text
stringlengths 0
1.73k
| source
stringlengths 35
119
| category
stringclasses 2
values |
|---|---|---|
Figure 7. Pipeline Bubble: , and denote forward, backward, and the optimizer update of micro-batch on device , respectively. The total bubble size in each iteration is times per micro-batch forward and backward cost.
| https://pytorch.org/blog/pipetransformer-automated-elastic-pipelining/ | pytorch blogs |
Additionally, such a technique can also speed up training by shrinking the size of pipeline bubbles. To explain bubble sizes in a pipeline, Figure 7 depicts how 4 micro-batches run through a 4-device pipeline . In general, the total bubble size is times per micro-batch forward and backward cost. Therefore, it is clear that shorter pipelines have smaller bubble sizes.
Dynamic Number of Micro-Batches | https://pytorch.org/blog/pipetransformer-automated-elastic-pipelining/ | pytorch blogs |
Prior pipeline parallel systems use a fixed number of micro-batches per mini-batch ( ). GPipe suggests , where is the number of partitions (pipeline length). However, given that PipeTransformer dynamically configures , we find it to be sub-optimal to maintain a static during training. Moreover, when integrated with DDP, the value of also has an impact on the efficiency of DDP gradient synchronizations. Since DDP must wait for the last micro-batch to finish its backward computation on a parameter before launching its gradient synchronization, finer micro-batches lead to a smaller overlap between computation and communication. Hence, instead of using a static value, PipeTransformer searches for optimal on the fly in the hybrid of DDP environment by enumerating values ranging from to . For a specific training environment, the profiling needs only to be done once (see Algorithm 1 line 35). | https://pytorch.org/blog/pipetransformer-automated-elastic-pipelining/ | pytorch blogs |
For the complete source code, please refer to https://github.com/Distributed-AI/PipeTransformer/blob/master/pipe_transformer/pipe/auto_pipe.py.
AutoDP: Spawning More Pipeline Replicas
As AutoPipe compresses the same pipeline into fewer GPUs, AutoDP can automatically spawn new pipeline replicas to increase data-parallel width.
Despite the conceptual simplicity, subtle dependencies on communications and states require careful design. The challenges are threefold:
DDP Communication: Collective communications in PyTorch DDP requires static membership, which prevents new pipelines from connecting with existing ones;
State Synchronization: newly activated processes must be consistent with existing pipelines in the training progress (e.g., epoch number and learning rate), weights and optimizer states, the boundary of frozen layers, and pipeline GPU range;
| https://pytorch.org/blog/pipetransformer-automated-elastic-pipelining/ | pytorch blogs |
Dataset Redistribution: the dataset should be re-balanced to match a dynamic number of pipelines. This not only avoids stragglers but also ensures that gradients from all DDP processes are equally weighted.
Figure 8. AutoDP: handling dynamical data-parallel with messaging between double process groups (Process 0-7 belong to machine 0, while process 8-15 belong to machine 1)
To tackle these challenges, we create double communication process groups for DDP. As in the example shown in Figure 8, the message process group (purple) is responsible for light-weight control messages and covers all processes, while the active training process group (yellow) only contains active processes and serves as a vehicle for heavy-weight tensor communications during training. The message group remains static, whereas the training group is dismantled and reconstructed to match active processes. | https://pytorch.org/blog/pipetransformer-automated-elastic-pipelining/ | pytorch blogs |
In T0, only processes 0 and 8 are active. During the transition to T1, process 0 activates processes 1 and 9 (newly added pipeline replicas) and synchronizes necessary information mentioned above using the message group. The four active processes then form a new training group, allowing static collective communications adaptive to dynamic memberships.
To redistribute the dataset, we implement a variant of DistributedSampler that can seamlessly adjust data samples to match the number of active pipeline replicas.
The above design also naturally helps to reduce DDP communication overhead. More specifically, when transitioning from T0 to T1, processes 0 and 1 destroy the existing DDP instances, and active processes construct a new DDP training group using a cached pipelined model (AutoPipe stores frozen model and cached model separately).
We use the following APIs to implement the design above.
```python
import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel as DDP | https://pytorch.org/blog/pipetransformer-automated-elastic-pipelining/ | pytorch blogs |
initialize the process group (this must be called in the initialization of PyTorch DDP)
dist.init_process_group(init_method='tcp://' + str(self.config.master_addr) + ':' +
str(self.config.master_port), backend=Backend.GLOO, rank=self.global_rank, world_size=self.world_size)
...
create active process group (yellow color)
self.active_process_group = dist.new_group(ranks=self.active_ranks, backend=Backend.NCCL, timeout=timedelta(days=365))
...
create message process group (yellow color)
self.comm_broadcast_group = dist.new_group(ranks=[i for i in range(self.world_size)], backend=Backend.GLOO, timeout=timedelta(days=365))
...
create DDP-enabled model when the number of data-parallel workers is changed. Note:
1. The process group to be used for distributed data all-reduction.
If None, the default process group, which is created by torch.distributed.init_process_group, will be used.
In our case, we set it as self.active_process_group | https://pytorch.org/blog/pipetransformer-automated-elastic-pipelining/ | pytorch blogs |
2. device_ids should be set when the pipeline length = 1 (the model resides on a single CUDA device).
self.pipe_len = gpu_num_per_process
if gpu_num_per_process > 1:
model = DDP(model, process_group=self.active_process_group, find_unused_parameters=True)
else:
model = DDP(model, device_ids=[self.local_rank], process_group=self.active_process_group, find_unused_parameters=True)
to broadcast message among processes, we use dist.broadcast_object_list
def dist_broadcast(object_list, src, group):
"""Broadcasts a given object to all parties."""
dist.broadcast_object_list(object_list, src, group=group)
return object_list
``
For the complete source code, please refer tohttps://github.com/Distributed-AI/PipeTransformer/blob/master/pipe_transformer/dp/auto_dp.py`.
Experiments
This section first summarizes experiment setups and then evaluates PipeTransformer using computer vision and natural language processing tasks. | https://pytorch.org/blog/pipetransformer-automated-elastic-pipelining/ | pytorch blogs |
Hardware. Experiments were conducted on 2 identical machines connected by InfiniBand CX353A (GB/s), where each machine is equipped with 8 NVIDIA Quadro RTX 5000 (16GB GPU memory). GPU-to-GPU bandwidth within a machine (PCI 3.0, 16 lanes) is GB/s.
Implementation. We used PyTorch Pipe as a building block. The BERT model definition, configuration, and related tokenizer are from HuggingFace 3.5.0. We implemented Vision Transformer using PyTorch by following its TensorFlow implementation. More details can be found in our source code. | https://pytorch.org/blog/pipetransformer-automated-elastic-pipelining/ | pytorch blogs |
Models and Datasets. Experiments employ two representative Transformers in CV and NLP: Vision Transformer (ViT) and BERT. ViT was run on an image classification task, initialized with pre-trained weights on ImageNet21K and fine-tuned on ImageNet and CIFAR-100. BERT was run on two tasks, text classification on the SST-2 dataset from the General Language Understanding Evaluation (GLUE) benchmark, and question answering on the SQuAD v1.1 Dataset (Stanford Question Answering), which is a collection of 100k crowdsourced question/answer pairs. | https://pytorch.org/blog/pipetransformer-automated-elastic-pipelining/ | pytorch blogs |
Training Schemes. Given that large models normally would require thousands of GPU-days {\emph{e.g.}, GPT-3) if trained from scratch, fine-tuning downstream tasks using pre-trained models has become a trend in CV and NLP communities. Moreover, PipeTransformer is a complex training system that involves multiple core components. Thus, for the first version of PipeTransformer system development and algorithmic research, it is not cost-efficient to develop and evaluate from scratch using large-scale pre-training. Therefore, the experiments presented in this section focuses on pre-trained models. Note that since the model architectures in pre-training and fine-tuning are the same, PipeTransformer can serve both. We discussed pre-training results in the Appendix. | https://pytorch.org/blog/pipetransformer-automated-elastic-pipelining/ | pytorch blogs |
Baseline. Experiments in this section compare PipeTransformer to the state-of-the-art framework, a hybrid scheme of PyTorch Pipeline (PyTorch’s implementation of GPipe) and PyTorch DDP. Since this is the first paper that studies accelerating distributed training by freezing layers, there are no perfectly aligned counterpart solutions yet.
Hyper-parameters. Experiments use ViT-B/16 (12 transformer layers, input patch size) for ImageNet and CIFAR-100, BERT-large-uncased (24 layers) for SQuAD 1.1, and BERT-base-uncased (12 layers) for SST-2. With PipeTransformer, ViT and BERT training can set the per-pipeline batch size to around 400 and 64, respectively. Other hyperparameters (e.g., epoch, learning rate) for all experiments are presented in Appendix.
