apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-12-24 13:28:13 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

[Iluvatar] Support V1_KVCACHE_SCHEDULER and paddleocr-vl rope mode (#5555)

Browse Source
This commit is contained in:
yzwu
2025-12-18 18:14:25 +08:00
committed by GitHub
parent 48f3e9797e
commit ac013803f3
24 changed files with 1212 additions and 1090 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
fastdeploy/model_executor/models/paddleocr_vl/paddleocr_vl.py
View File

@@ -40,6 +40,7 @@ from fastdeploy.model_executor.utils import (
default_weight_loader,
process_weights_after_loading,
)
from fastdeploy.platforms import current_platform
from .projector import Projector
from .siglip import SiglipVisionModel
@@ -101,12 +102,19 @@ class PaddleOCRVLModel(nn.Layer):
forward_meta: ForwardMeta,
):
hidden_states = input_embeddings
if current_platform.is_iluvatar() and forward_meta.attn_backend.mixed:
hidden_states = forward_meta.attn_backend.transpose(hidden_states)
residual = None
for i in range(self.num_layers):
hidden_states, residual = self.layers[i](forward_meta, hidden_states, residual)
out = self.norm(hidden_states, residual)[0]
if current_platform.is_iluvatar() and forward_meta.attn_backend.mixed:
out = forward_meta.attn_backend.reverse_transpose(out)
return out