integrations.kernels.plugin
integrations.kernels.plugin
Classes
| Name | Description |
|---|---|
| KernelsPlugin | Thin orchestrator: registers the expert-kernel backend and dispatches model-family |
KernelsPlugin
integrations.kernels.plugin.KernelsPlugin()Thin orchestrator: registers the expert-kernel backend and dispatches model-family
specifics to ModelAdapter subclasses (see adapters/).
Methods
| Name | Description |
|---|---|
| pre_model_load | Register the expert-kernel backend + generic capabilities, then run adapter hooks. |
pre_model_load
integrations.kernels.plugin.KernelsPlugin.pre_model_load(cfg)Register the expert-kernel backend + generic capabilities, then run adapter hooks.
Architecture-agnostic: routing stays in each model’s SparseMoEBlock; only the experts
call is dispatched through the registry. When EP is active the ExpertParallelPlugin owns
experts_implementation (a deep_ep_* composite), so we don’t overwrite it here.