor Serialization/Deserialization

TLDR

Extremely fast model loads from HTTP/HTTPS, Redis, and S3 endpoints. GPT-J (20GB) loads at wire-speed (~5GB/s) on a 40GbE network, and is only bottlenecked by the Linux kernel TCP stack.

Rationale

CoreWeave and our customers use KNative to deploy models as serverless functions. How long a model takes to load is a major factor in the latency of KNative scale-up. tensorizer is a tool to serialize models and their associated tensors into a single file that can be loaded quickly and efficiently off an HTTP/HTTPS or S3 endpoint.

By not embedding the model in the container image, we can reduce the container image size and the time it takes to load the model. This is especially important for models that are large in size, such as EleutherAI/gpt-neox-20B that weighs in at ~40GB.

This decoupling of the model from the container image also allows us to update the model without having to rebuild the container image. This allows us to quickly iterate on the model and deploy new versions without having to wait for the container image to build or for the container image cache to be populated.

tensorizer has S3 support, so we can store the serialized model in S3 object storage, and perform streaming loads from S3. This allows us to stream the model directly from S3 into the container without having to download the model to the container's local filesystem. This also pertains to HTTP/HTTPS endpoints, as S3 is just an HTTP/HTTPS endpoint.

tensorizer also has support for loading models from a local filesystem, so you can use it to serialize models locally and load them locally. This is extremely fast, as the same principles that make it fast for HTTP/HTTPS and S3 endpoints also apply to local filesystems.

tensorizer has preliminary support for Redis, but it is not recommended for model deployment due to the lack of distributed caching. It is intended for sharing state between inference pods, or for loading data on a per-request basis from a Redis cache.

Speed

tensorizer's deserialization speed is primarily network-bound.

The following graph presents data collected from the scripts and Kubernetes manifests in examples/benchmark_buffer_size comparing the various deserialization modes available in tensorizer release 2.5.0—along with the raw network speed, and the speed of torch.load().

Installation

From PyPI

tensorizer can be installed from PyPI with pip:

python -m pip install tensorizer

From Source

You can also install tensorizer from source using pip.

To clone the repository and install tensorizer in editable mode, run:

git clone https://github.com/coreweave/tensorizer
cd tensorizer
python -m pip install -e .

Or, run the following for pip to install tensorizer directly from GitHub:

python -m pip install git+https://github.com/coreweave/tensorizer

Basic Usage

Serialization is done with the TensorSerializer class. It takes a path_uri argument that can be a local filesystem path, an HTTP/HTTPS endpoint, or an S3 endpoint.

write_module is the main method of the TensorSerializer class. It takes a torch.nn.Module and serializes the tensors to the path_uri endpoint.

The below example serializes the EleutherAI/gpt-j-6B model to an S3 endpoint. It assumes that you have already configured your S3 credentials in ~/.s3cfg.

NOTE: Loading and serializing gpt-j-6B will take a lot of CPU RAM, up to ~20GB. Additionally, when loading gpt-j-6B into a GPU, you will need about ~16GB of VRAM. If you don't have that much RAM or VRAM, you can use the smaller gpt-neo-125M model instead.

NOTE2: The below examples require the transformers and accelerate libraries. You can install them with pip:

python -m pip install transformers accelerate

serialize.py

import torch
from tensorizer import TensorSerializer
from transformers import AutoModelForCausalLM

model_ref = "EleutherAI/gpt-j-6B"
# For less intensive requirements, swap above with the line below:
# model_ref = "EleutherAI/gpt-neo-125M"
model_name = model_ref.split("/")[-1]
# Change this to your S3 bucket.
s3_bucket = "bucket"
s3_uri = f"s3://{s3_bucket}/{model_name}.tensors"

model = AutoModelForCausalLM.from_pretrained(
    model_ref,
    revision="float16",
    torch_dtype=torch.float16,
    low_cpu_mem_usage=True,
)

serializer = TensorSerializer(s3_uri)
serializer.write_module(model)
serializer.close()

Conversely, deserialization is done with the TensorDeserializer class. It takes a path_uri argument that can be a local filesystem path, an HTTP/HTTPS endpoint, or an S3 endpoint.

load_into_module is the main method of the TensorDeserializer class. It takes a torch.nn.Module and loads the tensors from the path_uri endpoint into the torch.nn.Module.

