apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2025-09-26 20:41:53 +08:00
Code Issues Actions 5 Packages Projects Releases Wiki Activity

[Feature] support rl_tp_degree (#3934)
Some checks failed
CE Compile Job / ce_job_pre_check (push) Has been cancelled
Details
CE Compile Job / print_ce_job_pre_check_outputs (push) Has been cancelled
Details
CE Compile Job / FD-Clone-Linux (push) Has been cancelled
Details
CE Compile Job / Show Code Archive Output (push) Has been cancelled
Details
CE Compile Job / BUILD_SM8090 (push) Has been cancelled
Details
CE Compile Job / BUILD_SM8689 (push) Has been cancelled
Details
CE Compile Job / CE_UPLOAD (push) Has been cancelled
Details

Browse Source 
* [Feature] support rl_tp_degree

* add rl_tp_degree in lmhead

* add rl_tp_degree in bias

* fix split_axis=0 in bias

* fix split_axis in weight

* fix bias rl_tp_degree

* fix bias rl_tp_degree

* change attr to dict

---------

Co-authored-by: Jiang-Jia-Jun <163579578+Jiang-Jia-Jun@users.noreply.github.com>
This commit is contained in:
lizhenyun01
2025-09-08 16:20:32 +08:00
committed by GitHub
parent fa2369271d
commit d40a1046de
3 changed files with 31 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
fastdeploy/model_executor/layers/embeddings.py
View File

@@ -77,6 +77,11 @@ class VocabParallelEmbedding(nn.Layer):
)
if self.world_size > 1:
set_weight_attrs(self.embeddings.weight, {"output_dim": False})
set_weight_attrs(
self.embeddings.weight,
{"rl_need_attr": {"rl_tp_degree": fd_config.parallel_config.tensor_parallel_size}},
)
else:
# column cut embedding
self.embeddings = nn.Embedding(

18
fastdeploy/model_executor/layers/linear.py
View File

@@ -356,11 +356,21 @@ class ColumnParallelLinear(LinearBase):
)
if self.nranks > 0:
_set_var_distributed(self.weight, split_axis=-1)
if self.with_bias:
# col parallel
_set_var_distributed(self.bias, split_axis=1)
_set_var_distributed(self.bias, split_axis=0)
set_weight_attrs(self.bias, {"output_dim": True})
# set_rl_tp_degree
set_weight_attrs(
self.weight, {"rl_need_attr": {"rl_tp_degree": fd_config.parallel_config.tensor_parallel_size}}
)
if self.with_bias:
set_weight_attrs(
self.bias, {"rl_need_attr": {"rl_tp_degree": fd_config.parallel_config.tensor_parallel_size}}
)
class MergedColumnParallelLinear(ColumnParallelLinear):
"""
@@ -743,6 +753,7 @@ class RowParallelLinear(LinearBase):
model_format=fd_config.model_config.model_format,
)
if self.nranks > 0:
_set_var_distributed(self.weight, split_axis=0)
if self.with_bias:
# col parallel
_set_var_distributed(self.bias, split_axis=0)
@@ -755,6 +766,11 @@ class RowParallelLinear(LinearBase):
self.reduce_results = reduce_results
# set_rl_tp_degree
set_weight_attrs(
self.weight, {"rl_need_attr": {"rl_tp_degree": fd_config.parallel_config.tensor_parallel_size}}
)
def forward_cuda(self, x: paddle.Tensor) -> paddle.Tensor:
if self.fd_config.quant_config:
out = self.quant_method.apply(self, x)

9
fastdeploy/model_executor/layers/lm_head.py
View File

@@ -94,6 +94,12 @@ class ParallelLMHead(nn.Layer):
"model_format": self.fd_config.model_config.model_format,
},
)
if self.bias_key is not None:
set_weight_attrs(
self.linear.bias,
{"rl_need_attr": {"rl_tp_degree": fd_config.parallel_config.tensor_parallel_size}},
)
if self.nranks > 1:
set_weight_attrs(self.linear.weight, {"output_dim": True})
else:
@@ -116,6 +122,9 @@ class ParallelLMHead(nn.Layer):
if self.nranks > 1:
set_weight_attrs(self.linear.weight, {"output_dim": False})
set_weight_attrs(
self.linear.weight, {"rl_need_attr": {"rl_tp_degree": fd_config.parallel_config.tensor_parallel_size}}
)
def load_state_dict(self, state_dict: Dict[str, paddle.Tensor | np.ndarray]):
"""