apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2026-04-24 01:29:57 +08:00
Code Issues Actions 9 Packages Projects Releases Wiki Activity

Support Norm before Rope (#6332)

Browse Source 
Co-authored-by: K11OntheBoat <“ruianmaidanglao@163.com”>
This commit is contained in:
K11OntheBoat
2026-02-05 15:28:52 +08:00
committed by GitHub
parent 29a313a402
commit 116e2aea7a
4 changed files with 27 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files
fastdeploy/model_executor/layers/attention/append_attn_backend.py
+8 -4
View File
@@ -280,6 +280,10 @@ class AppendAttentionBackend(AttentionBackend):
sliding_window = layer.sliding_window
norm_after_rope_in_kernel = not getattr(layer, "qk_norm_before_rope", False)
q_norm_weight = getattr(layer, "q_norm_weight", None) if norm_after_rope_in_kernel else None
k_norm_weight = getattr(layer, "k_norm_weight", None) if norm_after_rope_in_kernel else None
if self.rope_3d:
assert len(forward_meta.rotary_embs.shape) == 6
else:
@@ -402,8 +406,8 @@ class AppendAttentionBackend(AttentionBackend):
layer.linear_smooth,
forward_meta.attn_mask_offsets,
metadata.kv_signal_data_list[layer.layer_id],
getattr(layer, "q_norm_weight", None),
getattr(layer, "k_norm_weight", None),
q_norm_weight,
k_norm_weight,
getattr(layer, "sinks", None),
getattr(layer, "rms_norm_eps", 1e-6),
metadata._fuse_kernel_compute_dtype,
@@ -458,8 +462,8 @@ class AppendAttentionBackend(AttentionBackend):
layer.linear_smooth,
forward_meta.attn_mask_offsets,
metadata.kv_signal_data_list[layer.layer_id],
getattr(layer, "q_norm_weight", None),
getattr(layer, "k_norm_weight", None),
q_norm_weight,
k_norm_weight,
getattr(layer, "sinks", None),
getattr(layer, "rms_norm_eps", 1e-6),
metadata._fuse_kernel_compute_dtype,