heavyball 0.17.3__tar.gz → 0.18.1__tar.gz

{heavyball-0.17.3 → heavyball-0.18.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: heavyball
-Version: 0.17.3
+Version: 0.18.1
 Summary: Efficient optimizers
 Home-page: https://github.com/clashluke/heavyball
 Author: Lucas Nestler

{heavyball-0.17.3 → heavyball-0.18.1}/heavyball/cached_delayed_psgd_kron.py

@@ -7,10 +7,9 @@ Source available at https://github.com/evanatyourservice/kron_torch/blob/97a2b5e
 from typing import Optional
 
 import torch
-from heavyball.utils import einsum_base
 
-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_, einsum_base, promote
+from .utils import update_param_, warmup, init_Q_exprs, trust_region_clip_, PSGDBase, split_p_and_g_in_group, \
+    line_to_triu, triu_to_line, set_, einsum_base, promote
 
 
 class ForeachCachedDelayedPSGDKron(PSGDBase):
@@ -42,7 +41,9 @@ class ForeachCachedDelayedPSGDKron(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: Optional[callable] = None, store_triu_as_line: bool = True,
-                 foreach: bool = True, q_dtype='float32'):
+                 foreach: bool = True, q_dtype='float32', stochastic_schedule: bool = True,  #
+                 # expert parameters
+                 precond_init_scale=1.0, precond_lr=0.1):
         if not 0.0 <= lr:
             raise ValueError(f"Invalid learning rate: {lr}")
         if not 0.0 <= beta < 1.0:
@@ -50,33 +51,17 @@ class ForeachCachedDelayedPSGDKron(PSGDBase):
         if not 0.0 <= weight_decay:
             raise ValueError(f"Invalid weight_decay value: {weight_decay}")
 
-        if preconditioner_update_probability is None:
-            preconditioner_update_probability = precond_update_prob_schedule()
         if clip_fn is None:
             clip_fn = lambda x: trust_region_clip_(x, 0.9, 1.5)
-        self.preconditioner_update_probability = preconditioner_update_probability
-        self.clip_fn = clip_fn
 
         defaults = dict(lr=lr, beta=beta, 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=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,
-                        store_triu_as_line=store_triu_as_line,
-                        q_dtype=q_dtype)
-        super().__init__(params, defaults, foreach)
-
-        self._prob_step = 0
+                        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)
+        super().__init__(params, defaults, foreach, stochastic_schedule, clip_fn, preconditioner_update_probability)
 
     def _step(self, group):
-        # update preconditioners all together
-        update_prob = self.preconditioner_update_probability
-        if callable(update_prob):
-            update_prob = update_prob(self._prob_step)
-        do_update = self.rng.random() < update_prob
-        self._prob_step += 1
-
         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']
@@ -128,11 +113,11 @@ class ForeachCachedDelayedPSGDKron(PSGDBase):
             q_orig = Q_list.pop(0)
             ea = exp_avg_list.pop(0)
 
-            if do_update:
+            if self.should_update(group):
                 q = line_to_triu(q_orig) if store_triu_as_line else q_orig
                 q32 = [promote(q_) for q_ in q]
-                self.balance([g], [q32])
-                self.do_update([p], [ea if momentum_into_precond_update else g], [q32], precond_lr, [q_orig], store_triu_as_line=store_triu_as_line)
+                self.do_update(group, [p], [ea if momentum_into_precond_update else g], [q32], precond_lr, [q_orig],
+                               store_triu_as_line)
                 for c_, q_ in zip(cached_q, q):
                     if q_.ndim == 2:
                         torch.matmul(q_.T.conj(), q_, out=c_)

{heavyball-0.17.3 → heavyball-0.18.1}/heavyball/cached_psgd_kron.py

@@ -7,10 +7,9 @@ Source available at https://github.com/evanatyourservice/kron_torch/blob/97a2b5e
 from typing import Optional
 
 import torch
-from heavyball.utils import einsum_base
 
-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_, einsum_base, promote
+from .utils import update_param_, warmup, init_Q_exprs, trust_region_clip_, PSGDBase, \
+    split_p_and_g_in_group, line_to_triu, triu_to_line, set_, einsum_base, promote
 
 
 class ForeachCachedPSGDKron(PSGDBase):
@@ -40,7 +39,9 @@ class ForeachCachedPSGDKron(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: Optional[callable] = None, store_triu_as_line: bool = True,
-                 foreach: bool = True, q_dtype='float32'):
+                 foreach: bool = True, q_dtype='float32', stochastic_schedule: bool = True,  #
+                 # expert parameters
+                 precond_init_scale=1.0, precond_lr=0.1):
         if not 0.0 <= lr:
             raise ValueError(f"Invalid learning rate: {lr}")
         if not 0.0 <= beta < 1.0:
@@ -48,33 +49,17 @@ class ForeachCachedPSGDKron(PSGDBase):
         if not 0.0 <= weight_decay:
             raise ValueError(f"Invalid weight_decay value: {weight_decay}")
 
