heavyball 0.18.2__tar.gz → 0.18.4__tar.gz

{heavyball-0.18.2 → heavyball-0.18.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: heavyball
-Version: 0.18.2
+Version: 0.18.4
 Summary: Efficient optimizers
 Home-page: https://github.com/clashluke/heavyball
 Author: Lucas Nestler

{heavyball-0.18.2 → heavyball-0.18.4}/heavyball/cached_delayed_psgd_kron.py

@@ -62,6 +62,7 @@ class ForeachCachedDelayedPSGDKron(PSGDBase):
         super().__init__(params, defaults, foreach, stochastic_schedule, clip_fn, preconditioner_update_probability)
 
     def _step(self, group):
+        should_update = self.should_update(group)
         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']
@@ -115,7 +116,7 @@ class ForeachCachedDelayedPSGDKron(PSGDBase):
 
             new = torch.einsum(self.state_(p)['cache_expr'], *cached_q, ea)
 
-            if self.should_update(group):
+            if should_update:
                 q = line_to_triu(q_orig) if store_triu_as_line else q_orig
                 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-0.18.2 → heavyball-0.18.4}/heavyball/cached_psgd_kron.py

@@ -8,8 +8,8 @@ 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, 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):
@@ -71,6 +71,7 @@ class ForeachCachedPSGDKron(PSGDBase):
         beta = group['beta']
         store_triu_as_line = group['store_triu_as_line']
         q_dtype = getattr(torch, group['q_dtype'])
+        should_update = self.should_update(group)
 
         vals = []
 
@@ -111,7 +112,7 @@ class ForeachCachedPSGDKron(PSGDBase):
             q_orig = Q_list.pop(0)
             ea = exp_avg_list.pop(0)
 
-            if self.should_update(group):
+            if should_update:
                 q = line_to_triu(q_orig) if store_triu_as_line else q_orig
                 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-0.18.2 → heavyball-0.18.4}/heavyball/delayed_psgd.py

@@ -62,6 +62,7 @@ class ForeachDelayedPSGD(PSGDBase):
 
 
     def _step(self, group):
+        should_update = self.should_update(group)
         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']
@@ -103,7 +104,7 @@ 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 self.should_update(group):
+            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], store_triu_as_line)
             set_(g, new)

{heavyball-0.18.2 → heavyball-0.18.4}/heavyball/p_adam.py

@@ -61,6 +61,7 @@ class ForeachPaLMPAdam(PSGDBase):
         super().__init__(params, defaults, foreach, stochastic_schedule, clip_fn, preconditioner_update_probability)
 
     def _step(self, group):
+        should_update = self.should_update(group)
         precond_init_scale = group['precond_init_scale']
         max_size_triangular = group['max_size_triangular']
         min_ndim_triangular = group['min_ndim_triangular']
@@ -94,7 +95,7 @@ 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 self.should_update(group):
+        if should_update:
             for g, p, q_, q_orig in zip(grad_list, p_list, Q_triu, Q_list):
                 q32 = [promote(qq_) for qq_ in q_]
                 self.do_update(group, [p], [g], [q32], precond_lr, [q_orig], store_triu_as_line)

{heavyball-0.18.2 → heavyball-0.18.4}/heavyball/psgd_kron.py

@@ -60,6 +60,7 @@ class ForeachPSGDKron(PSGDBase):
         super().__init__(params, defaults, foreach, stochastic_schedule, clip_fn, preconditioner_update_probability)
 
     def _step(self, group):
+        should_update = self.should_update(group)
         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']
@@ -101,7 +102,7 @@ class ForeachPSGDKron(PSGDBase):
             ea = exp_avg_list.pop(0)
             q = line_to_triu(q_orig) if store_triu_as_line else q_orig
 
-            if self.should_update(group):
+            if should_update:
                 q32 = [promote(q_) for q_ in q]
                 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))

