[Bugfix][KV Connector][Mooncake] Keep TP-sharded Mamba state out of the KV-head dedup (#49499)

Signed-off-by: Yifan Qiao <yifanqiao@inferact.ai>
Co-authored-by: Claude <noreply@anthropic.com>
This commit is contained in:
Yifan Qiao
2026-07-24 21:02:22 -07:00
committed by GitHub
co-authored by Claude
parent 318b527cc2
commit 70052fb924
3 changed files with 311 additions and 93 deletions
@@ -173,7 +173,6 @@ def test_e2e_swa_plus_full_save_then_lookup_hits():
worker = _build_worker_with_dict_store(vllm_config, cfg, store)
worker.tp_size = 1
worker.pp_size = 1
worker.put_step = 1
worker.num_kv_head = 8
# Register kv_caches using mocked thread classes so register_kv_caches
@@ -215,7 +214,7 @@ def test_e2e_swa_plus_full_save_then_lookup_hits():
block_size=worker.block_size,
coord=worker.coord,
tp_rank=worker.tp_rank,
put_step=worker.put_step,
group_put_steps=worker._group_tp_replication_factors,
kv_role=worker.kv_role,
ready_event=ready,
enable_kv_event=False,
@@ -90,7 +90,7 @@ def _make_store_sending_thread(
block_size=block_size,
coord=coord,
tp_rank=tp_rank,
put_step=put_step,
group_put_steps=[put_step] * len(token_databases),
kv_role="kv_producer",
ready_event=threading.Event(),
replicate_config=replicate_config,
@@ -527,6 +527,29 @@ def test_store_sending_thread_delta_strides_with_local_phase():
assert store.batch_put_from_multi_buffers.call_args.args[0] == keys
def test_tp_sharded_group_saves_every_block_on_every_rank():
"""Sharded ranks must write every block because peers hold different bytes."""
store = MagicMock()
store.batch_is_exist.side_effect = lambda keys: [0] * len(keys)
store.batch_put_from_multi_buffers.side_effect = lambda keys, *a: [256] * len(keys)
thread = _make_store_sending_thread(store, tp_rank=0, put_step=2)
thread.group_put_steps = [1]
thread.add_stored_request("req-a")
thread._handle_request(
ReqMeta(
req_id="req-a",
token_len_chunk=64,
block_ids=([0, 1, 2, 3],),
block_hashes=[b"a0", b"a1", b"a2", b"a3"],
can_save=True,
)
)
keys = store.batch_is_exist.call_args.args[0]
assert len(keys) == 4
def test_store_sending_thread_retries_skipped_range_after_pressure():
store = MagicMock()
store.batch_is_exist.side_effect = lambda keys: [0] * len(keys)
@@ -1274,7 +1297,8 @@ def test_worker_put_striding_covers_every_rank_get_namespace(
]
assert len(keys) == len(block_hashes)
# PUT side: mirrors KVCacheStoreSendingThread's striding slice.
put_keys.update(keys[w.tp_rank % w.put_step :: w.put_step])
put_step = w._group_tp_replication_factors[0]
put_keys.update(keys[w.tp_rank % put_step :: put_step])
# GET side: KVCacheStoreRecvingThread fetches every key.
get_keys_per_rank[tp_rank] = set(keys)
@@ -1613,6 +1637,15 @@ def _register_with_mocked_threads(
worker.register_kv_caches(kv_caches)
def _refresh_group_tp_replication_factors(
worker: mooncake_store_worker.MooncakeStoreWorker,
) -> None:
worker._group_tp_replication_factors = (
worker._compute_group_tp_replication_factors()
)
worker._init_lookup_key_prefixes()
def _make_bare_worker(
*,
num_gpu_blocks: int = 10,
@@ -1630,11 +1663,9 @@ def _make_bare_worker(
worker.cache_config.num_gpu_blocks = num_gpu_blocks
worker.store = MagicMock()
worker.store.register_buffer.return_value = 0
worker.use_mla = False
worker.kv_role = kv_role
worker.block_size = block_size
worker.tp_rank = 0
worker.put_step = 1
worker.enable_kv_events = False
worker.kv_send_thread = None
worker.kv_recv_threads = []
@@ -1664,7 +1695,6 @@ def _make_bare_worker(
worker.pcp_size = 1
worker.dcp_size = 1
worker.hash_block_size = block_size
worker.metadata = KeyMetadata("test-model", 0, 0, 0, 0)
