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 10, 06:30 UTC
[BUGFIX] Add missed remaping of the names of fp8 kv-scale (#32199)
Signed-off-by: Vadim Gimpelson <vadim.gimpelson@gmail.com>
Vadim Gimpelson committed
Jan 13, 2026 at 00:42 UTC
9f430c94bd0b5cf5a697ddeefd4507ae078bb0ed
1 file changed
+7
vllm/model_executor/models/qwen3_next.py
+7
@@ -64,6 +64,7 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
64
)
65
from vllm.model_executor.model_loader.weight_utils import (
66
default_weight_loader,
67
+ maybe_remap_kv_scale_name,
68
sharded_weight_loader,
69
)
70
from vllm.model_executor.models.qwen2_moe import Qwen2MoeMLP as Qwen3NextMLP
@@ -1065,6 +1066,12 @@ class Qwen3NextModel(nn.Module):
1066
if name.startswith("mtp."):
1067
continue
1068
1069
+ # Remapping the name of FP8 kv-scale.
1070
+ if name.endswith("scale"):
1071
+ name = maybe_remap_kv_scale_name(name, params_dict)
1072
+ if name is None:
1073
+ continue
1074
+
1075
for param_name, weight_name, shard_id in stacked_params_mapping:
1076
if weight_name not in name:
1077
continue