submission 553343

#!POPCORN leaderboard fp8_quant
#!POPCORN gpu B200_Nebius

from task import input_t, output_t

import torch
import helion
import helion.language as hl
from pathlib import Path


# Per-shape configs from autotuning. KernelBot does NOT have /opt/booster_pack/ - use minimal config (block_sizes, num_warps, num_stages only).
# Full autotuned configs with ACF-specific params (indexing, load_eviction_policies, etc.) fail without ACF.
SHAPE_CONFIGS: dict[tuple, helion.Config] = {
    # Test shapes
    (1, 256, 64): helion.Config(block_sizes=[16], num_warps=8, num_stages=1),
    (4, 512, 128): helion.Config(block_sizes=[32], num_warps=8, num_stages=2),
    (16, 1024, 64): helion.Config(block_sizes=[16], num_warps=8, num_stages=1),
    (1, 4096, 128): helion.Config(block_sizes=[8], num_warps=2, num_stages=1),
    (8, 4096, 128): helion.Config(block_sizes=[32], num_warps=4, num_stages=1),
    # Benchmark shapes
    (16, 4096, 128): helion.Config(block_sizes=[1], num_warps=4, num_stages=1),
    (256, 4096, 128): helion.Config(block_sizes=[8], num_warps=4, num_stages=3),
    (256, 8192, 128): helion.Config(block_sizes=[8], num_warps=2, num_stages=1),
    (4096, 7168, 128): helion.Config(block_sizes=[32], num_warps=8, num_stages=1),
}


def _make_kernel(config: helion.Config):
    @helion.kernel(static_shapes=True, config=config)
    def kernel(
        data: torch.Tensor,       # [N, G] input rows
        qout: torch.Tensor,       # [N, G] output buffer (writes in-place)
        scales_out: torch.Tensor,  # [N] output normalization factors
    ) -> None:
        nrows = data.size(0)
        ncols = hl.specialize(data.size(1))
        MAX_VAL = 448.0

        for rr in hl.tile(nrows):
            row = data[rr, :].to(torch.float32)
            amax = torch.amax(torch.abs(row), -1)
            amax = torch.clamp(amax, min=1e-10)
            scale = amax / MAX_VAL

            qout[rr, :] = torch.clamp(row / scale[:, None], min=-448.0, max=448.0)
            scales_out[rr] = scale

    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_q = x_q.reshape(N, gsz)
    flat_s = x_s.reshape(N)

    kernel(flat_in, flat_q, flat_s)
    return x_q, x_s