heavyball 0.21.8__tar.gz → 0.23.0__tar.gz

{heavyball-0.21.8 → heavyball-0.23.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: heavyball
-Version: 0.21.8
+Version: 0.23.0
 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, 0.22.1), the recommended stable optimizer is `PrecondSchedulePaLMSOAP` (see below). The
 recommended experimental optimizer is `DelayedPSGDKron` ([tuning guide](docs/psgd_efficiency.md)).
 
 ## Features

{heavyball-0.21.8 → heavyball-0.23.0}/README.md

@@ -8,7 +8,7 @@ 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, 0.22.1), the recommended stable optimizer is `PrecondSchedulePaLMSOAP` (see below). The
 recommended experimental optimizer is `DelayedPSGDKron` ([tuning guide](docs/psgd_efficiency.md)).
 
 ## Features

{heavyball-0.21.8 → heavyball-0.23.0}/heavyball/__init__.py

@@ -1,3 +1,4 @@
+from .cached_delayed_psgd_kron import ForeachCachedDelayedPSGDKron
 from .cached_psgd_kron import ForeachCachedPSGDKron
 from .delayed_psgd import ForeachDelayedPSGD
 from .foreach_adamw import ForeachAdamW
@@ -14,7 +15,6 @@ 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 .cached_delayed_psgd_kron import ForeachCachedDelayedPSGDKron
 
 PalmForEachSoap = PaLMForeachSOAP
 
@@ -39,7 +39,8 @@ CachedDelayedPSGDKron = ForeachCachedDelayedPSGDKron
 __all__ = ['PalmForEachSoap', 'PaLMForeachSFAdamW', 'PaLMForeachSOAP', 'SFPaLMForeachSOAP', 'PrecondScheduleSFPaLMSOAP',
            'ForeachSOAP', 'ForeachSFAdamW', 'ForeachLaProp', 'ForeachADOPT', 'PrecondScheduleForeachSOAP',
            'PrecondSchedulePaLMForeachSOAP', 'ForeachPSGDKron', 'ForeachAdamW', 'ForeachPurePSGD', 'ForeachPaLMPAdam',
-           'ForeachDelayedPSGD', 'ForeachCachedPSGDKron', 'ForeachCachedDelayedPSGDKron',  #
-           'PaLMSOAP', 'PaLMSFAdamW', 'PaLMSFSoap', 'PaLMSFAdamW', 'PrecondScheduleSFPaLMSOAP',
-           'SOAP', 'SFAdamW', 'LaProp', 'ADOPT', 'PSGDKron', 'AdamW', 'PurePSGD', 'PaLMPAdam', 'DelayedPSGD',
-           'CachedPSGDKron', 'CachedDelayedPSGDKron', 'PrecondScheduleSOAP', 'PrecondSchedulePaLMSOAP']
+           'ForeachDelayedPSGD', 'ForeachCachedPSGDKron', 'ForeachCachedDelayedPSGDKron',
+             #
+           'PaLMSOAP', 'PaLMSFAdamW', 'PaLMSFSoap', 'PaLMSFAdamW', 'PrecondScheduleSFPaLMSOAP', 'SOAP', 'SFAdamW',
+           'LaProp', 'ADOPT', 'PSGDKron', 'AdamW', 'PurePSGD', 'PaLMPAdam', 'DelayedPSGD', 'CachedPSGDKron',
+           'CachedDelayedPSGDKron', 'PrecondScheduleSOAP', 'PrecondSchedulePaLMSOAP']

{heavyball-0.21.8 → heavyball-0.23.0}/heavyball/cached_delayed_psgd_kron.py

@@ -9,7 +9,7 @@ from typing import Optional
 import torch
 from heavyball.utils import min_dtype, precond_grad_cached_
 
-from .utils import update_param_, warmup, init_Q_exprs, trust_region_clip_, PSGDBase, split_p_and_g_in_group, \
+from .utils import update_param_, warmup, init_Q_exprs, trust_region_clip_, PSGDBase,  \
     line_to_triu, triu_to_line, einsum_base, promote, stochastic_lerp_, beta_debias, precond_grad_cached_
 
