vidur

Vidur: LLM Inference System Simulator

Vidur is a high-fidelity and extensible LLM inference system simulator. It can help you with:

1. Study the system performance of models under different workloads and configurations.

   TTFT	TPOT	Request E2E Time	Batch Size

   Llama-3-8B running the AzureLLMInferenceTrace2023_conv trace on single A100 80GB at 6.45 QPS

2. Capacity planning and finding the best deployment configuration for your LLM deployments. Capacity per dollar for different deployment configurations vs TTFT-P90 and TBT-P99 for LLaMA2-70B.

3. Quickly test new research ideas like new scheduling algorithms, optimizations like speculative decoding, etc.

... all without access to GPUs except for a quick initial profiling phase 🎉. We highly recommend checking out our MLSys'24 paper and talk for more details.

Supported Models

Instructions on adding a new model to existing or new SKUs can be found here.

Model / Device	A100 80GB DGX	H100 DGX	4xA100 80GB Pairwise NVLink Node	8xA40 Pairwise NVLink Node
meta-llama/Meta-Llama-3-8B	✅	❌	✅	❌
meta-llama/Meta-Llama-3-70B	✅	❌	✅	❌
meta-llama/Llama-2-7b-hf	✅	✅	✅	✅
codellama/CodeLlama-34b-Instruct-hf"	✅	✅	✅	✅
meta-llama/Llama-2-70b-hf	✅	✅	✅	✅
internlm/internlm-20b	✅	✅	✅	✅
Qwen/Qwen-72B	✅	✅	✅	✅

• All models support a maximum context length of 4k except Llama3-8B and Llama3-70B which support 16k context length by passing additional CLI params:

  --random_forrest_execution_time_predictor_config_prediction_max_prefill_chunk_size 16384 \
  --random_forrest_execution_time_predictor_config_prediction_max_batch_size 512 \
  --random_forrest_execution_time_predictor_config_prediction_max_tokens_per_request 16384
  

• Pipeline parallelism is supported for all models. The PP dimension should divide the number of layers in the model.

• In DGX nodes, there are 8 GPUs, fully connected via NVLink. So TP1, TP2, TP4 and TP8 are supported.

• In 4x pairwise NVLink nodes, there are 4 GPUs, so TP1, TP2 and TP4 are supported. TP4 here is less performant than TP4 in DGX nodes because (GPU1, GPU2) are connected via NVLink and (GPU3, GPU4) are connected via NVLink. but between these layers, the interconnect is slower.

• You can use any combination of TP and PP. For example, you can run LLaMA2-70B on TP2-PP2 on a 4xA100 80GB Pairwise NVLink Node.

Setup

Using mamba

To run the simulator, create a mamba environment with the given dependency file.

mamba env create -p ./env -f ./environment.yml
mamba env update -f environment-dev.yml

Using venv

1. Ensure that you have Python 3.10 installed on your system. Refer https://www.bitecode.dev/p/installing-python-the-bare-minimum
2. cd into the repository root
3. Create a virtual environment using venv module using python3.10 -m venv .venv
4. Activate the virtual environment using source .venv/bin/activate
5. Install the dependencies using python -m pip install -r requirements.txt
6. Run deactivate to deactivate the virtual environment

Using conda (Least recommended)

To run the simulator, create a conda environment with the given dependency file.

conda env create -p ./env -f ./environment.yml
conda env update -f environment-dev.yml

Setting up wandb (Optional)

First, setup your account on https://<your-org>.wandb.io/ or public wandb, obtain the api key and then run the following command,

wandb login --host https://<your-org>.wandb.io

To opt out of wandb, pick any one of the following methods:

1. export WANDB_MODE=disabled in your shell or add this in ~/.zshrc or ~/.bashrc. Remember to reload using source ~/.zshrc.
2. Set wandb_project and wandb_group as "" in vidur/config/default.yml. Also, remove these CLI params from the shell command with which the simulator is invoked.

Running the simulator

To run the simulator, execute the following command from the repository root,

python -m vidur.main

or a big example with all the parameters,

python -m vidur.main  \
--replica_config_device a100 \
--replica_config_model_name meta-llama/Meta-Llama-3-8B \
--cluster_config_num_replicas 1 \
--replica_config_tensor_parallel_size 1 \
--replica_config_num_pipeline_stages 1 \
--request_generator_config_type synthetic \
--synthetic_request_generator_config_num_requests 512  \
--length_generator_config_type trace \
--trace_request_length_generator_config_max_tokens 16384 \
--trace_request_length_generator_config_trace_file ./data/processed_traces/splitwise_conv.csv \
--interval_generator_config_type poisson \
--poisson_request_interval_generator_config_qps 6.45 \
--replica_scheduler_config_type sarathi  \
--sarathi_scheduler_config_batch_size_cap 512  \
--sarathi_scheduler_config_chunk_size 512 \
--random_forrest_execution_time_predictor_config_prediction_max_prefill_chunk_size 16384 \
--random_forrest_execution_time_predictor_config_prediction_max_batch_size 512 \
--random_forrest_execution_time_predictor_config_prediction_max_tokens_per_request 16384

or to get information on all parameters,

python -m vidur.main -h

Simulator Output

• The metrics will be logged to wandb directly and a copy will be stored in the simulator_output/<TIMESTAMP> directory. A description of all the logged metrics can be found here.

• Vidur exports chrome traces of each simulation. The trace can be found in the simulator_output directory. The trace can be opened by navigating to chrome://tracing/ or edge://tracing/ and loading the trace.

  

Formatting Code

To format code, execute the following command:

make format

Using Canary Build

We have been working on several improvements for the simulator, including support for prefix caching, different routing policies, reducing memory requirements for the simulator, etc. However, there are some sharp edges that we are working on resolving. In the meantime, if you are looking for support for any of these features, please use the canary branch.

Contributing

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.

Trademarks

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.