Collected sources and patterns will appear here. Add from search, explore, or the patterns library.
Tensor<FP16> -> Tuple<Tensor<INT4>, Tensor<FP16>>
Quantize activation tensors at a per-thread granularity to maximize dynamic range representation without introducing global memory access overhead.
Problem it solves
Coarse-grained quantization (e.g., per-tensor or per-channel) degrades precision, while extremely fine-grained quantization typically incurs high scale-factor memory overhead.
Consumes
Emits
The real projects this mechanism was found in. Attribution is the point — this is how the best teams actually do it.