[Bugfix] Fix broken MTP weight loading for FP8 KV Scales (#27227)

Signed-off-by: Benjamin Chislett <bchislett@nvidia.com>

vllm/model_executor/models/deepseek_mtp.py

@@ -16,7 +16,10 @@ 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_kv_scale_name,
+)
 from vllm.platforms import current_platform
 from vllm.sequence import IntermediateTensors
 
@@ -278,6 +281,10 @@ class DeepSeekMTP(nn.Module, SupportsPP):
                     if name.endswith(".bias") and name not in params_dict:
                         continue
 
+                    name = maybe_remap_kv_scale_name(name, params_dict)
+                    if name is None:
+                        continue
+
                     # According to DeepSeek-V3 Technical Report, MTP modules
                     # shares embedding layer. We only load the first weights.
                     if (