apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2026-04-23 00:17:25 +08:00
Code Issues Actions 21 Packages Projects Releases Wiki Activity

[Intel HPU] enable MoE EP for hpu (#5855)

Browse Source 
* enable HPU MoE EP

* MoE intermediate_scale stack

* enable loader_v1 esp for tensor_wise_fp8 TP or EP

* modify activation_scale name
This commit is contained in:
Cheng Yanfei
2026-01-15 13:08:00 +08:00
committed by GitHub
parent 7c56041272
commit fbcccaa750
9 changed files with 177 additions and 11 deletions
Download Patch File Download Diff File Expand all files Collapse all files
fastdeploy/model_executor/load_weight_utils.py
+3
View File
@@ -278,6 +278,8 @@ def load_ep_checkpoint(cls: PretrainedModel, model_path: str, fd_config: FDConfi
down_proj_scale_key = f"ernie.{prefix_layer_name}.{i}.mlp.experts.{j}.down_proj.weight_scale"
down_proj_in_scale_key = f"ernie.{prefix_layer_name}.{i}.mlp.experts.{j}.down_proj.activation_scale"
# single up_gate_proj.activation_scale for all mlp.experts
up_gate_proj_in_scale_key = f"ernie.layers.{i}.mlp.experts.up_gate_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)
@@ -285,6 +287,7 @@ def load_ep_checkpoint(cls: PretrainedModel, model_path: str, fd_config: FDConfi
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)
num_local_ffn_keys.append(up_gate_proj_in_scale_key)
# for EP w4a8, we need all expert's activation_scale for up_gate_proj
num_experts = fd_config.model_config.moe_num_experts