NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16

NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16 has a checked-in NeMo AutoModel recipe for image-text-to-text LoRA fine-tuning. The recipe uses the model’s Hugging Face remote implementation and keeps every parallelism axis, including expert parallelism, at one.

Fine-Tune NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16

Follow the installation instructions, then run the recipe on one GB300 GPU from the repository root:

CUDA_VISIBLE_DEVICES=0 automodel \
  examples/vlm_finetune/nemotron_3_5_bolt/nemotron_3_5_bolt_cord_v2_lora.yaml \
  --nproc-per-node 1

Choose a Workflow

Goal	Start Here
LoRA fine-tuning on CORD-v2	Use the single-GPU recipe.

Configuration

Setting	Configuration
Hardware	1 NVIDIA GB300 GPU
Strategy	FSDP2; `dp_size=1`, `tp_size=1`, `pp_size=1`, `cp_size=1`, `ep_size=1`
Sequence length	4,096 tokens
Features	BF16 LoRA, reshard after forward, frozen vision encoder and projector
Advanced	Hugging Face remote implementation, eager attention

Model Reference

Model Architecture

Property	Value
Task	Image-text-to-text
Architecture	`NemotronH_Omni_Reasoning_V3`

Available Models

nvidia/NVIDIA-Nemotron-3.5-Bolt-74B-A9B-BF16

Vision-language model fine-tuning guide