TopK Bitonic Kimi K3 (256k)

"""Custom CUDA top-k for H100 (SM90).

Top-k over the last dim of a 2D fp32 tensor, values+indices sorted descending.

Two kernel strategies, plus a generic fallback:
  - small k (<= 8): per-thread register top-k (tile sort-8 + running merge
    with a whole-tile float reject), warp restore-merge tournament.
  - 16 <= k <= 64: per-warp distributed truncated bitonic top-W over register
    runs (strided layout, merge-truncate), block tree in smem.
  - k > 64: warp-cooperative sorted-list kernel (generic, robust fallback).

All selection is done on packed 64-bit keys (ordered-float32 << 32 | idx), so
values sort descending with a deterministic index tie-break.
"""
import os

os.environ["TORCH_CUDA_ARCH_LIST"] = "9.0"  # container pins many archs; H100 only

import torch
import torch.nn as nn
from torch.utils.cpp_extension import load_inline

OP_TYPE = "topk"
SUPPORTED_PRECISIONS = ["fp32"]
HARDWARE_REQUIRED = ["RTX_PRO_6000", "H100", "B200"]

_CUDA_SRC = r'''
#include <torch/extension.h>
#include <ATen/cuda/CUDAContext.h>
#include <ATen/cuda/EmptyTensor.h>
#include <cuda_runtime.h>
#include <math_constants.h>

using u32 = unsigned int;
using u64 = unsigned long long;

#define DEV __device__ __forceinline__

// ---------------- key packing ----------------
// Map fp32 to a monotone u32 (order-preserving). -0.0 < +0.0, NaN sorts above
// +inf (the usual "NaN largest" convention).
DEV u32 f2ord(float f) {
    u32 b = __float_as_uint(f);
    return (b & 0x80000000u) ? ~b : (b | 0x80000000u);
}
DEV float ord2f(u32 o) {
    u32 b = (o & 0x80000000u) ? (o ^ 0x80000000u) : ~o;
    return __uint_as_float(b);
}
// sort key: high 32 bits = ordered value, low 32 = element index (within row).
// Sorting keys descending == values descending (idx desc as tie-break).
DEV u64 pack(float f, u32 idx) { return ((u64)f2ord(f) << 32) | (u64)idx; }
// Pad key: below every real key (real keys are >= ((u64)0x007FFFFF << 32)).
#define PADKEY 0ULL

DEV void ce_desc(u64 &a, u64 &b) { u64 mx = a > b ? a : b; u64 mn = a > b ? b : a; a = mx; b = mn; }
DEV void ce_asc (u64 &a, u64 &b) { u64 mx = a > b ? a : b; u64 mn = a > b ? b : a; a = mn; b = mx; }

// ---------------- path A: per-thread insertion lists / tiles ----------------

template <int KP> DEV void list_insert(u64 (&l)[KP], u64 key) {
    if (key <= l[KP - 1]) return;
#pragma unroll
    for (int j = 0; j < KP; ++j) {
        bool gt = key > l[j];
        u64 t = l[j];
        l[j] = gt ? key : t;
        key  = gt ? t : key;
    }
}

template <int KP> DEV void clean_priv(u64 (&l)[KP]) {
#pragma unroll
    for (int d = KP / 2; d >= 1; d >>= 1)
#pragma unroll
        for (int j = 0; j < KP; ++j)
            if ((j & d) == 0 && (j + d) < KP) ce_desc(l[j], l[j + d]);
}
template <> DEV void clean_priv<1>(u64 (&)[1]) {}

template <int N> DEV void sortnet_priv(u64 (&l)[N]) {
#pragma unroll
    for (int s = 2; s <= N; s <<= 1)
#pragma unroll
        for (int d = s >> 1; d >= 1; d >>= 1)
#pragma unroll
            for (int i = 0; i < N; ++i)
                if ((i & d) == 0) {
                    bool desc = (i & s) == 0;
                    if (desc) ce_desc(l[i], l[i + d]);
                    else      ce_asc (l[i], l[i + d]);
                }
}

