import math
import pdb
import pytest
import torch
torch.set_printoptions(precision=4, profile="default", sci_mode=False)
import torch.nn.functional as F
from einops import rearrange, repeat
from flash_attn import (
    flash_attn_func,
    flash_attn_kvpacked_func,
    flash_attn_qkvpacked_func,
    flash_attn_varlen_func,
    flash_attn_varlen_kvpacked_func,
    flash_attn_varlen_qkvpacked_func,
    flash_attn_with_kvcache,
    flash_attn_func_blasst,
)
from flash_attn import flash_attn_func
from flash_attn.bert_padding import pad_input, unpad_input
from flash_attn.flash_attn_interface import _get_block_size_n
from flash_attn.layers.rotary import apply_rotary_emb

MAX_HEADDIM_SM8x = 192


is_sm75 = torch.cuda.get_device_capability("cuda") == (7, 5)
is_sm8x = torch.cuda.get_device_capability("cuda")[0] == 8
is_sm80 = torch.cuda.get_device_capability("cuda") == (8, 0)
is_sm90 = torch.cuda.get_device_capability("cuda") == (9, 0)


def attn_bias_from_alibi_slopes(
    slopes, seqlen_q, seqlen_k, query_padding_mask=None, key_padding_mask=None, causal=False, key_leftpad=None
):
    batch, nheads = slopes.shape
    device = slopes.device
    slopes = rearrange(slopes, "b h -> b h 1 1")
    if causal:
        return torch.arange(-seqlen_k + 1, 1, device=device, dtype=torch.float32) * slopes
    else:
        row_idx = rearrange(torch.arange(seqlen_q, device=device, dtype=torch.long), "s -> s 1")
        col_idx = torch.arange(seqlen_k, device=device, dtype=torch.long)
        if key_leftpad is not None:
            key_leftpad = rearrange(key_leftpad, "b -> b 1 1 1")
            col_idx = repeat(col_idx, "s -> b 1 1 s", b=key_leftpad.shape[0])
            col_idx = torch.where(col_idx >= key_leftpad, col_idx - key_leftpad, 2**32)
        sk = (
            seqlen_k
            if key_padding_mask is None
            else rearrange(key_padding_mask.sum(-1), "b -> b 1 1 1")
        )
        sq = (
            seqlen_q
            if query_padding_mask is None
            else rearrange(query_padding_mask.sum(-1), "b -> b 1 1 1")
        )
        relative_pos = torch.abs(row_idx + sk - sq - col_idx)
        return -slopes * relative_pos.to(dtype=slopes.dtype)


def generate_random_padding_mask(max_seqlen, batch_size, device, mode="random"):
    assert mode in ["full", "random", "third"]
    if mode == "full":
        lengths = torch.full((batch_size, 1), max_seqlen, device=device, dtype=torch.int32)
    elif mode == "random":
        lengths = torch.randint(
            max(1, max_seqlen - 20), max_seqlen + 1, (batch_size, 1), device=device
        )
    elif mode == "third":
        lengths = torch.randint(max_seqlen // 3, max_seqlen + 1, (batch_size, 1), device=device)
    padding_mask = (
        repeat(torch.arange(max_seqlen, device=device), "s -> b s", b=batch_size) < lengths
    )
    return padding_mask


def generate_qkv(
    q, k, v, query_padding_mask=None, key_padding_mask=None, kvpacked=False, qkvpacked=False
):
    """
    Arguments:
        q: (batch_size, seqlen_q, nheads, d)
        k: (batch_size, seqlen_k, nheads_k, d)
        v: (batch_size, seqlen_k, nheads_k, d)
        query_padding_mask: (batch_size, seqlen), bool
        key_padding_mask: (batch_size, seqlen), bool
    """
    assert not (kvpacked and qkvpacked)
    batch_size, seqlen_q, nheads, d = q.shape
    _, seqlen_k, nheads_k, _ = k.shape
    _, _, _, d_v = v.shape
    assert k.shape == (batch_size, seqlen_k, nheads_k, d)
    assert v.shape == (batch_size, seqlen_k, nheads_k, d_v)

    if query_padding_mask is not None:
        q_unpad, indices_q, cu_seqlens_q, max_seqlen_q = unpad_input(q, query_padding_mask)
        output_pad_fn = lambda output_unpad: pad_input(
            output_unpad, indices_q, batch_size, seqlen_q
        )
    else:
        q_unpad = rearrange(q, "b s h d -> (b s) h d")
        cu_seqlens_q = torch.arange(
            0, (batch_size + 1) * seqlen_q, step=seqlen_q, dtype=torch.int32, device=q_unpad.device
        )
        max_seqlen_q = seqlen_q
        output_pad_fn = lambda output_unpad: rearrange(
            output_unpad, "(b s) h d -> b s h d", b=batch_size
        )

    if key_padding_mask is not None:
        k_unpad, indices_k, cu_seqlens_k, max_seqlen_k = unpad_input(k, key_padding_mask)
        v_unpad, _, _, _ = unpad_input(v, key_padding_mask)
    else:
        k_unpad = rearrange(k, "b s h d -> (b s) h d")
        v_unpad = rearrange(v, "b s h d -> (b s) h d")
        cu_seqlens_k = torch.arange(
            0, (batch_size + 1) * seqlen_k, step=seqlen_k, dtype=torch.int32, device=k_unpad.device
        )
        max_seqlen_k = seqlen_k

