heavyball 0.16.0__py3-none-any.whl → 0.17.0__py3-none-any.whl

heavyball/__init__.py

@@ -14,6 +14,7 @@ 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
 
@@ -34,11 +35,12 @@ PurePSGD = ForeachPurePSGD
 PaLMPAdam = ForeachPaLMPAdam
 DelayedPSGD = ForeachDelayedPSGD
 CachedPSGDKron = ForeachCachedPSGDKron
+CachedDelayedPSGDKron
 
 __all__ = ['PalmForEachSoap', 'PaLMForeachSFAdamW', 'PaLMForeachSOAP', 'SFPaLMForeachSOAP', 'PrecondScheduleSFPaLMSOAP',
            'ForeachSOAP', 'ForeachSFAdamW', 'ForeachLaProp', 'ForeachADOPT', 'PrecondScheduleForeachSOAP',
            'PrecondSchedulePaLMForeachSOAP', 'ForeachPSGDKron', 'ForeachAdamW', 'ForeachPurePSGD', 'ForeachPaLMPAdam',
-           'ForeachDelayedPSGD', 'ForeachCachedPSGDKron',  #
+           'ForeachDelayedPSGD', 'ForeachCachedPSGDKron', 'ForeachCachedDelayedPSGDKron'  #
            'PaLMSOAP', 'PaLMSFAdamW', 'PaLMSFSoap', 'PaLMSFAdamW', 'PaLMForeachSOAP', 'PrecondScheduleSFPaLMSOAP',
            'SOAP', 'SFAdamW', 'LaProp', 'ADOPT', 'PSGDKron', 'AdamW', 'PurePSGD', 'PaLMPAdam', 'DelayedPSGD',
-           'CachedPSGDKron']
+           'CachedPSGDKron', 'CachedDelayedPSGDKron']

