libinephany 0.14.1__tar.gz → 0.15.1__tar.gz

libinephany-0.15.1/CODE_VERSION.cfg

@@ -0,0 +1 @@
+0.15.1

{libinephany-0.14.1/libinephany.egg-info → libinephany-0.15.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: libinephany
-Version: 0.14.1
+Version: 0.15.1
 Summary: Inephany library containing code commonly used by multiple subpackages.
 Author-email: Inephany <info@inephany.com>
 License: Apache 2.0

{libinephany-0.14.1 → libinephany-0.15.1}/libinephany/observations/observers/global_observers.py

@@ -36,7 +36,7 @@ class InitialHyperparameters(GlobalObserver):
 
         super().__init__(**kwargs)
 
-        force_skip = ["samples"]
+        force_skip = ["samples", "gradient_accumulation"]
         skip_hparams = force_skip if skip_hparams is None else skip_hparams + force_skip
         self.skip_hparams = [] if skip_hparams is None else skip_hparams
         self.pad_with = pad_with

{libinephany-0.14.1 → libinephany-0.15.1}/libinephany/observations/statistic_trackers.py

@@ -213,10 +213,10 @@ class Statistic(ABC):
 
         if torch_distributed_utils.is_scheduler_master_rank():
             if isinstance(statistic, torch.Tensor):
-                shape = statistic.shape
+                shape = statistic.view(-1).shape
 
             elif isinstance(statistic, TensorStatistics):
-                shape = statistic.to_tensor().shape
+                shape = statistic.to_tensor().view(-1).shape
 
             elif statistic is not None:
                 shape = torch.tensor([statistic]).shape
@@ -239,23 +239,21 @@ class Statistic(ABC):
         if not torch_distributed_utils.is_distributed():
             return statistic
 
-        if statistic is None:
-            shape = self._determine_reduction_shape(statistic=statistic)
-
-            if shape is None:
-                return statistic
+        shape = self._determine_reduction_shape(statistic=statistic)
 
-            to_reduce = torch.zeros(shape)
+        if statistic is None:
+            to_reduce = torch.zeros(shape, dtype=torch.float64)
 
         elif isinstance(statistic, torch.Tensor):
-            to_reduce = statistic.clone()
+            to_reduce = statistic.clone().to(torch.float64).view(-1)
 
         elif isinstance(statistic, TensorStatistics):
-            to_reduce = statistic.to_tensor()
+            to_reduce = statistic.to_tensor().to(torch.float64).view(-1)
 
         else:
-            to_reduce = torch.tensor([statistic])
+            to_reduce = torch.tensor([statistic], dtype=torch.float64)
 
+        to_reduce = to_reduce.to(torch_distributed_utils.get_local_device())
         dist.reduce(to_reduce, dst=MASTER_SCHEDULER_RANK, op=ReduceOp.SUM)
 
         if not torch_distributed_utils.is_scheduler_master_rank():
@@ -283,11 +281,13 @@ class Statistic(ABC):
 
         parameter_group = self._find_parameter_group(optimizer=optimizer)
         parameters = self._get_parameters(parameter_group=parameter_group)
+        self._sample_number += 1
 
         if self._sample_number % self.sample_frequency == 0:
             statistic = self._gather(
                 optimizer=optimizer, model=model, parameters=parameters, parameter_group=parameter_group
             )
+
             statistic = self._distributed_reduce(statistic=statistic)
 
             if not torch_distributed_utils.is_scheduler_master_rank():
@@ -303,9 +303,6 @@ class Statistic(ABC):
             elif statistic is not None:
                 self._data.append(statistic)  # type: ignore
 
-        if torch_distributed_utils.is_scheduler_master_rank():
-            self._sample_number += 1
-
     @final
     def fetch(self) -> TensorStatistics | float | None:
         """

{libinephany-0.14.1 → libinephany-0.15.1}/libinephany/pydantic_models/configs/hyperparameter_configs.py

@@ -4,7 +4,7 @@
 #
 # ======================================================================================================================
 
-from typing import Any
+from typing import Any, cast
 
 from pydantic import BaseModel, ConfigDict, ValidationError, field_serializer, field_validator, model_validator
 
@@ -232,6 +232,23 @@ class BatchSizeHParamConfig(HParamConfig):
     sample_discrete_values: list[float | int] | None = None
 
 
+class GradientAccumulationHParamConfig(HParamConfig):
+    max_hparam_value: float | int = 64
+    min_hparam_value: float | int = 1
+    hparam_dtype: type[float | int] = int
+    initial_value: int = 1
+    initial_delta: float = 0.0
+    scale: float = 1.0
+
+    sampler: str = "DiscreteRangeSampler"
+    sample_initial_values: bool = False
+    sample_lower_bound: int = 1
+    sample_upper_bound: int = 64
+    sample_step: int = 1
+    sample_discrete_values: list[float | int] | None = None
+    force_limit: float | int = 64
+
+
 class EpochsHParamConfig(HParamConfig):
     max_hparam_value: float | int = 16
     min_hparam_value: float | int = 1
@@ -289,6 +306,7 @@ class HParamConfigs(BaseModel):
     sgd_momentum_config: HParamConfig = SGDMomentumHParamConfig()
 
     batch_size_config: HParamConfig = BatchSizeHParamConfig()
+    gradient_accumulation_config: GradientAccumulationHParamConfig = GradientAccumulationHParamConfig()
     epochs_config: HParamConfig = EpochsHParamConfig()
     token_config: HParamConfig = TokensHParamConfig()
     samples_config: HParamConfig = SamplesHParamConfig()
@@ -351,6 +369,9 @@ class HParamConfigs(BaseModel):
             case AgentTypes.BatchSize:
                 self.batch_size_config = hparam_config
 
+            case AgentTypes.GradientAccumulationAgent:
+                self.gradient_accumulation_config = cast(GradientAccumulationHParamConfig, hparam_config)
+
             case AgentTypes.Epochs:
                 self.epochs_config = hparam_config
 
@@ -400,6 +421,9 @@ class HParamConfigs(BaseModel):
             case AgentTypes.BatchSize:
                 return self.batch_size_config
 
+            case AgentTypes.GradientAccumulationAgent:
+                return self.gradient_accumulation_config
+
             case AgentTypes.Epochs:
                 return self.epochs_config
 

{libinephany-0.14.1 → libinephany-0.15.1}/libinephany/pydantic_models/schemas/inner_task_profile.py

@@ -140,6 +140,40 @@ class InnerTaskProfiles(BaseModel):
 
         return sum(self.compiled_action_sizes.values())
 
+    @property
+    def max_total_observation_size(self) -> int:
+        """
+        :return: The summed observation size of all agents with the task that has the most layers.
+        """
+
+        if not self.profiles:
+            raise ValueError(
+                "No profiles to calculate max total observation size. Ensure profiles have been "
+                "added before executing the training loop"
+            )
+
+        largest_task_name = max(self.profiles, key=lambda k: self.profiles[k].number_of_layers)
+        largest_task = self.profiles[largest_task_name]
+
+        return sum(largest_task.observation_space_sizes.values())
+
+    @property
+    def max_total_action_size(self) -> int:
+        """
+        :return: The summed action size of all agents with the task that has the most layers.
+        """
+
+        if not self.profiles:
+            raise ValueError(
+                "No profiles to calculate max total action size. Ensure profiles have been "
+                "added before executing the training loop"
+            )
+
+        largest_task_name = max(self.profiles, key=lambda k: self.profiles[k].number_of_layers)
+        largest_task = self.profiles[largest_task_name]
+
+        return sum(largest_task.action_space_sizes.values())
+
     @staticmethod
     def _compile_gym_space_sizes(spaces: dict[str, dict[str, int]]) -> dict[str, int]:
         """

{libinephany-0.14.1 → libinephany-0.15.1}/libinephany/pydantic_models/states/hyperparameter_states.py

@@ -20,6 +20,7 @@ from libinephany.utils.constants import (
     DROPOUT,
     EPOCHS,
     GRAD_NORM_CLIP,
+    GRADIENT_ACCUMULATION,
     LEARNING_RATE,
     SAMPLES,
     SGD_MOMENTUM,
@@ -60,6 +61,7 @@ class UpdateCallbacks(BaseModel):
     sgd_momentum: Callable[..., None]
 
     batch_size: Callable[..., None] | None
+    gradient_accumulation: Callable[..., None] | None
     epochs: Callable[..., None] | None
 
     def __getitem__(self, item: str) -> Callable[..., None] | None:
@@ -457,6 +459,7 @@ class ParameterGroupHParams(HyperparameterContainer):
 class GlobalHParams(HyperparameterContainer):
 
     batch_size: Hyperparameter
+    gradient_accumulation: Hyperparameter
     epochs: Hyperparameter
     tokens: Hyperparameter
     samples: Hyperparameter
@@ -550,6 +553,14 @@ class HyperparameterStates(BaseModel):
         """
         return self.global_hparams.batch_size
 
+    @computed_field  # type: ignore[misc]
+    @property
+    def gradient_accumulation(self) -> Hyperparameter:
+        """
+        :return: The gradient accumulation steps of the inner model.
+        """
+        return self.global_hparams.gradient_accumulation
+
     @computed_field  # type: ignore[misc]
     @property
     def epochs(self) -> Hyperparameter:
@@ -676,6 +687,7 @@ class HyperparameterStates(BaseModel):
 
         return {
             BATCH_SIZE: hparam_configs.batch_size_config,
+            GRADIENT_ACCUMULATION: hparam_configs.gradient_accumulation_config,
             EPOCHS: hparam_configs.epochs_config,
             TOKENS: hparam_configs.token_config,
             SAMPLES: hparam_configs.samples_config,

