adv-optm 1.2.dev13__tar.gz → 2.dev1__tar.gz

{adv_optm-1.2.dev13 → adv_optm-2.dev1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: adv_optm
-Version: 1.2.dev13
+Version: 2.dev1
 Summary: A family of highly efficient, lightweight yet powerful optimizers.
 Home-page: https://github.com/Koratahiu/Advanced_Optimizers
 Author: Koratahiu
@@ -52,7 +52,7 @@ This library integrates multiple state-of-the-art optimization techniques valida
 ### **Memory-Efficient Optimization (SMMF-inspired)**
 - **Paper**: [SMMF: Square-Matricized Momentum Factorization](https://arxiv.org/abs/2412.08894)
 - **Approach**: Uses rank-1 non-negative matrix factorization with reconstruction cycle (factor → reconstruct → update → factor)
-- **Innovation**: 
+- **Innovation**:
   - First moment split into **1-bit sign + absolute value**
   - Final storage: **four factored vectors + one 1-bit sign state**
   - Preserves Adam-like update quality with drastically reduced memory
@@ -110,7 +110,7 @@ This library integrates multiple state-of-the-art optimization techniques valida
 
 ## 🛠️ Comprehensive Feature Guide
 
-### A. Universal Safe Features  
+### A. Universal Safe Features
 *These features work with all optimizers and are generally safe to enable.*
 
 | Feature | Description | Recommended Usage | Performance Impact | Theoretical Basis | Compatibility |
@@ -165,7 +165,7 @@ This library integrates multiple state-of-the-art optimization techniques valida
 | `beta1` | 0.99 | Controls accumulator memory length:<br>• Small BS: **0.99–0.9999**<br>• Large BS: **0.9** |
 | `Grad α` | 100 | Most critical parameter:<br>• Inversely scales with batch size<br>• **100–10** for small BS (≤32)<br>• **1–0.1** for large BS (≥512) |
 
-> ⚠️ **Critical**: Requires **~100x smaller learning rate** than AdamW (e.g., 1e-6 vs 1e-4).  
+> ⚠️ **Critical**: Requires **~100x smaller learning rate** than AdamW (e.g., 1e-6 vs 1e-4).
 > For `Prodigy_Adv`, set `initial_d` to:
 > - **LoRA**: `1e-8`
 > - **Full FT**: `1e-10`
@@ -175,10 +175,10 @@ This library integrates multiple state-of-the-art optimization techniques valida
 
 #### Performance Validation
 
-**Small Batch Training (SDXL, BS=2, 1.8K steps)**  
+**Small Batch Training (SDXL, BS=2, 1.8K steps)**
 ![Training Comparison](https://github.com/user-attachments/assets/7eff0671-cc59-47fc-8b63-d5205456d649)
 
-- **🟢 Prodigy_Adv** (beta1=0.9, d0=1e-5): Final LR = 2.9e-4  
+- **🟢 Prodigy_Adv** (beta1=0.9, d0=1e-5): Final LR = 2.9e-4
 - **🔵 Prodigy_Adv + Simplified_AdEMAMix** (beta1=0.99, α=100, d0=1e-7): Final LR = 5.8e-6
 
 **Results**:
@@ -202,8 +202,8 @@ This library integrates multiple state-of-the-art optimization techniques valida
 
 Instead of using a fixed β₂ (e.g., 0.999 or 0.95), it **dynamically modulates β₂ per layer** based on a bounded *sunspike ratio*:
 
-- **During gradient bursts** → β₂ ↓ toward `Lower β₂` → faster reaction  
-- **During calm phases** → β₂ ↑ toward `The Selected β₂` → stronger smoothing  
+- **During gradient bursts** → β₂ ↓ toward `Lower β₂` → faster reaction
+- **During calm phases** → β₂ ↑ toward `The Selected β₂` → stronger smoothing
 
 This is especially effective for **noisy training, small batch sizes, and high learning rates**, where gradient norms shift abruptly due to noise or aggressive LR schedules.
 