# Pre-build a single-group token_dbs so lookup-only tests don't have to
# call register_kv_caches.
worker.token_dbs = [
@@ -1679,7 +1709,7 @@ def _make_bare_worker(
scheduler_block_size=block_size,
hash_block_size=block_size,
)
worker._init_lookup_key_prefixes()
_refresh_group_tp_replication_factors(worker)
return worker
@@ -1688,9 +1718,8 @@ def test_lookup_key_prefixes_cover_dcp_rank_namespaces():
worker.tp_size = 4
worker.num_kv_head = 1
worker.dcp_size = 4
worker._init_lookup_key_prefixes()
_refresh_group_tp_replication_factors(worker)
assert worker._lookup_expected_per_key == 4
assert worker._lookup_key_prefixes[0] == (
"test-model@tp_rank:0@pcp0@dcp0@pp_rank:0@group:0",
"test-model@tp_rank:1@pcp0@dcp1@pp_rank:0@group:0",
@@ -1705,21 +1734,192 @@ def test_lookup_key_prefixes_cover_pcp_rank_namespaces():
worker.num_kv_head = 1
worker.pcp_size = 2
worker.dcp_size = 1
worker._init_lookup_key_prefixes()
_refresh_group_tp_replication_factors(worker)
assert worker._lookup_expected_per_key == 2
assert worker._lookup_key_prefixes[0] == (
"test-model@tp_rank:0@pcp0@dcp0@pp_rank:0@group:0",
"test-model@tp_rank:0@pcp1@dcp0@pp_rank:0@group:0",
)
def test_lookup_key_prefixes_expand_tp_sharded_groups_per_rank():
"""Replicated attention needs one namespace; sharded Mamba needs every rank."""
from vllm.v1.kv_cache_interface import (
FullAttentionSpec,
KVCacheGroupSpec,
MambaSpec,
)
worker = _make_bare_worker(block_size=16)
worker.tp_size = 2
worker.num_kv_head = 1
fa = FullAttentionSpec(block_size=16, num_kv_heads=8, head_size=64, dtype=None)
mamba = MambaSpec(
block_size=16,
shapes=((1, 1),),
dtypes=(torch.float32,),
mamba_cache_mode="align",
)
worker._kv_cache_groups = [
KVCacheGroupSpec(["l0"], fa),
KVCacheGroupSpec(["l1"], mamba),
]
worker.token_dbs = [
ChunkedTokenDatabase(
KeyMetadata("test-model", 0, 0, 0, 0, group_id=0), block_size=16
),
ChunkedTokenDatabase(
KeyMetadata("test-model", 1, 0, 0, 0, group_id=1), block_size=16
),
]
_refresh_group_tp_replication_factors(worker)
assert worker._lookup_key_prefixes[0] == (
"test-model@tp_rank:0@pcp0@dcp0@pp_rank:0@group:0",
)
assert worker._lookup_key_prefixes[1] == (
"test-model@tp_rank:0@pcp0@dcp0@pp_rank:0@group:1",
"test-model@tp_rank:1@pcp0@dcp0@pp_rank:0@group:1",
)
def test_group_tp_replication_factors_mixed_mla_gqa_mamba():
from vllm.v1.kv_cache_interface import (
FullAttentionSpec,
KVCacheGroupSpec,
MambaSpec,
MLAAttentionSpec,
)
worker = _make_bare_worker(block_size=16)
worker.tp_size = 4
worker.num_kv_head = 2
mla = MLAAttentionSpec(block_size=16, num_kv_heads=1, head_size=64, dtype=None)
gqa = FullAttentionSpec(block_size=16, num_kv_heads=8, head_size=64, dtype=None)
mamba = MambaSpec(
block_size=16,
