flash-attention/hopper/flash_api.cpp

/******************************************************************************
 * Copyright (c) 2024, Jay Shah, Ganesh Bikshandi, Ying Zhang, Vijay Thakkar, Pradeep Ramani, Tri Dao.
 ******************************************************************************/

// Include these 2 headers instead of torch/extension.h since we don't need all of the torch headers.
#include <torch/python.h>
#include <torch/nn/functional.h>
#include <ATen/cuda/CUDAContext.h>
#include <c10/cuda/CUDAGuard.h>

#include <cutlass/numeric_types.h>

#include "flash.h"
#include "static_switch.h"

#define CHECK_DEVICE(x) TORCH_CHECK(x.is_cuda(), #x " must be on CUDA")
#define CHECK_SHAPE(x, ...) TORCH_CHECK(x.sizes() == torch::IntArrayRef({__VA_ARGS__}), #x " must have shape (" #__VA_ARGS__ ")")
#define CHECK_CONTIGUOUS(x) TORCH_CHECK(x.is_contiguous(), #x " must be contiguous")


void set_params_fprop(Flash_fwd_params &params,
                      // sizes
                      const size_t b,
                      const size_t seqlen_q,
                      const size_t seqlen_k,
                      const size_t seqlen_q_rounded,
                      const size_t seqlen_k_rounded,
                      const size_t h,
                      const size_t h_k,
                      const size_t d,
                      const size_t d_rounded,
                      // device pointers
                      const at::Tensor q,
                      const at::Tensor k,
                      const at::Tensor v,
                      at::Tensor out,
                      void *cu_seqlens_q_d,
                      void *cu_seqlens_k_d,
                      void *seqused_k,
                      void *p_d,
                      void *softmax_lse_d,
                      float p_dropout,
                      float softmax_scale,
                      int window_size_left,
                      int window_size_right,
                      bool seqlenq_ngroups_swapped=false) {

    // Reset the parameters
    params = {};

    params.is_bf16 = q.dtype() == torch::kBFloat16;
    params.is_e4m3 = q.dtype() == torch::kFloat8_e4m3fn;

    // Set the pointers and strides.
    params.q_ptr = q.data_ptr();
    params.k_ptr = k.data_ptr();
    params.v_ptr = v.data_ptr();
    // All stride are in elements, not bytes.
    params.q_row_stride = q.stride(-3);
    params.k_row_stride = k.stride(-3);
    params.v_row_stride = v.stride(-3);
    params.q_head_stride = q.stride(-2);
    params.k_head_stride = k.stride(-2);
    params.v_head_stride = v.stride(-2);
    params.o_ptr = out.data_ptr();
    params.o_row_stride = out.stride(-3);
    params.o_head_stride = out.stride(-2);

    if (cu_seqlens_q_d == nullptr) {
        params.q_batch_stride = q.stride(0);
        params.k_batch_stride = k.stride(0);
        params.v_batch_stride = v.stride(0);
        params.o_batch_stride = out.stride(0);
        if (seqlenq_ngroups_swapped) {
             params.q_batch_stride *= seqlen_q;
             params.o_batch_stride *= seqlen_q;
        }
    }

    params.cu_seqlens_q = static_cast<int *>(cu_seqlens_q_d);
    params.cu_seqlens_k = static_cast<int *>(cu_seqlens_k_d);
    params.seqused_k = static_cast<int *>(seqused_k);

    // P = softmax(QK^T)
    params.p_ptr = p_d;

    // Softmax sum
    params.softmax_lse_ptr = softmax_lse_d;

    // Set the dimensions.
    params.b = b;
    params.h = h;
    params.h_k = h_k;
    params.h_h_k_ratio = h / h_k;
    params.seqlen_q = seqlen_q;
    params.seqlen_k = seqlen_k;
    params.seqlen_q_rounded = seqlen_q_rounded;
    params.seqlen_k_rounded = seqlen_k_rounded;
    params.d = d;
    params.d_rounded = d_rounded;

