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

[fix] w4a8 model loading and hadamard config (#3013)

Browse Source
This commit is contained in:
Yuan Xiaolan
2025-07-28 18:17:59 +08:00
committed by GitHub
parent c8bf8b3913
commit b1d787a272
4 changed files with 32 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
fastdeploy/model_executor/load_weight_utils.py
View File

@@ -101,12 +101,20 @@ def load_ep_checkpoint(model_path: str, fd_config: FDConfig, return_numpy: bool
up_gate_proj_scale_key = f"ernie.layers.{i}.mlp.experts.{j}.up_gate_proj.weight_scale"
down_proj_scale_key = f"ernie.layers.{i}.mlp.experts.{j}.down_proj.weight_scale"
down_proj_in_scale_key = f"ernie.layers.{i}.mlp.experts.{j}.down_proj.activation_scale"
num_local_ffn_keys.append(up_gate_proj_key)
num_local_ffn_keys.append(down_proj_key)
num_local_ffn_keys.append(up_gate_proj_quant_key)
num_local_ffn_keys.append(down_proj_quant_key)
num_local_ffn_keys.append(up_gate_proj_scale_key)
num_local_ffn_keys.append(down_proj_scale_key)
num_local_ffn_keys.append(down_proj_in_scale_key)
# for EP w4a8, we need all expert's activation_scale for up_gate_proj
for j in range(fd_config.model_config.moe_num_experts):
up_gate_proj_in_scale_key = f"ernie.layers.{i}.mlp.experts.{j}.up_gate_proj.activation_scale"
num_local_ffn_keys.append(up_gate_proj_in_scale_key)
for k in num_local_ffn_keys:
if k in weight_list: