Remove unnecessary prints in PagedAttention (pytorch#8374)

Signed-off-by: Woosuk Kwon <[email protected]>

torch_xla/experimental/custom_kernel.py

@@ -490,7 +490,6 @@ def _multi_queries_paged_attention_nonkernel(
     lengths,  # seq_lengths, [batch_size]. nb batch_size = len(seq_lens)
     page_indices,  # [batch_size, pages_per_sequence]
 ) -> torch.Tensor:  # [batch_size, query_len, num_heads, head_dim]
-  print('Running the nonkernel version of multi-queries paged attention.')
   batch_size, query_len, num_query_heads, head_size = q.shape
   num_kv_heads, total_num_pages, page_size, _ = k_pages.shape
   num_query_per_kv = num_query_heads // num_kv_heads
@@ -561,7 +560,6 @@ def multi_queries_paged_attention(
         lengths,
         page_indices,
     )
-  print('Running the kernel version of multi-queries paged attention.')
 
   # Import JAX within the function such that we don't need to call the jax_import_guard()
   # in the global scope which could cause problems for xmp.spawn.
@@ -612,8 +610,6 @@ def paged_attention(q,
                     pages_per_compute_block,
                     megacore_mode: str = None,
                     attn_logits_soft_cap: float = None):
-  print('Running the single-query paged attention.')
-
   # Import JAX within the function such that we don't need to call the jax_import_guard()
   # in the global scope which could cause problems for xmp.spawn.
   jax_import_guard()