[Bugfix][Model] Fix GraniteMoeHybrid per-expert quantized weight loading (#54052)
This commit is contained in:
@@ -30,7 +30,10 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
|
||||
ParallelLMHead,
|
||||
VocabParallelEmbedding,
|
||||
)
|
||||
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
|
||||
from vllm.model_executor.model_loader.weight_utils import (
|
||||
default_weight_loader,
|
||||
maybe_remap_moe_expert_param_name,
|
||||
)
|
||||
from vllm.sequence import IntermediateTensors
|
||||
|
||||
from .granitemoe import GraniteMoeMoE
|
||||
@@ -479,6 +482,9 @@ class GraniteMoeHybridModel(nn.Module):
|
||||
continue
|
||||
|
||||
name_mapped = name.replace(weight_name, param_name)
|
||||
name_mapped = maybe_remap_moe_expert_param_name(
|
||||
name_mapped, params_dict
|
||||
)
|
||||
|
||||
# Skip layers on other devices.
|
||||
if is_pp_missing_parameter(name_mapped, self):
|
||||
|
||||
Reference in New Issue
Block a user