-        if preconditioner_update_probability is None:
-            preconditioner_update_probability = precond_update_prob_schedule()
         if clip_fn is None:
             clip_fn = lambda x: trust_region_clip_(x, 0.9, 1.5)
-        self.preconditioner_update_probability = preconditioner_update_probability
-        self.clip_fn = clip_fn
 
         defaults = dict(lr=lr, beta=beta, 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=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,
-                        store_triu_as_line=store_triu_as_line,
-                        q_dtype=q_dtype)
-        super().__init__(params, defaults, foreach)
-
-        self._prob_step = 0
+                        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)
+        super().__init__(params, defaults, foreach, stochastic_schedule, clip_fn, preconditioner_update_probability)
 
     def _step(self, group):
-        # update preconditioners all together
-        update_prob = self.preconditioner_update_probability
-        if callable(update_prob):
-            update_prob = update_prob(self._prob_step)
-        do_update = self.rng.random() < update_prob
-        self._prob_step += 1
-
         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']
@@ -128,11 +113,11 @@ class ForeachCachedPSGDKron(PSGDBase):
 
             new = torch.einsum(self.state_(p)['cache_expr'], *cached_q, ea)
 
-            if do_update:
+            if self.should_update(group):
                 q = line_to_triu(q_orig) if store_triu_as_line else q_orig
                 q32 = [promote(q_) for q_ in q]
-                self.balance([g], [q32])
-                self.do_update([p], [ea if momentum_into_precond_update else g], [q32], precond_lr, [q_orig], store_triu_as_line=store_triu_as_line)
+                self.do_update(group, [p], [ea if momentum_into_precond_update else g], [q32], precond_lr, [q_orig],
+                               store_triu_as_line)
                 for c_, q_ in zip(cached_q, q):
                     if q_.ndim == 2:
                         torch.matmul(q_.T.conj(), q_, out=c_)

{heavyball-0.17.3 → heavyball-0.18.1}/heavyball/delayed_psgd.py

@@ -39,7 +39,9 @@ class ForeachDelayedPSGD(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'):
+                 foreach: bool = True, q_dtype='float32', stochastic_schedule: bool = True,  #
+                 # expert parameters
+                 precond_init_scale=1.0, precond_lr=0.1):
         if not 0.0 <= lr:
             raise ValueError(f"Invalid learning rate: {lr}")
         if not 0.0 <= beta < 1.0:
@@ -47,32 +49,19 @@ class ForeachDelayedPSGD(PSGDBase):
         if not 0.0 <= weight_decay:
             raise ValueError(f"Invalid weight_decay value: {weight_decay}")
 
-        if preconditioner_update_probability is None:
-            preconditioner_update_probability = precond_update_prob_schedule()
         if clip_fn is None:
             clip_fn = lambda x: trust_region_clip_(x, 0.9, 1.5)
-        self.preconditioner_update_probability = preconditioner_update_probability
-        self.clip_fn = clip_fn
 
         defaults = dict(lr=lr, beta=beta, 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=0.1,
-                        # precond lr hardcoded to 0.1
-                        precond_init_scale=1.0,  # precond init scale hardcoded to 1.0
+                        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)
-        super().__init__(params, defaults, foreach)
+        super().__init__(params, defaults, foreach, stochastic_schedule, clip_fn, preconditioner_update_probability)
 
-        self._prob_step = 0
 
     def _step(self, group):
-        # update preconditioners all together
-        update_prob = self.preconditioner_update_probability
-        if callable(update_prob):
-            update_prob = update_prob(self._prob_step)
-        do_update = self.rng.random() < update_prob
-        self._prob_step += 1
-
         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']
@@ -114,10 +103,9 @@ class ForeachDelayedPSGD(PSGDBase):
             ea = exp_avg_list.pop(0)
             q = line_to_triu(q_orig) if store_triu_as_line else q_orig
             new = psgd_precond_grad(q, self.state_(p)["exprs"], ea)
-            if do_update:
+            if self.should_update(group):
                 q32 = [promote(q_) for q_ in q]
-                self.do_update([p], [ea if momentum_into_precond_update else g], [q32], precond_lr, [q_orig], store_triu_as_line=store_triu_as_line)
-                self.balance([g], [q32])
+                self.do_update(group,[p], [ea if momentum_into_precond_update else g], [q32], precond_lr, [q_orig], store_triu_as_line)
             set_(g, new)
 
         grad_list = self.clip_fn(grad_list)

{heavyball-0.17.3 → heavyball-0.18.1}/heavyball/p_adam.py

@@ -5,10 +5,10 @@ Source available at https://github.com/evanatyourservice/kron_torch/blob/97a2b5e
 """
 
 import torch
-from heavyball.utils import triu_to_line, line_to_triu
 
-from .utils import update_param_, warmup, psgd_precond_grad, init_Q_exprs, PSGDBase, precond_update_prob_schedule, \
-    exp_avg_sq_, beta_debias, split_p_and_g_in_group, promote
+from heavyball.utils import triu_to_line, line_to_triu, identity
+from .utils import update_param_, warmup, psgd_precond_grad, init_Q_exprs, PSGDBase, exp_avg_sq_, beta_debias, \
+    split_p_and_g_in_group, promote
 
 
 class ForeachPaLMPAdam(PSGDBase):
@@ -38,8 +38,10 @@ class ForeachPaLMPAdam(PSGDBase):
                  max_size_triangular=2048, min_ndim_triangular=2, memory_save_mode=None,
                  momentum_into_precond_update=True, warmup_steps: int = 1, betas=(None, None), beta: float = 0.9,
                  beta2_scale: float = 0.8, merge_dims: bool = False, split: bool = False, clip_fn: callable = None,
-                 store_triu_as_line: bool = True,
-                 foreach: bool = True, q_dtype='float32'):
+                 store_triu_as_line: bool = True, foreach: bool = True, q_dtype='float32',
+                 stochastic_schedule: bool = True,  #
+                 # expert parameters
+                 precond_init_scale=1.0, precond_lr=0.1):
         if not 0.0 <= lr:
             raise ValueError(f"Invalid learning rate: {lr}")
         if not 0.0 <= weight_decay:
@@ -47,32 +49,18 @@ class ForeachPaLMPAdam(PSGDBase):
         if betas[0] is not None:
             beta = betas[0]
 
-        if preconditioner_update_probability is None:
-            preconditioner_update_probability = precond_update_prob_schedule()
         if clip_fn is None:
-            clip_fn = lambda x: x
-        self.preconditioner_update_probability = preconditioner_update_probability
-        self.clip_fn = clip_fn
+            clip_fn = identity
 
         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=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, beta=beta, beta2_scale=beta2_scale, merge_dims=merge_dims,
-                        split=split, store_triu_as_line=store_triu_as_line, q_dtype=q_dtype)
-        super().__init__(params, defaults, foreach)
-
-        self._prob_step = 0
+                        momentum_into_precond_update=momentum_into_precond_update, precond_lr=precond_lr,
+                        precond_init_scale=precond_init_scale, step=0, warmup_steps=warmup_steps, beta=beta,
+                        beta2_scale=beta2_scale, merge_dims=merge_dims, split=split,
+                        store_triu_as_line=store_triu_as_line, q_dtype=q_dtype)
+        super().__init__(params, defaults, foreach, stochastic_schedule, clip_fn, preconditioner_update_probability)
 
     def _step(self, group):
-        # update preconditioners all together
-        update_prob = self.preconditioner_update_probability
-        if callable(update_prob):
-            update_prob = update_prob(self._prob_step)
-        do_update = self.rng.random() < update_prob
-        self._prob_step += 1
-
         precond_init_scale = group['precond_init_scale']
         max_size_triangular = group['max_size_triangular']
         min_ndim_triangular = group['min_ndim_triangular']
@@ -91,8 +79,8 @@ class ForeachPaLMPAdam(PSGDBase):
             if 'Q' not in state:
                 state['exp_avg'] = torch.zeros_like(g)
                 state['exp_avg_sq'] = torch.zeros_like(g)
-                Q, state["exprs"] = init_Q_exprs(p, precond_init_scale, max_size_triangular,
-                                                 min_ndim_triangular, memory_save_mode, dtype=q_dtype)
+                Q, state["exprs"] = init_Q_exprs(p, precond_init_scale, max_size_triangular, min_ndim_triangular,
+                                                 memory_save_mode, dtype=q_dtype)
                 state['Q'] = triu_to_line(Q) if store_triu_as_line else Q
 
             vals.append((p, g, state["Q"], state['exp_avg'], state['exp_avg_sq']))
@@ -106,11 +94,10 @@ class ForeachPaLMPAdam(PSGDBase):
         group["step"] += 1
 