    // Set the different scale values.
    params.scale_softmax = softmax_scale;
    params.scale_softmax_log2 = softmax_scale * M_LOG2E;
    __half scale_softmax_log2_half = __float2half(params.scale_softmax_log2);
    __half2 scale_softmax_log2_half2 = __half2(scale_softmax_log2_half, scale_softmax_log2_half);
    params.scale_softmax_log2_half2 = reinterpret_cast<uint32_t&>(scale_softmax_log2_half2);

    // Set this to probability of keeping an element to simplify things.
    params.p_dropout = 1.f - p_dropout;
    // Convert p from float to int so we don't have to convert the random uint to float to compare.
    // [Minor] We want to round down since when we do the comparison we use <= instead of <
    // params.p_dropout_in_uint = uint32_t(std::floor(params.p_dropout * 4294967295.0));
    // params.p_dropout_in_uint16_t = uint16_t(std::floor(params.p_dropout * 65535.0));
    params.p_dropout_in_uint8_t = uint8_t(std::floor(params.p_dropout * 255.0));
    params.rp_dropout = 1.f / params.p_dropout;
    params.scale_softmax_rp_dropout = params.rp_dropout * params.scale_softmax;
    TORCH_CHECK(p_dropout < 1.f);
    #ifdef FLASHATTENTION_DISABLE_DROPOUT
        TORCH_CHECK(p_dropout == 0.0f, "This flash attention build does not support dropout.");
    #endif

    // Causal is the special case where window_size_right == 0 and window_size_left < 0.
    // Local is the more general case where window_size_right >= 0 or window_size_left >= 0.
    params.is_causal = window_size_left < 0 && window_size_right == 0;

    if (window_size_left < 0 && window_size_right >= 0) { window_size_left = seqlen_k; }
    if (window_size_left >= 0 && window_size_right < 0) { window_size_right = seqlen_k; }
    params.window_size_left = window_size_left;
    params.window_size_right = window_size_right;

    #ifdef FLASHATTENTION_DISABLE_LOCAL
        TORCH_CHECK(params.is_causal || (window_size_left < 0 && window_size_right < 0),
            "This flash attention build does not support local attention.");
    #endif

    params.is_seqlens_k_cumulative = true;

    #ifdef FLASHATTENTION_DISABLE_UNEVEN_K
        TORCH_CHECK(d == d_rounded, "This flash attention build does not support headdim not being a multiple of 32.");
    #endif
}

void set_params_dgrad(Flash_bwd_params &params,
                      // sizes
                      const size_t b,
                      const size_t seqlen_q,
                      const size_t seqlen_k,
                      const size_t seqlen_q_rounded,
                      const size_t seqlen_k_rounded,
                      const size_t h,
                      const size_t h_k,
                      const size_t d,
                      const size_t d_rounded,
                      // device pointers
                      const at::Tensor q,
                      const at::Tensor k,
                      const at::Tensor v,
                      const at::Tensor out,
                      const at::Tensor dout,
                      at::Tensor dq,
                      at::Tensor dk,
                      at::Tensor dv,
                      void *cu_seqlens_q_d,
                      void *cu_seqlens_k_d,
                      void *dq_accum_d,
                      void *dk_accum_d,
                      void *dv_accum_d,
                      void *softmax_lse_d,
                      void *dsoftmax_sum_d,
                      float p_dropout,
                      float softmax_scale,
                      int window_size_left,
                      int window_size_right,
                      bool deterministic) {

    set_params_fprop(params,
                     b, seqlen_q, seqlen_k, seqlen_q_rounded, seqlen_k_rounded, h, h_k, d, d_rounded,
                     q, k, v, out,
                     cu_seqlens_q_d,
                     cu_seqlens_k_d,
                     nullptr,
                     nullptr,
                     softmax_lse_d,
                     p_dropout,
                     softmax_scale,
                     window_size_left,
                     window_size_right);

