Module infer 

TinyTimeMixer inference engine.

Architecture: MLP-Mixer with adaptive patching.

• StdScaler → Patchify → Linear patcher → 3 adaptive levels → Linear adapter → 2 decoder layers → Flatten → Linear head → inverse scale
• No attention. Each “mixer layer” = PatchMixerBlock + FeatureMixerBlock.
• GatedAttention: softmax(linear(x)) * x applied after each MLP.

Structs

TtmModel

TtmModelBuilder

Fluent constructor for TtmModel. TtmConfig (from a parsed config.json) doubles as this model’s load-time config — there’s no separate InferConfig to map onto.