Overall Training Acceleration
| https://pytorch.org/blog/pipetransformer-automated-elastic-pipelining/ | pytorch blogs |
We summarize the overall experimental results in the table above. Note that the speedup we report is based on a conservative value that can obtain comparable or even higher accuracy. A more aggressive (, ) can obtain a higher speedup but may lead to a slight loss in accuracy. Note that the model size of BERT (24 layers) is larger than ViT-B/16 (12 layers), thus it takes more time for communication.
Performance Analysis
Speedup Breakdown
This section presents evaluation results and analyzes the performance of different components in \autopipe. More experimental results can be found in the Appendix.
| https://pytorch.org/blog/pipetransformer-automated-elastic-pipelining/ | pytorch blogs |
Figure 9. Speedup Breakdown (ViT on ImageNet)
To understand the efficacy of all four components and their impacts on training speed, we experimented with different combinations and used their training sample throughput (samples/second) and speedup ratio as metrics. Results are illustrated in Figure 9. Key takeaways from these experimental results are:
the main speedup is the result of elastic pipelining which is achieved through the joint use of AutoPipe and AutoDP;
AutoCache's contribution is amplified by AutoDP;
freeze training alone without system-wise adjustment even downgrades the training speed.
Tuning in Freezing Algorithm
| https://pytorch.org/blog/pipetransformer-automated-elastic-pipelining/ | pytorch blogs |
Figure 10. Tuning in Freezing Algorithm
We ran experiments to show how the in the freeze algorithms influences training speed. The result clearly demonstrates that a larger (excessive freeze) leads to a greater speedup but suffers from a slight performance degradation. In the case shown in Figure 10, where , freeze training outperforms normal training and obtains a -fold speedup. We provide more results in the Appendix.
Optimal Chunks in the elastic pipeline
Figure 11. Optimal chunk number in the elastic pipeline
| https://pytorch.org/blog/pipetransformer-automated-elastic-pipelining/ | pytorch blogs |
We profiled the optimal number of micro-batches for different pipeline lengths . Results are summarized in Figure 11. As we can see, different values lead to different optimal , and the throughput gaps across different M values are large (as shown when ), which confirms the necessity of an anterior profiler in elastic pipelining.
Understanding the Timing of Caching
Figure 12. the timing of caching
| https://pytorch.org/blog/pipetransformer-automated-elastic-pipelining/ | pytorch blogs |
Figure 12. the timing of caching
To evaluate AutoCache, we compared the sample throughput of training that activates AutoCache from epoch (blue) with the training job without AutoCache (red). Figure 12 shows that enabling caching too early can slow down training, as caching can be more expensive than the forward propagation on a small number of frozen layers. After more layers are frozen, caching activations clearly outperform the corresponding forward propagation. As a result, AutoCache uses a profiler to determine the proper timing to enable caching. In our system, for ViT (12 layers), caching starts from 3 frozen layers, while for BERT (24 layers), caching starts from 5 frozen layers.
For more detailed experimental analysis, please refer to our paper.
Summarization | https://pytorch.org/blog/pipetransformer-automated-elastic-pipelining/ | pytorch blogs |
Summarization
This blog introduces PipeTransformer, a holistic solution that combines elastic pipeline-parallel and data-parallel for distributed training using PyTorch Distributed APIs. More specifically, PipeTransformer incrementally freezes layers in the pipeline, packs remaining active layers into fewer GPUs, and forks more pipeline replicas to increase the data-parallel width. Evaluations on ViT and BERT models show that compared to the state-of-the-art baseline, PipeTransformer attains up to 2.83× speedups without accuracy loss.
Reference
[1] Li, S., Zhao, Y., Varma, R., Salpekar, O., Noordhuis, P., Li,T., Paszke, A., Smith, J., Vaughan, B., Damania, P., et al. Pytorch Distributed: Experiences on Accelerating Dataparallel Training. Proceedings of the VLDB Endowment,13(12), 2020
[2] Devlin, J., Chang, M. W., Lee, K., and Toutanova, K. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. In NAACL-HLT, 2019 | https://pytorch.org/blog/pipetransformer-automated-elastic-pipelining/ | pytorch blogs |
[3] Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., et al. An image is Worth 16x16 words: Transformers for Image Recognition at Scale.
[4] Brown, T. B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Neelakantan, A., Shyam, P., Sastry, G., Askell, A., et al. Language Models are Few-shot Learners.
[5] Lepikhin, D., Lee, H., Xu, Y., Chen, D., Firat, O., Huang, Y., Krikun, M., Shazeer, N., and Chen, Z. Gshard: Scaling Giant Models with Conditional Computation and Automatic Sharding.
[6] Li, M., Andersen, D. G., Park, J. W., Smola, A. J., Ahmed, A., Josifovski, V., Long, J., Shekita, E. J., and Su, B. Y. Scaling Distributed Machine Learning with the Parameter Server. In 11th {USENIX} Symposium on Operating Systems Design and Implementation ({OSDI} 14), pp. 583–598, 2014. | https://pytorch.org/blog/pipetransformer-automated-elastic-pipelining/ | pytorch blogs |
[7] Jiang, Y., Zhu, Y., Lan, C., Yi, B., Cui, Y., and Guo, C. A Unified Architecture for Accelerating Distributed DNN Training in Heterogeneous GPU/CPU Clusters. In 14th USENIX Symposium on Operating Systems Design and Implementation (OSDI 20), pp. 463–479. USENIX Association, November 2020. ISBN 978-1-939133-19- 9.
[8] Kim, S., Yu, G. I., Park, H., Cho, S., Jeong, E., Ha, H., Lee, S., Jeong, J. S., and Chun, B. G. Parallax: Sparsity-aware Data Parallel Training of Deep Neural Networks. In Proceedings of the Fourteenth EuroSys Conference 2019, pp. 1–15, 2019.
[9] Kim, C., Lee, H., Jeong, M., Baek, W., Yoon, B., Kim, I., Lim, S., and Kim, S. TorchGPipe: On-the-fly Pipeline Parallelism for Training Giant Models.
[10] Huang, Y., Cheng, Y., Bapna, A., Firat, O., Chen, M. X., Chen, D., Lee, H., Ngiam, J., Le, Q. V., Wu, Y., et al. Gpipe: Efficient Training of Giant Neural Networks using Pipeline Parallelism. | https://pytorch.org/blog/pipetransformer-automated-elastic-pipelining/ | pytorch blogs |
[11] Park, J. H., Yun, G., Yi, C. M., Nguyen, N. T., Lee, S., Choi, J., Noh, S. H., and ri Choi, Y. Hetpipe: Enabling Large DNN Training on (whimpy) Heterogeneous GPU Clusters through Integration of Pipelined Model Parallelism and Data Parallelism. In 2020 USENIX Annual Technical Conference (USENIX ATC 20), pp. 307–321. USENIX Association, July 2020. ISBN 978-1-939133- 14-4.
[12] Narayanan, D., Harlap, A., Phanishayee, A., Seshadri, V., Devanur, N. R., Ganger, G. R., Gibbons, P. B., and Zaharia, M. Pipedream: Generalized Pipeline Parallelism for DNN Training. In Proceedings of the 27th ACM Symposium on Operating Systems Principles, SOSP ’19, pp. 1–15, New York, NY, USA, 2019. Association for Computing Machinery. ISBN 9781450368735. doi: 10.1145/3341301.3359646.
[13] Lepikhin, D., Lee, H., Xu, Y., Chen, D., Firat, O., Huang, Y., Krikun, M., Shazeer, N., and Chen, Z. Gshard: Scaling Giant Models with Conditional Computation and Automatic Sharding. | https://pytorch.org/blog/pipetransformer-automated-elastic-pipelining/ | pytorch blogs |
[14] Shazeer, N., Cheng, Y., Parmar, N., Tran, D., Vaswani, A., Koanantakool, P., Hawkins, P., Lee, H., Hong, M., Young, C., Sepassi, R., and Hechtman, B. Mesh-Tensorflow: Deep Learning for Supercomputers. In Bengio, S., Wallach, H., Larochelle, H., Grauman, K., Cesa-Bianchi, N., and Garnett, R. (eds.), Advances in Neural Information Processing Systems, volume 31, pp. 10414–10423. Curran Associates, Inc., 2018.
[15] Shoeybi, M., Patwary, M., Puri, R., LeGresley, P., Casper, J., and Catanzaro, B. Megatron-LM: Training Multi-billion Parameter Language Models using Model Parallelism.
