submission 552497

from task import input_t, output_t

import torch
import helion
import helion.language as hl

FP8_MAX = 448.0
FP8_MIN = -448.0
FP8_EPS = 1e-10

OPTIMIZED = helion.Config(block_sizes=[32], num_warps=4, num_stages=2)
SHAPE_CONFIGS: dict[tuple, helion.Config] = {
    (1, 256, 64): OPTIMIZED,
    (4, 512, 128): OPTIMIZED,
    (16, 1024, 64): OPTIMIZED,
    (1, 4096, 128): OPTIMIZED,
    (8, 4096, 128): OPTIMIZED,
    (16, 4096, 128): OPTIMIZED,
    (256, 4096, 128): OPTIMIZED,
    (256, 8192, 128): OPTIMIZED,
    (4096, 7168, 128): OPTIMIZED,
}


def _make_kernel(config: helion.Config):
    @helion.kernel(static_shapes=True, config=config)
    def kernel(
        data: torch.Tensor,
        scales_out: torch.Tensor,
    ) -> torch.Tensor:
        nrows = data.size(0)
        ncols = hl.specialize(data.size(1))

        qout = torch.empty(nrows, ncols, dtype=torch.float32, device=data.device)

        for rr in hl.tile(nrows):
            row = data[rr, :].to(torch.float32)
            amax = torch.amax(torch.abs(row), dim=-1).clamp(min=FP8_EPS)
            scale = amax / FP8_MAX
            qout[rr, :] = (row / scale[:, None]).clamp(FP8_MIN, FP8_MAX)
            scales_out[rr] = scale

        return qout

    return kernel


_KERNELS = {shape: _make_kernel(cfg) for shape, cfg in SHAPE_CONFIGS.items()}


def custom_kernel(data: input_t) -> output_t:
    x, x_q, x_s = data
    T, H = x.shape
    G = x_s.shape[1]
    gsz = H // G
    N = T * G

    kernel = _KERNELS[(T, H, gsz)]

    flat_in = x.reshape(N, gsz)
    flat_s = x_s.reshape(N)

    flat_q = kernel(flat_in, flat_s)

    x_q[...] = flat_q.reshape(T, H)
    x_s[...] = flat_s.reshape(T, G)
    return x_q, x_s