# SPDX-License-Identifier: Apache-2.0
# Derived from vLLM; modified by DeepSeek.

from __future__ import annotations

import torch
import torch.nn as nn

from deepinfer.model_executor.layers.linear import get_fp8_cast_params
from deepinfer.model_executor.layers.tilelang_ops.norm import norm
from deepinfer.model_executor.utils import reconstruct_param


class RMSNorm(nn.Module):
    """Root mean square normalization.

    Computes x -> w * x / sqrt(E[x^2] + eps) where w is the learned weight.
    Refer to https://arxiv.org/abs/1910.07467
    """

    def __init__(
        self,
        hidden_size: int,
        eps: float = 1e-6,
        weight: torch.Tensor | None = None,
        use_fp8: bool = False,
        out_scale: float = 1.0,
    ) -> None:
        super().__init__()
        if weight is None:
            weight = torch.ones(hidden_size)
        assert weight.shape == (hidden_size,)
        self.weight = reconstruct_param(weight)
        self.variance_epsilon = eps
        self.use_fp8 = use_fp8
        self.out_scale = out_scale

    def forward(
        self,
        x: torch.Tensor,
        residual: torch.Tensor | None = None,
        out: torch.Tensor | tuple[torch.Tensor, torch.Tensor] | None = None,
        return_bf16: bool = False,
        fp8_cast_params: dict | None = None,
        extra_sf_layout: str | None = None,
        extra_sf_out: torch.Tensor | None = None,
        extra_sf_block_m: int = 0,
    ) -> torch.Tensor | tuple[torch.Tensor, torch.Tensor]:
        if fp8_cast_params is None:
            fp8_cast_params = get_fp8_cast_params() if self.use_fp8 else {}
        return norm(
            x,
            self.weight,
            residual=residual,
            eps=self.variance_epsilon,
            out=out,
            **fp8_cast_params,
            extra_sf_layout=extra_sf_layout,
            extra_sf_out=extra_sf_out,
            extra_sf_block_m=extra_sf_block_m,
            return_bf16=return_bf16,
            out_scale=self.out_scale,
        )