[16] Rajbhandari, S., Rasley, J., Ruwase, O., and He, Y. ZERO: Memory Optimization towards Training a Trillion Parameter Models.
[17] Raghu, M., Gilmer, J., Yosinski, J., and Sohl Dickstein, J. Svcca: Singular Vector Canonical Correlation Analysis for Deep Learning Dynamics and Interpretability. In NIPS, 2017. | https://pytorch.org/blog/pipetransformer-automated-elastic-pipelining/ | pytorch blogs |
[18] Morcos, A., Raghu, M., and Bengio, S. Insights on Representational Similarity in Neural Networks with Canonical Correlation. In Bengio, S., Wallach, H., Larochelle, H., Grauman, K., Cesa-Bianchi, N., and Garnett, R. (eds.), Advances in Neural Information Processing Systems 31, pp. 5732–5741. Curran Associates, Inc., 2018. | https://pytorch.org/blog/pipetransformer-automated-elastic-pipelining/ | pytorch blogs |
layout: blog_detail
title: 'Model Serving in PyTorch'
author: Jeff Smith
redirect_from: /2019/05/08/model-serving-in-pyorch.html
PyTorch has seen a lot of adoption in research, but people can get confused about how well PyTorch models can be taken into production. This blog post is meant to clear up any confusion people might have about the road to production in PyTorch.
Usually when people talk about taking a model “to production,” they usually mean performing inference, sometimes called model evaluation or prediction or serving. At the level of a function call, in PyTorch, inference looks something like this:
In Python
module(input)
In traced modules
module(input)
In C++
at::Tensor output = module->forward(inputs).toTensor();
Since we at Facebook perform inference operations using PyTorch hundreds of trillions of times per day, we've done a lot to make sure that inference runs as efficiently as possible.
Serving Strategies | https://pytorch.org/blog/model-serving-in-pyorch/ | pytorch blogs |
Serving Strategies
That zoomed-in view of how you use models in inference isn't usually the whole story, though. In a real world machine learning system, you often need to do more than just run a single inference operation in the REPL or Jupyter notebook. Instead, you usually need to integrate your model into a larger application in some way. Depending on what you need to do, you can usually take one of the following approaches.
Direct embedding
In application settings like mobile, we often just directly call the model as part of a larger program. This isn't just for apps; usually this is how robotics and dedicated devices work as well. At a code-level, the call to the model is exactly the same as what is shown above in the section about inference shown above. A key concern is often that a Python interpreter is not present in such environments, which is why PyTorch allows you to call your models from C++ and ship a model without the need for a Python runtime.
Model microservices | https://pytorch.org/blog/model-serving-in-pyorch/ | pytorch blogs |
Model microservices
If you're using your model in a server side context and you're managing multiple models, you might choose to treat each individual model (or each individual model version) as a separate service, usually using some sort of packaging mechanism like a Docker container. Then that service is often made network accessible via some sort of service, either using JSON over HTTP or an RPC technology like gRPC. The key characteristic of this approach is that you're defining a service with a single endpoint that just calls your model. Then you do do all of your model management (promotion, rollback, etc.) via whatever system you already use to manage your services (e.g. kubernetes, ECS).
Model servers | https://pytorch.org/blog/model-serving-in-pyorch/ | pytorch blogs |
Model servers
An additional possible solution is to use a model server. This is an application built to manage and serve models. It allows you to upload multiple models and get distinct prediction endpoints for each of them. Typically such systems include a number of other features to help solve more of the whole problem of managing and serving models. This can include things like metrics, visualization, data pre-processing, and more. Even something as simple as having a system for automatically versioning models can make building important features like model rollbacks much easier.
Evolving Patterns | https://pytorch.org/blog/model-serving-in-pyorch/ | pytorch blogs |
Evolving Patterns
The above is a somewhat arbitrary breakdown of different approaches based on a snapshot in time. Design patterns are still evolving. Recently, model server designs have started to adopt more of the technologies of general service infrastructure such as Docker containers and kubernetes, so many model servers have started to share properties of the model microservice design discussed above. For a deeper dive into the general concepts of model server designs, you can check out my book on machine learning systems.
Serving PyTorch Models
So, if you're a PyTorch user, what should you use if you want to take your models to production?
If you're on mobile or working on an embedded system like a robot, direct embedding in your application is often the right choice.
For mobile specifically, your use case might be served by the ONNX export functionality. | https://pytorch.org/blog/model-serving-in-pyorch/ | pytorch blogs |
Note that ONNX, by its very nature, has limitations and doesn't support all of the functionality provided by the larger PyTorch project.
You can check out this tutorial on deploying PyTorch models to mobile using ONNX to see if this path might suit your use case.
That said, we've heard that there's a lot more that PyTorch users want to do on mobile, so look for more mobile-specific functionality in PyTorch in the future.
For other embedded systems, like robots, running inference on a PyTorch model from the C++ API could be the right solution.
If you can't use the cloud or prefer to manage all services using the same technology, you can follow this example to build a simple model microservice using the Flask web framework. | https://pytorch.org/blog/model-serving-in-pyorch/ | pytorch blogs |
If you want to manage multiple models within a non-cloud service solution, there are teams developing PyTorch support in model servers like MLFlow, Kubeflow, and RedisAI. We're excited to see innovation from multiple teams building OSS model servers, and we'll continue to highlight innovation in the PyTorch ecosystem in the future. | https://pytorch.org/blog/model-serving-in-pyorch/ | pytorch blogs |
If you can use the cloud for your application, there are several great choices for working with models in the cloud. For AWS Sagemaker, you can start find a guide to all of the resources from AWS for working with PyTorch, including docs on how to use the Sagemaker Python SDK. You can also see some talks we've given on using PyTorch on Sagemaker. Finally, if you happen to be using PyTorch via FastAI, then they've written a really simple guide to getting up and running on Sagemaker. | https://pytorch.org/blog/model-serving-in-pyorch/ | pytorch blogs |
The story is similar across other major clouds. On Google Cloud, you can follow these instructions to get access to a Deep Learning VM with PyTorch pre-installed. On Microsoft Azure, you have a number of ways to get started from Azure Machine Learning Service to Azure Notebooks showing how to use PyTorch.
Your Models
Whichever approach you take to bringing your PyTorch models to production, we want to support you and enable your success. Do you love one of the options above? Are you having difficulty with that one crucial feature you can't find support for? We'd love to discuss more on the deployment category on the PyTorch Discuss forums. We'd love to help, and where you're seeing success, amplify your story. | https://pytorch.org/blog/model-serving-in-pyorch/ | pytorch blogs |
layout: blog_detail
title: 'Accelerating PyTorch with CUDA Graphs'
author: Vinh Nguyen, Michael Carilli, Sukru Burc Eryilmaz, Vartika Singh, Michelle Lin, Natalia Gimelshein, Alban Desmaison, Edward Yang
featured-img: 'assets/images/cudagraphs-pytorch.png'
| https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
Today, we are pleased to announce a new advanced CUDA feature, CUDA Graphs, has been brought to PyTorch. Modern DL frameworks have complicated software stacks that incur significant overheads associated with the submission of each operation to the GPU. When DL workloads are strong-scaled to many GPUs for performance, the time taken by each GPU operation diminishes to just a few microseconds and, in these cases, the high work submission latencies of frameworks often lead to low utilization of the GPU. As GPUs get faster and workloads are scaled to more devices, the likelihood of workloads suffering from these launch-induced stalls increases. To overcome these performance overheads, NVIDIA engineers worked with PyTorch developers to enable CUDA graph execution natively in PyTorch. This design was instrumental in scaling NVIDIA’s MLPerf workloads (implemented in PyTorch) to over 4000 GPUs in order to achieve record-breaking performance. | https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
|
CUDA graphs support in PyTorch is just one more example of a long collaboration between NVIDIA and Facebook engineers. torch.cuda.amp, for example, trains with half precision while maintaining the network accuracy achieved with single precision and automatically utilizing tensor cores wherever possible. AMP delivers up to 3X higher performance than FP32 with just a few lines of code change. Similarly, NVIDIA’s Megatron-LM was trained using PyTorch on up to 3072 GPUs. In PyTorch, one of the most performant methods to scale-out GPU training is with torch.nn.parallel.DistributedDataParallel coupled with the NVIDIA Collective Communications Library (NCCL) backend.
