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

[Loader V1] modify layername for DeepSeekV3 (#3336)

Browse Source 
Co-authored-by: Yuanle Liu <yuanlehome@163.com>
Co-authored-by: YUNSHEN XIE <1084314248@qq.com>
This commit is contained in:
Zero Rains
2025-08-13 15:47:06 +08:00
committed by GitHub
parent f702a675a1
commit be94bdd0b0
1 changed files with 4 additions and 5 deletions
Download Patch File Download Diff File Expand all files Collapse all files

9
fastdeploy/model_executor/models/deepseek_v3.py
View File

@@ -539,7 +539,7 @@ class DeepSeekV3Model(nn.Layer):
prefix="deepseek_v3.embed_tokens",
)
self.decoder_layers = nn.LayerList(
self.layers = nn.LayerList(
[
DeepSeekV3DecoderLayer(
fd_config,
@@ -564,7 +564,7 @@ class DeepSeekV3Model(nn.Layer):
self.norm.load_state_dict(state_dict)
for i in range(self.num_layers):
logger.info(f"Start load layer {i}")
self.decoder_layers[i].load_state_dict(state_dict)
self.layers[i].load_state_dict(state_dict)
def forward(
self,
@@ -578,7 +578,7 @@ class DeepSeekV3Model(nn.Layer):
residual = None
for i in range(self.num_layers):
hidden_states, residual = self.decoder_layers[i](
hidden_states, residual = self.layers[i](
forward_meta,
hidden_states,
residual,
@@ -658,12 +658,11 @@ class DeepseekV3ForCausalLM(ModelForCasualLM):
for loaded_weight_name, loaded_weight in weights_iterator:
loaded_weight_name = loaded_weight_name.replace("deepseek_v3", "model")
loaded_weight_name = loaded_weight_name.replace("layers", "decoder_layers")
for param_name, weight_name, shard_id in stacked_params_mapping:
if weight_name not in loaded_weight_name:
continue
if "mlp.experts." in loaded_weight_name and loaded_weight_name not in params_dict:
if "mlp.experts." in loaded_weight_name:
continue
model_param_name = loaded_weight_name.replace(weight_name, param_name)