squall/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

Fix double free

Browse Source
This commit is contained in:
Woosuk Kwon 2023-02-14 09:14:05 +00:00
parent c80ed212a1
commit c128c2ed03
1 changed files with 2 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
cacheflow/master/block_manager.py
View File

@ -115,13 +115,13 @@ class BlockSpaceManager:
last_block = block_table[-1]
assert last_block.device == Device.GPU
if last_block.ref_count == 1:
# Append.
# Not shared with other sequences. Appendable.
return None
else:
# The last block is shared with other sequences.
# Copy on Write: Allocate a new block and copy the tokens.
new_block = self.gpu_allocator.allocate()
block_table.append(new_block)
block_table[-1] = new_block
self.gpu_allocator.free(last_block)
return last_block.block_number, new_block.block_number