heavyball 0.21.3__py3-none-any.whl → 0.21.4__py3-none-any.whl

heavyball/cached_delayed_psgd_kron.py

@@ -120,7 +120,7 @@ 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)
+            precond_grad_cached_(cached_q, ea, self.state_(p)['cache_expr'], p, lr, weight_decay, self.clip_fn)
 
             if should_update:
                 q = line_to_triu(q_orig) if store_triu_as_line else q_orig

heavyball/cached_psgd_kron.py

@@ -128,4 +128,4 @@ 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)
+            precond_grad_cached_(cached_q, ea, self.state_(p)['cache_expr'], p, lr, weight_decay, self.clip_fn)

heavyball/delayed_psgd.py

@@ -12,9 +12,9 @@ from .utils import update_param_, warmup, psgd_precond_grad, init_Q_exprs, trust
 
 
 @torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
-def _compilable_psgd_precond_grad_(q, exprs, ea, p, lr, weight_decay):
+def _compilable_psgd_precond_grad_(q, exprs, ea, p, lr, weight_deca, clip_fn):
     new = psgd_precond_grad(q, exprs, ea)
-    update_param_([p], self.clip_fn([new]), lr, weight_decay)
+    update_param_([p], clip_fn([new]), lr, weight_decay)
 
 
 class ForeachDelayedPSGD(PSGDBase):
@@ -62,7 +62,7 @@ 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)
+                        split=split, store_triu_as_line=store_triu_as_line, q_dtype=q_dtype, storage_dtype=storage_dtype)
         super().__init__(params, defaults, foreach, stochastic_schedule, clip_fn, preconditioner_update_probability)
 
     def _step(self, group):
@@ -111,7 +111,7 @@ 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, state["exprs"], ea, p, lr, weight_decay)
+            _compilable_psgd_precond_grad_(q, self.state_(p)["exprs"], ea, p, lr, weight_decay, self.clip_fn)
             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/p_adam.py

@@ -6,7 +6,7 @@ Source available at https://github.com/evanatyourservice/kron_torch/blob/97a2b5e
 
 import torch
 
-from heavyball.utils import triu_to_line, line_to_triu, identity
+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
 
@@ -100,7 +100,7 @@ class ForeachPaLMPAdam(PSGDBase):
             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)
-        torch._foreach_lerp_(exp_avg, grad_list, 1 - beta_debias(group['beta'], group['step']))
+        stochastic_lerp_(exp_avg, grad_list, 1 - beta_debias(group['beta'], group['step']))
 
         beta2 = 1 - group['step'] ** -group['beta2_scale']
 

heavyball/utils.py

@@ -966,17 +966,17 @@ class PSGDBase(StatefulOptimizer):
 
 
 @torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
-def _compilable_precond_grad_cached_(cached_q, ea, expr, param, lr, weight_decay):
+def _compilable_precond_grad_cached_(cached_q, ea, expr, param, lr, weight_decay, clip_fn):
     md = min_dtype(cached_q + [ea])
     new = torch.einsum(expr, *[c_.to(md) for c_ in cached_q], ea.to(md)).to(torch.float32)
-    update_param_([param], self.clip_fn([new]), lr, weight_decay)
+    update_param_([param], clip_fn([new]), lr, weight_decay)
 
 
 def precond_grad_cached_(cached_q: List[torch.Tensor], ea: torch.Tensor, expr: str, param: torch.Tensor, lr: float,
-                         weight_decay: float):
+                         weight_decay: float, clip_fn):
     if isinstance(lr, float):
         lr = torch.empty((), dtype=torch.float32, device=param.device).fill_(lr)
-    _compilable_precond_grad_cached_(cached_q, ea, expr, param, lr, weight_decay)
+    _compilable_precond_grad_cached_(cached_q, ea, expr, param, lr, weight_decay, clip_fn)
 
 
 def precond_update_prob_schedule(max_prob=1.0, min_prob=0.03, decay=0.001, flat_start=250):

{heavyball-0.21.3.dist-info → heavyball-0.21.4.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: heavyball
-Version: 0.21.3
+Version: 0.21.4
 Summary: Efficient optimizers
 Home-page: https://github.com/clashluke/heavyball
 Author: Lucas Nestler