CUDA Graphs | https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
CUDA Graphs
CUDA Graphs, which made its debut in CUDA 10, let a series of CUDA kernels to be defined and encapsulated as a single unit, i.e., a graph of operations, rather than a sequence of individually-launched operations. It provides a mechanism to launch multiple GPU operations through a single CPU operation, and hence reduces the launching overheads. | https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
The benefits of CUDA graphs can be demonstrated with the simple example in Figure 1. On the top, a sequence of short kernels is launched one-by-one by the CPU. The CPU launching overhead creates a significant gap in between the kernels. If we replace this sequence of kernels with a CUDA graph, initially we will need to spend a little extra time on building the graph and launching the whole graph in one go on the first occasion, but subsequent executions will be very fast, as there will be very little gap between the kernels. The difference is more pronounced when the same sequence of operations is repeated many times, for example, overy many training steps. In that case, the initial costs of building and launching the graph will be amortized over the entire number of training iterations. For a more comprehensive introduction on the topic, see our blog | https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
Getting Started with CUDA Graphs and GTC talk Effortless CUDA Graphs.
Figure 1. Benefits of using CUDA graphs
NCCL support for CUDA graphs
The previously mentioned benefits of reducing launch overheads also extend to NCCL kernel launches. NCCL enables GPU-based collective and P2P communications. With NCCL support for CUDA graphs, we can eliminate the NCCL kernel launch overhead. | https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
Additionally, kernel launch timing can be unpredictable due to various CPU load and operating system factors. Such time skews can be harmful to the performance of NCCL collective operations. With CUDA graphs, kernels are clustered together so that performance is consistent across ranks in a distributed workload. This is especially useful in large clusters where even a single slow node can bring down overall cluster level performance.
For distributed multi-GPU workloads, NCCL is used for collective communications. If we look at training a neural network that leverages data parallelism, without NCCL support for CUDA graphs, we’ll need a separate launch for each of forward/back propagation and NCCL AllReduce. By contrast, with NCCL support for CUDA graphs, we can reduce launch overhead by lumping together the forward/backward propagation and NCCL AllReduce all in a single graph launch.
| https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
Figure 2. Looking at a typical neural network, all the kernel launches for NCCL AllReduce can be bundled into a graph to reduce overhead launch time.
PyTorch CUDA Graphs
From PyTorch v1.10, the CUDA graphs functionality is made available as a set of beta APIs.
API overview | https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
API overview
PyTorch supports the construction of CUDA graphs using stream capture, which puts a CUDA stream in capture mode. CUDA work issued to a capturing stream doesn’t actually run on the GPU. Instead, the work is recorded in a graph. After capture, the graph can be launched to run the GPU work as many times as needed. Each replay runs the same kernels with the same arguments. For pointer arguments this means the same memory addresses are used. By filling input memory with new data (e.g., from a new batch) before each replay, you can rerun the same work on new data. | https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
Replaying a graph sacrifices the dynamic flexibility of typical eager execution in exchange for greatly reduced CPU overhead. A graph’s arguments and kernels are fixed, so a graph replay skips all layers of argument setup and kernel dispatch, including Python, C++, and CUDA driver overheads. Under the hood, a replay submits the entire graph’s work to the GPU with a single call to cudaGraphLaunch. Kernels in a replay also execute slightly faster on the GPU, but eliding CPU overhead is the main benefit.
You should try CUDA graphs if all or part of your network is graph-safe (usually this means static shapes and static control flow, but see the other constraints) and you suspect its runtime is at least somewhat CPU-limited.
API example | https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
API example
PyTorch exposes graphs via a raw torch.cuda.CUDAGraphclass and two convenience wrappers, torch.cuda.graph and torch.cuda.make_graphed_callables. | https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
torch.cuda.graph is a simple, versatile context manager that captures CUDA work in its context. Before capture, warm up the workload to be captured by running a few eager iterations. Warmup must occur on a side stream. Because the graph reads from and writes to the same memory addresses in every replay, you must maintain long-lived references to tensors that hold input and output data during capture. To run the graph on new input data, copy new data to the capture’s input tensor(s), replay the graph, then read the new output from the capture’s output tensor(s).
If the entire network is capture safe, one can capture and replay the whole network as in the following example.
```python
N, D_in, H, D_out = 640, 4096, 2048, 1024
model = torch.nn.Sequential(torch.nn.Linear(D_in, H),
torch.nn.Dropout(p=0.2),
torch.nn.Linear(H, D_out), | https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
torch.nn.Dropout(p=0.1)).cuda()
loss_fn = torch.nn.MSELoss()
optimizer = torch.optim.SGD(model.parameters(), lr=0.1)
Placeholders used for capture
static_input = torch.randn(N, D_in, device='cuda')
static_target = torch.randn(N, D_out, device='cuda')
warmup
Uses static_input and static_target here for convenience,
but in a real setting, because the warmup includes optimizer.step()
you must use a few batches of real data.
s = torch.cuda.Stream()
s.wait_stream(torch.cuda.current_stream())
with torch.cuda.stream(s):
for i in range(3):
optimizer.zero_grad(set_to_none=True)
y_pred = model(static_input)
loss = loss_fn(y_pred, static_target)
loss.backward()
optimizer.step()
torch.cuda.current_stream().wait_stream(s)
capture
g = torch.cuda.CUDAGraph()
Sets grads to None before capture, so backward() will create
.grad attributes with allocations from the graph's private pool
optimizer.zero_grad(set_to_none=True) | https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
optimizer.zero_grad(set_to_none=True)
with torch.cuda.graph(g):
static_y_pred = model(static_input)
static_loss = loss_fn(static_y_pred, static_target)
static_loss.backward()
optimizer.step()
real_inputs = [torch.rand_like(static_input) for _ in range(10)]
real_targets = [torch.rand_like(static_target) for _ in range(10)]
for data, target in zip(real_inputs, real_targets):
# Fills the graph's input memory with new data to compute on
static_input.copy_(data)
static_target.copy_(target)
# replay() includes forward, backward, and step.
# You don't even need to call optimizer.zero_grad() between iterations
# because the captured backward refills static .grad tensors in place.
g.replay()
# Params have been updated. static_y_pred, static_loss, and .grad
# attributes hold values from computing on this iteration's data.
``` | https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
```
If some of your network is unsafe to capture (e.g., due to dynamic control flow, dynamic shapes, CPU syncs, or essential CPU-side logic), you can run the unsafe part(s) eagerly and use torch.cuda.make_graphed_callables to graph only the capture-safe part(s). This is demonstrated next. | https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
make_graphed_callables accepts callables (functions or nn.Module and returns graphed versions. By default, callables returned by make_graphed_callables are autograd-aware, and can be used in the training loop as direct replacements for the functions or nn.Module you passed. make_graphed_callables internally creates CUDAGraph objects, runs warm up iterations, and maintains static inputs and outputs as needed. Therefore, (unlike with torch.cuda.graph) you don’t need to handle those manually. | https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
In the following example, data-dependent dynamic control flow means the network isn’t capturable end-to-end, but make_graphed_callables() lets us capture and run graph-safe sections as graphs regardless:
```python
N, D_in, H, D_out = 640, 4096, 2048, 1024
module1 = torch.nn.Linear(D_in, H).cuda()
module2 = torch.nn.Linear(H, D_out).cuda()
module3 = torch.nn.Linear(H, D_out).cuda()
loss_fn = torch.nn.MSELoss()
optimizer = torch.optim.SGD(chain(module1.parameters(),
module2.parameters(),
module3.parameters()),
lr=0.1)
Sample inputs used for capture
requires_grad state of sample inputs must match
requires_grad state of real inputs each callable will see.
x = torch.randn(N, D_in, device='cuda')
h = torch.randn(N, H, device='cuda', requires_grad=True) | https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
module1 = torch.cuda.make_graphed_callables(module1, (x,))
module2 = torch.cuda.make_graphed_callables(module2, (h,))
module3 = torch.cuda.make_graphed_callables(module3, (h,))
real_inputs = [torch.rand_like(x) for _ in range(10)]
real_targets = [torch.randn(N, D_out, device="cuda") for _ in range(10)]
for data, target in zip(real_inputs, real_targets):
optimizer.zero_grad(set_to_none=True)
tmp = module1(data) # forward ops run as a graph
if tmp.sum().item() > 0:
tmp = module2(tmp) # forward ops run as a graph
else:
tmp = module3(tmp) # forward ops run as a graph
loss = loss_fn(tmp, target)
# module2's or module3's (whichever was chosen) backward ops,
# as well as module1's backward ops, run as graphs
loss.backward()
optimizer.step()
```
Example use cases
MLPerf v1.0 training workloads | https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
MLPerf v1.0 training workloads
The PyTorch CUDA graphs functionality was instrumental in scaling NVIDIA’s MLPerf training v1.0 workloads (implemented in PyTorch) to over 4000 GPUs, setting new records across the board. We illustrate below two MLPerf workloads where the most significant gains were observed with the use of CUDA graphs, yielding up to ~1.7x speedup.
