# Copyright (c) 2025 NVIDIA CORPORATION.
# Licensed under the MIT license.

# Adapted from https://github.com/NVlabs/VILA/tree/main under the Apache 2.0 license.
# LICENSE is in incl_licenses directory.

import operator
from typing import Optional

import torch
import triton
import triton.language as tl

from .utils import compare_version, element_mul_kernel, is_hip

if compare_version("triton", operator.ge, "3.0.0"):
    try:
        # typical import path with dispatch available
        from triton.language.extra.libdevice import tanh
    except ModuleNotFoundError:
        # for working with NGC containers
        from triton.language.extra.cuda.libdevice import tanh
else:
    from triton.language.math import tanh

_TRUE = tl.constexpr(1)
_FALSE = tl.constexpr(0)


@triton.jit
def liger_cross_entropy_kernel(
    X_ptr,
    X_stride,
    Y_ptr,
    Y_stride,
    loss_ptr,
    z_loss_ptr,
    loss_stride,
    n_cols,
    n_non_ignore,
    ignore_index,
    lse_square_scale: tl.constexpr,
    label_smoothing: tl.constexpr,
    reduction: tl.constexpr,  # set it as constexpr since reduction is always known at compile time
    softcap,
    RETURN_Z_LOSS: tl.constexpr,
    BLOCK_SIZE: tl.constexpr,
    HAS_SOFTCAPPING: tl.constexpr,
):
    """
    This kernel computes both cross entropy loss and the gradient of the input.
    We only consider hard label + mean reduction for now. Please refer to https://pytorch.org/docs/stable/generated/torch.nn.CrossEntropyLoss.html for the math.

    Parameters:
    X_ptr: Pointer to input tensor.
    X_stride (int): The stride of the input tensor.
    Y_ptr: Pointer to target tensor.
    Y_stride (int): The stride of the target tensor.
    loss_ptr: Pointer to tensor to store the loss.
    z_loss_ptr: Pointer to tensor to store the z loss. No operation if RETURN_Z_LOSS is 0.
    loss_stride (int): The stride of the loss tensor.
    n_cols (int): The number of columns in the input tensor.
    n_non_ignore (int): The number of non-ignored elements in the batch.
    ignore_index (int): The index to ignore in the target.
    label_smoothing (float): The amount of smoothing when computing the loss, where 0.0 means no smoothing.
    lse_square_scale (float): The scaler of (logsumexp(_input)) ^ 2 adding to the loss for the stability of training.
    RETURN_Z_LOSS (int): The boolean value to decide whether storing z loss to z_loss_ptr or not. It must be 0 or 1.
    reduction (str): The string for the reduction to apply
    softcap (float): The upper threshold for scaling logits to the range (-softcap, +softcap).
    BLOCK_SIZE (int): The block size for Triton operations.
    HAS_SOFTCAPPING (bool): The boolean value to determine whether applying soft-capping or not.
    """

    # https://github.com/triton-lang/triton/issues/1058
    # If B*T*V is too large, program_id * stride will overflow out of int32, so we convert to int64
    program_id = tl.program_id(0).to(tl.int64)

    # 1. Load Y_ptr first because if the target is ignore_index, we can return right away
    Y_ptr += program_id * Y_stride
    y = tl.load(Y_ptr)

    # 2. locate the start index
    X_ptr += program_id * X_stride

    if y == ignore_index:
        # set all X_ptr as 0
        for i in range(0, n_cols, BLOCK_SIZE):
            X_offsets = i + tl.arange(0, BLOCK_SIZE)
            tl.store(X_ptr + X_offsets, 0.0, mask=X_offsets < n_cols)
        return

    loss_ptr += program_id * loss_stride
    z_loss_ptr += program_id * loss_stride

    # Online softmax: 2 loads + 1 store (compared with 3 loads + 1 store for the safe softmax)
    # Refer to Algorithm 3 in the paper: https://arxiv.org/pdf/1805.02867

    # 3. [Online softmax] first pass: find max + sum
    m = float("-inf")  # m is the max value. use the notation from the paper
    d = 0.0  # d is the sum. use the notation from the paper
    ori_X_y = tl.load(X_ptr + y)  # we need to store the original value of X_y for the loss calculation
    if HAS_SOFTCAPPING:
        ori_X_y = softcap * tanh(ori_X_y / softcap)

