diff --git a/python/sglang/srt/layers/attention/aiter_backend.py b/python/sglang/srt/layers/attention/aiter_backend.py
index 44c868eb994f..130209e753ea 100755
--- a/python/sglang/srt/layers/attention/aiter_backend.py
+++ b/python/sglang/srt/layers/attention/aiter_backend.py
@@ -186,8 +186,15 @@ def __init__(
                 )
 
         # aiter kernel related initialization
+        # Cap effective sequence length by actual KV cache capacity to avoid
+        # over-allocating the workspace buffer on memory-constrained GPUs.
+        # No single sequence can exceed max_total_num_tokens.
+        effective_max_seq_len = min(
+            self.max_context_len,
+            getattr(model_runner, "max_total_num_tokens", self.max_context_len),
+        )
         self.max_num_partitions = (
-            self.max_context_len + _AITER_PARTITION_SIZE_ROCM - 1
+            effective_max_seq_len + _AITER_PARTITION_SIZE_ROCM - 1
         ) // _AITER_PARTITION_SIZE_ROCM
 
         nbyes_per_qo_elem = torch.finfo(torch.float32).bits // 8