// Warp tournament: every lane holds its own sorted-desc list[KP]; afterwards
// every lane holds the warp's top-KP (redundant).
template <int KP> DEV void warp_topk_priv(u64 (&l)[KP]) {
#pragma unroll
    for (int s = 1; s < 32; s <<= 1) {
#pragma unroll
        for (int j = 0; j < KP; ++j) {
            u64 p = __shfl_xor_sync(0xffffffffu, l[KP - 1 - j], s);
            l[j] = l[j] > p ? l[j] : p;
        }
        clean_priv<KP>(l);
    }
}

// ---------------- width-W distributed state (strided layout) ----------------

template <int RV, int W> DEV void dist_clean(u64 (&st)[RV], int lane, u32 mask) {
#pragma unroll
    for (int d = W / 2; d >= 1; d >>= 1) {
        if (d >= 32) {
            const int t = d / 32;
#pragma unroll
            for (int r = 0; r < RV; ++r)
                if ((r & t) == 0 && (r + t) < RV) ce_desc(st[r], st[r + t]);
        } else {
#pragma unroll
            for (int r = 0; r < RV; ++r) {
                u64 pv = __shfl_xor_sync(mask, st[r], d);
                bool low = (lane & d) == 0;
                st[r] = low ? (st[r] > pv ? st[r] : pv)
                            : (st[r] > pv ? pv : st[r]);
            }
        }
    }
}

template <int RV, int W>
DEV void dist_merge_mem(u64 (&st)[RV], const u64* __restrict__ partner,
                        int lane, u32 mask) {
#pragma unroll
    for (int r = 0; r < RV; ++r) {
        if (lane + 32 * r < W) {
            int p = lane + 32 * r;
            u64 pv = partner[W - 1 - p];
            st[r] = st[r] > pv ? st[r] : pv;
        }
    }
    dist_clean<RV, W>(st, lane, mask);
}

template <int RV, int W>
DEV void dist_merge_reg(u64 (&st)[RV], const u64 (&pt)[RV], int lane, u32 mask) {
#pragma unroll
    for (int r = 0; r < RV; ++r) {
        int p = lane + 32 * r;
        int q = W - 1 - p;
        u64 pv = __shfl_sync(mask, pt[q / 32], q % 32);
        st[r] = st[r] > pv ? st[r] : pv;
    }
    dist_clean<RV, W>(st, lane, mask);
}

// Sort R register runs (32 lane-strided keys each) jointly, descending.
template <int R> DEV void warp_sort32_multi(u64 (&v)[R], int lane) {
#pragma unroll
    for (int s = 2; s <= 32; s <<= 1) {
#pragma unroll
        for (int d = s >> 1; d >= 1; d >>= 1) {
            u64 pv[R];
#pragma unroll
            for (int r = 0; r < R; ++r) pv[r] = __shfl_xor_sync(0xffffffffu, v[r], d);
            int ilow = lane & ~d;
            bool desc = (ilow & s) == 0;
            bool low = (lane & d) == 0;
            bool keepmax = (low == desc);
#pragma unroll
            for (int r = 0; r < R; ++r)
                v[r] = keepmax ? (v[r] > pv[r] ? v[r] : pv[r])
                               : (v[r] > pv[r] ? pv[r] : v[r]);
        }
    }
}

