heavyball 0.14.7__py3-none-any.whl → 0.15.1__py3-none-any.whl

heavyball/precond_schedule_palm_foreach_soap.py

@@ -44,76 +44,61 @@ class PrecondSchedulePaLMForeachSOAP(StatefulOptimizer):
         self._data_format = data_format
         self.rng = random.Random(0x120983109)
 
-    @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 = group["beta"]
-
-            beta2 = 1 - max(step, 1) ** -group['beta2_scale']
-            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'])
-
-            update_precond = precond_schedule(step, group['precond_scheduler'], self.rng)
-            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
-                exp_avg_projected = exp_avg_projected / d
-                set_(d, project(exp_avg_projected, state['Q'], True))
-
-                update_preconditioner(g, state, max_precond_dim, precondition_1d, old_debiased2,
-                                      update_precond)
-
-            # 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 = group["beta"]
+
+        beta2 = 1 - max(step, 1) ** -group['beta2_scale']
+        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'])
+
+        update_precond = precond_schedule(step, group['precond_scheduler'], self.rng)
+        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
+            exp_avg_projected = exp_avg_projected / d
+            set_(d, project(exp_avg_projected, state['Q'], True))
+
+            update_preconditioner(g, state, max_precond_dim, precondition_1d, old_debiased2, update_precond)
+
+        # 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/precond_schedule_sfpsoap.py

@@ -54,85 +54,73 @@ class PrecondScheduleSFPaLMSOAP(ScheduleFree):
         self._data_format = data_format
         self.rng = random.Random(0x120983109)
 
-    @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 = []
-            max_precond_dim = group['max_precond_dim']
-            precondition_1d = group['precondition_1d']
-
-            step = group['step'] = group.get("step", -1) + 1
-
-            for p in group["params"]:
-                if p.grad is None:
-                    continue
-                grad = p.grad.float()
-                vals.append((p, grad))
-
-            p_list, grad = zip(*vals)
-            vals = []
-
-            # adaptive gradient clipping
-            adaptive_gradient_clipping_(p_list, grad, group["gradient_clip_val"], eps=group["eps"])
-
-            for p, g in split_p_and_g_in_group(group):
-                state = self.state_(p)
-
-                if "z" not in state:
-                    state["z"] = torch.clone(p.data)
-                    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)
-                z, exp_avg_sq = state["z"], state["exp_avg_sq"]
-                vals.append((p, g, grad_projected, z, exp_avg_sq))
-
-            if not vals:
+    def _step(self, group):
+        vals = []
+        max_precond_dim = group['max_precond_dim']
+        precondition_1d = group['precondition_1d']
+
+        step = group['step'] = group.get("step", -1) + 1
+
+        for p in group["params"]:
+            if p.grad is None:
                 continue
+            grad = p.grad.float()
+            vals.append((p, grad))
+
+        if not vals:
+            return
+
+        p_list, grad = zip(*vals)
+        vals = []
+
+        # adaptive gradient clipping
+        adaptive_gradient_clipping_(p_list, grad, group["gradient_clip_val"], eps=group["eps"])
+
+        for p, g in split_p_and_g_in_group(group):
+            state = self.state_(p)
+
+            if "z" not in state:
+                state["z"] = torch.clone(p.data)
+                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)
+            z, exp_avg_sq = state["z"], state["exp_avg_sq"]
+            vals.append((p, g, grad_projected, z, exp_avg_sq))
 
-            p_list, grad, grad_projected, z, exp_avg_sq = zip(*vals)
-            del vals
+        if not vals:
+            return
 
-            beta2 = 1 - max(step, 1) ** -group['beta2_scale']
-            old_debiased2 = beta_debias(beta2, step)
+        p_list, grad, grad_projected, z, exp_avg_sq = zip(*vals)
+        del vals
 
-            # Decay the first and second moment running average coefficient
-            # In-place operations to update the averages at the same time
-            denom = exp_avg_sq_(exp_avg_sq, grad_projected, old_debiased2, group['eps'])
-            torch._foreach_div_(grad_projected, denom)
+        beta2 = 1 - max(step, 1) ** -group['beta2_scale']
+        old_debiased2 = beta_debias(beta2, step)
 
-            update_precond = precond_schedule(step, group['precond_scheduler'], self.rng)
+        # Decay the first and second moment running average coefficient
+        # In-place operations to update the averages at the same time
+        denom = exp_avg_sq_(exp_avg_sq, grad_projected, old_debiased2, group['eps'])
+        torch._foreach_div_(grad_projected, denom)
 
-            for p, g, gp in zip(p_list, grad, grad_projected):
-                state = self.state_(p)
-                # Projecting back the preconditioned (by Adam) exponential moving average of gradients
-                # to the original space
-                set_(gp, project(gp, state['Q'], back=True))
+        update_precond = precond_schedule(step, group['precond_scheduler'], self.rng)
 
-                update_preconditioner(g, state, max_precond_dim, precondition_1d, old_debiased2,
-                                      update_precond)
+        for p, g, gp in zip(p_list, grad, grad_projected):
+            state = self.state_(p)
+            # Projecting back the preconditioned (by Adam) exponential moving average of gradients
+            # to the original space
+            set_(gp, project(gp, state['Q'], back=True))
 
-            # Weight decay calculated at y
-            if group["weight_decay"] > 0:
-                torch._foreach_add_(grad, p_list, alpha=group["weight_decay"])
+            update_preconditioner(g, state, max_precond_dim, precondition_1d, old_debiased2,
+                                  update_precond)
 
-            lr = warmup(group['lr'], step, group['warmup_steps'])
-            group['weight_sum'] = schedule_free_(lr, group['weight_lr_power'], group['weight_sum'], group['beta'],
-                                                 p_list, z, grad_projected, group['r'], step)
+        # Weight decay calculated at y
+        if group["weight_decay"] > 0:
+            torch._foreach_add_(grad, p_list, alpha=group["weight_decay"])
 
-        return loss
+        lr = warmup(group['lr'], step, group['warmup_steps'])
+        group['weight_sum'] = schedule_free_(lr, group['weight_lr_power'], group['weight_sum'], group['beta'],
+                                             p_list, z, grad_projected, group['r'], step)

heavyball/psgd_kron.py

@@ -4,9 +4,10 @@ Modified under Creative Commons Attribution 4.0 International
 Source available at https://github.com/evanatyourservice/kron_torch/blob/97a2b5ee8a1a4c29e4780bbf6c521e545189eff9/kron_torch/kron.py
 """
 
-import torch
 from typing import Optional
 
+import torch
+
 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, line_to_triu, triu_to_line, set_
 
@@ -37,7 +38,7 @@ class ForeachPSGDKron(PSGDBase):
     def __init__(self, params, lr=0.001, beta=0.9, weight_decay=0.0, preconditioner_update_probability=None,
                  max_size_triangular=2048, min_ndim_triangular=2, memory_save_mode=None,
                  momentum_into_precond_update=True, warmup_steps: int = 1, merge_dims: bool = False,
-                 split: bool = False, clip_fn: Optional[callable] = None):
+                 split: bool = False, clip_fn: Optional[callable] = None, store_triu_as_line: bool = True):
         if not 0.0 <= lr:
             raise ValueError(f"Invalid learning rate: {lr}")
         if not 0.0 <= beta < 1.0:
@@ -57,18 +58,13 @@ class ForeachPSGDKron(PSGDBase):
                         momentum_into_precond_update=momentum_into_precond_update, precond_lr=0.1,
                         # precond lr hardcoded to 0.1
                         precond_init_scale=1.0,  # precond init scale hardcoded to 1.0
-                        step=0, warmup_steps=warmup_steps, merge_dims=merge_dims, split=split)
+                        step=0, warmup_steps=warmup_steps, merge_dims=merge_dims, split=split,
+                        store_triu_as_line=store_triu_as_line)
         super().__init__(params, defaults)
 
         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):
@@ -76,54 +72,52 @@ class ForeachPSGDKron(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']
+        store_triu_as_line = group['store_triu_as_line']
 