         Q_triu = [line_to_triu(q) if store_triu_as_line else q for q in Q_list]
-        if do_update:
+        if self.should_update(group):
             for g, p, q_, q_orig in zip(grad_list, p_list, Q_triu, Q_list):
                 q32 = [promote(qq_) for qq_ in q_]
-                self.balance([g], [q32])
-                self.do_update([p], [g], [q32], precond_lr, [q_orig], store_triu_as_line=store_triu_as_line)
+                self.do_update(group, [p], [g], [q32], precond_lr, [q_orig], store_triu_as_line)
         torch._foreach_lerp_(exp_avg, grad_list, 1 - beta_debias(group['beta'], group['step']))
 
         beta2 = 1 - group['step'] ** -group['beta2_scale']

{heavyball-0.17.3 → heavyball-0.18.1}/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, \
-    precond_update_prob_schedule, split_p_and_g_in_group, line_to_triu, triu_to_line, set_, promote
+    split_p_and_g_in_group, line_to_triu, triu_to_line, set_, promote
 
 
 class ForeachPSGDKron(PSGDBase):
@@ -39,7 +39,9 @@ class ForeachPSGDKron(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: Optional[callable] = None, store_triu_as_line: bool = True,
-                 foreach: bool = True, q_dtype='float32'):
+                 foreach: bool = True, q_dtype='float32', stochastic_schedule: bool = True,  #
+                 # expert parameters
+                 precond_init_scale=1.0, precond_lr=0.1):
         if not 0.0 <= lr:
             raise ValueError(f"Invalid learning rate: {lr}")
         if not 0.0 <= beta < 1.0:
@@ -47,32 +49,17 @@ class ForeachPSGDKron(PSGDBase):
         if not 0.0 <= weight_decay:
             raise ValueError(f"Invalid weight_decay value: {weight_decay}")
 
-        if preconditioner_update_probability is None:
-            preconditioner_update_probability = precond_update_prob_schedule()
         if clip_fn is None:
             clip_fn = lambda x: trust_region_clip_(x, 0.9, 1.5)
-        self.preconditioner_update_probability = preconditioner_update_probability
-        self.clip_fn = clip_fn
 
         defaults = dict(lr=lr, beta=beta, 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=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,
-                        store_triu_as_line=store_triu_as_line, q_dtype=q_dtype)
-        super().__init__(params, defaults, foreach)
-
-        self._prob_step = 0
+                        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)
+        super().__init__(params, defaults, foreach, stochastic_schedule, clip_fn, preconditioner_update_probability)
 
     def _step(self, group):
-        # update preconditioners all together
-        update_prob = self.preconditioner_update_probability
-        if callable(update_prob):
-            update_prob = update_prob(self._prob_step)
-        do_update = self.rng.random() < update_prob
-        self._prob_step += 1
-
         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']
@@ -114,10 +101,9 @@ class ForeachPSGDKron(PSGDBase):
             ea = exp_avg_list.pop(0)
             q = line_to_triu(q_orig) if store_triu_as_line else q_orig
 
-            if do_update:
+            if self.should_update(group):
                 q32 = [promote(q_) for q_ in q]
-                self.balance([ea if momentum_into_precond_update else g], [q32])
-                self.do_update([p], [g], [q32], precond_lr, [q_orig], store_triu_as_line=store_triu_as_line)
+                self.do_update(group, [p], [g], [q32], precond_lr, [q_orig], store_triu_as_line)
             set_(g, psgd_precond_grad(q, self.state_(p)["exprs"], ea))
 
         grad_list = self.clip_fn(grad_list)

