For the complete documentation index, see llms.txt. Markdown versions of all pages are available by appending .md to any URL (e.g. /max/get-started.md).
Mojo struct
Struct_mxfp4_preshuffle_scale_4d_per_expert
struct Struct_mxfp4_preshuffle_scale_4d_per_expert
Per-step A-scale preshuffle for the AMD CDNA4 preb grouped matmul.
Takes row-major E8M0 A-scales [total_tokens, K_SCALES] and writes
cell-packed scales into per-expert fixed-stride slots of size
max_padded_M = align_up(max_num_tokens_per_expert, 32). The
mxfp4_grouped_matmul_amd_preb kernel reads slot e * max_padded_M
for expert slot e. Inactive slots and pad rows are left untouched
by this kernel; the matmul's per-expert tight V# bound guards
out-of-range reads.
Implemented traits
AnyType,
ImplicitlyDestructible
Methods
execute
static def execute[target: StringSlice[StaticConstantOrigin]](output: ManagedTensorSlice[Output, static_spec=output.static_spec], input: ManagedTensorSlice[Input, static_spec=input.static_spec], expert_start_indices: ManagedTensorSlice[Input, static_spec=expert_start_indices.static_spec], max_num_tokens_per_expert: UInt32, num_active_experts: UInt32, context: DeviceContext)
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!