apps/FastDeploy
1
0
Fork 0
mirror of https://github.com/PaddlePaddle/FastDeploy.git synced 2026-04-23 08:21:53 +08:00
Code Issues Actions 19 Packages Projects Releases Wiki Activity

[Feature] Support redundant expert for eplb (#5918)

Browse Source 
* [BugFix] support redundant expert for eplb

* support redundant expert for eplb

* support redundant expert for eplb

* update

* fix ci eplb
This commit is contained in:
xiaoxiaohehe001
2026-01-09 17:13:24 +08:00
committed by GitHub
parent e6cdea4492
commit 00a01ae024
11 changed files with 36 additions and 17 deletions
Download Patch File Download Diff File Expand all files Collapse all files
fastdeploy/model_executor/load_weight_utils.py
+5
View File
@@ -252,8 +252,13 @@ def load_ep_checkpoint(cls: PretrainedModel, model_path: str, fd_config: FDConfi
"mtp_block" if getattr(fd_config.speculative_config, "model_type", "main") == "mtp" else "layers"
)
moe_num_experts = fd_config.model_config.moe_num_experts
if isinstance(moe_num_experts, list):
moe_num_experts = moe_num_experts[0]
for i in range(fd_config.model_config.moe_layer_start_index, fd_config.model_config.num_hidden_layers):
for j in get_expert_ranges(fd_config):
# Map redundant expert IDs back to actual expert IDs for weight loading
j = j % moe_num_experts
up_gate_proj_key = f"ernie.{prefix_layer_name}.{i}.mlp.experts.{j}.up_gate_proj.weight"
down_proj_key = f"ernie.{prefix_layer_name}.{i}.mlp.experts.{j}.down_proj.weight"