[Bugfix][V1] Free out-of-window blocks on the processed-token basis under async scheduling (#47728)

Signed-off-by: Saddss <28726669061@qq.com>
Signed-off-by: Nick Hill <nickhill123@gmail.com>
Signed-off-by: Yifan Qiao <yifanqiao@inferact.ai>
Co-authored-by: Yifan Qiao <yifanqiao@inferact.ai>
Co-authored-by: Saddss <28726669061@qq.com>
Co-authored-by: Nick Hill <nickhill123@gmail.com>
This commit is contained in:
Saddss
2026-07-08 13:34:19 +01:00
committed by GitHub
co-authored by Yifan Qiao Saddss Nick Hill
parent 934eeaecfb
commit cd0de48d08
14 changed files with 549 additions and 143 deletions
+10 -11
View File
@@ -30,6 +30,7 @@ from vllm.v1.kv_cache_interface import (
FullAttentionSpec,
KVCacheConfig,
KVCacheGroupSpec,
KVCacheSpec,
)
from vllm.v1.request import Request
from vllm.v1.structured_output import StructuredOutputManager
@@ -63,6 +64,7 @@ def create_scheduler(
use_ec_connector: bool = False,
ec_role: str | None = None,
use_v2_model_runner: bool | None = None,
kv_cache_spec: KVCacheSpec | None = None,
) -> Scheduler | AsyncScheduler:
"""Create scheduler under test.
@@ -164,20 +166,17 @@ def create_scheduler(
speculative_config=speculative_config,
ec_transfer_config=ec_transfer_config,
)
if kv_cache_spec is None:
kv_cache_spec = FullAttentionSpec(
block_size=block_size,
num_kv_heads=1,
head_size=1,
dtype=torch.float32,
)
kv_cache_config = KVCacheConfig(
num_blocks=num_blocks, # A large number of blocks to hold all requests
kv_cache_tensors=[],
kv_cache_groups=[
KVCacheGroupSpec(
["layer"],
FullAttentionSpec(
block_size=block_size,
num_kv_heads=1,
head_size=1,
dtype=torch.float32,
),
)
],
kv_cache_groups=[KVCacheGroupSpec(["layer"], kv_cache_spec)],
)
cache_config.num_gpu_blocks = num_blocks
register_all_kvcache_specs(vllm_config)