heavyball 0.18.4__py3-none-any.whl → 0.18.6__py3-none-any.whl

heavyball/foreach_adamw.py

@@ -34,9 +34,9 @@ class ForeachAdamW(StatefulOptimizer):
 
         # Decay the first and second moment running average coefficient
         torch._foreach_lerp_(exp_avg, grad, 1 - beta_debias(group['betas'][0], k + 1))
-        denom = exp_avg_sq_(exp_avg_sq, grad, beta_debias(group['betas'][1], k + 1), eps)
+        denom = list(exp_avg_sq_(exp_avg_sq, grad, beta_debias(group['betas'][1], k + 1), eps))
 
         # Normalize grad in-place for memory efficiency
         lr = -warmup(group['lr'], k + 1, group['warmup_steps'])
-        update_param_(y, exp_avg, lr, decay, lambda p, e, l: torch._foreach_addcdiv_(p, e, denom, l))
+        update_param_(y, exp_avg, lr, decay, lambda p, e, l: p.addcdiv_(e, denom.pop(0), value=l))
         group['k'] = k + 1

heavyball/utils.py

@@ -38,6 +38,18 @@ 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)
+def _compilable_schedule_free_(p, z, ckp1, grad, lr, beta1):
+    p32 = p.float()
+    z32 = z.float()
+    p32.lerp_(end=z32, weight=1 - ckp1)
+    p32.add_(grad, alpha=lr * (beta1 * (1 - ckp1) - 1))
+    _guarded_copy_stochastic(p, p32)
+
+    z32.add_(grad, alpha=-lr)
+    _guarded_copy_stochastic(z, z32)
+
+
 def schedule_free_(lr: float, weight_lr_power: float, weight_sum: float, beta1: float, parameters: List[torch.Tensor],
                    z: List[torch.Tensor], grad: list[torch.Tensor], r: float = 0.0, step: int = 0):
     weight = lr ** weight_lr_power * max(step, 1) ** r
@@ -50,15 +62,10 @@ def schedule_free_(lr: float, weight_lr_power: float, weight_sum: float, beta1:
 
     # These operations update y in-place,
     # without computing x explicitly.
-    p32 = [promote(p) for p in parameters]
-    z32 = [promote(z_) for z_ in z]
-    torch._foreach_lerp_(p32, z32, weight=ckp1)
-    torch._foreach_add_(p32, grad, alpha=lr * (beta1 * (1 - ckp1) - 1))
-    copy_stochastic_list_(parameters, p32)
-
-    # z step
-    torch._foreach_sub_(z, grad, alpha=lr)
-    copy_stochastic_list_(z, z32)
+    lr_tensor = torch.empty((), dtype=torch.float32, device=parameters[0].device).fill_(lr)
+    ckp1_tensor = torch.empty((), dtype=torch.float32, device=parameters[0].device).fill_(ckp1)
+    for p, z_, g in zip(parameters, z, grad):
+        _compilable_schedule_free_(p, z_, ckp1_tensor, g, lr_tensor, beta1)
     return weight_sum
 
 
@@ -150,11 +157,11 @@ def adaptive_gradient_clipping_(parameters: List[torch.Tensor], gradients: List[
 
 
 def set_(dst: torch.Tensor, src: torch.Tensor):
-    if src.data_ptr() == dst.data_ptr():
+    if not torch.compiler.is_compiling() and src.data_ptr() == dst.data_ptr():
         return
     if src.shape != dst.shape:
         src = src.reshape_as(dst)
-    if src.is_contiguous() and dst.is_contiguous() and src.dtype == dst.dtype:
+    if not torch.compiler.is_compiling() and src.is_contiguous() and dst.is_contiguous() and src.dtype == dst.dtype:
         dst.set_(src)
     else:
         dst.copy_(src)
@@ -479,6 +486,12 @@ def copy_stochastic_list_(target: List[torch.Tensor], source: List[torch.Tensor]
         copy_stochastic_(t, s)
 
 
+def _guarded_copy_stochastic(target: torch.Tensor, source: torch.Tensor):
+    if target.dtype != torch.bfloat16 or source.dtype not in (torch.float16, torch.float32, torch.float64):
+        set_(target, source)
+    _compilable_copy_stochastic_(target, source)
+
+
 @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"""
@@ -498,23 +511,27 @@ def _compilable_copy_stochastic_(target: torch.Tensor, source: torch.Tensor):
 def copy_stochastic_(target: torch.Tensor, source: torch.Tensor):
     if target.data_ptr() == source.data_ptr():
         return
-    if target.dtype != torch.bfloat16 or source.dtype not in (torch.float16, torch.float32, torch.float64):
-        set_(target, source)
-        return
-    _compilable_copy_stochastic_(target, source)
+    _guarded_copy_stochastic(target, source)
 
