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 module

kernels

Registers core tensor graph ops (range, copy, reshape, and related utilities).

comptime values

logger

comptime logger = Logger(stdout, prefix=String(""), source_location=False)

Structs

• ​BundledAllReduceAddRMSNormQuantFP8: Registers the mo.composite.bundled.allreduce_add_rms_norm_quant_fp8 graph op with the graph compiler.
• ​BundledAllReduceSum: Registers the mo.bundled.allreduce.sum graph op with the graph compiler.
• ​CausalConv1DVarlenFwd: Varlen causal 1D convolution forward pass.
• ​Copy: Registers the copy graph op with the graph compiler.
• ​EaglePrefillShiftTokens: Registers the mo.eagle_prefill_shift_tokens graph op with the graph compiler.
• ​Fold: Registers the fold graph op with the graph compiler.
• ​GatedDeltaConv1dFwd: Gated DeltaNet causal conv1d forward pass (Pass 1 of two-pass prefill).
• ​GatedDeltaRecurrenceFwd: Gated DeltaNet recurrence forward pass (Pass 2 of two-pass prefill).
• ​GatedGroupRMSNorm: Fused silu-gate + group RMSNorm + weight-scale for the Mamba-2 mixer.
• ​InplaceMemcpy: Copies the contents of src into dst in place.
• ​IRFFT: Registers the irfft graph op with the graph compiler.
• ​LaunchHostFunc: Enqueues a pre-packed host callback on the device's default stream.
• ​Learnable2DInterpPosEmb: Registers the learnable_2d_interp_pos_emb graph op with the graph compiler.
• ​Mamba2SSDChunkScanVarlenFwd: Varlen Mamba-2 SSD chunked-scan prefill forward.
• ​Mamba2SSDChunkScanVarlenFwdInplace: Varlen Mamba-2 SSD chunked-scan: in-place SSM-pool write-back.
• ​MaxPool: Registers the mo.max_pool graph op with the graph compiler.
• ​MaxPoolCeilModeTrue: Registers the mo.max_pool_ceil_mode_true graph op with the graph compiler.
• ​NanCheckCountOp: Counts NaN/Inf values in a floating-point tensor.
• ​NanCheckRaiseOp: Raises an error if NaN or Inf counts are non-zero.
• ​NonMaximumSuppression: Registers the mo.non_maximum_suppression graph op with the graph compiler.
• ​PackConvTransposeFilterShape: Registers the pack_conv_transpose_filter_shape graph op with the graph compiler.
• ​RandomNormal: Registers the mo.random.normal graph op with the graph compiler.
• ​RandomUniform: Registers the mo.random.uniform graph op with the graph compiler.
• ​Range: Registers the mo.range graph op with the graph compiler.
• ​RepeatInterleave: Registers the repeat_interleave graph op with the graph compiler.
• ​ROIAlign: Registers the mo.roi_align graph op with the graph compiler.
• ​Sleep: Registers the mo.sleep graph op with the graph compiler.
• ​SpatialMerge: Registers the mo.spatial_merge graph op with the graph compiler.
• ​Struct_fused_token_sampling: Registers the sampler.fused_token_sampling graph op with the graph compiler.
• ​Struct_min_p_sampling: Registers the min_p_sampling graph op with the graph compiler.
• ​Struct_moe_create_indices: Registers the mo.moe.create.indices graph op with the graph compiler.
• ​Struct_moe_create_indices_with_scales_offset: Registers the mo.moe.create.indices.with.scales.offset graph op with the graph compiler.
• ​Struct_moe_eplb_remap: Registers the mo.moe.eplb.remap graph op with the graph compiler.
• ​Struct_moe_router_group_limited: Registers the mo.moe.router.group.limited graph op with the graph compiler.
• ​Struct_moe_single_group_router: Registers the mo.moe.single.group.router graph op with the graph compiler.
• ​Struct_moe_single_group_router_eplb: Registers the mo.moe.single.group.router.eplb graph op with the graph compiler.
• ​Struct_rope_ragged_paged: Registers the mo.rope.ragged graph op with the graph compiler.
• ​Struct_rope_ragged_paged_with_position_id: Registers the mo.rope.ragged.with_position_id graph op with the graph compiler.
• ​Struct_rope_split_store_ragged_paged: Registers the mo.rope_split_store.ragged.paged graph op with the graph compiler.
• ​Struct_rope_split_store_ragged_paged_with_position_id: Registers the mo.rope_split_store.ragged.paged.with_position_id graph op with the graph compiler.
• ​Struct_sampler_apply_penalties: Registers the sampler.apply_penalties graph op with the graph compiler.
• ​Struct_sampler_update_frequency_data: Registers the sampler.update_frequency_data graph op with the graph compiler.
• ​TPoolPatchMerger: Registers the tpool_patch_merger graph op with the graph compiler.
• ​WaitHostValue: Stalls the stream until a host-visible flag reaches a given value.
• ​WaitHostValueWithDep: Variant of mo.wait_host_value that takes a fake mutable dependency operand.

Functions

• ​causal_conv1d_varlen_fwd_shape: Computes the output shape for the causal_conv1d_varlen_fwd graph op.
• ​concat_from_list_shape_impl: Computes the concatenated output shape for a variadic list of input tensors.
• ​concat_shape_impl: Computes the concatenated output shape from the input tensor shapes.
• ​export: Keeps the managed-tensor-slice load/store entry points alive for export when this package is linked.
• ​fold_kernel_shape: Computes the output shape for the fold graph op.
• ​gated_delta_conv1d_fwd_shape: Computes the output shape for the gated_delta_conv1d_fwd graph op.
• ​gated_delta_recurrence_fwd_shape: Computes the output shape for the gated_delta_recurrence_fwd graph op.
• ​gated_group_rmsnorm_shape:
• ​generic_fused_qk_rope_bshd_paged_ragged_kernel_api: Applies fused rotary position embedding to Q and K, updating keys in the paged KV cache.
• ​generic_fused_qkv_matmul_kv_cache_bshd_paged_kernel_api: Implements the fused QKV matmul for BSHD inputs, writing the K and V projections directly into a paged KV cache.
• ​generic_fused_qkv_matmul_kv_cache_paged_ragged_kernel_api: Implements the fused QKV matmul for ragged inputs, writing the K and V projections directly into a paged KV cache.
• ​generic_fused_qkv_matmul_kv_cache_paged_ragged_kernel_api_bias: Implements the fused QKV matmul with bias for ragged inputs, writing the K and V projections directly into a paged KV cache.
• ​layout_transform_conv_filter_common: Packs a convolution filter into the layout expected by the conv kernels.
• ​layout_transform_conv_transpose_filter_common: Packs a transposed-convolution filter into the layout expected by the conv_transpose kernels.
• ​mamba2_ssd_chunk_scan_varlen_fwd_inplace_shape: Computes the output shape for the mamba2_ssd_chunk_scan_varlen_fwd_inplace graph op.
• ​mamba2_ssd_chunk_scan_varlen_fwd_shape: Computes the output shape for the mamba2_ssd_chunk_scan_varlen_fwd graph op.
• ​max_pool_ceil_mode_true_shape: Computes the output shape for the mo.max_pool_ceil_mode_true graph op.
• ​max_pool_shape: Computes the output shape for the mo.max_pool graph op.
• ​non_maximum_suppression_shape: Computes the output shape for the mo.non_maximum_suppression graph op.
• ​pack_conv_transpose_filter_shape_shape: Computes the output shape for the pack_conv_transpose_filter_shape graph op.
• ​print_kv_cache_paged_generic_kernel_api: Prints the contents of a paged KV cache for debugging.
• ​random_normal_shape: Computes the output shape for the mo.random.normal graph op.
• ​random_uniform_shape: Computes the output shape for the mo.random.uniform graph op.
• ​range_shape: Computes the output shape for the mo.range graph op.
• ​reduce_shape: Compute the output shape of a reduce operation, and assert the inputs are compatible.
• ​repeat_interleave_kernel_shape: Computes the output shape for the repeat_interleave graph op.
• ​roi_align_shape: Computes the output shape for the mo.roi_align graph op.
• ​tpool_patch_merger_shape: Computes the output shape for the tpool_patch_merger graph op.