adv-optm 1.2.dev18__py3-none-any.whl → 2.dev2__py3-none-any.whl

adv_optm/optim/Adopt_adv.py

@@ -56,13 +56,6 @@ class Adopt_adv(torch.optim.Optimizer):
             before it is added to the fast momentum term (`update = mt + alpha * mt_slow`).
             A higher value increases the stabilizing influence of the slow
             momentum. (default: 5.0)
-        t_alpha (Optional[int]): The number of steps for a linear warmup of the
-            `alpha` parameter (only used when `use_AdEMAMix` is `True`). This is
-            highly recommended to prevent instability at the beginning of training,
-            as it gradually introduces the stabilizing slow momentum term. During
-            the warmup, `alpha` ramps from 0 to its target value. If `None`,
-            the scheduler is disabled and the full `alpha` value is used from
-            the start. (default: None)
         Simplified_AdEMAMix (bool): whether to use the Simplified AdEMAMix update rule.
             This changes the EMA to accumulator and the update numerator to `alpha_grad * grad + mt`, which can be
             more responsive, especially for small batch sizes. Enabling this will
@@ -90,10 +83,10 @@ class Adopt_adv(torch.optim.Optimizer):
         k_logging (int): if > 0 and kourkoutas_beta=True, enables periodic console
             logging of Kourkoutas-β statistics (min, max, mean of `β₂` across layers)
             every logging steps. Useful for debugging and tuning. Set to 0 to disable
-            logging (default: 0). 
+            logging (default: 0).
         layer_key_fn (Optional[Callable]): A function that takes a parameter `p`
             and returns a unique, hashable key representing its "layer" or "bucket".
-            If `None`, parameters are bucketed by their memory ID (tensor-wise).
+            If `None`, parameters are bucketed by their shape.
             (default: None)
         nnmf_factor (bool): whether to use the factorization or disable it to use
             the uncompressed optimizer. (default: False)
@@ -107,7 +100,7 @@ class Adopt_adv(torch.optim.Optimizer):
         eps: float = 1e-6,
         weight_decay: float = 0.0,
         clip_lambda: Optional[Callable[[int], float]] = lambda step: step**0.25,
-        vector_reshape: bool = True,
+        vector_reshape: bool = False,
         stochastic_rounding: bool = True,
         use_atan2: bool = False,
         cautious_mask: bool = False,
@@ -116,7 +109,6 @@ class Adopt_adv(torch.optim.Optimizer):
         use_AdEMAMix: bool = False,
         beta3_ema: float = 0.9999,
         alpha: float = 5.0,
-        t_alpha: int | None = None,
         Simplified_AdEMAMix: bool = False,
         alpha_grad: float = 100.0,
         kourkoutas_beta: bool = False,
@@ -127,6 +119,8 @@ class Adopt_adv(torch.optim.Optimizer):
         k_logging: int = 0,
         layer_key_fn: Optional[Callable] = None,
         nnmf_factor: bool = False,
+        # Compiled
+        compiled_optimizer: bool = False,
     ):
         if not (lr >= 0.0):
             raise ValueError(f"Learning-rate should be >= 0.0. Got {lr}")
@@ -141,7 +135,8 @@ class Adopt_adv(torch.optim.Optimizer):
             cautious_mask = False
         if betas[0] == 0.0 and Simplified_AdEMAMix:
             raise ValueError(f"Beta1 cannot be 0.0 when using Simplified_AdEMAMix. Got {betas[0]}")
-        if kourkoutas_beta and not (betas[1] > beta2_min): raise ValueError(f"For Kourkoutas-β, betas[1] (as beta2_max) must be > beta2_min. Got {betas[1]} and {beta2_min}")
+        if kourkoutas_beta and not (betas[1] > beta2_min):
+            raise ValueError(f"For Kourkoutas-β, betas[1] (as beta2_max) must be > beta2_min. Got {betas[1]} and {beta2_min}")
         if use_AdEMAMix and Simplified_AdEMAMix:
             print("Warning: use_AdEMAMix is incompatible with Simplified_AdEMAMix, Disabling use_AdEMAMix.")
         if grams_moment and Simplified_AdEMAMix:
@@ -152,9 +147,10 @@ class Adopt_adv(torch.optim.Optimizer):
         defaults = {
             "lr": lr, "betas": betas, "eps": eps, "weight_decay": weight_decay,
             "vector_reshape": vector_reshape, "beta3_ema": beta3_ema, "alpha": alpha,
-            "t_alpha": t_alpha, "alpha_grad": alpha_grad, 
+            "alpha_grad": alpha_grad,
             "kourkoutas_beta": kourkoutas_beta, "beta2_min": beta2_min, "ema_alpha": ema_alpha,
             "tiny_spike": tiny_spike, "k_warmup_steps": k_warmup_steps, "k_logging": k_logging,
+            "compiled_optimizer": compiled_optimizer,
         }
         self.clip_lambda = clip_lambda
         self.stochastic_rounding = stochastic_rounding
@@ -169,9 +165,17 @@ class Adopt_adv(torch.optim.Optimizer):
         self.layer_key_fn = layer_key_fn
         super().__init__(params, defaults)
 