@@ -220,17 +220,17 @@ This is especially effective for **noisy training, small batch sizes, and high l
 
 #### 📊 Performance Validation
 
-**ADAMW_ADV - full SDXL finetuning (aggressive LR: 3e-5) (BS=4, 2.5K steps)**  
+**ADAMW_ADV - full SDXL finetuning (aggressive LR: 3e-5) (BS=4, 2.5K steps)**
 <img width="1460" height="382" alt="image" src="https://github.com/user-attachments/assets/007f278a-fbac-4f3d-9cc7-274c3b959cdd" />
 
-- 🟣 Fixed `beta2=0.999`  
-- 🟠 Auto K-beta  
+- 🟣 Fixed `beta2=0.999`
+- 🟠 Auto K-beta
 
-**Observations:**  
+**Observations:**
 - K-beta is clearly better and more robust/stable for high LRs.
 
-> 📚 **Reference**:  
-> - Paper: [Kourkoutas-β: A Sunspike-Driven Adam Optimizer with Desert Flair](https://arxiv.org/abs/2508.12996)  
+> 📚 **Reference**:
+> - Paper: [Kourkoutas-β: A Sunspike-Driven Adam Optimizer with Desert Flair](https://arxiv.org/abs/2508.12996)
 > - Code: [kbeta](https://github.com/sck-at-ucy/kbeta)
 
