jburtoft/qwen35-deltanet-neuron-kernels

⚠️ DEPRECATED AND EMPTIED — use jburtoft/qwen35-deltanet-tkg-full

This repository is deprecated and no longer contains any kernel code.

All Qwen3.5 Gated DeltaNet NKI kernels — inference and training — live at:

→ jburtoft/qwen35-deltanet-tkg-full

Every branch and every tag in this repository has been emptied of kernel code, including the historical tags v2.0-task007, v3.0-training, and v3.1-batched. Pinning to an old revision here will no longer give you a working kernel — this is deliberate, so that no one can accidentally load stale code from a retired location.

Migration

from kernels import get_kernel

# OLD -- no longer functional at any revision
# k = get_kernel("jburtoft/qwen35-deltanet-neuron-kernels", revision="v3.1-batched")

# NEW
k = get_kernel("jburtoft/qwen35-deltanet-tkg-full")

If you used the HF KernelConfig module-replacement form:

# OLD
# KernelConfig({"Qwen3_5GatedDeltaNet": "jburtoft/qwen35-deltanet-neuron-kernels:NeuronGatedDeltaNet"})

# NEW
KernelConfig({"Qwen3_5GatedDeltaNet": ("jburtoft/qwen35-deltanet-tkg-full", "NeuronGatedDeltaNet")})

Import names are unchanged at the new location.

Why this repo was retired

tkg-full supersedes it on every axis:

• Decode actually works. This repo's kernels ignored cache_params, so autoregressive decode was broken. tkg-full threads recurrent state correctly through decode.
• Chunked prefill. This repo used a per-token serial recurrence with a Python loop over (batch, head). tkg-full uses a chunked-parallel kernel batched over all slices in one launch.
• Fused elementwise work. The gated RMS norm and gating/normalization chains are computed inside the kernels rather than as separate PyTorch ops.
• A torch.compile regression at B >= 2 present in this repo's v3.1-batched is fixed.

License

Apache 2.0.