    # Label smoothing is a general case of normal cross entropy
    # See the full derivation at https://github.com/linkedin/Liger-Kernel/pull/198#issue-2503665310
    scaled_x_sum = 0.0
    eps = label_smoothing / n_cols

    for i in range(0, n_cols, BLOCK_SIZE):
        X_offsets = i + tl.arange(0, BLOCK_SIZE)
        X_block = tl.load(X_ptr + X_offsets, mask=X_offsets < n_cols, other=float("-inf"))
        if HAS_SOFTCAPPING:
            X_block = softcap * tanh(X_block / softcap)
        block_max = tl.max(X_block)
        if label_smoothing > 0:
            # scale X beforehand to avoid overflow
            scaled_x_sum += tl.sum(tl.where(X_offsets < n_cols, -eps * X_block, 0.0))
        m_new = tl.maximum(m, block_max)
        d = d * tl.exp(m - m_new) + tl.sum(tl.exp(X_block - m_new))
        m = m_new

    # log (sum(e^(X_i))) = log (sum(e ^ (max(X) * e ^ (X_i - max(X)))))
    #                    = log (e^(max(X)) * sum(e ^ (X_i - max(X))))
    #                    = max(X) + log (sum(e ^ (X_i - max(X)))) = m + log d
    lse = m + tl.log(d)

    # 4. [Online Softmax] Second pass: compute gradients
    # For 'mean' reduction, gradients are normalized by number of non-ignored elements (N)
    # dx_y = (softmax(x_y) - 1) / N
    # dx_i = softmax(x_i) / N, i != y
    # For label smoothing:
    # dx_i = (softmax(x_i) - label_smoothing / V) / N, V = n_cols, i != y
    # dx_y = (softmax(x_y) - label_smoothing / V - (1 - label_smoothing)) / N
    #      = dx_i - (1 - label_smoothing) / N
    # With Z loss:
    # dx_i = ((1 + 2 * lse_square_scale * lse) * softmax(x_i) - label_smoothing / V) / N, i != y
    # dx_y = dx_i - (1 - label_smoothing) / N
    # For 'sum' reduction, no normalization is applied:
    # dx_y = softmax(x_y) - 1
    # dx_i = softmax(x_i), for i ≠ y

    for i in range(0, n_cols, BLOCK_SIZE):
        X_offsets = i + tl.arange(0, BLOCK_SIZE)
        X_block = tl.load(X_ptr + X_offsets, mask=X_offsets < n_cols, other=float("-inf"))
        if HAS_SOFTCAPPING:
            intermediate = tanh(X_block / softcap)
            X_block = softcap * intermediate
        # softmax(x_i)
        X_block = tl.exp(X_block - m) / d
        # derivative of z-loss: 2 * lse_square_scale * lse * softmax(x_i)
        X_block += 2 * lse_square_scale * lse * X_block
        # smoothing term
        X_block += -eps
        # special handle dx_y
        X_block = tl.where(X_offsets != y, X_block, X_block - (1 - label_smoothing))
        # reduction scale
        if reduction == "mean":
            X_block = X_block / (n_non_ignore)
        # chain rule
        # d(softcap * tanh(x / softcap)) = (1 - tanh^2(x / softcap))
        if HAS_SOFTCAPPING:
            X_block = X_block * (1 - intermediate * intermediate)

        tl.store(X_ptr + X_offsets, X_block, mask=X_offsets < n_cols)

    # We need tl.debug_barrier() to ensure the new result of X_ptr is written as mentioned in
    # https://github.com/triton-lang/triton/blob/ba42a5c68fd0505f8c42f4202d53be0f8d9a5fe0/python/triton/ops/cross_entropy.py#L34
    tl.debug_barrier()

    # 5. Calculate the loss

    # loss = log (softmax(X_y)) = log ((e ^ (X_y - max(X)) / sum(e ^ (X - max(X))))
    #      = (X_y - max(X)) - log(sum(e ^ (X - max(X))))
    #      = X_y - m - log d = X_y - lse
    # sum(e ^ (X - max(X))) must >= 1 because the max term is e ^ 0 = 1
    # So we can safely calculate log (softmax(X_y)) without overflow
    loss = lse - ori_X_y

    # Original loss = H(q, p),  with label smoothing regularization = H(q', p) and (label_smoothing / V) = eps
    # H(q', p) = (1 - label_smoothing) * H(q, p) + label_smoothing * H(u, p)
    #          = (1 - label_smoothing) * H(q, p) + eps * sum(logsoftmax(x_i))
    # By using m (global max of xi) and d (sum of e^(xi-m)), we can simplify as:
    #          = (1 - label_smoothing) * H(q, p) + (sum(-eps * x_i) + label_smoothing * (m + logd))
    # Refer to H(q', p) in section 7 of the paper: https://arxiv.org/pdf/1512.00567
    # pytorch: https://github.com/pytorch/pytorch/blob/2981534f54d49fa3a9755c9b0855e7929c2527f0/aten/src/ATen/native/LossNLL.cpp#L516
    # See full derivation at https://github.com/linkedin/Liger-Kernel/pull/198#issuecomment-2333753087
    if label_smoothing > 0:
        smooth_loss = scaled_x_sum + label_smoothing * lse
        loss = loss * (1 - label_smoothing) + smooth_loss