    if qkvpacked:
        assert (query_padding_mask == key_padding_mask).all()
        assert nheads == nheads_k
        qkv_unpad = torch.stack([q_unpad, k_unpad, v_unpad], dim=1)
        qkv = torch.stack([q, k, v], dim=2)
        if query_padding_mask is not None:
            dqkv_pad_fn = lambda dqkv_unpad: pad_input(dqkv_unpad, indices_q, batch_size, seqlen_q)
        else:
            dqkv_pad_fn = lambda dqkv_unpad: rearrange(
                dqkv_unpad, "(b s) t h d -> b s t h d", b=batch_size
            )
        return (
            qkv_unpad.detach().requires_grad_(),
            cu_seqlens_q,
            max_seqlen_q,
            qkv.detach().requires_grad_(),
            output_pad_fn,
            dqkv_pad_fn,
        )
    elif kvpacked:
        kv_unpad = torch.stack([k_unpad, v_unpad], dim=1)
        kv = torch.stack([k, v], dim=2)
        dq_pad_fn = output_pad_fn
        if key_padding_mask is not None:
            dkv_pad_fn = lambda dkv_unpad: pad_input(dkv_unpad, indices_k, batch_size, seqlen_k)
        else:
            dkv_pad_fn = lambda dkv_unpad: rearrange(
                dkv_unpad, "(b s) t h d -> b s t h d", b=batch_size
            )
        return (
            q_unpad.detach().requires_grad_(),
            kv_unpad.detach().requires_grad_(),
            cu_seqlens_q,
            cu_seqlens_k,
            max_seqlen_q,
            max_seqlen_k,
            q.detach().requires_grad_(),
            kv.detach().requires_grad_(),
            output_pad_fn,
            dq_pad_fn,
            dkv_pad_fn,
        )
    else:
        dq_pad_fn = output_pad_fn
        if key_padding_mask is not None:
            dk_pad_fn = lambda dk_unpad: pad_input(dk_unpad, indices_k, batch_size, seqlen_k)
        else:
            dk_pad_fn = lambda dk_unpad: rearrange(dk_unpad, "(b s) h d -> b s h d", b=batch_size)
        return (
            q_unpad.detach().requires_grad_(),
            k_unpad.detach().requires_grad_(),
            v_unpad.detach().requires_grad_(),
            cu_seqlens_q,
            cu_seqlens_k,
            max_seqlen_q,
            max_seqlen_k,
            q.detach().requires_grad_(),
            k.detach().requires_grad_(),
            v.detach().requires_grad_(),
            output_pad_fn,
            dq_pad_fn,
            dk_pad_fn,
        )


def construct_local_mask(
    seqlen_q,
    seqlen_k,
    window_size=(-1, -1),  # -1 means infinite window size
    query_padding_mask=None,
    key_padding_mask=None,
    device=None,
    key_leftpad=None,
):
    row_idx = rearrange(torch.arange(seqlen_q, device=device, dtype=torch.long), "s -> s 1")
    col_idx = torch.arange(seqlen_k, device=device, dtype=torch.long)
    if key_leftpad is not None:
        key_leftpad = rearrange(key_leftpad, "b -> b 1 1 1")
        col_idx = repeat(col_idx, "s -> b 1 1 s", b=key_leftpad.shape[0])
        col_idx = torch.where(col_idx >= key_leftpad, col_idx - key_leftpad, 2**32)
    sk = (
        seqlen_k
        if key_padding_mask is None
        else rearrange(key_padding_mask.sum(-1), "b -> b 1 1 1")
    )
    sq = (
        seqlen_q
        if query_padding_mask is None
        else rearrange(query_padding_mask.sum(-1), "b -> b 1 1 1")
    )
    if window_size[0] < 0:
        return col_idx > row_idx + sk - sq + window_size[1]
    else:
        sk = torch.full_like(col_idx, seqlen_k) if key_padding_mask is None else sk
        return torch.logical_or(
            col_idx > torch.minimum(row_idx + sk - sq + window_size[1], sk),
            col_idx < row_idx + sk - sq - window_size[0],
        )