shapes=((1, 1),),
dtypes=(torch.float32,),
mamba_cache_mode="align",
)
worker._kv_cache_groups = [
KVCacheGroupSpec(["l0"], mla),
KVCacheGroupSpec(["l1"], gqa),
KVCacheGroupSpec(["l2"], mamba),
]
worker.token_dbs = [
ChunkedTokenDatabase(
KeyMetadata("test-model", 0, 0, 0, 0, group_id=g_idx), block_size=16
)
for g_idx in range(3)
]
_refresh_group_tp_replication_factors(worker)
assert worker._group_tp_replication_factors == (4, 2, 1)
assert worker._lookup_key_prefixes[0] == (
"test-model@tp_rank:0@pcp0@dcp0@pp_rank:0@group:0",
)
assert worker._lookup_key_prefixes[1] == (
"test-model@tp_rank:0@pcp0@dcp0@pp_rank:0@group:1",
"test-model@tp_rank:1@pcp0@dcp0@pp_rank:0@group:1",
)
assert worker._lookup_key_prefixes[2] == (
"test-model@tp_rank:0@pcp0@dcp0@pp_rank:0@group:2",
"test-model@tp_rank:1@pcp0@dcp0@pp_rank:0@group:2",
"test-model@tp_rank:2@pcp0@dcp0@pp_rank:0@group:2",
"test-model@tp_rank:3@pcp0@dcp0@pp_rank:0@group:2",
)
@pytest.mark.parametrize("spec_order", [("mla", "gqa"), ("gqa", "mla")])
def test_uniform_group_uses_common_inner_replication_factor(spec_order):
from vllm.v1.kv_cache_interface import (
FullAttentionSpec,
KVCacheGroupSpec,
MLAAttentionSpec,
UniformTypeKVCacheSpecs,
)
worker = _make_bare_worker(block_size=16)
worker.tp_size = 4
worker.num_kv_head = 2
specs_by_name = {
"mla": MLAAttentionSpec(
block_size=16, num_kv_heads=1, head_size=64, dtype=None
),
"gqa": FullAttentionSpec(
block_size=16, num_kv_heads=1, head_size=64, dtype=None
),
}
inner_specs = {name: specs_by_name[name] for name in spec_order}
uniform_spec = UniformTypeKVCacheSpecs(
block_size=16,
kv_cache_specs=inner_specs,
)
worker._kv_cache_groups = [
KVCacheGroupSpec(list(inner_specs), uniform_spec),
]
_refresh_group_tp_replication_factors(worker)
assert worker._group_tp_replication_factors == (2,)
assert worker._lookup_key_prefixes[0] == (
"test-model@tp_rank:0@pcp0@dcp0@pp_rank:0@group:0",
"test-model@tp_rank:1@pcp0@dcp0@pp_rank:0@group:0",
)
def test_lookup_rejects_boundary_missing_one_mamba_shard():
from vllm.v1.kv_cache_interface import (
FullAttentionSpec,
KVCacheGroupSpec,
MambaSpec,
)
worker = _make_bare_worker(block_size=16)
worker.tp_size = 2
worker.num_kv_head = 1
fa = FullAttentionSpec(block_size=16, num_kv_heads=8, head_size=64, dtype=None)
mamba = MambaSpec(
block_size=16,
shapes=((1, 1),),
dtypes=(torch.float32,),
mamba_cache_mode="align",
)
worker._kv_cache_groups = [
KVCacheGroupSpec(["l0"], fa),
KVCacheGroupSpec(["l1"], mamba),
]
worker.token_dbs = [
ChunkedTokenDatabase(
KeyMetadata("test-model", 0, 0, 0, 0, group_id=0), block_size=16
),
ChunkedTokenDatabase(
KeyMetadata("test-model", 1, 0, 0, 0, group_id=1), block_size=16
),
]
worker.coord = mooncake_store_worker.MooncakeStoreCoordinator(
