Enable modelopt gemma3 nvfp4/fp8, make workflow more robust (#22771)
Signed-off-by: Zhiyu Cheng <zhiyuc@nvidia.com> Signed-off-by: Michael Goin <mgoin64@gmail.com> Co-authored-by: Michael Goin <mgoin64@gmail.com>
This commit is contained in:
@@ -446,6 +446,22 @@ class Gemma3Model(nn.Module):
|
||||
weight_loader(param, loaded_weight)
|
||||
loaded_params.add(scale_name)
|
||||
continue
|
||||
|
||||
# Check if this is a scale parameter that needs remapping first
|
||||
if name.endswith(
|
||||
(".k_scale", ".v_scale", ".q_scale", ".prob_scale")):
|
||||
# Try to remap the scale name first
|
||||
remapped_name = maybe_remap_kv_scale_name(name, params_dict)
|
||||
if remapped_name is not None and remapped_name in params_dict:
|
||||
# Successfully remapped, use the remapped name
|
||||
param = params_dict[remapped_name]
|
||||
weight_loader = getattr(param, "weight_loader",
|
||||
default_weight_loader)
|
||||
weight_loader(param, loaded_weight)
|
||||
loaded_params.add(remapped_name)
|
||||
continue
|
||||
# If remapping failed, continue with normal processing
|
||||
|
||||
for (param_name, shard_name, shard_id) in stacked_params_mapping:
|
||||
if shard_name not in name:
|
||||
continue
|
||||
|
||||
Reference in New Issue
Block a user