diff --git a/test/test_tpu_paged_attention_kernel.py b/test/test_tpu_paged_attention_kernel.py
index 746439ba4d0..74a287611cc 100644
--- a/test/test_tpu_paged_attention_kernel.py
+++ b/test/test_tpu_paged_attention_kernel.py
@@ -110,7 +110,7 @@ def setUp(self):
       q_kv_head_ratio=(1, 4, 8),
       head_dim=(128, 256),
       num_queries_per_compute_block=(16, 32),
-      block_kv_size=(128, 256),
+      block_kv_size=(128, 192, 256),
   )
   def test_paged_attention_without_query_padding(
       self,
@@ -206,7 +206,7 @@ def test_paged_attention_without_query_padding(
       q_kv_head_ratio=(1, 4, 8),
       head_dim=(128, 256),
       num_queries_per_compute_block=(16, 32),
-      block_kv_size=(128, 256),
+      block_kv_size=(128, 192, 256),
   )
   def test_paged_attention_with_query_padding(
       self,
diff --git a/torch_xla/experimental/pallas_kernels/multi_queries_paged_attention_kernel.py b/torch_xla/experimental/pallas_kernels/multi_queries_paged_attention_kernel.py
index 557f8ad5ec3..14b432f92e6 100644
--- a/torch_xla/experimental/pallas_kernels/multi_queries_paged_attention_kernel.py
+++ b/torch_xla/experimental/pallas_kernels/multi_queries_paged_attention_kernel.py
@@ -475,10 +475,6 @@ def paged_attention(
     raise ValueError(
         f"{num_kv_pages_per_compute_block=} should be smaller or equal to {pages_per_sequence=}"
     )
-  if pages_per_sequence % num_kv_pages_per_compute_block != 0:
-    raise ValueError(
-        "num_kv_pages_per_compute_block must be divisible by pages per sequence. Got"
-        f" {pages_per_sequence=} and {num_kv_pages_per_compute_block=}.")
   if num_q_heads % num_kv_heads != 0:
     raise ValueError(
         "Number of Q heads must be divisible by number of KV heads. Got"
@@ -491,8 +487,8 @@ def paged_attention(
       num_kv_heads,
       pl.cdiv(query_len, num_queries_per_compute_block
              ),  # how many compute blocks we need to loop the query_len
-      pages_per_sequence //
-      num_kv_pages_per_compute_block,  # how many compute blocks we need to loop the kv_len
+      pl.cdiv(pages_per_sequence, num_kv_pages_per_compute_block
+             ),  # how many compute blocks we need to loop the kv_len
   )
 
   # out_shape