heavyball/cached_delayed_psgd_kron.py

@@ -0,0 +1,146 @@
+"""
+Originally from Evan Walters and Omead Pooladzandi, 2024
+Modified under Creative Commons Attribution 4.0 International
+Source available at https://github.com/evanatyourservice/kron_torch/blob/97a2b5ee8a1a4c29e4780bbf6c521e545189eff9/kron_torch/kron.py
+"""
+
+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
+
+
+class ForeachCachedDelayedPSGDKron(PSGDBase):
+    """
+    Implements PSGD with off-by-one preconditioning (akin to ADOPT and SOAP) with cached preconditioners.
+
+
+    Args:
+        params (iterable): Iterable of parameters to optimize or dicts defining
+            parameter groups.
+        lr (float): Learning rate.
+        b1 (float): Momentum parameter.
+        weight_decay (float): Weight decay (L2 penalty).
+        preconditioner_update_probability (callable or float, optional): Probability of
+            updating the preconditioner. If None, defaults to a schedule that anneals
+            from 1.0 to 0.03 by 4000 steps.
+        max_size_triangular (int): Max size for dim's preconditioner to be triangular.
+        min_ndim_triangular (int): Minimum number of dimensions a layer needs
+            to have triangular preconditioners.
+        memory_save_mode: (string, optional), None, 'one_diag', or 'all_diag', None is default
+            to set all preconditioners to be triangular, 'one_diag' sets the largest
+            or last dim to be diagonal per layer, and 'all_diag' sets all preconditioners
+            to be diagonal.
+        momentum_into_precond_update: (bool), whether to send momentum into preconditioner
+            update instead of raw gradients.
+    """
+
+    def __init__(self, params, lr=0.001, beta=0.9, weight_decay=0.0, preconditioner_update_probability=None,
+                 max_size_triangular=2048, min_ndim_triangular=2, memory_save_mode=None,
+                 momentum_into_precond_update=True, warmup_steps: int = 1, merge_dims: bool = False,
+                 split: bool = False, clip_fn: Optional[callable] = None, store_triu_as_line: bool = True,
+                 foreach: bool = True, q_dtype='float32'):
+        if not 0.0 <= lr:
+            raise ValueError(f"Invalid learning rate: {lr}")
+        if not 0.0 <= beta < 1.0:
+            raise ValueError(f"Invalid beta parameter: {beta}")
+        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
+
+    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']
+        min_ndim_triangular = group['min_ndim_triangular']
+        memory_save_mode = group['memory_save_mode']
+        precond_lr = group['precond_lr']
+        weight_decay = group['weight_decay']
+        lr = group['lr']
+        beta = group['beta']
+        store_triu_as_line = group['store_triu_as_line']
+        q_dtype = getattr(torch, group['q_dtype'])
+
+        vals = []
+
+        for p, g in split_p_and_g_in_group(group):
+            state = self.state_(p)
+
+            if 'Q' not in state:
+                state["exp_avg"] = torch.zeros_like(g)
+                Q, state["exprs"] = init_Q_exprs(p, precond_init_scale, max_size_triangular, min_ndim_triangular,
+                                                 memory_save_mode, dtype=q_dtype)
+                state['Q'] = triu_to_line(Q) if store_triu_as_line else Q
+                state['Q_cache'] = [torch.empty_like(q) for q in Q]
+
+                expr = [f'{c.upper()}{c}' if q_.ndim == 2 else c for c, q_ in zip(einsum_base, Q)]
+                expr = ','.join(expr)
+                grad_expr = ''.join(c for c, _ in zip(einsum_base, g.shape))
+                out_expr = ''.join(c.upper() if c.upper() in expr else c for c in grad_expr)
+                expr = f'{expr},{grad_expr}->{out_expr}'
+
+                state['cache_expr'] = expr
+
+            vals.append((p, g, state["exp_avg"], state["Q"], state['Q_cache']))
+
+        if not vals:
+            return
+
+        p_list, grad_list, exp_avg_list, Q_list, Q_cache_list = zip(*vals)
+        del vals
+
+        group["step"] += 1
+
+        torch._foreach_lerp_(exp_avg_list, grad_list, (1 - beta) / (1 - beta ** group["step"]))
+
+        grad_list, Q_list, Q_cache_list, exp_avg_list = list(grad_list), list(Q_list), list(Q_cache_list), list(
+            exp_avg_list)
+        for i, (p, g) in enumerate(zip(p_list, grad_list)):
+            cached_q = Q_cache_list.pop(0)
+            q_orig = Q_list.pop(0)
+            ea = exp_avg_list.pop(0)
+
+            if do_update:
+                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)
+                for c_, q_ in zip(cached_q, q):
+                    if q_.ndim == 2:
+                        torch.matmul(q_.T.conj(), q_, out=c_)
+                    else:
+                        torch.mul(q_.conj(), q_, out=c_)
+
+            set_(g, torch.einsum(self.state_(p)['cache_expr'], *cached_q, ea))
+        grad_list = self.clip_fn(grad_list)
+
+        lr = -warmup(lr, group['step'], group['warmup_steps'])
+        update_param_(p_list, grad_list, lr, weight_decay)

heavyball/cached_psgd_kron.py

@@ -10,7 +10,7 @@ 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
+    precond_update_prob_schedule, split_p_and_g_in_group, line_to_triu, triu_to_line, set_, einsum_base, promote
 
 
 class ForeachCachedPSGDKron(PSGDBase):
@@ -40,7 +40,7 @@ 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):
+                 foreach: bool = True, q_dtype='float32'):
         if not 0.0 <= lr:
             raise ValueError(f"Invalid learning rate: {lr}")
         if not 0.0 <= beta < 1.0:
@@ -61,7 +61,8 @@ class ForeachCachedPSGDKron(PSGDBase):
                         # 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)
+                        store_triu_as_line=store_triu_as_line,
+                        q_dtype=q_dtype)
         super().__init__(params, defaults, foreach)
 
         self._prob_step = 0
@@ -84,6 +85,7 @@ class ForeachCachedPSGDKron(PSGDBase):
         lr = group['lr']
         beta = group['beta']
         store_triu_as_line = group['store_triu_as_line']
+        q_dtype = getattr(torch, group['q_dtype'])
 
         vals = []
 
