adv-optm 0.1.6__py3-none-any.whl → 0.1.7__py3-none-any.whl

adv_optm/__init__.py

@@ -2,6 +2,7 @@ from .optim import (
     AdamW_adv,
     Prodigy_adv,
     Adopt_adv,
+    Simplified_AdEMAMix,
     Lion_adv,
     Lion_Prodigy_adv,
 )
@@ -10,8 +11,9 @@ __all__ = [
     "AdamW_adv",
     "Prodigy_adv",
     "Adopt_adv",
+    "Simplified_AdEMAMix",
     "Lion_adv",
     "Lion_Prodigy_adv",
 ]
 
-__version__ = "0.1.6"
+__version__ = "0.1.7"

adv_optm/optim/Prodigy_adv.py

@@ -52,7 +52,17 @@ class Prodigy_adv(torch.optim.Optimizer):
             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 th
+            the scheduler is disabled.
+        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
+            automatically disable `use_AdEMAMix`, `use_cautious`, `use_grams`,
+            and `use_atan2`. (default: False)
+        alpha_grad (float): Mixing coefficient for the Simplified AdEMAMix update rule
+            (only used when `Simplified_AdEMAMix` is `True`). Controls the weight of the
+            current gradient. For small batch sizes, use high values (e.g., 10-100) to be
+            more responsive. For large batch sizes, use low values (e.g., 0-1) for
+            stability. (default: 100.0)
         factored (bool): whether to use the factorization or disable it to use
             the uncompressed optimizer. (default: True)
         d0 (float):
@@ -91,6 +101,8 @@ class Prodigy_adv(torch.optim.Optimizer):
         beta3_ema: float = 0.9999,
         alpha: float = 5.0,
         t_alpha: int | None = None,
+        Simplified_AdEMAMix: bool = False,
+        alpha_grad: float = 100.0,
         factored: bool = True,
         # prodigy parameters
         beta3: float = None,
@@ -109,6 +121,17 @@ class Prodigy_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 betas[0] == 0.0 and Simplified_AdEMAMix:
+            raise ValueError(f"Beta 1 cannot be 0.0 when using Simplified_AdEMAMix. Got {betas[0]}")
+        if use_AdEMAMix and Simplified_AdEMAMix:
+            print("Warning: use_AdEMAMix is incompatible with Simplified_AdEMAMix, Disabling use_AdEMAMix.")
+        if use_grams and Simplified_AdEMAMix:
+            print("Warning: use_grams is incompatible with Simplified_AdEMAMix, Disabling use_grams.")
+        if use_cautious and Simplified_AdEMAMix:
+            print("Warning: use_cautious is incompatible with Simplified_AdEMAMix, Disabling use_cautious.")
+        if use_atan2 and Simplified_AdEMAMix:
+            print("Warning: use_atan2 is incompatible with Simplified_AdEMAMix. Disabling use_atan2.")
+            use_atan2 = False
 
         defaults = {
             "lr": lr, "betas": betas, "eps": eps, "weight_decay": weight_decay,
@@ -118,11 +141,13 @@ class Prodigy_adv(torch.optim.Optimizer):
             "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,
             "fsdp_in_use": fsdp_in_use,
+            "alpha_grad": alpha_grad, 
         }
         self.stochastic_rounding = stochastic_rounding
-        self.use_cautious = use_cautious
-        self.use_grams = use_grams
-        self.use_AdEMAMix = use_AdEMAMix
+        self.use_cautious = use_cautious and not Simplified_AdEMAMix
+        self.use_grams = use_grams and not Simplified_AdEMAMix
+        self.use_AdEMAMix = use_AdEMAMix and not Simplified_AdEMAMix
+        self.Simplified_AdEMAMix = Simplified_AdEMAMix
         self.factored = factored
         self.fsdp_in_use = fsdp_in_use
         super().__init__(params, defaults)
@@ -229,6 +254,8 @@ class Prodigy_adv(torch.optim.Optimizer):
             alpha_t = alpha
             if t_alpha is not None and t_alpha > 0 and current_step < t_alpha:
                 alpha_t = min(current_step * alpha / t_alpha, alpha)
+        if self.Simplified_AdEMAMix:
+            alpha_grad = group["alpha_grad"]
 