{libinephany-0.14.1 → libinephany-0.15.1}/libinephany/utils/constants.py

@@ -21,6 +21,7 @@ ADAM_BETA_TWO = "adam_beta_two"
 ADAM_EPS = "adam_eps"
 SGD_MOMENTUM = "sgd_momentum"
 BATCH_SIZE = "batch_size"
+GRADIENT_ACCUMULATION = "gradient_accumulation"
 EPOCHS = "epochs"
 TOKENS = "tokens"
 SAMPLES = "samples"
@@ -41,6 +42,7 @@ AGENT_PREFIX_EPS = "adam-eps"
 AGENT_PREFIX_SGD_MOMENTUM = "sgd-momentum"
 
 AGENT_BATCH_SIZE = "batch-size"
+AGENT_GRADIENT_ACCUMULATION = "gradient-accumulation"
 
 AGENT_BANDIT_SUFFIX = "bandit-agent"
 
@@ -53,6 +55,7 @@ AGENT_TYPES = [
     ADAM_BETA_TWO,
     ADAM_EPS,
     SGD_MOMENTUM,
+    GRADIENT_ACCUMULATION,
 ]
 SUFFIXES = [AGENT_BANDIT_SUFFIX]
 PREFIXES = [
@@ -64,6 +67,7 @@ PREFIXES = [
     AGENT_PREFIX_BETA_TWO,
     AGENT_PREFIX_EPS,
     AGENT_PREFIX_SGD_MOMENTUM,
+    AGENT_GRADIENT_ACCUMULATION,
 ]
 PREFIXES_TO_HPARAMS = {
     AGENT_PREFIX_LR: LEARNING_RATE,
@@ -74,4 +78,5 @@ PREFIXES_TO_HPARAMS = {
     AGENT_PREFIX_BETA_TWO: ADAM_BETA_TWO,
     AGENT_PREFIX_EPS: ADAM_EPS,
     AGENT_PREFIX_SGD_MOMENTUM: SGD_MOMENTUM,
+    AGENT_GRADIENT_ACCUMULATION: GRADIENT_ACCUMULATION,
 }

{libinephany-0.14.1 → libinephany-0.15.1}/libinephany/utils/enums.py

@@ -14,6 +14,7 @@ from libinephany.utils.constants import (
     DROPOUT,
     EPOCHS,
     GRAD_NORM_CLIP,
+    GRADIENT_ACCUMULATION,
     LEARNING_RATE,
     SAMPLES,
     SGD_MOMENTUM,
@@ -69,6 +70,7 @@ class AgentTypes(EnumWithIndices):
     AdamBetaTwoAgent = ADAM_BETA_TWO
     AdamEpsAgent = ADAM_EPS
     SGDMomentumAgent = SGD_MOMENTUM
+    GradientAccumulationAgent = GRADIENT_ACCUMULATION
 
     # Deprecated or Non-Agent
     BatchSize = BATCH_SIZE

{libinephany-0.14.1 → libinephany-0.15.1}/libinephany/utils/torch_distributed_utils.py

@@ -4,8 +4,10 @@
 #
 # ======================================================================================================================
 
+import os
 from typing import Any
 
+import torch
 import torch.distributed as dist
 
 # ======================================================================================================================
@@ -14,7 +16,11 @@ import torch.distributed as dist
 #
 # ======================================================================================================================
 
+CUDA = "cuda"
+CPU = "cpu"
+CUDA_PREFIX = f"{CUDA}:"
 MASTER_SCHEDULER_RANK = 0
+LOCAL_RANK = "LOCAL_RANK"
 
 # ======================================================================================================================
 #
@@ -48,7 +54,10 @@ def get_local_rank() -> int:
     :return: Distributed computing rank of this process.
     """
 
-    return dist.get_rank() if is_distributed() else MASTER_SCHEDULER_RANK
+    if not is_distributed():
+        return MASTER_SCHEDULER_RANK
+
+    return dist.get_rank()
 
 
 def is_scheduler_master_rank() -> bool:
@@ -83,3 +92,15 @@ def barrier() -> None:
 
     if is_distributed():
         dist.barrier()
+
+
+def get_local_device() -> torch.device:
+    """
+    :return: Local device of the current rank.
+    """
+
+    if not is_distributed():
+        return torch.device(CUDA if torch.cuda.is_available() else CPU)
+
+    local_device_rank = os.environ.get(LOCAL_RANK, MASTER_SCHEDULER_RANK)
+    return torch.device(f"{CUDA_PREFIX}{local_device_rank}" if torch.cuda.is_available() else CPU)

{libinephany-0.14.1 → libinephany-0.15.1/libinephany.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: libinephany
-Version: 0.14.1
+Version: 0.15.1
 Summary: Inephany library containing code commonly used by multiple subpackages.
 Author-email: Inephany <info@inephany.com>
 License: Apache 2.0

libinephany-0.14.1/CODE_VERSION.cfg

@@ -1 +0,0 @@
-0.14.1