[Bugfix][KV Cache][MLA] Align packed block strides for V3.2 sparse MLA (#55528)

Signed-off-by: lz <145014769+200lz@users.noreply.github.com>
Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
Co-authored-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: Robert Shaw <114415538+robertgshaw2-redhat@users.noreply.github.com>
Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
This commit is contained in:
lz
2026-09-25 05:33:36 +00:00
committed by GitHub
co-authored by OpenAI Codex Lucas Wilkinson Robert Shaw mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
parent 3a58aeb1a1
commit 29468dde8b
6 changed files with 187 additions and 27 deletions
@@ -10,6 +10,7 @@ block); the allocation is the same either way.
"""
from dataclasses import replace
from math import lcm
from unittest.mock import MagicMock
import pytest
@@ -26,9 +27,11 @@ from vllm.v1.core.kv_cache_utils import (
get_kv_cache_groups,
resolve_kv_cache_block_sizes,
)
from vllm.v1.hisparse.layout import _build_hisparse_kv_cache_tensors
from vllm.v1.kv_cache_interface import (
CircularBufferSpec,
FullAttentionSpec,
HiSparseHotSpec,
KVCacheGroupSpec,
KVCacheLayout,
KVCacheSpec,
@@ -79,6 +82,30 @@ def _mock_vllm_config(layout: str | None):
return config
def test_packed_alignment_preserves_hot_pages_and_generic_stride():
"""One shared stride must satisfy both constraints, regardless of rounding order."""
spec = replace(_mla(128), block_stride_alignment=1152)
hot = HiSparseHotSpec(block_size=16, page_size=5120, blocks_per_request=2)
groups = [KVCacheGroupSpec(["attention"], spec), KVCacheGroupSpec(["hot"], hot)]
alignment = lcm(1152, hot.page_size_bytes)
stride = _pool_bytes_per_block(groups)
assert stride == alignment
layout = KVCacheLayout.BLHNC
tensors = _build_hisparse_kv_cache_tensors(groups, 3, 3 * stride, layout, stride)
assert {tensor.block_stride for tensor in tensors} == {alignment}
assert {tensor.size for tensor in tensors} == {3 * alignment}
@pytest.mark.parametrize("spec", [_full(), _mla(128)])
def test_block_alignment_survives_merge_and_rejects_incompatible_specs(spec):
aligned = replace(spec, block_stride_alignment=512)
assert type(spec).merge([aligned, aligned]).block_stride_alignment == 512
with pytest.raises(AssertionError):
type(spec).merge([aligned, spec])
with pytest.raises(ValueError, match="must be positive"):
replace(spec, block_stride_alignment=0)
def _pages(groups) -> dict[str, int]:
return {
name: group.kv_cache_spec.kv_cache_specs[name].page_size_bytes