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 class

Nvfp4Scales

Nvfp4Scales

class max.nn.moe.Nvfp4Scales(gate_up_input, down_input, gate_up_expert, down_expert)

source

Bases: object

Bundled scales for NVFP4 quantization.

Note: gate_up_input is broadcast-uniform (all experts share the global max), whereas down_input contains genuinely per-expert scales. The non-EP path must account for this asymmetry when converting per-expert scales to a single global activation scale.

Parameters:

• gate_up_input (TensorValue)
• down_input (TensorValue)
• gate_up_expert (TensorValue)
• down_expert (TensorValue)

down_expert

down_expert: TensorValue

source

down_input

down_input: TensorValue

source

gate_up_expert

gate_up_expert: TensorValue

source

gate_up_input

gate_up_input: TensorValue

source