EngineX-Hygon/sglang
5
0
Fork 0
Code Issues Pull Requests Actions 7 Projects Releases Wiki Activity

Fix bench_one_batch producing unnatural results for expert parallel (#5149)

Browse Source
This commit is contained in:
fzyzcjy
2025-04-20 15:38:04 +08:00
committed by GitHub
parent c555d794f7
commit d07e797ace
1 changed files with 1 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
python/sglang/bench_one_batch.py
View File

@@ -207,7 +207,7 @@ def prepare_extend_inputs_for_correctness_test(
def prepare_synthetic_inputs_for_latency_test(batch_size, input_len): def prepare_synthetic_inputs_for_latency_test(batch_size, input_len):
input_ids = np.ones((batch_size, input_len), dtype=np.int32) input_ids = np.random.randint(0, 10000, (batch_size, input_len), dtype=np.int32)
sampling_params = SamplingParams( sampling_params = SamplingParams(
temperature=0, temperature=0,
max_new_tokens=BenchArgs.output_len, max_new_tokens=BenchArgs.output_len,