GDN-HIP (Gated Delta Net, RDNA4/gfx1201)
Native HIP + rocWMMA kernels for the Gated Delta Net linear-attention path: gdn_prefill/decode (+wmma/chunked), causal_conv1d, gated RMSNorm, with native backward + differentiable training wrappers.
Built with kernel-builder for AMD RDNA4 (gfx1201).
Load with the kernels library:
from kernels import get_kernel
kernel = get_kernel("pat883/gdn-hip")
Requires a ROCm PyTorch build (torch 2.10 / ROCm 7.x) on an RDNA4 card. Built variants:
torch210-cxx11-rocm70andtorch210-cxx11-rocm71.
- Downloads last month
- 20
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support