// Truncated top-W of R register runs (N = 32R candidates, strided layout).
template <int W, int R> DEV void dist_top(u64 (&st)[W / 32], u64 (&v)[R], int lane) {
    constexpr int RV = W / 32;
    warp_sort32_multi<R>(v, lane);
    if constexpr (RV == 1) {
        st[0] = v[0];
#pragma unroll
        for (int i = 1; i < R; ++i) {
            u64 pv = __shfl_sync(0xffffffffu, v[i], 31 - lane);
            st[0] = st[0] > pv ? st[0] : pv;
            dist_clean<1, 32>(st, lane, 0xffffffffu);
        }
    } else {
        st[0] = v[0];
        st[1] = __shfl_sync(0xffffffffu, v[1], 31 - lane);
        dist_clean<2, 64>(st, lane, 0xffffffffu);
#pragma unroll
        for (int i = 2; i < R; i += 2) {
            u64 w2[2];
            w2[0] = v[i];
            if (i + 1 < R) w2[1] = __shfl_sync(0xffffffffu, v[i + 1], 31 - lane);
            else w2[1] = PADKEY;
            dist_clean<2, 64>(w2, lane, 0xffffffffu);
            dist_merge_reg<2, 64>(st, w2, lane, 0xffffffffu);
        }
    }
}

// ---------------- shared block finalization ----------------
template <int RV, int W, int TPB>
DEV void block_finalize(u64 (&st)[RV], int row, int part, int parts,
                        int k, u64* __restrict__ smem,
                        u64* __restrict__ partials, u32* __restrict__ counters,
                        float* __restrict__ out_v, int64_t* __restrict__ out_i) {
    const int tid = threadIdx.x;
    const int lane = tid & 31;
    const int warp = tid >> 5;
    const int NW = TPB / 32;
    const u32 mask = (W >= 32) ? 0xffffffffu : ((1u << W) - 1u);

    if (NW > 1) {
#pragma unroll
        for (int r = 0; r < RV; ++r)
            if (lane + 32 * r < W) smem[warp * W + lane + 32 * r] = st[r];
        __syncthreads();
#pragma unroll
        for (int off = NW / 2; off >= 1; off >>= 1) {
            if (warp < off) {
                const u64* partner = smem + (warp + off) * W;
                dist_merge_mem<RV, W>(st, partner, lane, mask);
#pragma unroll
                for (int r = 0; r < RV; ++r)
                    if (lane + 32 * r < W) smem[warp * W + lane + 32 * r] = st[r];
            }
            __syncthreads();
        }
    }

    if (parts > 1) {
        if (warp == 0) {
            u64* dst = partials + ((long)row * parts + part) * W;
#pragma unroll
            for (int r = 0; r < RV; ++r)
                if (lane + 32 * r < W) dst[lane + 32 * r] = st[r];
        }
        __syncthreads();
        __shared__ int is_last;
        if (tid == 0) {
            __threadfence();
            unsigned int old = atomicAdd(counters + row, 1u);
            is_last = (old == (unsigned int)(parts - 1)) ? 1 : 0;
            if (is_last) counters[row] = 0u;  // self-reset for the next call
        }
        __syncthreads();
        if (!is_last) return;

        // this block merges all `parts` partials: buffer partner keys first
        const u64* base = partials + (long)row * parts * W;
        u64 buf[8][RV];
        int cnt = 0;
        for (int p = warp; p < parts && cnt < 8; p += NW, ++cnt) {
#pragma unroll
            for (int r = 0; r < RV; ++r) {
                int q = lane + 32 * r;
                if (q < W) buf[cnt][r] = base[(long)p * W + (W - 1 - q)];
            }
        }
#pragma unroll
        for (int r = 0; r < RV; ++r) st[r] = PADKEY;
        for (int c = 0; c < cnt; ++c) {
#pragma unroll
            for (int r = 0; r < RV; ++r) st[r] = st[r] > buf[c][r] ? st[r] : buf[c][r];
            dist_clean<RV, W>(st, lane, mask);
        }
        for (int p = warp + 8 * NW; p < parts; p += NW) {
            u64 a[RV];
#pragma unroll
            for (int r = 0; r < RV; ++r) {
                int q = lane + 32 * r;
                if (q < W) a[r] = base[(long)p * W + (W - 1 - q)];
            }
#pragma unroll
            for (int r = 0; r < RV; ++r) st[r] = st[r] > a[r] ? st[r] : a[r];
            dist_clean<RV, W>(st, lane, mask);
        }
        __syncthreads();
#pragma unroll
        for (int r = 0; r < RV; ++r)
            if (lane + 32 * r < W) smem[warp * W + lane + 32 * r] = st[r];
        __syncthreads();
#pragma unroll
        for (int off = NW / 2; off >= 1; off >>= 1) {
            if (warp < off) {
                const u64* partner = smem + (warp + off) * W;
                dist_merge_mem<RV, W>(st, partner, lane, mask);
#pragma unroll
                for (int r = 0; r < RV; ++r)
                    if (lane + 32 * r < W) smem[warp * W + lane + 32 * r] = st[r];
            }
            __syncthreads();
        }
    }

