[KV Offload] Support packed HMA KV cache layout (#46205)

Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
Co-authored-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
This commit is contained in:
Lucas Wilkinson
2026-06-20 21:19:40 +00:00
committed by GitHub
co-authored by OpenAI Codex Tyler Michael Smith
parent 77148992cf
commit cc22621b51
6 changed files with 175 additions and 21 deletions
+95 -2
View File
@@ -1,16 +1,23 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Tests for contiguous KV cache packing in _get_kv_cache_config_deepseek_v4."""
"""Tests for contiguous KV cache packing."""
from unittest.mock import MagicMock
import pytest
import torch
from vllm.v1.core.kv_cache_utils import _get_kv_cache_config_deepseek_v4
from vllm import envs
from vllm.v1.core.kv_cache_utils import (
_get_kv_cache_config_deepseek_v4,
get_kv_cache_config_from_groups,
)
from vllm.v1.kv_cache_interface import (
FullAttentionSpec,
KVCacheGroupSpec,
KVCacheTensor,
MLAAttentionSpec,
SlidingWindowSpec,
UniformTypeKVCacheSpecs,
)
@@ -28,6 +35,25 @@ def _make_mla_spec(page_size: int, block_size: int = 256) -> MLAAttentionSpec:
)
def _make_full_spec() -> FullAttentionSpec:
return FullAttentionSpec(
block_size=16,
num_kv_heads=2,
head_size=64,
dtype=torch.float16,
)
def _make_sw_spec() -> SlidingWindowSpec:
return SlidingWindowSpec(
block_size=16,
num_kv_heads=2,
head_size=64,
dtype=torch.float16,
sliding_window=128,
)
def _make_groups(n_c4, n_c128, n_swa):
PS_C4_MLA = 37440
PS_C4_IDX = 8640
@@ -130,6 +156,73 @@ class TestInterleavedPacking:
for i, v in enumerate(views):
assert (v == i + 1).all(), f"View {i} was corrupted"
def test_hma_attention_groups_keep_default_backing(self, monkeypatch):
monkeypatch.setattr(envs, "VLLM_USE_PACKED_HMA_KV_CACHE", False, raising=False)
full = _make_full_spec()
sw = _make_sw_spec()
page_size = full.page_size_bytes
groups = [
KVCacheGroupSpec(["full.0", "full.1"], full),
KVCacheGroupSpec(["sw.0", "sw.2"], sw),
KVCacheGroupSpec(["sw.1", "sw.3"], sw),
]
config = get_kv_cache_config_from_groups(
_mock_vllm_config(), groups, available_memory=page_size * 2 * 32
)
assert config.num_blocks == 32
assert sum(t.size for t in config.kv_cache_tensors) == page_size * 2 * 32
assert config.kv_cache_tensors == [
KVCacheTensor(size=page_size * 32, shared_by=["full.0", "sw.0", "sw.1"]),
KVCacheTensor(size=page_size * 32, shared_by=["full.1", "sw.2", "sw.3"]),
]
def test_hma_attention_groups_use_packed_backing_with_flag(self, monkeypatch):
monkeypatch.setattr(envs, "VLLM_USE_PACKED_HMA_KV_CACHE", True, raising=False)
full = _make_full_spec()
sw = _make_sw_spec()
page_size = full.page_size_bytes
groups = [
KVCacheGroupSpec(["full.0", "full.1"], full),
KVCacheGroupSpec(["sw.0", "sw.2"], sw),
KVCacheGroupSpec(["sw.1", "sw.3"], sw),
]
config = get_kv_cache_config_from_groups(
_mock_vllm_config(), groups, available_memory=page_size * 2 * 32
)
assert config.num_blocks == 32
assert {t.size for t in config.kv_cache_tensors} == {page_size * 2 * 32}
assert config.kv_cache_tensors == [
KVCacheTensor(
size=page_size * 2 * 32,
shared_by=["full.0", "sw.0", "sw.1"],
offset=0,
block_stride=page_size * 2,
),
KVCacheTensor(
size=page_size * 2 * 32,
shared_by=["full.1", "sw.2", "sw.3"],
offset=page_size,
block_stride=page_size * 2,
),
]
def test_single_group_attention_keeps_unpacked_layout(self):
spec = _make_full_spec()
groups = [KVCacheGroupSpec(["full.0", "full.1"], spec)]
config = get_kv_cache_config_from_groups(
