adv-optm 1.2.dev19__py3-none-any.whl → 2.dev2__py3-none-any.whl

adv_optm/optim/Lion_adv.py

@@ -27,10 +27,6 @@ class Lion_adv(torch.optim.Optimizer):
         stochastic_rounding (bool, optional): whether to use stochastic
             rounding for BF16 parameter updates (default: True).
         cautious_mask (bool): whether to use the cautious masking technique. (default: False).
-        clip_threshold (float, optional): whether to clip the gradients norm
-            per-parameter as proposed in the paper `Lions and Muons: Optimization via
-            Stochastic Frank-Wolfe` (https://arxiv.org/abs/2506.04192) to make Lion more stable
-            (default: 0.0).
         nnmf_factor (bool): whether to use the factorization or use the
             uncompressed optimizer. (default: True)
     """
@@ -41,12 +37,13 @@ class Lion_adv(torch.optim.Optimizer):
         lr: float = 1e-4,
         betas: Tuple[float, float] = (0.9, 0.99),
         weight_decay: float = 0.0,
-        vector_reshape: bool = True,
+        vector_reshape: bool = False,
         stochastic_rounding: bool = True,
         orthogonal_gradient: bool = False,
         cautious_mask: bool = False,
-        clip_threshold: float = 0.0,
         nnmf_factor: bool = True,
+        # Compiled
+        compiled_optimizer: bool = False,
     ):
         if not lr > 0.0:
             raise ValueError(f"Learning rate must be > 0.0, but got {lr}")
@@ -61,13 +58,20 @@ class Lion_adv(torch.optim.Optimizer):
             weight_decay=weight_decay,
             vector_reshape=vector_reshape,
             orthogonal_gradient=orthogonal_gradient,
-            clip_threshold=clip_threshold,
+            compiled_optimizer=compiled_optimizer,
         )
         self.stochastic_rounding = stochastic_rounding
         self.cautious_mask = cautious_mask
         self.factored = nnmf_factor
         super().__init__(params, defaults)
 
+        self.init_step()
+
+        if compiled_optimizer:
+            torch._dynamo.config.cache_size_limit = 8192
+            self.compile(fullgraph=True)
+
+
     @property
     def supports_fused_back_pass(self) -> bool:
         return True
@@ -80,50 +84,50 @@ class Lion_adv(torch.optim.Optimizer):
     def supports_flat_params(self) -> bool:
         return False
 
-    @torch.no_grad()
-    def step_parameter(self, p: torch.Tensor, group: dict, i: Optional[int] = None):
-        """Performs a single optimization step on a single parameter."""
-        if p.grad is None:
-            return
+    def init_step(self):
+        for group in self.param_groups:
+            for i, p in enumerate(group['params']):
+                self.__init_state(p, group)
 
-        grad = p.grad
-        if grad.dtype != torch.float32 and self.factored:
-            grad = grad.float()
-        if group["clip_threshold"] > 0.0:
-            grad_norm = torch.norm(grad.detach())
-            if grad_norm > group["clip_threshold"]:
-                clip_coef = group["clip_threshold"] / grad_norm
-                grad.mul_(clip_coef)
-        if group["orthogonal_gradient"]:
-            grad = _orthogonalize_gradient(p, grad)
+    @torch.no_grad()
+    def __init_state(self, p, group):
         state = self.state[p]
 
-        # State Initialization
-        if 'step' not in state:
-            state['step'] = 0
+        if len(state) == 0:
 
-            should_factor = (
+            state['factored'] = (
                 self.factored and
                 not (len(p.shape) == 1 and not group['vector_reshape'])
             )
 
-            state['factored'] = should_factor
-
             dtype = torch.float32 if self.factored else p.dtype
 
             if state['factored']:
                 state['effective_shape'] = _get_effective_shape(p.numel())
                 d1, d2 = state['effective_shape']
-                state['mu_m_nmf'] = torch.zeros(d1, device=p.device, dtype=dtype) 
+                state['mu_m_nmf'] = torch.zeros(d1, device=p.device, dtype=dtype)
                 state['mv_m_nmf'] = torch.zeros(d2, device=p.device, dtype=dtype)
                 packed_d2 = (d2 + 7) // 8
                 state['sign'] = torch.zeros((d1, packed_d2), dtype=torch.uint8, device=p.device)
             else: # Fallback to standard Lion
                 state['exp_avg'] = torch.zeros_like(p, device=p.device, dtype=dtype)
 
-        state['step'] += 1
+    @torch.no_grad()
+    def __step_parameter(self, p: torch.Tensor, group: dict, lr: torch.Tensor | float):
+        """Performs a single optimization step on a single parameter."""
+        if p.grad is None:
+            return
+
+        grad = p.grad
+        if grad.dtype != torch.float32 and self.factored:
+            grad = grad.float()
+        if group["orthogonal_gradient"]:
+            grad = _orthogonalize_gradient(p, grad)
+
+        state = self.state[p]
+
+
         beta1, beta2 = group["betas"]
-        lr = group["lr"]
 
         if state['factored']:
             # Factored Path
@@ -138,16 +142,16 @@ class Lion_adv(torch.optim.Optimizer):
                 exp_avg = exp_avg.float()
 
