heavyball 0.14.6__tar.gz → 0.15.0__tar.gz

{heavyball-0.14.6 → heavyball-0.15.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: heavyball
-Version: 0.14.6
+Version: 0.15.0
 Summary: Efficient optimizers
 Home-page: https://github.com/clashluke/heavyball
 Author: Lucas Nestler

{heavyball-0.14.6 → heavyball-0.15.0}/heavyball/delayed_psgd.py

@@ -5,8 +5,8 @@ Source available at https://github.com/evanatyourservice/kron_torch/blob/97a2b5e
 """
 
 import torch
-
 from heavyball.utils import copy_stochastic_list_
+
 from .utils import update_param_, warmup, psgd_precond_grad, init_Q_exprs, trust_region_clip_, PSGDBase, \
     precond_update_prob_schedule, split_p_and_g_in_group, triu_to_line, line_to_triu, set_
 
@@ -63,13 +63,7 @@ class ForeachDelayedPSGD(PSGDBase):
 
         self._prob_step = 0
 
-    @torch.no_grad()
-    def step(self, closure=None):
-        loss = None
-        if closure is not None:
-            with torch.enable_grad():
-                loss = closure()
-
+    def _step(self, group):
         # update preconditioners all together
         update_prob = self.preconditioner_update_probability
         if callable(update_prob):
@@ -77,55 +71,52 @@ class ForeachDelayedPSGD(PSGDBase):
         do_update = self.rng.random() < update_prob
         self._prob_step += 1
 
-        for group in self.param_groups:
-            momentum_into_precond_update = group.get("momentum_into_precond_update", True)
-            precond_init_scale = group['precond_init_scale']
-            max_size_triangular = group['max_size_triangular']
-            min_ndim_triangular = group['min_ndim_triangular']
-            memory_save_mode = group['memory_save_mode']
-            precond_lr = group['precond_lr']
-            weight_decay = group['weight_decay']
-            lr = group['lr']
-            beta = group['beta']
-
-            vals = []
+        momentum_into_precond_update = group.get("momentum_into_precond_update", True)
+        precond_init_scale = group['precond_init_scale']
+        max_size_triangular = group['max_size_triangular']
+        min_ndim_triangular = group['min_ndim_triangular']
+        memory_save_mode = group['memory_save_mode']
+        precond_lr = group['precond_lr']
+        weight_decay = group['weight_decay']
+        lr = group['lr']
+        beta = group['beta']
 
-            for p, g in split_p_and_g_in_group(group):
-                state = self.state_(p)
+        vals = []
 
-                if 'Q' not in state:
-                    state["exp_avg"] = torch.zeros_like(g)
-                    Q, state["exprs"] = init_Q_exprs(p, precond_init_scale, max_size_triangular, min_ndim_triangular,
-                                                     memory_save_mode, dtype=g.dtype)
-                    state["Q"] = triu_to_line(Q)
+        for p, g in split_p_and_g_in_group(group):
+            state = self.state_(p)
 
-                vals.append((p, g, state["exp_avg"], state["Q"]))
+            if 'Q' not in state:
+                state["exp_avg"] = torch.zeros_like(g)
+                Q, state["exprs"] = init_Q_exprs(p, precond_init_scale, max_size_triangular, min_ndim_triangular,
+                                                 memory_save_mode, dtype=g.dtype)
+                state["Q"] = triu_to_line(Q)
 
-            if not vals:
-                continue
+            vals.append((p, g, state["exp_avg"], state["Q"]))
 
-            p_list, grad_list, exp_avg_list, Q_list = zip(*vals)
-            del vals
+        if not vals:
+            return
 
-            group["step"] += 1
+        p_list, grad_list, exp_avg_list, Q_list = zip(*vals)
+        del vals
 
-            torch._foreach_lerp_(exp_avg_list, grad_list, (1 - beta) / (1 - beta ** group["step"]))
+        group["step"] += 1
 
-            Q_list, exp_avg_list = list(Q_list), list(exp_avg_list)
-            for i, (p, g) in enumerate(zip(p_list, grad_list)):
-                q_orig = Q_list.pop(0)
-                ea = exp_avg_list.pop(0)
-                q = line_to_triu(q_orig)
-                self.balance(do_update, [g], [q])
-                new = psgd_precond_grad(q, self.state_(p)["exprs"], ea)
+        torch._foreach_lerp_(exp_avg_list, grad_list, (1 - beta) / (1 - beta ** group["step"]))
 
-                if do_update:
-                    self.do_update([p], [ea if momentum_into_precond_update else g], [q], precond_lr, [q_orig])
-                set_(g, new)
+        Q_list, exp_avg_list = list(Q_list), list(exp_avg_list)
+        for i, (p, g) in enumerate(zip(p_list, grad_list)):
+            q_orig = Q_list.pop(0)
+            ea = exp_avg_list.pop(0)
+            q = line_to_triu(q_orig)
+            self.balance(do_update, [g], [q])
+            new = psgd_precond_grad(q, self.state_(p)["exprs"], ea)
 
-            grad_list = self.clip_fn(grad_list)
+            if do_update:
+                self.do_update([p], [ea if momentum_into_precond_update else g], [q], precond_lr, [q_orig])
+            set_(g, new)
 
-            lr = -warmup(lr, group['step'], group['warmup_steps'])
-            update_param_(p_list, grad_list, lr, weight_decay)
+        grad_list = self.clip_fn(grad_list)
 
-        return loss
+        lr = -warmup(lr, group['step'], group['warmup_steps'])
+        update_param_(p_list, grad_list, lr, weight_decay)

heavyball-0.15.0/heavyball/foreach_adamw.py

@@ -0,0 +1,41 @@
+import torch
+import torch.optim
+
+from .utils import warmup, exp_avg_sq_, beta_debias, update_param_, StatefulOptimizer
+
+
+class ForeachAdamW(StatefulOptimizer):
+    def __init__(self, params, lr=0.0025, betas=(0.9, 0.99), eps=1e-8, weight_decay=0, warmup_steps=0):
+        defaults = dict(lr=lr, betas=betas, eps=eps, k=0, warmup_steps=warmup_steps, train_mode=True, weight_sum=0.0,
+                        lr_max=-1.0, weight_decay=weight_decay)
+        super().__init__(params, defaults)
+
+    def _step(self, group):
+        eps = group['eps']
+        decay = group['weight_decay']
+        k = group['k']
+
+        if not group['train_mode']:
+            raise Exception("Not in train mode!")
+
+        active_p = [p for p in group['params'] if p.grad is not None]
+
+        if not active_p:
+            return
+
+        for p in active_p:
+            if 'exp_avg' not in self.state_(p):
+                self.state_(p)['exp_avg'] = torch.zeros_like(p.data, dtype=torch.float32)
+                self.state_(p)['exp_avg_sq'] = torch.zeros_like(p.data, dtype=torch.float32)
+
+        y, grad, exp_avg_sq, exp_avg = zip(
+            *[(p.data, p.grad.float(), self.state_(p)['exp_avg_sq'], self.state_(p)['exp_avg']) for p in active_p])
+
+        # Decay the first and second moment running average coefficient
+        torch._foreach_lerp_(exp_avg, grad, 1 - beta_debias(group['betas'][0], k + 1))
+        denom = exp_avg_sq_(exp_avg_sq, grad, beta_debias(group['betas'][1], k + 1), eps)
+
+        # Normalize grad in-place for memory efficiency
+        lr = -warmup(group['lr'], k + 1, group['warmup_steps'])
+        update_param_(y, exp_avg, lr, decay, lambda p, e, l: torch._foreach_addcdiv_(p, e, denom, l))
+        group['k'] = k + 1

heavyball-0.15.0/heavyball/foreach_adopt.py

@@ -0,0 +1,51 @@
+import torch
+import torch.optim
+
+from .utils import warmup, beta_debias, update_param_, StatefulOptimizer
+
+
+class ForeachADOPT(StatefulOptimizer):
+
+    def __init__(self, params, lr=0.0025, betas=(0.9, 0.99), eps=1e-8, weight_decay=0, warmup_steps=0):
+        defaults = dict(lr=lr, betas=betas, eps=eps, k=0, warmup_steps=warmup_steps, train_mode=True, weight_sum=0.0,
+                        lr_max=-1.0, weight_decay=weight_decay)
+        super().__init__(params, defaults)
+
+    def _step(self, group):
+        eps = group['eps']
+        decay = group['weight_decay']
+        k = group['k']
+
+        if not group['train_mode']:
+            raise Exception("Not in train mode!")
+
+        active_p = [p for p in group['params'] if p.grad is not None]
+
+        if not active_p:
+            return
+
+        for p in active_p:
+            if 'exp_avg' not in self.state_(p):
+                self.state_(p)['exp_avg'] = torch.zeros_like(p.data, dtype=torch.float32)
+                self.state_(p)['exp_avg_sq'] = torch.zeros_like(p.data, dtype=torch.float32)
+
+        y, grad, exp_avg_sq, exp_avg = zip(
+            *[(p.data, p.grad.float(), self.state_(p)['exp_avg_sq'], self.state_(p)['exp_avg']) for p in active_p])
+
+        if k > 1:
+            lr = -warmup(group['lr'], k - 1, group['warmup_steps'])
+
+            update_param_(y, exp_avg, lr, decay)
+        if k > 0:
+            beta1 = beta_debias(group['betas'][0], k)
+            denom = torch._foreach_sqrt(exp_avg_sq)
+            torch._foreach_maximum_(denom, eps)
+            torch._foreach_mul_(exp_avg, beta1)
+            torch._foreach_addcdiv_(exp_avg, grad, denom, 1 - beta1)
+
+        beta2 = beta_debias(group['betas'][1], k + 1)
+        torch._foreach_mul_(exp_avg_sq, beta2)
+        torch._foreach_addcmul_(exp_avg_sq, grad, grad, value=1 - beta2)
+        del grad
+
+        group['k'] = k + 1

heavyball-0.15.0/heavyball/foreach_laprop.py

@@ -0,0 +1,46 @@
+import torch
+import torch.optim
+
+from .utils import warmup, exp_avg_sq_, beta_debias, update_param_, StatefulOptimizer
+
+
+class ForeachLaProp(StatefulOptimizer):
+
+    def __init__(self, params, lr=0.0025, betas=(0.9, 0.99), eps=1e-8, weight_decay=0, warmup_steps=1):
+        defaults = dict(lr=lr, betas=betas, eps=eps, k=0, warmup_steps=warmup_steps, train_mode=True, weight_sum=0.0,
+                        lr_max=-1.0, weight_decay=weight_decay)
+        super().__init__(params, defaults)
+
+    def _step(self, group):
+        eps = group['eps']
+        decay = group['weight_decay']
+        k = group['k']
+
+        if not group['train_mode']:
+            raise Exception("Not in train mode!")
+
+        active_p = [p for p in group['params'] if p.grad is not None]
+
+        if not active_p:
+            return
+
+        for p in active_p:
+            if 'exp_avg' not in self.state_(p):
+                self.state_(p)['exp_avg'] = torch.zeros_like(p.data, dtype=torch.float32)
+                self.state_(p)['exp_avg_sq'] = torch.zeros_like(p.data, dtype=torch.float32)
+
+        y, grad, exp_avg_sq, exp_avg = zip(
+            *[(p.data, p.grad.float(), self.state_(p)['exp_avg_sq'], self.state_(p)['exp_avg']) for p in active_p])
+
+        # Decay the first and second moment running average coefficient
+        denom = exp_avg_sq_(exp_avg_sq, grad, beta_debias(group['betas'][1], k + 1), eps)
+        beta1 = beta_debias(group['betas'][0], k + 1)
+        torch._foreach_mul_(exp_avg, beta1)
+        torch._foreach_addcdiv_(exp_avg, grad, denom, 1 - beta1)
+        del grad
+
+        # Normalize grad in-place for memory efficiency
+        lr = -warmup(group['lr'], k + 1, group['warmup_steps'])
+        update_param_(y, exp_avg, lr, decay)
+
+        group['k'] = k + 1

heavyball-0.15.0/heavyball/foreach_sfadamw.py

@@ -0,0 +1,54 @@
+import torch
+import torch.optim
+
+from .utils import schedule_free_, warmup, ScheduleFree, exp_avg_sq_, beta_debias
+
+
+class ForeachSFAdamW(ScheduleFree):
+    def __init__(self, params, lr=0.0025, betas=(0.9, 0.99), eps=1e-8, weight_decay=0, warmup_steps=0, r=0.0,
+                 weight_lr_power=2.0, foreach=hasattr(torch, "_foreach_mul_")):
+
+        defaults = dict(lr=lr, betas=betas, eps=eps, r=r, k=0, warmup_steps=warmup_steps, train_mode=True,
+                        weight_sum=0.0, lr_max=-1.0, weight_lr_power=weight_lr_power, weight_decay=weight_decay,
+                        foreach=foreach)
+        super().__init__(params, defaults)
+
+    def _step(self, group):
+        eps = group['eps']
+        decay = group['weight_decay']
+        k = group['k']
+
+        if not group['train_mode']:
+            raise Exception("Not in train mode!")
+
+        active_p = [p for p in group['params'] if p.grad is not None]
+
+        if not active_p:
+            return
+
+        for p in active_p:
+            if 'z' not in self.state_(p):
+                self.state_(p)['z'] = torch.clone(p.data)
+                self.state_(p)['exp_avg_sq'] = torch.zeros_like(p.data, dtype=torch.float32)
+
+        y, grad, exp_avg_sq, z = zip(
+            *[(p.data, p.grad.float(), self.state_(p)['exp_avg_sq'], self.state_(p)['z']) for p in active_p])
+
+        # Decay the first moment running average coefficient
+        old_debiased = beta_debias(group['betas'][1], k + 1)
+
+        # Decay the first and second moment running average coefficient
+        denom = exp_avg_sq_(exp_avg_sq, grad, old_debiased, eps)
+
+        # Normalize grad in-place for memory efficiency
+        torch._foreach_div_(grad, denom)
+
+        # Weight decay calculated at y
+        if decay != 0:
+            torch._foreach_add_(grad, y, alpha=decay)
+
+        lr = warmup(group['lr'], k + 1, group['warmup_steps'])
+        group['weight_sum'] = schedule_free_(lr, group['weight_lr_power'], group['weight_sum'], group['betas'][0],
+                                             y, z, grad, group['r'], k + 1)
+
+        group['k'] = k + 1

heavyball-0.15.0/heavyball/foreach_soap.py

@@ -0,0 +1,92 @@
+import torch
+
+from .utils import init_preconditioner, update_preconditioner, project, beta_debias, exp_avg_sq_, update_param_, set_, \
+    split_p_and_g_in_group, StatefulOptimizer
+
+
+class ForeachSOAP(StatefulOptimizer):
+    """
+    SFPaLMForeachSOAP
+
+    Sources:
+        Baseline SOAP:
+            SOAP: Improving and Stabilizing Shampoo using Adam
+            Nikhil Vyas, Depen Morwani, Rosie Zhao, Itai Shapira, David Brandfonbrener, Lucas Janson, Sham Kakade
+            https://arxiv.org/abs/2409.11321
+            https://github.com/nikhilvyas/SOAP
+
+        ScheduleFree:
+            The Road Less Scheduled
+            Aaron Defazio, Xingyu Alice Yang, Harsh Mehta, Konstantin Mishchenko, Ahmed Khaled, Ashok Cutkosky
+            https://arxiv.org/abs/2405.15682
+            https://github.com/facebookresearch/schedule_free
+    """
+
+    def __init__(self, params, lr: float = 3e-3, betas=(0.9, 0.95), shampoo_beta: float = 0.95, eps: float = 1e-8,
+                 weight_decay: float = 0.01, precondition_frequency: int = 2, max_precond_dim: int = 2048,  #
+                 merge_dims: bool = True, precondition_1d: bool = False, normalize_grads: bool = False,
+                 data_format: str = "channels_first", correct_bias: bool = True, warmup_steps: int = 1,
+                 split: bool = False):
+        defaults = {"lr": lr, "betas": betas, "shampoo_beta": shampoo_beta, "eps": eps, "weight_decay": weight_decay,
+                    "precondition_frequency": precondition_frequency, "max_precond_dim": max_precond_dim,
+                    "merge_dims": merge_dims, "precondition_1d": precondition_1d, "normalize_grads": normalize_grads,
+                    "correct_bias": correct_bias, 'warmup_steps': warmup_steps, 'split': split}
+        super().__init__(params, defaults)
+        self._data_format = data_format
+
+    def _step(self, group):
+        vals = []
+        step = 0
+
+        max_precond_dim = group['max_precond_dim']
+        precondition_1d = group['precondition_1d']
+
+        for p, g in split_p_and_g_in_group(group):
+            state = self.state_(p)
+            step = state['step'] = state.get("step", -1) + 1
+
+            if "exp_avg" not in state:
+                state["exp_avg"] = torch.zeros_like(g, dtype=torch.float32)
+                state["exp_avg_sq"] = torch.zeros_like(g, dtype=torch.float32)
+                init_preconditioner(g, state, max_precond_dim, precondition_1d)
+                update_preconditioner(g, state, max_precond_dim, precondition_1d, 0, True)
+                continue  # first step is skipped so that we never use the current gradients in the projection.
+
+            # Projecting gradients to the eigenbases of Shampoo's preconditioner
+            # i.e. projecting to the eigenbases of matrices in state['GG']
+            grad_projected = project(g, state['Q'], False)
+            exp_avg, exp_avg_sq = state["exp_avg"], state["exp_avg_sq"]
+            vals.append((p, g, grad_projected, exp_avg, exp_avg_sq))
+
+        if not vals:
+            return 
+
+        p_list, grad, grad_projected, exp_avg, exp_avg_sq = zip(*vals)
+        beta1, beta2 = group["betas"]
+
+        old_debiased1 = beta_debias(beta1, step)
+        old_debiased2 = beta_debias(beta2, step)
+
+        # Decay the first and second moment running average coefficient
+        # In-place operations to update the averages at the same time
+        torch._foreach_mul_(exp_avg, old_debiased1)
+        torch._foreach_add_(exp_avg, grad, alpha=1 - old_debiased1)
+        denom = exp_avg_sq_(exp_avg_sq, grad_projected, old_debiased2, group['eps'])
+
+        for p, g, ea, d in zip(p_list, grad, exp_avg, denom):
+            state = self.state_(p)
+            # Projecting the exponential moving average of gradients to the eigenbases of Shampoo's preconditioner
+            # i.e. projecting to the eigenbases of matrices in state['GG']
+            exp_avg_projected = project(ea, state['Q'], False)
+
+            # Projecting back the preconditioned (by Adam) exponential moving average of gradients
+            # to the original space
+            # CANT DO /= HERE AS EXP_AVG MAY POINT TO THE BUFFER
+            set_(d, project(exp_avg_projected / d, state['Q'], True))
+
+            update_preconditioner(g, state, max_precond_dim, precondition_1d, old_debiased2,
+                                  step > 0 and step % group['precondition_frequency'] == 0)
+
+        # Why does this have to be rebiased here?
+        step_size = -group["lr"] * min(step / group['warmup_steps'], 1)
+        update_param_(p_list, denom, step_size, group["weight_decay"])

{heavyball-0.14.6 → heavyball-0.15.0}/heavyball/p_adam.py

@@ -62,13 +62,7 @@ class ForeachPaLMPAdam(PSGDBase):
 
         self._prob_step = 0
 
-    @torch.no_grad()
-    def step(self, closure=None):
-        loss = None
-        if closure is not None:
-            with torch.enable_grad():
-                loss = closure()
-
+    def _step(self, group):
         # update preconditioners all together
         update_prob = self.preconditioner_update_probability
         if callable(update_prob):
@@ -76,57 +70,54 @@ class ForeachPaLMPAdam(PSGDBase):
         do_update = self.rng.random() < update_prob
         self._prob_step += 1
 
-        for group in self.param_groups:
-            precond_init_scale = group['precond_init_scale']
-            max_size_triangular = group['max_size_triangular']
-            min_ndim_triangular = group['min_ndim_triangular']
-            memory_save_mode = group['memory_save_mode']
-            precond_lr = group['precond_lr']
-            weight_decay = group['weight_decay']
-            lr = group['lr']
-
-            vals = []
+        precond_init_scale = group['precond_init_scale']
+        max_size_triangular = group['max_size_triangular']
+        min_ndim_triangular = group['min_ndim_triangular']
+        memory_save_mode = group['memory_save_mode']
+        precond_lr = group['precond_lr']
+        weight_decay = group['weight_decay']
+        lr = group['lr']
 
-            for p, g in split_p_and_g_in_group(group):
-                state = self.state_(p)
+        vals = []
 
-                if 'Q' not in state:
-                    state['exp_avg'] = torch.zeros_like(g)
-                    state['exp_avg_sq'] = torch.zeros_like(g)
-                    state["Q"], state["exprs"] = init_Q_exprs(p, precond_init_scale, max_size_triangular,
-                                                              min_ndim_triangular, memory_save_mode, dtype=g.dtype)
+        for p, g in split_p_and_g_in_group(group):
+            state = self.state_(p)
 
-                vals.append((p, g, state["Q"], state['exp_avg'], state['exp_avg_sq']))
+            if 'Q' not in state:
+                state['exp_avg'] = torch.zeros_like(g)
+                state['exp_avg_sq'] = torch.zeros_like(g)
+                state["Q"], state["exprs"] = init_Q_exprs(p, precond_init_scale, max_size_triangular,
+                                                          min_ndim_triangular, memory_save_mode, dtype=g.dtype)
 
-            if not vals:
-                continue
+            vals.append((p, g, state["Q"], state['exp_avg'], state['exp_avg_sq']))
 
-            p_list, grad_list, Q_list, exp_avg, exp_avg_sq = zip(*vals)
-            del vals
+        if not vals:
+            return
 
-            group["step"] += 1
+        p_list, grad_list, Q_list, exp_avg, exp_avg_sq = zip(*vals)
+        del vals
 
-            self.balance(do_update, grad_list, Q_list)
-            if do_update:
-                self.do_update(p_list, grad_list, Q_list, precond_lr)
+        group["step"] += 1
 
-            torch._foreach_lerp_(exp_avg, grad_list, 1 - beta_debias(group['beta'], group['step']))
+        self.balance(do_update, grad_list, Q_list)
+        if do_update:
+            self.do_update(p_list, grad_list, Q_list, precond_lr)
 
-            beta2 = 1 - group['step'] ** -group['beta2_scale']
+        torch._foreach_lerp_(exp_avg, grad_list, 1 - beta_debias(group['beta'], group['step']))
 
-            for p, Q, g, ea, eas in zip(p_list, Q_list, grad_list, exp_avg, exp_avg_sq):
-                psgd_precond_grad(Q, self.state_(p)["exprs"], g, inplace=True)
-                ea = psgd_precond_grad(Q, self.state_(p)["exprs"], ea)
-                exp_avg_sq_(eas, g, beta_debias(beta2, group['step']), 1e-8, out=g)
-                torch.div(ea, g, out=g)
-                """
-                divide by g here, because g == denom (from exp_avg_sq_(out=g)), avoids denom allocation
-                divide into g so we can deallocate ea, avoids one allocation (-> less memory than equivalent foreach)
-                """
+        beta2 = 1 - group['step'] ** -group['beta2_scale']
 
