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

[GCU] Enable gcu CI (#3190)
Some checks failed
Deploy GitHub Pages / deploy (push) Has been cancelled
Details

Browse Source 
* [GCU] Update to the latest version

* [GCU] Enable CI
This commit is contained in:
EnflameGCU
2025-08-13 11:48:24 +08:00
committed by GitHub
parent ce9180241e
commit d1a92e3e17
6 changed files with 87 additions and 45 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
fastdeploy/model_executor/layers/backends/gcu/attention/flash_attn_backend.py
View File

@@ -76,6 +76,8 @@ class GCUFlashAttnBackend(AttentionBackend):
kv_num_heads: int,
num_heads: int,
head_dim: int,
encoder_block_shape_q: int = -1,
decoder_block_shape_q: int = -1,
):
"""
GCUFlashAttnBackend __init__
@@ -94,7 +96,7 @@ class GCUFlashAttnBackend(AttentionBackend):
self.head_dim = head_dim
self.scaling = 1.0 / (self.head_dim**0.5)
self.num_layers = fd_config.model_config.num_hidden_layers
self.position_ids_base = paddle.arange(self.max_seq_len)
self.position_ids_base = np.arange(self.max_seq_len)
# TODO(zhengjun): Need to adapt the allocation logic and
# temporarily allocate according to fixed size

4
fastdeploy/model_executor/layers/backends/gcu/attention/mem_efficient_attn_backend.py
View File

@@ -74,6 +74,8 @@ class GCUMemEfficientAttnBackend(AttentionBackend):
kv_num_heads: int,
num_heads: int,
head_dim: int,
encoder_block_shape_q: int = -1,
decoder_block_shape_q: int = -1,
):
"""
GCUMemEfficientAttnBackend __init__
@@ -92,7 +94,7 @@ class GCUMemEfficientAttnBackend(AttentionBackend):
self.head_dim = head_dim
self.scaling = 1.0 / (self.head_dim**0.5)
self.num_layers = fd_config.model_config.num_hidden_layers
self.position_ids_base = paddle.arange(self.max_seq_len)
self.position_ids_base = np.arange(self.max_seq_len)
# TODO(zhengjun): Need to adapt the allocation logic and
# temporarily allocate according to fixed size