 ---
@@ -258,7 +258,7 @@ settings:
   - factored: False  # Can be true or false, quality should not degrade due to Simplified_AdEMAMix’s high tolerance to 1-bit factorization.
 ```
 
-> ✅ **Why it works**:  
+> ✅ **Why it works**:
 > - `Kourkoutas-β` handles beta2 values
 > - `Simplified_AdEMAMix` ensures responsiveness in small-batch noise
 > - `OrthoGrad` prevents overfitting without weight decay
@@ -267,9 +267,9 @@ settings:
 
 ## 📚 References
 
-1. [Revisiting BFloat16 Training](https://arxiv.org/abs/2010.06192)  
-2. [SMMF: Square-Matricized Momentum Factorization](https://arxiv.org/abs/2412.08894)  
-3. [The AdEMAMix Optimizer](https://arxiv.org/abs/2409.03137)  
-4. [Connections between Schedule-Free Optimizers, AdEMAMix, and Accelerated SGD](https://arxiv.org/abs/2502.02431)  
-5. [AdaMeM: Memory Efficient Momentum for Adafactor](https://openreview.net/forum?id=fZqMVTz7K5)  
+1. [Revisiting BFloat16 Training](https://arxiv.org/abs/2010.06192)
+2. [SMMF: Square-Matricized Momentum Factorization](https://arxiv.org/abs/2412.08894)
+3. [The AdEMAMix Optimizer](https://arxiv.org/abs/2409.03137)
+4. [Connections between Schedule-Free Optimizers, AdEMAMix, and Accelerated SGD](https://arxiv.org/abs/2502.02431)
+5. [AdaMeM: Memory Efficient Momentum for Adafactor](https://openreview.net/forum?id=fZqMVTz7K5)
 6. [Kourkoutas-β: A Sunspike-Driven Adam Optimizer with Desert Flair](https://arxiv.org/abs/2508.12996)

{adv_optm-1.2.dev13 → adv_optm-2.dev1}/README.md

@@ -21,7 +21,7 @@ This library integrates multiple state-of-the-art optimization techniques valida
 ### **Memory-Efficient Optimization (SMMF-inspired)**
 - **Paper**: [SMMF: Square-Matricized Momentum Factorization](https://arxiv.org/abs/2412.08894)
 - **Approach**: Uses rank-1 non-negative matrix factorization with reconstruction cycle (factor → reconstruct → update → factor)
-- **Innovation**: 
+- **Innovation**:
   - First moment split into **1-bit sign + absolute value**
   - Final storage: **four factored vectors + one 1-bit sign state**
   - Preserves Adam-like update quality with drastically reduced memory
@@ -79,7 +79,7 @@ This library integrates multiple state-of-the-art optimization techniques valida
 
 ## 🛠️ Comprehensive Feature Guide
 
-### A. Universal Safe Features  
+### A. Universal Safe Features
 *These features work with all optimizers and are generally safe to enable.*
 
 | Feature | Description | Recommended Usage | Performance Impact | Theoretical Basis | Compatibility |
@@ -134,7 +134,7 @@ This library integrates multiple state-of-the-art optimization techniques valida
 | `beta1` | 0.99 | Controls accumulator memory length:<br>• Small BS: **0.99–0.9999**<br>• Large BS: **0.9** |
 | `Grad α` | 100 | Most critical parameter:<br>• Inversely scales with batch size<br>• **100–10** for small BS (≤32)<br>• **1–0.1** for large BS (≥512) |
 
-> ⚠️ **Critical**: Requires **~100x smaller learning rate** than AdamW (e.g., 1e-6 vs 1e-4).  
+> ⚠️ **Critical**: Requires **~100x smaller learning rate** than AdamW (e.g., 1e-6 vs 1e-4).
 > For `Prodigy_Adv`, set `initial_d` to:
 > - **LoRA**: `1e-8`
 > - **Full FT**: `1e-10`
@@ -144,10 +144,10 @@ This library integrates multiple state-of-the-art optimization techniques valida
 
 #### Performance Validation
 
-**Small Batch Training (SDXL, BS=2, 1.8K steps)**  
+**Small Batch Training (SDXL, BS=2, 1.8K steps)**
 ![Training Comparison](https://github.com/user-attachments/assets/7eff0671-cc59-47fc-8b63-d5205456d649)
 
-- **🟢 Prodigy_Adv** (beta1=0.9, d0=1e-5): Final LR = 2.9e-4  
+- **🟢 Prodigy_Adv** (beta1=0.9, d0=1e-5): Final LR = 2.9e-4
 - **🔵 Prodigy_Adv + Simplified_AdEMAMix** (beta1=0.99, α=100, d0=1e-7): Final LR = 5.8e-6
 
 **Results**:
@@ -171,8 +171,8 @@ This library integrates multiple state-of-the-art optimization techniques valida
 
 Instead of using a fixed β₂ (e.g., 0.999 or 0.95), it **dynamically modulates β₂ per layer** based on a bounded *sunspike ratio*:
 
-- **During gradient bursts** → β₂ ↓ toward `Lower β₂` → faster reaction  
-- **During calm phases** → β₂ ↑ toward `The Selected β₂` → stronger smoothing  
+- **During gradient bursts** → β₂ ↓ toward `Lower β₂` → faster reaction
+- **During calm phases** → β₂ ↑ toward `The Selected β₂` → stronger smoothing
 
 This is especially effective for **noisy training, small batch sizes, and high learning rates**, where gradient norms shift abruptly due to noise or aggressive LR schedules.
 
@@ -189,17 +189,17 @@ This is especially effective for **noisy training, small batch sizes, and high l
 
 #### 📊 Performance Validation
 
-**ADAMW_ADV - full SDXL finetuning (aggressive LR: 3e-5) (BS=4, 2.5K steps)**  
+**ADAMW_ADV - full SDXL finetuning (aggressive LR: 3e-5) (BS=4, 2.5K steps)**
 <img width="1460" height="382" alt="image" src="https://github.com/user-attachments/assets/007f278a-fbac-4f3d-9cc7-274c3b959cdd" />
 
-- 🟣 Fixed `beta2=0.999`  
-- 🟠 Auto K-beta  
+- 🟣 Fixed `beta2=0.999`
+- 🟠 Auto K-beta
 
-**Observations:**  
+**Observations:**
 - K-beta is clearly better and more robust/stable for high LRs.
 
-> 📚 **Reference**:  
-> - Paper: [Kourkoutas-β: A Sunspike-Driven Adam Optimizer with Desert Flair](https://arxiv.org/abs/2508.12996)  
+> 📚 **Reference**:
+> - Paper: [Kourkoutas-β: A Sunspike-Driven Adam Optimizer with Desert Flair](https://arxiv.org/abs/2508.12996)
 > - Code: [kbeta](https://github.com/sck-at-ucy/kbeta)
 
 ---
@@ -227,7 +227,7 @@ settings:
   - factored: False  # Can be true or false, quality should not degrade due to Simplified_AdEMAMix’s high tolerance to 1-bit factorization.
 ```
 
