From f381cf2302460fb921d67fc1356b9d1b1a93e960 Mon Sep 17 00:00:00 2001 From: Benjamin Chislett Date: Tue, 21 Oct 2025 01:51:44 -0400 Subject: [PATCH] [Bugfix] Fix broken MTP weight loading for FP8 KV Scales (#27227) Signed-off-by: Benjamin Chislett --- vllm/model_executor/models/deepseek_mtp.py | 9 ++++++++- 1 file changed, 8 insertions(+), 1 deletion(-) diff --git a/vllm/model_executor/models/deepseek_mtp.py b/vllm/model_executor/models/deepseek_mtp.py index 576977b00e..aa176ef05f 100644 --- a/vllm/model_executor/models/deepseek_mtp.py +++ b/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 (