worker._kv_cache_groups,
scheduler_block_size=16,
hash_block_size=16,
)
_refresh_group_tp_replication_factors(worker)
worker.store.batch_is_exist.side_effect = lambda keys: [1] * len(keys)
assert worker.lookup(32, [b"h0", b"h1"]) == 32
worker.store.batch_is_exist.side_effect = lambda keys: [
0 if "tp_rank:1" in k and "group:1" in k else 1 for k in keys
]
assert worker.lookup(32, [b"h0", b"h1"]) == 0
def test_lookup_requires_all_dcp_rank_namespaces():
worker = _make_bare_worker(block_size=16)
worker.tp_size = 4
worker.num_kv_head = 1
worker.dcp_size = 4
worker._init_lookup_key_prefixes()
_refresh_group_tp_replication_factors(worker)
worker.store.batch_is_exist.return_value = [1, 1, 0, 1]
assert worker.lookup(16, [b"a0"]) == 0
@@ -1850,7 +2050,7 @@ def test_lookup_checks_all_potential_swa_hit_boundaries():
hash_block_size=8,
retention_interval=0,
)
worker._init_lookup_key_prefixes()
_refresh_group_tp_replication_factors(worker)
# Candidate order: 3 full-attention chunks, then SWA chunks 3, 7, 11.
# Only the first full chunk and the SWA chunk ending at token 32 exist, so
# lookup should recover a 32-token external prefix hit. A sparse
@@ -1911,7 +2111,7 @@ def test_lookup_applies_swa_mask_before_accessing_hashes():
hash_block_size=8,
retention_interval=0,
)
worker._init_lookup_key_prefixes()
_refresh_group_tp_replication_factors(worker)
block_hashes = _RecordingBlockHashes([f"h{i}".encode() for i in range(12)])
accessed_before_rpc: list[int] = []
@@ -66,7 +66,15 @@ from vllm.v1.core.kv_cache_utils import (
maybe_convert_block_hash,
resolve_kv_cache_block_sizes,
)
from vllm.v1.kv_cache_interface import KVCacheConfig, KVCacheGroupSpec
from vllm.v1.kv_cache_interface import (
KVCacheConfig,
KVCacheGroupSpec,
KVCacheSpec,
MambaSpec,
MLAAttentionSpec,
SlidingWindowMLASpec,
UniformTypeKVCacheSpecs,
)
from .metrics import MooncakeStoreConnectorStats
@@ -448,7 +456,7 @@ class KVCacheStoreSendingThread(KVTransferThread):
token_databases: list[ChunkedTokenDatabase],
block_size: int,
tp_rank: int,
put_step: int,
group_put_steps: Sequence[int],
kv_role: str,
ready_event: threading.Event,
enable_kv_event: bool = False,
@@ -464,7 +472,8 @@ class KVCacheStoreSendingThread(KVTransferThread):
name="KVCacheStoreSendingThread",
record_operation=record_operation,
)
self.put_step = put_step
# Only ranks with identical group bytes may stripe PUTs (e.g., MLA).
self.group_put_steps = group_put_steps
self.coord = coord
self.kv_role = kv_role
self.stored_requests: defaultdict[str, int] = defaultdict(int)
@@ -568,13 +577,14 @@ class KVCacheStoreSendingThread(KVTransferThread):
group_indices: list[int] = []
for g_idx, db in enumerate(self.token_databases):
