hailin
/
inference
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity
inference/sglang/benchmark/mmmu
History
hailin 466c38a2d4 first commit 2025-05-27 11:42:20 +08:00
..
README.md first commit 2025-05-27 11:42:20 +08:00
bench_hf.py first commit 2025-05-27 11:42:20 +08:00
bench_sglang.py first commit 2025-05-27 11:42:20 +08:00
data_utils.py first commit 2025-05-27 11:42:20 +08:00
eval_utils.py first commit 2025-05-27 11:42:20 +08:00
prompt_format.yaml first commit 2025-05-27 11:42:20 +08:00

README.md

Run evaluation

Evaluate sglang

Host the VLM:

python -m sglang.launch_server --model-path Qwen/Qwen2-VL-7B-Instruct --chat-template qwen2-vl --port 30000

Benchmark:

python benchmark/mmmu/bench_sglang.py --port 30000

It's recommended to reduce the memory usage by appending something ike --mem-fraction-static 0.6 to the command above.

Evaluate hf

python benchmark/mmmu/bench_hf.py --model-path Qwen/Qwen2-VL-7B-Instruct

Some popular model results:

  1. Qwen/Qwen2-VL-2B-Instruct: 0.241
  2. Qwen/Qwen2-VL-7B-Instruct: 0.255
  3. Qwen/Qwen2.5-VL-3B-Instruct: 0.245
  4. Qwen/Qwen2.5-VL-7B-Instruct: 0.242