diff --git a/examples/configs/grpo_math_8B_megatron.yaml b/examples/configs/grpo_math_8B_megatron.yaml
index 41ce4412fc..fc839c8239 100644
--- a/examples/configs/grpo_math_8B_megatron.yaml
+++ b/examples/configs/grpo_math_8B_megatron.yaml
@@ -34,7 +34,8 @@ policy:
     empty_unused_memory_level: 0
     converter_type: "LlamaForCausalLM"
     tensor_model_parallel_size: 1
-    pipeline_model_parallel_size: 1
+    # On H100, can run with pp=1 for better performance with expandable segments (which is enabled by default)
+    pipeline_model_parallel_size: 2
     context_parallel_size: 1
     pipeline_dtype: ${policy.precision}
 
@@ -66,7 +67,7 @@ policy:
     stop_strings: null
     vllm_cfg:
       tensor_parallel_size: 1
-      gpu_memory_utilization: 0.95
+      gpu_memory_utilization: 0.8
       max_model_len: ${policy.max_total_sequence_length}
 
 cluster: