[Testing] Fix test_sharded_state_loader (#51736)
Signed-off-by: Richard Zou <zou3519@gmail.com>
Richard Zou committed
Aug 11, 2026 at 17:16 UTC
cb30f6f8ddca45fd2e2acc3c43735a7c3b2ecdd3
1 file changed
+4
-4
tests/model_executor/model_loader/test_sharded_state_loader.py
+4
-4
@@ -12,7 +12,6 @@ import torch
12
13
from vllm import LLM, SamplingParams
14
from vllm.model_executor.model_loader import ShardedStateLoader
15
-from vllm.platforms import current_platform
15
from vllm.transformers_utils.repo_utils import hf_api
16
17
prompts = [
@@ -96,9 +95,10 @@ def test_sharded_state_loader(
95
input_dir = llama_3p2_1b_files
96
ctx = mp.get_context("spawn")
97
99
- platform_args = {}
100
- if current_platform.is_rocm() or current_platform.is_xpu():
101
- platform_args["max_num_seqs"] = 1
98
+ # Keep batching deterministic: this test compares exact greedy outputs
99
+ # across separate engine processes, whose schedulers may otherwise form
100
+ # different batches depending on prompt-processing timing.
101
+ platform_args = {"max_num_seqs": 1}
102
103
# Run in separate processes for memory & CUDA isolation
104
with TemporaryDirectory() as output_dir: