[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
+6 -5
View File
@@ -56,6 +56,7 @@ from vllm.utils.torch_utils import STR_DTYPE_TO_TORCH_DTYPE
from vllm.v1.core.sched.output import GrammarOutput, SchedulerOutput
from vllm.v1.kv_cache_interface import KVCacheConfig, MambaSpec
from vllm.v1.outputs import DraftTokenIds, ModelRunnerOutput
from vllm.v1.worker.block_table import get_block_table_width
from vllm.v1.worker.cp_utils import check_attention_cp_compatibility
from vllm.v1.worker.gpu import pcp_manager as pcp
from vllm.v1.worker.gpu.async_utils import AsyncOutput, AsyncPoolingOutput
@@ -456,16 +457,16 @@ class GPUModelRunner(LoRAModelRunnerMixin):
max_num_blocks = cdiv(
block_table_max_model_len, spec.block_size * self.dcp_size
)
# Align to a multiple of (128 / block_size) as required by some attention
# backends such as TRTLLM (#39324)
if spec.block_size <= 128:
alignment = 128 // spec.block_size
max_num_blocks = cdiv(max_num_blocks, alignment) * alignment
# For Mamba/Hybrid Model, KVCaches need extra blocks for speculative tokens
if isinstance(spec, MambaSpec):
max_num_blocks = (
max_num_blocks if self.cache_config.enable_prefix_caching else 1
) + spec.num_speculative_blocks
max_num_blocks = get_block_table_width(
max_num_blocks, spec.block_size, token_alignment=None
)
else:
max_num_blocks = get_block_table_width(max_num_blocks, spec.block_size)
max_num_blocks_per_group.append(max_num_blocks)
self.attn_groups, attn_cg_support, self.kernel_block_sizes = init_attn_backend(