    if (warp == 0) {
        float* ov = out_v + (long)row * k;
        int64_t* oi = out_i + (long)row * k;
#pragma unroll
        for (int r = 0; r < RV; ++r) {
            int p = lane + 32 * r;
            if (p < W && p < k) {
                u64 key = st[r];
                ov[p] = ord2f((u32)(key >> 32));
                oi[p] = (int64_t)(u32)key;
            }
        }
    }
}

// ---------------- kernel A: small k ----------------
template <int KP, int TPB, bool TILED>
__global__ void topk_kernel_a(const float* __restrict__ x, int n, int k,
                              int parts, int elems_per_part,
                              float* __restrict__ out_v, int64_t* __restrict__ out_i,
                              u64* __restrict__ partials, u32* __restrict__ counters) {
    extern __shared__ u64 smem[];
    const int row = blockIdx.x;
    const int part = blockIdx.y;
    const int tid = threadIdx.x;
    const int lane = tid & 31;

    const long rowbase = (long)row * n;
    int begin = part * elems_per_part;
    int end = begin + elems_per_part; if (end > n) end = n;

    u64 l[KP];
    if constexpr (TILED) {
        u64 st[8];
#pragma unroll
        for (int j = 0; j < 8; ++j) st[j] = PADKEY;
        float thrF = -CUDART_INF_F;
        int idx = begin + tid;
        while (idx < end) {
            float vf[8];
#pragma unroll
            for (int c = 0; c < 8; ++c) {
                int id = idx + c * TPB;
                vf[c] = (id < end) ? __ldg(x + rowbase + id) : __int_as_float(0xff800000);
            }
            float fm = vf[0];
#pragma unroll
            for (int c = 1; c < 8; ++c) fm = fmaxf(fm, vf[c]);
            if (!(fm < thrF)) {
                u64 t[8];
#pragma unroll
                for (int c = 0; c < 8; ++c) {
                    int id = idx + c * TPB;
                    t[c] = (id < end) ? pack(vf[c], (u32)id) : PADKEY;
                }
                sortnet_priv<8>(t);
#pragma unroll
                for (int j = 0; j < 8; ++j) st[j] = st[j] > t[7 - j] ? st[j] : t[7 - j];
                clean_priv<8>(st);
                thrF = (st[KP - 1] == PADKEY) ? -CUDART_INF_F : ord2f((u32)(st[KP - 1] >> 32));
            }
            idx += 8 * TPB;
        }
#pragma unroll
        for (int j = 0; j < KP; ++j) l[j] = st[j];
    } else {
#pragma unroll
        for (int j = 0; j < KP; ++j) l[j] = PADKEY;
        int idx = begin + tid;
        while (idx < end) {
            float v[4];
            int id0 = idx;
#pragma unroll
            for (int c = 0; c < 4; ++c) {
                int id = id0 + c * TPB;
                v[c] = (id < end) ? __ldg(x + rowbase + id) : __int_as_float(0xff800000);
            }
#pragma unroll
            for (int c = 0; c < 4; ++c) {
                int id = id0 + c * TPB;
                if (id < end) list_insert<KP>(l, pack(v[c], (u32)id));
            }
            idx += 4 * TPB;
        }
    }

    warp_topk_priv<KP>(l);

    constexpr int W = KP;
    u64 st[1];
    st[0] = l[0];
#pragma unroll
    for (int j = 1; j < KP; ++j) st[0] = (lane == j) ? l[j] : st[0];

    block_finalize<1, W, TPB>(st, row, part, parts, k, smem, partials, counters,
                              out_v, out_i);
}

