heavyball 0.14.7__py3-none-any.whl → 0.15.1__py3-none-any.whl

heavyball/schedule_free_palm_foreach_soap.py

@@ -46,84 +46,73 @@ class SFPaLMForeachSOAP(ScheduleFree):
         self._data_format = data_format
         self.rng = random.Random(0x120983109)
 
-    @torch.no_grad()
-    def step(self, closure=None):
-        """
-        Performs a single optimization step.
-
-        Arguments:
-            closure (`Callable`, *optional*): A closure that reevaluates the model and returns the loss.
-        """
-        if closure is None:
-            loss = None
-        else:
-            loss = closure()
-
-        for group in self.param_groups:
-            vals = []
-            max_precond_dim = group['max_precond_dim']
-            precondition_1d = group['precondition_1d']
-
-            step = group['step'] = group.get("step", -1) + 1
-
-            for p in group["params"]:
-                if p.grad is None:
-                    continue
-                grad = p.grad.float()
-                vals.append((p, grad))
-
-            p_list, grad = zip(*vals)
-
-            adaptive_gradient_clipping_(p_list, grad, group["gradient_clip_val"], eps=group["eps"])
-
-            vals = []
-
-            for p, g in split_p_and_g_in_group(group):
-                state = self.state_(p)
-
-                if "z" not in state:
-                    state["z"] = torch.clone(p).float()
-                    state["exp_avg_sq"] = torch.zeros_like(g, dtype=torch.float32)
-                    init_preconditioner(g, state, max_precond_dim, precondition_1d)
-                    update_preconditioner(g, state, max_precond_dim, precondition_1d, 0, True)
-                    continue  # first step is skipped so that we never use the current gradients in the projection.
-
-                # Projecting gradients to the eigenbases of Shampoo's preconditioner
-                # i.e. projecting to the eigenbases of matrices in state['GG']
-                grad_projected = project(g, state['Q'], False)
-                z, exp_avg_sq = state["z"], state["exp_avg_sq"]
-                vals.append((p, g, grad_projected, z, exp_avg_sq))
-
-            if not vals:
+    def _step(self, group):
+        vals = []
+        max_precond_dim = group['max_precond_dim']
+        precondition_1d = group['precondition_1d']
+
+        step = group['step'] = group.get("step", -1) + 1
+
+        for p in group["params"]:
+            if p.grad is None:
                 continue
+            grad = p.grad.float()
+            vals.append((p, grad))
+
+        if not vals:
+            return
+
+        p_list, grad = zip(*vals)
+
+        adaptive_gradient_clipping_(p_list, grad, group["gradient_clip_val"], eps=group["eps"])
+
+        vals = []
+
+        for p, g in split_p_and_g_in_group(group):
+            state = self.state_(p)
+
+            if "z" not in state:
+                state["z"] = torch.clone(p).float()
+                state["exp_avg_sq"] = torch.zeros_like(g, dtype=torch.float32)
+                init_preconditioner(g, state, max_precond_dim, precondition_1d)
+                update_preconditioner(g, state, max_precond_dim, precondition_1d, 0, True)
+                continue  # first step is skipped so that we never use the current gradients in the projection.
+
+            # Projecting gradients to the eigenbases of Shampoo's preconditioner
+            # i.e. projecting to the eigenbases of matrices in state['GG']
+            grad_projected = project(g, state['Q'], False)
+            z, exp_avg_sq = state["z"], state["exp_avg_sq"]
+            vals.append((p, g, grad_projected, z, exp_avg_sq))
+
+        if not vals:
+            return
 
-            p_list, grad, grad_projected, z, exp_avg_sq = zip(*vals)
+        p_list, grad, grad_projected, z, exp_avg_sq = zip(*vals)
 
-            beta2 = 1 - max(step, 1) ** -group['beta2_scale']
-            new_debiased2 = beta_debias(beta2, step)
+        beta2 = 1 - max(step, 1) ** -group['beta2_scale']
+        new_debiased2 = beta_debias(beta2, step)
 