# Rotate the stride phase per group to balance load across ranks.
put_step_rank = (self.tp_rank + g_idx) % self.put_step
put_step = self.group_put_steps[g_idx]
put_step_rank = (self.tp_rank + g_idx) % put_step
for start, end, block_hash in db.process_tokens(
token_len,
req_meta.block_hashes,
mask_num=save_start,
chunk_mask=store_masks[g_idx],
put_step=self.put_step,
put_step=put_step,
put_step_rank=put_step_rank,
):
starts.append(start)
@@ -1002,44 +1012,7 @@ class MooncakeStoreWorker:
)
self.num_layers = model_config.get_num_layers(parallel_config)
self.use_mla = False
if (
hasattr(model_config, "use_mla")
and isinstance(model_config.use_mla, bool)
and model_config.use_mla
):
self.use_mla = True
if self.use_mla:
self.num_kv_head = 1
else:
self.num_kv_head = model_config.get_total_num_kv_heads()
if self.num_kv_head < self.tp_size and self.dcp_size <= 1:
# Dedup: TP ranks holding the same KV heads stripe PUTs across
# one shared key namespace. DCP splits the TP group, so with
# DCP>1 those ranks have different `@dcpN` namespaces and
# striping would leave keys unwritten (OBJECT_NOT_FOUND on
# GET). PCP is outer to TP (pcp_rank is constant within a TP
# group), so it needs no guard.
self.put_step = self.tp_size // self.num_kv_head
self.head_or_tp_rank = self.tp_rank // self.put_step
else:
self.head_or_tp_rank = self.tp_rank
self.put_step = 1
self.metadata = KeyMetadata(
model_name=model_config.model.rstrip("/").split("/")[-1],
tp_rank=self.head_or_tp_rank,
pcp_rank=self.pcp_rank,
dcp_rank=self.dcp_rank,
pp_rank=self.pp_rank,
cache_prefix=str(
vllm_config.kv_transfer_config.kv_connector_extra_config.get(
"cache_prefix", ""
)
),
)
self.num_kv_head = model_config.get_total_num_kv_heads()
# Initialize MooncakeDistributedStore with its own TransferEngine
store_config = MooncakeStoreConfig.load_from_config()
@@ -1156,10 +1129,32 @@ class MooncakeStoreWorker:
retention_interval=envs.VLLM_PREFIX_CACHE_RETENTION_INTERVAL,
)
# One ChunkedTokenDatabase per group; addresses populated in
# register_kv_caches once the kv-cache layout is known.
# register_kv_caches once the kv-cache layout is known. Each group's
# key namespace is its TP shard id: ranks holding identical bytes
# (MLA / shared GQA KV heads) share a namespace, TP-sharded Mamba
# state gets one namespace per rank.
metadata = KeyMetadata(
model_name=model_config.model.rstrip("/").split("/")[-1],
tp_rank=self.tp_rank,
pcp_rank=self.pcp_rank,
dcp_rank=self.dcp_rank,
pp_rank=self.pp_rank,
cache_prefix=str(
vllm_config.kv_transfer_config.kv_connector_extra_config.get(
"cache_prefix", ""
)
),
)
self._group_tp_replication_factors: tuple[int, ...] = (
self._compute_group_tp_replication_factors()
)
self.token_dbs: list[ChunkedTokenDatabase] = [
ChunkedTokenDatabase(
dataclasses.replace(self.metadata, group_id=g_idx),
dataclasses.replace(
metadata,
group_id=g_idx,
tp_rank=self.tp_rank // self._group_tp_replication_factors[g_idx],
),
g.kv_cache_spec.block_size,
hash_block_size=self.hash_block_size,
)
@@ -1167,27 +1162,50 @@ class MooncakeStoreWorker:
]
self._init_lookup_key_prefixes()
def _init_lookup_key_prefixes(self) -> None:
"""Prepare per-group key prefixes across parallel rank namespaces."""
# (tp_rank, pcp_rank, dcp_rank, pp_rank) namespaces
def _spec_tp_replication_factor(self, spec: KVCacheSpec) -> int:
if self.dcp_size > 1:
# DCP reuses the TP workers and splits each TP group into
# contiguous DCP groups, so dcp_rank == tp_rank % dcp_size.
# Store/load paths do not apply KV-head dedup under DCP
rank_namespaces = tuple(
(tp_rank, pcp_rank, tp_rank % self.dcp_size, pp_rank)
return 1
inner_specs = (
tuple(spec.kv_cache_specs.values())
if isinstance(spec, UniformTypeKVCacheSpecs)
else (spec,)
)
# Any rank-specific state makes the whole packed value rank-specific.
if any(isinstance(inner, MambaSpec) for inner in inner_specs):
return 1
# A pure MLA packed value is replicated on every TP rank.
if all(
isinstance(inner, (MLAAttentionSpec, SlidingWindowMLASpec))
for inner in inner_specs
):
return self.tp_size
return max(1, self.tp_size // self.num_kv_head)
def _compute_group_tp_replication_factors(self) -> tuple[int, ...]:
"""Return the number of byte-identical TP replicas per cache group.
