submission 552589

#!POPCORN leaderboard gated_deltanet_chunk_fwd_o
#!POPCORN gpu B200_Nebius

from task import input_t, output_t

import torch
import helion
import helion.language as hl


# Per-shape configs: map (B, T, H, K, V) to optimized helion.Config objects.
# Autotuned config from B200
_TUNED = helion.Config(block_sizes=[], indexing=['tensor_descriptor', 'pointer', 'pointer', 'pointer', 'pointer', 'pointer'], l2_groupings=[1], load_eviction_policies=['', '', '', '', ''], loop_orders=[[1, 0]], num_stages=1, num_warps=16, pid_type='flat', range_flattens=[None], range_multi_buffers=[None], range_num_stages=[0], range_unroll_factors=[0], range_warp_specializes=[None])

SHAPE_CONFIGS: dict[tuple, helion.Config] = {
    # Test shapes
    (1, 64, 2, 64, 64): _TUNED,
    (2, 128, 4, 64, 64): _TUNED,
    (1, 256, 4, 64, 128): _TUNED,
    # Benchmark shapes
    (1, 64, 1, 64, 64): _TUNED,
    (2, 512, 3, 64, 64): _TUNED,
    (2, 1024, 3, 64, 64): _TUNED,
    (3, 1024, 4, 100, 100): _TUNED,
    (4, 1024, 4, 128, 128): _TUNED,
    (2, 1536, 4, 128, 128): _TUNED,
    (4, 2048, 8, 64, 64): _TUNED,
}


def _make_kernel(config: helion.Config):
    @helion.kernel(static_shapes=True, dot_precision="ieee", config=config)
    def kernel(
        q: torch.Tensor,     # [B, T, H, K]
        k: torch.Tensor,     # [B, T, H, K]
        v: torch.Tensor,     # [B, T, H, V]
        h: torch.Tensor,     # [B, NT, H, K, V]
        g: torch.Tensor,     # [B, T, H]
        scale: float,
    ) -> torch.Tensor:
        B, T, H, K = q.shape
        V = v.shape[-1]
        C = 64
        K = hl.specialize(K)
        V = hl.specialize(V)

        out = torch.empty_like(v)

        BH = B * H
        for flat_bh, tile_t in hl.tile([BH, T], block_size=[1, C]):
            b_idx = flat_bh.begin // H
            h_idx = flat_bh.begin % H
            c_idx = tile_t.begin // C

            g_vals = g[b_idx, tile_t, h_idx].to(torch.float32)  # [C]
            q_chunk = q[b_idx, tile_t, h_idx, :].to(torch.float32)  # [C, K]
            k_chunk = k[b_idx, tile_t, h_idx, :].to(torch.float32)  # [C, K]
            v_chunk = v[b_idx, tile_t, h_idx, :]  # [C, V]

            # Local: qk = q @ k^T * exp(g_i - g_j), masked causal
            qk = hl.dot(q_chunk, k_chunk.T)
            # Apply gating: exp(g_i - g_j) -- numerically stable since g_i - g_j <= 0 for i >= j
            g_diff = g_vals[:, None] - g_vals[None, :]
            qk = qk * torch.exp(g_diff)
            # Causal mask
            idx = hl.arange(tile_t.block_size)
            mask = idx[:, None] >= idx[None, :]
            qk = torch.where(mask, qk, 0.0)
            # local_out = qk @ v
            local_out = hl.dot(qk.to(v.dtype), v_chunk)

            # Global: (q @ h) * exp(g) -- g is negative cumsum so exp(g) <= 1
            q_g = q_chunk * torch.exp(g_vals)[:, None]
            global_out = hl.dot(q_g, h[b_idx, c_idx, h_idx, :, :])

            out[b_idx, tile_t, h_idx, :] = ((global_out + local_out) * scale).to(out.dtype)

        return out

    return kernel


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


def custom_kernel(data: input_t) -> output_t:
    q, k, v_new, h, g = data
    B, T, H, K = q.shape
    V = v_new.shape[-1]
    scale = K ** -0.5
    kernel = _KERNELS[(B, T, H, K, V)]
    return kernel(q, k, v_new, h, g, scale)