gpuActivationFor function

({GpuActivationReason? reason, ActivationDataType? type}) gpuActivationFor({
  1. required LlmModelFamily? family,
  2. required bool? declaresActivation,
  3. required int maxNumTokens,
  4. required bool android,
})

The activation type a GPU load runs at when the caller set none: float32 for Gemma 4 with a context above gemma4Fp16MaxContextTokens (its float16 rotary positions stop being distinct there, LiteRT-LM #4001), and for a FunctionGemma bundle that declares no prefer_activation_type (converted before litetune 0.1.4: at float16 it answers with <pad> or nothing). family is the bundle's llm_model_type; declaresActivation whether it declares a type (null: unread). Null leaves the bundle's own.

On Android (android) Gemma 4 keeps the bundle's float16 at any context, with GpuActivationReason.gemma4LongContextAndroid above gemma4Fp16MaxContextTokens so the engine says so: float32 fails there on Adreno 8xx with this native build. An explicit float32 is still honoured.

Implementation

({ActivationDataType? type, GpuActivationReason? reason}) gpuActivationFor({
  required LlmModelFamily? family,
  required bool? declaresActivation,
  required int maxNumTokens,
  required bool android,
}) => switch (family) {
  LlmModelFamily.gemma4
      when maxNumTokens > gemma4Fp16MaxContextTokens && android =>
    (type: null, reason: GpuActivationReason.gemma4LongContextAndroid),
  LlmModelFamily.gemma4 when maxNumTokens > gemma4Fp16MaxContextTokens => (
    type: ActivationDataType.float32,
    reason: GpuActivationReason.gemma4LongContext,
  ),
  LlmModelFamily.functionGemma when declaresActivation == false => (
    type: ActivationDataType.float32,
    reason: GpuActivationReason.functionGemmaWithoutKey,
  ),
  _ => (type: null, reason: null),
};