apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-10-05 00:33:03 +08:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

Revert "[Feature] block sparse attention (#3209)" (#3647)

Browse Source 
This reverts commit 646a0c2fd8.
This commit is contained in:
Jiang-Jia-Jun
2025-08-27 17:35:04 +08:00
committed by GitHub
parent b2afdf4fc6
commit c694fa2879
31 changed files with 10 additions and 6507 deletions
Download Patch File Download Diff File Expand all files Collapse all files

41
fastdeploy/model_executor/layers/attention/attention.py
View File

@@ -28,11 +28,6 @@ from fastdeploy.model_executor.layers.quantization.quant_base import QuantMethod
if TYPE_CHECKING:
from fastdeploy.model_executor.forward_meta import ForwardMeta
import os
from safetensors import safe_open
from fastdeploy.model_executor.layers.utils import get_tensor
@@ -118,42 +113,6 @@ class Attention(nn.Layer):
self.k_norm_key = f"{self.prefix}.k_norm"
self.init_weight()
if fd_config.moba_attention_config is not None:
mlp_weight_path = os.path.join(
fd_config.model_config.model, fd_config.moba_attention_config.mlp_weight_name
)
self.moba_use_mlp = mlp_weight_path is not None and os.path.exists(mlp_weight_path)
moba_block_size = fd_config.moba_attention_config.moba_block_size
moba_max_seq_length = fd_config.moba_attention_config.moba_max_seq_length
if self.moba_use_mlp:
mlp_weight = {}
with safe_open(mlp_weight_path, framework="np", device="cpu") as f:
for key_name in f.keys():
weight = f.get_tensor(key_name)
weight = paddle.Tensor(weight, zero_copy=True)
weight = weight._copy_to(paddle.framework._current_expected_place(), False)
mlp_weight[key_name] = weight
if self.layer_id < fd_config.model_config.num_hidden_layers - 1:
self.attn_gate_weight = mlp_weight[
f"ernie.layers.{self.layer_id}.self_attn.attn_gate.weight"
].astype(paddle.get_default_dtype())[
fd_config.parallel_config.tensor_parallel_rank
* self.kv_num_heads : (fd_config.parallel_config.tensor_parallel_rank + 1)
* self.kv_num_heads
]
assert self.attn_gate_weight.shape[1] % moba_block_size == 0
self.cache_k_block_means = paddle.zeros(
[
fd_config.parallel_config.max_num_seqs,
moba_max_seq_length // moba_block_size,
self.kv_num_heads,
self.head_dim,
],
dtype=paddle.get_default_dtype(),
)
def init_weight(self):
self.q_norm_weight = self.create_parameter(
shape=[self.qk_head_dim],