// ---------------- kernel B: distributed truncated bitonic top-W ----------------
template <int W, int R, int TPB>
__global__ void topk_kernel_b(const float* __restrict__ x, int n, int k,
                              int parts, int elems_per_part,
                              float* __restrict__ out_v, int64_t* __restrict__ out_i,
                              u64* __restrict__ partials, u32* __restrict__ counters) {
    extern __shared__ u64 smem[];
    const int row = blockIdx.x;
    const int part = blockIdx.y;
    const int tid = threadIdx.x;
    const int lane = tid & 31;

    const long rowbase = (long)row * n;
    int begin = part * elems_per_part;
    int end = begin + elems_per_part; if (end > n) end = n;

    constexpr int RV = (W + 31) / 32;
    u64 v[R];
#pragma unroll
    for (int j = 0; j < R; ++j) {
        int idx = begin + tid + j * TPB;
        u64 key = PADKEY;
        if (idx < end) key = pack(__ldg(x + rowbase + idx), (u32)idx);
        v[j] = key;
    }
    static_assert(32 * R >= W, "sort must cover state width");
    u64 st[RV];
    dist_top<W, R>(st, v, lane);

    block_finalize<RV, W, TPB>(st, row, part, parts, k, smem, partials, counters,
                               out_v, out_i);
}

// ---------------- kernel F: generic warp list kernel (k > 64) ----------------
// Each warp keeps a sorted-desc list of k keys in smem; warp-cooperative
// insertion for any element above the list minimum. One block per row.
// Robustness path (rare/hypothetical shapes); simplicity over speed.
__global__ void topk_kernel_f(const float* __restrict__ x, int n, int k,
                              float* __restrict__ out_v, int64_t* __restrict__ out_i) {
    extern __shared__ u64 smem[];
    const int row = blockIdx.x;
    const int tid = threadIdx.x;
    const int lane = tid & 31;
    const int warp = tid >> 5;
    const int NW = blockDim.x >> 5;
    const long rowbase = (long)row * n;
    u64* list = smem + (long)warp * (k + 32);

    for (int i = lane; i < k; i += 32) list[i] = PADKEY;
    __syncwarp();

    // uniform trip count: lane idx = base + lane; guards per element
    for (int base = warp * 32; base < n; base += blockDim.x) {
        int idx = base + lane;
        u64 key = (idx < n) ? pack(__ldg(x + rowbase + idx), (u32)idx) : PADKEY;
        unsigned ballot = __ballot_sync(0xffffffffu, key > list[k - 1]);
        while (ballot) {
            int src = __ffs(ballot) - 1;
            ballot &= ballot - 1;
            u64 cand = __shfl_sync(0xffffffffu, key, src);
            int lo = 0, hi = k - 1;
            while (lo < hi) {
                int mid = (lo + hi) >> 1;
                if (list[mid] < cand) hi = mid; else lo = mid + 1;
            }
            for (int j = k - 1; j > lo; --j) list[j] = list[j - 1];
            list[lo] = cand;
            __syncwarp();
        }
    }
    __syncthreads();
    // merge the NW warp lists into warp 0's list (single-lane linear merge)
    if (warp == 0 && lane == 0) {
        u64* a = list;
        for (int w = 1; w < NW; ++w) {
            u64* b = smem + (long)w * (k + 32);
            u64* tmp = smem + (long)NW * (k + 32);  // scratch area
            int i = 0, j = 0;
            for (int t = 0; t < k; ++t) {
                if (j >= k || (i < k && a[i] >= b[j])) tmp[t] = a[i++];
                else                                   tmp[t] = b[j++];
            }
            for (int t = 0; t < k; ++t) a[t] = tmp[t];
        }
    }
    __syncthreads();
    if (warp == 0) {
        float* ov = out_v + (long)row * k;
        int64_t* oi = out_i + (long)row * k;
        for (int j = lane; j < k; j += 32) {
            ov[j] = ord2f((u32)(list[j] >> 32));
            oi[j] = (int64_t)(u32)list[j];
        }
    }
}