-            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 store_triu_as_line else 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
 
-            grad_list, Q_list, exp_avg_list = list(grad_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)
+        torch._foreach_lerp_(exp_avg_list, grad_list, (1 - beta) / (1 - beta ** group["step"]))
 
-                self.balance(do_update, [g], [q])
-                if do_update:
-                    self.do_update([p], [ea if momentum_into_precond_update else g], [q], precond_lr, [q_orig])
-                set_(g, psgd_precond_grad(q, self.state_(p)["exprs"], ea))
+        grad_list, Q_list, exp_avg_list = list(grad_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) if store_triu_as_line else q_orig
 
-            grad_list = self.clip_fn(grad_list)
+            if do_update:
+                self.balance([g], [q])
+                self.do_update([p], [ea if momentum_into_precond_update else g], [q], precond_lr, [q_orig] if store_triu_as_line else None)
+            set_(g, psgd_precond_grad(q, self.state_(p)["exprs"], ea))
 
-            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/pure_psgd.py

@@ -36,7 +36,7 @@ class ForeachPurePSGD(PSGDBase):
     def __init__(self, params, lr=0.001, weight_decay=0.0, preconditioner_update_probability=None,
                  max_size_triangular=2048, min_ndim_triangular=2, memory_save_mode=None,
                  momentum_into_precond_update=True, warmup_steps: int = 1, merge_dims: bool = False,
-                 split: bool = False, clip_fn: callable = None):
+                 split: bool = False, clip_fn: callable = None, store_triu_as_line: bool = True):
         if not 0.0 <= lr:
             raise ValueError(f"Invalid learning rate: {lr}")
         if not 0.0 <= weight_decay:
@@ -54,18 +54,13 @@ class ForeachPurePSGD(PSGDBase):
                         momentum_into_precond_update=momentum_into_precond_update, precond_lr=0.1,
                         # precond lr hardcoded to 0.1
                         precond_init_scale=1.0,  # precond init scale hardcoded to 1.0
-                        step=0, warmup_steps=warmup_steps, merge_dims=merge_dims, split=split)
+                        step=0, warmup_steps=warmup_steps, merge_dims=merge_dims, split=split,
+                        store_triu_as_line=store_triu_as_line)
         super().__init__(params, defaults)
 
         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):
@@ -73,48 +68,46 @@ class ForeachPurePSGD(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']
+        store_triu_as_line = group['store_triu_as_line']
 
-            for p, g in split_p_and_g_in_group(group):
-                state = self.state_(p)
+        vals = []
 
-                if 'Q' not in state:
-                    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["Q"]))
+            if 'Q' not in state:
+                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 store_triu_as_line else Q
 
-            if not vals:
-                continue
+            vals.append((p, g, state["Q"]))
 
-            p_list, grad_list, Q_list = zip(*vals)
-            del vals
+        if not vals:
+            return
 
-            group["step"] += 1
+        p_list, grad_list, Q_list = zip(*vals)
+        del vals
 
-            Q_list = list(Q_list)
-            for i, (p, g) in enumerate(zip(p_list, grad_list)):
-                q_orig = Q_list.pop(0)
-                q = line_to_triu(q_orig)
+        group["step"] += 1
 
-                self.balance(do_update, [g], [q])
-                if do_update:
-                    self.do_update([p], [g], [q], precond_lr, [q_orig])
-                psgd_precond_grad(q, self.state_(p)["exprs"], g, inplace=True)
+        Q_list = list(Q_list)
+        for i, (p, g) in enumerate(zip(p_list, grad_list)):
+            q_orig = Q_list.pop(0)
+            q = line_to_triu(q_orig) if store_triu_as_line else q_orig
 
-            grad_list = self.clip_fn(grad_list)
+            if do_update:
+                self.balance([g], [q])
+                self.do_update([p], [g], [q], precond_lr, [q_orig] if store_triu_as_line else None)
+            psgd_precond_grad(q, self.state_(p)["exprs"], g, inplace=True)
 
-            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)