diff --git a/tests/model_executor/model_loader/test_sharded_state_loader.py b/tests/model_executor/model_loader/test_sharded_state_loader.py index 48672d4314..06354fe3c5 100644 --- a/tests/model_executor/model_loader/test_sharded_state_loader.py +++ b/tests/model_executor/model_loader/test_sharded_state_loader.py @@ -12,7 +12,6 @@ import torch from vllm import LLM, SamplingParams from vllm.model_executor.model_loader import ShardedStateLoader -from vllm.platforms import current_platform from vllm.transformers_utils.repo_utils import hf_api prompts = [ @@ -96,9 +95,10 @@ def test_sharded_state_loader( input_dir = llama_3p2_1b_files ctx = mp.get_context("spawn") - platform_args = {} - if current_platform.is_rocm() or current_platform.is_xpu(): - platform_args["max_num_seqs"] = 1 + # Keep batching deterministic: this test compares exact greedy outputs + # across separate engine processes, whose schedulers may otherwise form + # different batches depending on prompt-processing timing. + platform_args = {"max_num_seqs": 1} # Run in separate processes for memory & CUDA isolation with TemporaryDirectory() as output_dir: