"""Original Wanda unstructured pruning for nn.Linear (Sun et al., 2023)."""

from __future__ import annotations

import torch
import torch.nn as nn


class WrappedGPT:
    """Accumulate per-input-channel squared L2, matching others/wanda WrappedGPT."""

    def __init__(self, layer: nn.Linear):
        self.layer = layer
        self.dev = layer.weight.device
        self.columns = layer.weight.data.shape[1]
        self.scaler_row = torch.zeros((self.columns), device=self.dev)
        self.nsamples = 0

    def add_batch(self, inp: torch.Tensor, out: torch.Tensor):
        if len(inp.shape) == 2:
            inp = inp.unsqueeze(0)
        tmp = inp.shape[0]
        if isinstance(self.layer, nn.Linear):
            if len(inp.shape) == 3:
                inp = inp.reshape((-1, inp.shape[-1]))
            inp = inp.t()
        self.scaler_row *= self.nsamples / (self.nsamples + tmp)
        self.nsamples += tmp
        inp = inp.type(torch.float32)
        self.scaler_row += torch.norm(inp, p=2, dim=1) ** 2 / self.nsamples


def prune_linear_wanda(linear: nn.Linear, scaler_row: torch.Tensor, sparsity: float) -> float:
    """Zero the lowest |W| * sqrt(scaler) entries in each output row. Return realized sparsity."""
    if not 0.0 <= float(sparsity) < 1.0:
        raise ValueError(f"sparsity must be in [0, 1), got {sparsity}")
    w = linear.weight.data
    metric = torch.abs(w) * torch.sqrt(
        scaler_row.reshape((1, -1)).to(device=w.device, dtype=torch.float32)
    )
    n_prune = int(metric.shape[1] * float(sparsity))
    mask = torch.zeros_like(metric, dtype=torch.bool)
    if n_prune > 0:
        indices = torch.sort(metric, dim=-1, stable=True)[1][:, :n_prune]
        mask.scatter_(1, indices, True)
    w[mask] = 0
    return float(mask.sum()) / float(mask.numel())