diff --git a/sharktank/sharktank/examples/export_paged_llm_v1.py b/sharktank/sharktank/examples/export_paged_llm_v1.py
index 43242e0c0..900c1a9ae 100644
--- a/sharktank/sharktank/examples/export_paged_llm_v1.py
+++ b/sharktank/sharktank/examples/export_paged_llm_v1.py
@@ -74,7 +74,7 @@ def main():
         tensor_parallelism_size=tensor_parallelism_size,
         use_hf=False,
         static_tables=False,  # Rely on the compiler for hoisting tables.
-        kv_cache_type="paged" if args.bs == [1] else "paged",
+        kv_cache_type="direct" if args.bs == [1] else "paged",
         attention_kernel=args.attention_kernel,
     )
     llama_config.fake_quant = args.fake_quant
diff --git a/sharktank/sharktank/layers/configs/llm_configs.py b/sharktank/sharktank/layers/configs/llm_configs.py
index 6cf79402e..88f5c344c 100644
--- a/sharktank/sharktank/layers/configs/llm_configs.py
+++ b/sharktank/sharktank/layers/configs/llm_configs.py
@@ -144,7 +144,7 @@ class LlamaModelConfig:
 
     # Block sequence stride for a paged KV cache. This must divide evenly
     # into the context length.
-    block_seq_stride: int = 32
+    block_seq_stride: int = 16
 
     # Either "paged" or "direct".
     kv_cache_type: str = "paged"