adv-optm 1.2.dev18__py3-none-any.whl → 1.2.dev19__py3-none-any.whl

adv_optm/__init__.py

@@ -20,4 +20,4 @@ __all__ = [
     "AdaMuon_adv",
 ]
 
-__version__ = "1.2.dev18"
+__version__ = "1.2.dev19"

adv_optm/optim/AdaMuon_adv.py

@@ -351,11 +351,7 @@ class AdaMuon_adv(torch.optim.Optimizer):
                 mean_squared_update = torch.mean(update.square(), dim=1)
                 v_t.mul_(beta2).add_(mean_squared_update, alpha=1 - beta2)
                 # Normalize update
-                if group['use_atan2']:
-                    a = 1.2732395
-                    update.atan2_(v_t.sqrt().unsqueeze(1)).mul_(a)
-                else:
-                    update.div_(v_t.sqrt().unsqueeze(1).add_(group['eps']))
+                update.div_(v_t.sqrt().unsqueeze(1).add_(group['eps']))
                 # Scale learning rate
                 update_norm = torch.linalg.vector_norm(update)
                 scaled_lr = group['rms_target'] * lr * (p.numel()**0.5) / update_norm.add_(group['eps'])
@@ -460,11 +456,7 @@ class AdaMuon_adv(torch.optim.Optimizer):
                     mean_squared_update = torch.mean(update.square(), dim=1)
                     v_t.mul_(beta2).add_(mean_squared_update, alpha=1 - beta2)
                     # Normalize update
-                    if group['use_atan2']:
-                        a = 1.2732395
-                        update.atan2_(v_t.sqrt().unsqueeze(1)).mul_(a)
-                    else:
-                        update.div_(v_t.sqrt().unsqueeze(1).add_(group['eps']))
+                    update.div_(v_t.sqrt().unsqueeze(1).add_(group['eps']))
                     # Scale learning rate
                     update_norm = torch.linalg.vector_norm(update)
                     scaled_lr = group['rms_target'] * lr * (p.numel()**0.5) / update_norm.add_(group['eps'])

adv_optm/optim/Muon_adv.py

@@ -60,7 +60,6 @@ class Muon_adv(torch.optim.Optimizer):
         normuon_eps (float): Epsilon for NorMuon normalization stability. (default: 1e-8)
         normuon_lr_scale (float): Scaling factor for the NorMuon learning rate.
             (default: 0.2)
-        normuon_atan2 (bool): whether to use the atan2 for NorMuon. (default: False)
         accelerated_ns (bool): If True, enables Chebyshev-accelerated Newton-Schulz, which
             dynamically calculates optimal 3rd-order polynomial coefficients. (default: False)
         cns_a_bound (float): Initial lower bound for singular values for CANS. (default: 1e-4)
@@ -105,7 +104,6 @@ class Muon_adv(torch.optim.Optimizer):
         beta2_normuon: float = 0.95,
         normuon_eps: float = 1e-8,
         normuon_lr_scale: float = 0.2,
-        normuon_atan2: bool = False,
         # CANS
         accelerated_ns: bool = False,
         cns_a_bound: float = 1e-4,
@@ -158,7 +156,6 @@ class Muon_adv(torch.optim.Optimizer):
             # NorMuon
             "normuon_variant": normuon_variant, "beta2_normuon": beta2_normuon,
             "normuon_eps": normuon_eps, "normuon_lr_scale": normuon_lr_scale,
-            "normuon_atan2": normuon_atan2,
             # CANS
             "accelerated_ns": accelerated_ns, "cns_a_bound": cns_a_bound,
             # AdamW_adv defaults
@@ -366,11 +363,7 @@ class Muon_adv(torch.optim.Optimizer):
                 mean_squared_update = torch.mean(update.square(), dim=1)
                 v_t.mul_(beta2_normuon).add_(mean_squared_update, alpha=1 - beta2_normuon)
                 # Normalize update
-                if group['normuon_atan2']:
-                    a = 1.2732395
-                    update.atan2_(v_t.sqrt().unsqueeze(1)).mul_(a)
-                else:
-                    update.div_(v_t.sqrt().unsqueeze(1).add_(group['normuon_eps']))
+                update.div_(v_t.sqrt().unsqueeze(1).add_(group['normuon_eps']))
                 # Scale learning rate
                 update_norm = torch.linalg.vector_norm(update)
 
@@ -471,11 +464,7 @@ class Muon_adv(torch.optim.Optimizer):
                     mean_squared_update = torch.mean(update.square(), dim=1)
                     v_t.mul_(beta2_normuon).add_(mean_squared_update, alpha=1 - beta2_normuon)
                     # Normalize update
-                    if group['normuon_atan2']:
-                        a = 1.2732395
-                        update.atan2_(v_t.sqrt().unsqueeze(1)).mul_(a)
-                    else:
-                        update.div_(v_t.sqrt().unsqueeze(1).add_(group['normuon_eps']))
+                    update.div_(v_t.sqrt().unsqueeze(1).add_(group['normuon_eps']))
                     # Scale learning rate
                     update_norm = torch.linalg.vector_norm(update)
                     scaled_lr = group['normuon_lr_scale'] * lr * (p.numel()**0.5) / update_norm.add_(group['normuon_eps'])

