- [2023/11] DeepSpeed-FastGen: High-throughput Text Generation for LLMs via MII and DeepSpeed-Inference
- [2022/11] Stable Diffusion Image Generation under 1 second w. DeepSpeed MII
- [2022/10] Announcing DeepSpeed Model Implementations for Inference (MII)
Introducing MII, an open-source Python library designed by DeepSpeed to democratize powerful model inference with a focus on high-throughput, low latency, and cost-effectiveness.
- MII v0.1 introduces several features such as blocked KV-caching, continuous batching, Dynamic SplitFuse, tensor parallelism, and high-performance CUDA kernels to support fast high throughput text-generation for LLMs such as Llama-2-70B. MII delivers up to 2.3 times higher effective throughput compared to leading systems such as vLLM. For detailed performance results please see our DeepSpeed-FastGen blog.
- We first announced MII in 2022, which covers all prior releases up to v0.0.9. In addition to language models, we also support accelerating text2image models like Stable Diffusion. For more details on our previous releases please see our legacy APIs.
MII provides accelerated text-generation inference through the use of four key technologies:
- Blocked KV Caching
- Continuous Batching
- Dynamic SplitFuse
- High Performance CUDA Kernels
For a deeper dive into understanding these features please refer to our blog which also includes a detailed performance analysis.
In the past, MII introduced several key performance optimizations for low-latency serving scenarios:
- DeepFusion for Transformers
- Multi-GPU Inference with Tensor-Slicing
- ZeRO-Inference for Resource Constrained Systems
- Compiler Optimizations
Figure 1: MII architecture, showing how MII automatically optimizes OSS models using DS-Inference before deploying them. DeepSpeed-FastGen optimizations in the figure have been published in our blog post.
Under-the-hood MII is powered by DeepSpeed-Inference. Based on the model architecture, model size, batch size, and available hardware resources, MII automatically applies the appropriate set of system optimizations to minimize latency and maximize throughput.
MII currently supports over 13,000 models across three popular model architectures. We plan to add additional models in the near term, if there are specific model architectures you would like supported please file an issue and let us know. All current models leverage Hugging Face in our backend to provide both the model weights and the model's corresponding tokenizer. For our current release we support the following model architectures:
model family | size range | ~model count |
---|---|---|
llama | 7B - 65B | 11,000 |
llama-2 | 7B - 70B | 800 |
mistral | 7B | 1,100 |
opt | 0.1B - 66B | 900 |
MII Legacy APIs support over 50,000 different models including BERT, RoBERTa, Stable Diffusion, and other text-generation models like Bloom, GPT-J, etc. For a full list please see our legacy supported models table.
DeepSpeed-MII allows users to create non-persistent and persistent deployments for supported models in just a few lines of code.
The fasest way to get started is with our PyPI release of DeepSpeed-MII which means you can get started within minutes via:
pip install deepspeed-mii
For ease of use and significant reduction in lengthy compile times that many projects require in this space we distribute a pre-compiled python wheel covering the majority of our custom kernels through a new library called DeepSpeed-Kernels. We have found this library to be very portable across environments with NVIDIA GPUs with compute capabilities 8.0+ (Ampere+), CUDA 11.6+, and Ubuntu 20+. In most cases you shouldn't even need to know this library exists as it is a dependency of DeepSpeed-MII and will be installed with it. However, if for whatever reason you need to compile our kernels manually please see our advanced installation docs.
A non-persistent pipeline is a great way to try DeepSpeed-MII. Non-persistent pipelines are only around for the duration of the python script you are running. The full example for running a non-persistent pipeline deployment is only 4 lines. Give it a try!
import mii
pipe = mii.pipeline("mistralai/Mistral-7B-v0.1")
response = pipe("DeepSpeed is", max_new_tokens=128)
print(response)
Taking advantage of multi-GPU systems for greater performance is easy with MII. When run with the deepspeed
launcher, tensor parallelism is automatically controlled by the --num_gpus
flag:
# Run on a single GPU
deepspeed --num_gpus 1 mii-example.py
# Run on multiple GPUs
deepspeed --num_gpus 2 mii-example.py
While only the model name or path is required to stand up a non-persistent pipeline deployment, we offer customization options to our users:
mii.pipeline()
Options:
model_name_or_path: str
Name or local path to a HuggingFace model.max_length: int
Sets the default maximum token length for the prompt + response.all_rank_output: bool
When enabled, all ranks return the generated text. By default, only rank 0 will return text.
