FP8 GRPO running on consumer RTX cards is a gamechanger for anyone doing RL work at home. Getting that down to 5GB of VRAM for Qwen3 means you can actualy experiment without needing datacenter hardware. The 60% VRAM savings make this feasible for way more people to test ideas localy.
Thank you for everything that you bring to the community! Do you have plans to support Nvidia Jetson Thor devices on top of DGX Spark? It would be greatly appreciated 🙏🏻
Both are working yes (same versions as Spark), the difference is that it uses tcgen05 pipeline like a B100, that’s why I was wondering if there is anything else special to add support :)
FP8 GRPO running on consumer RTX cards is a gamechanger for anyone doing RL work at home. Getting that down to 5GB of VRAM for Qwen3 means you can actualy experiment without needing datacenter hardware. The 60% VRAM savings make this feasible for way more people to test ideas localy.
Thank you for everything that you bring to the community! Do you have plans to support Nvidia Jetson Thor devices on top of DGX Spark? It would be greatly appreciated 🙏🏻
Thanks for the support. At the moment we're unsure. If PyTorch and Triton works on it then yes, Unsloth will work on it. :)
Both are working yes (same versions as Spark), the difference is that it uses tcgen05 pipeline like a B100, that’s why I was wondering if there is anything else special to add support :)
Brilliant! Keep up the great work on Unsloth.