heavyball 0.21.7__py3-none-any.whl → 0.22.0__py3-none-any.whl

heavyball/precond_schedule_palm_foreach_soap.py

@@ -3,7 +3,7 @@ import random
 import torch
 
 from .utils import init_preconditioner, update_preconditioner, project, beta_debias, exp_avg_, update_param_, \
-    precond_schedule, set_, split_p_and_g_in_group, StatefulOptimizer
+    precond_schedule, set_, StatefulOptimizer
 
 
 class PrecondSchedulePaLMForeachSOAP(StatefulOptimizer):
@@ -33,14 +33,15 @@ class PrecondSchedulePaLMForeachSOAP(StatefulOptimizer):
                  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,
                  precond_scheduler=(1 / 3, 9), betas=(None, None), beta2_scale: float = 0.8, split: bool = False,
-                 foreach: bool = True):
+                 foreach: bool = True, mars: bool = False, caution: bool = False, mars_gamma: float = 0.0025):
         if betas[0] is not None:
             beta = betas[0]
         defaults = {"lr": lr, "beta": beta, "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, 'precond_scheduler': precond_scheduler,
-                    'beta2_scale': beta2_scale, 'split': split}
+                    'beta2_scale': beta2_scale, 'split': split, 'mars': mars, 'caution': caution,
+                    'mars_gamma': mars_gamma}
         super().__init__(params, defaults, foreach)
         self._data_format = data_format
         self.rng = random.Random(0x120983109)
@@ -52,7 +53,7 @@ class PrecondSchedulePaLMForeachSOAP(StatefulOptimizer):
         max_precond_dim = group['max_precond_dim']
         precondition_1d = group['precondition_1d']
 
-        for p, g in split_p_and_g_in_group(group):
+        for p, g in self.split_p_and_g_in_group(group, beta1=group['beta']):
             state = self.state_(p)
             step = state['step'] = state.get("step", -1) + 1
 
@@ -86,6 +87,8 @@ class PrecondSchedulePaLMForeachSOAP(StatefulOptimizer):
         denom = exp_avg_(exp_avg, exp_avg_sq, grad, grad_projected, beta1, beta2, step_tensor)
 
         update_precond = precond_schedule(step, group['precond_scheduler'], self.rng)
+        step_size = -group["lr"] * min(step / group['warmup_steps'], 1)
+
         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
@@ -96,10 +99,7 @@ class PrecondSchedulePaLMForeachSOAP(StatefulOptimizer):
             # 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))
+            precond = 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"])
+            update_param_([p], [precond], step_size, group["weight_decay"], caution=group['caution'], grad=[g])

heavyball/precond_schedule_sfpsoap.py

@@ -3,11 +3,11 @@ import random
 import torch
 
 from .utils import init_preconditioner, update_preconditioner, project, set_, adaptive_gradient_clipping_, exp_avg_sq_, \
-    beta_debias, schedule_free_, warmup, ScheduleFree, precond_schedule, split_p_and_g_in_group, copy_stochastic_list_, \
+    beta_debias, schedule_free_, warmup, ScheduleFree, precond_schedule, copy_stochastic_list_, \
     promote
 
 
-@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
 def _compilable_exp_avg_sq_(exp_avg_sq, grad_projected, old_debiased2, eps):
     eas32, gp32 = [list(map(promote, x)) for x in (exp_avg_sq, grad_projected)]
     denom = exp_avg_sq_(eas32, gp32, old_debiased2, eps)
@@ -52,15 +52,20 @@ class PrecondScheduleSFPaLMSOAP(ScheduleFree):
                  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, r=0.0,
                  weight_lr_power=2.0, gradient_clip_val: float = 0.1, precond_scheduler=(1 / 3, 9), betas=(None, None),
-                 split: bool = False, foreach: bool = True):
+                 split: bool = False, foreach: bool = True, mars: bool = False, caution: bool = False,
+                 mars_gamma: float = 0.0025):
         if betas[0] is not None:
             beta = betas[0]
+
+        assert not caution, "Caution is not implemented in ScheduleFree optimizers"
+
         defaults = {"lr": lr, "beta": beta, "beta2_scale": beta2_scale, "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, 'r': r,
                     'weight_lr_power': weight_lr_power, 'train_mode': True, 'step': -1, 'weight_sum': 0,
-                    'gradient_clip_val': gradient_clip_val, 'precond_scheduler': precond_scheduler, 'split': split}
+                    'gradient_clip_val': gradient_clip_val, 'precond_scheduler': precond_scheduler, 'split': split,
+                    'mars': mars, 'caution': caution, 'mars_gamma': mars_gamma}
         super().__init__(params, defaults, foreach)
         self._data_format = data_format
         self.rng = random.Random(0x120983109)
@@ -87,7 +92,7 @@ class PrecondScheduleSFPaLMSOAP(ScheduleFree):
         # 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):
+        for p, g in self.split_p_and_g_in_group(group, beta1=group['beta']):
             state = self.state_(p)
 
