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

[V1] Revert uncache_blocks and support recaching full blocks (#12415)

Browse Source 
Signed-off-by: Cody Yu <hao.yu.cody@gmail.com>
This commit is contained in:
Cody Yu
2025-02-03 15:04:53 -08:00
committed by GitHub
parent cf58b9c4ca
commit 5095e96606
2 changed files with 16 additions and 55 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
tests/v1/core/test_prefix_caching.py
View File

@@ -629,33 +629,3 @@ def test_reset_prefix_cache():
assert manager.reset_prefix_cache()
assert not manager.cached_block_hash_to_block
assert all([blk.block_hash is None for blk in manager.block_pool])
def test_uncache_blocks():
manager = KVCacheManager(
block_size=16,
num_gpu_blocks=10,
max_model_len=8192,
sliding_window=None,
enable_caching=True,
num_preallocate_tokens=0,
)
req0 = make_request("0", list(range(30)))
blocks = manager.allocate_slots(req0, 30)
assert [b.block_id for b in blocks] == [0, 1]
assert len(manager.cached_block_hash_to_block) == 1
req0.num_computed_tokens = 30
# Simulate speculative tokens.
for _ in range(5):
req0.append_output_token_ids(8)
manager.allocate_slots(req0, 5)
assert len(manager.cached_block_hash_to_block) == 2
# After sampling, assuming only 1 token is accepted.
req0.num_computed_tokens = 31
num_uncached_blocks = manager.uncache_blocks(req0)
assert num_uncached_blocks == 1
assert len(manager.cached_block_hash_to_block) == 1