{heavyball-0.21.3.dist-info → heavyball-0.21.4.dist-info}/RECORD

@@ -1,13 +1,13 @@
 heavyball/__init__.py,sha256=iqP428JWwwx-XDOZ0nUdbCkOLEyfoqVyWZLQLAcwxaw,2214
-heavyball/cached_delayed_psgd_kron.py,sha256=jBMA5MI4EpOIKww0iQ0zkRO4g1EPPvWHpNWS5ndgDo8,6660
-heavyball/cached_psgd_kron.py,sha256=5qENxAs2N0YRWI0CHFK30bdffxVjOagfAonODKnkDnI,6578
-heavyball/delayed_psgd.py,sha256=b0Fd-3IoqC5Q0eSCGUGkOmKOPHsQke_Ozv4uWVEAZoU,5929
+heavyball/cached_delayed_psgd_kron.py,sha256=Nyxl-G-o6greKwDN-vLiw5W02GXO2LRvknc0OzvzFnE,6674
+heavyball/cached_psgd_kron.py,sha256=HzD6se0AYb-W5hpydUxcR9uqrpe_54PBwgL1VWX3DHU,6592
+heavyball/delayed_psgd.py,sha256=T4OzqGgiycbxuTYJyMSCEI3PBWqPdC6g29KgrNg-JHg,5984
 heavyball/foreach_adamw.py,sha256=Rb5U80cgUcEqlEbUU250UTWdoqA7nyiqkV5w1U4bWX4,2445
 heavyball/foreach_adopt.py,sha256=ecdi1fKg9i087OGjtKWVbE_DD6Yf4pvpzv4ELCcusvQ,3211
 heavyball/foreach_laprop.py,sha256=vi6C_gfjXxw5uN0KHgzxI9itUI1dcgOf3ufoO_VVMp0,2471
 heavyball/foreach_sfadamw.py,sha256=rLZORmCIMu9G09FdDgMSiI6pNq34IVoxsPVWtmeDdbQ,2753
 heavyball/foreach_soap.py,sha256=4mWSMWYTdjgiXiboI5DwdigecruDtNGKylGAFAVhCRA,4562
-heavyball/p_adam.py,sha256=J5QqFAlyLTQ1eQzM0LGxPdv4fEtZikIv9mJ_SSkO3ZY,6033
+heavyball/p_adam.py,sha256=Xyxsavwtw-t0OyTHitYQXZSmF9UJlMDzDAURge-MbbQ,6047
 heavyball/palm_foreach_sfadamw.py,sha256=JbNrcoquBGGUI5XNMFouDjpNurVHUW9DbX1A3tSrtno,3025
 heavyball/palm_foreach_soap.py,sha256=GzAwM8kOt1X0QCmUZDTdHwPxbJwjH8ic43dyAK5BYCA,6015
 heavyball/precond_schedule_foreach_soap.py,sha256=HcObXLfSNN_lKNb4nmC6tkdHcqDIMNX6hILpHKScqLc,4744
@@ -16,9 +16,9 @@ heavyball/precond_schedule_sfpsoap.py,sha256=PNneiOkrRyV1yIZn91lPmYofd1_OiLqJTDy
 heavyball/psgd_kron.py,sha256=RSLJi5FnSmYjvYBufNDClnYRm-eN_Kpa1Ar2tNP6-X0,5824
 heavyball/pure_psgd.py,sha256=LZK0qmvZkBF8g00evaVLtW-sIUJmdoxag1K7O26AqEo,4820
 heavyball/schedule_free_palm_foreach_soap.py,sha256=_AqrnChY6iDlQUkF2YUxS7eLjSWCIuvEUOHvMHVM1yY,6873
-heavyball/utils.py,sha256=uwzxBwlc5JOBBnjle4DiVoXR9HZQX8OG71_HR96BnMg,37187
-heavyball-0.21.3.dist-info/LICENSE,sha256=CGdGJim64YifGmUVPaeyRsxkvyExtClswhRNIp8FY_U,1322
-heavyball-0.21.3.dist-info/METADATA,sha256=6LEa-C-RwfG1sjpIIZhIm1iTiKpxS7VSh8RMXrLV-SE,11926
-heavyball-0.21.3.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
-heavyball-0.21.3.dist-info/top_level.txt,sha256=SzCxSVg_qCUPA4kZObW3Zyo4v-d_mMOD-p7a-WXTl2E,10
-heavyball-0.21.3.dist-info/RECORD,,
+heavyball/utils.py,sha256=L9OaPVWBGl6hyXbt9cYsq4QRUsX4tUIgokeGuywHT84,37209
+heavyball-0.21.4.dist-info/LICENSE,sha256=CGdGJim64YifGmUVPaeyRsxkvyExtClswhRNIp8FY_U,1322
+heavyball-0.21.4.dist-info/METADATA,sha256=ApqLqW-YQ9iqvfRMQu-Y9CEo0DMySpMxUdpV57Xilb4,11926
+heavyball-0.21.4.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
+heavyball-0.21.4.dist-info/top_level.txt,sha256=SzCxSVg_qCUPA4kZObW3Zyo4v-d_mMOD-p7a-WXTl2E,10
+heavyball-0.21.4.dist-info/RECORD,,