         if state['factored']:
             d1, d2 = state['effective_shape']
@@ -243,7 +270,10 @@ class Prodigy_adv(torch.optim.Optimizer):
                     torch.where(unpacked_sign, mt, -mt, out=mt)
                     del unpacked_sign
                 # Update momentum in full-size
-                mt.mul_(self.beta1).add_(grad_reshaped, alpha=self.d * (1.0 - self.beta1))
+                if self.Simplified_AdEMAMix:
+                    mt.mul_(self.beta1).add_(grad_reshaped, alpha=self.d)
+                else:
+                    mt.mul_(self.beta1).add_(grad_reshaped, alpha=self.d * (1.0 - self.beta1))
                 if self.use_grams:
                     mt.copy_(grad_reshaped.sign() * mt.abs())
                 elif self.use_cautious:
@@ -264,6 +294,8 @@ class Prodigy_adv(torch.optim.Optimizer):
                 del unpacked_sign_slow
                 mt_slow.mul_(beta3_ema).add_(grad_reshaped, alpha=self.d * (1.0 - beta3_ema))
                 update = mt + (alpha_t * mt_slow) if self.beta1 > 0 else grad_reshaped + (alpha_t * mt_slow)
+            elif self.Simplified_AdEMAMix:
+                update = torch.add(mt, grad_reshaped, alpha=alpha_grad * self.d)
             else:
                 update = mt.clone() if self.beta1 > 0 else grad_reshaped.clone()
             del grad_reshaped
@@ -297,7 +329,10 @@ class Prodigy_adv(torch.optim.Optimizer):
 
             if self.beta1 > 0:
                 exp_avg = state['exp_avg']
-                exp_avg.mul_(self.beta1).add_(grad, alpha=self.d * (1.0 - self.beta1))
+                if self.Simplified_AdEMAMix:
+                    exp_avg.mul_(self.beta1).add_(grad, alpha=self.d)
+                else:
+                    exp_avg.mul_(self.beta1).add_(grad, alpha=self.d * (1.0 - self.beta1))
                 if self.use_grams:
                     exp_avg = grad.sign() * exp_avg.abs()
                 elif self.use_cautious:
@@ -310,6 +345,8 @@ class Prodigy_adv(torch.optim.Optimizer):
                 exp_avg_slow = state['exp_avg_slow']
                 exp_avg_slow.mul_(beta3_ema).add_(grad, alpha=self.d * (1.0 - beta3_ema))
                 update = exp_avg + (alpha_t * exp_avg_slow) if self.beta1 > 0 else grad + (alpha_t * exp_avg_slow)
+            elif self.Simplified_AdEMAMix:
+                update = torch.add(exp_avg, grad, alpha=alpha_grad * self.d)
             else:
                 update = exp_avg.clone() if self.beta1 > 0 else grad.clone()
 

adv_optm/optim/Simplified_AdEMAMix.py

@@ -0,0 +1,246 @@
+import torch
+
+import math
+
+from ..util.BF16_Stochastic_Rounding import add_stochastic_
+from ..util.Effective_Shape import _get_effective_shape
+from ..util.NNMF import _nnmf,_unnmf
+from ..util.OrthoGrad import _orthogonalize_gradient
+from ..util.One_Bit_Boolean import _pack_bools, _unpack_bools
+
+# A little helper from the original simplified_AdEMAMix
+def linear_hl_warmup_scheduler(step, beta_end, beta_start=0, warmup=1):
+
+    def f(beta, eps=1e-8):
+        return math.log(0.5)/math.log(beta+eps)-1
+
+    def f_inv(t):
+        return math.pow(0.5, 1/(t+1))
+
+    if step < warmup:
+        a = step / float(warmup)
+        return f_inv((1.0-a) * f(beta_start) + a * f(beta_end))
+    return beta_end
+
+class Simplified_AdEMAMix(torch.optim.Optimizer):
+    """
+    Implements the Simplified AdEMAMix algorithm.
+    Refactored from:
+    https://github.com/DepenM/Simplified-AdEMAMix/blob/main/simplified_AdEMAMix.py
+
+    Args:
+        params (iterable): iterable of parameters to optimize or dicts defining
+            parameter groups
+        lr (float): learning rate (default: 1e-5)
+        betas (tuple[float, float]): coefficients used for computing running
+            averages of gradient and its square (default: (0.99, 0.999))
+        eps (float): term added to the denominator to improve
+            numerical stability (default: 1e-8)
+        weight_decay (float): weight decay (L2 penalty) (default: 0).
+        alpha_grad (float): Coeficient for mixing the current gradient and EMA. for small batch
+            sizes set it to high values, up to 100. And for large batch sized set it to small
+            value, down to 0. (default: 100)
+        beta1_warmup (int, optional): number of warmup steps used to increase beta1 (default: None)
+        min_beta1 (float, optional): minimum value of beta1 to start from (default 0.9)
+        vector_reshape (bool): whether to reshape 1D vectors into 2D
+            matrices to apply low-rank compression (default: True).
+        stochastic_rounding (bool): whether to use stochastic
+            rounding for BF16 parameter updates (default: True).
+        use_orthograd (bool): whether to use OrthoGrad. (default: False)
+        factored (bool): whether to use the factorization or disable it to use
+            the uncompressed optimizer. (default: False)
+    """
+
+    def __init__(
+        self,
+        params,
+        lr: float = 1e-5,
+        betas: tuple[float, float] = (0.99, 0.999),
+        eps: float = 1e-8,
+        weight_decay: float = 0.0,
+        alpha_grad: float = 100.0,
+        beta1_warmup: int | None = None,
+        min_beta1: float | None = 0.9,
+        use_bias_correction: bool = True,
+        vector_reshape: bool = True,
+        stochastic_rounding: bool = True,
+        use_orthograd: bool = False,
+        factored: bool = False,
+    ):
+        if not (lr >= 0.0):
+            raise ValueError(f"Learning-rate should be >= 0.0. Got {lr}")
+        if not (0.0 <= betas[0] < 1.0 and 0.0 <= betas[1] < 1.0):
+            raise ValueError(f"Betas should be in [0.0, 1.0). Got {betas}")
+        if not (eps >= 0.0):
+            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 not 0.0 <= alpha_grad:
+            raise ValueError("Invalid alpha value: {}".format(alpha_grad))
+
+        defaults = {
+            "lr": lr, "betas": betas, "eps": eps, "weight_decay": weight_decay,
+            "alpha_grad": alpha_grad, "beta1_warmup": beta1_warmup, "min_beta1": min_beta1,
+            "vector_reshape": vector_reshape,
+            "use_orthograd": use_orthograd, "use_bias_correction": use_bias_correction,
+        }
+        self.stochastic_rounding = stochastic_rounding
+        self.factored = factored
+        super().__init__(params, defaults)
+
+    @property
+    def supports_fused_back_pass(self):
+        return True
+
+    @property
+    def supports_memory_efficient_fp16(self):
+        return True
+
+    @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
+
+        grad = p.grad
+        if grad.dtype != torch.float32 and self.factored:
+            grad = grad.float()
+        if group["use_orthograd"]:
+            grad = _orthogonalize_gradient(p, grad)
+        state = self.state[p]
+
+        # State Initialization
+        if len(state) == 0:
+            state['step'] = 0
+
+            should_factor = (
+                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
