Resolve the issue of getting different results from transformers evaluation and vLLM evaluation

[RadoslawPlawecki]

Description

The different results were obtained depending on the evaluation method. To solve this issue, environment variables were added to interference_vllm.py.

os.environ["VLLM_USE_V1"] = "0"
os.environ["CUBLAS_WORKSPACE_CONFIG"]=":4096:8"
os.environ["VLLM_ENABLE_V1_MULTIPROCESSING"] = "0"

After introducing the changes, all the tests were passed successfully.

[DzmitryPihulski]

The main issue here is not using the chat template, but I will add it in future versions

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!