For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /get-started.md).

Python module

max.nn.moe

Mixture of Experts (MoE) module.

Mixture of experts

GatedActivation	The gated activation between an MoE's gate-up and down projections.
SigmoidTopKRouter	MoE gate with sigmoid scores and a correction bias (noaux_tc).
StackedMoE	Stacked Mixture of Experts layer with configurable components.

Quantization strategies

Fp8Strategy	FP8 quantization for MoE.
GateUpFormat	Specifies the format of the combined gate/up projection weights.
Nvfp4Scales	Bundled scales for NVFP4 quantization.
NvMxf4f8Strategy	NVIDIA NVFP4/MXFP4/MXFP8 quantization for MoE.
QuantStrategy	Quantization strategy for MoE layers.

Functions

interleaved_block_scales_shape	Returns the SM100 interleaved layout's shape for [rows, cols] scales.
make_concatenated_gated_activation_fn	Builds a gated activation for concatenated [gate | up] projections.
make_interleaved_gated_activation_fn	Builds a gated activation for interleaved [g0, u0, g1, u1, ...] projections.
make_stacked_gated_activation_fn	Builds a gated activation for split (gate, up) projections.