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

[Sync] Update to latest code (#2679)

Browse Source 
* [Sync] Update to latest code

* Add new code files

* Add new code files

* update code

* Try to fix build.sh

* Try to fix build.sh

* Update code

* Update requirements.txt

* Update code

---------

Co-authored-by: Jiang-Jia-Jun <jiangjiajun@baidu.com>
This commit is contained in:
Jiang-Jia-Jun
2025-07-03 15:43:53 +08:00
committed by GitHub
parent d222248d00
commit 05c670e593
95 changed files with 9916 additions and 1312 deletions
Download Patch File Download Diff File Expand all files Collapse all files

56
fastdeploy/model_executor/layers/embeddings.py
View File

@@ -14,10 +14,15 @@
# limitations under the License.
"""
from typing import Dict
import numpy as np
import paddle
from paddle import nn
from paddle.distributed import fleet
from fastdeploy.config import FDConfig
from .utils import get_tensor
@@ -28,12 +33,12 @@ class VocabParallelEmbedding(nn.Layer):
def __init__(
self,
fd_config,
num_embeddings,
embedding_dim=768,
params_dtype="bfloat16",
fd_config: FDConfig,
num_embeddings: int,
embedding_dim: int = 768,
params_dtype: str = "bfloat16",
prefix="",
):
) -> None:
"""
Initialize the VocabParallelEmbedding layer for the model.
@@ -41,28 +46,28 @@ class VocabParallelEmbedding(nn.Layer):
fd_config (FDConfig): Arguments related to inference, containing
attributes such as weight_dtype, act_dtype, mp_size, hidden_size, head_dim,
num_attention_heads, and ffn_hidden_size.
num_embeddings : vocabulary size.
embedding_dim : size of hidden state.
params_dtype : data type of parameters.
prefix (str): Unique name of the layer, used for naming internal attributes,
you can give it any name you like.
num_embeddings (int) : vocabulary size.
embedding_dim (int) : size of hidden state.
params_dtype (str) : data type of parameters.
prefix (str): The name of current layer. Defaults to "".
"""
super().__init__()
self.fd_config = fd_config
hcg = fleet.get_hybrid_communicate_group()
self.mp_rank = hcg.get_model_parallel_rank()
self.column_cut = fd_config.parallel_config.column_cut
self.world_size = hcg.get_model_parallel_world_size()
self.ring_id = hcg.get_model_parallel_group().id
self.use_rope = fd_config.model_config.use_rope
self.rope_head_dim = fd_config.model_config.rope_head_dim
self.use_ep = fd_config.parallel_config.use_ep
self.hidden_dropout_prob = fd_config.model_config.hidden_dropout_prob
self.initializer_range = fd_config.model_config.initializer_range
self.sequence_parallel = fd_config.parallel_config.sequence_parallel
self.max_position_embeddings = fd_config.model_config.max_position_embeddings
self.freeze_embedding = fd_config.model_config.freeze_embedding
self.tie_word_embeddings = fd_config.model_config.tie_word_embeddings
self.mp_rank: int = hcg.get_model_parallel_rank()
self.column_cut = False
self.world_size: int = hcg.get_model_parallel_world_size()
self.ring_id: int = hcg.get_model_parallel_group().id
self.use_rope: bool = fd_config.model_config.use_rope
self.rope_head_dim: int = fd_config.model_config.rope_head_dim
self.use_ep: bool = fd_config.parallel_config.use_ep
self.hidden_dropout_prob: float = fd_config.model_config.hidden_dropout_prob
self.initializer_range: float = fd_config.model_config.initializer_range
self.sequence_parallel: bool = fd_config.parallel_config.sequence_parallel
self.max_position_embeddings: int = fd_config.model_config.max_position_embeddings
self.freeze_embedding: bool = fd_config.model_config.freeze_embedding
self.tie_word_embeddings: bool = fd_config.model_config.tie_word_embeddings
self.params_dtype: str = params_dtype
if self.use_ep:
self.word_embeddings = nn.Embedding(
@@ -109,7 +114,8 @@ class VocabParallelEmbedding(nn.Layer):
self.rope_head_dim_shape_tensor = paddle.ones((self.rope_head_dim),
dtype="int8")
def load_state_dict(self, state_dict):
def load_state_dict(self, state_dict: Dict[str,
paddle.Tensor | np.ndarray]):
"""
Load the checkpoint state dictionary into the layer.
@@ -125,7 +131,7 @@ class VocabParallelEmbedding(nn.Layer):
get_tensor(state_dict.pop(self.prefix + ".weight")).astype(
paddle.get_default_dtype()))
def forward(self, ids_remove_padding=None):
def forward(self, ids_remove_padding=None) -> paddle.Tensor:
"""
Defines the forward computation of the layer.