+        self.init_step()
+
         if self.kourkoutas_beta:
             self.kourkoutas_helper = KourkoutasHelper(self)
 
+        self.global_step = 0
+
+        if compiled_optimizer:
+            torch._dynamo.config.cache_size_limit = 8192
+            self.compile(fullgraph=True)
+
     @property
     def supports_fused_back_pass(self): return True
     @property
@@ -179,29 +183,22 @@ class Adopt_adv(torch.optim.Optimizer):
     @property
     def supports_flat_params(self): return False
 
-    @torch.no_grad()
-    def step_parameter(self, p: torch.Tensor, group: dict, i: int | None = None):
-        if p.grad is None:
-            return
+    def init_step(self):
+        for group in self.param_groups:
+            for p in group['params']:
+                self.__init_state(p, group)
 
-        grad = p.grad
-        if self.factored and grad.dtype != torch.float32:
-            grad = grad.float()
-        if self.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']:
@@ -210,55 +207,75 @@ class Adopt_adv(torch.optim.Optimizer):
 
                 # m_0 = 0
                 if group['betas'][0] > 0:
-                    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)
                     if not self.grams_moment:
                         packed_d2 = (d2 + 7) // 8
                         state['sign'] = torch.zeros((d1, packed_d2), dtype=torch.uint8, device=p.device)
                 if self.use_AdEMAMix:
-                    state['mu_m_slow_nmf'] = torch.zeros(d1, device=p.device, dtype=dtype) 
+                    state['mu_m_slow_nmf'] = torch.zeros(d1, device=p.device, dtype=dtype)
                     state['mv_m_slow_nmf'] = torch.zeros(d2, device=p.device, dtype=dtype)
                     packed_d2 = (d2 + 7) // 8
                     state['sign_slow'] = torch.zeros((d1, packed_d2), dtype=torch.uint8, device=p.device)
-                # v_0 = g_0^2 (SMMF_ADOPT NMF storage)
-                vt_init = grad.view(d1, d2).square_()
-                # Allocate NMF factors for v
-                state['mu_v_nmf'] = torch.zeros(d1, device=p.device, dtype=dtype) 
-                state['mv_v_nmf'] = torch.zeros(d2, device=p.device, dtype=dtype)
-                # Initialize v_0 using NMF
-                _nnmf(vt_init, out=(state['mu_v_nmf'], state['mv_v_nmf']))
+
             else: # Fallback for non-factored tensors
                 if group['betas'][0] > 0:
                     state['exp_avg'] = torch.zeros_like(p, dtype=dtype) # m_0
                 if self.use_AdEMAMix:
                     state['exp_avg_slow'] = torch.zeros_like(p, dtype=dtype)
-                state['exp_avg_sq'] = grad.square()   # v_0
+
+    @torch.no_grad()
+    def __init_step(self, p, group):
+        if p.grad is None:
+            return
+        
+        state = self.state[p]
+
+        if 'exp_avg_sq' in state or 'mu_v_nmf' in state:
+            return
+
+        grad = p.grad
+        dtype = torch.float32 if self.factored else p.dtype
+
+        if state['factored']:
+            d1, d2 = state['effective_shape']
+            # v_0 = g_0^2 (SMMF_ADOPT NMF storage)
+            vt_init = grad.view(d1, d2).square_()
+            # Allocate NMF factors for v
+            state['mu_v_nmf'] = torch.zeros(d1, device=p.device, dtype=dtype)
+            state['mv_v_nmf'] = torch.zeros(d2, device=p.device, dtype=dtype)
+            # Initialize v_0 using NMF
+            _nnmf(vt_init, out=(state['mu_v_nmf'], state['mv_v_nmf']))
+            del vt_init
+        else:
+            state['exp_avg_sq'] = grad.square()   # v_0
+
+
+    @torch.no_grad()
+    def __step_parameter(self, p: torch.Tensor, group: dict, lr: torch.Tensor | float):
+        if p.grad is None:
+            return
+
+        grad = p.grad
+        if self.factored and grad.dtype != torch.float32:
+            grad = grad.float()
+        if self.orthogonal_gradient:
+            grad = _orthogonalize_gradient(p, grad)
+        state = self.state[p]
+
 
         beta1, beta2 = group['betas']
 
-        current_step = state['step']
         if group.get('kourkoutas_beta', False):
-            # Call prepare_step() once at the beginning of the step for all params
-            self.kourkoutas_helper.maybe_prepare_step(current_step)
             # Accumulate current grad's norm for the *next* step
             self.kourkoutas_helper.accumulate_gradient_sq_norm(p, grad)
             # Get the dynamic beta2 calculated in prepare_step()
-            beta2 = self.kourkoutas_helper.get_beta2(p, group, current_step)
-
-        # The first step is for initialization only (skip when use_atan2 as it's scale invariant).
-        if state['step'] == 0 and not self.use_atan2:
-            state['step'] += 1
-            return
+            beta2 = self.kourkoutas_helper.get_beta2(p, group)
 
         if self.use_AdEMAMix:
             beta3_ema = group['beta3_ema']
             alpha = group['alpha']
-            t_alpha = group['t_alpha']
-            # Use step+1 for 1-based step count in scheduler
-            alpha_step = state['step'] + 1
-            alpha_t = alpha
-            if t_alpha is not None and t_alpha > 0 and alpha_step < t_alpha:
-                alpha_t = min(alpha_step * alpha / t_alpha, alpha)
+
         if self.Simplified_AdEMAMix:
             alpha_grad = group["alpha_grad"]
 
@@ -296,7 +313,7 @@ class Adopt_adv(torch.optim.Optimizer):
             else:
                 normalized_grad = grad_reshaped / denom.add_(group['eps'])
                 if self.clip_lambda is not None:
-                    clip_val = self.clip_lambda(state['step'])
+                    clip_val = self.clip_lambda(self.global_step)
                     normalized_grad.clamp_(-clip_val, clip_val)
             del denom
 
@@ -317,9 +334,9 @@ class Adopt_adv(torch.optim.Optimizer):
             if self.use_AdEMAMix:
                 mt_slow.mul_(beta3_ema).add_(normalized_grad, alpha=1.0 - beta3_ema)
                 if beta1 > 0:
-                    update = torch.add(mt, mt_slow, alpha=alpha_t)
+                    update = torch.add(mt, mt_slow, alpha=alpha)
                 else:
-                    update = torch.add(normalized_grad, mt_slow, alpha=alpha_t)
+                    update = torch.add(normalized_grad, mt_slow, alpha=alpha)
             elif self.Simplified_AdEMAMix:
                 update = torch.add(mt, normalized_grad, alpha=alpha_grad)
             else:
@@ -328,9 +345,9 @@ class Adopt_adv(torch.optim.Optimizer):
             update = update.view(p.shape)
 
             if self.use_atan2:
-                update.mul_(group['lr'] * 1.2732395447351628)
+                update.mul_(lr * 1.2732395447351628)
             else:
-                update.mul_(group['lr']) 
+                update.mul_(lr)
 
             # Update second moment v_t for the *next* step using raw g_t
             vt.mul_(beta2).addcmul_(grad_reshaped, grad_reshaped, value=1.0 - beta2)
@@ -353,7 +370,7 @@ class Adopt_adv(torch.optim.Optimizer):
             del vt
 
         else: # Standard ADOPT logic for non-factored tensors
-            v = state['exp_avg_sq'] # v_{t-1}                
+            v = state['exp_avg_sq'] # v_{t-1}
 
             # ADOPT Step A: Decorrelate g_t using v_{t-1}
             denom = v.sqrt()
@@ -363,7 +380,7 @@ class Adopt_adv(torch.optim.Optimizer):
             else:
                 normalized_grad = grad / denom.add_(group['eps'])
                 if self.clip_lambda is not None:
-                    clip_val = self.clip_lambda(state['step'])
+                    clip_val = self.clip_lambda(self.global_step)
                     normalized_grad.clamp_(-clip_val, clip_val)
             del denom
 
@@ -387,18 +404,18 @@ class Adopt_adv(torch.optim.Optimizer):
                 m_slow = state['exp_avg_slow']
                 m_slow.mul_(beta3_ema).add_(normalized_grad, alpha=1.0 - beta3_ema)
                 if beta1 > 0:
-                    update = torch.add(m, m_slow, alpha=alpha_t)
+                    update = torch.add(m, m_slow, alpha=alpha)
                 else:
-                    update = torch.add(normalized_grad, m_slow, alpha=alpha_t)
+                    update = torch.add(normalized_grad, m_slow, alpha=alpha)
             elif self.Simplified_AdEMAMix:
                 update = torch.add(m, normalized_grad, alpha=alpha_grad)
             else:
                 update = m.clone() if beta1 > 0 else normalized_grad
 
             if self.use_atan2:
-                update.mul_(group['lr'] * 1.2732395447351628)
+                update.mul_(lr * 1.2732395447351628)
             else:
-                update.mul_(group['lr']) 
+                update.mul_(lr)
 
             # Update second moment v_t for the next step using raw g_t
             v.mul_(beta2).addcmul_(grad, grad.conj(), value=1 - beta2)
@@ -406,9 +423,9 @@ class Adopt_adv(torch.optim.Optimizer):
         # Parameter Update
         if group["weight_decay"] != 0:
             if p.dtype == torch.bfloat16 and self.stochastic_rounding:
-                add_stochastic_(p.data, p.data, alpha=-group["weight_decay"] * group["lr"])
+                add_stochastic_(p.data, p.data, alpha=-group["weight_decay"] * lr)
             else:
-                p.data.add_(p.data, alpha=-group["weight_decay"] * group["lr"])
+                p.data.add_(p.data, alpha=-group["weight_decay"] * lr)
 
         if p.dtype == torch.bfloat16 and self.stochastic_rounding:
             add_stochastic_(p.data, -update)
@@ -416,7 +433,33 @@ class Adopt_adv(torch.optim.Optimizer):
             p.data.add_(-update)
         del update
 
-        state['step'] += 1
+    @torch.no_grad()
+    def step_parameter(self, p: torch.Tensor, group: dict, i: int | None = None):
+        if self.global_step is None and 'step' in self.state[p]:
+            # For backward compatibility
+            self.global_step = self.state[p]['step']
+        
+        if self.global_step == 0:
+            self.__init_step(p, group)
+
+        # The first step is for initialization only (skip when use_atan2 as it's scale invariant).
+        if self.global_step == 0 and not self.use_atan2:
+            self.global_step += 1
+            return
+
+        if group.get('kourkoutas_beta', False):
+            # Prepare Kourkoutas-β once per step using the global step counter.
+            self.kourkoutas_helper.maybe_prepare_step(self.global_step)
+
+        if not group.get('compiled_optimizer', False):
+            self.__step_parameter(p, group, group['lr'])
+        else:
+            lr_tensor = torch.tensor(group['lr'], device=p.device)
+            self._compiled_step_parameter(p, group, lr_tensor)
+
+    def compile(self, *args, **kwargs):
+        self._compiled_step_parameter = torch.compile(self.__step_parameter, *args, **kwargs)
+
 
