diff --git a/tests/v1/e2e/spec_decode/draft_model/test_lora.py b/tests/v1/e2e/spec_decode/draft_model/test_lora.py index 4ffc07f88c..460ac4e596 100644 --- a/tests/v1/e2e/spec_decode/draft_model/test_lora.py +++ b/tests/v1/e2e/spec_decode/draft_model/test_lora.py @@ -8,6 +8,7 @@ This script contains: import pytest import torch +import vllm.envs as envs from vllm import SamplingParams from vllm.config import CompilationConfig from vllm.lora.request import LoRARequest @@ -64,9 +65,8 @@ def test_batch_inference_correctness( model_setup: (method, model_name, spec_model_name, lora_path, tp_size) """ with monkeypatch.context() as m: - # Disable randomness - if current_platform.is_cuda(): - m.setenv("CUBLAS_WORKSPACE_CONFIG", ":4096:8") + m.setenv("VLLM_BATCH_INVARIANT", "1") + m.setattr(envs, "VLLM_BATCH_INVARIANT", True) set_random_seed(SEED) m.setattr(torch.backends.cudnn, "benchmark", False) m.setattr(torch.backends.cudnn, "deterministic", True)