Collected sources and patterns will appear here. Add from search or the patterns library.
ActivationTensor -> QuantizedActivationTensor
Dynamically compute scale factors and quantize activation tensors on a per-token basis during the forward pass.
Problem it solves
Outlier activations in large language models cause severe degradation in accuracy when using static quantization strategies.
Consumes
Emits
The real projects this mechanism was found in. Attribution is the point — this is how the best teams actually do it.