alma-torch 0.2.5

A package for benchmarking the speed of different PyTorch conversion options

alma

A Python library for benchmarking PyTorch model speed for different conversion options 🚀

The motivation of alma is to make it easy for people to benchmark their models for different conversion options, e.g. eager, tracing, scripting, torch.compile, torch.export, ONNX, Tensort, etc. The library is designed to be simple to use, with benchmarking provided via a single API call, and to be easily extensible for adding new conversion options.

Beyond just benchmarking, alma is designed to be a one-stop-shop for all model conversion options, so that one can learn about the different conversion options, how to implement them, and how they affect model speed and performance.

Table of Contents

• Getting Started
  • Installation
  • Docker
• Basic Usage
• Examples
• Advanced Features and Design Decisions
• Conversion Options
• Future Work
• How to Contribute

Getting Started

Installation

alma is available as a Python package.

One can install the package from python package index by running

pip install alma-torch

Alternatively, it can be installed from the root of this repository (save level as this README) by running:

pip install -e .

Docker

We recommend that you build the provided Dockerfile to ensure an easy installation of all of the system dependencies and the alma pip packages.

1. Build the Docker Image

   bash scripts/build_docker.sh
   

2. Run the Docker Container
   Create and start a container named alma:

   bash scripts/run_docker.sh
   

3. Access the Running Container
   Enter the container's shell:

   docker exec -it alma bash
   

4. Mount Your Repository
   By default, the run_docker.sh script mounts your /home directory to /home inside the container.
   If your alma repository is in a different location, update the bind mount, for example:

   -v /Users/myuser/alma:/home/alma
   

Basic usage

The core API is benchmark_model, which is used to benchmark the speed of a model for different conversion options. The usage is as follows:

from alma import benchmark_model
from alma.benchmark.log import display_all_results

device = torch.device("cuda") if torch.cuda.is_available() else torch.device("cpu")

# Load the model
model = ...

# Load the dataloader used in benchmarking
data_loader = ...

# Set the configuration
config = {
    "batch_size": 64,
    "n_samples": 2048,
    "device": device,
}

# Choose with conversions to benchmark:
conversions = ["EAGER", "EXPORT+EAGER"]

# Benchmark the model
results = benchmark_model(model, config, conversions, data_loader=data_loader)

# Print all results
display_all_results(results)

The results will look like this, depending on one's model, dataloader, and hardware.

EAGER results:
Device: cuda
Total elapsed time: 0.0211 seconds
Total inference time (model only): 0.0073 seconds
Total samples: 2048 - Batch size: 64
Throughput: 282395.70 samples/second


EXPORT+EAGER results:
Device: cuda
Total elapsed time: 0.0209 seconds
Total inference time (model only): 0.0067 seconds
Total samples: 2048 - Batch size: 64
Throughput: 305974.83 samples/second

Examples:

For extensive examples on how to use alma, as well as simple examples on how train a model and quantize it, see the MNIST example directory. This contains code examples for all of the different alma features, and is where one can find examples on every feature.

For a short working example on a simple Linear+ReLU, see the linear example.

Advanced Features and Design Decisions

alma is designed to be simple to use, with a single API call to benchmark a model for different conversion options. Below are some features we have produced and some design decisions we have made, which are all configurable by the user. For examples on how to use these features, see the MNIST example.

Implicitly initialise a data loader inside of `benchmark_model`

Rather than initializing and feeding in a data loader like in the above example, one can also just pass in a `data` tensor (with no batch dimension), and `benchmark_model` will automatically create a dataloader that produces random tensors of the same shape as the input tensor, with the batch size controlled via the `config` dictionary. This can be convenient if one does not want to create a data loader.

See here for details.

Pre-defined argparser for easy control and experimentation

We provide an argparser that allows one to easily select conversion methods by numerical index or string name. It also allows one to set the batch size, number of samples, and device easily, as well as other commonly used parameters like model weights path.

See here for details.

Graceful or fast failure

By default, `alma` will fail fast if any conversion method fails. This is because we want to know if a conversion method fails, so that we can fix it. However, if one wants to continue benchmarking other options even if a conversion method fails, one can set `fail_on_error` to False in the config dictionary. `alma` will then fail gracefully for that method. One can then access the associated error messages and full tracebacks for the failed methods from the returned object.

See here for details.

Isolated environments for each conversion method via multi-processing

By default, `alma` will run each conversion method in a separate process (one at a time), so that one can benchmark each conversion method in isolation. This ensures that each conversion method is benchmarked in a fair and isolated environment, and is relevant because some of the methods (e.g. optimum quanto) can affect the global torch state and break other methods (e.g. by overwriting tensor defaults in the C++ backend).

To disable multiprocessing, set multiprocessing to False in the config dictionary.

See here for details and discussion.

Logging, debugging, and CI integration

A lot of the conversion methods have verbose internal logging. We have opted to mostly silence those logs. However, if one wants access to those logs, one should use the `setup_logging` function and set the debugging level to `DEBUG`.

