submission 128999

#!POPCORN leaderboard nvfp4_gemm

# This is a submission template for popcorn leaderboard 'nvfp4_gemm'.
# Your task is as follows:
# >
# > You will implement a block scaled matrix-matrix multiplication kernel optimized for NVIDIA B200.
# > To be explicit, you will be given a tuple of tensors:
# > ```
# > (a, b, sfa, sfb, c)
# > ```
# > where:
# > * `a` is M x K x L in K-major order in nvfp4(e2m1)
# > * `b` is N x K x L in K-major order in nvfp4(e2m1)
# > * `sfa` is M x (K // 16) x L in K-major order in fp8(e4m3fnuz)
# > * `sfb` is N x (K // 16) x L in K-major order in fp8(e4m3fnuz)
# > * `c` is M x N x L in fp16
# >
# > Matrix sizes `M` is divisible by mma_tiler_mn[0], `N` is divisible by mma_tiler_mn[1], `K` is divisible by 256.
# > The ranking criteria is the geometric mean of the benchmark results.
# > For the grand price, your kernel will be evaluated against the speed of light analysis
# > and the solution closest to the speed of light will be awarded the grand price.
# > ```
# > The speed of light analysis based on the max(FP4 Tensor Core math throughput, DRAM memory throughput) of B200 and tested under 1.5Ghz clock:
# > M   N    K   L time[us]
# > 128 7168 16384 1 8.994
# > 128 4096 7168  1 2.354
# > 128 7168 2048  1 1.333
# > ```
# The deadline for this leaderboard is 2025-12-20 06:59:00+00:00

# You can automatically route this file to specific GPUs by adding a line
# `#!POPCORN gpus <GPUs>` to the header of this file.
# Happy hacking!

from task import input_t, output_t
import torch


def ceil_div(a, b):
    """Helper for ceiling division"""
    return (a + b - 1) // b


def to_blocked(input_matrix):
    """Convert scale factor tensor to blocked format for torch._scaled_mm"""
    rows, cols = input_matrix.shape

    n_row_blocks = ceil_div(rows, 128)
    n_col_blocks = ceil_div(cols, 4)

    padded = input_matrix
    blocks = padded.view(n_row_blocks, 128, n_col_blocks, 4).permute(0, 2, 1, 3)
    rearranged = blocks.reshape(-1, 4, 32, 4).transpose(1, 2).reshape(-1, 32, 16)

    return rearranged.flatten()


def custom_kernel(data: input_t) -> output_t:
    """
    Block-scaled FP4 GEMM using torch._scaled_mm
    Args:
        data: Tuple that expands to:
            a: torch.Tensor[float4e2m1fn] of shape [m, k, l],
            b: torch.Tensor[float4e2m1fn] of shape [n, k, l],
            sfa: torch.Tensor[float8_e4m3fnuz] of shape [m, k // 16, l],
            sfb: torch.Tensor[float8_e4m3fnuz] of shape [n, k // 16, l],
            sfa_permuted: torch.Tensor[float8_e4m3fnuz] of shape [32, 4, rest_m, 4, rest_k, l],
            sfb_permuted: torch.Tensor[float8_e4m3fnuz] of shape [32, 4, rest_n, 4, rest_k, l],
            c: torch.Tensor[float16] of shape [m, n, l]
    Returns:
        Tensor containing output in float16
        c: torch.Tensor[float16] of shape [m, n, l]
    """

    # c: [m, n, l] is pre-allocated memory to avoid timing allocation overhead.
    a, b, sfa, sfb, sfa_permuted, sfb_permuted, c = data

    # Get dimensions
    m, k, l = a.shape
    n = b.shape[0]

    # Process each layer using torch._scaled_mm
    for layer in range(l):
        # Convert scale factors to blocked format
        scale_a_blocked = to_blocked(sfa[:, :, layer])
        scale_b_blocked = to_blocked(sfb[:, :, layer])

        # Use torch._scaled_mm for FP4 GEMM
        # (m, k) @ (n, k).T -> (m, n)
        result = torch._scaled_mm(
            a[:, :, layer],
            b[:, :, layer].transpose(0, 1),
            scale_a_blocked.cuda()
            if scale_a_blocked.device.type != "cuda"
            else scale_a_blocked,
            scale_b_blocked.cuda()
            if scale_b_blocked.device.type != "cuda"
            else scale_b_blocked,
            bias=None,
            out_dtype=torch.float16,
        )

        c[:, :, layer] = result

    return c