-> ✅ **Why it works**:  
+> ✅ **Why it works**:
 > - `Kourkoutas-β` handles beta2 values
 > - `Simplified_AdEMAMix` ensures responsiveness in small-batch noise
 > - `OrthoGrad` prevents overfitting without weight decay
@@ -236,9 +236,9 @@ settings:
 
 ## 📚 References
 
-1. [Revisiting BFloat16 Training](https://arxiv.org/abs/2010.06192)  
-2. [SMMF: Square-Matricized Momentum Factorization](https://arxiv.org/abs/2412.08894)  
-3. [The AdEMAMix Optimizer](https://arxiv.org/abs/2409.03137)  
-4. [Connections between Schedule-Free Optimizers, AdEMAMix, and Accelerated SGD](https://arxiv.org/abs/2502.02431)  
-5. [AdaMeM: Memory Efficient Momentum for Adafactor](https://openreview.net/forum?id=fZqMVTz7K5)  
+1. [Revisiting BFloat16 Training](https://arxiv.org/abs/2010.06192)
+2. [SMMF: Square-Matricized Momentum Factorization](https://arxiv.org/abs/2412.08894)
+3. [The AdEMAMix Optimizer](https://arxiv.org/abs/2409.03137)
+4. [Connections between Schedule-Free Optimizers, AdEMAMix, and Accelerated SGD](https://arxiv.org/abs/2502.02431)
+5. [AdaMeM: Memory Efficient Momentum for Adafactor](https://openreview.net/forum?id=fZqMVTz7K5)
 6. [Kourkoutas-β: A Sunspike-Driven Adam Optimizer with Desert Flair](https://arxiv.org/abs/2508.12996)

{adv_optm-1.2.dev13 → adv_optm-2.dev1}/adv_optm/__init__.py

@@ -20,4 +20,4 @@ __all__ = [
     "AdaMuon_adv",
 ]
 
-__version__ = "1.2.dev13"
+__version__ = "2.dev1"

