[Fix] Align fused moe lora_b shape with peft (#31534)

Signed-off-by: bk-201 <joy25810@foxmail.com>
This commit is contained in:
B-201
2025-12-31 09:44:59 +08:00
committed by GitHub
parent e1ee11b2a5
commit ecd49ce7e6
3 changed files with 9 additions and 9 deletions
+5 -5
View File
@@ -679,12 +679,12 @@ class FusedMoE3DWithLoRA(FusedMoEWithLoRA):
# (num_experts,rank,input_size)
w13_lora_a = w13_lora_a.reshape(num_experts, -1, w13_lora_a.shape[-1])
w2_lora_a = w2_lora_a.reshape(num_experts, -1, w2_lora_a.shape[-1])
# (output_size,num_experts,rank)
w13_lora_b = w13_lora_b.reshape(w13_lora_b.shape[0], num_experts, -1)
w2_lora_b = w2_lora_b.reshape(w2_lora_b.shape[0], num_experts, -1)
# (output_size,rank,num_experts)
w13_lora_b = w13_lora_b.reshape(w13_lora_b.shape[0], -1, num_experts)
w2_lora_b = w2_lora_b.reshape(w2_lora_b.shape[0], -1, num_experts)
# (num_experts,output_size,rank)
w13_lora_b = w13_lora_b.permute(1, 0, 2)
w2_lora_b = w2_lora_b.permute(1, 0, 2)
w13_lora_b = w13_lora_b.permute(2, 0, 1)
w2_lora_b = w2_lora_b.permute(2, 0, 1)
sliced_w13_lora_a = self._slice_w13_a(w13_lora_a)
sliced_w13_lora_b = self._slice_w13_b(w13_lora_b)