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

Minor

Browse Source
This commit is contained in:
Woosuk Kwon 2023-02-09 11:30:55 +00:00
parent 0a11a2e5ca
commit 38ed06a846
1 changed files with 3 additions and 3 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
cacheflow/master/block_manager.py
View File

@ -119,10 +119,10 @@ class BlockSpaceManager:
else:
# The last block is shared with other sequences.
# Copy on Write: Allocate a new block and copy the tokens.
block = self.gpu_allocator.allocate()
block_table.append(block)
new_block = self.gpu_allocator.allocate()
block_table.append(new_block)
self.gpu_allocator.free(last_block)
return last_block.block_number, block.block_number
return last_block.block_number, new_block.block_number
def fork(self, src_seq: Sequence, child_seq: Sequence) -> None:
# NOTE: fork does not allocate a new physical block.