apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-05 08:37:06 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

[Executor] Move forward_meta.py to fastdeploy/model_executor (#2774)

Browse Source 
* Use PEP 563 in attention.py and fix conflict

* merge commit

* Change what was left out last time
This commit is contained in:
littledgg
2025-07-10 20:36:51 +08:00
committed by GitHub
parent 8c660a0dfb
commit 59071268b6
27 changed files with 53 additions and 55 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
fastdeploy/model_executor/layers/backends/gcu/attention/flash_attn_backend.py
View File

@@ -24,15 +24,12 @@ import paddle
import numpy as np
if TYPE_CHECKING:
from paddle._typing.dtype_like import _DTypeLiteral
from fastdeploy.config import FDConfig
from fastdeploy.model_executor.layers.attention.attention import Attention
from fastdeploy.model_executor.layers.attention.base_attention_backend import (
AttentionBackend, AttentionMetadata)
from fastdeploy.worker.forward_meta import ForwardMeta, ForwardMode
if TYPE_CHECKING:
from fastdeploy.model_executor.forward_meta import ForwardMeta, ForwardMode
from fastdeploy.model_executor.ops.gcu import (fused_rotary_embedding,
mem_efficient_attention,
@@ -47,7 +44,7 @@ class GCUFlashAttnMetadata(AttentionMetadata):
"""
forward_mode: ForwardMode = ForwardMode.MIXED
_dtype: _DTypeLiteral = paddle.bfloat16
_dtype: paddle.dtype = paddle.bfloat16
seq_lens_encoder: Optional[paddle.Tensor] = None
seq_lens_decoder: Optional[paddle.Tensor] = None

10
fastdeploy/model_executor/layers/backends/gcu/attention/mem_efficient_attn_backend.py
View File

@@ -25,28 +25,26 @@ import paddle
import numpy as np
import math
if TYPE_CHECKING:
from paddle._typing.dtype_like import _DTypeLiteral
from fastdeploy.config import FDConfig
from fastdeploy.model_executor.layers.attention.attention import Attention
from fastdeploy.model_executor.layers.attention.base_attention_backend import (
AttentionBackend, AttentionMetadata)
from fastdeploy.worker.forward_meta import ForwardMeta, ForwardMode
from fastdeploy.model_executor.ops.gcu import (fused_rotary_embedding,
mem_efficient_attention,
flash_attn_var_len)
from paddleformers.utils.log import logger
if TYPE_CHECKING:
from fastdeploy.model_executor.forward_meta import ForwardMeta, ForwardMode
@dataclass
class GCUMemEfficientAttnMetadata(AttentionMetadata):
"""
GCUMemEfficientAttnMetadata
"""
forward_mode: ForwardMode = ForwardMode.MIXED
_dtype: _DTypeLiteral = paddle.bfloat16
_dtype: paddle.dtype = paddle.bfloat16
seq_lens_encoder: Optional[paddle.Tensor] = None
seq_lens_decoder: Optional[paddle.Tensor] = None