_mock_vllm_config(), groups, available_memory=spec.page_size_bytes * 2 * 32
)
assert sum(t.size for t in config.kv_cache_tensors) == (
spec.page_size_bytes * 2 * 32
)
assert [t.block_stride for t in config.kv_cache_tensors] == [0, 0]
if __name__ == "__main__":
pytest.main([__file__, "-v"])
@@ -50,7 +50,8 @@ class OffloadingConnectorWorker:
def register_kv_caches(
self, kv_caches: dict[str, torch.Tensor | list[torch.Tensor]]
):
num_blocks = self.spec.kv_cache_config.num_blocks
kv_cache_config = self.spec.kv_cache_config
num_blocks = kv_cache_config.num_blocks
# layer_name -> (num_blocks, page_size_bytes) tensor
tensors_per_block: dict[str, tuple[torch.Tensor, ...]] = {}
@@ -58,7 +59,7 @@ class OffloadingConnectorWorker:
unpadded_page_size_bytes: dict[str, int] = {}
# layer_name -> size of page in bytes
page_size_bytes: dict[str, int] = {}
for kv_cache_group in self.spec.kv_cache_config.kv_cache_groups:
for kv_cache_group in kv_cache_config.kv_cache_groups:
group_layer_names = kv_cache_group.layer_names
group_kv_cache_spec = kv_cache_group.kv_cache_spec
if isinstance(group_kv_cache_spec, UniformTypeKVCacheSpecs):
@@ -122,9 +123,35 @@ class OffloadingConnectorWorker:
else:
raise NotImplementedError
packed_kv_cache_tensor = next(
(t for t in kv_cache_config.kv_cache_tensors if t.block_stride), None
)
is_dsv4 = all(
isinstance(group.kv_cache_spec, UniformTypeKVCacheSpecs)
for group in kv_cache_config.kv_cache_groups
)
if packed_kv_cache_tensor is not None and not is_dsv4:
(tensor,) = tensors_per_block[packed_kv_cache_tensor.shared_by[0]]
block_stride = tensor.stride(0)
packed_tensor = tensor.as_strided(
(num_blocks, block_stride),
(block_stride, 1),
storage_offset=0,
)
self._register_handlers(
CanonicalKVCaches(
[CanonicalKVCacheTensor(packed_tensor, block_stride)],
[
[CanonicalKVCacheRef(0, block_stride)]
for _ in kv_cache_config.kv_cache_groups
],
)
)
return
block_tensors: list[CanonicalKVCacheTensor] = []
block_data_refs: dict[str, list[CanonicalKVCacheRef]] = defaultdict(list)
for kv_cache_tensor in self.spec.kv_cache_config.kv_cache_tensors:
for kv_cache_tensor in kv_cache_config.kv_cache_tensors:
# Filter to layers that were actually processed above.
# _get_kv_cache_config_deepseek_v4 emits KVCacheTensor entries for
# every (tuple_idx, page_size) slot; slots where no group has a
@@ -166,7 +193,7 @@ class OffloadingConnectorWorker:
)
group_data_refs: list[list[CanonicalKVCacheRef]] = []
for kv_cache_group in self.spec.kv_cache_config.kv_cache_groups:
for kv_cache_group in kv_cache_config.kv_cache_groups:
group_refs: list[CanonicalKVCacheRef] = []
for layer_name in kv_cache_group.layer_names:
group_refs += block_data_refs[layer_name]
+6
View File
@@ -209,6 +209,7 @@ if TYPE_CHECKING:
VLLM_EXECUTE_MODEL_TIMEOUT_SECONDS: int = 300
VLLM_WORKER_SHUTDOWN_TIMEOUT_SECONDS: int = 5
VLLM_KV_CACHE_LAYOUT: Literal["NHD", "HND"] | None = None
VLLM_USE_PACKED_HMA_KV_CACHE: bool = False
VLLM_SSM_CONV_STATE_LAYOUT: Literal["SD", "DS"] | None = None
VLLM_COMPUTE_NANS_IN_LOGITS: bool = False
VLLM_ROCM_QUICK_REDUCE_QUANTIZATION: Literal[
@@ -1608,6 +1609,11 @@ environment_variables: dict[str, Callable[[], Any]] = {
"VLLM_KV_CACHE_LAYOUT": env_with_choices(
"VLLM_KV_CACHE_LAYOUT", None, ["NHD", "HND"]
),
# Opt into packed per-block KV cache allocation for multi-group
# attention-only HMA models (e.g. gpt-oss, Gemma 3/4).
"VLLM_USE_PACKED_HMA_KV_CACHE": lambda: bool(
int(os.getenv("VLLM_USE_PACKED_HMA_KV_CACHE", "0"))
