[Bugfix] Universally align block table width to 128 tokens (#50302)

Signed-off-by: Matthew Bonanni <mbonanni@redhat.com>
This commit is contained in:
Matthew Bonanni
2026-07-31 08:52:17 -07:00
committed by GitHub
parent 94e9ef0768
commit a0cd2b69b3
8 changed files with 90 additions and 30 deletions
+12 -1
View File
@@ -36,6 +36,7 @@ from vllm.v1.kv_cache_interface import (
MambaSpec,
UniformTypeKVCacheSpecs,
)
from vllm.v1.worker.block_table import get_block_table_width
logger = init_logger(__name__)
@@ -237,12 +238,22 @@ class AttentionGroup:
if kernel_block_size is not None
else self.kv_cache_spec
)
builder_cls = self.backend.get_builder_cls()
builder_kwargs = {}
if builder_cls.requires_block_table_width:
max_num_blocks = self.kv_cache_spec.max_num_blocks_per_req(
vllm_config, vllm_config.model_config.max_model_len
)
builder_kwargs["block_table_width"] = get_block_table_width(
max_num_blocks, self.kv_cache_spec.block_size, kernel_block_size
)
self.metadata_builders = [
self.backend.get_builder_cls()(
builder_cls(
kv_cache_spec_builder,
self.layer_names,
vllm_config,
device,
**builder_kwargs,
)
for _ in range(num_metadata_builders)
]