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_dispatch_sm90_fp8

def matmul_dispatch_sm90_fp8[c_type: DType, a_type: DType, b_type: DType, //, transpose_b: Bool = True, elementwise_lambda_fn: Optional[def[dtype: DType, width: SIMDLength, *, alignment: Int = Int(1)](IndexList[Int(2)], SIMD[dtype, width]) capturing thin -> None] = None, elementwise_compute_lambda_fn: Optional[def[dtype: DType, width: SIMDLength, *, alignment: Int = Int(1)](IndexList[Int(2)], SIMD[dtype, width]) capturing thin -> SIMD[dtype, width]] = None, pdl_level: PDLLevel = PDLLevel()](c: TileTensor[c_type, Storage=c.Storage, address_space=c.address_space, linear_idx_type=c.linear_idx_type], a: TileTensor[a_type, Storage=a.Storage, address_space=a.address_space, linear_idx_type=a.linear_idx_type], b: TileTensor[b_type, Storage=b.Storage, address_space=b.address_space, linear_idx_type=b.linear_idx_type], ctx: DeviceContext) -> Int

Dispatches an FP8 (e4m3fn) rank-2 matmul to the SM90 warp-specialized kernel.

Searches the llama-405B and llama-8B FP8 tuning tables for matching static N and K, then falls back to a generic kernel sized by _find_largest_bn_for_sm90_matmul for other shapes. Honors the AUTOTUNING_MODE compile-time flag to launch a single autotuning config.

Parameters:

c_type (DType): Element type of the output tensor c (inferred).
a_type (DType): Element type of the input tensor a; must be float8_e4m3fn (inferred).
b_type (DType): Element type of the input tensor b; must be float8_e4m3fn (inferred).
transpose_b (Bool): Whether b is stored transposed (defaults to True); the SM90 kernel requires this to be True to dispatch.
elementwise_lambda_fn (Optional[def[dtype: DType, width: SIMDLength, *, alignment: Int = Int(1)](IndexList[Int(2)], SIMD[dtype, width]) capturing thin -> None]): Epilogue applied to each output tile after the matmul, consuming the tile and returning nothing (defaults to None for no epilogue).
elementwise_compute_lambda_fn (Optional[def[dtype: DType, width: SIMDLength, *, alignment: Int = Int(1)](IndexList[Int(2)], SIMD[dtype, width]) capturing thin -> SIMD[dtype, width]]): Compute lambda transforming each output value before the epilogue, returning the transformed value (defaults to None for no transform).
pdl_level (PDLLevel): Programmatic dependent launch level for overlapping this kernel with prior work (defaults to PDLLevel()).

Args:

c (TileTensor[c_type, Storage=c.Storage, address_space=c.address_space, linear_idx_type=c.linear_idx_type]): Rank-2 output tensor of shape (M, N).
a (TileTensor[a_type, Storage=a.Storage, address_space=a.address_space, linear_idx_type=a.linear_idx_type]): Rank-2 input tensor of shape (M, K).
b (TileTensor[b_type, Storage=b.Storage, address_space=b.address_space, linear_idx_type=b.linear_idx_type]): Rank-2 input tensor of shape (K, N) when transpose_b is True.
ctx (DeviceContext): Device context for the kernel launch.

Returns:

Int: DISPATCH_HIT (1) if the kernel was launched, DISPATCH_MISS (0) otherwise.