def attention_ref(
    q,
    k,
    v,
    query_padding_mask=None,
    key_padding_mask=None,
    attn_bias=None,
    dropout_p=0.0,
    dropout_mask=None,
    causal=False,
    window_size=(-1, -1),  # -1 means infinite window size
    softcap=0.0,
    upcast=True,
    reorder_ops=False,
    key_leftpad=None,
    return_lse=False,
):
    if causal:
        window_size = (window_size[0], 0)
    dtype_og = q.dtype
    if upcast:
        q, k, v = q.float(), k.float(), v.float()
    seqlen_q, seqlen_k = q.shape[1], k.shape[1]
    k = repeat(k, "b s h d -> b s (h g) d", g=q.shape[2] // k.shape[2])
    v = repeat(v, "b s h d -> b s (h g) d", g=q.shape[2] // v.shape[2])
    d = q.shape[-1]
    if not reorder_ops:
        scores = torch.einsum("bthd,bshd->bhts", q / math.sqrt(d), k)
    else:
        scores = torch.einsum("bthd,bshd->bhts", q, k / math.sqrt(d))
    if softcap > 0:
        scores = scores / softcap
        scores = scores.tanh()
        scores = scores * softcap
    if key_padding_mask is not None:
        scores.masked_fill_(rearrange(~key_padding_mask, "b s -> b 1 1 s"), float("-inf"))
    if window_size[0] >= 0 or window_size[1] >= 0:
        local_mask = construct_local_mask(
            seqlen_q,
            seqlen_k,
            window_size,
            query_padding_mask,
            key_padding_mask,
            q.device,
            key_leftpad=key_leftpad,
        )
        scores.masked_fill_(local_mask, float("-inf"))
    if attn_bias is not None:
        scores = scores + attn_bias
    attention = torch.softmax(scores, dim=-1).to(v.dtype)
    # Some rows might be completely masked out so we fill them with zero instead of NaN
    if window_size[0] >= 0 or window_size[1] >= 0:
        attention = attention.masked_fill(torch.all(local_mask, dim=-1, keepdim=True), 0.0)
    # We want to mask here so that the attention matrix doesn't have any NaNs
    # Otherwise we'll get NaN in dV
    if query_padding_mask is not None:
        attention = attention.masked_fill(rearrange(~query_padding_mask, "b s -> b 1 s 1"), 0.0)
    dropout_scaling = 1.0 / (1 - dropout_p)
    # attention_drop = attention.masked_fill(~dropout_mask, 0.0) * dropout_scaling
    # output = torch.einsum('bhts,bshd->bthd', attention_drop , v)
    if dropout_mask is not None:
        attention_drop = attention.masked_fill(~dropout_mask, 0.0)
    else:
        attention_drop = attention
    output = torch.einsum("bhts,bshd->bthd", attention_drop, v * dropout_scaling)
    if query_padding_mask is not None:
        output.masked_fill_(rearrange(~query_padding_mask, "b s -> b s 1 1"), 0.0)
    if not return_lse:
        return output.to(dtype=dtype_og), attention.to(dtype=dtype_og)
    else:
        return output.to(dtype=dtype_og), attention.to(dtype=dtype_og), scores.logsumexp(dim=-1)

def ceil_div(a, b):
    return (a + b - 1) // b

def attention_blasst_ref(
    q, k, v, blockM = 128, blockN = 64, 
    blasst_scale_factor : float = 1.0e-4,
    query_padding_mask=None,
    key_padding_mask=None,
    attn_bias=None,
    dropout_p=0.0,
    dropout_mask=None,
    causal=False,
    window_size=(-1, -1),  # -1 means infinite window size
    softcap=0.0,
    upcast=True,
    reorder_ops=False,
    key_leftpad=None,
    return_lse=False,
):
    # return None, None
    if causal:
        window_size = (window_size[0], 0)
    if upcast:
        q, k, v = q.float(), k.float(), v.float()
    k = repeat(k, "b s h d -> b s (h g) d", g=q.shape[2] // k.shape[2])
    v = repeat(v, "b s h d -> b s (h g) d", g=q.shape[2] // v.shape[2])
    b, s_q, h, d = q.shape
    s_kv = k.shape[1]
    skip_softmax_threshold = blasst_scale_factor / s_kv
    loop_m = ceil_div(s_q, blockM)
    loop_n = ceil_div(s_kv, blockN)
    scores = torch.empty([b, h, s_q, s_kv], dtype = torch.float, device=q.device)
    exp_s = torch.empty([b, h, s_q, s_kv], dtype = torch.float, device=q.device)
    out = torch.empty_like(q, dtype = torch.float)
    def compute_single_tile(b, h, q_idx):
        total_blocks, skip_blocks = 0, 0
        global_row_sum , global_row_max = 0, float("-inf")
        softmax_scale = 1 / math.sqrt(d)
        scale_softmax_log2 = softmax_scale *  math.log2(math.e)
        scores_bh = scores[b, h]
        exp_s_bh = exp_s[b, h]
        ms, me = q_idx * blockM, (q_idx + 1) * blockM
        me = me if me < s_q else s_q
        for n in range(loop_n - 1, -1, -1):
            ns, ne = n * blockN, (n + 1) * blockN
            ns = ns if -ns < s_kv else -s_kv
            skip = False
            if n == loop_n - 1:
                total_blocks += 1
                # pdb.set_trace()
                scores_bh[ms:me, ns:ne] = torch.mm(q[b, ms:me, h, :], k[b, ns:ne, h, :].t())
                global_row_max = torch.max(scores_bh[ms:me, ns:ne], -1)
                exp_s_bh[ms:me, ns:ne] = torch.exp2(scores_bh[ms:me, ns:ne] * scale_softmax_log2 - global_row_max.values[:, None] * scale_softmax_log2)
                global_row_sum = torch.sum(exp_s_bh[ms:me, ns:ne], -1)
                acc_o = torch.mm(exp_s_bh[ms:me, ns:ne].to(v.dtype), v[b, ns:ne, h, :])
                acc_o_prev = acc_o.clone().detach()
            else:
                total_blocks += 1
                scores_bh[ms:me, ns:ne] = torch.mm(q[b, ms:me, h, :], k[b, ns:ne, h, :].t())
                row_max = torch.max(scores_bh[ms:me, ns:ne], -1)
                skip = torch.exp2((row_max.values[:, None] - global_row_max.values[:, None]) * scale_softmax_log2) < skip_softmax_threshold
                # pdb.set_trace()
                row_max = torch.max(torch.cat((scores_bh[ms:me, ns:ne], global_row_max.values[:, None]), 1), -1)
                if all(skip): 
                    skip_blocks += 1
                    continue
                exp_s_bh[ms:me, ns:ne] = torch.exp2(scores_bh[ms:me, ns:ne] * scale_softmax_log2 - row_max.values[:, None] * scale_softmax_log2)
                scores_scale = torch.exp2((global_row_max.values[:, None] - row_max.values[:, None]) * scale_softmax_log2)
                row_sum = torch.sum(exp_s_bh[ms:me, ns:ne], -1)
                global_row_max = row_max
                global_row_sum = (global_row_sum[:, None] * scores_scale).squeeze() + row_sum
                acc_o = torch.mm(exp_s_bh[ms:me, ns:ne].to(v.dtype), v[b, ns:ne, h, :]) + acc_o_prev * scores_scale
                acc_o_prev = acc_o.clone().detach()
        inv_sum = 1 / global_row_sum
        return acc_o * inv_sum[:, None], total_blocks, skip_blocks