 
@@ -43,7 +43,8 @@ class ForeachCachedDelayedPSGDKron(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:
@@ -61,7 +62,7 @@ class ForeachCachedDelayedPSGDKron(PSGDBase):
                         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)
+                        storage_dtype=storage_dtype, caution=caution, mars_gamma=mars_gamma, mars=mars)
         super().__init__(params, defaults, foreach, stochastic_schedule, clip_fn, preconditioner_update_probability)
 
     def _step(self, group):
@@ -81,7 +82,7 @@ class ForeachCachedDelayedPSGDKron(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:
@@ -120,7 +121,7 @@ class ForeachCachedDelayedPSGDKron(PSGDBase):
             q_orig = Q_list.pop(0)
             ea = exp_avg_list.pop(0)
 
-            precond_grad_cached_(cached_q, ea, self.state_(p)['cache_expr'], p, lr, weight_decay, self.clip_fn)
+            precond_grad_cached_(cached_q, ea, self.state_(p)['cache_expr'], p, lr, weight_decay, self.clip_fn, group['caution'], g)
 
             if should_update:
                 q = line_to_triu(q_orig) if store_triu_as_line else q_orig

{heavyball-0.21.8 → heavyball-0.23.0}/heavyball/cached_psgd_kron.py

@@ -8,7 +8,7 @@ from typing import Optional
 
 import torch
 
-from .utils import update_param_, warmup, init_Q_exprs, trust_region_clip_, PSGDBase, split_p_and_g_in_group, \
+from .utils import update_param_, warmup, init_Q_exprs, trust_region_clip_, PSGDBase, \
     line_to_triu, triu_to_line, einsum_base, promote, stochastic_lerp_, beta_debias, precond_grad_cached_
 
 
@@ -40,7 +40,8 @@ class ForeachCachedPSGDKron(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:
@@ -58,7 +59,7 @@ class ForeachCachedPSGDKron(PSGDBase):
                         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)
+                        storage_dtype=storage_dtype, caution=caution, mars_gamma=mars_gamma, mars=mars)
         super().__init__(params, defaults, foreach, stochastic_schedule, clip_fn, preconditioner_update_probability)
 
     def _step(self, group):
@@ -78,7 +79,7 @@ class ForeachCachedPSGDKron(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:
@@ -128,4 +129,5 @@ class ForeachCachedPSGDKron(PSGDBase):
                     else:
                         torch.mul(q_.conj(), q_, out=c_)
 
-            precond_grad_cached_(cached_q, ea, self.state_(p)['cache_expr'], p, lr, weight_decay, self.clip_fn)
+            precond_grad_cached_(cached_q, ea, self.state_(p)['cache_expr'], p, lr, weight_decay, self.clip_fn,
+                                 group['caution'], g)

{heavyball-0.21.8 → heavyball-0.23.0}/heavyball/delayed_psgd.py

@@ -5,17 +5,16 @@ Source available at https://github.com/evanatyourservice/kron_torch/blob/97a2b5e
 """
 
 import torch
-from heavyball.utils import stochastic_lerp_, beta_debias
+from heavyball.utils import stochastic_lerp_, beta_debias, stochastic_add_
 
 from .utils import update_param_, warmup, psgd_precond_grad, init_Q_exprs, trust_region_clip_, PSGDBase, \
-    split_p_and_g_in_group, triu_to_line, line_to_triu, promote
+    triu_to_line, line_to_triu, promote,_compilable_update_
+
 
-# TODO: E1123 00:51:55.423000 159394 site-packages/torch/_guards.py:283] [5/0] Error while creating guard:
-# E1123 00:51:55.423000 159394 site-packages/torch/_guards.py:283] [5/0] Name: "G['psgd_precond_grad'].__defaults__[0]"
 @torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
-def _compilable_psgd_precond_grad_(q, exprs, ea, p, lr, weight_decay, clip_fn):
-    new = psgd_precond_grad(q, exprs, ea)
-    update_param_([p], clip_fn([new]), lr, weight_decay)
+def _compilable_psgd_precond_grad_(q, exprs, ea, p, lr, weight_decay, clip_fn, caution, grad):
+    new = psgd_precond_grad(False, exprs, ea, *q)
+    _compilable_update_([p], clip_fn([new]), weight_decay, stochastic_add_, lr, caution, [grad])
 
 
 class ForeachDelayedPSGD(PSGDBase):
@@ -46,7 +45,8 @@ 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', stochastic_schedule: bool = True, storage_dtype: str = 'float32',  #
+                 q_dtype='float32', stochastic_schedule: bool = True, 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:
@@ -63,7 +63,9 @@ class ForeachDelayedPSGD(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,
+                        caution=caution, mars_gamma=mars_gamma, mars=mars)
         super().__init__(params, defaults, foreach, stochastic_schedule, clip_fn, preconditioner_update_probability)
 
     def _step(self, group):
@@ -83,7 +85,7 @@ class ForeachDelayedPSGD(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:
@@ -112,7 +114,8 @@ class ForeachDelayedPSGD(PSGDBase):
             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
-            _compilable_psgd_precond_grad_(q, self.state_(p)["exprs"], ea, p, lr, weight_decay, self.clip_fn)
+            _compilable_psgd_precond_grad_(q, self.state_(p)["exprs"][-1], ea, p, lr, weight_decay, self.clip_fn, group['caution'],
+                                           g)
             if should_update:
                 q32 = [promote(q_) for q_ in q]
                 self.do_update(group, [p], [ea if momentum_into_precond_update else g], [q32], precond_lr, [q_orig],

{heavyball-0.21.8 → heavyball-0.23.0}/heavyball/foreach_adamw.py

@@ -1,18 +1,19 @@
 import torch
 import torch.optim
-
 from heavyball.utils import copy_stochastic_list_
+
 from .utils import warmup, exp_avg_sq_, beta_debias, update_param_, StatefulOptimizer, promote
 
 
-@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
-def _compilable_step_(y, grad, exp_avg_sq, exp_avg, beta1, beta2, step, lr, eps, decay):
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
+def _compilable_step_(y, grad, exp_avg_sq, exp_avg, beta1, beta2, step, lr, eps, decay, caution):
     g32, exp_avg32, exp_avg_sq32 = [list(map(promote, x)) for x in [grad, exp_avg, exp_avg_sq]]
 
     torch._foreach_lerp_(exp_avg32, g32, 1 - beta_debias(beta1, step + 1))
     denom = list(exp_avg_sq_(exp_avg_sq32, g32, beta_debias(beta2, step + 1), eps))
 
-    update_param_(y, exp_avg32, lr, decay, lambda p, e, l: p.addcdiv_(e, denom.pop(0), value=l))
+    update_param_(y, exp_avg32, lr, decay, lambda p, e, l: p.addcdiv_(e, denom.pop(0), value=l), caution=caution,
+                  grad=g32)
 
     copy_stochastic_list_(exp_avg, exp_avg32)
     copy_stochastic_list_(exp_avg_sq, exp_avg_sq32)
@@ -20,9 +21,11 @@ def _compilable_step_(y, grad, exp_avg_sq, exp_avg, beta1, beta2, step, lr, eps,
 
 class ForeachAdamW(StatefulOptimizer):
     def __init__(self, params, lr=0.0025, betas=(0.9, 0.99), eps=1e-8, weight_decay=0, warmup_steps=0,
-                 foreach: bool = True, storage_dtype: str = 'float32'):
+                 foreach: bool = True, storage_dtype: str = 'float32', mars: bool = False, caution: bool = False,
+                 mars_gamma: float = 0.0025):
         defaults = dict(lr=lr, betas=betas, eps=eps, k=0, warmup_steps=warmup_steps, train_mode=True, weight_sum=0.0,
-                        lr_max=-1.0, weight_decay=weight_decay, storage_dtype=storage_dtype)
+                        lr_max=-1.0, weight_decay=weight_decay, storage_dtype=storage_dtype, mars=mars, caution=caution,
+                        mars_gamma=mars_gamma)
         super().__init__(params, defaults, foreach)
 
     def _step(self, group):
@@ -48,9 +51,13 @@ class ForeachAdamW(StatefulOptimizer):
         y, grad, exp_avg_sq, exp_avg = zip(
             *[(p.data, p.grad, self.state_(p)['exp_avg_sq'], self.state_(p)['exp_avg']) for p in active_p])
 
+        if group['mars']:
+            self.mars_correct_list(group, y, grad, group['mars_gamma'], group['betas'][0])
+
         lr = -warmup(group['lr'], k + 1, group['warmup_steps'])
         lr = torch.empty((), dtype=torch.float32, device=y[0].device).fill_(lr)
         step = torch.empty((), dtype=torch.int32, device=y[0].device).fill_(k)
-        _compilable_step_(y, grad, exp_avg_sq, exp_avg, group['betas'][0], group['betas'][1], step, lr, eps, decay)
+        _compilable_step_(y, grad, exp_avg_sq, exp_avg, group['betas'][0], group['betas'][1], step, lr, eps, decay,
+                          group['caution'])
 
         group['k'] = k + 1

{heavyball-0.21.8 → heavyball-0.23.0}/heavyball/foreach_adopt.py

@@ -5,10 +5,10 @@ from heavyball.utils import copy_stochastic_list_
 from .utils import warmup, beta_debias, update_param_, StatefulOptimizer, promote
 
 
-@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
-def _compilable_step_(y, grad, exp_avg_sq, exp_avg, beta1, beta2, step, lr, eps, decay):
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
+def _compilable_step_(y, grad, exp_avg_sq, exp_avg, beta1, beta2, step, lr, eps, decay, caution):
     g32, exp_avg32, exp_avg_sq32 = [list(map(promote, x)) for x in [grad, exp_avg, exp_avg_sq]]
-    update_param_(y, exp_avg, lr, decay)
+    update_param_(y, exp_avg, lr, decay, caution=caution, grad=g32)
 
     beta1 = beta_debias(beta1, step)
     denom = torch._foreach_sqrt(exp_avg_sq32)
@@ -27,9 +27,11 @@ def _compilable_step_(y, grad, exp_avg_sq, exp_avg, beta1, beta2, step, lr, eps,
 class ForeachADOPT(StatefulOptimizer):
 
     def __init__(self, params, lr=0.0025, betas=(0.9, 0.99), eps=1e-8, weight_decay=0, warmup_steps=0,
-                 foreach: bool = True, storage_dtype: str = 'float32'):
+                 foreach: bool = True, storage_dtype: str = 'float32', mars: bool = False, caution: bool = False,
+                 mars_gamma: float = 0.0025):
         defaults = dict(lr=lr, betas=betas, eps=eps, k=0, warmup_steps=warmup_steps, train_mode=True, weight_sum=0.0,
-                        lr_max=-1.0, weight_decay=weight_decay, storage_dtype=storage_dtype)
+                        lr_max=-1.0, weight_decay=weight_decay, storage_dtype=storage_dtype, mars=mars, caution=caution,
+                        mars_gamma=mars_gamma)
         super().__init__(params, defaults, foreach)
 
     def _step(self, group):
@@ -57,11 +59,14 @@ class ForeachADOPT(StatefulOptimizer):
 
         group['k'] = k + 1
 
+        if group['mars']:
+            self.mars_correct_list(group, y, grad, group['mars_gamma'], group['betas'][0])
+
         if k > 1:
             lr = -warmup(group['lr'], k - 1, group['warmup_steps'])
             lr = torch.empty((), dtype=torch.float32, device=y[0].device).fill_(lr)
             k = torch.empty((), dtype=torch.int32, device=y[0].device).fill_(k)
-            _compilable_step_(y, grad, exp_avg_sq, exp_avg, group['betas'][0], group['betas'][1], k, lr, eps, decay)
+            _compilable_step_(y, grad, exp_avg_sq, exp_avg, group['betas'][0], group['betas'][1], k, lr, eps, decay, group['caution'])
             return
 
         grad = [promote(g) for g in grad]

{heavyball-0.21.8 → heavyball-0.23.0}/heavyball/foreach_laprop.py

@@ -4,8 +4,8 @@ import torch.optim
 from .utils import warmup, exp_avg_sq_, beta_debias, update_param_, StatefulOptimizer, promote, copy_stochastic_list_
 
 
-@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
-def _compilable_step_(y, grad, exp_avg_sq, exp_avg, beta1, beta2, step, lr, eps, decay):
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
+def _compilable_step_(y, grad, exp_avg_sq, exp_avg, beta1, beta2, step, lr, eps, decay, caution):
     g32, exp_avg32, exp_avg_sq32 = [list(map(promote, x)) for x in [grad, exp_avg, exp_avg_sq]]
 
     denom = exp_avg_sq_(exp_avg_sq32, g32, beta_debias(beta2, step), eps)
@@ -14,7 +14,7 @@ def _compilable_step_(y, grad, exp_avg_sq, exp_avg, beta1, beta2, step, lr, eps,
     torch._foreach_mul_(exp_avg32, beta1)
     [ea32.addcdiv_(g, d, value=1 - beta1) for ea32, g, d in zip(exp_avg32, g32, denom)]
 
-    update_param_(y, exp_avg32, lr, decay)
+    update_param_(y, exp_avg32, lr, decay, caution=caution, grad=g32)
 
     copy_stochastic_list_(exp_avg, exp_avg32)
     copy_stochastic_list_(exp_avg_sq, exp_avg_sq32)
@@ -23,9 +23,11 @@ def _compilable_step_(y, grad, exp_avg_sq, exp_avg, beta1, beta2, step, lr, eps,
 class ForeachLaProp(StatefulOptimizer):
 
     def __init__(self, params, lr=0.0025, betas=(0.9, 0.99), eps=1e-8, weight_decay=0, warmup_steps=1,
-                 foreach: bool = True, storage_dtype: str = 'float32'):
+                 foreach: bool = True, storage_dtype: str = 'float32', mars: bool = False, caution: bool = False,
+                 mars_gamma: float = 0.0025):
         defaults = dict(lr=lr, betas=betas, eps=eps, k=0, warmup_steps=warmup_steps, train_mode=True, weight_sum=0.0,
-                        lr_max=-1.0, weight_decay=weight_decay, storage_dtype=storage_dtype)
+                        lr_max=-1.0, weight_decay=weight_decay, storage_dtype=storage_dtype, mars=mars, caution=caution,
+                        mars_gamma=mars_gamma)
         super().__init__(params, defaults, foreach)
 
     def _step(self, group):
@@ -52,10 +54,14 @@ class ForeachLaProp(StatefulOptimizer):
             *[(p.data, p.grad, self.state_(p)['exp_avg_sq'], self.state_(p)['exp_avg'])  #
               for p in active_p])
 
+        if group['mars']:
+            self.mars_correct_list(group, y, grad, group['mars_gamma'], group['betas'][0])
+
         lr = -warmup(group['lr'], k + 1, group['warmup_steps'])
         lr = torch.empty((), dtype=torch.float32, device=y[0].device).fill_(lr)
         step = torch.empty((), dtype=torch.int32, device=y[0].device).fill_(k + 1)
 
-        _compilable_step_(y, grad, exp_avg_sq, exp_avg, group['betas'][0], group['betas'][1], step, lr, eps, decay)
+        _compilable_step_(y, grad, exp_avg_sq, exp_avg, group['betas'][0], group['betas'][1], step, lr, eps, decay,
+                          group['caution'])
 
         group['k'] = k + 1

{heavyball-0.21.8 → heavyball-0.23.0}/heavyball/foreach_sfadamw.py

@@ -5,7 +5,7 @@ from heavyball.utils import get_ckp1, copy_stochastic_list_
 from .utils import warmup, ScheduleFree, exp_avg_sq_, beta_debias, promote, _compilable_schedule_free_
 
 
-@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
 def _compilable_step_(y, grad, exp_avg_sq, z, beta1, beta2, step, ckp1, eps, decay, lr):
     old_debiased2 = beta_debias(beta2, step)
 
@@ -21,13 +21,17 @@ def _compilable_step_(y, grad, exp_avg_sq, z, beta1, beta2, step, ckp1, eps, dec
 
     copy_stochastic_list_(exp_avg_sq, exp_avg_sq32)
 
+
 class ForeachSFAdamW(ScheduleFree):
     def __init__(self, params, lr=0.0025, betas=(0.9, 0.99), eps=1e-8, weight_decay=0, warmup_steps=0, r=0.0,
-                 weight_lr_power=2.0, foreach: bool = True, storage_dtype: str = 'float32'):
+                 weight_lr_power=2.0, foreach: bool = True, storage_dtype: str = 'float32', mars: bool = False,
+                 caution: bool = False, mars_gamma: float = 0.0025):
+
+        assert not caution, "Caution not implemented for SFAdamW"
 
         defaults = dict(lr=lr, betas=betas, eps=eps, r=r, k=0, warmup_steps=warmup_steps, train_mode=True,
                         weight_sum=0.0, lr_max=-1.0, weight_lr_power=weight_lr_power, weight_decay=weight_decay,
-                        foreach=foreach, storage_dtype=storage_dtype)
+                        foreach=foreach, storage_dtype=storage_dtype, mars=mars, caution=caution, mars_gamma=mars_gamma)
         super().__init__(params, defaults, foreach)
 
     def _step(self, group):
@@ -53,6 +57,9 @@ class ForeachSFAdamW(ScheduleFree):
         y, grad, exp_avg_sq, z = zip(*[(p.data, p.grad, self.state_(p)['exp_avg_sq'], self.state_(p)['z'])  #
                                        for p in active_p])
 
+        if group['mars']:
+            self.mars_correct_list(group, y, grad, group['mars_gamma'], group['betas'][0])
+
         lr = warmup(group['lr'], k + 1, group['warmup_steps'])
         ckp1, group['weight_sum'] = get_ckp1(lr, group['weight_lr_power'], group['weight_sum'], group['r'], k + 1)
 

{heavyball-0.21.8 → heavyball-0.23.0}/heavyball/foreach_soap.py

@@ -1,7 +1,7 @@
 import torch
 
 from .utils import init_preconditioner, update_preconditioner, project, beta_debias, exp_avg_sq_, update_param_, set_, \
-    split_p_and_g_in_group, StatefulOptimizer, exp_avg_
+    StatefulOptimizer, exp_avg_
 
 
 class ForeachSOAP(StatefulOptimizer):
@@ -26,11 +26,13 @@ class ForeachSOAP(StatefulOptimizer):
                  weight_decay: float = 0.01, precondition_frequency: int = 2, max_precond_dim: int = 2048,  #
                  merge_dims: bool = True, precondition_1d: bool = False, normalize_grads: bool = False,
                  data_format: str = "channels_first", correct_bias: bool = True, warmup_steps: int = 1,
-                 split: bool = False, foreach: bool = True):
+                 split: bool = False, foreach: bool = True, mars: bool = False, caution: bool = False,
+                 mars_gamma: float = 0.0025):
         defaults = {"lr": lr, "betas": betas, "shampoo_beta": shampoo_beta, "eps": eps, "weight_decay": weight_decay,
                     "precondition_frequency": precondition_frequency, "max_precond_dim": max_precond_dim,
                     "merge_dims": merge_dims, "precondition_1d": precondition_1d, "normalize_grads": normalize_grads,
-                    "correct_bias": correct_bias, 'warmup_steps': warmup_steps, 'split': split}
+                    "correct_bias": correct_bias, 'warmup_steps': warmup_steps, 'split': split, 'mars': mars,
+                    'caution': caution, 'mars_gamma': mars_gamma}
         super().__init__(params, defaults, foreach)
         self._data_format = data_format
 
@@ -41,7 +43,7 @@ class ForeachSOAP(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['betas'][0]):
             state = self.state_(p)
             step = state['step'] = state.get("step", -1) + 1
 
@@ -71,6 +73,8 @@ class ForeachSOAP(StatefulOptimizer):
         step_tensor = torch.empty((), dtype=torch.int32, device=p_list[0].device).fill_(step)
         denom = exp_avg_(exp_avg, exp_avg_sq, grad, grad_projected, beta1, beta2, step_tensor)
 
+        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
@@ -80,11 +84,9 @@ class ForeachSOAP(StatefulOptimizer):
             # 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))
+            precond = 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"])
+            update_param_([p], [precond], step_size, group["weight_decay"], caution=group['caution'], grad=[g])

{heavyball-0.21.8 → heavyball-0.23.0}/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, identity, stochastic_lerp_
+
 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
+    promote
 
 
 class ForeachPaLMPAdam(PSGDBase):
@@ -39,7 +39,8 @@ class ForeachPaLMPAdam(PSGDBase):
                  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',
-                 stochastic_schedule: bool = True,  storage_dtype:str ='float32',#
+                 stochastic_schedule: bool = True, 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 +58,8 @@ class ForeachPaLMPAdam(PSGDBase):
                         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, storage_dtype=storage_dtype)
+                        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):
@@ -75,7 +77,7 @@ class ForeachPaLMPAdam(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=group['beta']):
             state = self.state_(p)
 
             if 'Q' not in state:
@@ -107,13 +109,13 @@ class ForeachPaLMPAdam(PSGDBase):
         lr = -warmup(lr, group['step'], group['warmup_steps'])
 
         for p, Q, g, ea, eas in zip(p_list, Q_triu, grad_list, exp_avg, exp_avg_sq):
-            psgd_precond_grad(Q, self.state_(p)["exprs"], g, inplace=True)
-            ea = psgd_precond_grad(Q, self.state_(p)["exprs"], ea)
+            gc = g.clone() if group['caution'] else None
+            psgd_precond_grad(True, self.state_(p)["exprs"][-1], g, *Q)
+            ea = psgd_precond_grad(False, self.state_(p)["exprs"][-1], ea, *Q)
             exp_avg_sq_(eas, g, beta_debias(beta2, group['step']), 1e-8, out=g)
             torch.div(ea, g, out=g)
             """
             divide by g here, because g == denom (from exp_avg_sq_(out=g)), avoids denom allocation
             divide into g so we can deallocate ea, avoids one allocation (-> less memory than equivalent foreach)
             """
-            update_param_([p], self.clip_fn([g]), lr, weight_decay)
-
+            update_param_([p], self.clip_fn([g]), lr, weight_decay, caution=group['caution'], grad=gc)

{heavyball-0.21.8 → heavyball-0.23.0}/heavyball/palm_foreach_sfadamw.py

@@ -5,7 +5,7 @@ from .utils import warmup, ScheduleFree, exp_avg_sq_, beta_debias, get_ckp1, pro
     _compilable_schedule_free_, copy_stochastic_list_
 
 
-@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
 def _compilable_step_(y, grad, exp_avg_sq, z, beta1, beta2, step, ckp1, eps, decay, lr):
     old_debiased2 = beta_debias(beta2, step)
 