The below example loads the EleutherAI/gpt-j-6B model from an S3 endpoint.

deserialize.py

import time
import torch
from tensorizer import TensorDeserializer
from tensorizer.utils import no_init_or_tensor, convert_bytes, get_mem_usage

from transformers import AutoModelForCausalLM, AutoTokenizer, AutoConfig

model_ref = "EleutherAI/gpt-j-6B"
# To run this at home, swap this with the line below for a smaller example:
# model_ref = "EleutherAI/gpt-neo-125M"
model_name = model_ref.split("/")[-1]
# Change this to your S3 bucket.
s3_bucket = "bucket"
s3_uri = f"s3://{s3_bucket}/{model_name}.tensors"

config = AutoConfig.from_pretrained(model_ref)

# This ensures that the model is not initialized.
with no_init_or_tensor():
    model = AutoModelForCausalLM.from_config(config)

before_mem = get_mem_usage()

# Lazy load the tensors from S3 into the model.
start = time.time()
deserializer = TensorDeserializer(s3_uri)
deserializer.load_into_module(model)
end = time.time()

# Brag about how fast we are.
total_bytes_str = convert_bytes(deserializer.total_tensor_bytes)
duration = end - start
per_second = convert_bytes(deserializer.total_tensor_bytes / duration)
after_mem = get_mem_usage()
deserializer.close()
print(f"Deserialized {total_bytes_str} in {end - start:0.2f}s, {per_second}/s")
print(f"Memory usage before: {before_mem}")
print(f"Memory usage after: {after_mem}")

# Tokenize and generate
model.eval()
tokenizer = AutoTokenizer.from_pretrained(model_ref)
eos = tokenizer.eos_token_id
input_ids = tokenizer.encode(
    "¡Hola! Encantado de conocerte. hoy voy a", return_tensors="pt"
).to("cuda")

with torch.no_grad():
    output = model.generate(
        input_ids, max_new_tokens=50, do_sample=True, pad_token_id=eos
    )

print(f"Output: {tokenizer.decode(output[0], skip_special_tokens=True)}")

It should produce output similar to the following, with GPT-J-6B:

Deserialized model in 6.25 seconds
Test Output: ¡Hola! Encantado de conocerte. hoy voy a comentar por primera
vez una teoría de trineo, que quizá te parezca
algo desconocido, ya que en este mundo han
llegado a dominar tantos

More practical examples for the usage of tensorizer can be found in examples/hf_serialization.py, where df_main() serializes models from HuggingFace Diffusers and hf_main() serializes HuggingFace Transformers models.

Tensor Weight Encryption

tensorizer supports fast tensor weight encryption and decryption during serialization and deserialization, respectively.

Be aware that metadata (tensor names, dtypes, shapes, etc.) are not encrypted, only the weights themselves.

To use tensorizer encryption, a recent version of libsodium must be installed. Install libsodium with apt-get install libsodium23 on Ubuntu or Debian, or follow the instructions in libsodium's documentation for other platforms.

Quick Encryption Example

The following outline demonstrates how to encrypt and decrypt a tensorized model with a randomly-generated encryption key:

from tensorizer import (
    EncryptionParams, DecryptionParams, TensorDeserializer, TensorSerializer
)

# Serialize and encrypt a model:

encryption_params = EncryptionParams.random()

serializer = TensorSerializer("model.tensors", encryption=encryption_params)
serializer.write_module(...)  # or write_state_dict(), etc.
serializer.close()

# Save the randomly-generated encryption key somewhere
with open("tensor.key", "wb") as key_file:
    key_file.write(encryption_params.key)


# Then decrypt it again:

# Load the randomly-generated key from where it was saved
with open("tensor.key", "rb") as key_file:
    key: bytes = key_file.read()
 
decryption_params = DecryptionParams.from_key(key)

deserializer = TensorDeserializer("model.tensors", encryption=decryption_params)
deserializer.load_into_module(...)
deserializer.close()

For more detail, refer to docs/encryption.md. A complete example is also available as examples/encryption.py. The EncryptionParams and DecryptionParams class docstrings additionally contain some usage information for quick reference from an IDE.

An example command line tool to add or remove encryption from existing serialized models is also available as examples/encryption.py.

Benchmarks

You can run your own benchmarks on CoreWeave or your own Kubernetes cluster by using the benchmark.yaml file in the examples/benchmark_buffer_size directory. Please see the README.

Available Pre-Tensorized Models on the CoreWeave Cloud

The following models are available on the CoreWeave Cloud for free, and can be used with the TensorDeserializer class. The S3 support defaults to the accel-object.ord1.coreweave.com endpoint, and the bucket to use as tensorized.

We name the keys in the S3 bucket after the HuggingFace model identifier, and append the /fp16 suffix for the half-precision version.

For example, the S3 URI for the EleutherAI/gpt-j-6B model is: s3://tensorized/EleutherAI/gpt-j-6B/fp16/model.tensors

The below table shows the available models and their S3 URIs.

Large Language Models

Generative Diffusion Models

S3 Usage Notes

tensorizer uses the boto3 library to interact with S3. The easiest way to use tensorizer with S3 is to configure your S3 credentials in ~/.s3cfg.

If you don't want to use ~/.s3cfg, or wish to use a .s3cfg config file saved at a nonstandard location (e.g. under /var/run), you can also specify your S3 credentials using the tensorizer.stream_io.open_stream() function, and then pass that into the TensorSerializer or TensorDeserializer constructor.

