set paged attn block size as a env parameter (#1109)

* set default block size

Signed-off-by: jiqing-feng <[email protected]>

* decoding use single query

Signed-off-by: jiqing-feng <[email protected]>

* set block size as a env parameter

Signed-off-by: jiqing-feng <[email protected]>

* set different default value for block size based on device

Signed-off-by: jiqing-feng <[email protected]>

---------

Signed-off-by: jiqing-feng <[email protected]>

Author: jiqing-feng

optimum/exporters/ipex/cache_utils.py

@@ -1,3 +1,4 @@
+import os
 from typing import List, Optional, Tuple
 
 import torch
@@ -44,7 +45,8 @@ def __init__(
         self.batch_size = batch_size
         # Used in `generate` to keep tally of how many tokens the cache has seen
         self._seen_tokens = torch.zeros([batch_size], dtype=torch.int32, device=device)
-        self.block_size = 64
+        default_block_size = 16 if device.type == "cpu" else 64
+        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)) * batch_size
         self.block_tables = -1 * torch.ones([self.num_blocks], dtype=torch.int32, device=device).reshape(
             batch_size, -1