heavyball 0.14.6__py3-none-any.whl → 0.15.0__py3-none-any.whl

heavyball/__init__.py

@@ -1,3 +1,4 @@
+from .delayed_psgd import ForeachDelayedPSGD
 from .foreach_adamw import ForeachAdamW
 from .foreach_adopt import ForeachADOPT
 from .foreach_laprop import ForeachLaProp
@@ -12,7 +13,6 @@ from .precond_schedule_sfpsoap import PrecondScheduleSFPaLMSOAP
 from .psgd_kron import ForeachPSGDKron
 from .pure_psgd import ForeachPurePSGD
 from .schedule_free_palm_foreach_soap import SFPaLMForeachSOAP
-from .delayed_psgd import ForeachDelayedPSGD
 
 PalmForEachSoap = PaLMForeachSOAP
 

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/foreach_adamw.py

@@ -10,42 +10,32 @@ class ForeachAdamW(StatefulOptimizer):
                         lr_max=-1.0, weight_decay=weight_decay)
         super().__init__(params, defaults)
 
-    def step(self, closure=None):
-        """Performs a single optimization step.
+    def _step(self, group):
+        eps = group['eps']
+        decay = group['weight_decay']
+        k = group['k']
 
-        Arguments:
-            closure (callable, optional): A closure that reevaluates the model
-                and returns the loss.
-        """
+        if not group['train_mode']:
+            raise Exception("Not in train mode!")
 
-        loss = None
-        if closure is not None:
-            loss = closure()
+        active_p = [p for p in group['params'] if p.grad is not None]
 
-        for group in self.param_groups:
-            eps = group['eps']
-            decay = group['weight_decay']
-            k = group['k']
+        if not active_p:
+            return
 
-            if not group['train_mode']:
-                raise Exception("Not in train mode!")
+        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)
 
-            active_p = [p for p in group['params'] if p.grad is not None]
+        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])
 
-            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)
+        # 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)
 
-            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
-        return loss
+        # 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/foreach_adopt.py

@@ -11,51 +11,41 @@ class ForeachADOPT(StatefulOptimizer):
                         lr_max=-1.0, weight_decay=weight_decay)
         super().__init__(params, defaults)
 
-    def step(self, closure=None):
-        """Performs a single optimization step.
-
-        Arguments:
-            closure (callable, optional): A closure that reevaluates the model
-                and returns the loss.
-        """
-
-        loss = None
-        if closure is not None:
-            loss = closure()
-
-        for group in self.param_groups:
-            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]
-
-            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
-        return loss
+    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/foreach_laprop.py

@@ -11,46 +11,36 @@ class ForeachLaProp(StatefulOptimizer):
                         lr_max=-1.0, weight_decay=weight_decay)
         super().__init__(params, defaults)
 
-    def step(self, closure=None):
-        """Performs a single optimization step.
+    def _step(self, group):
+        eps = group['eps']
+        decay = group['weight_decay']
+        k = group['k']
 
-        Arguments:
-            closure (callable, optional): A closure that reevaluates the model
-                and returns the loss.
-        """
+        if not group['train_mode']:
+            raise Exception("Not in train mode!")
 
-        loss = None
-        if closure is not None:
-            loss = closure()
+        active_p = [p for p in group['params'] if p.grad is not None]
 
-        for group in self.param_groups:
-            eps = group['eps']
-            decay = group['weight_decay']
-            k = group['k']
+        if not active_p:
+            return
 
-            if not group['train_mode']:
-                raise Exception("Not in train mode!")
+        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)
 
-            active_p = [p for p in group['params'] if p.grad is not None]
+        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])
 
-            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)
+        # 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
 
-            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])
+        # Normalize grad in-place for memory efficiency
+        lr = -warmup(group['lr'], k + 1, group['warmup_steps'])
+        update_param_(y, exp_avg, lr, decay)
 
-            # 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
-        return loss
+        group['k'] = k + 1

heavyball/foreach_sfadamw.py

@@ -13,52 +13,42 @@ class ForeachSFAdamW(ScheduleFree):
                         foreach=foreach)
         super().__init__(params, defaults)
 
-    def step(self, closure=None):
-        """Performs a single optimization step.
+    def _step(self, group):
+        eps = group['eps']
+        decay = group['weight_decay']
+        k = group['k']
 
-        Arguments:
-            closure (callable, optional): A closure that reevaluates the model
-                and returns the loss.
-        """
+        if not group['train_mode']:
+            raise Exception("Not in train mode!")
 
-        loss = None
-        if closure is not None:
-            loss = closure()
+        active_p = [p for p in group['params'] if p.grad is not None]
 
-        for group in self.param_groups:
-            eps = group['eps']
-            decay = group['weight_decay']
-            k = group['k']
+        if not active_p:
+            return
 
-            if not group['train_mode']:
-                raise Exception("Not in train mode!")
+        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)
 
-            active_p = [p for p in group['params'] if p.grad is not None]
+        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])
 
-            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)
+        # Decay the first moment running average coefficient
+        old_debiased = beta_debias(group['betas'][1], k + 1)
 
-            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 and second moment running average coefficient
+        denom = exp_avg_sq_(exp_avg_sq, grad, old_debiased, eps)
 
-            # Decay the first moment running average coefficient
-            old_debiased = beta_debias(group['betas'][1], k + 1)
+        # Normalize grad in-place for memory efficiency
+        torch._foreach_div_(grad, denom)
 
-            # Decay the first and second moment running average coefficient
-            denom = exp_avg_sq_(exp_avg_sq, grad, old_debiased, eps)
+        # Weight decay calculated at y
+        if decay != 0:
+            torch._foreach_add_(grad, y, alpha=decay)
 
-            # Normalize grad in-place for memory efficiency
-            torch._foreach_div_(grad, denom)
+        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)
 
-            # 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
-        return loss
+        group['k'] = k + 1

heavyball/foreach_soap.py

@@ -34,73 +34,59 @@ class ForeachSOAP(StatefulOptimizer):
         super().__init__(params, defaults)
         self._data_format = data_format
 
-    @torch.no_grad()
-    def step(self, closure=None):
-        """
-        Performs a single optimization step.
-
-        Arguments:
-            closure (`Callable`, *optional*): A closure that reevaluates the model and returns the loss.
-        """
-        if closure is None:
-            loss = None
-        else:
-            loss = closure()
-
-        for group in self.param_groups:
-            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:
-                continue
-
-            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"])
-        return loss
+    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"])