    // Set the pointers and strides.
    params.do_ptr = dout.data_ptr();
    params.do_row_stride = dout.stride(-3);
    params.do_head_stride = dout.stride(-2);
    params.dq_ptr = dq.data_ptr();
    params.dk_ptr = dk.data_ptr();
    params.dv_ptr = dv.data_ptr();
    params.dq_row_stride = dq.stride(-3);
    params.dk_row_stride = dk.stride(-3);
    params.dv_row_stride = dv.stride(-3);
    params.dq_head_stride = dq.stride(-2);
    params.dk_head_stride = dk.stride(-2);
    params.dv_head_stride = dv.stride(-2);

    if (cu_seqlens_q_d == nullptr) {
        params.do_batch_stride = dout.stride(0);
        params.dq_batch_stride = dq.stride(0);
        params.dk_batch_stride = dk.stride(0);
        params.dv_batch_stride = dv.stride(0);
    }

    params.dq_accum_ptr = dq_accum_d;
    params.dk_accum_ptr = dk_accum_d;
    params.dv_accum_ptr = dv_accum_d;

    // Softmax sum
    params.dsoftmax_sum = dsoftmax_sum_d;

    params.deterministic = deterministic;
}

void run_mha_fwd(Flash_fwd_params &params, cudaStream_t stream, bool force_split_kernel=false) {
    // HEADDIM_SWITCH(params.d, [&] {
    //     run_mha_fwd_<cutlass::half_t, kHeadSize>(params, stream);
    // });
    if (!params.is_e4m3) {
        if (params.is_bf16) {
            if (params.d == 64) {
                run_mha_fwd_<cutlass::bfloat16_t, 64>(params, stream);
            } else if (params.d == 128) {
                run_mha_fwd_<cutlass::bfloat16_t, 128>(params, stream);
            } else {
                run_mha_fwd_<cutlass::bfloat16_t, 256>(params, stream);
            }
        } else {
            if (params.d == 64) {
                run_mha_fwd_<cutlass::half_t, 64>(params, stream);
            } else if (params.d == 128) {
                run_mha_fwd_<cutlass::half_t, 128>(params, stream);
            } else {
                run_mha_fwd_<cutlass::half_t, 256>(params, stream);
            }
        }
    } else {
        // run_mha_fwd_<cutlass::float_e4m3_t, 128>(params, stream);
    }
}

std::vector<at::Tensor>
mha_fwd(at::Tensor &q,         // batch_size x seqlen_q x num_heads x head_size
        const at::Tensor &k,         // batch_size x seqlen_k x num_heads_k x head_size
        const at::Tensor &v,         // batch_size x seqlen_k x num_heads_k x head_size
        c10::optional<at::Tensor> &out_,             // batch_size x seqlen_q x num_heads x head_size
        const float softmax_scale,
        bool is_causal) {

    auto dprops = at::cuda::getCurrentDeviceProperties();
    bool is_sm90 = dprops->major == 9 && dprops->minor == 0;
    TORCH_CHECK(is_sm90, "FlashAttention only supports Hopper GPUs or newer.");

    auto q_dtype = q.dtype();
    TORCH_CHECK(q_dtype == torch::kFloat16 || q_dtype == torch::kBFloat16,
                "FlashAttention only support fp16 and bf16 data type for now");
    // TODO: will add e4m3 later
    // TORCH_CHECK(q_dtype == torch::kFloat16 || q_dtype == torch::kFloat8_e4m3fn,
                // "FlashAttention only support fp16 and bf16 data type");
                // "FlashAttention only support fp16 and fp8 (e4m3) data type for now");
    TORCH_CHECK(k.dtype() == q_dtype, "query and key must have the same dtype");
    TORCH_CHECK(v.dtype() == q_dtype, "query and value must have the same dtype");

    CHECK_DEVICE(q); CHECK_DEVICE(k); CHECK_DEVICE(v);