@@ -93,7 +95,7 @@ class ForeachCachedPSGDKron(PSGDBase):
             if 'Q' not in state:
                 state["exp_avg"] = torch.zeros_like(g)
                 Q, state["exprs"] = init_Q_exprs(p, precond_init_scale, max_size_triangular, min_ndim_triangular,
-                                                 memory_save_mode, dtype=g.dtype)
+                                                 memory_save_mode, dtype=q_dtype)
                 state['Q'] = triu_to_line(Q) if store_triu_as_line else Q
                 state['Q_cache'] = [torch.empty_like(q) for q in Q]
 
@@ -124,18 +126,21 @@ class ForeachCachedPSGDKron(PSGDBase):
             q_orig = Q_list.pop(0)
             ea = exp_avg_list.pop(0)
 
+            new = torch.einsum(self.state_(p)['cache_expr'], *cached_q, ea)
+
             if do_update:
                 q = line_to_triu(q_orig) if store_triu_as_line else q_orig
-                self.balance([g], [q])
-                self.do_update([p], [ea if momentum_into_precond_update else g], [q], precond_lr,
-                               [q_orig] if store_triu_as_line else None)
+                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)
                 for c_, q_ in zip(cached_q, q):
                     if q_.ndim == 2:
                         torch.matmul(q_.T.conj(), q_, out=c_)
                     else:
                         torch.mul(q_.conj(), q_, out=c_)
 
-            set_(g, torch.einsum(self.state_(p)['cache_expr'], *cached_q, ea))
+            set_(g, new)
+
         grad_list = self.clip_fn(grad_list)
 
         lr = -warmup(lr, group['step'], group['warmup_steps'])

heavyball/delayed_psgd.py

@@ -8,7 +8,7 @@ import torch
 from heavyball.utils import copy_stochastic_list_
 
 from .utils import update_param_, warmup, psgd_precond_grad, init_Q_exprs, trust_region_clip_, PSGDBase, \
-    precond_update_prob_schedule, split_p_and_g_in_group, triu_to_line, line_to_triu, set_
+    precond_update_prob_schedule, split_p_and_g_in_group, triu_to_line, line_to_triu, set_, promote
 
 
 class ForeachDelayedPSGD(PSGDBase):
@@ -39,7 +39,7 @@ 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):
+                 foreach: bool = True, q_dtype='float32'):
         if not 0.0 <= lr:
             raise ValueError(f"Invalid learning rate: {lr}")
         if not 0.0 <= beta < 1.0:
@@ -60,7 +60,7 @@ class ForeachDelayedPSGD(PSGDBase):
                         # 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)
+                        store_triu_as_line=store_triu_as_line, q_dtype=q_dtype)
         super().__init__(params, defaults, foreach)
 
         self._prob_step = 0
@@ -83,6 +83,7 @@ class ForeachDelayedPSGD(PSGDBase):
         lr = group['lr']
         beta = group['beta']
         store_triu_as_line = group['store_triu_as_line']
+        q_dtype = getattr(torch, group['q_dtype'])
 
         vals = []
 
@@ -92,7 +93,7 @@ class ForeachDelayedPSGD(PSGDBase):
             if 'Q' not in state:
                 state["exp_avg"] = torch.zeros_like(g)
                 Q, state["exprs"] = init_Q_exprs(p, precond_init_scale, max_size_triangular, min_ndim_triangular,
-                                                 memory_save_mode, dtype=g.dtype)
+                                                 memory_save_mode, dtype=q_dtype)
                 state["Q"] = triu_to_line(Q) if store_triu_as_line else Q
 
             vals.append((p, g, state["exp_avg"], state["Q"]))
@@ -114,9 +115,9 @@ class ForeachDelayedPSGD(PSGDBase):
             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:
-                self.do_update([p], [ea if momentum_into_precond_update else g], [q], precond_lr,
-                               [q_orig] if store_triu_as_line else None)
-                self.balance([g], [q])
+                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])
             set_(g, new)
 
         grad_list = self.clip_fn(grad_list)

heavyball/p_adam.py

@@ -8,7 +8,7 @@ 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
+    exp_avg_sq_, beta_debias, split_p_and_g_in_group, promote
 
 
 class ForeachPaLMPAdam(PSGDBase):
@@ -39,7 +39,7 @@ 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):
+                 foreach: bool = True, q_dtype='float32'):
         if not 0.0 <= lr:
             raise ValueError(f"Invalid learning rate: {lr}")
         if not 0.0 <= weight_decay:
@@ -60,7 +60,7 @@ class ForeachPaLMPAdam(PSGDBase):
                         # 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)
+                        split=split, store_triu_as_line=store_triu_as_line, q_dtype=q_dtype)
         super().__init__(params, defaults, foreach)
 
         self._prob_step = 0
@@ -81,6 +81,7 @@ class ForeachPaLMPAdam(PSGDBase):
         weight_decay = group['weight_decay']
         lr = group['lr']
         store_triu_as_line = group['store_triu_as_line']
+        q_dtype = getattr(torch, group['q_dtype'])
 
         vals = []
 
@@ -91,7 +92,7 @@ class ForeachPaLMPAdam(PSGDBase):
                 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=g.dtype)
+                                                 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']))
@@ -106,9 +107,10 @@ class ForeachPaLMPAdam(PSGDBase):
 
         Q_triu = [line_to_triu(q) if store_triu_as_line else q for q in Q_list]
         if do_update:
-            self.balance(grad_list, Q_triu)
-            self.do_update(p_list, grad_list, Q_triu, precond_lr, Q_list if store_triu_as_line else None)
-
+            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)
         torch._foreach_lerp_(exp_avg, grad_list, 1 - beta_debias(group['beta'], group['step']))
 
         beta2 = 1 - group['step'] ** -group['beta2_scale']

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_
+    precond_update_prob_schedule, split_p_and_g_in_group, line_to_triu, triu_to_line, set_, promote
 
 
 class ForeachPSGDKron(PSGDBase):
@@ -39,7 +39,7 @@ 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):
+                 foreach: bool = True, q_dtype='float32'):
         if not 0.0 <= lr:
             raise ValueError(f"Invalid learning rate: {lr}")
         if not 0.0 <= beta < 1.0:
@@ -60,7 +60,7 @@ class ForeachPSGDKron(PSGDBase):
                         # 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)
+                        store_triu_as_line=store_triu_as_line, q_dtype=q_dtype)
         super().__init__(params, defaults, foreach)
 
         self._prob_step = 0
@@ -83,6 +83,7 @@ class ForeachPSGDKron(PSGDBase):
         lr = group['lr']
         beta = group['beta']
         store_triu_as_line = group['store_triu_as_line']
+        q_dtype = getattr(torch, group['q_dtype'])
 
         vals = []
 
@@ -92,7 +93,7 @@ class ForeachPSGDKron(PSGDBase):
             if 'Q' not in state:
                 state["exp_avg"] = torch.zeros_like(g)
                 Q, state["exprs"] = init_Q_exprs(p, precond_init_scale, max_size_triangular, min_ndim_triangular,
-                                                 memory_save_mode, dtype=g.dtype)
+                                                 memory_save_mode, dtype=q_dtype)
                 state['Q'] = triu_to_line(Q) if store_triu_as_line else Q
 
             vals.append((p, g, state["exp_avg"], state["Q"]))
@@ -114,9 +115,9 @@ class ForeachPSGDKron(PSGDBase):
             q = line_to_triu(q_orig) if store_triu_as_line else q_orig
 
