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 5, 03:30 UTC
FIX: support language_model.backbone naming in NemotronH Nano VL quantization config (#39901)
Signed-off-by: <> Co-authored-by: root <root@lyris0144.lyris.clusters.nvidia.com>
danielafrimi committed
Apr 15, 2026 at 16:49 UTC
8b5531933a7b11965b58b21c19ea853487d3f78d
1 file changed
+7
vllm/model_executor/models/nano_nemotron_vl.py
+7
@@ -37,6 +37,7 @@ from vllm.model_executor.models.nemotron_h import NemotronHForCausalLM
37
from vllm.model_executor.models.parakeet import ParakeetExtractor, ProjectedParakeet
38
from vllm.model_executor.models.radio import RadioModel, calc_seq_lens
39
from vllm.model_executor.models.utils import (
40
+ WeightsMapper,
41
init_vllm_registered_model,
42
maybe_prefix,
43
)
@@ -903,6 +904,12 @@ class NemotronH_Nano_VL_V2(
904
requires_sequential_video_encoding = True
905
"""Temporarily needed for dynamic res video w/ conv3d, doesn't support bs>1 yet"""
906
907
+ hf_to_vllm_mapper = WeightsMapper(
908
+ orig_to_new_prefix={
909
+ "language_model.backbone": "language_model.model",
910
+ },
911
+ )
912
+
913
@classmethod
914
def get_placeholder_str(cls, modality: str, i: int) -> str | None:
915
if modality.startswith("image"):