Add GPTQ support (#916)
This commit is contained in:
@@ -153,7 +153,7 @@ class MixtralMoE(nn.Module):
|
||||
self.gate = ReplicatedLinear(config.hidden_size,
|
||||
self.num_total_experts,
|
||||
bias=False,
|
||||
linear_method=linear_method)
|
||||
linear_method=None)
|
||||
|
||||
def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
|
||||
batch_size, sequence_length, hidden_dim = hidden_states.shape
|
||||
@@ -418,11 +418,18 @@ class MixtralForCausalLM(nn.Module):
|
||||
for (param_name, weight_name, shard_id) in stacked_params_mapping:
|
||||
if weight_name not in name:
|
||||
continue
|
||||
param = params_dict[name.replace(weight_name, param_name)]
|
||||
name = name.replace(weight_name, param_name)
|
||||
# Skip loading extra bias for GPTQ models.
|
||||
if name.endswith(".bias") and name not in params_dict:
|
||||
continue
|
||||
param = params_dict[name]
|
||||
weight_loader = param.weight_loader
|
||||
weight_loader(param, loaded_weight, shard_id)
|
||||
break
|
||||
else:
|
||||
# Skip loading extra bias for GPTQ models.
|
||||
if name.endswith(".bias") and name not in params_dict:
|
||||
continue
|
||||
param = params_dict[name]
|
||||
weight_loader = getattr(param, "weight_loader",
|
||||
default_weight_loader)
|
||||
|
||||
Reference in New Issue
Block a user