[Bugfix] Fix GLM4-MoE weight loading for NVFP4 quantized checkpoints (#41755)
Signed-off-by: Stoyan Yanev <stoyan.yanev@cleverpine.com>
This commit is contained in:
@@ -506,16 +506,24 @@ class Glm4MoeModel(nn.Module):
|
|||||||
# for mlp.experts[0].gate_gate_up_proj, which breaks load.
|
# for mlp.experts[0].gate_gate_up_proj, which breaks load.
|
||||||
if ("mlp.experts." in name) and name not in params_dict:
|
if ("mlp.experts." in name) and name not in params_dict:
|
||||||
continue
|
continue
|
||||||
|
|
||||||
name = name.replace(weight_name, param_name)
|
name = name.replace(weight_name, param_name)
|
||||||
# Skip loading extra bias for GPTQ models.
|
# Skip loading extra bias for GPTQ models.
|
||||||
if name.endswith(".bias") and name not in params_dict:
|
if name.endswith(".bias") and name not in params_dict:
|
||||||
continue
|
continue
|
||||||
|
|
||||||
|
name = maybe_remap_kv_scale_name(name, params_dict)
|
||||||
|
if name is None:
|
||||||
|
continue
|
||||||
if is_pp_missing_parameter(name, self):
|
if is_pp_missing_parameter(name, self):
|
||||||
continue
|
continue
|
||||||
|
|
||||||
param = params_dict[name]
|
param = params_dict[name]
|
||||||
weight_loader = param.weight_loader
|
weight_loader = getattr(param, "weight_loader", default_weight_loader)
|
||||||
weight_loader(param, loaded_weight, shard_id)
|
if weight_loader == default_weight_loader:
|
||||||
|
weight_loader(param, loaded_weight)
|
||||||
|
else:
|
||||||
|
weight_loader(param, loaded_weight, shard_id)
|
||||||
break
|
break
|
||||||
else:
|
else:
|
||||||
is_expert_weight = False
|
is_expert_weight = False
|
||||||
|
|||||||
Reference in New Issue
Block a user