|
Mila
Deep Neural Network Library
|
Configuration for mixture-of-experts routing: how many experts, and how many run per token. More...
#include <stdexcept>#include <string>#include <sstream>#include <format>#include <utility>#include <cstdint>import Serialization.Metadata;import Dnn.ComponentConfig;import Dnn.TensorTypes;Classes | |
| class | Mila::Dnn::RouterConfig |
Namespaces | |
| namespace | Mila |
| Mila main API namespace. | |
Configuration for mixture-of-experts routing: how many experts, and how many run per token.
Gemma 4 26B-A4B routes each 2816-wide token to 8 of 128 experts. See Specifications/Gemma4MoE.md Phase 5.