             if "z" not in state:

heavyball/psgd_kron.py

@@ -9,7 +9,7 @@ from typing import Optional
 import torch
 
 from .utils import update_param_, warmup, psgd_precond_grad, init_Q_exprs, trust_region_clip_, PSGDBase, \
-    split_p_and_g_in_group, line_to_triu, triu_to_line, promote, stochastic_lerp_, beta_debias
+    line_to_triu, triu_to_line, promote, stochastic_lerp_, beta_debias
 
 
 class ForeachPSGDKron(PSGDBase):
@@ -40,7 +40,8 @@ class ForeachPSGDKron(PSGDBase):
                  momentum_into_precond_update=True, warmup_steps: int = 1, merge_dims: bool = False,
                  split: bool = False, clip_fn: Optional[callable] = None, store_triu_as_line: bool = True,
                  foreach: bool = True, q_dtype='float32', stochastic_schedule: bool = True,
-                 storage_dtype: str = 'float32',  #
+                 storage_dtype: str = 'float32', mars: bool = False, caution: bool = False, mars_gamma: float = 0.0025,
+                 #
                  # expert parameters
                  precond_init_scale=1.0, precond_lr=0.1):
         if not 0.0 <= lr:
@@ -57,7 +58,9 @@ class ForeachPSGDKron(PSGDBase):
                         min_ndim_triangular=min_ndim_triangular, memory_save_mode=memory_save_mode,
                         momentum_into_precond_update=momentum_into_precond_update, precond_lr=precond_lr,
                         precond_init_scale=precond_init_scale, step=0, warmup_steps=warmup_steps, merge_dims=merge_dims,
-                        split=split, store_triu_as_line=store_triu_as_line, q_dtype=q_dtype, storage_dtype=storage_dtype)
+                        split=split, store_triu_as_line=store_triu_as_line, q_dtype=q_dtype,
+                        storage_dtype=storage_dtype,
+                        mars=mars, caution=caution, mars_gamma=mars_gamma)
         super().__init__(params, defaults, foreach, stochastic_schedule, clip_fn, preconditioner_update_probability)
 
     def _step(self, group):
@@ -77,7 +80,7 @@ class ForeachPSGDKron(PSGDBase):
 
         vals = []
 
-        for p, g in split_p_and_g_in_group(group, should_promote=False):
+        for p, g in self.split_p_and_g_in_group(group, should_promote=False, beta1=beta):
             state = self.state_(p)
 
             if 'Q' not in state:
@@ -114,4 +117,4 @@ class ForeachPSGDKron(PSGDBase):
                 self.do_update(group, [p], [ea if momentum_into_precond_update else g], [q32], precond_lr, [q_orig],
                                store_triu_as_line)
             g = psgd_precond_grad(q, self.state_(p)["exprs"], ea)
-            update_param_([p], self.clip_fn([g]), lr, weight_decay)
+            update_param_([p], self.clip_fn([g]), lr, weight_decay, caution=group['caution'], grad=[g])

heavyball/pure_psgd.py

@@ -5,9 +5,9 @@ Source available at https://github.com/evanatyourservice/kron_torch/blob/97a2b5e
 """
 
 import torch
-
 from heavyball.utils import identity
-from .utils import update_param_, warmup, psgd_precond_grad, init_Q_exprs, PSGDBase, split_p_and_g_in_group, \
+
+from .utils import update_param_, warmup, psgd_precond_grad, init_Q_exprs, PSGDBase, \
     line_to_triu, triu_to_line, promote
 
 
@@ -38,7 +38,8 @@ class ForeachPurePSGD(PSGDBase):
                  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, store_triu_as_line: bool = True, foreach: bool = True,
-                 q_dtype='float32', stochastic_schedule: bool = True,  #
+                 q_dtype='float32', stochastic_schedule: bool = True, mars: bool = False, caution: bool = False,
+                 mars_gamma: float = 0.0025,  #
                  # expert parameters
                  precond_init_scale=1.0, precond_lr=0.1):
         if not 0.0 <= lr:
@@ -49,11 +50,14 @@ class ForeachPurePSGD(PSGDBase):
         if clip_fn is None:
             clip_fn = identity
 
+        assert not mars, "MARS is not supported in this optimizer"
+
         defaults = dict(lr=lr, weight_decay=weight_decay, max_size_triangular=max_size_triangular,
                         min_ndim_triangular=min_ndim_triangular, memory_save_mode=memory_save_mode,
                         momentum_into_precond_update=momentum_into_precond_update, precond_lr=precond_lr,
                         precond_init_scale=precond_init_scale, step=0, warmup_steps=warmup_steps, merge_dims=merge_dims,
-                        split=split, store_triu_as_line=store_triu_as_line, q_dtype=q_dtype)
+                        split=split, store_triu_as_line=store_triu_as_line, q_dtype=q_dtype, mars=mars, caution=caution,
+                        mars_gamma=mars_gamma)
         super().__init__(params, defaults, foreach, stochastic_schedule, clip_fn, preconditioner_update_probability)
 
     def _step(self, group):
@@ -70,7 +74,7 @@ class ForeachPurePSGD(PSGDBase):
 
         vals = []
 
-        for p, g in split_p_and_g_in_group(group, should_promote=False):
+        for p, g in self.split_p_and_g_in_group(group, should_promote=False, beta1=0.0):
             state = self.state_(p)
 
             if 'Q' not in state:
@@ -98,4 +102,4 @@ class ForeachPurePSGD(PSGDBase):
                 q32 = [promote(q_) for q_ in q]
                 self.do_update(group, [p], [g], [q32], precond_lr, [q_orig], store_triu_as_line)
             psgd_precond_grad(q, self.state_(p)["exprs"], g, inplace=True)
-            update_param_([p], self.clip_fn([g]), lr, weight_decay)
+            update_param_([p], self.clip_fn([g]), lr, weight_decay, caution=group['caution'], grad=[g])

heavyball/schedule_free_palm_foreach_soap.py

@@ -1,12 +1,13 @@
 import random
 
 import torch
+from heavyball.utils import mars_correction
 
 from .utils import init_preconditioner, update_preconditioner, project, set_, adaptive_gradient_clipping_, exp_avg_sq_, \
-    beta_debias, schedule_free_, warmup, ScheduleFree, split_p_and_g_in_group, copy_stochastic_list_, promote
+    beta_debias, schedule_free_, warmup, ScheduleFree, copy_stochastic_list_, promote
 
 
-@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
 def _compilable_exp_avg_sq_(exp_avg_sq, grad_projected, old_debiased2, eps):
     eas32, gp32 = [list(map(promote, x)) for x in (exp_avg_sq, grad_projected)]
     denom = exp_avg_sq_(eas32, gp32, old_debiased2, eps)
@@ -44,15 +45,19 @@ class SFPaLMForeachSOAP(ScheduleFree):
                  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, r=0.0,
                  weight_lr_power=2.0, gradient_clip_val: float = 0.1, betas=(None, None), split: bool = False,
-                 foreach: bool = True):
+                 foreach: bool = True, mars: bool = False, caution: bool = False, mars_gamma: float = 0.0025):
         if betas[0] is not None:
             beta = betas[0]
+
+        assert not caution, "Caution is not implemented in ScheduleFree optimizers"
+
         defaults = {"lr": lr, "beta": beta, "beta2_scale": beta2_scale, "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, 'r': r,
                     'weight_lr_power': weight_lr_power, 'train_mode': True, 'step': -1,
-                    'gradient_clip_val': gradient_clip_val, 'weight_sum': 0, 'split': split}
+                    'gradient_clip_val': gradient_clip_val, 'weight_sum': 0, 'split': split, 'mars': mars,
+                    'caution': caution, 'mars_gamma': mars_gamma}
         super().__init__(params, defaults, foreach)
         self._data_format = data_format
         self.rng = random.Random(0x120983109)
@@ -61,6 +66,7 @@ class SFPaLMForeachSOAP(ScheduleFree):
         vals = []
         max_precond_dim = group['max_precond_dim']
         precondition_1d = group['precondition_1d']
+        mars = group['mars']
 
         step = group['step'] = group.get("step", 0) + 1
 
@@ -79,12 +85,14 @@ class SFPaLMForeachSOAP(ScheduleFree):
 
         vals = []
 
-        for p, g in split_p_and_g_in_group(group):
+        for p, g in self.split_p_and_g_in_group(group, beta1=group['beta']):
             state = self.state_(p)
 
             if "z" not in state:
                 state["z"] = torch.clone(p).float()
                 state["exp_avg_sq"] = torch.zeros_like(g, dtype=torch.float32)
+                if mars:
+                    state['mars_prev_grad'] = g.clone()
                 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.

heavyball/utils.py

@@ -38,7 +38,7 @@ def warmup(lr: float, step: int, warmup_steps: int):
     return lr * step / warmup_steps
 
 
-@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
 def _compilable_schedule_free_(p, z, ckp1, grad, lr, beta1):
     p32 = promote(p)
     z32 = promote(z)
@@ -141,19 +141,27 @@ def beta_debias(beta, step):
     return 1 - (1 - beta) / (1 - beta ** step)
 
 
-@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
-def exp_avg_sq_(state, grad, beta2, eps, out=None):
-    if isinstance(state, torch.Tensor):
-        state.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
-        return torch.sqrt(state, out=out).clamp_(min=eps)
-
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
+def _compilable_exp_avg_sq_(state, grad, beta2, eps, out=None):
     torch._foreach_mul_(state, beta2)
     [s.addcmul_(g, g, value=1 - beta2) for s, g in zip(state, grad)]
     denom = torch._foreach_sqrt(state)
-    torch._foreach_maximum_(denom, eps)
+    [denom.clamp_(min=eps) for denom in denom]
+    if out is not None:
+        copy_stochastic_list_(out, denom)
+        return out
+
     return denom
 
 
+def exp_avg_sq_(state, grad, beta2, eps, out=None):
+    state, grad = list_guard(state), list_guard(grad)
+    if not isinstance(beta2, torch.Tensor):
+        beta2 = torch.empty((), dtype=torch.float32, device=state[0].device).fill_(beta2)
+    if not isinstance(eps, torch.Tensor):
+        eps = torch.empty((), dtype=torch.float32, device=state[0].device).fill_(eps)
+    return _compilable_exp_avg_sq_(state, grad, beta2, eps, out)
+
 def adaptive_gradient_clipping_(parameters: List[torch.Tensor], gradients: List[torch.Tensor], clip_val: float,
                                 minimum: float = 1e-3, eps: float = 1e-8):
     if clip_val <= 0:
@@ -168,12 +176,19 @@ def adaptive_gradient_clipping_(parameters: List[torch.Tensor], gradients: List[
     torch._foreach_mul_(gradients, p_norm)
 
 
+def is_compiling():
+    try:
+        return torch.compiler.is_compiling()
+    except AttributeError:
+        return True
+
+
 def set_(dst: torch.Tensor, src: torch.Tensor):
-    if not torch.compiler.is_compiling() and src.data_ptr() == dst.data_ptr():
+    if not is_compiling() and src.data_ptr() == dst.data_ptr():
         return
     if src.shape != dst.shape:
         src = src.reshape_as(dst)
-    if not torch.compiler.is_compiling() and src.is_contiguous() and dst.is_contiguous() and src.dtype == dst.dtype:
+    if not is_compiling() and src.is_contiguous() and dst.is_contiguous() and src.dtype == dst.dtype:
         dst.set_(src)
     else:
         dst.copy_(src)
@@ -328,7 +343,7 @@ def get_orthogonal_matrix(mat):
     return final
 
 
-@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
 def _compilable_stochastic_lerp_(x: List[torch.Tensor], y: List[torch.Tensor], a: Union[float, int, torch.Tensor]):
     for x_, y_ in zip(x, y):
         x32 = promote(x_)
@@ -338,12 +353,19 @@ def _compilable_stochastic_lerp_(x: List[torch.Tensor], y: List[torch.Tensor], a
 
 
 def stochastic_lerp_(x: List[torch.Tensor], y: List[torch.Tensor], a: Union[float, int, torch.Tensor]):
+    x, y = list_guard(x), list_guard(y)
     if not isinstance(a, torch.Tensor):
         a = torch.empty((), dtype=torch.float32, device=x[0].device).fill_(a)
     _compilable_stochastic_lerp_(x, y, a)
 
