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 package

linalg

Provides CPU and GPU implementations of linear algebra functions.

Packages

• ​arch: Provides architecture specific utility functions.
• ​matmul: Provides the backend implementation for matmuls.

Modules

• ​accumulate: Provides register-file accumulator structures used by the matmul inner loop.
• ​block_scaled_quantization: Provides block-scaled quantization kernels for NVFP4, MXFP4, and MXFP8.
• ​block_scaled_utils: Provides helpers shared by every MX block-scaled format (MXFP4 and MXFP8).
• ​bmm: Provides batched matrix multiplication (BMM) for CPU and GPU targets.
• ​fp4_utils: Provides low-level FP4 encode/decode utilities and scale-factor helpers.
• ​fp6_quantization: MXFP6 activation quantization.
• ​fp6_utils: Provides low-level FP6 encode/decode utilities and scale-factor helpers.
• ​fp8_quantization: Provides FP8 quantization kernels supporting static, dynamic, and blockwise scaling.
• ​gemv: Provides general matrix-vector (GEMV) and general vector-matrix (GEVM) kernels for CPU and GPU.
• ​grouped_matmul: Provides grouped matrix multiplication kernels for CPU, AMD, and NVIDIA GPU targets.
• ​grouped_matmul_block_scaled_dispatch: General dispatch for grouped block-scaled matmul.
• ​grouped_matmul_sm100: Provides warp-specialized persistent grouped GEMM kernels for SM100 (B200) GPUs.
• ​grouped_matmul_sm100_1d1d: Provides block-scaled grouped GEMM for 1D×1D scale layouts on SM100 (B200) GPUs.
• ​grouped_matmul_sm100_blockwise_fp8: Provides blockwise-scaled FP8 grouped GEMM kernels for SM100 (B200) GPUs.
• ​grouped_matmul_tile_scheduler: Provides a persistent tile scheduler for grouped matmul GPU kernels.
• ​lora: Provides LoRA (Low-Rank Adaptation) grouped matmul kernels for SM100 GPUs.
• ​matrix_band_part: The module implements matrix band part functions.
• ​mx_format: Vendor-neutral element encodings of an OCP microscaling (MX) operand.
• ​mxfp4_dequant: MXFP4 dequantization kernel for H100 (SM90).
• ​mxfp4_matmul_sm90: MXFP4 matmul on H100 (SM90) via dequant-to-FP8 + FP8 GEMM.
• ​mxfp6_dequant: MXFP6 dequantization kernel.
• ​packing: Provides matrix packing routines that reorder B tiles into cache-friendly layouts for matmul.
• ​qr_factorization: Provides QR factorization of matrices using Householder reflections.
• ​structuring: Provides structured sparsity utilities for scatter-gather operations on GPU.
• ​transpose: The module implements Transpose functions.
• ​utils: Provides shared CPU matmul utilities including tile-consumer traits, kernel shape selection, and partial SIMD load/store helpers.
• ​utils_gpu: Provides GPU matmul configuration selection, block-swizzling, and Hilbert-curve tile-ordering utilities.