NVIDIA DGX Station GB300 for Fine-Tuning: LoRA, QLoRA and Enterprise Model Customization

On one DGX Station GB300, LoRA fine-tuning of a 70B model fits comfortably in the 252 GB of HBM3e, QLoRA reaches up to about 405B at the very edge of that memory, and a full fine-tune fits only for models up to roughly 8–15B before you start offloading to the slower LPDDR5X. Those are estimates from Hugging Face and Unsloth, before long-sequence activations. Two tools are documented on the Station itself: Unsloth published an official DGX Station guide, and a Cornell research group ran TRL on one. Axolotl and NVIDIA NeMo may work but have no Station-specific documentation. The pipeline that follows takes you from data preparation through training and evaluation to serving the adapter in vLLM. SourceIT has not run fine-tuning jobs on a Station; the figures and steps come from the named publishers.

What fits in the GB300's memory for each method?

Source: Hugging Face Llama 3.1 blog (23 Jul 2024, "estimated values and may vary"); Unsloth requirements page ("the absolute minimum", accessed 9 Nov 2026); ZeRO paper for the 16 bytes per parameter Adam accounting. CALC = our arithmetic. Figures exclude activations for long sequences, so real jobs need headroom. SourceIT has not run these jobs.

Which fine-tuning tools are documented on the Station?

Unsloth has an official guide, "Fine-Tuning LLMs on NVIDIA DGX Station with Unsloth", written on early-access hardware. It reports that fine-tuning gpt-oss-120b used "around 72 GB of GPU memory" and took "about 10 minutes" on its example notebook. Its install steps start like this:

sudo apt install python3.12-dev
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130
pip install unsloth

Unsloth warns that installing PyTorch without the CUDA 13 index can give you a CPU-only build. One detail needs care: its optional flash-attention build sets TORCH_CUDA_ARCH_LIST="10.0", while NVIDIA lists the Station GB300 as compute capability 10.3. Arch-specific builds may need the 10.3 target, so test before relying on them.

TRL (Hugging Face's training library) was used by Cornell's Kilian group on a GB300 Station for GRPO reinforcement-learning runs on Qwen3 models from 0.6B to 7B. They ran Qwen3-4B at batch size 32 with 16 rollouts, against batch 8 and 4 rollouts on four H100s with DeepSpeed offload, and estimated about three times fewer GPU-hours for Qwen3-0.6B. They also wrote that wheels were only "just shipping" in February 2026 and recommended NGC containers.

Axolotl documents CUDA 13 and PyTorch 2.11 or later for Blackwell, and notes that CUDA 12.8 "cannot compile for sm_103a", but doesn't mention aarch64. NVIDIA NeMo has a DGX Spark playbook but none for the Station. Both are unverified here. On any tool, PyTorch 2.11.0 and later ship CUDA wheels for aarch64 Linux by default, which removes one common source of CPU-only installs.

Can the 748 GB of coherent memory help with bigger jobs?

Only slowly. The 496 GB of LPDDR5X is reached over NVLink-C2C, which StorageReview measured at about 390 GB/s to HBM, against 7.1 TB/s for the HBM itself. A full fine-tune of a 30B-class model (about 512 GB by our calculation) would need optimizer and parameter offload to the Grace memory, for example with DeepSpeed's CPU offload, and would run at the speed of that link. A 70B full fine-tune at roughly 1.1 TB exceeds even 748 GB without offloading to NVMe. No published run of either on a Station was found, so treat them as experiments, not plans. NVIDIA's "up to 1 trillion parameters" claim for the Station refers to running models, not training them.

Rule of thumb for one GB300: LoRA or QLoRA for anything 32B and up, full fine-tuning only for small models, and a pilot run before you promise a timeline.

What does a practical fine-tuning pipeline look like?

First, decide whether fine-tuning is the right answer. If the goal is to answer questions from company documents, retrieval usually comes first; our RAG guide covers that route. Fine-tuning suits changing a model's style, format or behaviour on a narrow task. Our GB10 fine-tuning post explains the difference between LoRA, QLoRA and full fine-tuning in more detail.

1. Prepare the data. Convert examples into the chat or instruction format your base model expects, remove personal data you don't need, and hold back a test set the model never trains on.
2. Measure the baseline. Run the untouched base model on the test set and record the scores. Without this, nobody can tell whether training helped.
3. Train small first. Start with a short LoRA run on a sample, check the loss falls and outputs improve, then scale up.
4. Evaluate before and after. Compare against the baseline on the held-out set, and have a domain expert review a sample of answers.
5. Version the adapters. Name each adapter with the base model, dataset version and date, store it on the NAS with its training settings, and keep the evaluation results beside it.

Checkpoints are the storage surprise. A resumable full-fine-tune checkpoint is roughly 12–16 bytes per parameter by a common rule of thumb, so plan where they go using our storage planning worksheet.

How do you serve the result and share the GPU?

vLLM can serve LoRA adapters on top of a base model, so several fine-tuned variants can share one copy of the base weights; check the vLLM LoRA documentation for the version you deploy, and see our vLLM deployment guide for the serving setup on the Station. Training and serving compete for the same 252 GB of HBM, so most teams run training at night or at weekends, or reserve a MIG slice. The GB300 supports up to seven MIG instances, with profiles such as 3g.126gb, but MIG layouts don't survive a reboot and MIG splits only the HBM, not the LPDDR5X or CPU. For overnight batch work, our 24/7 AI factory post shows how to plan the schedule, and what you can run on a DGX Station covers the full-precision arithmetic for 70B.

Buying in Singapore

SourceIT lists DGX Station GB300 systems from six OEMs, supplied through each OEM's channel: the ASUS ExpertCenter Pro ET900N G3, Dell Pro Max with GB300, HP ZGX Fury G1n, MSI XpertStation WS300, GIGABYTE W775-V10-L01 and Supermicro ARS-511GD-NB-LCC. All are price on request; that is a supplier condition rather than a SourceIT choice. Tell us the model sizes and methods you plan to use, and we can help check the configuration and storage before you order. Use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.

Frequently asked questions

Can a DGX Station GB300 fine-tune a 70B model? With LoRA or QLoRA, yes: published estimates are about 160 GB for LoRA and 41–48 GB for QLoRA, within the 252 GB of HBM3e. A full 70B fine-tune needs about 500 GB or more and does not fit in HBM.

Which fine-tuning tool works on the Station? Unsloth has an official DGX Station guide, and TRL has been used on one by a Cornell group. Axolotl and NeMo have no Station-specific documentation yet.

How long does a fine-tuning job take? It depends on model, data and method. Unsloth reports about 10 minutes for its gpt-oss-120b example notebook; no published LoRA throughput figures for the Station exist yet.

Can I fine-tune and serve users at the same time? Both compete for the same HBM. Most teams schedule training out of hours or reserve a MIG slice, remembering that MIG layouts reset at reboot.

Do I need the LPDDR5X for fine-tuning? Not for LoRA or QLoRA on most models. It can hold offloaded optimizer state for larger full fine-tunes, but at a fraction of HBM bandwidth.