    skip_blocks_info = torch.zeros([b, h, 2], dtype=torch.int, device=q.device)
    for bi in range(b):
        for hi in range(h):
            for mi in range(loop_m):
                ms, me = mi * blockM, (mi + 1) * blockM
                out[bi, ms:me, hi], total_blocks, skip_blocks = compute_single_tile(bi, hi, mi)
                skip_blocks_info[bi, hi, 0] += total_blocks
                skip_blocks_info[bi, hi, 1] += skip_blocks
    if return_lse:
        return out, exp_s, scores.logsumexp(dim=-1), skip_blocks_info
    return out, exp_s, skip_blocks_info


def attention_kvpacked_ref(
    q,
    kv,
    query_padding_mask=None,
    key_padding_mask=None,
    attn_bias=None,
    dropout_p=0.0,
    dropout_mask=None,
    causal=False,
    window_size=(-1, -1),  # -1 means infinite window size
    softcap=0.0,
    upcast=True,
    reorder_ops=False,
    key_leftpad=None,
):
    return attention_ref(
        q,
        kv[:, :, 0],
        kv[:, :, 1],
        query_padding_mask,
        key_padding_mask,
        attn_bias,
        dropout_p,
        dropout_mask,
        upcast=upcast,
        causal=causal,
        window_size=window_size,
        softcap=softcap,
        reorder_ops=reorder_ops,
        key_leftpad=key_leftpad,
    )


def attention_qkvpacked_ref(
    qkv,
    key_padding_mask=None,
    attn_bias=None,
    dropout_p=0.0,
    dropout_mask=None,
    causal=False,
    window_size=(-1, -1),  # -1 means infinite window size
    softcap=0.0,
    upcast=True,
    reorder_ops=False,
):
    return attention_ref(
        qkv[:, :, 0],
        qkv[:, :, 1],
        qkv[:, :, 2],
        key_padding_mask,
        key_padding_mask,
        attn_bias,
        dropout_p,
        dropout_mask,
        upcast=upcast,
        causal=causal,
        window_size=window_size,
        softcap=softcap,
        reorder_ops=reorder_ops,
    )


def generate_sparsity_mask(seqlen, sparsity=0.3):
    repeats = seqlen // 16 // 2
    # mask = torch.stack([torch.tensor([1, 0] * repeats, dtype=torch.bool, device='cuda'),
    #                     torch.tensor([0, 1] * repeats, dtype=torch.bool, device='cuda')], dim=-1)
    # mask = torch.stack([torch.tensor([1, 1] * repeats, dtype=torch.bool, device='cuda'),
    #                     torch.tensor([1, 1] * repeats, dtype=torch.bool, device='cuda')], dim=-1)
    # mask = torch.stack([torch.tensor([1, 1] * repeats, dtype=torch.bool, device='cuda')], dim=-1)
    # mask = torch.stack([torch.tensor([1, 0] * repeats, dtype=torch.bool, device='cuda')], dim=-1)
    nrow, ncol = seqlen // 16, seqlen // 256
    mask = torch.rand(nrow, ncol, device="cuda") < sparsity
    return mask


def attention_blocksparse_ref(qkv, blockmask, attn_mask, dropout_p, dropout_mask):
    """
    Arguments:
        qkv: (batch_size, seqlen, 3, nheads, head_dim)
        blockmask: (seqlen / 16, seqlen / 256)
        attn_mask: (batch_size, seqlen)
        dropout_p: float
        dropout_mask: (batch_size, nheads, seqlen, seqlen)
    Output:
        output: (batch_size, seqlen, nheads, head_dim)
        attention: softmax after dropout
    """
    q, k, v = qkv.float().unbind(dim=2)
    d = qkv.shape[-1]
    seqlen = qkv.shape[1]
    scores = torch.einsum("bthd,bshd->bhts", q / math.sqrt(d), k)
    scores.masked_fill_(rearrange(~attn_mask, "b s -> b 1 1 s"), float("-inf"))
    blockmask = repeat(blockmask, "s_16 s_256 -> (s_16 16) (s_256 256)")
    blockmask = blockmask[:seqlen, :seqlen]
    scores.masked_fill_(rearrange(~blockmask, "t s -> 1 1 t s"), float("-inf"))
    attention = torch.softmax(scores, dim=-1)
    attention = attention.masked_fill(rearrange(~attn_mask, "b s -> b 1 s 1"), 0.0)
    attention = attention.masked_fill_(rearrange(~blockmask, "t s -> 1 1 t s"), 0.0)
    attention_drop = attention.masked_fill(~dropout_mask, 0.0) / (1 - dropout_p)
    output = torch.einsum("bhts,bshd->bthd", attention_drop, v)
    output.masked_fill_(rearrange(~attn_mask, "b s -> b s 1 1"), 0)
    return output.to(dtype=qkv.dtype), attention.to(dtype=qkv.dtype)