The stream_io.open_stream() function takes a path_uri argument, which can be an s3:// URI, and accepts the following keyword arguments:

• s3_access_key_id: S3 access key ID
• s3_secret_access_key: S3 secret access key
• s3_endpoint: S3 endpoint

Or,

• s3_config_path: Alternative filesystem path to a .s3cfg config file

For example:

TensorSerializer(
    open_stream(s3_uri,
                "wb",
                s3_access_key_id=ACCESS_KEY,
                s3_secret_access_key=SECRET_KEY,
                s3_endpoint="object.ord1.coreweave.com"))

and...

TensorDeserializer(
    open_stream(s3_uri,
                "rb",
                s3_access_key_id=ACCESS_KEY,
                s3_secret_access_key=SECRET_KEY,
                s3_endpoint="object.ord1.coreweave.com"))

NOTE: For faster object downloads in the CoreWeave Cloud, you can use the accel-object.ord1.coreweave.com endpoint. This endpoint is optimized for object downloads, and will be faster than the object.ord1.coreweave.com endpoint once the object is cached.

NOTE2: The cache above does not get invalidated when the object is updated in S3. If you update an object in S3, you will need to wait for the cache to expire before you can download the updated object. This takes 24 hours since the last download.

For this reason, it is recommended to use a unique S3 key for each version of a model if you use the accel-object.ord1.coreweave.com endpoint.

Additional Features

tensorizer has a few additional features that make it more useful than just a serialization/deserialization tool.

Plaid Mode

tensorizer has a plaid_mode argument that can be passed to the TensorDeserializer class when deserializing tensors to the GPU. When plaid_mode is True, tensorizer will load tensors extremely fast by reusing a single small CPU buffer as a staging area to load tensors before transferring them to the GPU through DMA. This can also greatly decrease CPU RAM usage during deserialization.

plaid_mode is enabled by default when available.

state_dict Support

The TensorDeserializer object can be used as-is as a state_dict for torch.nn.Module.load_state_dict. This is useful for loading the tensors into a torch.nn.Module that is already initialized, or for inspection.

Keep in mind that load_state_dict is not a fast operation, and will likely be much slower than load_into_module.

The state_dict can also be used to initialize a HuggingFace Transformers AutoModel. But HuggingFace Transformers performs three or more copies of the data, so memory use will explode.

bfloat16 Support

Tensorizer supports models using the bfloat16 data type. However, tensorizer uses numpy to save the tensors as binary and numpy doesn't support bfloat16. This means that special conversions need to be applied.

To be saved, the torch tensor is cast to int16 before being converted to numpy, which doesn't change any of the underlying data. When serialized, the original bfloat16 datatype string is also saved so that it will be cast back to bfloat16 during the deserialization process.

The complex32 datatype is supported in a similar way, by casting to int32. The quantized datatypes (qint8, qint32, etc.) are not currently supported by tensorizer as they would require supplemental quantization parameters to be deserialized correctly.

NOTE: The exact choice of intermediate types as int16 and int32 is considered an implementation detail, and is subject to change, so they should not be relied upon.

NOTE2: This does not interfere with storing actual int datatypes used in tensors in tensorized files.

Numpy Support

Tensorizer can be used with numpy directly to read and write numpy.ndarrays.

The serializer's write_tensor function handles supplying both torch.Tensors and numpy.ndarrays.

The deserializer has a separate function read_numpy_arrays that will return the data as numpy.ndarrays.

As explained above in bfloat16 support, tensorizer uses special conversions to write "opaque" datatypes, those not supported by numpy. Therefore, special considerations need to be taken when loading such data as numpy.ndarrays.

By default, the TensorDeserializer.read_numpy_arrays function sets its allow_raw_data parameter to False. This means that if a file contains opaque datatypes, a ValueError will be raised during deserialization.

If you want to return the raw data regardless, set allow_raw_data to True. Otherwise, the file may be read with TensorDeserializer.read_tensors instead, which yields torch.Tensor objects of the correct datatype.

A fifth and sixth variable are also returned by the read_numpy_arrays generator. The fifth is a bool that indicates whether the returned array has an opaque datatype and requires special handling (only legal when allow_raw_data=True). The sixth is a string describing the true, non-numpy datatype that the raw data should be interpreted as in such cases. For all other datatypes that require no special handling, these are returned as False and None, respectively. The exact numpy datatypes used by the returned opaque numpy.ndarray objects is not guaranteed, and should not be relied upon.

Running Tests

tensorizer uses unittest for testing. The tests have their own set of dependencies, which can be installed with pip install -r tests/requirements.txt.

Some tests require a GPU, and will be skipped if no GPU is available. To run the tests, run the following in the root of the repository:

python -m pip install -e .
python -m pip install -r tests/requirements.txt
python -m unittest discover tests/ --verbose