fix deepcopy(tp_group) in spec (#3648)

This commit is contained in:
lzy
2025-08-29 16:08:21 +08:00
committed by GitHub
parent 45f81b34f0
commit 48d760539b
3 changed files with 12 additions and 2 deletions

View File

@@ -58,7 +58,7 @@ class Qwen3MoeBlock(nn.Layer):
self.tp_group = fd_config.parallel_config.tp_group
self.use_ep = self.expert_parallel_size > 1
self.us_tp = self.tensor_parallel_size > 1
self.use_tp = self.tensor_parallel_size > 1
weight_key_map = {
"up_gate_proj_expert_weight_key": f"{prefix}.experts.{{}}.up_gate_proj.weight",