diff --git a/training/configs/external_generalisation/math_curriculum_qwen_7b.yaml b/training/configs/external_generalisation/math_curriculum_qwen_7b.yaml
index b8291b3f..f313bccc 100644
--- a/training/configs/external_generalisation/math_curriculum_qwen_7b.yaml
+++ b/training/configs/external_generalisation/math_curriculum_qwen_7b.yaml
@@ -134,7 +134,7 @@ actor_rollout_ref:
     enforce_eager: True
     free_cache_engine: True
     load_format: dummy_dtensor
-    tensor_model_parallel_size: 4
+    tensor_model_parallel_size: 2
     max_num_batched_tokens: 12288
     max_num_seqs: 1024
     log_prob_micro_batch_size: null # will be deprecated, use log_prob_micro_batch_size_per_gpu