-            # Decay the first and second moment running average coefficient
-            # In-place operations to update the averages at the same time
-            denom = exp_avg_sq_(exp_avg_sq, grad, new_debiased2, group["eps"])
-            torch._foreach_div_(grad_projected, denom)
+        # Decay the first and second moment running average coefficient
+        # In-place operations to update the averages at the same time
+        denom = exp_avg_sq_(exp_avg_sq, grad, new_debiased2, group["eps"])
+        torch._foreach_div_(grad_projected, denom)
 
-            update_precond = group['step'] > 0 and group['step'] % group['precondition_frequency'] == 0
+        update_precond = group['step'] > 0 and group['step'] % group['precondition_frequency'] == 0
 
-            for p, g, gp in zip(p_list, grad, grad_projected):
-                state = self.state_(p)
-                # 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_(gp, project(gp, state['Q'], back=True))
+        for p, g, gp in zip(p_list, grad, grad_projected):
+            state = self.state_(p)
+            # 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_(gp, project(gp, state['Q'], back=True))
 
-                update_preconditioner(g, state, max_precond_dim, precondition_1d, 1 - new_debiased2,
-                                      update_precond)
+            update_preconditioner(g, state, max_precond_dim, precondition_1d, 1 - new_debiased2,
+                                  update_precond)
 
-            # Weight decay calculated at y
-            if group["weight_decay"] > 0:
-                torch._foreach_add_(grad, p_list, alpha=group["weight_decay"])
+        # Weight decay calculated at y
+        if group["weight_decay"] > 0:
+            torch._foreach_add_(grad, p_list, alpha=group["weight_decay"])
 
-            lr = warmup(group['lr'], step, group['warmup_steps'])
-            group['weight_sum'] = schedule_free_(lr, group['weight_lr_power'], group['weight_sum'], group['beta'],
-                                                 p_list, z, grad_projected, group['r'], step)
-        return loss
+        lr = warmup(group['lr'], step, group['warmup_steps'])
+        group['weight_sum'] = schedule_free_(lr, group['weight_lr_power'], group['weight_sum'], group['beta'],
+                                             p_list, z, grad_projected, group['r'], step)

heavyball/utils.py

@@ -3,7 +3,7 @@ import gc
 import math
 import random
 import string
-from typing import List, Optional, Tuple
+from typing import List, Optional, Tuple, Callable
 
 import numpy as np
 import torch
@@ -29,7 +29,7 @@ def decorator(func):
     return _fn
 
 
-_einsum_base = string.ascii_lowercase + string.ascii_uppercase
+einsum_base = string.ascii_lowercase + string.ascii_uppercase
 
 
 def warmup(lr: float, step: int, warmup_steps: int):
@@ -317,8 +317,8 @@ def compute_ggt(grad, GG, max_precond_dim, precondition_1d, beta):
     for idx, sh in enumerate(grad.shape):
         if sh > max_precond_dim:
             continue
-        b = _einsum_base[idx]
-        g0 = _einsum_base[:grad.dim()]
+        b = einsum_base[idx]
+        g0 = einsum_base[:grad.dim()]
         g1 = g0.replace(b, b.upper())
         outer_product = torch.einsum(f'{g0},{g1}->{b + b.upper()}', grad, grad)
         GG[idx].lerp_(promote(outer_product), 1 - beta)
@@ -374,7 +374,7 @@ def project(grad, Q, back: bool):
     :param back: whether to project to Shampoo eigenbases or back to original space
     :return:
     """
-    param = _einsum_base[:grad.dim()]
+    param = einsum_base[:grad.dim()]
     preconditioners = ",".join([(g + g.upper())[::-1 if back else 1] for m, g in zip(Q, param) if len(m) > 0])
     if preconditioners:
         out = ''.join([c.upper() if c.upper() in preconditioners else c for c in param])
@@ -399,6 +399,20 @@ class StatefulOptimizer(torch.optim.Optimizer):
                 tree_map(_add, self.state_(p))
         return total_bytes
 
+    def _step(self, group):
+        raise NotImplementedError
+
+    def step(self, closure: Optional[Callable] = None):
+        if closure is None:
+            loss = None
+        else:
+            with torch.enable_grad():
+                loss = closure()
+        with torch.no_grad():
+            for group in self.param_groups:
+                self._step(group)
+        return loss
+
 
 class ScheduleFree(StatefulOptimizer):
     def eval(self):
@@ -684,9 +698,11 @@ def a_law_compress(x, A=87.6):
     torch._foreach_mul_(xa, 1 / (1 + math.log(A)))
     return xa
 