     @torch.no_grad()
     def step(self, closure=None):
@@ -430,4 +473,6 @@ class Adopt_adv(torch.optim.Optimizer):
             for i, p in enumerate(group['params']):
                 self.step_parameter(p, group, i)
 
-        return loss
+        self.global_step += 1
+
+        return loss

adv_optm/optim/Lion_Prodigy_adv.py

@@ -27,17 +27,13 @@ class Lion_Prodigy_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)
         d0 (float):
             Initial D estimate for D-adaptation (default 1e-6). Rarely needs changing.
         d_coef (float):
             Coefficient in the expression for the estimate of d (default 1.0).
-            Values such as 0.5 and 2.0 typically work as well. 
+            Values such as 0.5 and 2.0 typically work as well.
             Changing this parameter is the preferred way to tune the method.
         growth_rate (float):
             prevent the D estimate from growing faster than this multiplicative rate.
@@ -47,8 +43,8 @@ class Lion_Prodigy_adv(torch.optim.Optimizer):
             If you're using sharded parameters, this should be set to True. The optimizer
             will attempt to auto-detect this, but if you're using an implementation other
             than PyTorch's builtin version, the auto-detection won't work.
-        slice_p (int): Reduce memory usage by calculating LR adaptation statistics on only every 
-            pth entry of each tensor. For values greater than 1 this an an approximation to standard 
+        slice_p (int): Reduce memory usage by calculating LR adaptation statistics on only every
+            pth entry of each tensor. For values greater than 1 this an an approximation to standard
             Prodigy. Values ~11 are reasonable (default 11).
         prodigy_steps (int): If greater than zero, disable Prodigy's stepsize adjustments
             after the specified optimiser step and release all state memory required by Prodigy
@@ -64,11 +60,10 @@ class Lion_Prodigy_adv(torch.optim.Optimizer):
         lr: float = 1,
         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 = False,
         # prodigy parameters
         beta3: float = None,
@@ -80,6 +75,8 @@ class Lion_Prodigy_adv(torch.optim.Optimizer):
         slice_p: int = 11,
         prodigy_steps: int = 0,
         d_limiter: bool = True,
+        # Compiled
+        compiled_optimizer: bool = False,
     ):
         if not lr > 0.0:
             raise ValueError(f"Learning rate must be > 0.0, but got {lr}")
@@ -94,21 +91,28 @@ class Lion_Prodigy_adv(torch.optim.Optimizer):
             weight_decay=weight_decay,
             vector_reshape=vector_reshape,
             orthogonal_gradient=orthogonal_gradient,
-            clip_threshold=clip_threshold,
             beta3=beta3, d=d0, d0=d0, d_max=d0, d_numerator=0.0, d_coef=d_coef,
-            growth_rate=growth_rate, safeguard_warmup=safeguard_warmup, k=0, slice_p=slice_p,
+            growth_rate=growth_rate, safeguard_warmup=safeguard_warmup, slice_p=slice_p,
             fsdp_in_use=fsdp_in_use,
             prodigy_steps=prodigy_steps,
             d_limiter=d_limiter,
+            compiled_optimizer=compiled_optimizer,
         )
         self.stochastic_rounding = stochastic_rounding
         self.cautious_mask = cautious_mask
         self.factored = nnmf_factor
         self.fsdp_in_use = fsdp_in_use
         super().__init__(params, defaults)
-        # Global state for accumulating metrics across parameter updates within a single step.
+        # Use the device of the first parameter to avoid hardcoding '.cuda()'
+        self.device = self.param_groups[0]['params'][0].device
+
+        self.global_step = 0
         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
@@ -124,14 +128,13 @@ class Lion_Prodigy_adv(torch.optim.Optimizer):
     def init_step(self):
         """Resets accumulators and calculates dlr for the upcoming step."""
         self.d_denom = 0.0
-        
+
         g_group = self.param_groups[0]
         self.beta1, self.beta2 = g_group['betas']
         self.beta3 = g_group['beta3']
         if self.beta3 is None:
             self.beta3 = math.sqrt(self.beta2)
-        
-        k = g_group['k']
+
         self.d = g_group['d']
         lr = g_group['lr']
 
