[Bugfix][KV Connector] Fix DecodeBench DCP block selection (#54679)
Signed-off-by: Jeff Ma <jeffjma@umich.edu> Co-authored-by: OpenAI Codex <noreply@openai.com>
This commit is contained in:
co-authored by
OpenAI Codex
parent
16f16876fd
commit
18c53727ce
@@ -22,18 +22,26 @@ from vllm.distributed.kv_transfer.kv_connector.v1.decode_bench_connector import
|
||||
)
|
||||
from vllm.forward_context import ForwardContext
|
||||
from vllm.utils.hashing import sha256
|
||||
from vllm.v1.core.kv_cache_utils import get_request_block_hasher, init_none_hash
|
||||
from vllm.v1.core.kv_cache_manager import KVCacheBlocks
|
||||
from vllm.v1.core.kv_cache_utils import (
|
||||
KVCacheBlock,
|
||||
get_request_block_hasher,
|
||||
init_none_hash,
|
||||
)
|
||||
from vllm.v1.core.sched.output import SchedulerOutput
|
||||
from vllm.v1.core.sched.scheduler import Scheduler
|
||||
from vllm.v1.kv_cache_interface import (
|
||||
FullAttentionSpec,
|
||||
KVCacheConfig,
|
||||
KVCacheGroupSpec,
|
||||
SlidingWindowSpec,
|
||||
)
|
||||
from vllm.v1.request import Request
|
||||
|
||||
from .utils import (
|
||||
EOS_TOKEN_ID,
|
||||
create_model_runner_output,
|
||||
create_request,
|
||||
create_scheduler,
|
||||
create_vllm_config,
|
||||
)
|
||||
@@ -276,6 +284,63 @@ def test_decode_bench_connector_uses_per_group_block_sizes():
|
||||
assert num_tokens == 17
|
||||
|
||||
|
||||
def test_decode_bench_connector_uses_per_group_dcp_block_sizes():
|
||||
"""DCP scales full-attention blocks but not sliding-window blocks."""
|
||||
block_size = 16
|
||||
dcp_world_size = 2
|
||||
vllm_config = create_vllm_config(
|
||||
block_size=block_size,
|
||||
kv_connector="DecodeBenchConnector",
|
||||
)
|
||||
vllm_config.parallel_config.decode_context_parallel_size = dcp_world_size
|
||||
full_attention = FullAttentionSpec(
|
||||
block_size=block_size,
|
||||
num_kv_heads=1,
|
||||
head_size=1,
|
||||
dtype=torch.float32,
|
||||
)
|
||||
sliding_window = SlidingWindowSpec(
|
||||
block_size=block_size,
|
||||
num_kv_heads=1,
|
||||
head_size=1,
|
||||
dtype=torch.float32,
|
||||
sliding_window=2 * block_size,
|
||||
)
|
||||
connector = DecodeBenchConnector(
|
||||
vllm_config,
|
||||
KVConnectorRole.SCHEDULER,
|
||||
KVCacheConfig(
|
||||
num_blocks=8,
|
||||
kv_cache_tensors=[],
|
||||
kv_cache_groups=[
|
||||
KVCacheGroupSpec(["full_attention"], full_attention),
|
||||
KVCacheGroupSpec(["sliding_window"], sliding_window),
|
||||
],
|
||||
),
|
||||
)
|
||||
num_external_tokens = block_size * dcp_world_size + 1
|
||||
request = create_request(
|
||||
request_id=1,
|
||||
num_tokens=num_external_tokens + 1,
|
||||
block_size=block_size,
|
||||
)
|
||||
blocks = KVCacheBlocks(
|
||||
(
|
||||
[KVCacheBlock(block_id=i) for i in (1, 2, 3)],
|
||||
[KVCacheBlock(block_id=i) for i in (5, 6, 7, 8)],
|
||||
)
|
||||
)
|
||||
|
||||
connector.update_state_after_alloc(request, blocks, num_external_tokens)
|
||||
metadata = connector.build_connector_meta(SchedulerOutput.make_empty())
|
||||
|
||||
assert isinstance(metadata, DecodeBenchConnectorMetadata)
|
||||
assert metadata.reqs_to_fill[request.request_id] == (
|
||||
([1, 2], [5, 6, 7]),
|
||||
num_external_tokens,
|
||||
)
|
||||
|
||||
|
||||
def test_decode_bench_connector_no_refill():
|
||||
"""Test that DecodeBenchConnector only fills once per request."""
