[misc] benchmark_throughput : Add LoRA

[varun-sundar-rabindranath]

Update benchmark_throughput.py to support LoRA benchmarking.

Examples:
Machine : 1xA100

python3 benchmarks/benchmark_throughput.py --model meta-llama/Llama-2-7b-hf --backend vllm --dataset ./ShareGPT_V3_unfiltered_cleaned_split.json --num-prompts 1000

Throughput: 11.45 requests/s, 5509.93 total tokens/s, 2693.62 output tokens/s

python3 benchmarks/benchmark_throughput.py --model meta-llama/Llama-2-7b-hf --backend vllm --dataset ./ShareGPT_V3_unfiltered_cleaned_split.json --num-prompts ${num_prompts} --max-loras 1 --max-lora-rank 8 --enable-lora --lora-path "yard1/llama-2-7b-sql-lora-test"

Throughput: 7.60 requests/s, 3656.98 total tokens/s, 1787.78 output tokens/s

python3 benchmarks/benchmark_throughput.py --model meta-llama/Llama-2-7b-hf --backend vllm --dataset ./ShareGPT_V3_unfiltered_cleaned_split.json --num-prompts ${num_prompts} --max-loras 4 --max-lora-rank 8 --enable-lora --lora-path "yard1/llama-2-7b-sql-lora-test"

Throughput: 7.61 requests/s, 3664.67 total tokens/s, 1791.53 output tokens/s

python3 benchmarks/benchmark_throughput.py --model meta-llama/Llama-2-7b-hf --backend vllm --dataset ./ShareGPT_V3_unfiltered_cleaned_split.json --num-prompts ${num_prompts} --async-engine

Throughput: 11.31 requests/s, 5441.21 total tokens/s, 2660.02 output tokens/s

python3 benchmarks/benchmark_throughput.py --model meta-llama/Llama-2-7b-hf --backend vllm --dataset ./ShareGPT_V3_unfiltered_cleaned_split.json --num-prompts ${num_prompts} --async-engine --max-loras 1 --max-lora-rank 8 --enable-lora --lora-path "yard1/llama-2-7b-sql-lora-test"

Throughput: 7.59 requests/s, 3654.39 total tokens/s, 1786.51 output tokens/s

python3 benchmarks/benchmark_throughput.py --model meta-llama/Llama-2-7b-hf --backend vllm --dataset ./ShareGPT_V3_unfiltered_cleaned_split.json --num-prompts ${num_prompts} --async-engine --max-loras 4 --max-lora-rank 8 --enable-lora --lora-path "yard1/llama-2-7b-sql-lora-test"

Throughput: 7.55 requests/s, 3634.68 total tokens/s, 1776.87 output tokens/s

cc @mgoin @robertgshaw2-neuralmagic @jeejeelee

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.
Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can do one of these:

• Add ready label to the PR
• Enable auto-merge.

🚀

[jeejeelee]

Not recommended to use default , as LoRA is strongly coupled bound to the model. We also need to add help to describe this arg

[varun-sundar-rabindranath]

You are correct. I have removed the default 👍 Thanks !

[jeejeelee]

There are too many LoRA weights that don't include tokenizer files, which leads to request_tokenizer being None. We need to consider this

[varun-sundar-rabindranath]

Nice catch ! Added a fallback to use the base tokenizer.

[jeejeelee]

overall lgtm, thanks