# Copyright 2023-2024 SGLang Team
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Fused operators for normalization layers."""

import logging
from typing import Optional, Tuple, Union

import torch
import torch.nn as nn
import torch.nn.functional as F

from sglang.srt.batch_invariant_ops import (
    is_batch_invariant_mode_enabled,
    rms_norm_batch_invariant,
)
from sglang.srt.environ import envs
from sglang.srt.layers.utils import MultiPlatformOp
from sglang.srt.server_args import get_global_server_args
from sglang.srt.utils import (
    cpu_has_amx_support,
    get_bool_env_var,
    is_cpu,
    is_cuda,
    is_flashinfer_available,
    is_hip,
    is_npu,
    is_xpu,
)

_is_cuda = is_cuda()
_is_flashinfer_available = is_flashinfer_available()
_is_hip = is_hip()
_is_npu = is_npu()
_use_aiter = get_bool_env_var("SGLANG_USE_AITER") and _is_hip
_is_cpu_amx_available = cpu_has_amx_support()
_is_cpu = is_cpu()
_is_xpu = is_xpu()
_flashinfer_layernorm_available = False

if _is_cuda or _is_xpu:
    if _is_flashinfer_available:
        try:
            from flashinfer.norm import layernorm

            _flashinfer_layernorm_available = True
        except (ImportError, AttributeError):
            _flashinfer_layernorm_available = False
    else:
        _flashinfer_layernorm_available = False

    from sgl_kernel import (
        fused_add_rmsnorm,
        gemma_fused_add_rmsnorm,
        gemma_rmsnorm,
        rmsnorm,
    )
_has_vllm_rms_norm = False
if _use_aiter:
    from aiter import rmsnorm2d_fwd as rms_norm
    from aiter import rmsnorm2d_fwd_with_add as fused_add_rms_norm

    _has_vllm_rms_norm = True  # aiter provides the rms_norm functions
elif _is_hip:
    try:
        from vllm._custom_ops import fused_add_rms_norm, rms_norm

        _has_vllm_rms_norm = True
    except ImportError:
        # Fallback: vllm not available, will use forward_native
        _has_vllm_rms_norm = False

logger = logging.getLogger(__name__)

if _is_npu:
    import torch_npu


class RMSNorm(MultiPlatformOp):
    def __init__(
        self,
        hidden_size: int,
        eps: float = 1e-6,
        var_hidden_size: Optional[int] = None,
        cast_x_before_out_mul: bool = False,
        fp32_residual: bool = False,
        has_weight: bool = True,
        weight_dtype: Optional = None,
        override_orig_dtype: Optional = None,
    ) -> None:
        super().__init__()
        self.has_weight = has_weight
        self.cast_x_before_out_mul = cast_x_before_out_mul
        self.fp32_residual = fp32_residual
        self.override_orig_dtype = override_orig_dtype
        if self.has_weight:
            self.weight = nn.Parameter(torch.ones(hidden_size, dtype=weight_dtype))
        else:
            self.weight = torch.ones(hidden_size, dtype=weight_dtype)
        self.variance_epsilon = eps
        self.hidden_size = hidden_size
        self.variance_size_override = (
            None if var_hidden_size == hidden_size else var_hidden_size
        )
        if _use_aiter:
            self._forward_method = self.forward_aiter

    def forward_cuda(
        self,
        x: torch.Tensor,
        residual: Optional[torch.Tensor] = None,
        post_residual_addition: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
        if x.numel() == 0:
            return x
        if self.variance_size_override is not None:
            return self.forward_native(x, residual, post_residual_addition)
        if is_batch_invariant_mode_enabled():
            if (
                residual is not None
                or get_global_server_args().rl_on_policy_target == "fsdp"
            ):
                return self.forward_native(x, residual, post_residual_addition)
            return rms_norm_batch_invariant(
                x,
                self.weight.data,
                self.variance_epsilon,
            )
        if residual is not None:
            # TODO: Ideally we want to have (hidden_states+residual)+post_residual_addition.
            # but right now we can only have hidden_states+(residual+post_residual_addition).
            # (hidden_states+residual)+post_residual_addition != hidden_states+(residual+post_residual_addition),
            # we probably need to add another parameter to fused_add_rmsnorm
            if post_residual_addition is not None:
                residual = residual + post_residual_addition
            fused_add_rmsnorm(x, residual, self.weight.data, self.variance_epsilon)
            return x, residual
        out = rmsnorm(x, self.weight.data, self.variance_epsilon)
        return out

    def forward_npu(
        self,
        x: torch.Tensor,
        residual: Optional[torch.Tensor] = None,
        post_residual_addition: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
        if residual is not None:
            if post_residual_addition is not None:
                residual = residual + post_residual_addition
            out, _, residual_out = torch_npu.npu_add_rms_norm(
                residual, x, self.weight.data, self.variance_epsilon
            )
            return out, residual_out
        return torch_npu.npu_rms_norm(x, self.weight.data, self.variance_epsilon)[0]

    def forward_aiter(
        self,
        x: torch.Tensor,
        residual: Optional[torch.Tensor] = None,
        post_residual_addition: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
        if residual is not None:
            residual_out = torch.empty_like(x)
            output = torch.empty_like(x)
            if post_residual_addition is not None:
                residual = residual + post_residual_addition
            fused_add_rms_norm(
                output,
                x,
                residual,
                residual_out,
                self.weight.data,
                self.variance_epsilon,
            )
            return output, residual_out
        return rms_norm(x, self.weight.data, self.variance_epsilon)

    def forward_hip(
        self,
        x: torch.Tensor,
        residual: Optional[torch.Tensor] = None,
        post_residual_addition: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
        # Fallback to native implementation if vllm is not available
        if not _has_vllm_rms_norm:
            return self.forward_native(x, residual, post_residual_addition)

        if not x.is_contiguous():
            # NOTE: Remove this if aiter kernel supports discontinuous input
            x = x.contiguous()
        if residual is not None:
            out = torch.empty_like(x)
            residual_out = torch.empty_like(x)
            if post_residual_addition is not None:
                residual = residual + post_residual_addition
            fused_add_rms_norm(
                out, x, residual_out, residual, self.weight.data, self.variance_epsilon
            )
            return out, residual_out
        out = torch.empty_like(x)
        rms_norm(out, x, self.weight.data, self.variance_epsilon)
        return out

    def forward_native(
        self,
        x: torch.Tensor,
        residual: Optional[torch.Tensor] = None,
        post_residual_addition: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
        if not x.is_contiguous():
            x = x.contiguous()
        orig_dtype = self.override_orig_dtype or x.dtype
        x = x.to(torch.float32)
        if residual is not None:
            x = x + residual.to(torch.float32)
            if post_residual_addition is not None:
                x = x + post_residual_addition.to(torch.float32)
            if self.fp32_residual:
                residual = x.clone()
            else:
                residual = x.to(orig_dtype)

        hidden_size = x.shape[-1]
        if hidden_size != self.hidden_size:
            raise ValueError(
                "Expected hidden_size to be "
                f"{self.hidden_size}, but found: {hidden_size}"
            )

        if self.variance_size_override is None:
            x_var = x
        else:
            if hidden_size < self.variance_size_override:
                raise ValueError(
                    "Expected hidden_size to be at least "
                    f"{self.variance_size_override}, but found: {hidden_size}"
                )

            x_var = x[..., : self.variance_size_override]

        variance = x_var.pow(2).mean(dim=-1, keepdim=True)
        x = x * torch.rsqrt(variance + self.variance_epsilon)

        if self.cast_x_before_out_mul:
            x = self.weight * x.to(orig_dtype)
        else:
            x = (x * self.weight).to(orig_dtype)

        if residual is None:
            return x
        else:
            return x, residual

    def forward_cpu(
        self,
        x: torch.Tensor,
        residual: Optional[torch.Tensor] = None,
        post_residual_addition: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
        if _is_cpu_amx_available:
            if residual is not None:
                if post_residual_addition is not None:
                    residual = residual + post_residual_addition
                torch.ops.sgl_kernel.fused_add_rmsnorm_cpu(
                    x, residual, self.weight.data, self.variance_epsilon
                )
                return x, residual
            return torch.ops.sgl_kernel.rmsnorm_cpu(
                x, self.weight.data, self.variance_epsilon
            )
        else:
            return self.forward_native(x, residual, post_residual_addition)

    def forward_xpu(
        self,
        x: torch.Tensor,
        residual: Optional[torch.Tensor] = None,
        post_residual_addition: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
        if self.variance_size_override is not None:
            return self.forward_native(x, residual, post_residual_addition)
        if residual is not None:
            if post_residual_addition is not None:
                residual = residual + post_residual_addition
            fused_add_rmsnorm(x, residual, self.weight.data, self.variance_epsilon)
            return x, residual
        out = rmsnorm(x, self.weight.data, self.variance_epsilon)
        return out

    def forward_with_allreduce_fusion(
        self,
        x: torch.Tensor,
        residual: Optional[torch.Tensor] = None,
        post_residual_addition: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
        """
        Forward method with allreduce fusion, prioritizing flashinfer fused operations
        """
        if residual is not None:
            from sglang.srt.distributed import get_tensor_model_parallel_world_size
            from sglang.srt.layers.flashinfer_comm_fusion import (
                flashinfer_allreduce_residual_rmsnorm,
            )

            if get_tensor_model_parallel_world_size() > 1:
                if post_residual_addition is not None:
                    residual = residual + post_residual_addition
                fused_result = flashinfer_allreduce_residual_rmsnorm(
                    input_tensor=x,
                    residual=residual,
                    weight=self.weight,
                    eps=self.variance_epsilon,
                )
                if fused_result[0] is not None:
                    return fused_result

        return self.forward(x, residual, post_residual_addition)


class LayerNorm(MultiPlatformOp):
    def __init__(
        self,
        hidden_size: int,
        eps: float = 1e-6,
        elementwise_affine: bool = True,
        bias: bool = True,
        dtype: torch.dtype = torch.float32,
    ) -> None:
        super().__init__()
        self.hidden_size = hidden_size
        self.variance_epsilon = eps
        self.elementwise_affine = elementwise_affine
        self.use_bias = bias
        self.dtype = dtype

        self.bias = nn.Parameter(torch.zeros(hidden_size, dtype=self.dtype))
        self.weight = nn.Parameter(torch.ones(hidden_size, dtype=self.dtype))

    def forward_cuda(
        self,
        x: torch.Tensor,
    ) -> torch.Tensor:
        if (
            _flashinfer_layernorm_available
            and x.dtype == torch.bfloat16
            and self.dtype == torch.float32
        ):
            return layernorm(x, self.weight, self.bias, self.variance_epsilon)
        else:
            return self.forward_native(x)

    def forward_native(
        self,
        x: torch.Tensor,
    ) -> torch.Tensor:
        weight = self.weight if self.elementwise_affine else None
        bias = self.bias if self.use_bias else None
        orig_dtype = x.dtype
        x = x.to(self.dtype)
        return F.layer_norm(
            x,
            (self.hidden_size,),
            weight=weight,
            bias=bias,
            eps=self.variance_epsilon,
        ).to(orig_dtype)

    def forward_hip(
        self,
        x: torch.Tensor,
    ) -> torch.Tensor:
        return self.forward_native(x)

    def forward_npu(
        self,
        x: torch.Tensor,
    ) -> torch.Tensor:
        return self.forward_native(x)

    def forward_cpu(
        self,
        x: torch.Tensor,
    ) -> torch.Tensor:
        if _is_cpu_amx_available:
            return torch.ops.sgl_kernel.layernorm_cpu(
                x, self.weight.data, self.variance_epsilon
            )
        else:
            return self.forward_native(x)


class GemmaRMSNorm(MultiPlatformOp):
    def __init__(
        self,
        hidden_size: int,
        eps: float = 1e-6,
    ) -> None:
        super().__init__()
        self.weight = nn.Parameter(torch.zeros(hidden_size))
        self.variance_epsilon = eps