{heavyball-0.17.3 → heavyball-0.18.1}/heavyball/pure_psgd.py

@@ -5,10 +5,10 @@ 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, PSGDBase, precond_update_prob_schedule, \
-    split_p_and_g_in_group, line_to_triu, triu_to_line, promote
+from heavyball.utils import identity
+from .utils import update_param_, warmup, psgd_precond_grad, init_Q_exprs, PSGDBase, split_p_and_g_in_group, \
+    line_to_triu, triu_to_line, promote
 
 
 class ForeachPurePSGD(PSGDBase):
@@ -37,39 +37,27 @@ 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, store_triu_as_line: bool = True,
-                 foreach: bool = True, q_dtype='float32'):
+                 split: bool = False, clip_fn: callable = None, store_triu_as_line: bool = True, foreach: bool = True,
+                 q_dtype='float32', stochastic_schedule: bool = True,  #
+                 # expert parameters
+                 precond_init_scale=1.0, precond_lr=0.1):
         if not 0.0 <= lr:
             raise ValueError(f"Invalid learning rate: {lr}")
         if not 0.0 <= weight_decay:
             raise ValueError(f"Invalid weight_decay value: {weight_decay}")
 
-        if preconditioner_update_probability is None:
-            preconditioner_update_probability = precond_update_prob_schedule()
         if clip_fn is None:
-            clip_fn = lambda x: x
-        self.preconditioner_update_probability = preconditioner_update_probability
-        self.clip_fn = clip_fn
+            clip_fn = identity
 
         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=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,
-                        store_triu_as_line=store_triu_as_line, q_dtype=q_dtype)
-        super().__init__(params, defaults, foreach)
-
-        self._prob_step = 0
+                        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)
+        super().__init__(params, defaults, foreach, stochastic_schedule, clip_fn, preconditioner_update_probability)
 
     def _step(self, group):
         # update preconditioners all together
-        update_prob = self.preconditioner_update_probability
-        if callable(update_prob):
-            update_prob = update_prob(self._prob_step)
-        do_update = self.rng.random() < update_prob
-        self._prob_step += 1
-
         precond_init_scale = group['precond_init_scale']
         max_size_triangular = group['max_size_triangular']
         min_ndim_triangular = group['min_ndim_triangular']
@@ -105,10 +93,9 @@ class ForeachPurePSGD(PSGDBase):
             q_orig = Q_list.pop(0)
             q = line_to_triu(q_orig) if store_triu_as_line else q_orig
 
-            if do_update:
+            if self.should_update(group):
                 q32 = [promote(q_) for q_ in q]
-                self.balance([g], [q32])
-                self.do_update([p], [g], [q32], precond_lr, [q_orig], store_triu_as_line=store_triu_as_line)
+                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)
 
         grad_list = self.clip_fn(grad_list)

{heavyball-0.17.3 → heavyball-0.18.1}/heavyball/utils.py

@@ -793,21 +793,35 @@ def update_triu_(q_state, materialised):
 class PSGDBase(StatefulOptimizer):
     balance_probability: float = 0.01
 
-    def __init__(self, parameters, groups, foreach: bool = True):
-        super().__init__(parameters, groups, foreach)
+    def __init__(self, parameters, groups, foreach: bool, stochastic_schedule: bool, clip_fn,
+                 preconditioner_update_probability):
+        super().__init__(parameters, {**groups, 'stochastic_schedule': stochastic_schedule}, foreach)
         self.rng = random.Random(0x1923213)
         self._tiny = torch.finfo(torch.bfloat16).tiny
-
-    def balance(self, grad_list, Q_list):
-        if self.rng.random() > self.balance_probability:
-            return
-
-        for g, q in zip(grad_list, Q_list):
-            if g.dim() > 1:
-                psgd_balance_Q(q)
-
-    def do_update(self, p_list, grad_list, q_list, precond_lr, original_q: Optional[List] = None,
+        if clip_fn is None:
+            clip_fn = identity
+        if preconditioner_update_probability is None:
+            preconditioner_update_probability = precond_update_prob_schedule()
+        self.clip_fn = clip_fn
+        self.preconditioner_update_probability = preconditioner_update_probability
+
+    def should_update(self, group, prob: Optional[float] = None, name: str = 'cumulative_prob'):
+        group[f'{name}_prob_step'] = group.get(f'{name}_prob_step', 0) + 1
+        if prob is None:
+            prob = self.preconditioner_update_probability(group[f'{name}_prob_step'])
+        if group['stochastic_schedule']:
+            return self.rng.random() < prob
+        cumulative_prob = group.get(name, 0)
+        group[name] = cumulative_prob + prob
+        return int(group[name]) > int(cumulative_prob)
+
+    def do_update(self, group, p_list, grad_list, q_list, precond_lr, original_q: Optional[List] = None,
                   store_triu_as_line=False):
+        if self.should_update(group, self.balance_probability, 'balance_prob'):
+            for g, q in zip(grad_list, q_list):
+                if g.dim() > 1:
+                    psgd_balance_Q(q)
+
         for i, (p, grad, Q) in enumerate(zip(p_list, grad_list, q_list)):
             psgd_update_precond(Q, self.state_(p)["exprs"], torch.randn_like(grad), grad, precond_lr, self._tiny)
             if original_q:

{heavyball-0.17.3 → heavyball-0.18.1}/heavyball.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: heavyball
-Version: 0.17.3
+Version: 0.18.1
 Summary: Efficient optimizers
 Home-page: https://github.com/clashluke/heavyball
 Author: Lucas Nestler

{heavyball-0.17.3 → heavyball-0.18.1}/heavyball.egg-info/SOURCES.txt

@@ -32,4 +32,5 @@ test/test_memory.py
 test/test_merge.py
 test/test_no_grad.py
 test/test_psgd.py
-test/test_soap.py
+test/test_soap.py
+test/test_stochastic_updates.py

{heavyball-0.17.3 → heavyball-0.18.1}/setup.py

@@ -10,7 +10,7 @@ setuptools.setup(
     name='heavyball',
     license='BSD',
     description='Efficient optimizers',
-    version='0.17.3',
+    version='0.18.1',
     long_description=README,
     url='https://github.com/clashluke/heavyball',
     packages=setuptools.find_packages(),

{heavyball-0.17.3 → heavyball-0.18.1}/test/test_bf16_q.py

@@ -28,11 +28,11 @@ def test_foreach(opt, size, depth: int, iterations: int = 128, outer_iterations:
     losses = []
 
     for q_dtype in ['float32', 'bfloat16']:
+        torch.manual_seed(0x2131290)
         peaks.append([])
         losses.append([])
 
         for i in range(outer_iterations):
-            torch.manual_seed(0x2131290)
             model = nn.Sequential(*[nn.Linear(size, size) for _ in range(depth)]).cuda()
             o = get_optim(opt, model.parameters(), lr=1e-3, q_dtype=q_dtype)
 

{heavyball-0.17.3 → heavyball-0.18.1}/test/test_foreach.py

@@ -26,11 +26,11 @@ def test_foreach(opt, size, depth: int, iterations: int = 5, outer_iterations: i
     losses = []
 
     for foreach in [True, False]:
+        torch.manual_seed(0x2131290)
         peaks.append([])
         losses.append([])
 
         for i in range(outer_iterations):
-            torch.manual_seed(0x2131290)
             clean()
             model = nn.Sequential(*[nn.Linear(size, size) for _ in range(depth)]).cuda()
             clean()

heavyball-0.18.1/test/test_stochastic_updates.py

@@ -0,0 +1,52 @@
+import heavyball
+import heavyball.utils
+import pytest
+import torch
+from benchmark.utils import get_optim
+from heavyball.utils import clean, set_torch, PSGDBase
+from torch import nn
+
+
+def get_memory():
+    clean()
+    torch.cuda.synchronize()
+    clean()
+    torch.cuda.synchronize()
+    return torch.cuda.memory_allocated()
+
+
+@pytest.mark.parametrize("opt", heavyball.__all__)
+@pytest.mark.parametrize("size,depth", [(128, 2)])
+def test_foreach(opt, size, depth: int, iterations: int = 8192, outer_iterations: int = 3):
+    set_torch()
+
+    opt = getattr(heavyball, opt)
+    if not issubclass(opt, PSGDBase):
+        raise pytest.skip('Only PSGD is supported')
+
+    peaks = []
+    losses = []
+
+    for stochastic in [False, True]:
+        torch.manual_seed(0x2131290)
+        peaks.append([])
+        losses.append([])
+
+        for i in range(outer_iterations):
+            model = nn.Sequential(*[nn.Linear(size, size) for _ in range(depth)]).cuda()
+            o = get_optim(opt, model.parameters(), lr=1e-3, stochastic_schedule=stochastic)
+
+            for _ in range(iterations):
+                loss = model(torch.randn((128, size)).cuda()).square().mean()
+                loss.backward()
+                o.step()
+                o.zero_grad()
+                losses[-1].append(loss.detach())
+
+            del model, o
+            clean()
+
+    stochastic = sum([l.item() for l in losses[1]])
+    deterministic = sum([l.item() for l in losses[0]])
+    print(f"{deterministic=}, {stochastic=}")
+    assert deterministic < stochastic