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 function

matmul_qint4

def matmul_qint4[group_size: Int, elementwise_lambda_fn: Optional[def[dtype: DType, width: SIMDLength, *, alignment: Int = Int(1)](IndexList[Int(2)], SIMD[dtype, width]) capturing thin -> None] = None](a_tt: TileTensor[DType.float32, Storage=a_tt.Storage, linear_idx_type=a_tt.linear_idx_type], b_tt: TileTensor[DType.uint8, Storage=b_tt.Storage, linear_idx_type=b_tt.linear_idx_type], c_tt: TileTensor[DType.float32, Storage=c_tt.Storage, linear_idx_type=c_tt.linear_idx_type], ctx: Optional[DeviceContext] = None)

Computes a matrix multiply of a float32 A matrix against block-wise quantized int4 B weights, producing a float32 result.

Dispatches to an architecture-specific kernel (VNNI, AVX2, NEON i8mm, or NEON dotprod) at compile time.

Parameters:

group_size (Int): Number of elements per quantization group.
elementwise_lambda_fn (Optional[def[dtype: DType, width: SIMDLength, *, alignment: Int = Int(1)](IndexList[Int(2)], SIMD[dtype, width]) capturing thin -> None]): Optional epilogue applied to each output element.

Args:

a_tt (TileTensor[DType.float32, Storage=a_tt.Storage, linear_idx_type=a_tt.linear_idx_type]): Input A tensor in float32.
b_tt (TileTensor[DType.uint8, Storage=b_tt.Storage, linear_idx_type=b_tt.linear_idx_type]): Input B tensor holding packed uint8 int4 weights.
c_tt (TileTensor[DType.float32, Storage=c_tt.Storage, linear_idx_type=c_tt.linear_idx_type]): Output C tensor in float32.
ctx (Optional[DeviceContext]): Optional device context for parallel execution.