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).

Mojo struct

PreShuffledBGroupedGEMM

struct PreShuffledBGroupedGEMM[cu_count: Int, wg_per_cu: Int = Int(2), lane_bytes: Int = Int(16)]

Grouped GEMM for MXFP4 on AMD CDNA4 with pre-shuffled weights.

This grouped GEMM operates on weights B that have been pre-shuffled into a layout enabling coalesced reads from shared memory and direct MFMA usage. It offers a persistent kernel (grid-stride over work tiles with XCD-aware work-group swizzling) and a direct kernel (one block per output tile, expert dispatched via block_idx.z), selected at launch time by the persistent comptime flag.

Parameters

• ​cu_count (Int): Number of compute units on the target device.
• ​wg_per_cu (Int): Work groups per compute unit (default 2).
• ​lane_bytes (Int): Operand bytes per lane per MFMA — 16 (MXFP4) or 32 (MXFP8). BK_ELEMS counts ELEMENTS, so a given BK_ELEMS costs twice the registers and LDS at MXFP8.

Implemented traits

AnyType, Deinitable, Movable

comptime members

elems_per_byte

comptime elems_per_byte = (Int(32) // lane_bytes)

num_xcd

comptime num_xcd = 8

total_wg

comptime total_wg = (cu_count * wg_per_cu)

wg_per_xcd

comptime wg_per_xcd = (PreShuffledBGroupedGEMM[cu_count, wg_per_cu, lane_bytes].total_wg // Int(8))

Methods

to_swizzled_idx

static def to_swizzled_idx(linear_idx: Int) -> Int

Returns:

Int

persistent_kernel

static def persistent_kernel[BM: Int, BN: Int, BK_ELEMS: Int, WN: Int, out_dtype: DType, LayoutC: TensorLayout, LayoutA: TensorLayout, LayoutBPre: TensorLayout, LayoutSFA: TensorLayout, LayoutSFB: TensorLayout, AOffsetsLayout: TensorLayout, ExpertIdsLayout: TensorLayout, N: Int, K_BYTES: Int, b_cache_policy: CacheOperation = CacheOperation.ALWAYS, dram_to_lds: Bool = False, cluster_drain_sched: Bool = False, mfma_cluster: Int = Int(4), deep_prime: Bool = False, pipeline_depth: Int = Int(2), waves_per_eu: Int = Int(0)](c_tensor: TileTensor[out_dtype, LayoutC, MutAnyOrigin], a_tensor: TileTensor[DType.uint8, LayoutA, ImmutAnyOrigin], b_pre_tensor: TileTensor[DType.uint8, LayoutBPre, ImmutAnyOrigin], sfa_tensor: TileTensor[DType.float8_e8m0fnu, LayoutSFA, ImmutAnyOrigin], sfb_tensor: TileTensor[DType.float8_e8m0fnu, LayoutSFB, ImmutAnyOrigin], a_offsets: TileTensor[DType.uint32, AOffsetsLayout, ImmutAnyOrigin], expert_ids: TileTensor[DType.int32, ExpertIdsLayout, ImmutAnyOrigin], num_active_experts: Int32, max_padded_M: Int32)

kernel

static def kernel[BM: Int, BN: Int, BK_ELEMS: Int, WN: Int, out_dtype: DType, LayoutC: TensorLayout, LayoutA: TensorLayout, LayoutBPre: TensorLayout, LayoutSFA: TensorLayout, LayoutSFB: TensorLayout, AOffsetsLayout: TensorLayout, ExpertIdsLayout: TensorLayout, N: Int, K_BYTES: Int, b_cache_policy: CacheOperation = CacheOperation.ALWAYS, dram_to_lds: Bool = False, cluster_drain_sched: Bool = False, mfma_cluster: Int = Int(4), deep_prime: Bool = False, pipeline_depth: Int = Int(2), waves_per_eu: Int = Int(0)](c_tensor: TileTensor[out_dtype, LayoutC, MutAnyOrigin], a_tensor: TileTensor[DType.uint8, LayoutA, ImmutAnyOrigin], b_pre_tensor: TileTensor[DType.uint8, LayoutBPre, ImmutAnyOrigin], sfa_tensor: TileTensor[DType.float8_e8m0fnu, LayoutSFA, ImmutAnyOrigin], sfb_tensor: TileTensor[DType.float8_e8m0fnu, LayoutSFB, ImmutAnyOrigin], a_offsets: TileTensor[DType.uint32, AOffsetsLayout, ImmutAnyOrigin], expert_ids: TileTensor[DType.int32, ExpertIdsLayout, ImmutAnyOrigin], num_active_experts: Int32, max_padded_M: Int32)

launch

static def launch[BM: Int, BN: Int, BK_ELEMS: Int, WN: Int, persistent: Bool, b_cache_policy: CacheOperation = CacheOperation.ALWAYS, dram_to_lds: Bool = False, cluster_drain_sched: Bool = False, mfma_cluster: Int = Int(4), deep_prime: Bool = False, pipeline_depth: Int = Int(2), waves_per_eu: Int = Int(0), static_grid_z: Bool = False](c: TileTensor[Storage=c.Storage, address_space=c.address_space, linear_idx_type=c.linear_idx_type], a: TileTensor[DType.uint8, Storage=a.Storage, address_space=a.address_space, linear_idx_type=a.linear_idx_type], b_pre: TileTensor[DType.uint8, Storage=b_pre.Storage, address_space=b_pre.address_space, linear_idx_type=b_pre.linear_idx_type], a_scales: TileTensor[DType.float8_e8m0fnu, Storage=a_scales.Storage, address_space=a_scales.address_space, linear_idx_type=a_scales.linear_idx_type], b_scales: TileTensor[DType.float8_e8m0fnu, Storage=b_scales.Storage, address_space=b_scales.address_space, linear_idx_type=b_scales.linear_idx_type], a_offsets: TileTensor[DType.uint32, Storage=a_offsets.Storage, linear_idx_type=a_offsets.linear_idx_type], expert_ids: TileTensor[DType.int32, Storage=expert_ids.Storage, linear_idx_type=expert_ids.linear_idx_type], max_num_tokens_per_expert: Int, num_active_experts: Int, ctx: DeviceContext, grid_m_cap: Int = Int(-1))