// ---------------- host ----------------
static inline int next_pow2(int x) { int p = 1; while (p < x) p <<= 1; return p; }

struct Workspace {
    torch::Tensor buf;
    u64* partials = nullptr;
    u32* counters = nullptr;
    long partial_capacity = 0;
    long counter_capacity = 0;
};
static Workspace g_ws;

static void ensure_workspace(const torch::TensorOptions& opt, long partial_keys, long rows) {
    if (partial_keys <= g_ws.partial_capacity && rows <= g_ws.counter_capacity) return;
    long pk = std::max(partial_keys, g_ws.partial_capacity * 2);
    long cr = std::max(rows, g_ws.counter_capacity * 2);
    if (pk < 1 << 16) pk = 1 << 16;
    if (cr < 1024) cr = 1024;
    long bytes = pk * 8 + cr * 4;
    g_ws.buf = at::zeros({bytes}, opt.dtype(at::kByte));
    g_ws.partials = reinterpret_cast<u64*>(g_ws.buf.data_ptr());
    g_ws.counters = reinterpret_cast<u32*>(reinterpret_cast<char*>(g_ws.buf.data_ptr()) + pk * 8);
    g_ws.partial_capacity = pk;
    g_ws.counter_capacity = cr;
}

std::tuple<torch::Tensor, torch::Tensor> run(torch::Tensor x, long k_) {
    TORCH_CHECK(x.is_cuda(), "x must be a CUDA tensor");
    TORCH_CHECK(x.scalar_type() == at::kFloat, "x must be fp32");
    TORCH_CHECK(x.dim() == 2, "x must be 2D");
    auto xc = x.is_contiguous() ? x : x.contiguous();
    const int rows = (int)xc.size(0);
    const int n = (int)xc.size(1);
    const int k = (int)k_;
    TORCH_CHECK(k >= 1 && k <= n, "need 1 <= k <= n");

    auto stream = at::cuda::getCurrentCUDAStream();
    auto vals = at::detail::empty_cuda({rows, k}, at::kFloat, x.device(), ::std::nullopt);
    auto idxs = at::detail::empty_cuda({rows, k}, at::kLong,  x.device(), ::std::nullopt);
    float* ov = vals.data_ptr<float>();
    int64_t* oi = idxs.data_ptr<int64_t>();
    const float* xp = xc.data_ptr<float>();