 
-@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
+def list_guard(x):
+    if isinstance(x, (list, tuple)):
+        return x
+    return [x]
+
+
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
 def _compilable_stochastic_add_(x: List[torch.Tensor], y: List[torch.Tensor], alpha: Union[float, int, torch.Tensor]):
     for x_, y_ in zip(x, y):
         x32 = promote(x_)
@@ -353,6 +375,7 @@ def _compilable_stochastic_add_(x: List[torch.Tensor], y: List[torch.Tensor], al
 
 
 def stochastic_add_(x: List[torch.Tensor], y: List[torch.Tensor], alpha: Union[float, int, torch.Tensor]):
+    x, y = list_guard(x), list_guard(y)
     if not isinstance(alpha, torch.Tensor):
         alpha = torch.empty((), dtype=torch.float32, device=x[0].device).fill_(alpha)
     _compilable_stochastic_add_(x, y, alpha)
@@ -463,6 +486,43 @@ class StatefulOptimizer(torch.optim.Optimizer):
     def state_(self, arg: torch.Tensor):
         return self.state[self.key(arg)]
 
+    def mars_correct_list(self, group, p_list, g_list, mars_gamma, beta):
+        for p, g in zip(p_list, g_list):
+            state = self.state_(p)
+            if 'mars_old_grad' not in state:
+                state['mars_old_grad'] = torch.zeros_like(g)
+        old_gs = [self.state_(p)['mars_old_grad'] for p in p_list]
+        mars_correction(g_list, old_gs, mars_gamma, beta)
+
+    def split_p_and_g_in_group(self, group: dict, skip_none: bool = True, should_promote: bool = True,
+                               beta1: float = -1.0):
+        for p in group["params"]:
+            if skip_none and p.grad is None:
+                continue
+
+            if p.grad is None:
+                grad = None
+            else:
+                if should_promote:
+                    grad = promote(p.grad)
+                else:
+                    grad = p.grad
+                if beta1 >= 0 and group.get('mars', False):
+                    self.mars_correct_list(group, [p], [grad], group['mars_gamma'], beta1)
+
+                p.grad = None
+
+            p_views = merge_group(group, p)
+            if grad is not None:
+                grad = merge_group(group, grad)
+            if isinstance(p_views, torch.Tensor):
+                yield p_views, grad
+                continue
+            if grad is None:
+                yield from zip(p_views, [None] * len(p_views))
+                continue
+            yield from zip(p_views, grad)
+
     def state_size(self) -> int:
         total_bytes = 0
 
@@ -472,7 +532,7 @@ class StatefulOptimizer(torch.optim.Optimizer):
                 total_bytes += x.numel() * x.element_size()
 
         for group in self.param_groups:
-            for p, _ in split_p_and_g_in_group(group, skip_none=False):
+            for p, _ in self.split_p_and_g_in_group(group, skip_none=False):
                 tree_map(_add, self.state_(p))
         return total_bytes
 
@@ -581,7 +641,7 @@ def copy_stochastic_list_(target: List[torch.Tensor], source: List[torch.Tensor]
         copy_stochastic_(t, s)
 
 
-@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
 def _compilable_exp_avg_(exp_avg, exp_avg_sq, grad, grad_projected, beta1, beta2, step):
     beta1 = beta_debias(beta1, step)
     beta2 = beta_debias(beta2, step)
@@ -625,22 +685,24 @@ def _compilable_copy_stochastic_(target: torch.Tensor, source: torch.Tensor):
 
 
 def copy_stochastic_(target: torch.Tensor, source: torch.Tensor):
-    if not torch.compiler.is_compiling() and target.data_ptr() == source.data_ptr():
+    if not is_compiling() and target.data_ptr() == source.data_ptr():
         return
     if target.dtype != torch.bfloat16 or source.dtype not in (torch.float16, torch.float32, torch.float64):
         set_(target, source)
     _compilable_copy_stochastic_(target, source)
 
 
-@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
-def _compilable_update_(p, u, decay, add_fn, lr):
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
+def _compilable_update_(p, u, decay, add_fn, lr, caution, g):
     u = [u_.view_as(p_) for u_, p_ in zip(u, p)]
-    p32, u32 = [list(map(promote, x)) for x in [p, u]]
+    p32, u32, g32 = [list(map(promote, x)) for x in [p, u, g]]
 
     if decay > 0:
         torch._foreach_mul_(p32, 1 - decay * lr)
 