Number of GPUs
Speedup from CUDA-graphs
Mask R-CNN
272
1.70×
BERT
4096
1.12×
Table 1. MLPerf training v1.0 performance improvement with PyTorch CUDA graph.
Mask R-CNN | https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
Mask R-CNN
Deep learning frameworks use GPUs to accelerate computations, but a significant amount of code still runs on CPU cores. CPU cores process meta-data like tensor shapes in order to prepare arguments needed to launch GPU kernels. Processing meta-data is a fixed cost while the cost of the computational work done by the GPUs is positively correlated with batch size. For large batch sizes, CPU overhead is a negligible percentage of total run time cost, but at small batch sizes CPU overhead can become larger than GPU run time. When that happens, GPUs go idle between kernel calls. This issue can be identified on an NSight timeline plot in Figure 3. The plot below shows the “backbone” portion of Mask R-CNN with per-gpu batch size of 1 before graphing. The green portion shows CPU load while the blue portion shows GPU load. In this profile we see that the CPU is maxed out at 100% load while GPU is idle most of the time, there is a lot of empty space between GPU kernels.
| https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
Figure 3: NSight timeline plot of Mask R-CNN
CUDA graphs can automatically eliminate CPU overhead when tensor shapes are static. A complete graph of all the kernel calls is captured during the first step, in subsequent steps the entire graph is launched with a single op, eliminating all the CPU overhead, as observed in Figure 4..
Figure 4: CUDA graphs optimization
| https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
Figure 4: CUDA graphs optimization
With graphing, we see that the GPU kernels are tightly packed and GPU utilization remains high. The graphed portion now runs in 6 ms instead of 31ms, a speedup of 5x. We did not graph the entire model, mostly just the resnet backbone, which resulted in an overall speedup of ~1.7x.
In order to increase the scope of the graph, we made some changes in the software stack to eliminate some of the CPU-GPU synchronization points. In MLPerf v1.0, this work included changing the implementation of torch.randperm function to use CUB instead of Thrust because the latter is a synchronous C++ template library. These improvements are available in the latest NGC container.
BERT | https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
Similarly, by graph capturing the model, we eliminate CPU overhead and accompanying synchronization overhead. CUDA graphs implementation results in a 1.12x performance boost for our max-scale BERT configuration. To maximize the benefits from CUDA graphs, it is important to keep the scope of the graph as large as possible. To achieve this, we modified the model script to remove CPU-GPU synchronizations during the execution such that the full model can be graph captured. Furthermore, we also made sure that the tensor sizes during the execution are static within the scope of the graph. For instance, in BERT, only a specific subset of total tokens contribute to loss function, determined by a pre-generated mask tensor. Extracting the indices of valid tokens from this mask, and using these indices to gather the tokens that contribute to the loss, results in a tensor with a dynamic shape, i.e. with shape that is not constant across iterations. In order to make sure tensor sizes are static, instead of using the dynamic-shape tensors in the loss computation, we used static shape tensors where a mask is used to indicate which elements are valid. As a result, all tensor shapes are static. Dynamic shapes also require CPU-GPU synchronization since it has to involve the framework’s memory management on the CPU side. With static-only shapes, no CPU-GPU synchronizations are necessary. This is shown in Figure 5. | https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
Figure 5. By using a fixed size tensor and a boolean mask as described in the text, we are able to eliminate CPU synchronizations needed for dynamic sized tensors
CUDA graphs in NVIDIA DL examples collection
Single GPU use cases can also benefit from using CUDA Graphs. This is particularly true for workloads launching many short kernels with small batches. A good example is training and inference for recommender systems. Below we present preliminary benchmark results for NVIDIA's implementation of the Deep Learning Recommendation Model (DLRM) from our Deep Learning Examples collection. Using CUDA graphs for this workload provides significant speedups for both training and inference. The effect is particularly visible when using very small batch sizes, where CPU overheads are more pronounced. | https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
CUDA graphs are being actively integrated into other PyTorch NGC model scripts and the NVIDIA Github deep learning examples. Stay tuned for more examples on how to use it.
Figure 6: CUDA graphs optimization for the DLRM model.
Call to action: CUDA Graphs in PyTorch v1.10 | https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
Call to action: CUDA Graphs in PyTorch v1.10
CUDA graphs can provide substantial benefits for workloads that comprise many small GPU kernels and hence bogged down by CPU launch overheads. This has been demonstrated in our MLPerf efforts, optimizing PyTorch models. Many of these optimizations, including CUDA graphs, have or will eventually be integrated into our PyTorch NGC model scripts collection and the NVIDIA Github deep learning examples. For now, check out our open-source MLPerf training v1.0 implementation which could serve as a good starting point to see CUDA graph in action. Alternatively, try the PyTorch CUDA graphs API on your own workloads.
We thank many NVIDIAN’s and Facebook engineers for their discussions and suggestions: | https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
Karthik Mandakolathur US,
Tomasz Grel,
PLJoey Conway,
Arslan Zulfiqar US
Authors bios
Vinh Nguyen
DL Engineer, NVIDIA
Vinh is a Deep learning engineer and data scientist, having published more than 50 scientific articles attracting more than 2500 citations. At NVIDIA, his work spans a wide range of deep learning and AI applications, including speech, language and vision processing, and recommender systems.
Michael Carilli
Senior Developer Technology Engineer, NVIDIA | https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
Senior Developer Technology Engineer, NVIDIA
Michael worked at the Air Force Research Laboratory optimizing CFD code for modern parallel architectures. He holds a PhD in computational physics from the University of California, Santa Barbara. A member of the PyTorch team, he focuses on making GPU training fast, numerically stable, and easy(er) for internal teams, external customers, and Pytorch community users.
Sukru Burc Eryilmaz
Senior Architect in Dev Arch, NVIDIA
Sukru received his PhD from Stanford University, and B.S from Bilkent University. He currently works on improving the end-to-end performance of neural network training both at single-node scale and supercomputer scale.
Vartika Singh
Tech Partner Lead for DL Frameworks and Libraries, NVIDIA | https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
Vartika has led teams working in confluence of cloud and distributed computing, scaling and AI, influencing the design and strategy of major corporations. She currently works with the major frameworks and compiler organizations and developers within and outside NVIDIA, to help the design to work efficiently and optimally on NVIDIA hardware.
Michelle Lin
Product Intern, NVIDIA
Michelle is currently pursuing an undergraduate degree in Computer Science and Business Administration at UC Berkeley. She is currently managing execution of projects such as conducting market research and creating marketing assets for Magnum IO.
Natalia Gimelshein
Applied Research Scientist, Facebook
Natalia Gimelshein worked on GPU performance optimization for deep learning workloads at NVIDIA and Facebook. She is currently a member of the PyTorch core team, working with partners to seamlessly support new software and hardware features. | https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
Alban Desmaison
Research Engineer, Facebook
Alban studied engineering and did a PhD in Machine Learning and Optimization, during which he was an OSS contributor to PyTorch prior to joining Facebook. His main responsibilities are maintaining some core library and features (autograd, optim, nn) and working on making PyTorch better in general.
Edward Yang
Research Engineer, Facebook
Edward studied CS at MIT and then Stanford before starting at Facebook. He is a part of the PyTorch core team and is one of the leading contributors to PyTorch. | https://pytorch.org/blog/accelerating-pytorch-with-cuda-graphs/ | pytorch blogs |
layout: blog_detail
title: 'PyTorch 1.10 Release, including CUDA Graphs APIs, Frontend and Compiler Improvements'
author: Team PyTorch
We are excited to announce the release of PyTorch 1.10. This release is composed of over 3,400 commits since 1.9, made by 426 contributors. We want to sincerely thank our community for continuously improving PyTorch.
PyTorch 1.10 updates are focused on improving training and performance of PyTorch, and developer usability. The full release notes are available here. Highlights include:
1. CUDA Graphs APIs are integrated to reduce CPU overheads for CUDA workloads.
2. Several frontend APIs such as FX, torch.special, and nn.Module Parametrization, have moved from beta to stable.
3. Support for automatic fusion in JIT Compiler expands to CPUs in addition to GPUs.
4. Android NNAPI support is now available in beta. | https://pytorch.org/blog/pytorch-1.10-released/ | pytorch blogs |
Along with 1.10, we are also releasing major updates to the PyTorch libraries, which you can read about in this blog post.
Frontend APIs
(Stable) Python code transformations with FX
FX provides a Pythonic platform for transforming and lowering PyTorch programs. It is a toolkit for pass writers to facilitate Python-to-Python transformation of functions and nn.Module instances. This toolkit aims to support a subset of Python language semantics—rather than the whole Python language—to facilitate ease of implementation of transforms. With 1.10, FX is moving to stable.