@@ -139,38 +142,21 @@ class Lion_Prodigy_adv(torch.optim.Optimizer):
 
         self.d_numerator = g_group.get('d_numerator', 0.0) * self.beta3
 
-    @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
-
-        if hasattr(p, "_fsdp_flattened"):
-            self.fsdp_in_use = True
+        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
 
             slice_p = group['slice_p']
@@ -185,13 +171,28 @@ class Lion_Prodigy_adv(torch.optim.Optimizer):
             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)
 
+    @torch.no_grad()
+    def __step_parameter(self, p: torch.Tensor, group: dict, d: torch.Tensor | float, dlr: 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]
+
+
         if state['factored']:
             # Factored Path
             d1, d2 = state['effective_shape']
@@ -205,7 +206,7 @@ class Lion_Prodigy_adv(torch.optim.Optimizer):
                 exp_avg = exp_avg.float()
 
             # Compute update term c_t = β1*m_{t-1} + (1-β1)*g_t
-            signed_update = exp_avg.clone().mul_(self.beta1).add_(grad_reshaped, alpha=self.d * (1-self.beta1)).sign_()
+            signed_update = exp_avg.clone().mul_(self.beta1).add_(grad_reshaped, alpha=d * (1-self.beta1)).sign_()
 
             if self.cautious_mask:
                 mask = (signed_update * grad_reshaped > 0).to(grad_reshaped.dtype)
@@ -214,10 +215,10 @@ class Lion_Prodigy_adv(torch.optim.Optimizer):
                 del mask
 
             # Parameter update: p_t = p_{t-1} - lr * sign(c_t)
-            update_for_param = signed_update.view(p.shape).mul(self.dlr)
+            update_for_param = signed_update.view(p.shape).mul(dlr)
 
             # Update momentum m_t = β2*m_{t-1} + (1-β2)*lr*g_t
-            exp_avg.mul_(self.beta2).add_(grad_reshaped, alpha=self.d * (1 - self.beta2))
+            exp_avg.mul_(self.beta2).add_(grad_reshaped, alpha=d * (1 - self.beta2))
             del grad_reshaped
 
             # Compress new momentum m_t and store factors
@@ -232,7 +233,7 @@ class Lion_Prodigy_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_(self.beta1).add_(grad, alpha=self.d * (1-self.beta1)).sign_()
+            signed_update = exp_avg.clone().mul_(self.beta1).add_(grad, alpha=d * (1-self.beta1)).sign_()
 
             if self.cautious_mask:
                 mask = (signed_update * grad > 0).to(grad.dtype)
@@ -240,41 +241,18 @@ class Lion_Prodigy_adv(torch.optim.Optimizer):
                 signed_update.mul_(mask)
                 del mask
 
-            update_for_param = signed_update.mul(self.dlr)
-
-            # Update momentum 
-            exp_avg.mul_(self.beta2).add_(grad, alpha=self.d * (1 - self.beta2))
-
-        prodigy_steps = group['prodigy_steps']
-        if prodigy_steps <= 0 or group['k'] < prodigy_steps:
-            # --- Accumulate Prodigy stats ---
-            d0, safeguard_warmup, slice_p = group['d0'], group['safeguard_warmup'], group['slice_p']
-            s, p0 = state['s'], state['p0']
-            grad_flat = grad.flatten().float()
-            p_flat = p.data.flatten().float()
-            p0 = p0.float()
-
-            self.d_numerator += (self.d / d0) * self.dlr * torch.dot(grad_flat[::slice_p], p0.data - p_flat[::slice_p]).item()
-
-            alpha = ((self.d / d0) * self.d) if safeguard_warmup else ((self.d / d0) * self.dlr)
-            s.mul_(self.beta3).add_(grad_flat[::slice_p], alpha=alpha)
-            self.d_denom += s.abs().sum().item()
+            update_for_param = signed_update.mul(dlr)
 
