biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Core] Optimize block_manager_v2 vs block_manager_v1 (to make V2 default) (#5602)

Browse Source
This commit is contained in:
Alexander Matveev
2024-07-01 23:10:37 -04:00
committed by GitHub
parent 54600709b6
commit 3476ed0809
19 changed files with 1182 additions and 525 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
tests/core/block/test_block_table.py
View File

@@ -373,8 +373,9 @@ def test_cow(block_size: int, sequence_len: int, append_len: int,
block_size) - (sequence_len // block_size)
original_block_table.allocate(token_ids=token_ids, device=Device.GPU)
original_block_ids = original_block_table.physical_block_ids
original_block_ids = original_block_table.physical_block_ids[:]
print("original_block_ids = {}".format(original_block_ids))
forked_block_table = original_block_table.fork()
# Expect no additional allocation (copy on _write_).
@@ -457,7 +458,7 @@ def test_cow_lookahead_simple(block_size: int, sequence_len: int,
# Allocate lookahead slots.
original_block_table.ensure_num_empty_slots(lookahead_slots)
original_block_ids = original_block_table.physical_block_ids
original_block_ids = original_block_table.physical_block_ids[:]
forked_block_table = original_block_table.fork()