),
# SSM conv state layout used for Mamba models.
# - SD: (state_len, dim) — dim contiguous (default)
# - DS: (dim, state_len) — TP-sharded dim on dim1,
+25 -13
View File
@@ -938,9 +938,7 @@ def _pool_bytes_per_block(kv_cache_groups: list[KVCacheGroupSpec]) -> int:
kv_cache_groups[0].kv_cache_spec, UniformTypeKVCacheSpecs
):
return kv_cache_groups[0].kv_cache_spec.page_size_bytes
if all(
isinstance(g.kv_cache_spec, UniformTypeKVCacheSpecs) for g in kv_cache_groups
):
if _use_packed_kv_cache_groups(kv_cache_groups):
# buckets = {page_size: [[layer_names], [layer_names], ...]}
buckets = _bucket_layers_by_page_size(kv_cache_groups)
return sum(ps * len(slots) for ps, slots in buckets.items())
@@ -1218,16 +1216,29 @@ def _bucket_layers_by_page_size(
return buckets
def _get_kv_cache_config_deepseek_v4(
def _use_packed_kv_cache_groups(
kv_cache_groups: list[KVCacheGroupSpec],
) -> bool:
is_dsv4 = all(
isinstance(group.kv_cache_spec, UniformTypeKVCacheSpecs)
for group in kv_cache_groups
)
return is_dsv4 or (
bool(envs.VLLM_USE_PACKED_HMA_KV_CACHE) and len(kv_cache_groups) > 1
)
def _get_kv_cache_config_packed(
vllm_config: VllmConfig,
kv_cache_groups: list[KVCacheGroupSpec],
available_memory: int,
) -> tuple[int, list[KVCacheTensor]]:
"""DeepseekV4 KV cache tensor layout planning.
"""Plan a packed per-block KV cache tensor layout.
Emit one KVCacheTensor per (slot_idx, page_size). Layers from different
groups at the same slot share a tensor (they have independent block
tables so block-id namespaces never collide).
tables so block-id namespaces never collide). Each emitted tensor aliases
one physical backing allocation, with per-block data laid out contiguously.
"""
# buckets = {page_size: [[layer_names], [layer_names], ...]}
buckets = _bucket_layers_by_page_size(kv_cache_groups)
@@ -1255,6 +1266,9 @@ def _get_kv_cache_config_deepseek_v4(
return num_blocks, kv_cache_tensors
_get_kv_cache_config_deepseek_v4 = _get_kv_cache_config_packed
def get_kv_cache_config_from_groups(
vllm_config: VllmConfig,
kv_cache_groups: list[KVCacheGroupSpec],
@@ -1299,13 +1313,11 @@ def get_kv_cache_config_from_groups(
)
for layer_name in kv_cache_groups[0].layer_names
]
elif all(
isinstance(group.kv_cache_spec, UniformTypeKVCacheSpecs)
for group in kv_cache_groups
):
# DeepseekV4: UniformTypeKVCacheSpecs but multiple groups.
# Delegate to the DeepseekV4-specific allocator.
num_blocks, kv_cache_tensors = _get_kv_cache_config_deepseek_v4(
elif _use_packed_kv_cache_groups(kv_cache_groups):
# DeepSeek V4 keeps the existing packed layout. Other multi-group
# attention-only HMA layouts can opt in with
# VLLM_USE_PACKED_HMA_KV_CACHE=1.
num_blocks, kv_cache_tensors = _get_kv_cache_config_packed(
vllm_config, kv_cache_groups, available_memory
)
else:
+9 -1
View File
@@ -58,7 +58,15 @@ class CPUOffloadingSpec(OffloadingSpec):
self.cpu_page_size_per_worker = 0
assert kv_cache_config is not None
if kv_cache_config.num_blocks > 0 and world_size > 0:
total_gpu_kv_bytes = sum(t.size for t in kv_cache_config.kv_cache_tensors)
is_packed = any(t.block_stride for t in kv_cache_config.kv_cache_tensors)
assert not is_packed or all(
t.block_stride for t in kv_cache_config.kv_cache_tensors
)
total_gpu_kv_bytes = (
kv_cache_config.kv_cache_tensors[0].size
if is_packed
else sum(t.size for t in kv_cache_config.kv_cache_tensors)
)
kv_bytes_per_block = (
total_gpu_kv_bytes // kv_cache_config.num_blocks
) * world_size
+9 -1
View File
@@ -187,7 +187,13 @@ class SimpleCPUOffloadScheduler:
assert len(gpu_config.kv_cache_tensors) > 0
gpu_total_bytes = sum(t.size for t in gpu_config.kv_cache_tensors)
is_packed = any(t.block_stride for t in gpu_config.kv_cache_tensors)
assert not is_packed or all(t.block_stride for t in gpu_config.kv_cache_tensors)
gpu_total_bytes = (
gpu_config.kv_cache_tensors[0].size
if is_packed
else sum(t.size for t in gpu_config.kv_cache_tensors)
)
num_gpu_blocks = gpu_config.num_blocks
num_cpu_blocks = max(1, num_gpu_blocks * cpu_capacity_bytes // gpu_total_bytes)
# Create CPU kv_cache_tensors mirroring GPU by scaling size proportionally.
@@ -195,6 +201,8 @@ class SimpleCPUOffloadScheduler:
KVCacheTensor(
size=t.size // num_gpu_blocks * num_cpu_blocks,
shared_by=list(t.shared_by),
offset=t.offset,
block_stride=t.block_stride,
)
for t in gpu_config.kv_cache_tensors
]