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]>
huggingface · Jan 16, 2025 · 7b4044d · 7b4044d
1 parent 726191f
commit 7b4044d
Showing 1 changed file with 3 additions and 1 deletion.
diff --git a/optimum/exporters/ipex/cache_utils.py b/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