heavyball 1.7.0__tar.gz → 1.7.2__tar.gz

{heavyball-1.7.0 → heavyball-1.7.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: heavyball
-Version: 1.7.0
+Version: 1.7.2
 Summary: Efficient Optimizers
 Author-email: HeavyBall Authors <github.heavyball@nestler.sh>
 Project-URL: source, https://github.com/HomebrewML/HeavyBall

{heavyball-1.7.0 → heavyball-1.7.2}/heavyball/__init__.py

@@ -1,4 +1,5 @@
 import functools
+import math
 from typing import Optional
 
 from . import chainable as C
@@ -564,6 +565,10 @@ class ForeachCachedNewtonPSGD(ForeachCachedPSGDKron):
     hessian_approx = True
 
 
+class NewtonHybrid2PSGDKron(ForeachCachedNewtonPSGD):
+    hvp_interval = 2
+
+
 class ForeachPSGDLRA(C.BaseOpt):
     """
     Originally from Evan Walters and Omead Pooladzandi, 2024
@@ -582,7 +587,7 @@ class ForeachPSGDLRA(C.BaseOpt):
         weight_decay=0.0,
         preconditioner_update_probability=None,
         momentum_into_precond_update=True,
-        rank: int = 4,
+        rank: Optional[int] = None,
         warmup_steps: int = 0,
         foreach: bool = True,
         q_dtype="float32",
@@ -608,6 +613,14 @@ class ForeachPSGDLRA(C.BaseOpt):
         )
         params = defaults.pop("params")
 
+        if rank is None:
+            utils.warn_once(
+                f"{rank=}. It will be set to log2(param_count). This requires `params` to be of type list. Currently, {type(params)=}"
+            )
+            params = list(params)
+            defaults["rank"] = round(math.log2(sum(p.numel() for p in params)))
+            utils.warn_once(f"rank was set to {defaults['rank']}")
+
         delayed = C.default(delayed, self.delayed)
         exp_avg_input = C.default(exp_avg_input, self.exp_avg_input)
         update_clipping = C.default(update_clipping, utils.trust_region_clip_)
@@ -632,6 +645,10 @@ class ForeachNewtonPSGDLRA(ForeachPSGDLRA):
     hessian_approx = True
 
 
+class NewtonHybrid2PSGDLRA(ForeachNewtonPSGDLRA):
+    hvp_interval = 2
+
+
 PalmForEachSoap = PaLMForeachSOAP
 PaLMSOAP = PaLMForeachSOAP
 PaLMSFAdamW = PaLMForeachSFAdamW
@@ -696,4 +713,6 @@ __all__ = [
     "DelayedPSGD",
     "PSGDLRA",
     "NewtonPSGDLRA",
+    "NewtonHybrid2PSGDLRA",
+    "NewtonHybrid2PSGDKron",
 ]

{heavyball-1.7.0 → heavyball-1.7.2}/heavyball/chainable.py

@@ -1,4 +1,5 @@
 import functools
+import math
 import random
 from typing import List, Literal, Optional, Union
 
@@ -43,7 +44,7 @@ class FunctionTransform:
         raise NotImplementedError
 
     def get_fn(self):
-        if hasattr(self.fn, "get_fn"):
+        if utils.hasattr_none(self.fn, "get_fn"):
             return self.fn.get_fn()
         return self.fn
 
@@ -426,7 +427,7 @@ def _store_std(state, group, update, grad, param):
     state["init_std"] = torch.std(grad, dim=0)
 
 
-@general_guard("init_std", init_fn=_store_std)
+@general_guard("init_std", init_fn=_store_std, skip_first=False)
 @no_state
 def mup_approx(group, updates, grads, params, init_std):
     _updates = [(u, i) for u, i in zip(updates, init_std) if u.ndim > 1]
@@ -435,6 +436,40 @@ def mup_approx(group, updates, grads, params, init_std):
     return updates
 
 
+def _init_delta(state, group, update, grad, param, log_space: bool):
+    val = group["initial_d"]
+    state["delta"] = torch.full((), math.log(val) if log_space else val, dtype=param.dtype, device=param.device)
+
+
+def _init_full_delta(state, group, update, grad, param, log_space: bool):
+    val = group["initial_d"]
+    state["delta"] = torch.full_like(param, math.log(val) if log_space else val)
+
+
+@zero_guard("state")
+@general_guard("delta", init_fn=functools.partial(_init_delta, log_space=False), skip_first=False)
+@no_state
+def scale_by_d_adaptation(group, update, grad, param, state, delta):
+    utils.d_adaptation(grad, update, state, delta)
+    return update
+
+
+@zero_guard("state")
+@general_guard("delta", init_fn=functools.partial(_init_delta, log_space=True), skip_first=False)
+@no_state
+def scale_by_lr_adaptation(group, update, grad, param, state, delta):
+    utils.lr_adaptation(grad, update, state, delta, group["lr_lr"])
+    return update
+
+
+@zero_guard("state")
+@general_guard("delta", init_fn=functools.partial(_init_full_delta, log_space=True), skip_first=False)
+@no_state
+def scale_by_pointwise_lr_adaptation(group, update, grad, param, state, delta):
+    utils.pointwise_lr_adaptation(grad, update, state, delta, group["lr_lr"])
+    return update
+
+
 @zero_guard("momentum")
 @no_state
 def heavyball_momentum(group, updates, grads, params, momentum):
@@ -484,18 +519,22 @@ def _update_psgd_precond(cached, Q_cache, group, param, grad, Q_mat, Q, exprs, p
     if not group["is_preconditioning"]:
         return Q_mat
 
+    if utils.hasattr_none(param, "vector"):
+        vector, hessian_vector = param.vector, param.hessian_vector
+        del param.vector
+        del param.hessian_vector
+    else:
+        vector, hessian_vector = utils.dampen_grad(grad)
+
     utils.psgd_update_precond(
         Q_mat,
         exprs,
-        getattr(param, "hessian_vector", grad),
+        hessian_vector,
         group["precond_lr"],
         Q,
         group["store_triu_as_line"],
-        getattr(param, "vector", None),
+        vector,
     )
-    if hasattr(param, "vector"):
-        del param.vector
-        del param.hessian_vector
 
     if grad.dim() > 1 and precond_schedule(group, balance_probability, f"balance_prob_{id(Q)}"):
         if group["store_triu_as_line"]:
@@ -566,9 +605,12 @@ def _update_lra(
     if not group["is_preconditioning"]:
         return utils.flatten(U, 1), utils.flatten(V, 1), utils.flatten(d)
 
-    if hasattr(params[0], "hessian_vector") and params[0].hessian_vector is not None:
+    if utils.hasattr_none(params[0], "hessian_vector"):
         vector = utils.flatten([p.vector for p in params])
         hessian_vector = utils.flatten([p.hessian_vector for p in params])
+        for p in params:
+            del p.vector
+            del p.hessian_vector
     else:
         vector, hessian_vector = utils.dampen_multiple(grads)
     return utils.update_lra_precond_(U, V, d, vector, hessian_vector, group["eps"], group["precond_lr"], delayed)