-            del s, p0, grad_flat, p_flat, alpha
-        else:
-            # Free memory if prodigy_steps is reached
-            if 's' in state:
-                del state['s']
-            if 'p0' in state:
-                del state['p0']
+            # Update momentum
+            exp_avg.mul_(self.beta2).add_(grad, alpha=d * (1 - self.beta2))
 
         if group["weight_decay"] != 0:
             if p.dtype == torch.bfloat16 and self.stochastic_rounding:
                 add_stochastic_(p.data, p.data,
-                                alpha=-group["weight_decay"] * self.dlr)
+                                alpha=-group["weight_decay"] * dlr)
             else:
                 p.data.add_(
-                    p.data, alpha=-group["weight_decay"] * self.dlr
+                    p.data, alpha=-group["weight_decay"] * dlr
                 )
 
         if p.dtype == torch.bfloat16 and self.stochastic_rounding:
@@ -284,6 +262,29 @@ class Lion_Prodigy_adv(torch.optim.Optimizer):
 
         del update_for_param
 
+    @torch.no_grad()
+    def step_parameter(self, p: torch.Tensor, group: dict, i: int | None = None):
+        if hasattr(p, "_fsdp_flattened"):
+            self.fsdp_in_use = True
+
+        if self.global_step is None and 'step' in self.state[p]:
+            # For backward compatibility
+            self.global_step = self.state[p]['step']
+
+        if isinstance(self.d_numerator, float):
+            self.d_numerator = torch.tensor(self.d_numerator, device=p.device)
+            self.d_denom = torch.tensor(self.d_denom, device=p.device)
+
+        if not group.get('compiled_optimizer', False):
+            self.__step_parameter(p, group, self.d, self.dlr)
+        else:
+            d_tensor = torch.tensor(self.d, device=p.device)
+            dlr_tensor = torch.tensor(self.dlr, device=p.device)
+            self._compiled_step_parameter(p, group, d_tensor, dlr_tensor)
+
+    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):
         """Performs a single optimization step."""
@@ -306,21 +307,19 @@ class Lion_Prodigy_adv(torch.optim.Optimizer):
         """Calculates the new `d` based on the accumulated stats."""
         g_group = self.param_groups[0]
         # Only perform d-adaptation if prodigy_steps has not been reached
-        prodigy_active = not (g_group.get('prodigy_steps', 0) > 0 and g_group['k'] >= g_group['prodigy_steps'])
+        prodigy_active = not (g_group.get('prodigy_steps', 0) > 0 and self.global_step >= g_group['prodigy_steps'])
 
         if prodigy_active:
             d_max, d_coef, growth_rate = g_group['d_max'], g_group['d_coef'], g_group['growth_rate']
-            
+
             if self.fsdp_in_use and dist.is_available() and dist.is_initialized():
-                # Use the device of the first parameter to avoid hardcoding '.cuda()'
-                device = self.param_groups[0]['params'][0].device
-                dist_tensor = torch.tensor([self.d_numerator, self.d_denom], device=device)
+                dist_tensor = torch.stack([self.d_numerator, self.d_denom])
                 dist.all_reduce(dist_tensor, op=dist.ReduceOp.SUM)
                 global_d_numerator = dist_tensor[0].item()
                 global_d_denom = dist_tensor[1].item()
             else:
-                global_d_numerator = self.d_numerator
-                global_d_denom = self.d_denom
+                global_d_numerator = self.d_numerator.item()
+                global_d_denom = self.d_denom.item()
 
             d_hat = self.d
             if global_d_denom > 0:
@@ -337,5 +336,4 @@ class Lion_Prodigy_adv(torch.optim.Optimizer):
                 group['d'] = self.d
                 group['d_max'] = d_max
         # Increment step counter for all groups, regardless of whether d was updated
-        for group in self.param_groups:
-            group['k'] += 1
+        self.global_step += 1