{adv_optm-1.2.dev13 → adv_optm-2.dev1}/adv_optm/optim/AdamW_adv.py

@@ -49,12 +49,6 @@ class AdamW_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. (default: None)
         kourkoutas_beta (bool): whether to enable the layer-wise dynamic β₂ logic.
             If `False`, the optimizer behaves as standard AdamW. (default: False)
         beta2_min (float): The minimum value for dynamic β₂, used during periods of
@@ -72,11 +66,7 @@ class AdamW_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). 
-        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).
-            (default: None)
+            logging (default: 0).
         nnmf_factor (bool): whether to use the factorization or disable it to use
             the uncompressed optimizer. (default: False)
     """
@@ -89,7 +79,7 @@ class AdamW_adv(torch.optim.Optimizer):
         eps: float = 1e-8,
         weight_decay: float = 0.0,
         use_bias_correction: bool = True,
-        vector_reshape: bool = True,
+        vector_reshape: bool = False,
         stochastic_rounding: bool = True,
         use_atan2: bool = False,
         cautious_mask: bool = False,
@@ -98,15 +88,15 @@ class AdamW_adv(torch.optim.Optimizer):
         use_AdEMAMix: bool = False,
         beta3_ema: float = 0.9999,
         alpha: float = 5.0,
-        t_alpha: int | None = None,
         kourkoutas_beta: bool = False,
         beta2_min: float = 0.9,
         ema_alpha: float = 0.95,
         tiny_spike: float = 1e-9,
         k_warmup_steps: int = 0,
         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}")
@@ -116,7 +106,8 @@ class AdamW_adv(torch.optim.Optimizer):
             raise ValueError(f"Epsilon should be >= 0.0. Got {eps}")
         if not (weight_decay >= 0.0):
             raise ValueError(f"Weight-decay should be >= 0.0. Got {weight_decay}")
-        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 cautious_mask and grams_moment:
             print("Warning: cautious is incompatible with grams, Disabling cautious.")
@@ -126,9 +117,10 @@ class AdamW_adv(torch.optim.Optimizer):
             "lr": lr, "betas": betas, "eps": eps, "weight_decay": weight_decay,
             "vector_reshape": vector_reshape, "use_atan2": use_atan2,
             "orthogonal_gradient": orthogonal_gradient, "use_bias_correction": use_bias_correction,
-            "beta3_ema": beta3_ema, "alpha": alpha, "t_alpha": t_alpha,
+            "beta3_ema": beta3_ema, "alpha": alpha,
             "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.stochastic_rounding = stochastic_rounding
         self.cautious_mask = cautious_mask
@@ -136,12 +128,20 @@ class AdamW_adv(torch.optim.Optimizer):
         self.use_AdEMAMix = use_AdEMAMix
         self.factored = nnmf_factor
         self.kourkoutas_beta = kourkoutas_beta
-        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
@@ -154,29 +154,22 @@ class AdamW_adv(torch.optim.Optimizer):
     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 grad.dtype != torch.float32 and self.factored:
-            grad = grad.float()
-        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
             device = p.device
 
@@ -186,18 +179,18 @@ class AdamW_adv(torch.optim.Optimizer):
 
                 # First moment (m)
                 if group['betas'][0] > 0:
-                    state['mu_m_nmf'] = torch.zeros(d1, device=device, dtype=dtype) 
+                    state['mu_m_nmf'] = torch.zeros(d1, device=device, dtype=dtype)
                     state['mv_m_nmf'] = torch.zeros(d2, device=device, dtype=dtype)
                     if not self.grams_moment:
                         packed_d2 = (d2 + 7) // 8
                         state['sign'] = torch.zeros((d1, packed_d2), dtype=torch.uint8, device=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)
                 # Second moment (v)
-                state['mu_v_nmf'] = torch.zeros(d1, device=device, dtype=dtype) 
+                state['mu_v_nmf'] = torch.zeros(d1, device=device, dtype=dtype)
                 state['mv_v_nmf'] = torch.zeros(d2, device=device, dtype=dtype)
             else:  # Fallback to standard AdamW for non-factored tensors
                 if group['betas'][0] > 0:
@@ -206,37 +199,32 @@ class AdamW_adv(torch.optim.Optimizer):
                     state['exp_avg_slow'] = torch.zeros_like(p, device=device, dtype=dtype)
                 state['exp_avg_sq'] = torch.zeros_like(p, device=device, dtype=dtype)
 
+    @torch.no_grad()
+    def __step_parameter(self, p: torch.Tensor, group: dict, lr: torch.Tensor | float, bias_correction1: torch.Tensor | float, bias_correction2: torch.Tensor | float):
+        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']
 
-        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)
+            beta2 = self.kourkoutas_helper.get_beta2(p, group)
 
-        step = state['step'] + 1
-        if group['use_bias_correction']:
-            bias_correction1 = 1.0 - beta1 ** step
-            if group.get('kourkoutas_beta', False):
-                bias_correction2 = 1.0 - group['betas'][1] ** step
-                # Use beta2_max for bias correction
-            else:
-                bias_correction2 = 1.0 - beta2 ** step
-        else:
-            bias_correction1 = 1
-            bias_correction2 = 1
-        step_size = group['lr'] / bias_correction1
+        step_size = lr / bias_correction1
 