You can learn more about FX in the official documentation and GitHub examples of program transformations implemented using torch.fx.
(Stable) torch.special | https://pytorch.org/blog/pytorch-1.10-released/ | pytorch blogs |
(Stable) torch.special
A torch.special module, analogous to SciPy’s special module, is now available in stable. The module has 30 operations, including gamma, Bessel, and (Gauss) error functions.
Refer to this documentation for more details.
(Stable) nn.Module Parametrization
nn.Module parametrizaton, a feature that allows users to parametrize any parameter or buffer of an nn.Module without modifying the nn.Module itself, is available in stable. This release adds weight normalization (weight_norm), orthogonal parametrization (matrix constraints and part of pruning) and more flexibility when creating your own parametrization. | https://pytorch.org/blog/pytorch-1.10-released/ | pytorch blogs |
Refer to this tutorial and the general documentation for more details.
(Beta) CUDA Graphs APIs Integration
PyTorch now integrates CUDA Graphs APIs to reduce CPU overheads for CUDA workloads.
CUDA Graphs greatly reduce the CPU overhead for CPU-bound cuda workloads and thus improve performance by increasing GPU utilization. For distributed workloads, CUDA Graphs also reduce jitter, and since parallel workloads have to wait for the slowest worker, reducing jitter improves overall parallel efficiency.
Integration allows seamless interop between the parts of the network captured by cuda graphs, and parts of the network that cannot be captured due to graph limitations. | https://pytorch.org/blog/pytorch-1.10-released/ | pytorch blogs |
Read the note for more details and examples, and refer to the general documentation for additional information.
[Beta] Conjugate View
PyTorch’s conjugation for complex tensors (torch.conj()) is now a constant time operation, and returns a view of the input tensor with a conjugate bit set as can be seen by calling torch.is_conj() . This has already been leveraged in various other PyTorch operations like matrix multiplication, dot product etc., to fuse conjugation with the operation leading to significant performance gain and memory savings on both CPU and CUDA.
Distributed Training
Distributed Training Releases Now in Stable | https://pytorch.org/blog/pytorch-1.10-released/ | pytorch blogs |
Distributed Training Releases Now in Stable
In 1.10, there are a number of features that are moving from beta to stable in the distributed package:
* (Stable) Remote Module: This feature allows users to operate a module on a remote worker like using a local module, where the RPCs are transparent to the user. Refer to this documentation for more details.
* (Stable) DDP Communication Hook: This feature allows users to override how DDP synchronizes gradients across processes. Refer to this documentation for more details. | https://pytorch.org/blog/pytorch-1.10-released/ | pytorch blogs |
(Stable) ZeroRedundancyOptimizer: This feature can be used in conjunction with DistributedDataParallel to reduce the size of per-process optimizer states. With this stable release, it now can handle uneven inputs to different data-parallel workers. Check out this tutorial. We also improved the parameter partition algorithm to better balance memory and computation overhead across processes. Refer to this documentation and this tutorial to learn more.
Performance Optimization and Tooling
[Beta] Profile-directed typing in TorchScript | https://pytorch.org/blog/pytorch-1.10-released/ | pytorch blogs |
[Beta] Profile-directed typing in TorchScript
TorchScript has a hard requirement for source code to have type annotations in order for compilation to be successful. For a long time, it was only possible to add missing or incorrect type annotations through trial and error (i.e., by fixing the type-checking errors generated by torch.jit.script one by one), which was inefficient and time consuming.
Now, we have enabled profile directed typing for torch.jit.script by leveraging existing tools like MonkeyType, which makes the process much easier, faster, and more efficient. For more details, refer to the documentation.
(Beta) CPU Fusion
In PyTorch 1.10, we've added an LLVM-based JIT compiler for CPUs that can fuse together sequences of torch library calls to improve performance. While we've had this capability for some time on GPUs, this release is the first time we've brought compilation to the CPU. | https://pytorch.org/blog/pytorch-1.10-released/ | pytorch blogs |
You can check out a few performance results for yourself in this Colab notebook.
(Beta) PyTorch Profiler
The objective of PyTorch Profiler is to target the execution steps that are the most costly in time and/or memory, and visualize the workload distribution between GPUs and CPUs. PyTorch 1.10 includes the following key features:
Enhanced Memory View: This helps you understand your memory usage better. This tool will help you avoid Out of Memory errors by showing active memory allocations at various points of your program run.
Enhanced Automated Recommendations: This helps provide automated performance recommendations to help optimize your model. The tools recommend changes to batch size, TensorCore, memory reduction technologies, etc.
Enhanced Kernel View: Additional columns show grid and block sizes as well as shared memory usage and registers per thread.
| https://pytorch.org/blog/pytorch-1.10-released/ | pytorch blogs |
Distributed Training: Gloo is now supported for distributed training jobs.
Correlate Operators in the Forward & Backward Pass: This helps map the operators found in the forward pass to the backward pass, and vice versa, in a trace view.
TensorCore: This tool shows the Tensor Core (TC) usage and provides recommendations for data scientists and framework developers.
NVTX: Support for NVTX markers was ported from the legacy autograd profiler.
Support for profiling on mobile devices: The PyTorch profiler now has better integration with TorchScript and mobile backends, enabling trace collection for mobile workloads.
Refer to this documentation for details. Check out this tutorial to learn how to get started with this feature.
PyTorch Mobile
(Beta) Android NNAPI Support in Beta | https://pytorch.org/blog/pytorch-1.10-released/ | pytorch blogs |
(Beta) Android NNAPI Support in Beta
Last year we released prototype support for Android’s Neural Networks API (NNAPI). NNAPI allows Android apps to run computationally intensive neural networks on the most powerful and efficient parts of the chips that power mobile phones, including GPUs (Graphics Processing Units) and NPUs (specialized Neural Processing Units).
Since the prototype we’ve added more op coverage, added support for load-time flexible shapes and ability to run the model on the host for testing. Try out this feature using the tutorial. | https://pytorch.org/blog/pytorch-1.10-released/ | pytorch blogs |
Additionally, Transfer Learning steps have been added to Object Detection examples. Check out this GitHub page to learn more. Please provide your feedback or ask questions on the forum. You can also check out this presentation to get an overview.
Thanks for reading. If you’re interested in these updates and want to join the PyTorch community, we encourage you to join the discussion forums and open GitHub issues. To get the latest news from PyTorch, follow us on Twitter, Medium, YouTube, and LinkedIn.
Cheers!
Team PyTorch | https://pytorch.org/blog/pytorch-1.10-released/ | pytorch blogs |
layout: blog_detail
title: "Ambient Clinical Intelligence: Generating Medical Reports with PyTorch"
author: Miguel Del-Agua, Principal Research Scientist, Nuance and Jeremy Jancsary, Senior Principal Research Scientist, Nuance
featured-img: ""
Introduction
Complete and accurate clinical documentation is an essential tool for tracking patient care. It allows for treatment plans to be shared among care teams to aid in continuity of care and ensures a transparent and effective process for reimbursement. | https://pytorch.org/blog/ambient-clinical-intelligence-generating-medical-reports-with-pytorch/ | pytorch blogs |
Physicians are responsible for documenting patient care. Traditional clinical documentation methods have resulted in a sub-par patient-provider experience, less time interacting with patients, and decreased work-life balance. A significant amount of physicians’ time is spent in front of the computer doing administrative tasks. As a result, patients are less satisfied with the overall experience, and physicians, who prepare for years studying medicine, cannot practice at the top of their license and are burned out. Every hour physicians provide direct clinical face time to patients results in nearly two additional hours spent on EHR and desk work within the clinic day. Outside office hours, physicians spend another 1 to 2 hours of personal time each night doing additional computer and other clerical work.
42% of all physicians reported having burnout. – Medscape
| https://pytorch.org/blog/ambient-clinical-intelligence-generating-medical-reports-with-pytorch/ | pytorch blogs |
The problem has grown worse due to the pandemic with 64% of U.S. physicians now reporting burnout. - AAFP
"Too many bureaucratic tasks e.g., charting and paperwork" is the leading contribution to burnout, increased computerization ranks 4th. - Medscape
75% of U.S. Consumers Wish Their Healthcare Experiences Were More Personalized,- Business Wire
| https://pytorch.org/blog/ambient-clinical-intelligence-generating-medical-reports-with-pytorch/ | pytorch blogs |
61% of patients would visit their healthcare provider more often if the communication experience felt more personalized. – Business Wire
Physician burnout is one of the primary causes for increased medical errors, malpractice suits, turnover, and decreased access to care. Burnout leads to an increase in healthcare costs and a decrease in overall patient satisfaction. Burnout costs the United States $4.6 billion a year. | https://pytorch.org/blog/ambient-clinical-intelligence-generating-medical-reports-with-pytorch/ | pytorch blogs |
What can we do to bring back trust, joy, and humanity to the delivery of healthcare? A significant portion of the administrative work consists of entering patient data into Electronic Health Records (EHRs) and creating clinical documentation. Clinical documentation is created from information already in the EHR as well as from the patient-provider encounter conversation.