+
+            if state['factored']:
+                state['effective_shape'] = _get_effective_shape(p.numel())
+                d1, d2 = state['effective_shape']
+
+                # First moment (m)
+                state['mu_m_nmf'] = torch.zeros(d1, device=device, dtype=dtype) 
+                state['mv_m_nmf'] = torch.zeros(d2, device=device, dtype=dtype)
+                packed_d2 = (d2 + 7) // 8
+                state['sign'] = torch.zeros((d1, packed_d2), dtype=torch.uint8, device=device)
+                # Second moment (v)
+                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 optimizer for non-factored tensors
+                state['exp_avg'] = torch.zeros_like(p, device=device, dtype=dtype)
+                state['exp_avg_sq'] = torch.zeros_like(p, device=device, dtype=dtype)
+            
+            if group['use_bias_correction']:
+                state['num_sum'] = 0.0
+                state['den_sum'] = 0.0
+            else:
+                state['num_sum'] = 1.0
+                state['den_sum'] = 1.0
+
+        beta1_final, beta2 = group["betas"]
+        beta1_warmup = group["beta1_warmup"]
+        alpha_grad = group["alpha_grad"]
+
+        if beta1_warmup is not None:
+            step = state['step'] + 1
+            beta1 = linear_hl_warmup_scheduler(step, beta_end=beta1_final, beta_start=group['min_beta1'], warmup=beta1_warmup)
+        else:
+            beta1 = beta1_final
+
+        if group['use_bias_correction']:
+            state['num_sum'] = beta1 * state['num_sum'] + 1.0
+            state['den_sum'] = beta2 * state['den_sum'] + (1.0 - beta2)
+
+        if state['factored']:
+            d1, d2 = state['effective_shape']
+
+            # Reconstruct momentum from previous step's factors
+            mt = _unnmf((state['mu_m_nmf'], state['mv_m_nmf']))
+            unpacked_sign = _unpack_bools(state['sign'], original_m=d2)
+            torch.where(unpacked_sign, mt, -mt, out=mt)
+            del unpacked_sign
+            # Update momentum in full-size
+            grad_reshaped = grad.view(d1, d2)
+            mt.mul_(beta1).add_(grad_reshaped, alpha=1.0)
+
+            vt = _unnmf((state['mu_v_nmf'], state['mv_v_nmf']))
+            vt.mul_(beta2).addcmul_(grad_reshaped, grad_reshaped, value=1.0 - beta2)
+
+            update = torch.add(mt, grad_reshaped, alpha=alpha_grad)
+            del grad_reshaped
+
+            denom = vt.sqrt().add_(group['eps'] * math.sqrt(state['den_sum']))
+            update.div_(denom)
+            del denom
+
+            if group['use_bias_correction']:
+                update = (update / state['num_sum']) * math.sqrt(state['den_sum'])
+
+            update = update.view(p.shape).mul_(group['lr'])
+
+            # Compress updated moments and store new factors
+            state['sign'] = _pack_bools(mt > 0)
+            _nnmf(mt.abs(), out=(state['mu_m_nmf'], state['mv_m_nmf']))
+            del mt
+            _nnmf(vt, out=(state['mu_v_nmf'], state['mv_v_nmf']))
+            del vt
+
+        else:  # Standard optimizer logic for non-factored tensors
+            exp_avg_sq = state['exp_avg_sq']
+
+            exp_avg = state['exp_avg']
+            exp_avg.mul_(beta1).add_(grad, alpha=1.0)
+
+            update = torch.add(exp_avg, grad, alpha=alpha_grad)
+
+            exp_avg_sq.mul_(beta2).addcmul_(grad, grad, value=1 - beta2)
+
+            denom = exp_avg_sq.sqrt().add_(group['eps'] * math.sqrt(state['den_sum']))
+            update.div_(denom)
+            del denom
+
+            if group['use_bias_correction']:
+                update = (update / state['num_sum']) * math.sqrt(state['den_sum'])
+
+            update.mul_(group['lr'])
+
+        # 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"])
+            else:
+                p.data.add_(p.data, alpha=-group["weight_decay"] * group["lr"])
+
+        if p.dtype == torch.bfloat16 and self.stochastic_rounding:
+            add_stochastic_(p.data, -update)
+        else:
+            p.data.add_(-update)
+        del update
+
+        state['step'] += 1
+
+    @torch.no_grad()
+    def step(self, closure=None):
+        """Performs a single optimization step."""
+        loss = None
+        if closure is not None:
+            with torch.enable_grad():
+                loss = closure()
+
+        for group in self.param_groups:
+            for i, p in enumerate(group['params']):
+                self.step_parameter(p, group, i)
+
+        return loss