See here for details.

Conversion Options

Naming conventions

The naming convention for conversion options is to use short but descriptive names, e.g. EAGER, EXPORT+EAGER, EXPORT+TENSORRT, etc. If multiple "techniques" are used in a single conversion option, then the names are separated by a + sign in chronological order of operation. Underscores _ are used within each technique name to seperate the words for readability, e.g. EXPORT+AOT_INDUCTOR, where EXPORT and AOT_INDUCTOR are considered seperate steps. All conversion options are located in the src/alma/conversions/ directory. Within this directory:

Code

All conversion options are located in the src/alma/conversions/ directory. In this directory:

• The options/ subdirectory contains one Python file per conversion option (or a closely related family of options, e.g. torch.compile backends).
• The main selection logic for these options is found in select.py. This is just a glorified match-case statement that returns the forward calls of each model conversion option, which is returned to the benchmarking loop. It is that simple!

At the risk of some code duplication, we have chosen to keep the conversion options separate, so that one can easily add new conversion options without having to modify the existing ones. It also makes it easier for the user to see what conversion options are available, and to understand what each conversion option does.

Options Summary

Below is a table summarizing the currently supported conversion options and their identifiers:

ID	Conversion Option
0	EAGER
1	EXPORT+EAGER
2	ONNX_CPU
3	ONNX_GPU
4	ONNX+DYNAMO_EXPORT
5	COMPILE_CUDAGRAPHS
6	COMPILE_INDUCTOR_DEFAULT
7	COMPILE_INDUCTOR_REDUCE_OVERHEAD
8	COMPILE_INDUCTOR_MAX_AUTOTUNE
9	COMPILE_INDUCTOR_EAGER_FALLBACK
10	COMPILE_ONNXRT
11	COMPILE_OPENXLA
12	COMPILE_TVM
13	EXPORT+AI8WI8_FLOAT_QUANTIZED
14	EXPORT+AI8WI8_FLOAT_QUANTIZED+RUN_DECOMPOSITION
15	EXPORT+AI8WI8_STATIC_QUANTIZED
16	EXPORT+AI8WI8_STATIC_QUANTIZED+RUN_DECOMPOSITION
17	EXPORT+AOT_INDUCTOR
18	EXPORT+COMPILE_CUDAGRAPHS
19	EXPORT+COMPILE_INDUCTOR_DEFAULT
20	EXPORT+COMPILE_INDUCTOR_REDUCE_OVERHEAD
21	EXPORT+COMPILE_INDUCTOR_MAX_AUTOTUNE
22	EXPORT+COMPILE_INDUCTOR_DEFAULT_EAGER_FALLBACK
23	EXPORT+COMPILE_ONNXRT
24	EXPORT+COMPILE_OPENXLA
25	EXPORT+COMPILE_TVM
26	NATIVE_CONVERT_AI8WI8_STATIC_QUANTIZED
27	NATIVE_FAKE_QUANTIZED_AI8WI8_STATIC
28	COMPILE_TENSORRT
29	EXPORT+COMPILE_TENSORRT
30	JIT_TRACE
31	TORCH_SCRIPT
32	OPTIMIM_QUANTO_AI8WI8
33	OPTIMIM_QUANTO_AI8WI4
34	OPTIMIM_QUANTO_AI8WI2
35	OPTIMIM_QUANTO_WI8
36	OPTIMIM_QUANTO_WI4
37	OPTIMIM_QUANTO_WI2
38	OPTIMIM_QUANTO_Wf8E4M3N
39	OPTIMIM_QUANTO_Wf8E4M3NUZ
40	OPTIMIM_QUANTO_Wf8E5M2
41	OPTIMIM_QUANTO_Wf8E5M2+COMPILE_CUDAGRAPHS

These conversion options are also all hard-coded in the alma/conversions/select.py file, which is the source of truth.

## Testing

We use pytest for testing. Simply run:

pytest

We currently don't have extensive tests, but we are working on adding more tests to ensure that the conversion options are working as expected in known environments (e.g. the Docker container).

Future work:

• Add more conversion options. This is a work in progress, and we are always looking for more conversion options.
• Multi-device benchmarking. Currently alma only supports single-device benchmarking, but ideally a model could be split across multiple devices.
• Integrating conversion options beyond PyTorch, e.g. HuggingFace, JAX, llama.cpp, etc.

How to contribute:

Contributions are welcome! If you have a new conversion option, feature, or other you would like to add, so that the whole community can benefit, please open a pull request! We are always looking for new conversion options, and we are happy to help you get started with adding a new conversion option/feature!

See the CONTRIBUTING.md file for more detailed information on how to contribute.

Citation

@Misc{alma,
  title =        {Alma: PyTorch model speed benchmarking across all conversion types},
  author =       {Oscar Savolainen and Saif Haq},
  howpublished = {\url{https://github.com/saifhaq/alma}},
  year =         {2024}
}