|
||||
block_size = 16
|
||||
|
||||
@@ -47,6 +47,10 @@ from vllm.distributed.kv_transfer.kv_connector.v1.base import (
|
||||
from vllm.logger import init_logger
|
||||
from vllm.utils.math_utils import cdiv
|
||||
from vllm.v1.attention.backend import AttentionMetadata
|
||||
from vllm.v1.core.kv_cache_utils import (
|
||||
dcp_world_size_for_kv_cache_spec,
|
||||
resolve_dcp_kv_block_size,
|
||||
)
|
||||
|
||||
if TYPE_CHECKING:
|
||||
from vllm.config import VllmConfig
|
||||
@@ -187,9 +191,16 @@ class DecodeBenchConnectorScheduler:
|
||||
"""Scheduler-side implementation for DecodeBenchConnector."""
|
||||
|
||||
def __init__(self, vllm_config: "VllmConfig", kv_cache_config: "KVCacheConfig"):
|
||||
self.vllm_config = vllm_config
|
||||
dcp_world_size = vllm_config.parallel_config.decode_context_parallel_size
|
||||
self.group_block_sizes = tuple(
|
||||
group.kv_cache_spec.block_size for group in kv_cache_config.kv_cache_groups
|
||||
resolve_dcp_kv_block_size(
|
||||
group.kv_cache_spec,
|
||||
dcp_world_size_for_kv_cache_spec(
|
||||
group.kv_cache_spec,
|
||||
dcp_world_size,
|
||||
),
|
||||
)
|
||||
for group in kv_cache_config.kv_cache_groups
|
||||
)
|
||||
|
||||
# Track which requests have already been filled
|
||||
@@ -253,7 +264,7 @@ class DecodeBenchConnectorScheduler:
|
||||
# block_groups is a tuple of lists, one per KV cache group
|
||||
block_groups = blocks.get_block_ids()
|
||||
|
||||
# Extract the blocks covering num_external_tokens from each group
|
||||
# Extract the blocks covering the external tokens from each group
|
||||
block_ids_per_group = tuple(
|
||||
group_blocks[: cdiv(num_external_tokens, group_block_size)]
|
||||
for group_blocks, group_block_size in zip(
|
||||
@@ -269,12 +280,14 @@ class DecodeBenchConnectorScheduler:
|
||||
)
|
||||
self._filled_requests.add(req_id)
|
||||
|
||||
block_counts = tuple(len(group) for group in block_ids_per_group)
|
||||
logger.debug(
|
||||
"DecodeBenchConnector: Allocated %s blocks across %d KV cache groups "
|
||||
"for request %s",
|
||||
tuple(len(block_ids) for block_ids in block_ids_per_group),
|
||||
"DecodeBenchConnector: Selected %d total blocks across %d KV cache "
|
||||
"groups for request %s (per-group counts: %s)",
|
||||
sum(block_counts),
|
||||
len(block_groups),
|
||||
req_id,
|
||||
", ".join(map(str, block_counts)),
|
||||
)
|
||||
|
||||
def build_connector_meta(
|
||||
@@ -302,9 +315,6 @@ class DecodeBenchConnectorWorker:
|
||||
"""Worker-side implementation for DecodeBenchConnector."""
|
||||
|
||||
def __init__(self, vllm_config: "VllmConfig", kv_cache_config: "KVCacheConfig"):
|
||||
self.vllm_config = vllm_config
|
||||
self.block_size = vllm_config.cache_config.block_size
|
||||
|
||||
# Get fill parameters from extra config
|
||||
kv_transfer_config = vllm_config.kv_transfer_config
|
||||
assert kv_transfer_config is not None
|
||||
@@ -348,13 +358,15 @@ class DecodeBenchConnectorWorker:
|
||||
for group_idx, block_ids in enumerate(block_ids_per_group):
|
||||
self._fill_blocks(group_idx, block_ids, num_tokens)
|
||||
|
||||
block_counts = tuple(len(group) for group in block_ids_per_group)
|
||||
logger.debug(
|
||||
"DecodeBenchConnector: Filled %d blocks (%d tokens) across %d groups "
|
||||
"for request %s",
|
||||
len(block_ids_per_group[0]) if block_ids_per_group else 0,
|
||||
"DecodeBenchConnector: Filled %d total blocks (%d tokens) across "
|
||||
"%d groups for request %s (per-group counts: %s)",
|
||||
sum(block_counts),
|
||||
num_tokens,
|
||||
len(block_ids_per_group),
|
||||
req_id,
|
||||
", ".join(map(str, block_counts)),
|
||||
)
|
||||
|
||||
def _fill_blocks(self, group_idx: int, block_ids: list[int], num_tokens: int):
|
||||
|
||||
Reference in New Issue
Block a user