libinephany 0.15.0__tar.gz → 0.15.2__tar.gz

libinephany-0.15.2/CODE_VERSION.cfg

@@ -0,0 +1 @@
+0.15.2

{libinephany-0.15.0/libinephany.egg-info → libinephany-0.15.2}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: libinephany
-Version: 0.15.0
+Version: 0.15.2
 Summary: Inephany library containing code commonly used by multiple subpackages.
 Author-email: Inephany <info@inephany.com>
 License: Apache 2.0

{libinephany-0.15.0 → libinephany-0.15.2}/libinephany/observations/observation_utils.py

@@ -14,7 +14,7 @@ import torch
 import torch.optim as optim
 
 from libinephany.pydantic_models.schemas.tensor_statistics import TensorStatistics
-from libinephany.utils import optim_utils
+from libinephany.utils import optim_utils, torch_distributed_utils
 
 # ======================================================================================================================
 #
@@ -173,7 +173,14 @@ def tensor_on_local_rank(tensor: torch.Tensor | None) -> bool:
     :return: Whether the tensor is owned by the local rank.
     """
 
-    return tensor is not None and tensor.grad is not None and tensor.numel() > 0
+    valid_tensor = tensor is not None and tensor.grad is not None and tensor.numel() > 0
+
+    if valid_tensor and tensor.is_cuda:
+        local_rank = torch_distributed_utils.get_local_rank()
+
+        return tensor.device.index == local_rank
+
+    return valid_tensor
 
 
 def form_update_tensor(

{libinephany-0.15.0 → libinephany-0.15.2}/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-0.15.0 → libinephany-0.15.2}/libinephany/observations/statistic_trackers.py

@@ -193,13 +193,14 @@ class Statistic(ABC):
         Processes the tensor cache to build a TensorStatistic model.
         """
 
-        concatenated = torch.cat(self._tensor_cache)
-        self._tensor_cache = []
+        if self._tensor_cache:
+            concatenated = torch.cat(self._tensor_cache)
+            self._tensor_cache = []
 
-        statistics = TensorStatistics.build(
-            tensor=concatenated, skip_statistics=self.skip_statistics, sample_percentage=self.downsample_percent
-        )
-        self._data.append(statistics)  # type: ignore
+            statistics = TensorStatistics.build(
+                tensor=concatenated, skip_statistics=self.skip_statistics, sample_percentage=self.downsample_percent
+            )
+            self._data.append(statistics)  # type: ignore
 
     @staticmethod
     @final
@@ -213,10 +214,10 @@ 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 +240,21 @@ 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():
@@ -288,23 +287,21 @@ class Statistic(ABC):
             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():
-                return
+            statistic = self._distributed_reduce(statistic=statistic)
 
-            if isinstance(statistic, torch.Tensor):
-                statistic = statistic.view(-1)
-                self._tensor_cache.append(statistic)
+            if torch_distributed_utils.is_scheduler_master_rank():
+                if isinstance(statistic, torch.Tensor):
+                    statistic = statistic.view(-1)
+                    self._tensor_cache.append(statistic)
 
-                if len(self._tensor_cache) >= self.max_cache_size:
-                    self._process_tensor_cache()
+                    if len(self._tensor_cache) >= self.max_cache_size:
+                        self._process_tensor_cache()
 
-            elif statistic is not None:
-                self._data.append(statistic)  # type: ignore
+                elif statistic is not None:
+                    self._data.append(statistic)  # type: ignore
 
-        if torch_distributed_utils.is_scheduler_master_rank():
-            self._sample_number += 1
+        self._sample_number += 1
 
     @final
     def fetch(self) -> TensorStatistics | float | None:

{libinephany-0.15.0 → libinephany-0.15.2}/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-0.15.0 → libinephany-0.15.2/libinephany.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: libinephany
-Version: 0.15.0
+Version: 0.15.2
 Summary: Inephany library containing code commonly used by multiple subpackages.
 Author-email: Inephany <info@inephany.com>
 License: Apache 2.0

libinephany-0.15.0/CODE_VERSION.cfg

@@ -1 +0,0 @@
-0.15.0