[Bugfix][Model] Fix GraniteMoeHybrid per-expert quantized weight loading (#54052)

This commit is contained in:
Priyansh Jain
2026-09-01 12:38:53 +00:00
committed by GitHub
parent 40824284bc
commit 55aa766dc8
@@ -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):