ZenDNN Release v6.1.0

ZenDNN 6.1.0

ZenDNN 6.1.0 builds on the 6.0.0 LowOHA inference stack. It makes group-matmul custom kernels faster and adds FP16 and DA8W4 kernels, adds matmul custom kernels for DA8W4, DA8W8, INT8, and BF16, and adds GQA and sliding-window flash attention, a routed-MoE API, and in-place weight caching by default.

Highlights

Area	What's New
MoE / group matmul	Faster group-matmul custom kernels, and new FP16 and DA8W4 custom kernels. New routed-MoE API with pre-quantized INT8 sources and in-place caching.
Matmul	New custom kernels for DA8W4, DA8W8, INT8, and BF16 to optimize LLM workloads. The kernel is chosen dynamically from the problem shape.
SDPA	GQA and sliding-window flash attention. Dynamic INT8 quantization for BF16 flash attention, with INT8 matrix multiplies and FP32 softmax and output.
Quantization	Symmetric INT8 prepack, DA8W4, per-channel grouped dynamic quantization, and a direct per-token dynamic-quant kernel.
Caching	In-place weight caching is the default. Public APIs clear matmul and custom-kernel weight caches. Caches can be turned off globally.
Normalization	Fused layer-norm-add. Fused RMS norm with FP32 accumulation.
BenchDNN	Sweep mode, Qwen3.6 and Gemma-4 MoE fused inputs, and dynamic INT8 workloads.
Build and platform	GCC >= 11.2.0, GCC < 15.2.0, and Clang 14 are supported.