def convert_flash_attn_S_to_softmax(
    S,
    seqlen_q,
    seqlen_k,
    query_padding_mask,
    key_padding_mask,
    head_dim,
    is_dropout,
    causal=False,
    window_size=(-1, -1),  # -1 means infinite window size
):
    """FlashAttention stores the S matrix in a different way.
    Arguments:
        S: (batch_size, nheads, seqlen_q_rounded, seqlen_k_rounded)
        query_padding_mask: (batch_size, seqlen_q_rounded)
        key_padding_mask: (batch_size, seqlen_k_rounded)
    """
    if causal:
        window_size = (window_size[0], 0)
    seqlen_q_rounded, seqlen_k_rounded = S.shape[-2:]
    S_converted = S
    if window_size[0] >= 0 or window_size[1] >= 0:
        local_mask = construct_local_mask(
            seqlen_q,
            seqlen_k,
            window_size,
            query_padding_mask,
            key_padding_mask,
            S.device,
        )
        local_mask = F.pad(
            local_mask,
            (0, seqlen_k_rounded - seqlen_k, 0, seqlen_q_rounded - seqlen_q),
            value=True,
        )
        S_converted = S_converted.masked_fill(local_mask, 0.0)

    # Need to zero out things not in attention_mask in case S was initialized with random values
    # and some of those values aren't overwritten.
    seqlen_q_og = (
        query_padding_mask.shape[-1] if query_padding_mask is not None else seqlen_q_rounded
    )
    if query_padding_mask is not None:
        query_padding_mask = F.pad(query_padding_mask, (0, seqlen_q_rounded - seqlen_q_og))
        S_converted = S_converted.masked_fill(rearrange(~query_padding_mask, "b s -> b 1 s 1"), 0.0)
    seqlen_k_og = key_padding_mask.shape[-1] if key_padding_mask is not None else seqlen_k
    if key_padding_mask is not None:
        key_padding_mask = F.pad(key_padding_mask, (0, seqlen_k_rounded - seqlen_k_og))
        S_converted = S_converted.masked_fill(rearrange(~key_padding_mask, "b s -> b 1 1 s"), 0.0)
    S_converted = F.pad(S_converted, (0, 0, 0, seqlen_q_og - seqlen_q_rounded))
    S_converted = F.pad(S_converted, (0, seqlen_k_og - seqlen_k_rounded))
    return S_converted[:, :, :seqlen_q, :seqlen_k]


def normalize_flash_attn_S(
    attn_unnorm,
    q,
    k,
    v,
    query_padding_mask=None,
    key_padding_mask=None,
    attn_bias=None,
    is_dropout=False,
    causal=False,
    window_size=(-1, -1),  # -1 means infinite window size
    return_lse = False
):
    """
    Arguments:
        q: (batch_size, seqlen_q, nheads, head_dim)
        k, v: (batch_size, seqlen_k, nheads, head_dim)
        key_padding_mask: (batch_size, seqlen_q)
        attn_bias: broadcastable to (batch_size, nheads, seqlen_q, seqlen_k)
    Output:
        softmax_lse: (batch_size, nheads, seqlen_q)
        softmax_max: (batch_size, nheads, seqlen_q)
    """
    if causal:
        window_size = (window_size[0], 0)
    q, k, v = q.float(), k.float(), v.float()
    _, seqlen_q, _, head_dim = q.shape
    seqlen_k = k.shape[1]
    scores = torch.einsum("bthd,bshd->bhts", q / math.sqrt(head_dim), k)
    if key_padding_mask is not None:
        scores.masked_fill_(rearrange(~key_padding_mask, "b s -> b 1 1 s"), float("-inf"))
    if window_size[0] >= 0 or window_size[1] >= 0:
        local_mask = construct_local_mask(
            seqlen_q,
            seqlen_k,
            window_size,
            query_padding_mask,
            key_padding_mask,
            q.device,
        )
        scores.masked_fill_(local_mask, float("-inf"))
    if attn_bias is not None:
        scores = scores + attn_bias.to(dtype=scores.dtype)
    block_size_n = _get_block_size_n(scores.device, head_dim, is_dropout, causal)
    scores_block = scores.split(block_size_n, dim=-1)
    lse_block = torch.stack([torch.logsumexp(s, dim=-1) for s in scores_block], dim=-1)
    lse = torch.logsumexp(lse_block, dim=-1)
    # lse could be -inf (i.e. all values in scores are -inf), and we want to set those to inf
    # so that when we do torch.exp(m - lse), we get 0.0 instead of NaN.
    lse[lse == float("-inf")] = float("inf")
    scores_max_block = torch.stack([torch.amax(s, dim=-1) for s in scores_block], dim=-1)
    cummax_block = torch.cummax(scores_max_block.flip(-1), dim=-1).values.flip(-1).unbind(dim=-1)
    attn_unnorm_block = attn_unnorm.split(block_size_n, dim=-1)
    attn_norm = torch.cat(
        [
            a * rearrange(torch.exp(m - lse), "b h s -> b h s 1")
            for a, m in zip(attn_unnorm_block, cummax_block)
        ],
        dim=-1,
    )
    if query_padding_mask is not None:
        attn_norm.masked_fill_(rearrange(~query_padding_mask, "b s -> b 1 s 1"), 0.0)
    return attn_norm.to(dtype=attn_unnorm.dtype) if not return_lse else (attn_norm.to(dtype=attn_unnorm.dtype), lse)