This article will showcase how the Nuance Dragon Ambient eXperience (DAX), an AI-powered, voice-enabled, ambient clinical intelligence solution, automatically documents patient encounters accurately and efficiently at the point of care and the technologies that enable it.
Nuance DAX enhances the quality of care and patient experience, increases provider efficiency and satisfaction, and improves financial outcomes. It can be used in office and telehealth settings in all ambulatory specialties, including primary and urgent care.
| https://pytorch.org/blog/ambient-clinical-intelligence-generating-medical-reports-with-pytorch/ | pytorch blogs |
Natural Language Processing
Natural Language Processing (NLP) is one of the most challenging fields in Artificial Intelligence (AI). It comprehends a set of algorithms that allow computers to understand or generate the language used by humans. These algorithms can process and analyze vast amounts of natural language data from different sources (either sound or text) to build models that can understand, classify, or even generate natural language as humans would. Like other fields in AI, NLP has significantly progressed thanks to the advent of Deep Learning (DL), which has resulted in models that can obtain results on par with humans in some tasks. | https://pytorch.org/blog/ambient-clinical-intelligence-generating-medical-reports-with-pytorch/ | pytorch blogs |
These advanced NLP techniques are being applied in healthcare. During a typical patient-provider encounter, a conversation ensues where the doctor constructs, through questions and answers, a chronological description of the development of the patient's presenting illness or symptoms. A physician examines the patient and makes clinical decisions to establish a diagnosis and determine a treatment plan. This conversation, and data in the EHR, provide the required information for physicians to generate the clinical documentation, referred to as medical reports. | https://pytorch.org/blog/ambient-clinical-intelligence-generating-medical-reports-with-pytorch/ | pytorch blogs |
Two main NLP components play a role in automating the creation of clinical documentation. The first component, Automatic Speech Recognition (ASR), is used to translate speech into text. It takes the audio recording of the encounter and generates a conversation transcription (cf. Figure 2). The second component, Automatic Text Summarization, helps generate summaries from large text documents. This component is responsible for understanding and capturing the nuances and most essential aspects from the transcribed conversation into a final report in narrative form (cf. Figure 3), structured form, or a combination of both.
We will focus on this second component, Automatic Text Summarization, which is a difficult task with many challenges:
Its performance is tied to the ASR quality from multiple speakers (noisy input).
The input is conversational in nature and contains layman's terms.
Protected Health Information (PHI) regulations limit medical data access.
| https://pytorch.org/blog/ambient-clinical-intelligence-generating-medical-reports-with-pytorch/ | pytorch blogs |
The information for one output sentence is potentially spread across multiple conversation turns.
There is no explicit sentence alignment between input and output.
Various medical specialties, encounter types, and EHR systems constitute a broad and complex output space.
Physicians have different styles of conducting encounters and have their preferences for medical reports; there is no standard.
Standard summarization metrics might differ from human judgment of quality.
Figure 2: Transcript of a patient-doctor conversation
Figure 3: Excerpt of an AI-generated medical report. HPI stands for History of present illness.
Text Summarization with PyTorch and Fairseq | https://pytorch.org/blog/ambient-clinical-intelligence-generating-medical-reports-with-pytorch/ | pytorch blogs |
Text Summarization with PyTorch and Fairseq
PyTorch is an open-source machine learning framework developed by Facebook that helps researchers prototype Deep Learning models. The Fairseq toolkit is built on top of PyTorch and focuses on sequence generation tasks, such as Neural Machine Translation (NMT) or Text Summarization. Fairseq features an active community that is continuously providing reference implementations of state-of-the-art models. It contains many built-in components (model architectures, modules, loss functions, and optimizers) and is easily extendable with plugins. | https://pytorch.org/blog/ambient-clinical-intelligence-generating-medical-reports-with-pytorch/ | pytorch blogs |
Text summarization constitutes a significant challenge in NLP. We need models capable of generating a short version of a document while retaining the key points and avoiding uninformative content. These challenges can be addressed with different approaches. 1). Abstractive text summarization aimed at training models that can generate a summary in narrative form. 2). Extractive methods where the models are trained to select the most important parts from the input text. 3). A combination of the two, where the essential parts from the input are selected and then summarized in an abstractive fashion. Hence, summarization can be accomplished via a single end-to-end network or as a pipeline of extractive and abstractive components. To that end, Fairseq provides all the necessary tools to be successful in our endeavor. It features either end-to-end models such as the classical Transformer, different types of Language Models and pre-trained versions that enable researchers to focus on what matters most—to build state-of-the-art models that generate valuable reports. | https://pytorch.org/blog/ambient-clinical-intelligence-generating-medical-reports-with-pytorch/ | pytorch blogs |
However, we are not just summarizing the transcribed conversation; we generate high-quality medical reports, which have many considerations.
Every section of a medical report is different in terms of content, structure, fluency, etc.
All medical facts mentioned in the conversation should be present in the report, for example, a particular treatment or dosage.
In the healthcare domain, the vocabulary is extensive, and models need to deal with medical terminology.
Patient-doctor conversations are usually much longer than the final report.
| https://pytorch.org/blog/ambient-clinical-intelligence-generating-medical-reports-with-pytorch/ | pytorch blogs |
All these challenges require our researchers to run a battery of extensive experiments. Thanks to the flexibility of PyTorch and Fairseq, their productivity has greatly increased. Further, the ecosystem offers an easy path from ideation, implementation, experimentation, and final roll-out to production. Using multiple GPUs or CPUs is as simple as providing an additional argument to the tools, and because of the tight Python integration, PyTorch code can be easily debugged. | https://pytorch.org/blog/ambient-clinical-intelligence-generating-medical-reports-with-pytorch/ | pytorch blogs |
In our continuous effort to contribute to the open-source community, features have been developed at Nuance and pushed to the Fairseq GitHub repository. These try to overcome some of the challenges mentioned such as, facilitating copying of, especially rare or unseen, words from the input to summary, training speedups by improving Tensor Core utilization, and ensuring TorchScript compatibility of different Transformer configurations. Following, we will show an example of how to train a Transformer model with a Pointer Generator mechanism (Transformer-PG), which can copy words from the input.
How to build a Transformer model with a Pointer Generator mechanism
In this step-by-step guide, it is assumed the user has already installed PyTorch and Fairseq.
1. Create a vocabulary and extend it with source position markers:
These markers will allow the model to point to any word in the input sequence.
```python
vocab_size=
position_markers=512
export LC_ALL=C
cat train.src train.tgt | | https://pytorch.org/blog/ambient-clinical-intelligence-generating-medical-reports-with-pytorch/ | pytorch blogs |
export LC_ALL=C
cat train.src train.tgt |
tr -s '[:space:]' '\n' |
sort |
uniq -c |
sort -k1,1bnr -k2 |
head -n "$((vocab_size - 4))" |
awk '{ print $2 " " $1 }' > dict.pg.txt
python3 -c "[print(' 0'.format(n)) for n in range($position_markers)]" >> dict.pg.txt
This will create a file "dict.pg.txt" that contains the \<vocab_size> most frequent words followed by 512 position markers named from "\<unk-0>" to "\<unk-511>".
In case we have an input like
```python
src = "Hello, I'm The Dogtor"
it could happen that our model has been trained without the word "Dogtor" in its vocabulary. Therefore, when we feed this sequence into the model, it should be converted to:
src = "Hello, I'm The <unk-3>"
Now, "\" is part of our vocabulary and could be predicted by the model (this is where the pointer-generator comes in). In such a case, we will only need to post-process the output to replace "\" by the word at input position 3. | https://pytorch.org/blog/ambient-clinical-intelligence-generating-medical-reports-with-pytorch/ | pytorch blogs |
2. Preprocess the text data to replace unknown words by its positional markers:
We can use the scripts from https://github.com/pytorch/fairseq/tree/master/examples/pointer_generator.
