Fix w8a8_int8 model shared experts fusion load weights error (#5120)
This commit is contained in:
@@ -1480,14 +1480,24 @@ class DeepseekV2ForCausalLM(nn.Module):
|
|||||||
if self.n_share_experts_fusion is not None and self.n_share_experts_fusion > 0:
|
if self.n_share_experts_fusion is not None and self.n_share_experts_fusion > 0:
|
||||||
weights_list = list(weights)
|
weights_list = list(weights)
|
||||||
weights_dict = dict(weights_list)
|
weights_dict = dict(weights_list)
|
||||||
suffix_list = [
|
if self.quant_config.get_name() == "w8a8_int8":
|
||||||
"down_proj.weight",
|
suffix_list = [
|
||||||
"down_proj.weight_scale_inv",
|
"down_proj.weight",
|
||||||
"gate_proj.weight",
|
"down_proj.weight_scale",
|
||||||
"gate_proj.weight_scale_inv",
|
"gate_proj.weight",
|
||||||
"up_proj.weight",
|
"gate_proj.weight_scale",
|
||||||
"up_proj.weight_scale_inv",
|
"up_proj.weight",
|
||||||
]
|
"up_proj.weight_scale",
|
||||||
|
]
|
||||||
|
else:
|
||||||
|
suffix_list = [
|
||||||
|
"down_proj.weight",
|
||||||
|
"down_proj.weight_scale_inv",
|
||||||
|
"gate_proj.weight",
|
||||||
|
"gate_proj.weight_scale_inv",
|
||||||
|
"up_proj.weight",
|
||||||
|
"up_proj.weight_scale_inv",
|
||||||
|
]
|
||||||
names_to_remove = []
|
names_to_remove = []
|
||||||
for moe_layer in tqdm(
|
for moe_layer in tqdm(
|
||||||
range(
|
range(
|
||||||
|
|||||||
Reference in New Issue
Block a user