+
 def identity(x):
     return x
 
+
 def trust_region_clip_(grad, lerp: float = 0.9, scale: float = 1.5):
     torch._foreach_mul_(grad, 1 / scale)
     tanh = torch._foreach_tanh(grad)
@@ -743,8 +759,8 @@ class PSGDBase(StatefulOptimizer):
         self.rng = random.Random(0x1923213)
         self._tiny = torch.finfo(torch.bfloat16).tiny
 
-    def balance(self, do_update, grad_list, Q_list):
-        if not do_update or self.rng.random() > 0.01:
+    def balance(self, grad_list, Q_list):
+        if self.rng.random() > 0.01:
             return
 
         for g, q in zip(grad_list, Q_list):

{heavyball-0.14.7.dist-info → heavyball-0.15.1.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: heavyball
-Version: 0.14.7
+Version: 0.15.1
 Summary: Efficient optimizers
 Home-page: https://github.com/clashluke/heavyball
 Author: Lucas Nestler

heavyball-0.15.1.dist-info/RECORD

@@ -0,0 +1,23 @@
+heavyball/__init__.py,sha256=KbT0GMU0DKqZxq9laCrD7XgiqS9yxC1W52zhte5kjKs,2054
+heavyball/cached_psgd_kron.py,sha256=mXDtxq2WJST_aUJhrLr_xCCXSFaDvD5gCTSEveBUtac,6754
+heavyball/delayed_psgd.py,sha256=dN3NW1jmjxmUkgqxPwUVrqLY8nnBOFp4TVtJ_BhPDR4,5814
+heavyball/foreach_adamw.py,sha256=NSzoIgNm7eavzbJgkAF0k7TUEnWAgOpt9-4juIFoaSA,1729
+heavyball/foreach_adopt.py,sha256=WA07m5jocLfb1GPU8s6mJ2PteS-03ronkKm-VJrAm5I,1863
+heavyball/foreach_laprop.py,sha256=mE2NDGX9XgvRhsewcWnk_-FulZPqGA65ejYF_9-A1Xk,1768
+heavyball/foreach_sfadamw.py,sha256=ussHfPd99u3RTfMrCuu5oIbwNFLXK19wO1Fbz3JShlc,2097
+heavyball/foreach_soap.py,sha256=WWvssYKg607uoEJHftp8ag8mtKSKSeHrT0QTgqBucVg,4587
+heavyball/p_adam.py,sha256=ms7BoMHu3jKGsuztUeECrsXufGAwBpqGsxgZ5LBXLQg,6073
+heavyball/palm_foreach_sfadamw.py,sha256=wjUb_fNZNUmzWXyKvwB0unP9lvNMmaYSQo5YoeS5cj0,2200
+heavyball/palm_foreach_soap.py,sha256=2Sb4hUHQeexJcCgjHeQM_ENkZ6lG1DVxW72ryrvR6iY,5890
+heavyball/precond_schedule_foreach_soap.py,sha256=bHsDyh-UvHpHjumjqqy0PePoR1ZMsJV6o5wWvpLAA04,4815
+heavyball/precond_schedule_palm_foreach_soap.py,sha256=myLTJNQKLtZ3Xi3MVTB-RYtx_XeMRJw5CIMJW75ndUY,6163
+heavyball/precond_schedule_sfpsoap.py,sha256=xeNWetBzBEYqfOSzl98aAVJsHk43QkrUUhHH_YD_mS4,6740
+heavyball/psgd_kron.py,sha256=rMG5UPEgyfQs_n1MHSEicekVDpbbIzinlL8akEyY918,5795
+heavyball/pure_psgd.py,sha256=LLVJhUAb04hgAmT3BTz_faswwQEQUkLhm_VwGQmbBUo,5088
+heavyball/schedule_free_palm_foreach_soap.py,sha256=w0P7lMmoijTpL9V7NwOHcNBFJQ7S1TS9aCiwPhY2yVw,6319
+heavyball/utils.py,sha256=PWmwjZPL4oxMjK79a5R1e7JHykphNi5GdpYqO_xmmFU,27829
+heavyball-0.15.1.dist-info/LICENSE,sha256=CGdGJim64YifGmUVPaeyRsxkvyExtClswhRNIp8FY_U,1322
+heavyball-0.15.1.dist-info/METADATA,sha256=0wImMJNYM-Zg0akh9hRf7X8ofVW6zlmpyDGgAkK5GFA,11667
+heavyball-0.15.1.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
+heavyball-0.15.1.dist-info/top_level.txt,sha256=SzCxSVg_qCUPA4kZObW3Zyo4v-d_mMOD-p7a-WXTl2E,10
+heavyball-0.15.1.dist-info/RECORD,,