             if do_update:
-                self.balance([g], [q])
-                self.do_update([p], [ea if momentum_into_precond_update else g], [q], precond_lr,
-                               [q_orig] if store_triu_as_line else None)
+                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)
             set_(g, psgd_precond_grad(q, self.state_(p)["exprs"], ea))
 
         grad_list = self.clip_fn(grad_list)

heavyball/pure_psgd.py

@@ -5,9 +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
+    split_p_and_g_in_group, line_to_triu, triu_to_line, promote
 
 
 class ForeachPurePSGD(PSGDBase):
@@ -37,7 +38,7 @@ 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):
+                 foreach: bool = True, q_dtype='float32'):
         if not 0.0 <= lr:
             raise ValueError(f"Invalid learning rate: {lr}")
         if not 0.0 <= weight_decay:
@@ -56,7 +57,7 @@ class ForeachPurePSGD(PSGDBase):
                         # 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)
+                        store_triu_as_line=store_triu_as_line, q_dtype=q_dtype)
         super().__init__(params, defaults, foreach)
 
         self._prob_step = 0
@@ -77,6 +78,7 @@ class ForeachPurePSGD(PSGDBase):
         weight_decay = group['weight_decay']
         lr = group['lr']
         store_triu_as_line = group['store_triu_as_line']
+        q_dtype = getattr(torch, group['q_dtype'])
 
         vals = []
 
@@ -85,7 +87,7 @@ class ForeachPurePSGD(PSGDBase):
 
             if 'Q' not in state:
                 Q, state["exprs"] = init_Q_exprs(p, precond_init_scale, max_size_triangular, min_ndim_triangular,
-                                                 memory_save_mode, dtype=g.dtype)
+                                                 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"]))
@@ -104,8 +106,9 @@ class ForeachPurePSGD(PSGDBase):
             q = line_to_triu(q_orig) if store_triu_as_line else q_orig
 
             if do_update:
-                self.balance([g], [q])
-                self.do_update([p], [g], [q], precond_lr, [q_orig] if store_triu_as_line else None)
+                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)
             psgd_precond_grad(q, self.state_(p)["exprs"], g, inplace=True)
 
         grad_list = self.clip_fn(grad_list)

heavyball/utils.py

@@ -325,9 +325,9 @@ def compute_ggt(grad, GG, max_precond_dim, precondition_1d, beta):
 
 
 def promote(x):
-    if x is (torch.bfloat16, torch.float16):
+    if x in (torch.bfloat16, torch.float16):
         return torch.float32
-    if x.dtype in (torch.bfloat16, torch.float16):
+    if hasattr(x, 'dtype') and x.dtype in (torch.bfloat16, torch.float16):
         return x.float()
     return x
 
@@ -468,15 +468,15 @@ class ScheduleFree(StatefulOptimizer):
 
 def copy_stochastic_list_(target: List[torch.Tensor], source: List[torch.Tensor]):
     for t, s in zip(target, source):
-        if t.dtype == torch.bfloat16:
-            copy_stochastic_(t, s)
-        else:
-            set_(t, s)
+        copy_stochastic_(t, s)
 
 
 def copy_stochastic_(target: torch.Tensor, source: torch.Tensor):
     if target.data_ptr() == source.data_ptr():
         return
+    if target.dtype != torch.bfloat16:
+        set_(target, source)
+        return
 
     """Taken as-is from https://github.com/pytorch/pytorch/issues/120376#issuecomment-1974828905"""
     # create a random 16 bit integer
@@ -555,7 +555,7 @@ def init_Q_exprs(t, scale, max_size, min_ndim_triangular, memory_save_mode, dtyp
     for i, (size, dim_d) in enumerate(zip(shape, dim_diag)):
         if size == 1 or size > max_size or len(shape) < min_ndim_triangular or dim_d:
             # use diagonal matrix as preconditioner for this dim
-            Q.append(scale * torch.ones(size, dtype=dtype, device=t.device))
+            Q.append(scale * torch.ones(size, dtype=promote(dtype), device=t.device))
 
             piece1A.append(letters[i])
             piece2A = piece2A + letters[i]
@@ -669,11 +669,11 @@ def psgd_update_precond(Q, exprs, V, G, step, tiny):
 @decorator
 def psgd_precond_grad(Q, exprs, G, inplace: bool = False):
     """Precondition gradient G with preconditioner Q."""
-    out = torch.einsum(exprs[-1], *[q.conj() for q in Q], *Q, G)
+    out = torch.einsum(exprs[-1], *[q.conj() for q in Q], *Q, G.to(Q[0].dtype))
     if inplace:
         set_(G, out)
         return G
-    return out
+    return out.to(G.dtype)
 
 
 def norm_clip_(x, scale=None):
@@ -768,28 +768,33 @@ def line_to_triu(Q_list: List[Tuple[Optional[List[int]], torch.Tensor]]):
 def update_triu_(q_state, materialised):
     for (shape0, q), (shape1, m) in zip(q_state, triu_to_line(materialised)):
         assert shape0 == shape1
-        set_(q, m)
+        copy_stochastic_(q, m)
 
 
 class PSGDBase(StatefulOptimizer):
+    balance_probability: float = 0.01
+
     def __init__(self, parameters, groups, foreach: bool = True):
         super().__init__(parameters, groups, foreach)
         self.rng = random.Random(0x1923213)
         self._tiny = torch.finfo(torch.bfloat16).tiny
 
     def balance(self, grad_list, Q_list):
-        if self.rng.random() > 0.01:
+        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):
+    def do_update(self, p_list, grad_list, q_list, precond_lr, original_q: Optional[List] = None, store_triu_as_line=False):
         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:
-                update_triu_(original_q[i], Q)
+                if store_triu_as_line:
+                    update_triu_(original_q[i], Q)
+                else:
+                    copy_stochastic_(original_q[i], Q)
 
 
 def precond_update_prob_schedule(max_prob=1.0, min_prob=0.03, decay=0.001, flat_start=250):

{heavyball-0.16.0.dist-info → heavyball-0.17.0.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: heavyball
-Version: 0.16.0
+Version: 0.17.0
 Summary: Efficient optimizers
 Home-page: https://github.com/clashluke/heavyball
 Author: Lucas Nestler