    TORCH_CHECK(q.stride(-1) == 1, "Input tensor must have contiguous last dimension");
    TORCH_CHECK(k.stride(-1) == 1, "Input tensor must have contiguous last dimension");
    TORCH_CHECK(v.stride(-1) == 1, "Input tensor must have contiguous last dimension");

    TORCH_CHECK(q.is_contiguous(), "Input tensor must be contiguous");
    TORCH_CHECK(k.is_contiguous(), "Input tensor must be contiguous");
    TORCH_CHECK(v.is_contiguous(), "Input tensor must be contiguous");

    const auto sizes = q.sizes();

    const int batch_size = sizes[0];
    int seqlen_q = sizes[1];
    int num_heads = sizes[2];
    const int head_size_og = sizes[3];
    const int seqlen_k = k.size(1);
    const int num_heads_k = k.size(2);
    TORCH_CHECK(batch_size > 0, "batch size must be positive");
    TORCH_CHECK(head_size_og <= 256, "FlashAttention forward only supports head dimension at most 256");
    TORCH_CHECK(num_heads % num_heads_k == 0, "Number of heads in key/value must divide number of heads in query");

    TORCH_CHECK(head_size_og == 64 || head_size_og == 128 || head_size_og == 256, "Only support head size 64, 128, and 256 for now");

    CHECK_SHAPE(q, batch_size, seqlen_q, num_heads, head_size_og);
    CHECK_SHAPE(k, batch_size, seqlen_k, num_heads_k, head_size_og);
    CHECK_SHAPE(v, batch_size, seqlen_k, num_heads_k, head_size_og);

    at::Tensor q_padded, k_padded, v_padded;
    if (head_size_og % 8 != 0) {
        q_padded = torch::nn::functional::pad(q, torch::nn::functional::PadFuncOptions({0, 8 - head_size_og % 8}));
        k_padded = torch::nn::functional::pad(k, torch::nn::functional::PadFuncOptions({0, 8 - head_size_og % 8}));
        v_padded = torch::nn::functional::pad(v, torch::nn::functional::PadFuncOptions({0, 8 - head_size_og % 8}));
    } else {
        q_padded = q;
        k_padded = k;
        v_padded = v;
    }

    at::Tensor out;
    if (out_.has_value()) {
        out = out_.value();
        TORCH_CHECK(out.dtype() == q_dtype, "Output must have the same dtype as inputs");
        CHECK_DEVICE(out);
        TORCH_CHECK(out.stride(-1) == 1, "Output tensor must have contiguous last dimension");
        CHECK_SHAPE(out, batch_size, seqlen_q, num_heads, head_size_og);
        if (head_size_og % 8 != 0) { out = torch::empty_like(q_padded); }
    } else {
        out = torch::empty_like(q_padded);
    }

    auto round_multiple = [](int x, int m) { return (x + m - 1) / m * m; };
    const int head_size = round_multiple(head_size_og, 8);
    const int head_size_rounded = round_multiple(head_size, 32);
    const int seqlen_q_rounded = round_multiple(seqlen_q, 128);
    const int seqlen_k_rounded = round_multiple(seqlen_k, 128);

    // Otherwise the kernel will be launched from cuda:0 device
    // Cast to char to avoid compiler warning about narrowing
    at::cuda::CUDAGuard device_guard{(char)q.get_device()};

    auto opts = q.options();

    auto softmax_lse = torch::empty({batch_size, num_heads, seqlen_q}, opts.dtype(at::kFloat));
    at::Tensor p;

