[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
+28
View File
@@ -46,6 +46,11 @@ from vllm.v1.kv_cache_interface import (
from vllm.v1.outputs import EMPTY_MODEL_RUNNER_OUTPUT
from vllm.v1.sample.metadata import SamplingMetadata
from vllm.v1.spec_decode.metadata import SpecDecodeMetadata
from vllm.v1.worker.block_table import (
MultiGroupBlockTable,
SlotMappingMode,
get_block_table_width,
)
from vllm.v1.worker.gpu.lora_utils import LoraState
from vllm.v1.worker.gpu.mm.encoder_cache import EncoderCache
from vllm.v1.worker.gpu.mm.lora import set_active_mm_loras
@@ -1380,6 +1385,29 @@ def test_hybrid_block_table_initialization():
)
def test_get_block_table_width_aligns_to_128_tokens():
assert get_block_table_width(1875, 64) == 1876
def test_get_block_table_width_splits_virtual_blocks():
assert get_block_table_width(235, 256, 64) == 940
def test_mamba_state_table_width_is_not_aligned():
block_tables = MultiGroupBlockTable(
max_num_reqs=1,
max_num_batched_tokens=1,
pin_memory=False,
device=torch.device("cpu"),
block_sizes=[39664],
kernel_block_sizes=[39664],
max_num_blocks=[1],
slot_mapping_modes=[SlotMappingMode.NONE],
)
assert block_tables[0].max_num_blocks_per_req == 1
def test_input_batch_with_kernel_block_sizes():
"""Test InputBatch initialization with kernel_block_sizes parameter."""
max_num_reqs = 10