Unsloth Adds AMD Support: Train LLMs on 3 GB VRAM GPUs
Unsloth now supports AMD GPUs with custom ROCm‑optimized Triton kernels, delivering up to double the training speed and 70% lower memory usage, enabling over 500 LLMs to be trained on as little as 3 GB VRAM and providing detailed performance benchmarks on MI300X.
Unsloth announced official support for AMD GPUs, allowing large language model (LLM) training and inference on Radeon RX 6000/7000 series and integrated graphics.
Technical integration
The collaboration with AMD adapted Unsloth's custom Triton kernels and math algorithms to the ROCm stack, resulting in a two‑fold increase in training speed and a 70% reduction in VRAM consumption without sacrificing accuracy.
What can be done?
Train and run more than 500 models, including Qwen3.6, Gemma 4, DeepSeek‑V4, GLM‑5.2, and other recent releases.
Support for Windows, WSL, and Linux across Radeon, Instinct, Ryzen, and data‑center GPUs.
Only 3 GB VRAM is enough to fine‑tune Qwen3.5 or Gemma 4 (8 GB can handle 12B‑parameter models).
Reinforcement learning (GRPO) also runs, cutting VRAM usage by an additional 80%.
Installation
Installation is a single command: curl -fsSL https://unsloth.ai/install.sh | sh Windows users can use PowerShell: irm https://unsloth.ai/install.ps1 | iex The script automatically detects the AMD GPU and installs the appropriate ROCm‑enabled PyTorch, bitsandbytes, and llama.cpp pre‑compiled packages.
Performance results
On an MI300X, Unsloth trains Llama‑3.1‑8B (LoRA SFT) in 2.07 s/step, compared with 2.87 s/step for TRL + Flash Attention 2, a 1.39× speedup. Peak VRAM usage drops from 24.3 GB to 18.3 GB, a 1.33× reduction.
Inference also benefits: Unsloth adds tool‑call healing (boosting accuracy by 50%), code execution, web search, and Cloudflare HTTPS‑tunnel remote access. Training logs can be viewed on a phone, and the LLM can generate HTML games that run directly in mobile browsers.
Observations
Community feedback notes that the 3 GB VRAM requirement dramatically lowers the barrier for users with older RX 6600‑class cards, though multi‑GPU training on AMD is currently limited to Linux because ROCm lacks a collective communication backend for Windows, and GPUs older than gfx1030 have only partial support.
Unsloth also provides a free AMD developer notebook on an MI300X (192 GB VRAM) without registration.
More details are available in the official documentation:
Unsloth for AMD guide (https://unsloth.ai/docs/basics/amd)
AMD official blog (https://www.amd.com/en/developer/resources/technical-articles/2026/train-and-run-models-on-amd-gpus-with-unsloth.html)
Signed-in readers can open the original source through BestHub's protected redirect.
This article has been distilled and summarized from source material, then republished for learning and reference. If you believe it infringes your rights, please contactand we will review it promptly.
AI Engineering
Focused on cutting‑edge product and technology information and practical experience sharing in the AI field (large models, MLOps/LLMOps, AI application development, AI infrastructure).
How this landed with the community
Was this worth your time?
0 Comments
Thoughtful readers leave field notes, pushback, and hard-won operational detail here.