    Flash_fwd_params params;
    set_params_fprop(params,
                     batch_size,
                     seqlen_q, seqlen_k,
                     seqlen_q_rounded, seqlen_k_rounded,
                     num_heads, num_heads_k,
                     head_size, head_size_rounded,
                     q_padded, k_padded, v_padded, out,
                     /*cu_seqlens_q_d=*/nullptr,
                     /*cu_seqlens_k_d=*/nullptr,
                     /*seqused_k=*/nullptr,
                     nullptr,
                     softmax_lse.data_ptr(),
                     /*p_dropout=*/0.f,
                     softmax_scale,
                     /*window_size_left=*/-1,
                     /*window_size_right=*/is_causal ? 0 : -1);

    auto tile_count_semaphore = is_causal ? torch::zeros({1}, opts.dtype(torch::kInt32)) : torch::empty({1}, opts.dtype(torch::kInt32));
    params.tile_count_semaphore = tile_count_semaphore.data_ptr<int>();

    if (seqlen_k > 0) {
        auto stream = at::cuda::getCurrentCUDAStream().stream();
        run_mha_fwd(params, stream);
    } else {
        // If seqlen_k == 0, then we have an empty tensor. We need to set the output to 0.
        out.zero_();
        softmax_lse.fill_(std::numeric_limits<float>::infinity());
    }

    at::Tensor out_padded = out;
    if (head_size_og % 8 != 0) {
        out = out.index({"...", torch::indexing::Slice(torch::indexing::None, head_size_og)});
        if (out_.has_value()) { out_.value().copy_(out); }
    }

    return {out, q_padded, k_padded, v_padded, out_padded, softmax_lse, p};
}

void run_mha_bwd(Flash_bwd_params &params, cudaStream_t stream) {
    // FP16_SWITCH(!params.is_bf16, [&] {
    //     HEADDIM_SWITCH(params.d, [&] {
    //         run_mha_bwd_<elem_type, kHeadDim>(params, stream);
    //     });
    // });
    if (params.d == 64) {
      run_mha_bwd_<cutlass::half_t, 64>(params, stream);
    } else if (params.d == 128) {
      run_mha_bwd_<cutlass::half_t, 128>(params, stream);
    } else {
      run_mha_bwd_<cutlass::half_t, 256>(params, stream);
    }
}

std::vector<at::Tensor>
mha_bwd(const at::Tensor &dout,  // batch_size x seqlen_q x num_heads, x head_size_og
        const at::Tensor &q,   // batch_size x seqlen_q x num_heads x head_size
        const at::Tensor &k,   // batch_size x seqlen_k x num_heads_k x head_size
        const at::Tensor &v,   // batch_size x seqlen_k x num_heads_k x head_size
        const at::Tensor &out,   // batch_size x seqlen_q x num_heads x head_size
        const at::Tensor &softmax_lse,     // b x h x seqlen_q
        c10::optional<at::Tensor> &dq_,   // batch_size x seqlen_q x num_heads x head_size
        c10::optional<at::Tensor> &dk_,   // batch_size x seqlen_k x num_heads_k x head_size
        c10::optional<at::Tensor> &dv_,   // batch_size x seqlen_k x num_heads_k x head_size
        const float softmax_scale,
        const bool is_causal) {

    #ifdef FLASHATTENTION_DISABLE_BACKWARD
        TORCH_CHECK(false, "This flash attention build does not support backward.");
    #endif
    auto dprops = at::cuda::getCurrentDeviceProperties();
    bool is_sm9x = dprops->major == 9 && dprops->minor >= 0;
    TORCH_CHECK(is_sm9x, "FlashAttentionHopper only supports Hopper GPUs or newer.");

    auto stream = at::cuda::getCurrentCUDAStream().stream();

    auto q_dtype = q.dtype();
    TORCH_CHECK(q_dtype == torch::kFloat16,
                // "FlashAttention only support fp16 and bf16 data type");
                "FlashAttention only support fp16 data type for now");
    TORCH_CHECK(k.dtype() == q_dtype, "query and key must have the same dtype");
    TORCH_CHECK(v.dtype() == q_dtype, "query and value must have the same dtype");
    TORCH_CHECK(out.dtype() == q_dtype, "query and out must have the same dtype");
    TORCH_CHECK(dout.dtype() == q_dtype, "query and dout must have the same dtype");

    CHECK_DEVICE(q); CHECK_DEVICE(k); CHECK_DEVICE(v);
    CHECK_DEVICE(out); CHECK_DEVICE(dout); CHECK_DEVICE(softmax_lse);