-    for p32_, u32_ in zip(p32, u32):  # lr is data-dependent -> can't compile a foreach
+    for p32_, u32_, g32_ in zip(p32, u32, g32):  # lr is data-dependent -> can't compile a foreach
+        if caution:
+            _compilable_cautioning_(g32_, u32_)
         if add_fn is None:
             p32_.add_(u32_, alpha=lr)
         else:
@@ -650,9 +712,12 @@ def _compilable_update_(p, u, decay, add_fn, lr):
 
 
 def update_param_(param: List[torch.Tensor], update: List[torch.Tensor], lr: float, decay: float,
-                  add_fn: callable = None):
+                  add_fn: callable = None, caution: bool = False, grad: List[torch.Tensor] = None):
     lr_tensor = torch.empty((), dtype=torch.float32, device=param[0].device).fill_(lr)
-    _compilable_update_(param, update, decay, add_fn, lr_tensor)
+    param, update, grad = list_guard(param), list_guard(update), list_guard(grad)
+    if not caution:
+        grad = [None] * len(param)
+    _compilable_update_(param, update, decay, add_fn, lr_tensor, caution, grad)
 
 
 def precond_schedule(step, precond_scheduler, rng):
@@ -788,7 +853,7 @@ def psgd_lb(A, max_abs):
     return x
 
 
-@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
 def psgd_update_precond(Q, exprs, G, precond_lr, tiny, oq, store_triu_as_line):
     """Update Kronecker product preconditioner Q with pair (V, G)."""
     exprA, exprGs, _ = exprs
@@ -821,7 +886,7 @@ def psgd_update_precond(Q, exprs, G, precond_lr, tiny, oq, store_triu_as_line):
         stochastic_add_([o], [term1], -1)
 
 
-@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
 def psgd_precond_grad(Q, exprs, G, inplace: bool = False):
     """Precondition gradient G with preconditioner Q."""
     md = min_dtype(Q)
@@ -965,18 +1030,45 @@ class PSGDBase(StatefulOptimizer):
                         psgd_balance_Q(q)
 
 
-#@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
-def _compilable_precond_grad_cached_(cached_q, ea, expr, param, lr, weight_decay, clip_fn):
+# TODO: Figure out why this sometimes crashes
+# @torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
+def _compilable_precond_grad_cached_(cached_q, ea, expr, param, lr, weight_decay, clip_fn, caution, grad):
     md = min_dtype(cached_q + [ea])
     new = torch.einsum(expr, *[c_.to(md) for c_ in cached_q], ea.to(md)).to(torch.float32)
-    update_param_([param], clip_fn([new]), lr, weight_decay)
+    update_param_([param], clip_fn([new]), lr, weight_decay, caution=caution, grad=grad)
 
 
 def precond_grad_cached_(cached_q: List[torch.Tensor], ea: torch.Tensor, expr: str, param: torch.Tensor, lr: float,
-                         weight_decay: float, clip_fn):
+                         weight_decay: float, clip_fn, caution, grad):
     if isinstance(lr, float):
         lr = torch.empty((), dtype=torch.float32, device=param.device).fill_(lr)
-    _compilable_precond_grad_cached_(cached_q, ea, expr, param, lr, weight_decay, clip_fn)
+    _compilable_precond_grad_cached_(cached_q, ea, expr, param, lr, weight_decay, clip_fn, caution, grad)
+
+
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
+def _compilable_mars_correction_(g, old_g, a):
+    g_copy = [g_.clone() for g_ in g]
+    _compilable_stochastic_lerp_(g, old_g, a)
+    copy_stochastic_list_(old_g, g_copy)
+
+
+def mars_correction(g, old_g, beta1, gamma):
+    a = -gamma * beta1 / (1 - beta1)
+    g, old_g = list_guard(g), list_guard(old_g)
+    a = torch.empty((), dtype=torch.float32, device=g[0].device).fill_(a)
+    _compilable_mars_correction_(g, old_g, a)
+
+
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
+def _compilable_cautioning_(g, update):
+    mask = (g * update) > 0
+    update.masked_fill_(~mask, 0)
+    scale = mask.numel() / mask.sum().clamp(min=1)
+    update.mul_(scale)
+
+
+def caution(g, update):
+    _compilable_cautioning_(g, update)
 
 
 def precond_update_prob_schedule(max_prob=1.0, min_prob=0.03, decay=0.001, flat_start=250):
@@ -1013,29 +1105,3 @@ def merge_group(group, *tensors):
         append_or_extend(out, dim_merger(t, group['max_size_triangular'] if 'max_size_triangular' in group else group[
             'max_precond_dim'], group.get('split', False)))
     return out
-
-
-def split_p_and_g_in_group(group: dict, skip_none: bool = True, should_promote: bool = True):
-    for p in group["params"]:
-        if skip_none and p.grad is None:
-            continue
-
-        if p.grad is None:
-            grad = None
-        else:
-            if should_promote:
-                grad = promote(p.grad)
-            else:
-                grad = p.grad
-            p.grad = None
-
-        p_views = merge_group(group, p)
-        if grad is not None:
-            grad = merge_group(group, grad)
-        if isinstance(p_views, torch.Tensor):
-            yield p_views, grad
-            continue
-        if grad is None:
-            yield from zip(p_views, [None] * len(p_views))
-            continue
-        yield from zip(p_views, grad)

