adv-optm 1.2.dev15__tar.gz → 1.2.dev17__tar.gz

{adv_optm-1.2.dev15 → adv_optm-1.2.dev17}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: adv_optm
-Version: 1.2.dev15
+Version: 1.2.dev17
 Summary: A family of highly efficient, lightweight yet powerful optimizers.
 Home-page: https://github.com/Koratahiu/Advanced_Optimizers
 Author: Koratahiu

{adv_optm-1.2.dev15 → adv_optm-1.2.dev17}/adv_optm/__init__.py

@@ -20,4 +20,4 @@ __all__ = [
     "AdaMuon_adv",
 ]
 
-__version__ = "1.2.dev15"
+__version__ = "1.2.dev17"

{adv_optm-1.2.dev15 → adv_optm-1.2.dev17}/adv_optm/optim/Muon_adv.py

@@ -76,6 +76,7 @@ class Muon_adv(torch.optim.Optimizer):
         adam_beta3_ema (float): Beta3 for AdEMAMix.
         adam_alpha (float): Alpha for AdEMAMix.
         adam_kourkoutas_beta (bool): Kourkoutas-β for AdamW.
+        adam_nnmf_factor (bool): 1-bit factored for AdamW.
     """
 
     def __init__(
@@ -125,6 +126,7 @@ class Muon_adv(torch.optim.Optimizer):
         adam_ema_alpha: float = 0.95,
         adam_tiny_spike: float = 1e-9,
         adam_k_warmup_steps: int = 0,
+        adam_nnmf_factor: bool = False,
     ):
         if not (lr >= 0.0):
             raise ValueError(f"Learning-rate should be >= 0.0. Got {lr}")
@@ -165,6 +167,7 @@ class Muon_adv(torch.optim.Optimizer):
             "adam_kourkoutas_beta": adam_kourkoutas_beta, "adam_beta2_min": adam_beta2_min,
             "adam_ema_alpha": adam_ema_alpha, "adam_tiny_spike": adam_tiny_spike,
             "adam_k_warmup_steps": adam_k_warmup_steps,
+            "adam_nnmf_factor":adam_nnmf_factor,
         }
         self.stochastic_rounding = stochastic_rounding
         self.compiled_optimizer = compiled_optimizer

{adv_optm-1.2.dev15 → adv_optm-1.2.dev17}/adv_optm/util/Kourkoutas.py

@@ -155,4 +155,5 @@ class KourkoutasHelper:
         """
         layer_key = self.optimizer.layer_key_fn(p)
         # The default is the max value, which is correct for unmapped params or edge cases
-        return self.layer_state.get(layer_key, {}).get('dynamic_beta2', group['betas'][1])
+        beta2_default = group.get('betas', group.get('adam_betas'))[1] if group.get('betas', group.get('adam_betas')) else 0.999
+        return self.layer_state.get(layer_key, {}).get('dynamic_beta2', beta2_default)

{adv_optm-1.2.dev15 → adv_optm-1.2.dev17}/adv_optm.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: adv_optm
-Version: 1.2.dev15
+Version: 1.2.dev17
 Summary: A family of highly efficient, lightweight yet powerful optimizers.
 Home-page: https://github.com/Koratahiu/Advanced_Optimizers
 Author: Koratahiu

{adv_optm-1.2.dev15 → adv_optm-1.2.dev17}/setup.py

@@ -5,7 +5,7 @@ with open("README.md", "r", encoding="utf-8") as fh:
 
 setup(
     name="adv_optm",
-    version="1.2.dev15",
+    version="1.2.dev17",
     author="Koratahiu",
     author_email="hiuhonor@gmail.com",
     license='Apache 2.0',