@@ -24,12 +24,17 @@ def _compilable_step_(y, grad, exp_avg_sq, z, beta1, beta2, step, ckp1, eps, dec
 
 class PaLMForeachSFAdamW(ScheduleFree):
     def __init__(self, params, lr=0.0025, beta=0.9, betas=(None, None), eps=1e-8, weight_decay=0, warmup_steps=0, r=0.0,
-                 weight_lr_power=2.0, beta2_scale: float = 0.8, foreach: bool = True, storage_dtype: str = 'float32'):
+                 weight_lr_power=2.0, beta2_scale: float = 0.8, foreach: bool = True, storage_dtype: str = 'float32',
+                 mars: bool = False, caution: bool = False, mars_gamma: float = 0.0025):
         if betas[0] is not None:
             beta = betas[0]
+
+        assert not caution, "Caution not implemented for SFAdamW"
+
         defaults = dict(lr=lr, beta=beta, eps=eps, r=r, k=0, warmup_steps=warmup_steps, train_mode=True, weight_sum=0.0,
                         lr_max=-1.0, weight_lr_power=weight_lr_power, weight_decay=weight_decay,
-                        beta2_scale=beta2_scale, storage_dtype=storage_dtype)
+                        beta2_scale=beta2_scale, storage_dtype=storage_dtype, mars=mars, caution=caution,
+                        mars_gamma=mars_gamma)
         super().__init__(params, defaults, foreach)
 
     def _step(self, group):
@@ -58,6 +63,9 @@ class PaLMForeachSFAdamW(ScheduleFree):
         y, grad, exp_avg_sq, z = zip(*[(p.data, p.grad, self.state_(p)['exp_avg_sq'], self.state_(p)['z'])  #
                                        for p in active_p])
 
+        if group['mars']:
+            self.mars_correct_list(group, y, grad, group['mars_gamma'], group['betas'][0])
+
         lr = warmup(group['lr'], k + 1, group['warmup_steps'])
         ckp1, group['weight_sum'] = get_ckp1(lr, group['weight_lr_power'], group['weight_sum'], group['r'], k + 1)
 

{heavyball-0.21.8 → heavyball-0.23.0}/heavyball/palm_foreach_soap.py

@@ -1,8 +1,7 @@
 import torch
 
 from .utils import init_preconditioner, update_preconditioner, project, beta_debias, exp_avg_sq_, update_param_, set_, \
-    split_p_and_g_in_group, StatefulOptimizer, exp_avg_
-
+    StatefulOptimizer, exp_avg_
 
 
 class PaLMForeachSOAP(StatefulOptimizer):
@@ -33,14 +32,15 @@ class PaLMForeachSOAP(StatefulOptimizer):
                  max_precond_dim: int = 2048,  #
                  merge_dims: bool = True, precondition_1d: bool = False, normalize_grads: bool = False,
                  data_format: str = "channels_first", correct_bias: bool = True, warmup_steps: int = 1,
-                 beta2_scale: float = 0.8, split: bool = False, foreach: bool = True):
+                 beta2_scale: float = 0.8, 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]
         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, 'beta2_scale': beta2_scale,
-                    'split': split}
+                    'split': split, 'mars': mars, 'caution': caution, 'mars_gamma': mars_gamma}
         super().__init__(params, defaults, foreach)
         self._data_format = data_format
 
@@ -51,7 +51,7 @@ class PaLMForeachSOAP(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
 
@@ -82,6 +82,7 @@ class PaLMForeachSOAP(StatefulOptimizer):
         beta2 = torch.empty((), dtype=torch.float32, device=p_list[0].device).fill_(beta2)
         step_tensor = torch.empty((), dtype=torch.int32, device=p_list[0].device).fill_(step)
         denom = exp_avg_(exp_avg, exp_avg_sq, grad, grad_projected, beta1, beta2, step_tensor)
+        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)
@@ -92,11 +93,9 @@ class PaLMForeachSOAP(StatefulOptimizer):
             # 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))
+            precond = 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"])
+            update_param_([p], [precond], step_size, group["weight_decay"], caution=group['caution'], grad=[g])