    TORCH_CHECK(q.stride(-1) == 1, "Input tensor must have contiguous last dimension");
    TORCH_CHECK(k.stride(-1) == 1, "Input tensor must have contiguous last dimension");
    TORCH_CHECK(v.stride(-1) == 1, "Input tensor must have contiguous last dimension");
    TORCH_CHECK(out.stride(-1) == 1, "out tensor must have contiguous last dimension");
    TORCH_CHECK(dout.stride(-1) == 1, "dout tensor must have contiguous last dimension");

    TORCH_CHECK(q.is_contiguous(), "Input tensor must be contiguous");
    TORCH_CHECK(k.is_contiguous(), "Input tensor must be contiguous");
    TORCH_CHECK(v.is_contiguous(), "Input tensor must be contiguous");

    const auto sizes = q.sizes();

    const int batch_size = sizes[0];
    const int seqlen_q = sizes[1];
    const int num_heads = sizes[2];
    const int head_size_og = dout.size(3);
    const int head_size = sizes[3];
    const int seqlen_k = k.size(1);
    const int num_heads_k = k.size(2);
    TORCH_CHECK(batch_size > 0, "batch size must be positive");
    TORCH_CHECK(head_size % 8 == 0, "head_size should be a multiple of 8");
    TORCH_CHECK(head_size <= 256, "FlashAttention backward only supports head dimension at most 256");
    TORCH_CHECK(num_heads % num_heads_k == 0, "Number of heads in key/value must divide number of heads in query");

    TORCH_CHECK(head_size_og == 64 || head_size_og == 128, "Only support head size 64 and 128 for now");

    auto round_multiple = [](int x, int m) { return (x + m - 1) / m * m; };
    const int head_size_rounded = round_multiple(head_size, 32);
    const int seqlen_q_rounded = round_multiple(seqlen_q, 128);
    const int seqlen_k_rounded = round_multiple(seqlen_k, 128);

    TORCH_CHECK(head_size == round_multiple(head_size_og, 8), "head_size must be head_size_og rounded to a multiple of 8");

    CHECK_SHAPE(q, batch_size, seqlen_q, num_heads, head_size);
    CHECK_SHAPE(k, batch_size, seqlen_k, num_heads_k, head_size);
    CHECK_SHAPE(v, batch_size, seqlen_k, num_heads_k, head_size);
    CHECK_SHAPE(out, batch_size, seqlen_q, num_heads, head_size);
    CHECK_SHAPE(dout, batch_size, seqlen_q, num_heads, head_size_og);

    at::Tensor dq, dk, dv;
    if (dq_.has_value()) {
        dq = dq_.value();
        TORCH_CHECK(dq.dtype() == q_dtype, "dq must have the same dtype as q");
        CHECK_DEVICE(dq);
        TORCH_CHECK(dq.stride(-1) == 1, "dq must have contiguous last dimension");
        CHECK_SHAPE(dq, batch_size, seqlen_q, num_heads, head_size);
    } else {
        dq = torch::empty_like(q);
    }
    if (dk_.has_value()) {
        dk = dk_.value();
        TORCH_CHECK(dk.dtype() == q_dtype, "dk must have the same dtype as q");
        CHECK_DEVICE(dk);
        TORCH_CHECK(dk.stride(-1) == 1, "dk must have contiguous last dimension");
        CHECK_SHAPE(dk, batch_size, seqlen_k, num_heads_k, head_size);
    } else {
        dk = torch::empty_like(k);
    }
    if (dv_.has_value()) {
        dv = dv_.value();
        TORCH_CHECK(dv.dtype() == q_dtype, "dv must have the same dtype as q");
        CHECK_DEVICE(dv);
        TORCH_CHECK(dv.stride(-1) == 1, "dv must have contiguous last dimension");
        CHECK_SHAPE(dv, batch_size, seqlen_k, num_heads_k, head_size);
    } else {
        dv = torch::empty_like(v);
    }

    at::Tensor dout_padded;
    if (head_size_og % 8 != 0) {
        dout_padded = torch::nn::functional::pad(dout, torch::nn::functional::PadFuncOptions({0, 8 - head_size_og % 8}));
    } else {
        dout_padded = dout;
    }