def get_dropout_fraction(
    dropout_mask,
    query_padding_mask=None,
    key_padding_mask=None,
    causal=False,
    window_size=(-1, -1),  # -1 means infinite window size
):
    """
    dropout_mask: (batch_size, nheads, seqlen_q, seqlen_k), bool. True means keep, False means drop.
    query_padding_mask: (batch_size, seqlen_q)
    key_padding_mask: (batch_size, seqlen_k)
    """
    if causal:
        window_size = (window_size[0], 0)
    batch_size, nheads, seqlen_q, seqlen_k = dropout_mask.shape
    dropped = ~dropout_mask
    valid = torch.ones_like(dropout_mask)
    if query_padding_mask is not None:
        dropped.masked_fill_(rearrange(~query_padding_mask, "b s -> b 1 s 1"), False)
        valid.masked_fill_(rearrange(~query_padding_mask, "b s -> b 1 s 1"), False)
    if key_padding_mask is not None:
        dropped.masked_fill_(rearrange(~key_padding_mask, "b s -> b 1 1 s"), False)
        valid.masked_fill_(rearrange(~key_padding_mask, "b s -> b 1 1 s"), False)
    if window_size[0] >= 0 or window_size[1] >= 0:
        local_mask = construct_local_mask(
            seqlen_q,
            seqlen_k,
            window_size,
            query_padding_mask,
            key_padding_mask,
            dropout_mask.device,
        )
        dropped.masked_fill_(local_mask, False)
        valid.masked_fill_(local_mask, False)
    dropped_total = dropped.sum()
    return dropped.sum() / valid.sum()


# @pytest.mark.parametrize("dtype", ([torch.float16] if is_sm75 else [torch.float16, torch.bfloat16]))
@pytest.mark.parametrize("dtype", [torch.bfloat16])
# @pytest.mark.parametrize("mha_type", ["mha", "mqa", "gqa"])
@pytest.mark.parametrize("mha_type", ["mha"])
# @pytest.mark.parametrize("deterministic", [False, True])
@pytest.mark.parametrize("deterministic", [False])
# @pytest.mark.parametrize("alibi", [False, True])
@pytest.mark.parametrize("alibi", [False])
# @pytest.mark.parametrize("local", [False, True])
@pytest.mark.parametrize("local", [False])
# @pytest.mark.parametrize("causal", [False, True])
@pytest.mark.parametrize("causal", [False])
# @pytest.mark.parametrize("d", [32, 40, 59, 64, 96, 111, 128, 160, 192, 224, 256])
# @pytest.mark.parametrize("d", [32, 64, 96, 128, 160, 192, 224, 256])
# @pytest.mark.parametrize('d', [32, 40, 64, 80, 96, 128, 160, 192])
# @pytest.mark.parametrize('d', [32, 64, 96, 128, 160, 192])
# @pytest.mark.parametrize('d', [32, 40, 64, 80, 96, 128])
# @pytest.mark.parametrize('d', [56, 80])
@pytest.mark.parametrize("d", [128])
@pytest.mark.parametrize(
    "seqlen_q,seqlen_k",
    [
        (113, 203),
        (128, 217),
        (113, 211),
        (108, 256),
        (256, 512),
        (512, 256),
        (1024, 1024),
        (1023, 1024),
        (1024, 1023),
        (2048, 2048),
    ],
)
# @pytest.mark.parametrize('seqlen_q,seqlen_k', [(2048, 2048)])
# @pytest.mark.parametrize("dropout_p", [0.0, 0.17])
@pytest.mark.parametrize("dropout_p", [0.0])
# @pytest.mark.parametrize("softcap", [0.0, 50.0])
@pytest.mark.parametrize("softcap", [0.0])
@pytest.mark.parametrize("bhsd", [False])
def test_flash_attn_output_with_lse(
    seqlen_q, seqlen_k, d, dropout_p, causal, local, alibi, deterministic, mha_type, dtype, softcap, bhsd
):
    if (
        max(seqlen_q, seqlen_k) >= 2048
        and torch.cuda.get_device_properties("cuda").total_memory <= 16 * 2**30
    ):
        pytest.skip()  # Reference implementation OOM
    if softcap > 0.0 and dropout_p > 0.0:
        pytest.skip("Softcap and dropout not supported together")
    device = "cuda"
    # set seed
    torch.random.manual_seed(0)
    batch_size = 4
    nheads = 6 if softcap == 0.0 else 4  # softcap reference impl takes more memory
    nheads_k = nheads if mha_type == "mha" else (1 if mha_type == "mqa" else 2)
    assert nheads % nheads_k == 0
    window_size = (-1, -1) if not local else torch.randint(0, seqlen_k, (2,))
    blasst_threshold_scale_factor = 10000.0
    q = torch.randn(batch_size, seqlen_q, nheads, d, device=device, dtype=dtype, requires_grad=True)
    if bhsd:
        q_bhsd = q.transpose(2, 1).contiguous()
    if softcap > 0:
        # Ensure the values of qk are at least within softcap range.
        q = q * softcap
    
    k = torch.randn(
        batch_size, seqlen_k, nheads_k, d, device=device, dtype=dtype, requires_grad=True
    )
    v = torch.randn(
        batch_size, seqlen_k, nheads_k, d, device=device, dtype=dtype, requires_grad=True
    )
    if bhsd:
        k_bhsd = k.transpose(2, 1).contiguous()
        v_bhsd = v.transpose(2, 1).contiguous()
    if alibi:
        alibi_slopes = torch.rand(batch_size, nheads, device=device, dtype=torch.float32) * 0.3
        attn_bias = attn_bias_from_alibi_slopes(alibi_slopes, seqlen_q, seqlen_k, causal=causal)
    else:
        alibi_slopes, attn_bias = None, None
    
