Today we are announcing the release of ZeRO-3 Offload, a highly efficient and easy to use implementation of ZeRO Stage 3 and ZeRO Offload combined, geared towards our continued goal of democratizing AI by making efficient large-scale DL training available to everyone. The key benefits of ZeRO-3 Offload are:
While DeepSpeed supports training advanced large-scale models, using these trained models in the desired application scenarios is still challenging due to three major limitations in existing inference solutions: 1) lack of support for multi-GPU inference to fit large models and meet latency requirements, 2) limited GPU kernel performance when running inference with small batch sizes, and 3) difficulties in exploiting quantization, which includes both quantizing the model to reduce the model size and latency as well as supporting high-performance inference of quantized models without specialized hardware.
The field of Artificial Intelligence-Generated Content (AIGC) is rapidly growing, with the goal of making content creation more efficient and accessible. One of the most exciting areas of AIGC is the development of large-scale multi-modal models like [Flamingo](https://arxiv.org/abs/2204.14198), [BLIP](https://arxiv.org/abs/2301.12597), and [GPT4](https://arxiv.org/abs/2303.08774), which can accept inputs from multiple resources, e.g., image, text, audio, etc., and generate a variety of formats as outputs. For example, image creation can be made through stable diffusion and DALLE using the prompt text, and the new feature in the coming Office can create slides with texts, images, animations, etc., by leveraging the power of the new Microsoft Office Copilot.
Scaling up the model size is one common approach to boost usability and capability of AIGC tasks. However, simply scaling up dense architectures (e.g., from GPT-1 to GPT-3) is usually extremely resource-intensive and time-consuming for both model training and inference. One effective way to tackle this challenge is to apply mixture of experts (MoE). In particular, recent [text-based MoE](https://arxiv.org/abs/2201.05596) and [vision-based MoE](https://arxiv.org/abs/2106.05974) studies have demonstrated that MoE models can significantly reduce the training and resource cost as compared to a quality-equivalent dense model, or produce a higher quality model under the same training budget. Up to now, the effectiveness of jointly training MoE for multi-modal models remains not well understood. To explore this important capability, [DeepSpeed team](https://www.deepspeed.ai/) is proud to announce our first large-scale generative mixture-of-expert (MoE) multimodal model, named [VL-MoE](https://arxiv.org/abs/2303.07226).
*Figure 1: The new encoding process in our VL-MoE for various modality inputs, for which gray and colored blocks indicate non-activated and activated modules, respectively.*
Specifically, we incorporate the MoE structure into the classical single-tower multi-modal model by comprising of the following components: (1) a shared self-attention module across modalities, (2) a pool of modality-specific experts in the feed-forward network (FFN), and (3) a sparse gated MoE extended from the dense FFN. Subsequently, under the same amount of training resources as that used in [VLMO](https://arxiv.org/abs/2111.02358)(200k training steps), we demonstrate VL-MoE's advantages over the state-of-the-art dense counterparts in the following two aspects:
(1) **VL-MoE can achieve significant accuracy improvement in comparison to its dense counterparts.** Table 1 demonstrates that under the same training budget (i.e., have the same number of activated parameters for each token), VL-MoE Base with 32 experts achieves better accuracy than the VLMO-Base dense model on all four vision-language datasets.
(2) **VL-MoE achieves similar model quality with a much smaller activated number of parameters compared to its dense counterparts.** Our results show that the finetuning performance of our VL-MoE is similar to that of the 3.1X larger VLMO-Large dense model (i.e., 3.1X more activated number of parameters per token). This can directly translate to approximately 3.1X training cost reduction as the training FLOPs for transformers are proportional to the activated model size per token.
| | Param per Token (# Total Param) | VQA | NLVR2 | COCO | Flickr30K |
| | | test-dev / std | dev / test-P | TR / IR | TR / IR |
*Table 1: Comparison of finetuning accuracy results for different models used in vision-language classification tasks and image-text retrieval tasks.*
A sophisticated MoE model design requires a highly efficient and scalable training system that can support multi-dimensional parallelism and efficient memory management. [DeepSpeed MoE](https://www.microsoft.com/en-us/research/blog/deepspeed-advancing-moe-inference-and-training-to-power-next-generation-ai-scale/) training system offers such advanced capabilities including easy-to-use APIs enabling flexible combinations of data, tensor, and expert parallelism. Furthermore, DeepSpeed MoE enables larger model scale than state-of-the-art systems by exploiting expert parallelism and [ZeRO optimizations](https://arxiv.org/abs/1910.02054) together. By leveraging the DeepSpeed MoE system, VL-MoE Base with 32 experts achieves similar model quality as VLMO-dense Large with about 2.5x training speedup.