{heavyball-0.21.7.dist-info → heavyball-0.22.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: heavyball
-Version: 0.21.7
+Version: 0.22.0
 Summary: Efficient optimizers
 Home-page: https://github.com/clashluke/heavyball
 Author: Lucas Nestler
@@ -32,7 +32,7 @@ A simple package of efficient optimizers
 The goal is not to thrive for completeness, full maintenance or abstraction, but instead to provide a simple
 largely static alternative to `torch.optim` with more and better optimizers.
 
-Currently (2024-11-22, 0.21.0), the recommended stable optimizer is `PrecondSchedulePaLMSOAP` (see below). The
+Currently (2024-11-26, 0.22.0), the recommended stable optimizer is `PrecondSchedulePaLMSOAP` (see below). The
 recommended experimental optimizer is `DelayedPSGDKron` ([tuning guide](docs/psgd_efficiency.md)).
 
 ## Features

heavyball-0.22.0.dist-info/RECORD

@@ -0,0 +1,24 @@
+heavyball/__init__.py,sha256=icHYN-MGsmHkLUlHCMcZkOlwY7GT63_ayR_a5iPKmzM,2226
+heavyball/cached_delayed_psgd_kron.py,sha256=n3wIOhrop0Ls4MZ0kXpwGuImp1jzPs6VGdxIlPyoYdQ,6827
+heavyball/cached_psgd_kron.py,sha256=KCLsfvj9qh_2FNwRTdWM3zjnt2oGHfsf4Y341rPcceI,6778
+heavyball/delayed_psgd.py,sha256=CaG17zqorLsCSDeGEePOyb6n9ugv8W6gyRQqeQNq-e8,6272
+heavyball/foreach_adamw.py,sha256=uawSbGGUD2E1RtcwspP83yQNElERdGX-diqCI5e8FqE,2825
+heavyball/foreach_adopt.py,sha256=DFEaPswVzdHcbxC-mirsf_okM_HR6r34PDUTty5CrUE,3547
+heavyball/foreach_laprop.py,sha256=J4Vms0nAOMh3GQtAOPyrYOe5WtpzokVv25b9oDnwc2A,2833
+heavyball/foreach_sfadamw.py,sha256=HWbLekY5BloHDIgrN2J0a7IolZCt8Ah2xkLAU_-5oSc,3079
+heavyball/foreach_soap.py,sha256=7B_dP2Hm_xqwpBQiPYkv_c6eoRnU1dV2VZfvSoa4uJ8,4729
+heavyball/p_adam.py,sha256=F-id4qOkAaDTJaKTSNhSsonX-Js5IzIu1Bdj1S4qE2g,6306
+heavyball/palm_foreach_sfadamw.py,sha256=E8raxrBIkSmTEGFzwnfWxKwDJjBQE2vdsmyqfc8aL_A,3375
+heavyball/palm_foreach_soap.py,sha256=IknGm_CzrqDIFEoCkejxjoZ4sfIy6RSoInqlMUOYLB4,6156
+heavyball/precond_schedule_foreach_soap.py,sha256=bJ2ifPFa8zEP9GO8eBpqZzsmP7p_iQkkCkllNeEMHPU,4892
+heavyball/precond_schedule_palm_foreach_soap.py,sha256=4dT9f134-Faq2KuCMCHzMtrkMO-es5p_DYS1of5yF-s,6428
+heavyball/precond_schedule_sfpsoap.py,sha256=FOR-axwlkSN7IHZWYYUVFfjSFCLxc_NdiTlb-n5gmgs,7530
+heavyball/psgd_kron.py,sha256=achB23mQUT3F00IGhjjVf_8YW7VOTHR6YdoCDRyWxsI,6039
+heavyball/pure_psgd.py,sha256=dbYgkunFFA6EsO6fJEhaJRxTH0smi7qLX3Np9XTQ9E4,5079
+heavyball/schedule_free_palm_foreach_soap.py,sha256=0WT_gvTKymqLQzYT6ewDgCmpDq-HgMAewipw1QvyQYA,7267
+heavyball/utils.py,sha256=TVpyev0oL4a78px4cvtaGoGPJqfpfTKE-xBWkRCmzkw,39785
+heavyball-0.22.0.dist-info/LICENSE,sha256=CGdGJim64YifGmUVPaeyRsxkvyExtClswhRNIp8FY_U,1322
+heavyball-0.22.0.dist-info/METADATA,sha256=LqVR3tUgxpk21zsmKxfJAQCKLPzmaQz2PQiKvlvpQe8,11926
+heavyball-0.22.0.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
+heavyball-0.22.0.dist-info/top_level.txt,sha256=SzCxSVg_qCUPA4kZObW3Zyo4v-d_mMOD-p7a-WXTl2E,10
+heavyball-0.22.0.dist-info/RECORD,,