heavyball-0.14.7.dist-info/RECORD

@@ -1,22 +0,0 @@
-heavyball/__init__.py,sha256=ef7IWcPF8Uh3WQHzMiMqOFvUnU_LdG7BO9XVecJOph4,1156
-heavyball/delayed_psgd.py,sha256=Gfa1ogkFPPL7ohayYAwbugB8hyLRUI5FgcJfsK69KGI,5936
-heavyball/foreach_adamw.py,sha256=L727chOuVqdiVbYYzloy4g3oKH2FmQh40o_bqWeZtk8,2106
-heavyball/foreach_adopt.py,sha256=M4zZVcYlhGjqO6qekivCuYfX6JtMcp4cJi_RrSHT3H8,2268
-heavyball/foreach_laprop.py,sha256=htYGqgvlZsF_JpILdPMTnX72SqfrokBJ2J4nDeT0JVU,2157
-heavyball/foreach_sfadamw.py,sha256=KIGol7Phfq1DHE_nEle4wDuqNdbGsT3kUcMKzJX3msg,2498
-heavyball/foreach_soap.py,sha256=Ccz9Mc_xaHnrJ_7jUq9ZVxyR0WEqopzOXTUqUY-V8G8,5137
-heavyball/p_adam.py,sha256=jQgTkKekqnLj1XPA4-fgpWG8P_BtUq2976zEt2QymTo,6060
-heavyball/palm_foreach_sfadamw.py,sha256=8IGlRCdkfMzUqcSfmTM3Ce04NLNyrT2xfiBcPkrWwqc,2605
-heavyball/palm_foreach_soap.py,sha256=NEJ3Xeh7pqURUk3cAP2qJe8z2WzYKg60pQe4bsGiaY4,6441
-heavyball/precond_schedule_foreach_soap.py,sha256=H6Oc5IAL5MR-fgu92AboPs3Xm8mBmYUMPLsEcuJ12VI,5370
-heavyball/precond_schedule_palm_foreach_soap.py,sha256=v81hRjcqS6Mm-KxT5Rk3TEiKAE8WI2IbmVbSa-YfBkE,6760
-heavyball/precond_schedule_sfpsoap.py,sha256=7ScnN0in8u9hPiJE7QnOoZOH6Tn-6HeVy4f-bO3bHzY,7279
-heavyball/psgd_kron.py,sha256=AH8ugd_IxKGVtY9y_Ot7myVSxFDbLlRJIqr2bBlAYy8,5911
-heavyball/pure_psgd.py,sha256=jp5fnawUdgccEFlZDPrZr4ZbxYV85IIrev4tybZxBVU,5185
-heavyball/schedule_free_palm_foreach_soap.py,sha256=bV7H-FNNoH5WpposLrNhkqU7mBicMorqKEALBSdROEM,6853
-heavyball/utils.py,sha256=y5VAd9CQjcl_a1WUcORviAYf7Jz_c7n3-b7i5kLUJIA,27464
-heavyball-0.14.7.dist-info/LICENSE,sha256=CGdGJim64YifGmUVPaeyRsxkvyExtClswhRNIp8FY_U,1322
-heavyball-0.14.7.dist-info/METADATA,sha256=5QWB3nuNAp8YjeX0-Y5Uzkek_wjuGF3XG6UWrQk8R0c,11667
-heavyball-0.14.7.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
-heavyball-0.14.7.dist-info/top_level.txt,sha256=SzCxSVg_qCUPA4kZObW3Zyo4v-d_mMOD-p7a-WXTl2E,10
-heavyball-0.14.7.dist-info/RECORD,,