adv_optm/optim/__init__.py

@@ -1,6 +1,7 @@
 from .AdamW_adv import AdamW_adv
 from .Prodigy_adv import Prodigy_adv
 from .Adopt_adv import Adopt_adv
+from .Simplified_AdEMAMix import Simplified_AdEMAMix
 from .Lion_adv import Lion_adv
 from .Lion_Prodigy_adv import Lion_Prodigy_adv
 
@@ -8,6 +9,7 @@ __all__ = [
     "AdamW_adv",
     "Prodigy_adv",
     "Adopt_adv",
+    "Simplified_AdEMAMix",
     "Lion_adv",
     "Lion_Prodigy_adv",
 ]

{adv_optm-0.1.6.dist-info → adv_optm-0.1.7.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: adv_optm
-Version: 0.1.6
+Version: 0.1.7
 Summary: A family of highly efficient, lightweight yet powerful optimizers.
 Home-page: https://github.com/Koratahiu/Advanced_Optimizers
 Author: Koratahiu

{adv_optm-0.1.6.dist-info → adv_optm-0.1.7.dist-info}/RECORD

@@ -1,18 +1,19 @@
-adv_optm/__init__.py,sha256=EdoSvk5lJeXMyo8GvdB8HcV7dhgclGDoxyGSY6tn7nc,252
+adv_optm/__init__.py,sha256=CZ_tjWWk5d5D8q_R0rcr8vvwlZyY_44zyAcIAmN_SDY,306
 adv_optm/optim/AdamW_adv.py,sha256=ZeNzk2tWbyd2QDI5hp4InwG3iuHHfqLrlhr_VmcQfRM,13884
 adv_optm/optim/Adopt_adv.py,sha256=rzBWfFOPrMuC6vwETsw7QPKmVXcv4IJRDCTj-6eU1Qk,14798
 adv_optm/optim/Lion_Prodigy_adv.py,sha256=JMss9X8lRpIU4E34PfFpWMMal_XNvZ8Yuqc6i7R5wIQ,14588
 adv_optm/optim/Lion_adv.py,sha256=BA4bSEhJiQ7BhGLDRn9nuMlBrLVh-OMscbmSTeGgRmI,10137
-adv_optm/optim/Prodigy_adv.py,sha256=mGV-B_9fvG3ycuG_-ohyECCU3fEki74fH-ZBHji_lO8,17738
-adv_optm/optim/__init__.py,sha256=e5UighM92LDvDB2JJwj8gDsTpXEedpytScwqS6F2FR8,300
+adv_optm/optim/Prodigy_adv.py,sha256=gJL2r32R3xGD62jMR55ZyKxRv0yL70XHxj4FzEJbFc4,20196
+adv_optm/optim/Simplified_AdEMAMix.py,sha256=opIZjnGJ03-DDAIHTZyJBMReVfgusGDb8FZSWMU3-UM,9774
+adv_optm/optim/__init__.py,sha256=pcP865H2j1tut2VfTUhzQh7V8TF_tzPjqFnjMfFed2k,382
 adv_optm/util/BF16_Stochastic_Rounding.py,sha256=Q5H0BcogmE4atP65dLoI21HKSf50lRdsBDfeF6v9Tbg,1548
 adv_optm/util/Effective_Shape.py,sha256=TBvIk1V8IuTbbBsxuekJA4e_v8JlR5Nujtut8RTWAm4,318
 adv_optm/util/NNMF.py,sha256=yRf5IP5Sjq0Uf0DxN0Q8NxEGSdD-f1ULziLVDOjY8K4,639
 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=qoyIF0jcLjs_vSEcsv36clw5LFNBEbifyXrrVxMH-G4,349
-adv_optm-0.1.6.dist-info/licenses/LICENSE,sha256=HrhfyXIkWY2tGFK11kg7vPCqhgh5DcxleloqdhrpyMY,11558
-adv_optm-0.1.6.dist-info/METADATA,sha256=g0mWOWkYvoHtceMAWPdPAUw1vrgI5v9dnIIxLnD6aD0,5846
-adv_optm-0.1.6.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
-adv_optm-0.1.6.dist-info/top_level.txt,sha256=iNfBIIzu-lPrQ7jyC56WBCcbkRwitM2nJ15-MRQ_6fg,9
-adv_optm-0.1.6.dist-info/RECORD,,
+adv_optm-0.1.7.dist-info/licenses/LICENSE,sha256=HrhfyXIkWY2tGFK11kg7vPCqhgh5DcxleloqdhrpyMY,11558
+adv_optm-0.1.7.dist-info/METADATA,sha256=BEKyVG9zVdb9WThOw9YtgWZ_zqDmErumpY5Fr-AkbX0,5846
+adv_optm-0.1.7.dist-info/WHEEL,sha256=_zCd3N1l69ArxyTb8rzEoP9TpbYXkqRFSNOD5OuxnTs,91
+adv_optm-0.1.7.dist-info/top_level.txt,sha256=iNfBIIzu-lPrQ7jyC56WBCcbkRwitM2nJ15-MRQ_6fg,9
+adv_optm-0.1.7.dist-info/RECORD,,