             # Compute update term c_t
-            signed_update = exp_avg.clone().mul_(beta1).add_(grad_reshaped, alpha=(1-beta1)).sign_()
+            update = exp_avg.clone().mul_(beta1).add_(grad_reshaped, alpha=(1-beta1)).sign_()
 
             if self.cautious_mask:
-                mask = (signed_update * grad_reshaped > 0).to(grad_reshaped.dtype)
+                mask = (update * grad_reshaped > 0).to(grad_reshaped.dtype)
                 mask.div_(mask.mean().clamp_(min=1e-3))
-                signed_update.mul_(mask)
+                update.mul_(mask)
                 del mask
 
             # Parameter update
-            update_for_param = signed_update.view(p.shape).mul_(lr)
+            update = update.view(p.shape).mul_(lr)
 
             # Standard Lion momentum update
             exp_avg.mul_(beta2).add_(grad_reshaped, alpha=1-beta2)
@@ -165,15 +169,15 @@ class Lion_adv(torch.optim.Optimizer):
             # Compute update term and sign for the update
             if exp_avg.dtype != torch.float32 and self.factored:
                 exp_avg = exp_avg.float()
-            signed_update = exp_avg.clone().mul_(beta1).add_(grad, alpha=(1-beta1)).sign_()
+            update = exp_avg.clone().mul_(beta1).add_(grad, alpha=(1-beta1)).sign_()
 
             if self.cautious_mask:
-                mask = (signed_update * grad > 0).to(grad.dtype)
+                mask = (update * grad > 0).to(grad.dtype)
                 mask.div_(mask.mean().clamp_(min=1e-3))
-                signed_update.mul_(mask)
+                update.mul_(mask)
                 del mask
 
-            update_for_param = signed_update.mul_(lr)
+            update.mul_(lr)
 
             # Standard Lion momentum update
             exp_avg.mul_(beta2).add_(grad, alpha=1-beta2)
@@ -188,11 +192,23 @@ class Lion_adv(torch.optim.Optimizer):
                 )
 
         if p.dtype == torch.bfloat16 and self.stochastic_rounding:
-            add_stochastic_(p.data, -update_for_param)
+            add_stochastic_(p.data, -update)
+        else:
+            p.data.add_(-update)
+
+        del update
+
+    @torch.no_grad()
+    def step_parameter(self, p: torch.Tensor, group: dict, i: int | None = None):
+
+        if not group.get('compiled_optimizer', False):
+            self.__step_parameter(p, group, group["lr"])
         else:
-            p.data.add_(-update_for_param)
+            lr_tensor = torch.tensor(group["lr"], device=p.device)
+            self._compiled_step_parameter(p, group, lr_tensor)
 
-        del update_for_param
+    def compile(self, *args, **kwargs):
+        self._compiled_step_parameter = torch.compile(self.__step_parameter, *args, **kwargs)
 
     @torch.no_grad()
     def step(self, closure: Optional[callable] = None):
@@ -207,4 +223,4 @@ class Lion_adv(torch.optim.Optimizer):
                 if p.grad is not None:
                     self.step_parameter(p, group, i)
 
-        return loss
+        return loss

adv_optm/optim/Muon_adv.py

@@ -41,7 +41,6 @@ class Muon_adv(torch.optim.Optimizer):
             stability. (default: 100.0)
         stochastic_rounding (bool): whether to use stochastic rounding for
             BF16 parameter updates (default: True).
-        orthogonal_gradient (bool): whether to use OrthoGrad.  (default: False)
         vector_reshape_muon (bool): whether to reshape 1D vectors into 2D
             matrices for muon NewtonSchulz (default: False).
         vector_reshape (bool): whether to reshape 1D vectors into 2D
@@ -60,6 +59,7 @@ 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)
@@ -76,7 +76,6 @@ 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__(
@@ -92,7 +91,6 @@ class Muon_adv(torch.optim.Optimizer):
         Simplified_AdEMAMix: bool = False,
         alpha_grad: float = 100.0,
         stochastic_rounding: bool = True,
-        orthogonal_gradient: bool = False,
         vector_reshape_muon: bool = False,
         vector_reshape: bool = False,
         nnmf_factor: bool = False,
@@ -104,6 +102,7 @@ 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,
@@ -126,7 +125,6 @@ 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}")
@@ -149,13 +147,13 @@ class Muon_adv(torch.optim.Optimizer):
             "vector_reshape": vector_reshape,
             "vector_reshape_muon": vector_reshape_muon,
             "Simplified_AdEMAMix": Simplified_AdEMAMix, "alpha_grad": alpha_grad,
-            "orthogonal_gradient": orthogonal_gradient,
             'compiled_optimizer': compiled_optimizer,
             # Low-rank Ortho
             "low_rank_ortho": low_rank_ortho, "ortho_rank": ortho_rank,
             # 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
@@ -167,7 +165,6 @@ 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
@@ -293,10 +290,6 @@ class Muon_adv(torch.optim.Optimizer):
         nesterov = group['nesterov']
         Simplified_AdEMAMix = group['Simplified_AdEMAMix']
         alpha_grad = group['alpha_grad']
-        if grad.dtype != torch.float32 and state.get('factored', False):
-            grad = grad.float()
-        if group.get("orthogonal_gradient"):
-            grad = _orthogonalize_gradient(p, grad)
 
         if state['factored']: # Factored Muon
 
@@ -363,7 +356,11 @@ 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
-                update.div_(v_t.sqrt().unsqueeze(1).add_(group['normuon_eps']))
+                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']))
                 # Scale learning rate
                 update_norm = torch.linalg.vector_norm(update)
 
@@ -464,7 +461,11 @@ 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
-                    update.div_(v_t.sqrt().unsqueeze(1).add_(group['normuon_eps']))
+                    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']))
                     # 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'])