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 module

qmatmul_gpu

Provides GPU kernels for block-wise quantized int4 matrix multiplication.

Computes C = A @ B on the GPU, where B holds 4-bit quantized integer weights packed eight values per 32-bit word and A and C are bfloat16. The weights are dequantized in shared memory and fed through a multistage, software-pipelined tensor-core GEMM.

Limitations

• Requires an NVIDIA GPU target.
• B must be repacked before the matmul: use gpu_qint4_repack_Q4_0 for GGUF Q4_0 weights or gpu_qint4_repack_GPTQ for GPTQ weights (with optional activation-order permutation).

Functions

• ​gpu_qint4_repack_GPTQ: Launches the GPU kernel that repacks GPTQ weights into the packed GEMM layout.
• ​gpu_qint4_repack_Q4_0: Launches the GPU kernel that repacks Q4_0 weights into the packed GEMM layout.
• ​matmul_gpu_qint4: Launches a GPU int4 quantized matrix multiplication for the given tile tensors.
• ​matmul_gpu_qint4_impl: Dispatches a GPU int4 quantized matrix multiplication to the tuned kernel configuration for the runtime M dimension.

Was this page helpful?

Thank you! We'll create more content like this.

Thank you for helping us improve!

😔 What went wrong?

Some code doesn’t work

It includes inaccurate information

It's missing information I need

It was difficult to understand

Other

Submit