apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-04 08:16:42 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

make append_attn supports mask_offset (#3138)

Browse Source 
* make append_attn supports mask_offset

* add unittest
This commit is contained in:
lzy
2025-08-14 18:40:55 +08:00
committed by GitHub
parent 6031f9a5f5
commit 1e06b9fa6d
10 changed files with 88 additions and 20 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
fastdeploy/model_executor/layers/attention/append_attn_backend.py
View File

@@ -62,6 +62,7 @@ class AppendAttentionMetadata(AttentionMetadata):
block_tables: Optional[paddle.Tensor] = None
rotary_embs: Optional[paddle.Tensor] = None
attn_mask: Optional[paddle.Tensor] = None
mask_offset: Optional[paddle.Tensor] = None
_fuse_kernel_compute_dtype: str = "bf16"
# pd_disaggregation
@@ -261,6 +262,7 @@ class AppendAttentionBackend(AttentionBackend):
getattr(layer, "cache_v_zp", None),
layer.linear_shift,
layer.linear_smooth,
metadata.mask_offset,
metadata.kv_signal_data_list[layer.layer_id],
getattr(layer, "q_norm_weight", None),
getattr(layer, "k_norm_weight", None),