# activation_preconditioned_optimizer.py

import torch
import math
from torch.optim.optimizer import Optimizer

class ActivationPreconditionedOptimizer(Optimizer):
    """
    Implements an Activation-Preconditioned Optimizer with optional normalization,
    scaling, Nesterov momentum, and grafting, inspired by NormalizedSteepestDescent.
    """
    def __init__(self, params, lr=1e-3, betas=(0.9, 0.95), weight_decay=0.1, 
                 scale=0.2, eps=1e-8, use_nesterov=False, use_grafting=False):
        if not 0.0 <= lr:
            raise ValueError(f"Invalid learning rate: {lr}")
        if not 0.0 <= betas[0] < 1.0:
            raise ValueError(f"Invalid beta1 parameter: {betas[0]}")
        if not 0.0 <= betas[1] < 1.0:
            raise ValueError(f"Invalid beta2 parameter: {betas[1]}")
        if not 0.0 <= weight_decay:
            raise ValueError(f"Invalid weight_decay value: {weight_decay}")

        defaults = dict(lr=lr, betas=betas, weight_decay=weight_decay, scale=scale, 
                        eps=eps, use_nesterov=use_nesterov, use_grafting=use_grafting)
        super(ActivationPreconditionedOptimizer, self).__init__(params, defaults)

    @torch.no_grad()
    def step(self, closure=None):
        loss = None
        if closure is not None:
            with torch.enable_grad():
                loss = closure()

        for group in self.param_groups:
            lr = group['lr']
            beta1, beta2_precond = group['betas']
            # For grafting, we use the same beta2 as Adam, which is typically betas[1]
            beta2_grafting = group['betas'][1] 
            weight_decay = group['weight_decay']
            eps = group['eps']
            scale = group['scale']
            use_nesterov = group['use_nesterov']
            use_grafting = group['use_grafting']
            
            for p in group['params']:
                if p.grad is None:
                    continue
                
                if not hasattr(p, 'activation_second_moment'):
                    continue

                grad = p.grad
                if grad.is_sparse:
                    raise RuntimeError('This optimizer does not support sparse gradients.')

                state = self.state[p]

                # State initialization
                if len(state) == 0:
                    state['step'] = 0
                    state['momentum_buffer'] = torch.zeros_like(p, memory_format=torch.preserve_format)
                    state['preconditioner'] = torch.zeros((p.shape[1], p.shape[1]), device=p.device)
                    if use_grafting:
                        state['exp_avg_sq'] = torch.zeros_like(p, memory_format=torch.preserve_format)

                momentum_buffer = state['momentum_buffer']
                preconditioner = state['preconditioner']
                
                state['step'] += 1

                # --- Update momentum and preconditioner ---
                momentum_buffer.mul_(beta1).add_(grad, alpha=1 - beta1)
                
                xxt = p.activation_second_moment
                preconditioner.mul_(beta2_precond).add_(xxt, alpha=1 - beta2_precond)

                # --- Nesterov Momentum Logic ---
                if use_nesterov:
                    Gt = torch.lerp(grad, momentum_buffer, beta1)
                else:
                    Gt = momentum_buffer

                # --- Calculate and Normalize the Update Direction ---
                eps_I = torch.eye(preconditioner.size(0), device=preconditioner.device) * eps
                precond_inv = torch.linalg.inv(preconditioner + eps_I)

                print(f"Preconditioner shape: {preconditioner.shape}, Gt shape: {Gt.shape}")
                
                # update_direction = torch.matmul(precond_inv, Gt)
                update_direction = torch.matmul(Gt, precond_inv)
                update_norm = torch.linalg.norm(update_direction)
                update_direction.div_(update_norm + eps)
                
                # --- Calculate the Update Magnitude (alpha) ---
                alpha = 0.0
                if use_grafting:
                    exp_avg_sq = state['exp_avg_sq']
                    exp_avg_sq.mul_(beta2_grafting).addcmul_(grad, grad, value=1 - beta2_grafting)
                    
                    step = state['step']
                    bias_correction1 = 1 - beta1 ** step
                    bias_correction2 = 1 - beta2_grafting ** step
                    
                    denom = (exp_avg_sq.sqrt() / math.sqrt(bias_correction2)).add_(eps)
                    adam_update_magnitude = torch.linalg.norm((Gt / bias_correction1) / denom)
                    alpha = -lr * adam_update_magnitude
                else:
                    mn = p.numel()
                    alpha = -lr * scale * math.sqrt(mn)

                # --- Apply Updates ---
                if weight_decay > 0.0:
                    p.add_(p, alpha=-lr * weight_decay)
                
                p.add_(update_direction, alpha=alpha)

                delattr(p, 'activation_second_moment')

        return loss