    // bool loop = seqlen_k > blocksize_c;
    // TODO: change later, for now set to true for simplicity
    bool loop = true;

    // Otherwise the kernel will be launched from cuda:0 device
    // Cast to char to avoid compiler warning about narrowing
    at::cuda::CUDAGuard device_guard{(char)q.get_device()};

    auto opts = q.options();
    auto softmax_d = torch::empty({batch_size, num_heads, seqlen_q_rounded}, opts.dtype(at::kFloat));
    at::Tensor dq_accum;
    at::Tensor dk_accum, dv_accum;
    if (loop) {
        dq_accum = torch::empty({batch_size, seqlen_q_rounded, num_heads, head_size_rounded}, opts.dtype(at::kFloat));
        // dk_accum = torch::zeros({batch_size, seqlen_k_rounded, num_heads_k, head_size_rounded}, opts.dtype(at::kFloat));
        // dv_accum = torch::zeros({batch_size, seqlen_k_rounded, num_heads_k, head_size_rounded}, opts.dtype(at::kFloat));
    }

    at::Tensor dk_expanded, dv_expanded;
    if (num_heads_k != num_heads) {  // MQA / GQA
        dk_expanded = torch::empty({batch_size, seqlen_k, num_heads, head_size}, opts);
        dv_expanded = torch::empty({batch_size, seqlen_k, num_heads, head_size}, opts);
    } else {
        dk_expanded = dk;
        dv_expanded = dv;
    }

    Flash_bwd_params params;

    set_params_dgrad(params,
                     batch_size,
                     seqlen_q, seqlen_k,
                     seqlen_q_rounded, seqlen_k_rounded,
                     num_heads, num_heads_k,
                     head_size, head_size_rounded,
                     q, k, v, out,
                     dout_padded, dq, dk_expanded, dv_expanded,
                     nullptr,
                     nullptr,
                     loop ? dq_accum.data_ptr() : nullptr,
                     // loop ? dk_accum.data_ptr() : nullptr,
                     // loop ? dv_accum.data_ptr() : nullptr,
                     nullptr,
                     nullptr,
                     softmax_lse.data_ptr(),
                     softmax_d.data_ptr(),
                     /*p_dropout=*/0.f,
                     softmax_scale,
                     /*window_size_left=*/-1,
                     /*window_size_right=*/-1,
                     /*deterministic=*/false);

    at::Tensor dq_semaphore = torch::zeros({(seqlen_q + 64 - 1) / 64, batch_size, num_heads}, opts.dtype(torch::kInt32));
    params.dq_semaphore = dq_semaphore.data_ptr<int>();
    // printf("dq_semaphore: %p, [%d, %d, %d]\n", params.dq_semaphore, (seqlen_q + 64 - 1) / 64, batch_size, num_heads);

    auto launch = &run_mha_bwd;

    if (seqlen_q > 0) {
        launch(params, stream);
    } else {
        // If seqlen_q == 0, then we have an empty tensor. We need to set the output to 0.
        dk_expanded.zero_();
        dv_expanded.zero_();
        softmax_d.zero_();
    }

    if (head_size_og % 8 != 0) {
        dq = dq.index({"...", torch::indexing::Slice(torch::indexing::None, head_size_og)});
        dk = dk.index({"...", torch::indexing::Slice(torch::indexing::None, head_size_og)});
        dv = dv.index({"...", torch::indexing::Slice(torch::indexing::None, head_size_og)});
    }

    return { dq, dk, dv, softmax_d };
}


PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
    m.doc() = "FlashAttention";
    m.def("fwd", &mha_fwd, "Forward pass");
    m.def("bwd", &mha_bwd, "Backward pass");
}