[Bugfix][Model] Fix GraniteMoeHybrid per-expert quantized weight loading (#54052)
Priyansh Jain committed
Sep 1, 2026 at 18:08 UTC
55aa766dc8c1a7d739ef90c5580ca4b91050b35f
1 file changed
+7
-1
vllm/model_executor/models/granitemoehybrid.py
+7
-1
@@ -30,7 +30,10 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
30
ParallelLMHead,
31
VocabParallelEmbedding,
32
)
33
-from vllm.model_executor.model_loader.weight_utils import default_weight_loader
33
+from vllm.model_executor.model_loader.weight_utils import (
34
+ default_weight_loader,
35
+ maybe_remap_moe_expert_param_name,
36
+)
37
from vllm.sequence import IntermediateTensors
38
39
from .granitemoe import GraniteMoeMoE
@@ -479,6 +482,9 @@ class GraniteMoeHybridModel(nn.Module):
482
continue
483
484
name_mapped = name.replace(weight_name, param_name)
485
+ name_mapped = maybe_remap_moe_expert_param_name(
486
+ name_mapped, params_dict
487
+ )
488
489
# Skip layers on other devices.
490
if is_pp_missing_parameter(name_mapped, self):