[DeepSpeed MoE](https://www.microsoft.com/en-us/research/blog/deepspeed-advancing-moe-inference-and-training-to-power-next-generation-ai-scale/) system is already open-sourced and can be easily used as plug-and-play component to achieve high-performance low-cost training for any large-scale MoE models. The tutorial of how to use DeepSpeed MoE is available [here](https://www.deepspeed.ai/tutorials/mixture-of-experts/). VL-MoE is currently in the process of being integrated as a model example of [DeepSpeed Examples](https://github.com/microsoft/DeepSpeedExamples). Please stay tuned for our upcoming updates on this thread.
@@ -65,7 +66,7 @@ With automatic tensor parallelism, we do not need to provide the injection polic
# Example Script
We can observe performance improvement with automatic tensor parallelism using the [inference test suite](https://github.com/microsoft/DeepSpeedExamples/blob/master/inference/huggingface/text-generation/inference-test.py). The script includes per token latency, bandwidth, throughput and memory checks for comparison. See the [README](https://github.com/microsoft/DeepSpeedExamples/tree/master/inference/huggingface/text-generation#deepspeed-huggingface-text-generation-examples) for more information.
We can observe performance improvement with automatic tensor parallelism using the [inference test suite](https://github.com/microsoft/DeepSpeedExamples/blob/master/inference/huggingface/text-generation/inference-test.py). This script is for testing text-generation models and includes per token latency, bandwidth, throughput and memory checks for comparison. See the [README](https://github.com/microsoft/DeepSpeedExamples/tree/master/inference/huggingface/text-generation#deepspeed-huggingface-text-generation-examples) for more information.
## Launching
...
...
@@ -83,19 +84,31 @@ To enable tensor parallelism, you need to use the flag `ds_inference` for the co
@@ -201,7 +201,7 @@ the `--predict_batch_size` should also be 8.
For further details about the transformer kernel, please see our [usage
tutorial](/tutorials/transformer_kernel/) and [technical deep
dive](https://www.deepspeed.ai/news/2020/05/27/fastest-bert-training.html) on
dive](https://www.deepspeed.ai/2020/05/27/fastest-bert-training.html) on
the fastest BERT training.
...
...
@@ -302,7 +302,7 @@ Table 4. The setting of memory-optimization flags for a range of micro-batch siz
### FineTuning model pre-trained with DeepSpeed Transformer Kernels
Fine-tuning the model pre-trained using DeepSpeed Transformer and the recipe in [DeepSpeed Fast-Bert Training](https://www.deepspeed.ai/news/2020/05/27/fastest-bert-training.html) should yield F1 score of 90.5 and is expected to increase if you let the pre-training longer than suggested in the tutorial.
Fine-tuning the model pre-trained using DeepSpeed Transformer and the recipe in [DeepSpeed Fast-Bert Training](https://www.deepspeed.ai/2020/05/27/fastest-bert-training.html) should yield F1 score of 90.5 and is expected to increase if you let the pre-training longer than suggested in the tutorial.
To get these results, we do require some tuning of the dropout settings as described below:
Compared to SOTA, DeepSpeed significantly improves single GPU performance for transformer-based model like BERT. Figure above shows the single GPU throughput of training BertBERT-Large optimized through DeepSpeed, compared with two well-known Pytorch implementations, NVIDIA BERT and HuggingFace BERT. DeepSpeed reaches as high as 64 and 53 teraflops throughputs (corresponding to 272 and 52 samples/second) for sequence lengths of 128 and 512, respectively, exhibiting up to 28% throughput improvements over NVIDIA BERT and up to 62% over HuggingFace BERT. We also support up to 1.8x larger batch size without running out of memory.
For more details on how we achieve the record breaking BERT training time please check out deep dive into DeepSpeed BERT [Fastest BERT Training](https://www.deepspeed.ai/news/2020/05/18/bert-record.html)
For more details on how we achieve the record breaking BERT training time please check out deep dive into DeepSpeed BERT [Fastest BERT Training](https://www.deepspeed.ai/2020/05/18/bert-record.html)