 
-def update_param_(param: List[torch.Tensor], update: List[torch.Tensor], lr: float, decay: float,
-                  add_fn: callable = None):
-    param32 = [promote(p) for p in param]
-    update32 = [promote(u.view(p.shape)) for u, p in zip(update, param)]
+@torch.compile(mode='max-autotune-no-cudagraphs', fullgraph=True, dynamic=True)
+def _compilable_update_one_(p, u, decay, add_fn, lr):
+    p32 = p.float()
+    u32 = u.view(p.shape).float()
     if decay > 0:
-        torch._foreach_mul_(param32, 1 - decay * lr)
+        p32.mul_(1 - decay * lr)
     if add_fn is None:
-        torch._foreach_add_(param32, update32, alpha=lr)
+        p32.add_(u32, alpha=lr)
     else:
-        add_fn(param32, update32, lr)
-    copy_stochastic_list_(param, param32)
+        add_fn(p32, u32, lr)
+    _guarded_copy_stochastic(p, p32)
+
+
+def update_param_(param: List[torch.Tensor], update: List[torch.Tensor], lr: float, decay: float,
+                  add_fn: callable = None):
+    lr_tensor = torch.empty((), dtype=torch.float32, device=param[0].device).fill_(lr)
+    for p, u in zip(param, update):
+        _compilable_update_one_(p, u, decay, add_fn, lr_tensor)
 
 
 def precond_schedule(step, precond_scheduler, rng):
@@ -828,7 +845,10 @@ class PSGDBase(StatefulOptimizer):
 
         for g, q in zip(grad_list, original_q if original_q else q_list):
             if g.dim() > 1:
-                psgd_balance_Q(q)
+                if store_triu_as_line:
+                    psgd_balance_Q([q_ for _, q_ in q])
+                else:
+                    psgd_balance_Q(q)
 
 
 def precond_update_prob_schedule(max_prob=1.0, min_prob=0.03, decay=0.001, flat_start=250):

{heavyball-0.18.4.dist-info → heavyball-0.18.6.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: heavyball
-Version: 0.18.4
+Version: 0.18.6
 Summary: Efficient optimizers
 Home-page: https://github.com/clashluke/heavyball
 Author: Lucas Nestler
@@ -32,7 +32,7 @@ A simple package of efficient optimizers
 The goal is not to thrive for completeness, full maintenance or abstraction, but instead to provide a simple
 largely static alternative to `torch.optim` with more and better optimizers.
 
-Currently (2024-11-20, 0.17.0), the recommended stable optimizer is `PrecondSchedulePaLMSOAP` (see below). The
+Currently (2024-11-21, 0.18.6), the recommended stable optimizer is `PrecondSchedulePaLMSOAP` (see below). The
 recommended experimental optimizer is `DelayedPSGDKron` ([tuning guide](docs/psgd_efficiency.md)).
 
 ## Features

{heavyball-0.18.4.dist-info → heavyball-0.18.6.dist-info}/RECORD

@@ -2,7 +2,7 @@ heavyball/__init__.py,sha256=iqP428JWwwx-XDOZ0nUdbCkOLEyfoqVyWZLQLAcwxaw,2214
 heavyball/cached_delayed_psgd_kron.py,sha256=PQAER6UgVh5l87DGRZrJ8CVP9UhyCG5wJD9rPLnj_G8,6460
 heavyball/cached_psgd_kron.py,sha256=GaeneBp0irksCSBIrJY4D_0hCpZ-uSRPMhqVX_a-og8,6417
 heavyball/delayed_psgd.py,sha256=fhBWFLTSl1S2gHWCeYak-STaXRwpC56sWZGLFMKFEJM,5589
-heavyball/foreach_adamw.py,sha256=h_ar0ZZRM0q_wxuEkxEOEYe-2p-mB4OMgAHivrUnPl8,1777
+heavyball/foreach_adamw.py,sha256=CTg7rfUmlTSjihD5KY9xP0sT2dUKZyZ4-2V42Vlr28U,1780
 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
@@ -16,9 +16,9 @@ heavyball/precond_schedule_sfpsoap.py,sha256=vq7jd302refKPa_9X2lkOTOtCCcTBVByPdo
 heavyball/psgd_kron.py,sha256=u46dorOUXx-do1IYeno2wj-6l1zYKMQQC-N2Zr2PzLI,5476
 heavyball/pure_psgd.py,sha256=iUy7mMKWxwNiVUMYrQ7SBnreu3t_XSbnhTW3a1yw4m0,4835
 heavyball/schedule_free_palm_foreach_soap.py,sha256=zkcikH5wWbzq4kOrmBjilvY3iWzuUddcv2HNEPKr3MI,6366
-heavyball/utils.py,sha256=U8BX11BAhUCH_JO_tnVy4JLSkfDSBHFi_a5s8Pvsf-s,30437
-heavyball-0.18.4.dist-info/LICENSE,sha256=CGdGJim64YifGmUVPaeyRsxkvyExtClswhRNIp8FY_U,1322
-heavyball-0.18.4.dist-info/METADATA,sha256=kBGWYUMlF21_Cg2qtcHa4ajfzB3xlm5UUb24S3mUgKI,11810
-heavyball-0.18.4.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
-heavyball-0.18.4.dist-info/top_level.txt,sha256=SzCxSVg_qCUPA4kZObW3Zyo4v-d_mMOD-p7a-WXTl2E,10
-heavyball-0.18.4.dist-info/RECORD,,
+heavyball/utils.py,sha256=UN0nqXSF-NkRtvCZUdlrs4sRJzZy1v89BkixasQ-bZg,31304
+heavyball-0.18.6.dist-info/LICENSE,sha256=CGdGJim64YifGmUVPaeyRsxkvyExtClswhRNIp8FY_U,1322
+heavyball-0.18.6.dist-info/METADATA,sha256=Grc4Ok9z1BdjtZ7XqRHedKLGNNMcyAU8-QpcGZNl8Hk,11810
+heavyball-0.18.6.dist-info/WHEEL,sha256=GJ7t_kWBFywbagK5eo9IoUwLW6oyOeTKmQ-9iHFVNxQ,92
+heavyball-0.18.6.dist-info/top_level.txt,sha256=SzCxSVg_qCUPA4kZObW3Zyo4v-d_mMOD-p7a-WXTl2E,10
+heavyball-0.18.6.dist-info/RECORD,,