-            grad_list = self.clip_fn(grad_list)
+        for p, Q, g, ea, eas in zip(p_list, Q_list, grad_list, exp_avg, exp_avg_sq):
+            psgd_precond_grad(Q, self.state_(p)["exprs"], g, inplace=True)
+            ea = psgd_precond_grad(Q, self.state_(p)["exprs"], ea)
+            exp_avg_sq_(eas, g, beta_debias(beta2, group['step']), 1e-8, out=g)
+            torch.div(ea, g, out=g)
+            """
+            divide by g here, because g == denom (from exp_avg_sq_(out=g)), avoids denom allocation
+            divide into g so we can deallocate ea, avoids one allocation (-> less memory than equivalent foreach)
+            """
 
-            lr = -warmup(lr, group['step'], group['warmup_steps'])
-            update_param_(p_list, grad_list, lr, weight_decay)
+        grad_list = self.clip_fn(grad_list)
 
-        return loss
+        lr = -warmup(lr, group['step'], group['warmup_steps'])
+        update_param_(p_list, grad_list, lr, weight_decay)

heavyball-0.15.0/heavyball/palm_foreach_sfadamw.py

@@ -0,0 +1,56 @@
+import torch
+import torch.optim
+
+from .utils import schedule_free_, warmup, ScheduleFree, exp_avg_sq_, beta_debias
+
+
+class PaLMForeachSFAdamW(ScheduleFree):
+    def __init__(self, params, lr=0.0025, beta=0.9, betas=(None, None), eps=1e-8, weight_decay=0, warmup_steps=0, r=0.0,
+                 weight_lr_power=2.0, beta2_scale: float = 0.8):
+        if betas[0] is not None:
+            beta = betas[0]
+        defaults = dict(lr=lr, beta=beta, eps=eps, r=r, k=0, warmup_steps=warmup_steps, train_mode=True, weight_sum=0.0,
+                        lr_max=-1.0, weight_lr_power=weight_lr_power, weight_decay=weight_decay,
+                        beta2_scale=beta2_scale)
+        super().__init__(params, defaults)
+
+    def _step(self, group):
+        eps = group['eps']
+        decay = group['weight_decay']
+        k = group['k']
+
+        if not group['train_mode']:
+            raise Exception("Not in train mode!")
+
+        active_p = [p for p in group['params'] if p.grad is not None]
+
+        if not active_p:
+            return
+
+        for p in active_p:
+            if 'z' not in self.state_(p):
+                self.state_(p)['z'] = torch.clone(p.data)
+                self.state_(p)['exp_avg_sq'] = torch.zeros_like(p.data, dtype=torch.float32)
+
+        y, grad, exp_avg_sq, z = zip(
+            *[(p.data, p.grad.float(), self.state_(p)['exp_avg_sq'], self.state_(p)['z']) for p in active_p])
+
+        # Decay the first moment running average coefficient
+        beta2 = 1 - (k + 1) ** -group['beta2_scale']
+        old_debiased = beta_debias(beta2, k + 1)
+
+        # Decay the first and second moment running average coefficient
+        denom = exp_avg_sq_(exp_avg_sq, grad, old_debiased, eps)
+
+        # Normalize grad in-place for memory efficiency
+        torch._foreach_div_(grad, denom)
+
+        # Weight decay calculated at y
+        if decay != 0:
+            torch._foreach_add_(grad, y, alpha=decay)
+
+        lr = warmup(group['lr'], k + 1, group['warmup_steps'])
+        group['weight_sum'] = schedule_free_(lr, group['weight_lr_power'], group['weight_sum'], group['beta'], y, z,
+                                             grad, group['r'], k + 1)
+
+        group['k'] = k + 1