    # An auxiliary loss, z_loss
    # Refer to Page14 Loss function section in the paper PaLM: https://www.jmlr.org/papers/v24/22-1144.html
    z_loss = lse_square_scale * lse * lse
    loss += z_loss
    # Normalize the loss by the number of non-ignored elements if reduction is "mean"
    if reduction == "mean":
        z_loss = z_loss / n_non_ignore
        loss = loss / n_non_ignore

    tl.store(loss_ptr, loss)
    if RETURN_Z_LOSS == _TRUE:
        tl.store(z_loss_ptr, z_loss)


# The hard limit of TRITON_MAX_TENSOR_NUMEL is 1048576 https://github.com/triton-lang/triton/blob/ba42a5c68fd0505f8c42f4202d53be0f8d9a5fe0/python/triton/language/core.py#L19
# However, setting limit as 65536 as in LayerNorm tutorial is faster because of less register spilling
# The optimal maximum block size depends on your hardware, your kernel, and your dtype
MAX_FUSED_SIZE = 65536 // 2  # the best size we found by manually tuning


_bool_to_return_z_loss = {
    True: _TRUE.value,
    False: _FALSE.value,
}


def cross_entropy_forward(
    _input,
    target,
    ignore_index,
    lse_square_scale,
    label_smoothing,
    reduction,
    softcap,
    return_z_loss,
):
    if not isinstance(return_z_loss, int):
        assert return_z_loss in _bool_to_return_z_loss, f"return_z_loss must be True or False. Got: {return_z_loss}"
        return_z_loss = _bool_to_return_z_loss[return_z_loss]
    else:
        assert return_z_loss in _bool_to_return_z_loss, f"return_z_loss must be True or False. Got: {return_z_loss}"

    BT, V = _input.shape
    n_rows = BT

    BLOCK_SIZE = min(MAX_FUSED_SIZE, triton.next_power_of_2(V))

    # unreduced loss
    loss_1d = torch.zeros(n_rows, dtype=_input.dtype, device=_input.device)
    if return_z_loss == _TRUE.value:
        z_loss_1d = torch.zeros(n_rows, dtype=_input.dtype, device=_input.device)
    else:
        z_loss_1d = loss_1d  # dummy ptr when return_z_loss == False

    n_non_ignore = (target != ignore_index).sum().item()

    # ensure _input and target are contiguous in the last dimension
    if _input.stride(-1) != 1:
        _input = _input.contiguous()
    if target.stride(-1) != 1:
        target = target.contiguous()

    # Here we use a trick to store X_ptr gradient in X_ptr so we can save memory
    liger_cross_entropy_kernel[(n_rows,)](
        X_ptr=_input,
        X_stride=_input.stride(-2),
        Y_ptr=target,
        Y_stride=target.stride(-1),  # always 1
        loss_ptr=loss_1d,
        z_loss_ptr=z_loss_1d,
        loss_stride=loss_1d.stride(-1),  # always 1
        n_cols=V,
        n_non_ignore=n_non_ignore,
        ignore_index=ignore_index,
        lse_square_scale=lse_square_scale,
        label_smoothing=label_smoothing,
        reduction=reduction,
        softcap=softcap if softcap is not None else 0.0,
        RETURN_Z_LOSS=return_z_loss,
        BLOCK_SIZE=BLOCK_SIZE,
        HAS_SOFTCAPPING=True if softcap is not None else False,
        # TODO: 32 seems to give the best performance
        # Performance is quite sensitive to num_warps
        num_warps=32 if not is_hip() else 16,
    )

    loss = torch.sum(loss_1d)
    if return_z_loss == _TRUE.value:
        z_loss = torch.sum(z_loss_1d)
    else:
        z_loss = None

    return loss, z_loss, _input


def cross_entropy_backward(_input, grad_output):
    # If cross entropy is the last layer, grad_output is 1.0. Skip the mul to save time
    if torch.equal(grad_output, torch.tensor(1.0, device=grad_output.device)):
        pass

