[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
+32 -4
View File
@@ -1,6 +1,7 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import math
from enum import Enum
import numpy as np
@@ -16,6 +17,29 @@ from vllm.v1.utils import CpuGpuBuffer
logger = init_logger(__name__)
def get_block_table_width(
max_num_blocks: int,
block_size: int,
kernel_block_size: int | None = None,
*,
token_alignment: int | None = 128,
) -> int:
"""Return the width after optional alignment and virtual block splitting."""
if kernel_block_size is None:
kernel_block_size = block_size
if block_size % kernel_block_size != 0:
raise ValueError(
f"kernel_block_size {kernel_block_size} must divide "
f"block_size {block_size} evenly"
)
if token_alignment is not None:
if token_alignment <= 0:
raise ValueError("token_alignment must be positive")
block_alignment = token_alignment // math.gcd(token_alignment, block_size)
max_num_blocks = cdiv(max_num_blocks, block_alignment) * block_alignment
return max_num_blocks * block_size // kernel_block_size
class SlotMappingMode(Enum):
TOKEN_TO_KV_SLOT = "token_to_kv_slot"
NONE = "none"
@@ -277,11 +301,15 @@ class MultiGroupBlockTable:
f"must match block_sizes length ({len(block_sizes)})"
)
# Align to a multiple of (128 / block_size) as required
# by some attention backends such as TRTLLM (#39324)
max_num_blocks = [
cdiv(n, 128 // bs) * (128 // bs) if bs <= 128 else n
for n, bs in zip(max_num_blocks, block_sizes)
(
get_block_table_width(n, block_size, token_alignment=None)
if slot_mapping_mode == SlotMappingMode.NONE
else get_block_table_width(n, block_size)
)
for n, block_size, slot_mapping_mode in zip(
max_num_blocks, block_sizes, slot_mapping_modes
)
]
self.block_tables = [