    # print("q:", q)
    # print("k:", k)
    # print("v:", v)
    out, lse, S_dmask, skip_blocks_info = flash_attn_func_blasst(
        q if not bhsd else q_bhsd,
        k if not bhsd else k_bhsd,
        v if not bhsd else v_bhsd,
        dropout_p,
        causal=causal,
        window_size=window_size,
        softcap=softcap,
        alibi_slopes=alibi_slopes,
        deterministic=deterministic,
        return_attn_probs=True,
        bhsd=bhsd,
        blasst_threshold_scale_factor=blasst_threshold_scale_factor
    )
    # print("lse:", lse.shape, lse.stride(), lse)
    # print("S_dmask:", S_dmask)
    if dropout_p > 0.0:
        S_dmask_converted = convert_flash_attn_S_to_softmax(
            S_dmask,
            seqlen_q,
            seqlen_k,
            None,
            None,
            d,
            dropout_p > 0.0,
            causal=causal,
            window_size=window_size,
        )
        dropout_mask = S_dmask_converted >= 0
        attn_unnorm = S_dmask_converted.abs()
        k_rep = repeat(k, "b s h d -> b s (h g) d", g=nheads // nheads_k)
        v_rep = repeat(v, "b s h d -> b s (h g) d", g=nheads // nheads_k)
        attn, lse_ref = normalize_flash_attn_S(
            attn_unnorm,
            q,
            k_rep,
            v_rep,
            None,
            None,
            attn_bias,
            dropout_p > 0.0,
            causal=causal,
            window_size=window_size,
            return_lse=True,
        )
        dropout_fraction = get_dropout_fraction(
            dropout_mask, None, None, causal=causal, window_size=window_size
        ).item()
        print(f"Actual dropout fraction: {dropout_fraction}")
    else:
        dropout_mask = None
        '''
        S_dmask_converted = convert_flash_attn_S_to_softmax(
            S_dmask,
            seqlen_q,
            seqlen_k,
            None,
            None,
            d,
            dropout_p > 0.0,
            causal=causal,
            window_size=window_size,
        )
        dropout_mask = S_dmask_converted >= 0
        attn_unnorm = S_dmask_converted.abs()
        if kvpacked:
            kv_rep = repeat(kv, "b s two h d -> b s two (h g) d", g=nheads // nheads_k)
            k_rep, v_rep = kv_rep.unbind(dim=2)
        else:
            k_rep = repeat(k, "b s h d -> b s (h g) d", g=nheads // nheads_k)
            v_rep = repeat(v, "b s h d -> b s (h g) d", g=nheads // nheads_k)
        attn, lse_ref = normalize_flash_attn_S(
            attn_unnorm,
            q,
            k_rep,
            v_rep,
            None,
            None,
            attn_bias,
            dropout_p > 0.0,
            causal=causal,
            window_size=window_size,
            return_lse=True,
        )
        '''

    cu_count = torch.cuda.get_device_properties(0).multi_processor_count
    num_blocks_64 = batch_size * nheads * ((seqlen_q + 63) // 64)
    num_blocks_128 = batch_size * nheads * ((seqlen_q + 127) // 128)
    condition = num_blocks_64 < cu_count or (num_blocks_128 // cu_count == 1 and num_blocks_128 % cu_count != 0 and (num_blocks_64 + cu_count - 1) // cu_count <= 3)
    blockM = 64 if condition else 128
    # out_ref, attn_ref, lse_ref = attention_blasst_ref(
    out_ref, attn_ref, lse_ref, skip_blocks_info_ref = attention_blasst_ref(
        q,
        k,
        v,
        blockM = blockM, blockN = 64, 
        blasst_scale_factor=blasst_threshold_scale_factor,
        query_padding_mask=None,
        key_padding_mask=None,
        attn_bias=attn_bias,
        dropout_p=dropout_p,
        dropout_mask=dropout_mask,
        causal=causal,
        window_size=window_size,
        softcap=softcap,
        return_lse=True
    )
    out_pt, attn_pt, _ = attention_blasst_ref(
        q,
        k,
        v,
        blockM = 128, blockN = 64, 
        blasst_scale_factor=blasst_threshold_scale_factor,
        query_padding_mask=None,
        key_padding_mask=None,
        attn_bias=attn_bias,
        dropout_p=dropout_p,
        dropout_mask=dropout_mask,
        causal=causal,
        window_size=window_size,
        softcap=softcap,
        upcast=False,
        reorder_ops=True,
    )
    sparsity = torch.sum(torch.sum(skip_blocks_info, 0), 0)
    sparsity_ref = torch.sum(torch.sum(skip_blocks_info_ref, 0), 0)
    sparsity_ratio = sparsity[1] / sparsity[0]
    sparsity_ratio_ref = sparsity_ref[1] / sparsity_ref[0]
    print(f"Output sparsity: {sparsity_ratio:.2f}")
    print(f"Pytorch sparsity: {sparsity_ratio_ref:.2f}")
    if bhsd:
        out_bshd = out.transpose(2, 1).contiguous()
        print(f"Output max diff: {(out_bshd - out_ref).abs().max().item()}")
        print(f"Output mean diff: {(out_bshd - out_ref).abs().mean().item()}")
    else:
        print(f"Output max diff: {(out - out_ref).abs().max().item()}")
        print(f"Output mean diff: {(out - out_ref).abs().mean().item()}")
    print(f"Pytorch max diff: {(out_pt - out_ref).abs().max().item()}")
    print(f"Pytorch mean diff: {(out_pt - out_ref).abs().mean().item()}")
    '''
    if dropout_p > 0.0:
        print(f"Attention max diff: {(attn - attn_ref).abs().max().item()}")
        print(f"Attention Pytorch max diff: {(attn_pt - attn_ref).abs().max().item()}")