{adv_optm-1.2.dev18.dist-info → adv_optm-1.2.dev19.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: adv_optm
-Version: 1.2.dev18
+Version: 1.2.dev19
 Summary: A family of highly efficient, lightweight yet powerful optimizers.
 Home-page: https://github.com/Koratahiu/Advanced_Optimizers
 Author: Koratahiu

{adv_optm-1.2.dev18.dist-info → adv_optm-1.2.dev19.dist-info}/RECORD

@@ -1,10 +1,10 @@
-adv_optm/__init__.py,sha256=1UzgEkreoqaobiwUZ8yR-8Fnda7T7XiHQ4PhJKQocy4,380
-adv_optm/optim/AdaMuon_adv.py,sha256=VpNsw2CnU8bZThj9cJJ6HGIATPxv4VkIf3xTsUMXQAY,35027
+adv_optm/__init__.py,sha256=1AKxG--scx5Bl9G08tQcnfzAMaQVSgmW99uy3v2QWMw,380
+adv_optm/optim/AdaMuon_adv.py,sha256=7Had92OcsCiN1E9UJRyrpPV7VzHqmIvS-qM6OEcc24I,34671
 adv_optm/optim/AdamW_adv.py,sha256=jgMuRAfsnUh_2wUEZgYpJX5uwoT_kQjtMs2Xn2vJ3x0,17480
 adv_optm/optim/Adopt_adv.py,sha256=kbAeBG4bXWBvgj_qrE9W67J6c0swpEi4Erj2rfYrMXE,21252
 adv_optm/optim/Lion_Prodigy_adv.py,sha256=LEA3UYJpPeFnmxeniLNv1u2LKKj4ufx3Bq_MLw-nWXk,14617
 adv_optm/optim/Lion_adv.py,sha256=aGNAplZlyXYgVllYcV_s4bK8iC4fv6EizFoWIMNLdBc,8299
-adv_optm/optim/Muon_adv.py,sha256=0D4k8UfMSzITJwQEDfqpceD5H7HQvv0f8uyVKvdvkHo,33704
+adv_optm/optim/Muon_adv.py,sha256=tZY8K3pNBCGk1V09GbK05lJooFw92NfkF7_T548up3Q,33171
 adv_optm/optim/Prodigy_adv.py,sha256=k7f2J_RQpnrUXjwER_XOokISlQWpTSwGG-OL-bjMfBk,26061
 adv_optm/optim/Simplified_AdEMAMix.py,sha256=nEIA3yM11nBooKzHudB5l3x4UdFRBYRwiKVUkGmO0K8,12971
 adv_optm/optim/__init__.py,sha256=hpUWE6CKtt_rvMdgQVb3PtjhfZAvAxTq6hp8H8rIpBo,489
@@ -16,8 +16,8 @@ adv_optm/util/Newton_Schulz.py,sha256=bBboYw_jm5_FMf0Citl79uqNedkHOTjQnUI7rZgLBm
 adv_optm/util/One_Bit_Boolean.py,sha256=Wat49esdwohuN-OHOFMW8D0aOQgV9cP5Rl8z6yfmpos,1068
 adv_optm/util/OrthoGrad.py,sha256=NzInuBQGy_Ja__M1R9XbvqVaQ0fhGbtGgFE9YON7B3I,707
 adv_optm/util/__init__.py,sha256=CXzS703GB4gil85khZi7sgKOnbzXGBOltshIOSPqj18,435
-adv_optm-1.2.dev18.dist-info/licenses/LICENSE,sha256=HrhfyXIkWY2tGFK11kg7vPCqhgh5DcxleloqdhrpyMY,11558
-adv_optm-1.2.dev18.dist-info/METADATA,sha256=cfQdGhiRlf_-xnPKqwuCE8PR6faLYx_RC6MrkjYDqI8,14023
-adv_optm-1.2.dev18.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-adv_optm-1.2.dev18.dist-info/top_level.txt,sha256=iNfBIIzu-lPrQ7jyC56WBCcbkRwitM2nJ15-MRQ_6fg,9
-adv_optm-1.2.dev18.dist-info/RECORD,,
+adv_optm-1.2.dev19.dist-info/licenses/LICENSE,sha256=HrhfyXIkWY2tGFK11kg7vPCqhgh5DcxleloqdhrpyMY,11558
+adv_optm-1.2.dev19.dist-info/METADATA,sha256=pQm5WuMKvf5Xse10viziVK9ry1UufcYRDwOd55jad8Y,14023
+adv_optm-1.2.dev19.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+adv_optm-1.2.dev19.dist-info/top_level.txt,sha256=iNfBIIzu-lPrQ7jyC56WBCcbkRwitM2nJ15-MRQ_6fg,9
+adv_optm-1.2.dev19.dist-info/RECORD,,