A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Bugfix] Fix broken MTP weight loading for FP8 KV Scales (#27227)
Signed-off-by: Benjamin Chislett <bchislett@nvidia.com>
Benjamin Chislett committed
Oct 21, 2025 at 01:51 UTC
f381cf2302460fb921d67fc1356b9d1b1a93e960
1 file changed
+8
-1
vllm/model_executor/models/deepseek_mtp.py
+8
-1
@@ -16,7 +16,10 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
16
ParallelLMHead,
17
VocabParallelEmbedding,
18
)
19
-from vllm.model_executor.model_loader.weight_utils import default_weight_loader
19
+from vllm.model_executor.model_loader.weight_utils import (
20
+ default_weight_loader,
21
+ maybe_remap_kv_scale_name,
22
+)
23
from vllm.platforms import current_platform
24
from vllm.sequence import IntermediateTensors
25
@@ -278,6 +281,10 @@ class DeepSeekMTP(nn.Module, SupportsPP):
281
if name.endswith(".bias") and name not in params_dict:
282
continue
283
284
+ name = maybe_remap_kv_scale_name(name, params_dict)
285
+ if name is None:
286
+ continue
287
+
288
# According to DeepSeek-V3 Technical Report, MTP modules
289
# shares embedding layer. We only load the first weights.
290
if (