         if self.use_AdEMAMix:
             beta3_ema = group['beta3_ema']
             alpha = group['alpha']
-            t_alpha = group['t_alpha']
-            alpha_t = alpha
-            if t_alpha is not None and t_alpha > 0 and step < t_alpha:
-                alpha_t = min(step * alpha / t_alpha, alpha)
 
         if state['factored']:
             d1, d2 = state['effective_shape']
@@ -252,7 +240,7 @@ class AdamW_adv(torch.optim.Optimizer):
                 # Update momentum in full-size
                 mt.mul_(beta1).add_(grad_reshaped, alpha=1.0 - beta1)
                 if self.grams_moment:
-                    mt.copy_(grad_reshaped.sign() * mt.abs())
+                    mt = (grad_reshaped.sign().mul_(mt.abs()))
                 elif self.cautious_mask:
                     mask = (mt * grad_reshaped > 0).to(grad_reshaped.dtype)
                     mask.div_(mask.mean().clamp_(min=1e-3))
@@ -272,9 +260,9 @@ class AdamW_adv(torch.optim.Optimizer):
 
                 mt_slow.mul_(beta3_ema).add_(grad_reshaped, 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(grad_reshaped, mt_slow, alpha=alpha_t)
+                    update = torch.add(grad_reshaped, mt_slow, alpha=alpha)
             else:
                 update = mt.clone() if beta1 > 0 else grad_reshaped.clone()
             del grad_reshaped
@@ -310,7 +298,7 @@ class AdamW_adv(torch.optim.Optimizer):
                 exp_avg = state['exp_avg']
                 exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1)
                 if self.grams_moment:
-                    exp_avg = grad.sign() * exp_avg.abs()
+                    exp_avg = grad.sign().mul_(exp_avg.abs())
                 elif self.cautious_mask:
                     mask = (exp_avg * grad > 0).to(grad.dtype)
                     mask.div_(mask.mean().clamp_(min=1e-3))
@@ -321,9 +309,9 @@ class AdamW_adv(torch.optim.Optimizer):
                 exp_avg_slow = state['exp_avg_slow']
                 exp_avg_slow.mul_(beta3_ema).add_(grad, alpha=1 - beta3_ema)
                 if beta1 > 0:
-                    update = torch.add(exp_avg, exp_avg_slow, alpha=alpha_t)
+                    update = torch.add(exp_avg, exp_avg_slow, alpha=alpha)
                 else:
-                    update = torch.add(grad, exp_avg_slow, alpha=alpha_t)
+                    update = torch.add(grad, exp_avg_slow, alpha=alpha)
             else:
                 update = exp_avg.clone() if beta1 > 0 else grad.clone()
 
@@ -343,9 +331,9 @@ class AdamW_adv(torch.optim.Optimizer):
         # Decoupled weight decay
         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)
@@ -353,7 +341,38 @@ class AdamW_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 'exp_avg_sq' not in self.state[p] and 'mu_v_nmf' not in self.state[p]:
+#            return
+
+        if self.global_step is None and 'step' in self.state[p]:
+            # For backward compatibility
+            self.global_step = self.state[p]['step']
+
+        if group['use_bias_correction']:
+            current_step = self.global_step + 1
+            beta1, beta2 = group['betas']
+            bias_correction1 = 1.0 - beta1 ** current_step
+            bias_correction2 = 1.0 - beta2 ** current_step
+        else:
+            bias_correction1 = 1.0
+            bias_correction2 = 1.0
+
+        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'], bias_correction1, bias_correction2)
+        else:
+            lr_tensor = torch.tensor(group['lr'], device=p.device)
+            bias_correction1_tensor = torch.tensor(bias_correction1, device=p.device)
+            bias_correction2_tensor = torch.tensor(bias_correction2, device=p.device)
+            self._compiled_step_parameter(p, group, lr_tensor, bias_correction1_tensor, bias_correction2_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):
@@ -367,4 +386,6 @@ class AdamW_adv(torch.optim.Optimizer):
             for i, p in enumerate(group['params']):
                 self.step_parameter(p, group, i)
 
+        self.global_step += 1
+
         return loss