    if (k <= 8) {
        const int KP = next_pow2(k);
        const int T = 256;
        int B = next_pow2(std::max(1, std::min((64 + rows - 1) / rows, 1 << 30)));
        int maxB = (n + T - 1) / T; if (maxB < 1) maxB = 1;
        if (B > maxB) { int p = 1; while (p * 2 <= maxB) p *= 2; B = p; }
        if (B < 1) B = 1;
        int elems = (n + B - 1) / B;
        dim3 grid(rows, B);
        size_t smem = (T / 32) * KP * 8;
        u64* partials = nullptr; u32* counters = nullptr;
        if (B > 1) {
            ensure_workspace(x.options(), (long)rows * B * KP, rows);
            partials = g_ws.partials; counters = g_ws.counters;
        }
        #define LAUNCH_A(KK, TT) topk_kernel_a<KK, TT, (KK > 1)><<<grid, TT, smem, stream>>>(xp, n, k, B, elems, ov, oi, partials, counters)
        switch (KP) {
            case 1: LAUNCH_A(1, 256); break;
            case 2: LAUNCH_A(2, 256); break;
            case 4: LAUNCH_A(4, 256); break;
            case 8: LAUNCH_A(8, 256); break;
        }
        #undef LAUNCH_A
    } else if (k <= 64) {
        const int W = k <= 32 ? 32 : 64;
        const int T = 256;
        int denom = (W == 32 ? (T * 7 / 2) : (T * 8)); if (denom < 1) denom = 1;
        int B0 = (n + denom / 2) / denom; if (B0 < 1) B0 = 1;
        int pb = 1; while (pb * 2 <= B0) pb <<= 1;
        if (pb * 3 < B0 * 2) pb <<= 1;
        int B = pb; if (B > 64) B = 64; if (B < 1) B = 1;
        int elems = (n + B - 1) / B;
        int R = (elems + T - 1) / T; if (R < 4) R = 4;
        while (R > 32 && B < 64) { B *= 2; elems = (n + B - 1) / B; R = (elems + T - 1) / T; }
        if (R > 32) R = 32;
        R = next_pow2(R);
        dim3 grid(rows, B);
        size_t smem = (T / 32) * W * 8;
        u64* partials = nullptr; u32* counters = nullptr;
        if (B > 1) {
            ensure_workspace(x.options(), (long)rows * B * W, rows);
            partials = g_ws.partials; counters = g_ws.counters;
        }
        #define LAUNCH_B(WW, RR) topk_kernel_b<WW, RR, 256><<<grid, T, smem, stream>>>(xp, n, k, B, elems, ov, oi, partials, counters)
        switch (W * 100 + R) {
            case 3204: LAUNCH_B(32, 4); break;
            case 3208: LAUNCH_B(32, 8); break;
            case 3216: LAUNCH_B(32, 16); break;
            case 3232: LAUNCH_B(32, 32); break;
            case 6404: LAUNCH_B(64, 4); break;
            case 6408: LAUNCH_B(64, 8); break;
            case 6416: LAUNCH_B(64, 16); break;
            case 6432: LAUNCH_B(64, 32); break;
            default: TORCH_CHECK(false, "bad kernel B config");
        }
        #undef LAUNCH_B
    } else {
        TORCH_CHECK(k <= 2048, "k too large");
        const int T = 256;
        size_t smem = (size_t)(T / 32 + 1) * ((int)k + 32) * 8;
        cudaFuncSetAttribute(topk_kernel_f, cudaFuncAttributeMaxDynamicSharedMemorySize, 220000);
        topk_kernel_f<<<rows, T, smem, stream>>>(xp, n, k, ov, oi);
    }
    return {torch::Tensor(vals), torch::Tensor(idxs)};
}

PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.def("run", &run);
}
'''

_mod = load_inline(
    name="topk_h100",
    cpp_sources="",
    cuda_sources=_CUDA_SRC,
    verbose=False,
    extra_cuda_cflags=["-O3"],
)


class Model(nn.Module):
    """Top-k over the last dim of a 2D fp32 tensor.

    Input: x (batch, n) fp32. Output: values (batch, k) fp32 sorted
    descending, indices (batch, k) int64.
    """

    def __init__(self, batch: int, n: int, k: int):
        super().__init__()
        self.batch, self.n, self.k = batch, n, k
        self.register_buffer("_dummy", torch.zeros(1))
        self._run = _mod.run

    def forward(self, x: torch.Tensor):
        return self._run(x, self.k)

    # bypass nn.Module.__call__ hook machinery (hot path)
    __call__ = forward


# Module-level shims rebuilt by check.py / benchmark.py per shape.
batch = 64
n = 8192
k = 8


def get_inputs():
    x = torch.randn(batch, n, dtype=torch.float32)
    return [x]


def get_init_inputs():
    return [batch, n, k]