{heavyball-0.18.2 → heavyball-0.18.4}/heavyball/pure_psgd.py

@@ -57,7 +57,7 @@ class ForeachPurePSGD(PSGDBase):
         super().__init__(params, defaults, foreach, stochastic_schedule, clip_fn, preconditioner_update_probability)
 
     def _step(self, group):
-        # update preconditioners all together
+        should_update = self.should_update(group)
         precond_init_scale = group['precond_init_scale']
         max_size_triangular = group['max_size_triangular']
         min_ndim_triangular = group['min_ndim_triangular']
@@ -93,7 +93,7 @@ class ForeachPurePSGD(PSGDBase):
             q_orig = Q_list.pop(0)
             q = line_to_triu(q_orig) if store_triu_as_line else q_orig
 
-            if self.should_update(group):
+            if group:
                 q32 = [promote(q_) for q_ in q]
                 self.do_update(group, [p], [g], [q32], precond_lr, [q_orig], store_triu_as_line)
             psgd_precond_grad(q, self.state_(p)["exprs"], g, inplace=True)

{heavyball-0.18.2 → heavyball-0.18.4}/heavyball/utils.py

@@ -335,9 +335,7 @@ def promote(x):
 def min_dtype(xs: List[torch.Tensor]):
     dtypes = [x.dtype for x in xs]
     for d in (torch.float32, torch.bfloat16, torch.float16):
-        if all(d == x for x in dtypes):
-            return d
-        if all(d in (x, torch.float32, torch.float64) for x in dtypes):
+        if all(x in (d, torch.float32, torch.float64) for x in dtypes):
             return d
     return torch.float32
 
@@ -481,7 +479,7 @@ def copy_stochastic_list_(target: List[torch.Tensor], source: List[torch.Tensor]
         copy_stochastic_(t, s)
 
 
-@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
 def _compilable_copy_stochastic_(target: torch.Tensor, source: torch.Tensor):
     """Taken as-is from https://github.com/pytorch/pytorch/issues/120376#issuecomment-1974828905"""
     # create a random 16 bit integer
@@ -817,20 +815,21 @@ class PSGDBase(StatefulOptimizer):
 
     def do_update(self, group, 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)):
+        if original_q:
+            if store_triu_as_line:
+                update_fn = update_triu_
+            else:
+                update_fn = copy_stochastic_list_
+        else:
+            update_fn = lambda x, y: None
+        for i, (p, grad, Q, oq) in enumerate(zip(p_list, grad_list, q_list, original_q)):
             psgd_update_precond(Q, self.state_(p)["exprs"], torch.randn_like(grad), grad, precond_lr, self._tiny)
+            update_fn(oq, Q)
 
-        for g, q in zip(grad_list, q_list):
+        for g, q in zip(grad_list, original_q if original_q else q_list):
             if g.dim() > 1:
                 psgd_balance_Q(q)
 
-        if original_q:
-            for q in q_list:
-                if store_triu_as_line:
-                    update_triu_(original_q[i], Q)
-                else:
-                    copy_stochastic_list_(original_q[i], Q)
-
 
 def precond_update_prob_schedule(max_prob=1.0, min_prob=0.03, decay=0.001, flat_start=250):
     """Anneal preconditioner update probability during beginning of training.

{heavyball-0.18.2 → heavyball-0.18.4}/heavyball.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: heavyball
-Version: 0.18.2
+Version: 0.18.4
 Summary: Efficient optimizers
 Home-page: https://github.com/clashluke/heavyball
 Author: Lucas Nestler

{heavyball-0.18.2 → heavyball-0.18.4}/setup.py

@@ -10,7 +10,7 @@ setuptools.setup(
     name='heavyball',
     license='BSD',
     description='Efficient optimizers',
-    version='0.18.2',
+    version='0.18.4',
     long_description=README,
     url='https://github.com/clashluke/heavyball',
     packages=setuptools.find_packages(),