heavyball-0.21.7.dist-info/RECORD

@@ -1,24 +0,0 @@
-heavyball/__init__.py,sha256=iqP428JWwwx-XDOZ0nUdbCkOLEyfoqVyWZLQLAcwxaw,2214
-heavyball/cached_delayed_psgd_kron.py,sha256=Nyxl-G-o6greKwDN-vLiw5W02GXO2LRvknc0OzvzFnE,6674
-heavyball/cached_psgd_kron.py,sha256=HzD6se0AYb-W5hpydUxcR9uqrpe_54PBwgL1VWX3DHU,6592
-heavyball/delayed_psgd.py,sha256=m4c-OvcLMrRxSAPYs2l6Up21uCyF2kvHvpcnfe3nzGs,6212
-heavyball/foreach_adamw.py,sha256=Rb5U80cgUcEqlEbUU250UTWdoqA7nyiqkV5w1U4bWX4,2445
-heavyball/foreach_adopt.py,sha256=ecdi1fKg9i087OGjtKWVbE_DD6Yf4pvpzv4ELCcusvQ,3211
-heavyball/foreach_laprop.py,sha256=vi6C_gfjXxw5uN0KHgzxI9itUI1dcgOf3ufoO_VVMp0,2471
-heavyball/foreach_sfadamw.py,sha256=rLZORmCIMu9G09FdDgMSiI6pNq34IVoxsPVWtmeDdbQ,2753
-heavyball/foreach_soap.py,sha256=4mWSMWYTdjgiXiboI5DwdigecruDtNGKylGAFAVhCRA,4562
-heavyball/p_adam.py,sha256=Xyxsavwtw-t0OyTHitYQXZSmF9UJlMDzDAURge-MbbQ,6047
-heavyball/palm_foreach_sfadamw.py,sha256=JbNrcoquBGGUI5XNMFouDjpNurVHUW9DbX1A3tSrtno,3025
-heavyball/palm_foreach_soap.py,sha256=GzAwM8kOt1X0QCmUZDTdHwPxbJwjH8ic43dyAK5BYCA,6015
-heavyball/precond_schedule_foreach_soap.py,sha256=HcObXLfSNN_lKNb4nmC6tkdHcqDIMNX6hILpHKScqLc,4744
-heavyball/precond_schedule_palm_foreach_soap.py,sha256=xZ7CJvIfdu2RNAZt2g1S7Xb0Jyy1hNC4MowOFU3nWkk,6283
-heavyball/precond_schedule_sfpsoap.py,sha256=PNneiOkrRyV1yIZn91lPmYofd1_OiLqJTDy75RLpXJk,7273
-heavyball/psgd_kron.py,sha256=RSLJi5FnSmYjvYBufNDClnYRm-eN_Kpa1Ar2tNP6-X0,5824
-heavyball/pure_psgd.py,sha256=LZK0qmvZkBF8g00evaVLtW-sIUJmdoxag1K7O26AqEo,4820
-heavyball/schedule_free_palm_foreach_soap.py,sha256=_AqrnChY6iDlQUkF2YUxS7eLjSWCIuvEUOHvMHVM1yY,6873
-heavyball/utils.py,sha256=lyJRL-j_-LW6nVcbYTOrzcMACT3lIBmYioMseqgiexk,37211
-heavyball-0.21.7.dist-info/LICENSE,sha256=CGdGJim64YifGmUVPaeyRsxkvyExtClswhRNIp8FY_U,1322
-heavyball-0.21.7.dist-info/METADATA,sha256=rO1SbkbCdLKf2SD9aTxb_oDuQSZTOq1uh4NmcsjBt4g,11926
-heavyball-0.21.7.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
-heavyball-0.21.7.dist-info/top_level.txt,sha256=SzCxSVg_qCUPA4kZObW3Zyo4v-d_mMOD-p7a-WXTl2E,10
-heavyball-0.21.7.dist-info/RECORD,,