DCP and Mamba use 1; MLA uses ``tp_size``; GQA uses
``tp_size // num_kv_head``.
"""
return tuple(
self._spec_tp_replication_factor(group.kv_cache_spec)
for group in self._kv_cache_groups
)
def _init_lookup_key_prefixes(self) -> None:
def rank_namespaces(factor: int) -> tuple[tuple[int, int, int, int], ...]:
if self.dcp_size > 1:
# DCP is a TP subdivision: dcp_rank == tp_rank % dcp_size.
return tuple(
(tp_rank, pcp_rank, tp_rank % self.dcp_size, pp_rank)
for pcp_rank in range(self.pcp_size)
for tp_rank in range(self.tp_size)
for pp_rank in range(self.pp_size)
)
return tuple(
(shard_rank, pcp_rank, 0, pp_rank)
for pcp_rank in range(self.pcp_size)
for tp_rank in range(self.tp_size)
for pp_rank in range(self.pp_size)
)
else:
# Without DCP, TP ranks that share a KV head write identical KV, so
# lookup only needs one TP namespace per unique KV head.
tp_count = min(self.tp_size, self.num_kv_head)
rank_namespaces = tuple(
(tp_rank, pcp_rank, 0, pp_rank)
for pcp_rank in range(self.pcp_size)
for tp_rank in range(tp_count)
for shard_rank in range(self.tp_size // factor)
for pp_rank in range(self.pp_size)
)
@@ -1200,11 +1218,12 @@ class MooncakeStoreWorker:
dcp_rank=dcp_rank,
pp_rank=pp_rank,
)
for tp_rank, pcp_rank, dcp_rank, pp_rank in rank_namespaces
for tp_rank, pcp_rank, dcp_rank, pp_rank in rank_namespaces(
self._group_tp_replication_factors[g_idx]
)
)
for db in self.token_dbs
for g_idx, db in enumerate(self.token_dbs)
)
self._lookup_expected_per_key = len(rank_namespaces)
def register_cross_layers_kv_caches(self, kv_cache: torch.Tensor) -> None:
"""Register a cross-layers KV cache tensor.
@@ -1297,7 +1316,7 @@ class MooncakeStoreWorker:
self.token_dbs,
self.block_size,
self.tp_rank,
self.put_step,
self._group_tp_replication_factors,
self.kv_role,
ready_event_sending,
self.enable_kv_events,
@@ -1527,16 +1546,16 @@ class MooncakeStoreWorker:
logger.error("Remote connection failed in lookup: %s", e)
return 0
# A (group, hash) is "present" only when every TP*PP rank has it.
ranks_per_candidate = self._lookup_expected_per_key
exists_set = {
(g_idx, hash_bytes)
for i, (g_idx, hash_bytes) in enumerate(candidate_meta)
if all(
res[i * ranks_per_candidate + j] == 1
for j in range(ranks_per_candidate)
)
}
# A (group, hash) is "present" only when every namespace that will be
# loaded has it (per-group count: sharded groups need every rank's
# shard, replicated groups one namespace per unique KV head).
exists_set = set()
pos = 0
for g_idx, hash_bytes in candidate_meta:
count = len(self._lookup_key_prefixes[g_idx])
if all(res[pos + j] == 1 for j in range(count)):
exists_set.add((g_idx, hash_bytes))
pos += count
cached_block_pool = ExternalCachedBlockPool(
self.hash_block_size,