```python
Considering we have our data in:
train_src = /path/to/train.src
train_tgt = /path/to/train.tgt
valid_src = /path/to/valid.src
valid_tgt = /path/to/valid.tgt
./preprocess.py --source /path/to/train.src \
--target /path/to/train.tgt \
--vocab <(cut -d' ' -f1 dict.pg.txt) \
--source-out /path/to/train.pg.src \
--target-out /path/to/train.pg.tgt
./preprocess.py --source /path/to/valid.src \
--target /path/to/valid.tgt \
--vocab <(cut -d' ' -f1 dict.pg.txt) \
--source-out /path/to/valid.pg.src \
--target-out /path/to/valid.pg.tgt
./preprocess.py --source /path/to/test.src \ | https://pytorch.org/blog/ambient-clinical-intelligence-generating-medical-reports-with-pytorch/ | pytorch blogs |
./preprocess.py --source /path/to/test.src \
--vocab <(cut -d' ' -f1 dict.pg.txt) \
--source-out /path/to/test.pg.src
### 3. Now let's binarize the data, so that it can be processed faster:
```python
fairseq-preprocess --task "translation" \
--source-lang "pg.src" \
--target-lang "pg.tgt" \
--trainpref /path/to/train \
--validpref /path/to/valid \
--srcdict dict.pg.txt \
--cpu \
--joined-dictionary \
--destdir <data_dir>
You might notice the type of task is "translation". This is because there is no "summarization" task available; we could understand it as a kind of NMT task where the input and output languages are shared and the output (summary) is shorter than the input.
4. Now we can train the model:
```python
fairseq-train \
--save-dir \ | https://pytorch.org/blog/ambient-clinical-intelligence-generating-medical-reports-with-pytorch/ | pytorch blogs |
--save-dir \
--task "translation" \
--source-lang "src" \
--target-lang "tgt" \
--arch "transformer_pointer_generator" \
--max-source-positions 512 \
--max-target-positions 128 \
--truncate-source \
--max-tokens 2048 \
--required-batch-size-multiple 1 \
--required-seq-len-multiple 8 \
--share-all-embeddings \
--dropout 0.1 \
--criterion "cross_entropy" \
--optimizer adam \
--adam-betas '(0.9, 0.98)' \
--adam-eps 1e-9 \
--update-freq 4 \
--lr 0.004 \
# Pointer Generator
--alignment-layer -1 \
--alignment-heads 1 \
--source-position-markers 512
```
This configuration makes use of features Nuance has contributed back to Fairseq: | https://pytorch.org/blog/ambient-clinical-intelligence-generating-medical-reports-with-pytorch/ | pytorch blogs |
Transformer with a Pointer Generator mechanism to facilitate copying of words from the input.
Sequence length padded to a multiple of 8 to better use tensor cores and reduce training time.
5. Now let's take a look at how to generate a summary with our new medical report generation system:
```python
import torch
from examples.pointer_generator.pointer_generator_src.transformer_pg import TransformerPointerGeneratorModel
Patient-Doctor conversation
input = "[doctor] Lisa Simpson, thirty six year old female, presents to the clinic today because " \
"she has severe right wrist pain"
Load the model
model = TransformerPointerGeneratorModel.from_pretrained(data_name_or_path=,
model_name_or_path=,
checkpoint_file="checkpoint_best.pt")
result = model.translate([input], beam=2)
print(result[0]) | https://pytorch.org/blog/ambient-clinical-intelligence-generating-medical-reports-with-pytorch/ | pytorch blogs |
print(result[0])
Ms. is a 36-year-old female who presents to the clinic today for evaluation of her right wrist.
### 6. Alternatively, we can use fairseq-interactive and a postprocessing tool to substitute positional unknown tokens by its words from the input:
```python
fairseq-interactive <data_dir> \
--batch-size <batch_size> \
--task translation \
--source-lang src \
--target-lang tgt \
--path <model_dir>/checkpoint_last.pt \
--input /path/to/test.pg.src \
--buffer-size 20 \
--max-len-a 0 \
--max-len-b 128 \
--beam 2 \
--skip-invalid-size-inputs-valid-test | tee generate.out
grep "^H-" generate.out | cut -f 3- > generate.hyp
./postprocess.py \
--source <(awk 'NF<512' /path/to/test.pg.src) \
--target generate.hyp \
--target-out generate.hyp.processed
| https://pytorch.org/blog/ambient-clinical-intelligence-generating-medical-reports-with-pytorch/ | pytorch blogs |
--target-out generate.hyp.processed
```
Now we have the final set of reports in "generate.hyp.processed", with "\" replaced by the original word from the input sequence.
Model Deployment
PyTorch offers great flexibility in modeling and a rich surrounding ecosystem. However, while several recent articles have suggested that the use of PyTorch in research and academia may be close to surpassing TensorFlow, there seems to be an overall sense of TensorFlow being the preferred platform for deployment to production. Is this still the case in 2021? Teams looking to serve their PyTorch models in production have a few options.
Before describing our journey, let's take a brief detour and define the term model.
Models as computation graphs | https://pytorch.org/blog/ambient-clinical-intelligence-generating-medical-reports-with-pytorch/ | pytorch blogs |
A few years back, it was still common for machine learning toolkits to support only particular classes of models of a rather fixed and rigid structure, with only a few degrees of freedom (like the kernel of a support vector machine or the number of hidden layers of a neural network). Inspired by foundational work in Theano, toolkits like Microsoft's CNTK or Google's TensorFlow were among the first to popularize a more flexible view on models, as computation graphs with associated parameters that can be estimated from data. This view blurred the boundaries between popular types of models (such as DNNs or SVMs), as it became easy to blend the characteristics of each into your type of graph. Still, such a graph had to be defined upfront before estimating its parameters, and it was pretty static. This made it easy to save models to a self-contained bundle, like a TensorFlow SavedModel (such a bundle simply contains the structure of the graph, as well as the concrete values of the estimated parameters). However, debugging such models can be difficult because the statements in the Python code that build the graph are logically separate from the lines that execute it. Researchers also long for easier ways of expressing dynamic behavior, such as the computation steps of the forward pass of a model being conditionally dependent on its input data (or its previous output). | https://pytorch.org/blog/ambient-clinical-intelligence-generating-medical-reports-with-pytorch/ | pytorch blogs |
Most recently, the above limitations have led to a second revolution spearheaded by PyTorch and TensorFlow 2. The computation graph is no longer defined explicitly. Instead, it will be populated implicitly as the Python code executes operations on tensor arguments. An essential technique that powers this development is automatic differentiation. As the computation graph is being built implicitly while executing the steps of the forward pass, all the necessary data will be tracked for later computation of the gradient concerning the model parameters. This allows for great flexibility in training a model, but it raises an important question. If the computation happening inside a model is only implicitly defined through our Python code's steps as it executes concrete data, what is it that we want to save as a model? The answer – at least initially – was the Python code with all its dependencies, along with the estimated parameters. This is undesirable for practical reasons. For instance, there is a danger that the team working on model deployment does not exactly reproduce the Python code dependencies used during training, leading to subtly divergent behavior. The solution typically consists of combining two techniques, scripting and tracing, that is, extra annotations in your Python code and execution of your code on exemplary input data, allowing PyTorch to define and save the graph that should be executed during later inference on new, unseen data. This requires some discipline by whoever creates the model code (arguably voiding some of the original flexibility of eager execution), but it results in a self-contained model bundle in TorchScript format. The solution in TensorFlow 2 is remarkably similar. | https://pytorch.org/blog/ambient-clinical-intelligence-generating-medical-reports-with-pytorch/ | pytorch blogs |
Serving our report generation models | https://pytorch.org/blog/ambient-clinical-intelligence-generating-medical-reports-with-pytorch/ | pytorch blogs |
Our journey in deploying the report generation models reflects the above discussion. We started out serving our models by deploying the model code and its dependencies along with the parameter checkpoints in a custom Docker image exposing a gRPC service interface. However, we soon noticed that it became error-prone to replicate the exact code and environment used by the modeling team while estimating the parameters. Moreover, this approach prevented us from leveraging high-performance model serving frameworks like NVIDIA's Triton, which is written in C++ and requires self-contained models that can be used without a Python interpreter. At this stage, we were facing a choice between attempting to export our PyTorch models to ONNX or TorchScript format. ONNX is an open specification for representing machine learning models that increasingly finds adoption. It is powered by a high-performance runtime developed by Microsoft (ONNX Runtime). While we were able to achieve performance acceleration for our TensorFlow BERT-based model using ONNX Runtime, at the time one of our PyTorch model required some operators that weren’t yet supported in ONNX. Rather than implement these using custom operators, we decided to look into TorchScript for the time being. | https://pytorch.org/blog/ambient-clinical-intelligence-generating-medical-reports-with-pytorch/ | pytorch blogs |
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.