    g = torch.randn_like(out)
    if bhsd:
        g_bshd = g.transpose(2, 1).contiguous()
    # print("g:", g.shape, g)
    do_o = (g.float() * out.float()).sum(-1)
    if (d <= MAX_HEADDIM_SM8x or dropout_p == 0) or (is_sm80 or is_sm90):
        print("g:", g.shape)
        dq, dk, dv = torch.autograd.grad(out, (
            q if not bhsd else q_bhsd, 
            k if not bhsd else k_bhsd, 
            v if not bhsd else v_bhsd), g)
        dq_ref, dk_ref, dv_ref = torch.autograd.grad(out_ref, (q, k, v), g if not bhsd else g_bshd)
        dq_pt, dk_pt, dv_pt = torch.autograd.grad(out_pt, (q, k, v), g if not bhsd else g_bshd)
        if bhsd:
            dq = dq.transpose(2, 1).contiguous()
            dk = dk.transpose(2, 1).contiguous()
            dv = dv.transpose(2, 1).contiguous()
        print(f"dQ max diff: {(dq - dq_ref).abs().max().item()}")
        print(f"dK max diff: {(dk - dk_ref).abs().max().item()}")
        print(f"dV max diff: {(dv - dv_ref).abs().max().item()}")
        print(f"dQ mean diff: {(dq - dq_ref).abs().mean().item()}")
        print(f"dK mean diff: {(dk - dk_ref).abs().mean().item()}")
        print(f"dV mean diff: {(dv - dv_ref).abs().mean().item()}")
        print(f"dQ Pytorch max diff: {(dq_pt - dq_ref).abs().max().item()}")
        print(f"dK Pytorch max diff: {(dk_pt - dk_ref).abs().max().item()}")
        print(f"dV Pytorch max diff: {(dv_pt - dv_ref).abs().max().item()}")
        print(f"dQ Pytorch mean diff: {(dq_pt - dq_ref).abs().mean().item()}")
        print(f"dK Pytorch mean diff: {(dk_pt - dk_ref).abs().mean().item()}")
        print(f"dV Pytorch mean diff: {(dv_pt - dv_ref).abs().mean().item()}")
    # '''
    # Check that FlashAttention's numerical error is at most twice the numerical error
    # of a Pytorch implementation.
    # torch.nonzero((out - out_ref).abs()>0.01)
    # torch.nonzero((out1 - out_ref).abs()>0.01)
    # torch.nonzero((lse - lse_ref).abs()>0.01)
    # out_ref_ = torch.arange(0, d, device=device, dtype=torch.float).unsqueeze(0).unsqueeze(0).unsqueeze(0).repeat(batch_size, seqlen_q, nheads, 1) * 0.001
    # torch.nonzero((out - out_ref_).abs()>0.01)
    # pdb.set_trace()
    if bhsd:
        assert (out_bshd - out_ref).abs().max().item() <= 2 * (out_pt - out_ref).abs().max().item()
    else:
        assert (out - out_ref).abs().max().item() <= 2 * (out_pt - out_ref).abs().max().item()
    # print("dv:", dv.shape, dv[ 2,  0,  0, 70])
    # print("dv_ref:", dv_ref.shape, dv_ref[ 2,  0,  0, 70])
    # print("max:", torch.nonzero((dv - dv_ref).abs()==(dv - dv_ref).abs().max()))
    # torch.cuda.synchronize()
    return
    # if dropout_p > 0.0:
    #     assert (attn - attn_ref).abs().max().item() <= 2 * (attn_pt - attn_ref).abs().max().item()
    #     # With alibi, many of the prob values are 0.0 & -0.0 so dropout_fraction isn't accurate
    #     if not alibi:
    #         assert abs(dropout_fraction - dropout_p) <= (0.01 if not local else 0.025)
    if (d <= MAX_HEADDIM_SM8x or dropout_p == 0) or (is_sm80 or is_sm90):
        assert (dq - dq_ref).abs().max().item() <= 3 * (dq_pt - dq_ref).abs().max().item()
        assert (dk - dk_ref).abs().max().item() <= 3 * (dk_pt - dk_ref).abs().max().item()
        assert (dv - dv_ref).abs().max().item() <= 3 * (dv_pt - dv_ref).abs().max().item()