Convert formatting to use ruff instead of yapf + isort (#26247)
Signed-off-by: Harry Mellor <19981378+hmellor@users.noreply.github.com>
This commit is contained in:
@@ -14,6 +14,7 @@ class AttentionType:
|
||||
Attention type.
|
||||
Use string to be compatible with `torch.compile`.
|
||||
"""
|
||||
|
||||
DECODER = "decoder"
|
||||
"""Decoder attention between previous layer Q/K/V."""
|
||||
ENCODER = "encoder"
|
||||
@@ -26,6 +27,7 @@ class AttentionType:
|
||||
|
||||
class AttentionBackend(ABC):
|
||||
"""Abstract class for attention backends."""
|
||||
|
||||
# For some attention backends, we allocate an output tensor before
|
||||
# calling the custom op. When piecewise cudagraph is enabled, this
|
||||
# makes sure the output tensor is allocated inside the cudagraph.
|
||||
@@ -91,7 +93,6 @@ T = TypeVar("T", bound=AttentionMetadata)
|
||||
|
||||
|
||||
class AttentionLayer(Protocol):
|
||||
|
||||
_q_scale: torch.Tensor
|
||||
_k_scale: torch.Tensor
|
||||
_v_scale: torch.Tensor
|
||||
@@ -107,12 +108,10 @@ class AttentionLayer(Protocol):
|
||||
value: torch.Tensor,
|
||||
kv_cache: torch.Tensor,
|
||||
attn_metadata: AttentionMetadata,
|
||||
) -> torch.Tensor:
|
||||
...
|
||||
) -> torch.Tensor: ...
|
||||
|
||||
|
||||
class AttentionImpl(ABC, Generic[T]):
|
||||
|
||||
# Whether the attention impl can return the softmax lse for decode.
|
||||
# Some features like decode context parallelism require the softmax lse.
|
||||
can_return_lse_for_decode: bool = False
|
||||
@@ -129,14 +128,16 @@ class AttentionImpl(ABC, Generic[T]):
|
||||
self = super().__new__(cls)
|
||||
try:
|
||||
from vllm.distributed.parallel_state import get_dcp_group
|
||||
|
||||
self.dcp_world_size = get_dcp_group().world_size
|
||||
self.dcp_rank = get_dcp_group().rank_in_group
|
||||
except AssertionError:
|
||||
# DCP might not be initialized in testing
|
||||
self.dcp_world_size = 1
|
||||
self.dcp_rank = 0
|
||||
self.need_to_return_lse_for_decode = self.dcp_world_size > 1 \
|
||||
and self.can_return_lse_for_decode
|
||||
self.need_to_return_lse_for_decode = (
|
||||
self.dcp_world_size > 1 and self.can_return_lse_for_decode
|
||||
)
|
||||
return self
|
||||
|
||||
@abstractmethod
|
||||
@@ -183,7 +184,6 @@ class AttentionImpl(ABC, Generic[T]):
|
||||
|
||||
|
||||
class MLAAttentionImpl(AttentionImpl[T], Generic[T]):
|
||||
|
||||
@abstractmethod
|
||||
def forward(
|
||||
self,
|
||||
|
||||
@@ -1,6 +1,7 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
"""Attention backend utils"""
|
||||
|
||||
from dataclasses import dataclass
|
||||
from typing import Optional
|
||||
|
||||
|
||||
Reference in New Issue
Block a user