Users can also control the generation characteristics for individual prompts (i.e., when calling pipe()
) with the following options:
max_length: int
Sets the per-prompt maximum token length for prompt + response.min_new_tokens: int
Sets the minimum number of tokens generated in the response.max_length
will take precedence over this setting.max_new_tokens: int
Sets the maximum number of tokens generated in the response.ignore_eos: bool
(Defaults toFalse
) Setting toTrue
prevents generation from ending when the EOS token is encountered.top_p: float
(Defaults to0.9
) When set below1.0
, filter tokens and keep only the most probable, where token probabilities sum to ≥top_p
.top_k: int
(Defaults toNone
) WhenNone
, top-k filtering is disabled. When set, the number of highest probability tokens to keep.temperature: float
(Defaults toNone
) WhenNone
, temperature is disabled. When set, modulates token probabilities.do_sample: bool
(Defaults toTrue
) WhenTrue
, sample output logits. WhenFalse
, use greedy sampling.return_full_text: bool
(Defaults toFalse
) WhenTrue
, prepends the input prompt to the returned text
A persistent deployment is ideal for use with long-running and production applications. The persistent model uses a lightweight GRPC server that can be queried by multiple clients at once. The full example for running a persistent model is only 5 lines. Give it a try!
import mii
client = mii.serve("mistralai/Mistral-7B-v0.1")
response = client.generate("Deepspeed is", max_new_tokens=128)
print(response.response)
If we want to generate text from other processes, we can do that too:
client = mii.client("mistralai/Mistral-7B-v0.1")
response = client.generate("Deepspeed is", max_new_tokens=128)
When we no longer need a persistent deployment, we can shutdown the server from any client:
client.terminate_server()
Taking advantage of multi-GPU systems for better latency and throughput is also easy with the persistent deployments. Model parallelism is controlled by the tensor_parallel
input to mii.serve
:
client = mii.serve("mistralai/Mistral-7B-v0.1", tensor_parallel=2)
The resulting deployment will split the model across 2 GPUs to deliver faster inference and higher throughput than a single GPU.
We can also take advantage of multi-GPU (and multi-node) systems by setting up multiple model replicas and taking advantage of the load-balancing that DeepSpeed-MII provides:
client = mii.serve("mistralai/Mistral-7B-v0.1", replica_num=2)
The resulting deployment will load 2 model replicas (one per GPU) and load-balance incoming requests between the 2 model instances.
Model parallelism and replicas can also be combined to take advantage of systems with many more GPUs. In the example below, we run 2 model replicas, each split across 2 GPUs on a system with 4 GPUs:
client = mii.serve("mistralai/Mistral-7B-v0.1", tensor_parallel=2, replica_num=2)
The choice between model parallelism and model replicas for maximum performance will depend on the nature of the hardware, model, and workload. For example, with small models users may find that model replicas provide the lowest average latency for requests. Meanwhile, large models may achieve greater overall throughput when using only model parallelism.
While only the model name or path is required to stand up a persistent deployment, we offer customization options to our users.
mii.serve()
Options:
model_name_or_path: str
Name or local path to a HuggingFace model.max_length: int
Sets the default maximum token length for the prompt + response.deployment_name: str
A unique identifying string for the persistent model. If provided, client objects should be retrieved withclient = mii.client(deployment_name)
.tensor_parallel: int
Number of GPUs to split the model across.replica_num: int
The number of model replicas to stand up.
mii.client()
Options:
model_or_deployment_name: str
Name of the model ordeployment_name
passed tomii.serve()
Users can also control the generation characteristics for individual prompts (i.e., when calling client.generate()
) with the following options:
max_length: int
Sets the per-prompt maximum token length for prompt + response.min_new_tokens: int
Sets the minimum number of tokens generated in the response.max_length
will take precedence over this setting.max_new_tokens: int
Sets the maximum number of tokens generated in the response.ignore_eos: bool
(Defaults toFalse
) Setting toTrue
prevents generation from ending when the EOS token is encountered.top_p: float
(Defaults to0.9
) When set below1.0
, filter tokens and keep only the most probable, where token probabilities sum to ≥top_p
.top_k: int
(Defaults toNone
) WhenNone
, top-k filtering is disabled. When set, the number of highest probability tokens to keep.temperature: float
(Defaults toNone
) WhenNone
, temperature is disabled. When set, modulates token probabilities.do_sample: bool
(Defaults toTrue
) WhenTrue
, sample output logits. WhenFalse
, use greedy sampling.return_full_text: bool
(Defaults toFalse
) WhenTrue
, prepends the input prompt to the returned text
This project welcomes contributions and suggestions. Most contributions require you to agree to a Contributor License Agreement (CLA) declaring that you have the right to, and actually do, grant us the rights to use your contribution. For details, visit https://cla.opensource.microsoft.com.
When you submit a pull request, a CLA bot will automatically determine whether you need to provide a CLA and decorate the PR appropriately (e.g., status check, comment). Simply follow the instructions provided by the bot. You will only need to do this once across all repos using our CLA.
This project has adopted the Microsoft Open Source Code of Conduct. For more information see the Code of Conduct FAQ or contact [email protected] with any additional questions or comments.
This project may contain trademarks or logos for projects, products, or services. Authorized use of Microsoft trademarks or logos is subject to and must follow Microsoft's Trademark & Brand Guidelines. Use of Microsoft trademarks or logos in modified versions of this project must not cause confusion or imply Microsoft sponsorship. Any use of third-party trademarks or logos are subject to those third-party's policies.