    # We use a Triton kernel instead of a PyTorch operation because modifying inputs in-place
    # for gradient storage and backward multiple times causes anomalies with PyTorch but not with Triton.
    else:
        BT, V = _input.shape
        n_rows = BT
        BLOCK_SIZE = min(MAX_FUSED_SIZE, triton.next_power_of_2(V))

        element_mul_kernel[(n_rows,)](
            _input,
            _input.stride(-2),
            grad_output,
            V,
            BLOCK_SIZE=BLOCK_SIZE,
            num_warps=32 if not is_hip() else 16,
        )

    return _input


class LigerCrossEntropyFunction(torch.autograd.Function):
    """
    This class implements a custom autograd function for the Liger Cross Entropy loss.
    It overrides the forward and backward methods of the torch.autograd.Function class.
    """

    @staticmethod
    def forward(
        ctx,
        _input: torch.Tensor,
        target: torch.Tensor,
        ignore_index: int = -100,
        lse_square_scale: float = 0.0,
        label_smoothing: float = 0.0,
        reduction: str = "mean",
        softcap: Optional[float] = None,
        return_z_loss: bool = False,
    ):
        """
        The forward pass of the Liger Cross Entropy loss.

        Parameters:
        ctx : The context object.
        _input (tensor): The input tensor of shape (BT, V) where B is batch size, T is sequence length, V is vocab size.
        target (tensor): The target tensor of shape (BT) where each value is in [0, V-1].
        ignore_index (int): The index to ignore in the target.
        lse_square_scale (float): The scaler of (logsumexp(_input)) ^ 2 adding to the loss for the stability of training.
        label_smoothing (float): The amount of smoothing when computing the loss, where 0.0 means no smoothing.
        reduction (str): The reduction to apply to the output: "none" | "mean | "sum".
        softcap (Optional[float]): The upper threshold for scaling logits to the range (-softcap, +softcap).
        return_z_loss (bool): When `return_z_loss` is `True`, returns (loss, z_loss) instead of (loss, None). Default: `False`

        Returns:
        tuple: A tuple with the compouted losses with respect to loss and z loss. The elements are tensors or None.
        """
        loss, z_loss, _input = cross_entropy_forward(
            _input,
            target,
            ignore_index,
            lse_square_scale,
            label_smoothing,
            reduction,
            softcap,
            return_z_loss,
        )
        # TODO: investigation
        # If we don't detach the _input tensor, the memory will double
        # Not sure why but seems that there will be a time both grad and value exist but in different location
        ctx.save_for_backward(_input.detach())
        ctx.return_z_loss = return_z_loss

        return loss, z_loss

    @staticmethod
    def backward(ctx, grad_output, grad_ouput2):
        """
        The backward pass of the Liger Cross Entropy loss.

        Parameters:
        ctx : The context object with saved tensors.
        grad_output (tensor): The tensor containing the gradient of the loss with respect to the output.
        grad_output2 (tenosr): No use.
        Returns:
        tuple: A tuple with the gradients with respect to the inputs. The elements are tensors or None.
        """
        if ctx.return_z_loss:
            del grad_ouput2  # z_loss is only for logging

        (_input,) = ctx.saved_tensors
        _input = cross_entropy_backward(_input, grad_output)
        return (
            _input,
            None,
            None,
            None,
            None,
            None,
            None,
            None,
        )


def liger_fixed_cross_entropy(source, target, num_items_in_batch: int = None, ignore_index: int = -100, **kwargs):
    reduction = "sum" if num_items_in_batch is not None else "mean"
    # loss = nn.functional.cross_entropy(source, target, ignore_index=ignore_index, reduction=reduction)
    loss, _ = LigerCrossEntropyFunction.apply(source, target, ignore_index, 0.0, 0.0, reduction)
    if reduction == "sum":
        loss = loss / num_items_in_batch
    return loss


def LigerForCausalLMLoss(
    logits, labels, vocab_size: int, num_items_in_batch: int = None, ignore_index: int = -100, **kwargs
):
    # Upcast to float if we need to compute the loss to avoid potential precision issues
    logits = logits.float()
    # Shift so that tokens < n predict n
    shift_logits = logits[..., :-1, :].contiguous()
    shift_labels = labels[..., 1:].contiguous()

    # Flatten the tokens
    shift_logits = shift_logits.view(-1, vocab_size)
    shift_labels = shift_labels.view(-1)
    # Enable model parallelism
    shift_labels = shift_labels.to(shift_logits.device)
    loss = liger_fixed_cross_entropy(shift_logits, shift_labels, num_items_in_batch, ignore_index, **kwargs)
    return loss