Update .buildkite/lm-eval-harness/test_lm_eval_correctness.py

update gpu default usage

Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>

Author: zhewenl

.buildkite/lm-eval-harness/test_lm_eval_correctness.py

@@ -19,7 +19,7 @@
 def launch_lm_eval(eval_config, tp_size):
     trust_remote_code = eval_config.get("trust_remote_code", False)
     max_model_len = eval_config.get("max_model_len", 4096)
-    gpu_memory_utilization = eval_config.get("gpu_memory_utilization", 1.0)
+    gpu_memory_utilization = eval_config.get("gpu_memory_utilization", 0.9)
     batch_size = eval_config.get("batch_size", "auto")
     model_args = (
         f"pretrained={eval_config['model_name']},"