heavyball 0.21.6__tar.gz → 0.21.8__tar.gz

{heavyball-0.21.6 → heavyball-0.21.8}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: heavyball
-Version: 0.21.6
+Version: 0.21.8
 Summary: Efficient optimizers
 Home-page: https://github.com/clashluke/heavyball
 Author: Lucas Nestler

{heavyball-0.21.6 → heavyball-0.21.8}/heavyball/utils.py

@@ -38,7 +38,7 @@ def warmup(lr: float, step: int, warmup_steps: int):
     return lr * step / warmup_steps
 
 
-@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
 def _compilable_schedule_free_(p, z, ckp1, grad, lr, beta1):
     p32 = promote(p)
     z32 = promote(z)
@@ -141,7 +141,7 @@ def beta_debias(beta, step):
     return 1 - (1 - beta) / (1 - beta ** step)
 
 
-@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
 def exp_avg_sq_(state, grad, beta2, eps, out=None):
     if isinstance(state, torch.Tensor):
         state.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
@@ -328,7 +328,7 @@ def get_orthogonal_matrix(mat):
     return final
 
 
-@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
 def _compilable_stochastic_lerp_(x: List[torch.Tensor], y: List[torch.Tensor], a: Union[float, int, torch.Tensor]):
     for x_, y_ in zip(x, y):
         x32 = promote(x_)
@@ -343,7 +343,7 @@ def stochastic_lerp_(x: List[torch.Tensor], y: List[torch.Tensor], a: Union[floa
     _compilable_stochastic_lerp_(x, y, a)
 
 
-@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
 def _compilable_stochastic_add_(x: List[torch.Tensor], y: List[torch.Tensor], alpha: Union[float, int, torch.Tensor]):
     for x_, y_ in zip(x, y):
         x32 = promote(x_)
@@ -581,7 +581,7 @@ def copy_stochastic_list_(target: List[torch.Tensor], source: List[torch.Tensor]
         copy_stochastic_(t, s)
 
 
-@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
 def _compilable_exp_avg_(exp_avg, exp_avg_sq, grad, grad_projected, beta1, beta2, step):
     beta1 = beta_debias(beta1, step)
     beta2 = beta_debias(beta2, step)
@@ -632,7 +632,7 @@ def copy_stochastic_(target: torch.Tensor, source: torch.Tensor):
     _compilable_copy_stochastic_(target, source)
 
 
-@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
 def _compilable_update_(p, u, decay, add_fn, lr):
     u = [u_.view_as(p_) for u_, p_ in zip(u, p)]
     p32, u32 = [list(map(promote, x)) for x in [p, u]]
@@ -788,7 +788,7 @@ def psgd_lb(A, max_abs):
     return x
 
 
-@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
 def psgd_update_precond(Q, exprs, G, precond_lr, tiny, oq, store_triu_as_line):
     """Update Kronecker product preconditioner Q with pair (V, G)."""
     exprA, exprGs, _ = exprs
@@ -821,7 +821,7 @@ def psgd_update_precond(Q, exprs, G, precond_lr, tiny, oq, store_triu_as_line):
         stochastic_add_([o], [term1], -1)
 
 
-@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=False)
 def psgd_precond_grad(Q, exprs, G, inplace: bool = False):
     """Precondition gradient G with preconditioner Q."""
     md = min_dtype(Q)

{heavyball-0.21.6 → heavyball-0.21.8}/heavyball.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: heavyball
-Version: 0.21.6
+Version: 0.21.8
 Summary: Efficient optimizers
 Home-page: https://github.com/clashluke/heavyball
 Author: Lucas Nestler

{heavyball-0.21.6 → heavyball-0.21.8}/setup.py

@@ -10,7 +10,7 @@ setuptools.setup(
     name='heavyball',
     license='BSD',
     description='Efficient optimizers',
-    version='0.21.6',
+    version='0.21.8',
     long_description=README,
     url='https://github.com/clashluke/heavyball',
     packages=setuptools.find_packages(),

{heavyball-0.21.6 → heavyball-0.21.8}/test/test_bf16_params.py

@@ -1,15 +1,15 @@
-import pytest
-import torch
-from torch import nn
-from torch._dynamo import config
-
 import heavyball
 import heavyball.utils
+import pytest
+import torch
 from benchmark.utils import get_optim
 from heavyball.utils import clean, set_torch
+from torch import nn
+from torch._dynamo import config
 
 config.cache_size_limit = 128
 
+
 def get_memory():
     clean()
     torch.cuda.synchronize()
@@ -22,11 +22,10 @@ def get_memory():
 @pytest.mark.parametrize("size,depth", [(256, 2)])
 def test_foreach(opt, size, depth: int, iterations: int = 512, outer_iterations: int = 3):
     set_torch()
-    if 'psgd' not in opt.lower() and 'padam' not in opt.lower():
+    if 'psgd' not in opt.lower():
         raise pytest.skip('Only PSGD and PaLMPAdam are supported')
     opt = getattr(heavyball, opt)
 
-
     peaks = []
     losses = []
 
@@ -37,7 +36,10 @@ def test_foreach(opt, size, depth: int, iterations: int = 512, outer_iterations:
 
         for i in range(outer_iterations):
             model = nn.Sequential(*[nn.Linear(size, size) for _ in range(depth)]).cuda().to(dtype)
-            o = get_optim(opt, model.parameters(), lr=1e-3)
+            o = get_optim(opt, model.parameters(), lr=1e-3, weight_decay=1e-4, warmup_steps=16,
+                          max_size_triangular=2048, merge_dims=True, split=False, memory_save_mode='one_diag',
+                          store_triu_as_line=False, stochastic_schedule=False, storage_dtype='bfloat16',
+                          q_dtype='bfloat16')
 
             for _ in range(iterations):
                 loss = model(torch.randn((1024, size), device='cuda', dtype=dtype)).square().mean()