change block_size to 16 for xpu, as `single_query_cached_kv_attention…

…` API does not support 64

Signed-off-by: Liu, Kaixuan <[email protected]>

optimum/exporters/ipex/cache_utils.py

@@ -52,7 +52,7 @@ def __init__(
         # Used in `generate` to keep tally of how many tokens the cache has seen
 
         self._seen_tokens = torch.zeros([max_batch_size], dtype=torch.int32, device=device)
-        default_block_size = 16 if device.type == "cpu" else 64
+        default_block_size = 16
         self.block_size = int(os.environ.get("OI_PAGED_ATTN_BLOCK_SIZE", str(default_block_size)))
         self.num_blocks = (max_cache_len // self.block_size + (max_cache_len % self.block_size != 0)) * max_batch_size
         self.block_tables = -1 * torch.ones([self.num_blocks], dtype=torch.int32, device=device).reshape(