apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-05 16:48:03 +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

8
fastdeploy/model_executor/layers/attention/mla_attention_backend.py
View File

@@ -35,15 +35,13 @@ if current_platform.is_cuda() and not current_platform.is_dcu():
prefill_mla_write_cache)
if TYPE_CHECKING:
from paddle._typing.dtype_like import _DTypeLiteral
from fastdeploy.model_executor.forward_meta import ForwardMeta
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.model_executor.layers.attention.utils import \
init_rank_and_device_id
from fastdeploy.worker.forward_meta import ForwardMeta
from fastdeploy.model_executor.layers.attention.utils import init_rank_and_device_id
def yarn_get_mscale(scale=1, mscale=1):
@@ -71,7 +69,7 @@ class MLAAttentionMetadata(AttentionMetadata):
decoder_tile_ids_per_batch: paddle.Tensor = None
decoder_num_blocks: paddle.Tensor = None
_dtype: _DTypeLiteral = paddle.bfloat16
_dtype: paddle.dtype = paddle.bfloat16
encoder_max_partition_size: int = 32768
max_partition_size: int = 32768
block_tables: Optional[paddle.Tensor] = None