{heavyball-0.16.0.dist-info → heavyball-0.17.0.dist-info}/RECORD

@@ -1,23 +1,24 @@
-heavyball/__init__.py,sha256=KbT0GMU0DKqZxq9laCrD7XgiqS9yxC1W52zhte5kjKs,2054
-heavyball/cached_psgd_kron.py,sha256=vJuy639G-_ZLSRX3goSFMXALv-ucYjrxaEtpj0IHo-M,6802
-heavyball/delayed_psgd.py,sha256=sbwgAed5gmQpHNTPvuE7Si-gB-s0NVvN4d-4rNUJj4c,5893
+heavyball/__init__.py,sha256=mDHahP4u0fy2YKWA4FPMAp7jLPMt5WwUkEiOrwE4u3E,2199
+heavyball/cached_delayed_psgd_kron.py,sha256=DvjNNHzbnS-NDq965wve-VQ-ol7IFljYYGTuTwPHOhU,6971
+heavyball/cached_psgd_kron.py,sha256=xy3-yRKFUvRTstJb_asMVp-k-5Zuw_HyILPi7BsuMKQ,6974
+heavyball/delayed_psgd.py,sha256=rDDUj3miEn6HRJmKl-ZImsqkqBASSn8aC7MEV_06fzU,6017
 heavyball/foreach_adamw.py,sha256=h_ar0ZZRM0q_wxuEkxEOEYe-2p-mB4OMgAHivrUnPl8,1777
 heavyball/foreach_adopt.py,sha256=ogOw2JjwEQNj7AKlweAphQFdMJ_GcMDm-RyDvEzugoc,1911
 heavyball/foreach_laprop.py,sha256=yGVmGqWiSw8Y2Xj70ndkR8ZMygakTB4_iRwV02Svkqg,1816
 heavyball/foreach_sfadamw.py,sha256=15-n6-lx4PAHYsKYmXbugxsR5MnqaPYy2vUudPRiitg,2087
 heavyball/foreach_soap.py,sha256=h6ptMch7oaynvu3eIJtWnVXypDA_5JDVm3Zb3PNEma0,4634
-heavyball/p_adam.py,sha256=aCu4Qn0eHJETHuCGrfNKp2aygKk2ZoNQyxut3Vcqmoc,6112
+heavyball/p_adam.py,sha256=F2b-xGNROi9VfX7isa3kffWePojpBl5BI1n464w4tGQ,6334
 heavyball/palm_foreach_sfadamw.py,sha256=yvZbPyjDW8qd3r4qDXb6uTr5RozQ7JSDj4aYYRnKGLA,2248
 heavyball/palm_foreach_soap.py,sha256=g4hbiGRcti-J-a0SwAkP4ii5pU-aalsZH5bssyhroLk,5938
 heavyball/precond_schedule_foreach_soap.py,sha256=WLg5SzpJnKPZUvFyIvdwSZa1Umt5cpr3Kow_42orM-E,4863
 heavyball/precond_schedule_palm_foreach_soap.py,sha256=ammQrvRZFF-wc-wEiPEoFhS_7b8pdV61QfcLoQfimSo,6211
 heavyball/precond_schedule_sfpsoap.py,sha256=vq7jd302refKPa_9X2lkOTOtCCcTBVByPdojklrY8pA,6770
-heavyball/psgd_kron.py,sha256=iWTAViuzxTodtQGZnkLsEXrLG8tNU-BQB3KkTYAVcX4,5874
-heavyball/pure_psgd.py,sha256=EuCPNM8TX13cOop-mvvBFh6Uo1UjD1vsE053hvil92Q,5136
+heavyball/psgd_kron.py,sha256=2IpPj2TOExNGm8hSewi3er2GczJRNgC7r2J5yYSSA_0,5998
+heavyball/pure_psgd.py,sha256=uA7W9a3Qm1sxHQhtNxaUYrmE5x55lP5iJOKy_qT8XaQ,5341
 heavyball/schedule_free_palm_foreach_soap.py,sha256=zkcikH5wWbzq4kOrmBjilvY3iWzuUddcv2HNEPKr3MI,6366
-heavyball/utils.py,sha256=z6taEvpgszKTrscqgowKYqb0xIVpBDVDBNGgvTE4Pb8,28484
-heavyball-0.16.0.dist-info/LICENSE,sha256=CGdGJim64YifGmUVPaeyRsxkvyExtClswhRNIp8FY_U,1322
-heavyball-0.16.0.dist-info/METADATA,sha256=yjpldOTN2rXN2-KG7R9ytuyBfmSCDpznZeRuziANChE,11941
-heavyball-0.16.0.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
-heavyball-0.16.0.dist-info/top_level.txt,sha256=SzCxSVg_qCUPA4kZObW3Zyo4v-d_mMOD-p7a-WXTl2E,10
-heavyball-0.16.0.dist-info/RECORD,,
+heavyball/utils.py,sha256=Jqh7VdWGeiSdwaPtUNB9l14wuuFPSReLaTwJA3juFbM,28765
+heavyball-0.17.0.dist-info/LICENSE,sha256=CGdGJim64YifGmUVPaeyRsxkvyExtClswhRNIp8FY_U,1322
+heavyball-0.17.0.dist-info/METADATA,sha256=GIJQ4ha-fcYR6ltOs4WUO8L_LhWGiZv2UrEZcuJD0LI,11941
+heavyball-0.17.0.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
+heavyball-0.17.0.dist-info/top_level.txt,sha256=SzCxSVg_qCUPA4kZObW3Zyo4v-d_mMOD-p7a-WXTl2E,10
+heavyball-0.17.0.dist-info/RECORD,,