        # Re-dispatch
        if _is_hip:
            self._forward_method = self.forward_native

    def _forward_impl(
        self,
        x: torch.Tensor,
        residual: Optional[torch.Tensor] = None,
        post_residual_addition: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
        if residual is not None:
            if post_residual_addition is not None:
                residual = residual + post_residual_addition
            gemma_fused_add_rmsnorm(
                x, residual, self.weight.data, self.variance_epsilon
            )
            return x, residual
        out = gemma_rmsnorm(x, self.weight.data, self.variance_epsilon)
        return out

    def forward_native(
        self,
        x: torch.Tensor,
        residual: Optional[torch.Tensor] = None,
        post_residual_addition: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
        orig_dtype = x.dtype
        if residual is not None:
            if post_residual_addition is not None:
                residual = residual + post_residual_addition
            x = x + residual
            residual = x

        x = x.float()
        variance = x.pow(2).mean(dim=-1, keepdim=True)
        x = x * torch.rsqrt(variance + self.variance_epsilon)
        x = x * (1.0 + self.weight.float())
        x = x.to(orig_dtype)
        return x if residual is None else (x, residual)

    def forward_cuda(
        self,
        x: torch.Tensor,
        residual: Optional[torch.Tensor] = None,
        post_residual_addition: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
        return self._forward_impl(x, residual, post_residual_addition)

    def forward_cpu(
        self,
        x: torch.Tensor,
        residual: Optional[torch.Tensor] = None,
        post_residual_addition: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
        if _is_cpu_amx_available:
            if residual is not None:
                if post_residual_addition is not None:
                    residual = residual + post_residual_addition
                torch.ops.sgl_kernel.gemma_fused_add_rmsnorm_cpu(
                    x, residual, self.weight.data, self.variance_epsilon
                )
                return x, residual
            return torch.ops.sgl_kernel.gemma_rmsnorm_cpu(
                x, self.weight.data, self.variance_epsilon
            )
        return self.forward_native(x, residual, post_residual_addition)

    def forward_npu(
        self,
        x: torch.Tensor,
        residual: Optional[torch.Tensor] = None,
        post_residual_addition: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
        if envs.SGLANG_NPU_FORWARD_NATIVE_GEMMA_RMS_NORM.get():
            return self.forward_native(x, residual)
        if residual is not None:
            if post_residual_addition is not None:
                residual = residual + post_residual_addition
            x = x + residual
            residual = x

        x, _ = torch_npu.npu_gemma_rms_norm(x, self.weight, self.variance_epsilon)
        return x if residual is None else (x, residual)

    def forward_xpu(
        self,
        x: torch.Tensor,
        residual: Optional[torch.Tensor] = None,
        post_residual_addition: Optional[torch.Tensor] = None,
    ) -> Union[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
        return self._forward_impl(x, residual, post_residual_addition)


class Gemma3RMSNorm(MultiPlatformOp):
    def __init__(self, dim: int, eps: float = 1e-6):
        super().__init__()
        self.eps = eps
        self.weight = nn.Parameter(torch.zeros(dim))
        # Re-dispatch

    def _norm(self, x):
        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)

    def forward_native(self, x):
        output = self._norm(x.float())
        # Llama does x.to(float16) * w whilst Gemma3 is (x * w).to(float16)
        # See https://github.com/huggingface/transformers/pull/29402
        output = output * (1.0 + self.weight.float())
        return output.type_as(x)

    def forward_cpu(self, x):
        if _is_cpu_amx_available and x.stride(-1) == 1:
            return torch.ops.sgl_kernel.gemma3_rmsnorm_cpu(x, self.weight, self.eps)
        return self.forward_native(x)

    def forward_cuda(self, x):
        return self.forward_native(x)

    def forward_npu(self, x):
        output, _ = torch_npu.npu_gemma_rms_norm(x, self.weight, self.eps)
        return output

    def extra_repr(self):
        return f"{tuple(self.weight.shape)}, eps={self.eps}"