ipex-llm/docs/mddocs/Quickstart/benchmark_quickstart.md

Run Performance Benchmarking with IPEX-LLM

We can perform benchmarking for IPEX-LLM on Intel CPUs and GPUs using the benchmark scripts we provide.

Prepare The Environment

You can refer to here to install IPEX-LLM in your environment. The following dependencies are also needed to run the benchmark scripts.

pip install pandas
pip install omegaconf

Prepare The Scripts

Navigate to your local workspace and then download IPEX-LLM from GitHub. Modify the config.yaml under all-in-one folder for your benchmark configurations.

cd your/local/workspace
git clone https://github.com/intel-analytics/ipex-llm.git
cd ipex-llm/python/llm/dev/benchmark/all-in-one/

config.yaml

repo_id:
  - 'meta-llama/Llama-2-7b-chat-hf'
local_model_hub: 'path to your local model hub'
warm_up: 1 # must set >=2 when run "pipeline_parallel_gpu" test_api
num_trials: 3
num_beams: 1 # default to greedy search
low_bit: 'sym_int4' # default to use 'sym_int4' (i.e. symmetric int4)
batch_size: 1 # default to 1
in_out_pairs:
  - '32-32'
  - '1024-128'
  - '2048-256'
test_api:
  - "transformer_int4_gpu"   # on Intel GPU, transformer-like API, (qtype=int4)
cpu_embedding: False # whether put embedding to CPU
streaming: False # whether output in streaming way (only avaiable now for gpu win related test_api)
task: 'continuation' # task can be 'continuation', 'QA' and 'summarize'

Some parameters in the yaml file that you can configure:

• repo_id: The name of the model and its organization.
• local_model_hub: The folder path where the models are stored on your machine. Replace 'path to your local model hub' with /llm/models.
• warm_up: The number of warmup trials before performance benchmarking (must set to >= 2 when using "pipeline_parallel_gpu" test_api).
• num_trials: The number of runs for performance benchmarking (the final result is the average of all trials).
• low_bit: The low_bit precision you want to convert to for benchmarking.
• batch_size: The number of samples on which the models make predictions in one forward pass.
• in_out_pairs: Input sequence length and output sequence length combined by '-'.
• test_api: Different test functions for different machines.
  • transformer_int4_gpu on Intel GPU for Linux
  • transformer_int4_gpu_win on Intel GPU for Windows
  • transformer_int4 on Intel CPU
• cpu_embedding: Whether to put embedding on CPU (only available for windows GPU-related test_api).
• streaming: Whether to output in a streaming way (only available for GPU Windows-related test_api).
• use_fp16_torch_dtype: Whether to use fp16 for the non-linear layer (only available for "pipeline_parallel_gpu" test_api).
• n_gpu: Number of GPUs to use (only available for "pipeline_parallel_gpu" test_api).
• task: There are three tasks: continuation, QA and summarize. continuation refers to writing additional content based on prompt. QA refers to answering questions based on prompt. summarize refers to summarizing the prompt.

Note

If you want to benchmark the performance without warmup, you can set warm_up: 0 and num_trials: 1 in config.yaml, and run each single model and in_out_pair separately.

Run on Windows

Please refer to here to configure oneAPI environment variables. Choose corresponding commands base on your device.

• For Intel iGPU:

  set SYCL_CACHE_PERSISTENT=1
  set BIGDL_LLM_XMX_DISABLED=1
  
  python run.py
  

• For Intel Arc™ A300-Series or Pro A60:

  set SYCL_CACHE_PERSISTENT=1
  python run.py
  

• For Other Intel dGPU Series:

  # e.g. Arc™ A770
  python run.py
  

Run on Linux

Please choose corresponding commands base on your device.

• For Intel Arc™ A-Series and Intel Data Center GPU Flex:

  For Intel Arc™ A-Series Graphics and Intel Data Center GPU Flex Series, we recommend:

  ./run-arc.sh
  

• For Intel iGPU:

  For Intel iGPU, we recommend:

  ./run-igpu.sh
  

• For Intel Data Center GPU Max:

  Please note that you need to run conda install -c conda-forge -y gperftools=2.10 before running the benchmark script on Intel Data Center GPU Max Series.

  ./run-max-gpu.sh
  

• For Intel SPR:

  For Intel SPR machine, we recommend:

  ./run-spr.sh
  

  The scipt uses a default numactl strategy. If you want to customize it, please use lscpu or numactl -H to check how cpu indexs are assigned to numa node, and make sure the run command is binded to only one socket.

• For Intel HBM:

  For Intel HBM machine, we recommend:

  ./run-hbm.sh
  

  The scipt uses a default numactl strategy. If you want to customize it, please use numactl -H to check how the index of hbm node and cpu are assigned.

  For example:

  node   0   1   2   3
     0:  10  21  13  23
     1:  21  10  23  13
     2:  13  23  10  23
     3:  23  13  23  10
  

  here hbm node is the node whose distance from the checked node is 13, node 2 is node 0's hbm node.

  And make sure the run command is binded to only one socket.

Result

After the benchmarking is completed, you can obtain a CSV result file under the current folder. You can mainly look at the results of columns 1st token avg latency (ms) and 2+ avg latency (ms/token) for the benchmark results. You can also check whether the column actual input/output tokens is consistent with the column input/output tokens and whether the parameters you specified in config.yaml have been successfully applied in the benchmarking.