Developer Ports PyTorch 2.7.1 and Mamba-1 to Legacy NVIDIA GPUs
A developer announced a Windows port of an LLM training stack for NVIDIA GPUs from SM 3.5 through SM 7.5, including a custom PyTorch 2.7.1 build with CUDA 11.8, ported causal-conv1d, native-CUDA Mamba-1 and Windows NCCL. The author reports that Mamba-1 works across those architectures and achieved a maximum absolute error of 3.8e-06 against a PyTorch reference on a 2013 Tesla K40c. Mamba-2 and Mamba-3 require Turing or newer because they depend on Triton and TileLang, while older GPUs automatically fall back to Mamba-1.
The release provides a way to run modern PyTorch and Mamba tooling on older NVIDIA hardware that many current AI stacks no longer support, although its results remain self-reported and community testing is limited.
Hello!
I want to share an active development project. If you have a drawer full of old NVIDIA GPU's collecting dust because modern AI stacks refuse to compile for them, this is for you.
I have ported the modern LLM training stack on Windows to legacy CUDA GPUs (sm_35 through sm_75):
- PyTorch (v2.7.1) (custom build compiled with CUDA 11.8, keeping the legacy SASS) ([https://github.com/theIvanR/torch-on-clunkers])
- causal-conv1d (v1.7.0) (https://github.com/theIvanR/conv-on-clunkers)
- mamba_ssm (Mamba-1) with pure native CUDA kernels (https://github.com/theIvanR/mamba-on-clunkers)
- Windows NCCL (DDP) for multi-GPU training across Kepler cards (https://github.com/theIvanR/nccl-on-windows)
Status
- Mamba-1 (Pure CUDA): Fully functional with optimized execution across legacy GPU architectures (SM 3.5–SM 7.5). Tested on a 2013 NVIDIA Tesla K40c, achieving a maximum absolute error of
3.8e-06against the PyTorch reference implementation. - Mamba-2 & Mamba-3: Currently supported with optimized execution on Turing (SM 7.5) and newer architectures, due to their reliance on Triton and TileLang.
- Graceful Fallback: On older hardware, the package detects unavailable Triton-dependent modules and automatically falls back to the pure-CUDA Mamba-1 implementation.
#Mamba on Clunkers#NVIDIA#PyTorch#Mamba-1#Mamba-2#Mamba-3#causal-conv1d#NCCL#Triton#TileLang#Tesla K40c