adv-optm 0.1.2__tar.gz → 0.1.4__tar.gz

{adv_optm-0.1.2 → adv_optm-0.1.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: adv_optm
-Version: 0.1.2
+Version: 0.1.4
 Summary: A family of highly efficient, lightweight yet powerful optimizers.
 Home-page: https://github.com/Koratahiu/Advanced_Optimizers
 Author: Koratahiu

{adv_optm-0.1.2 → adv_optm-0.1.4}/adv_optm/__init__.py

@@ -14,4 +14,4 @@ __all__ = [
     "Lion_Prodigy_adv",
 ]
 
-__version__ = "0.1.2"
+__version__ = "0.1.4"

{adv_optm-0.1.2 → adv_optm-0.1.4}/adv_optm/optim/AdamW_adv.py

@@ -1,297 +1,297 @@
-import torch
-from typing import Optional
-
-from ..util.BF16_Stochastic_Rounding import add_stochastic_
-from ..util.Effective_Shape import _get_effective_shape
-from ..util.NNMF import _nnmf,_unnmf
-from ..util.OrthoGrad import _orthogonalize_gradient
-from ..util.One_Bit_Boolean import _pack_bools, _unpack_bools
-
-class AdamW_adv(torch.optim.Optimizer):
-    """
-    Implements a factored AdamW algorithm.
-    This is an advanced version of AdamW with optional features like
-    low-rank factorization of optimizer states (SMMF), OrthoGrad, etc.
-
-    Args:
-        params (iterable): iterable of parameters to optimize or dicts defining
-            parameter groups
-        lr (float): learning rate (default: 1e-3)
-        betas (tuple[float, float]): coefficients used for computing running
-            averages of gradient and its square (default: (0.9, 0.999))
-        eps (float): term added to the denominator to improve
-            numerical stability (default: 1e-8)
-        weight_decay (float): weight decay (L2 penalty) (default: 0)
-        vector_reshape (bool): whether to reshape 1D vectors into 2D
-            matrices to apply low-rank compression (default: True).
-        stochastic_rounding (bool): whether to use stochastic
-            rounding for BF16 parameter updates (default: True).
-        use_atan2 (bool): whether to use the atan2 update rule. (default: False)
-        use_grams (bool): whether to use Grams-style updates. (default: False)
-        use_cautious (bool):  whether to use cautious masking to align the gradient's
-            direction with the first moment's.  (default: False)
-        use_orthograd (bool): whether to use OrthoGrad.  (default: False)
-        use_AdEMAMix (bool): whether to enable the AdEMAMix feature. This adds
-            a second, slow-moving average of the momentum (`mt_slow`) which is
-            combined with the primary momentum (`mt`) to stabilize updates,
-            especially in noisy, small-batch settings. If `False`, the
-            optimizer behaves as standard AdamW. (default: False)
-        beta3_ema (float): The decay rate for the slow exponential moving average of
-            the momentum (only used when `use_AdEMAMix` is `True`). A higher
-            value (e.g., 0.9999) gives the EMA a longer memory, making it more
-            stable but slower to adapt. A lower value (e.g., 0.999) is often
-            better for shorter training runs. (default: 0.9999)
-        alpha (float): The mixing coefficient that scales the slow momentum term
-            before it is added to the fast momentum term (`update = mt + alpha * mt_slow`).
-            A higher value increases the stabilizing influence of the slow
-            momentum. (default: 5.0)
-        t_alpha (Optional[int]): The number of steps for a linear warmup of the
-            `alpha` parameter (only used when `use_AdEMAMix` is `True`). This is
-            highly recommended to prevent instability at the beginning of training,
-            as it gradually introduces the stabilizing slow momentum term. During
-            the warmup, `alpha` ramps from 0 to its target value. If `None`,
-            the scheduler is disabled and th
-        factored (bool): whether to use the factorization or disable it to use
-            the uncompressed optimizer. (default: True)
-    """
-
-    def __init__(
-        self,
-        params,
-        lr: float = 1e-3,
-        betas: tuple[float, float] = (0.9, 0.999),
-        eps: float = 1e-8,
-        weight_decay: float = 0.0,
-        vector_reshape: bool = True,
-        stochastic_rounding: bool = True,
-        use_atan2: bool = False,
-        use_cautious: bool = False,
-        use_grams: bool = False,
-        use_orthograd: bool = False,
-        use_AdEMAMix: bool = False,
-        beta3_ema: float = 0.9999,
-        alpha: float = 5.0,
-        t_alpha: int | None = None,
-        factored: bool = True,
-    ):
-        if not (lr >= 0.0):
-            raise ValueError(f"Learning-rate should be >= 0.0. Got {lr}")
-        if not (0.0 <= betas[0] < 1.0 and 0.0 <= betas[1] < 1.0):
-            raise ValueError(f"Betas should be in [0.0, 1.0). Got {betas}")
-        if not (eps >= 0.0):
-            raise ValueError(f"Epsilon should be >= 0.0. Got {eps}")
-        if not (weight_decay >= 0.0):
-            raise ValueError(f"Weight-decay should be >= 0.0. Got {weight_decay}")
-
-        defaults = {
-            "lr": lr, "betas": betas, "eps": eps, "weight_decay": weight_decay,
-            "vector_reshape": vector_reshape, "use_atan2": use_atan2,
-            "use_orthograd": use_orthograd,
-            "beta3_ema": beta3_ema, "alpha": alpha, "t_alpha": t_alpha,
-        }
-        self.stochastic_rounding = stochastic_rounding
-        self.use_cautious = use_cautious
-        self.use_grams = use_grams
-        self.use_AdEMAMix = use_AdEMAMix
-        self.factored = factored
-        super().__init__(params, defaults)
-
-    @property
-    def supports_fused_back_pass(self):
-        return True
-
-    @property
-    def supports_memory_efficient_fp16(self):
-        return True
-
-    @property
-    def supports_flat_params(self):
-        return False
-
-    @torch.no_grad()
-    def step_parameter(self, p: torch.Tensor, group: dict, i: int | None = None):
-        if p.grad is None:
-            return
-
-        grad = p.grad
-        if grad.dtype != torch.float32 and self.factored:
-            grad = grad.float()
-        if group["use_orthograd"]:
-            grad = _orthogonalize_gradient(p, grad)
-        state = self.state[p]
-
-        beta1, beta2 = group['betas']
-
-        # State Initialization
-        if len(state) == 0:
-            state['step'] = 0
-
-            should_factor = (
-                self.factored and
-                not (len(p.shape) == 1 and not group['vector_reshape'])
-            )
-
-            state['factored'] = should_factor
-
-            dtype = torch.float32 if self.factored else p.dtype
-            device = p.device
-
-            if state['factored']:
-                state['effective_shape'] = _get_effective_shape(p.numel())
-                d1, d2 = state['effective_shape']
-
-                # First moment (m)
-                if beta1 > 0:
-                    state['mu_m_nmf'] = torch.zeros(d1, device=device, dtype=dtype) 
-                    state['mv_m_nmf'] = torch.zeros(d2, device=device, dtype=dtype)
-                    if not self.use_grams:
-                        packed_d2 = (d2 + 7) // 8
-                        state['sign'] = torch.zeros((d1, packed_d2), dtype=torch.uint8, device=device)
-                if self.use_AdEMAMix:
-                    state['mu_m_slow_nmf'] = torch.zeros(d1, device=p.device, dtype=dtype) 
-                    state['mv_m_slow_nmf'] = torch.zeros(d2, device=p.device, dtype=dtype)
-                    packed_d2 = (d2 + 7) // 8
-                    state['sign_slow'] = torch.zeros((d1, packed_d2), dtype=torch.uint8, device=p.device)
-                # Second moment (v)
-                state['mu_v_nmf'] = torch.zeros(d1, device=device, dtype=dtype) 
-                state['mv_v_nmf'] = torch.zeros(d2, device=device, dtype=dtype)
-            else:  # Fallback to standard AdamW for non-factored tensors
-                if beta1 > 0:
-                    state['exp_avg'] = torch.zeros_like(p, device=device, dtype=dtype)
-                if self.use_AdEMAMix:
-                    state['exp_avg_slow'] = torch.zeros_like(p, dtype=dtype)
-                state['exp_avg_sq'] = torch.zeros_like(p, device=device, dtype=dtype)
-
-        if self.use_AdEMAMix:
-            beta3_ema = group['beta3_ema']
-            alpha = group['alpha']
-            t_alpha = group['t_alpha']
-            current_step = state['step'] + 1
-            alpha_t = alpha
-            if t_alpha is not None and t_alpha > 0 and current_step < t_alpha:
-                alpha_t = min(current_step * alpha / t_alpha, alpha)
-
-        if state['factored']:
-            d1, d2 = state['effective_shape']
-
-            # Reconstruct momentum from previous step's factors
-            if beta1 > 0:
-                mt = _unnmf((state['mu_m_nmf'], state['mv_m_nmf']))
-                if not self.use_grams:
-                    unpacked_sign = _unpack_bools(state['sign'], original_m=d2)
-                    torch.where(unpacked_sign, mt, -mt, out=mt)
-                    del unpacked_sign
-                # Update momentum in full-size
-                grad_reshaped = grad.view(d1, d2)
-                mt.mul_(beta1).add_(grad_reshaped, alpha=1.0 - beta1)
-                if self.use_grams:
-                    mt.copy_(grad_reshaped.sign() * mt.abs())
-                elif self.use_cautious:
-                    mask = (mt * grad_reshaped > 0).to(grad_reshaped.dtype)
-                    mask.div_(mask.mean().clamp_(min=1e-3))
-                    mt.mul_(mask)
-                    del mask
-
-            vt = _unnmf((state['mu_v_nmf'], state['mv_v_nmf']))
-            vt.mul_(beta2).addcmul_(grad_reshaped, grad_reshaped, value=1.0 - beta2)
-
-            if self.use_AdEMAMix:
-                mt_slow = _unnmf((state['mu_m_slow_nmf'], state['mv_m_slow_nmf']))
-                if state['sign_slow'].dtype != torch.uint8:
-                    state['sign_slow'] = state['sign_slow'].to(torch.uint8)
-                unpacked_sign_slow = _unpack_bools(state['sign_slow'], original_m=d2)
-                torch.where(unpacked_sign_slow, mt_slow, -mt_slow, out=mt_slow)
-                del unpacked_sign_slow
-
-                mt_slow.mul_(beta3_ema).add_(grad_reshaped, alpha=1.0 - beta3_ema)
-                update = mt + (alpha_t * mt_slow) if beta1 > 0 else grad_reshaped + (alpha_t * mt_slow)
-            else:
-                update = mt if beta1 > 0 else grad_reshaped
-            del grad_reshaped
-
-            if group['use_atan2']:
-                a = 1.2732395
-                denom = vt.sqrt()
-                update.atan2_(denom).mul_(a)
-            else:
-                denom = vt.sqrt()
-                update.div_(denom.add_(group['eps']))
-            del denom
-
-            update.view(p.shape).mul_(group['lr'])
-
-            # Compress updated moments and store new factors
-            if beta1 > 0:
-                if not self.use_grams:
-                    state['sign'] = _pack_bools(mt > 0)
-                _nnmf(mt.abs(), out=(state['mu_m_nmf'], state['mv_m_nmf']))
-                del mt
-            if self.use_AdEMAMix:
-                state['sign_slow'] = _pack_bools(mt_slow > 0)
-                _nnmf(mt_slow.abs(), out=(state['mu_m_slow_nmf'], state['mv_m_slow_nmf']))
-                del mt_slow
-            _nnmf(vt, out=(state['mu_v_nmf'], state['mv_v_nmf']))
-            del vt
-
-        else:  # Standard AdamW logic for non-factored tensors
-            exp_avg_sq = state['exp_avg_sq']
-
-            if beta1 > 0:
-                exp_avg = state['exp_avg']
-                exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1)
-                if self.use_grams:
-                    exp_avg = grad.sign() * exp_avg.abs()
-                elif self.use_cautious:
-                    mask = (exp_avg * grad > 0).to(grad.dtype)
-                    mask.div_(mask.mean().clamp_(min=1e-3))
-                    exp_avg.mul_(mask)
-                    del mask
-
-            if self.use_AdEMAMix:
-                exp_avg_slow = state['exp_avg_slow']
-                exp_avg_slow.mul_(beta3_ema).add_(grad, alpha=1 - beta3_ema)
-                update = exp_avg + (alpha_t * exp_avg_slow) if beta1 > 0 else grad + (alpha_t * exp_avg_slow)
-            else:
-                update = exp_avg if beta1 > 0 else grad
-
-            exp_avg_sq.mul_(beta2).addcmul_(grad, grad.conj(), value=1 - beta2)
-
-            if group['use_atan2']:
-                a = 1.2732395
-                denom = exp_avg_sq.sqrt()
-                update.atan2_(denom).mul_(a)
-            else:
-                denom = exp_avg_sq.sqrt()
-                update.div_(denom.add_(group['eps']))
-            del denom
-
-            update.mul_(group['lr'])
-
-        # Decoupled weight decay
-        if group["weight_decay"] != 0:
-            if p.dtype == torch.bfloat16 and self.stochastic_rounding:
-                add_stochastic_(p.data, p.data, alpha=-group["weight_decay"] * group["lr"])
-            else:
-                p.data.add_(p.data, alpha=-group["weight_decay"] * group["lr"])
-
-        if p.dtype == torch.bfloat16 and self.stochastic_rounding:
-            add_stochastic_(p.data, -update)
-        else:
-            p.data.add_(-update)
-        del update
-
-        state['step'] += 1
-
-    @torch.no_grad()
-    def step(self, closure=None):
-        """Performs a single optimization step."""
-        loss = None
-        if closure is not None:
-            with torch.enable_grad():
-                loss = closure()
-
-        for group in self.param_groups:
-            for i, p in enumerate(group['params']):
-                self.step_parameter(p, group, i)
-
+import torch
+from typing import Optional
+
+from ..util.BF16_Stochastic_Rounding import add_stochastic_
+from ..util.Effective_Shape import _get_effective_shape
+from ..util.NNMF import _nnmf,_unnmf
+from ..util.OrthoGrad import _orthogonalize_gradient
+from ..util.One_Bit_Boolean import _pack_bools, _unpack_bools
+
+class AdamW_adv(torch.optim.Optimizer):
+    """
+    Implements a factored AdamW algorithm.
+    This is an advanced version of AdamW with optional features like
+    low-rank factorization of optimizer states (SMMF), OrthoGrad, etc.
+
+    Args:
+        params (iterable): iterable of parameters to optimize or dicts defining
+            parameter groups
+        lr (float): learning rate (default: 1e-3)
+        betas (tuple[float, float]): coefficients used for computing running
+            averages of gradient and its square (default: (0.9, 0.999))
+        eps (float): term added to the denominator to improve
+            numerical stability (default: 1e-8)
+        weight_decay (float): weight decay (L2 penalty) (default: 0)
+        vector_reshape (bool): whether to reshape 1D vectors into 2D
+            matrices to apply low-rank compression (default: True).
+        stochastic_rounding (bool): whether to use stochastic
+            rounding for BF16 parameter updates (default: True).
+        use_atan2 (bool): whether to use the atan2 update rule. (default: False)
+        use_grams (bool): whether to use Grams-style updates. (default: False)
+        use_cautious (bool):  whether to use cautious masking to align the gradient's
+            direction with the first moment's.  (default: False)
+        use_orthograd (bool): whether to use OrthoGrad.  (default: False)
+        use_AdEMAMix (bool): whether to enable the AdEMAMix feature. This adds
+            a second, slow-moving average of the momentum (`mt_slow`) which is
+            combined with the primary momentum (`mt`) to stabilize updates,
+            especially in noisy, small-batch settings. If `False`, the
+            optimizer behaves as standard AdamW. (default: False)
+        beta3_ema (float): The decay rate for the slow exponential moving average of
+            the momentum (only used when `use_AdEMAMix` is `True`). A higher
+            value (e.g., 0.9999) gives the EMA a longer memory, making it more
+            stable but slower to adapt. A lower value (e.g., 0.999) is often
+            better for shorter training runs. (default: 0.9999)
+        alpha (float): The mixing coefficient that scales the slow momentum term
+            before it is added to the fast momentum term (`update = mt + alpha * mt_slow`).
+            A higher value increases the stabilizing influence of the slow
+            momentum. (default: 5.0)
+        t_alpha (Optional[int]): The number of steps for a linear warmup of the
+            `alpha` parameter (only used when `use_AdEMAMix` is `True`). This is
+            highly recommended to prevent instability at the beginning of training,
+            as it gradually introduces the stabilizing slow momentum term. During
+            the warmup, `alpha` ramps from 0 to its target value. If `None`,
+            the scheduler is disabled and th
+        factored (bool): whether to use the factorization or disable it to use
+            the uncompressed optimizer. (default: True)
+    """
+
+    def __init__(
+        self,
+        params,
+        lr: float = 1e-3,
+        betas: tuple[float, float] = (0.9, 0.999),
+        eps: float = 1e-8,
+        weight_decay: float = 0.0,
+        vector_reshape: bool = True,
+        stochastic_rounding: bool = True,
+        use_atan2: bool = False,
+        use_cautious: bool = False,
+        use_grams: bool = False,
+        use_orthograd: bool = False,
+        use_AdEMAMix: bool = False,
+        beta3_ema: float = 0.9999,
+        alpha: float = 5.0,
+        t_alpha: int | None = None,
+        factored: bool = True,
+    ):
+        if not (lr >= 0.0):
+            raise ValueError(f"Learning-rate should be >= 0.0. Got {lr}")
+        if not (0.0 <= betas[0] < 1.0 and 0.0 <= betas[1] < 1.0):
+            raise ValueError(f"Betas should be in [0.0, 1.0). Got {betas}")
+        if not (eps >= 0.0):
+            raise ValueError(f"Epsilon should be >= 0.0. Got {eps}")
+        if not (weight_decay >= 0.0):
+            raise ValueError(f"Weight-decay should be >= 0.0. Got {weight_decay}")
+
+        defaults = {
+            "lr": lr, "betas": betas, "eps": eps, "weight_decay": weight_decay,
+            "vector_reshape": vector_reshape, "use_atan2": use_atan2,
+            "use_orthograd": use_orthograd,
+            "beta3_ema": beta3_ema, "alpha": alpha, "t_alpha": t_alpha,
+        }
+        self.stochastic_rounding = stochastic_rounding
+        self.use_cautious = use_cautious
+        self.use_grams = use_grams
+        self.use_AdEMAMix = use_AdEMAMix
+        self.factored = factored
+        super().__init__(params, defaults)
+
+    @property
+    def supports_fused_back_pass(self):
+        return True
+
+    @property
+    def supports_memory_efficient_fp16(self):
+        return True
+
+    @property
+    def supports_flat_params(self):
+        return False
+
+    @torch.no_grad()
+    def step_parameter(self, p: torch.Tensor, group: dict, i: int | None = None):
+        if p.grad is None:
+            return
+
+        grad = p.grad
+        if grad.dtype != torch.float32 and self.factored:
+            grad = grad.float()
+        if group["use_orthograd"]:
+            grad = _orthogonalize_gradient(p, grad)
+        state = self.state[p]
+
+        beta1, beta2 = group['betas']
+
+        # State Initialization
+        if len(state) == 0:
+            state['step'] = 0
+
+            should_factor = (
+                self.factored and
+                not (len(p.shape) == 1 and not group['vector_reshape'])
+            )
+
+            state['factored'] = should_factor
+
+            dtype = torch.float32 if self.factored else p.dtype
+            device = p.device
+
+            if state['factored']:
+                state['effective_shape'] = _get_effective_shape(p.numel())
+                d1, d2 = state['effective_shape']
+
+                # First moment (m)
+                if beta1 > 0:
+                    state['mu_m_nmf'] = torch.zeros(d1, device=device, dtype=dtype) 
+                    state['mv_m_nmf'] = torch.zeros(d2, device=device, dtype=dtype)
+                    if not self.use_grams:
+                        packed_d2 = (d2 + 7) // 8
+                        state['sign'] = torch.zeros((d1, packed_d2), dtype=torch.uint8, device=device)
+                if self.use_AdEMAMix:
+                    state['mu_m_slow_nmf'] = torch.zeros(d1, device=p.device, dtype=dtype) 
+                    state['mv_m_slow_nmf'] = torch.zeros(d2, device=p.device, dtype=dtype)
+                    packed_d2 = (d2 + 7) // 8
+                    state['sign_slow'] = torch.zeros((d1, packed_d2), dtype=torch.uint8, device=p.device)
+                # Second moment (v)
+                state['mu_v_nmf'] = torch.zeros(d1, device=device, dtype=dtype) 
+                state['mv_v_nmf'] = torch.zeros(d2, device=device, dtype=dtype)
+            else:  # Fallback to standard AdamW for non-factored tensors
+                if beta1 > 0:
+                    state['exp_avg'] = torch.zeros_like(p, device=device, dtype=dtype)
+                if self.use_AdEMAMix:
+                    state['exp_avg_slow'] = torch.zeros_like(p, dtype=dtype)
+                state['exp_avg_sq'] = torch.zeros_like(p, device=device, dtype=dtype)
+
+        if self.use_AdEMAMix:
+            beta3_ema = group['beta3_ema']
+            alpha = group['alpha']
+            t_alpha = group['t_alpha']
+            current_step = state['step'] + 1
+            alpha_t = alpha
+            if t_alpha is not None and t_alpha > 0 and current_step < t_alpha:
+                alpha_t = min(current_step * alpha / t_alpha, alpha)
+
+        if state['factored']:
+            d1, d2 = state['effective_shape']
+
+            # Reconstruct momentum from previous step's factors
+            if beta1 > 0:
+                mt = _unnmf((state['mu_m_nmf'], state['mv_m_nmf']))
+                if not self.use_grams:
+                    unpacked_sign = _unpack_bools(state['sign'], original_m=d2)
+                    torch.where(unpacked_sign, mt, -mt, out=mt)
+                    del unpacked_sign
+                # Update momentum in full-size
+                grad_reshaped = grad.view(d1, d2)
+                mt.mul_(beta1).add_(grad_reshaped, alpha=1.0 - beta1)
+                if self.use_grams:
+                    mt.copy_(grad_reshaped.sign() * mt.abs())
+                elif self.use_cautious:
+                    mask = (mt * grad_reshaped > 0).to(grad_reshaped.dtype)
+                    mask.div_(mask.mean().clamp_(min=1e-3))
+                    mt.mul_(mask)
+                    del mask
+
+            vt = _unnmf((state['mu_v_nmf'], state['mv_v_nmf']))
+            vt.mul_(beta2).addcmul_(grad_reshaped, grad_reshaped, value=1.0 - beta2)
+
+            if self.use_AdEMAMix:
+                mt_slow = _unnmf((state['mu_m_slow_nmf'], state['mv_m_slow_nmf']))
+                if state['sign_slow'].dtype != torch.uint8:
+                    state['sign_slow'] = state['sign_slow'].to(torch.uint8)
+                unpacked_sign_slow = _unpack_bools(state['sign_slow'], original_m=d2)
+                torch.where(unpacked_sign_slow, mt_slow, -mt_slow, out=mt_slow)
+                del unpacked_sign_slow
+
+                mt_slow.mul_(beta3_ema).add_(grad_reshaped, alpha=1.0 - beta3_ema)
+                update = mt + (alpha_t * mt_slow) if beta1 > 0 else grad_reshaped + (alpha_t * mt_slow)
+            else:
+                update = mt if beta1 > 0 else grad_reshaped
+            del grad_reshaped
+
+            if group['use_atan2']:
+                a = 1.2732395
+                denom = vt.sqrt()
+                update.atan2_(denom).mul_(a)
+            else:
+                denom = vt.sqrt()
+                update.div_(denom.add_(group['eps']))
+            del denom
+
+            update.view(p.shape).mul_(group['lr'])
+
+            # Compress updated moments and store new factors
+            if beta1 > 0:
+                if not self.use_grams:
+                    state['sign'] = _pack_bools(mt > 0)
+                _nnmf(mt.abs(), out=(state['mu_m_nmf'], state['mv_m_nmf']))
+                del mt
+            if self.use_AdEMAMix:
+                state['sign_slow'] = _pack_bools(mt_slow > 0)
+                _nnmf(mt_slow.abs(), out=(state['mu_m_slow_nmf'], state['mv_m_slow_nmf']))
+                del mt_slow
+            _nnmf(vt, out=(state['mu_v_nmf'], state['mv_v_nmf']))
+            del vt
+
+        else:  # Standard AdamW logic for non-factored tensors
+            exp_avg_sq = state['exp_avg_sq']
+
+            if beta1 > 0:
+                exp_avg = state['exp_avg']
+                exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1)
+                if self.use_grams:
+                    exp_avg = grad.sign() * exp_avg.abs()
+                elif self.use_cautious:
+                    mask = (exp_avg * grad > 0).to(grad.dtype)
+                    mask.div_(mask.mean().clamp_(min=1e-3))
+                    exp_avg.mul_(mask)
+                    del mask
+
+            if self.use_AdEMAMix:
+                exp_avg_slow = state['exp_avg_slow']
+                exp_avg_slow.mul_(beta3_ema).add_(grad, alpha=1 - beta3_ema)
+                update = exp_avg + (alpha_t * exp_avg_slow) if beta1 > 0 else grad + (alpha_t * exp_avg_slow)
+            else:
+                update = exp_avg if beta1 > 0 else grad
+
+            exp_avg_sq.mul_(beta2).addcmul_(grad, grad.conj(), value=1 - beta2)
+
+            if group['use_atan2']:
+                a = 1.2732395
+                denom = exp_avg_sq.sqrt()
+                update.atan2_(denom).mul_(a)
+            else:
+                denom = exp_avg_sq.sqrt()
+                update.div_(denom.add_(group['eps']))
+            del denom
+
+            update.mul_(group['lr'])
+
+        # Decoupled weight decay
+        if group["weight_decay"] != 0:
+            if p.dtype == torch.bfloat16 and self.stochastic_rounding:
+                add_stochastic_(p.data, p.data, alpha=-group["weight_decay"] * group["lr"])
+            else:
+                p.data.add_(p.data, alpha=-group["weight_decay"] * group["lr"])
+
+        if p.dtype == torch.bfloat16 and self.stochastic_rounding:
+            add_stochastic_(p.data, -update)
+        else:
+            p.data.add_(-update)
+        del update
+
+        state['step'] += 1
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        """Performs a single optimization step."""
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        for group in self.param_groups:
+            for i, p in enumerate(group['params']):
+                self.step_parameter(p, group, i)
+
         return loss