From 55aa766dc8c1a7d739ef90c5580ca4b91050b35f Mon Sep 17 00:00:00 2001 From: Priyansh Jain <167848587+Priyjain-amd@users.noreply.github.com> Date: Tue, 1 Sep 2026 18:08:53 +0530 Subject: [PATCH] [Bugfix][Model] Fix GraniteMoeHybrid per-expert quantized weight loading (#54052) --- vllm/model_executor/models/granitemoehybrid.py | 8 +++++++- 1 file changed, 7 insertions(+), 1 deletion(-) diff --git a/vllm/model_executor/models/granitemoehybrid.py b/vllm/model_executor/models/granitemoehybrid.py index 810770aab4..8dc47c5226 100644 --- a/vllm/model_executor/models/granitemoehybrid.py +++ b/vllm/model_executor/models/granitemoehybrid.py @@ -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):