nshtrainer 1.0.0b42__py3-none-any.whl → 1.0.0b43__py3-none-any.whl

nshtrainer/_checkpoint/metadata.py

@@ -3,7 +3,6 @@ from __future__ import annotations
 import copy
 import datetime
 import logging
-from collections.abc import Callable
 from pathlib import Path
 from typing import TYPE_CHECKING, Any, ClassVar
 
@@ -115,7 +114,7 @@ def _metadata_path(checkpoint_path: Path):
     return checkpoint_path.with_suffix(CheckpointMetadata.PATH_SUFFIX)
 
 
-def _write_checkpoint_metadata(trainer: Trainer, checkpoint_path: Path):
+def write_checkpoint_metadata(trainer: Trainer, checkpoint_path: Path):
     metadata_path = _metadata_path(checkpoint_path)
     metadata = _generate_checkpoint_metadata(trainer, checkpoint_path, metadata_path)
 

nshtrainer/_checkpoint/saver.py

@@ -25,7 +25,10 @@ def link_checkpoint(
 
     if remove_existing:
         try:
-            if linkpath.exists():
+            if linkpath.exists(follow_symlinks=False):
+                # follow_symlinks=False is EXTREMELY important here
+                # Otherwise, we've already deleted the file that the symlink
+                # used to point to, so this always returns False
                 if linkpath.is_dir():
                     shutil.rmtree(linkpath)
                 else:

nshtrainer/trainer/trainer.py

@@ -18,7 +18,7 @@ from lightning.pytorch.trainer.states import TrainerFn
 from lightning.pytorch.utilities.types import _EVALUATE_OUTPUT, _PREDICT_OUTPUT
 from typing_extensions import Never, Unpack, assert_never, deprecated, override
 
-from .._checkpoint.metadata import _write_checkpoint_metadata
+from .._checkpoint.metadata import write_checkpoint_metadata
 from ..callbacks.base import resolve_all_callbacks
 from ..util._environment_info import EnvironmentConfig
 from ..util.bf16 import is_bf16_supported_no_emulation
@@ -478,7 +478,7 @@ class Trainer(LightningTrainer):
         metadata_path = None
         if self.hparams.save_checkpoint_metadata and self.is_global_zero:
             # Generate the metadata and write to disk
-            metadata_path = _write_checkpoint_metadata(self, filepath)
+            metadata_path = write_checkpoint_metadata(self, filepath)
 
         # Call the `on_checkpoint_saved` method on all callbacks
         from .. import _callback

nshtrainer/util/path.py

@@ -85,11 +85,16 @@ def try_symlink_or_copy(
     target_is_directory: bool = False,
     relative: bool = True,
     remove_existing: bool = True,
+    throw_on_invalid_target: bool = False,
 ):
     """
     Symlinks on Unix, copies on Windows.
     """
 
+    # Check if the target file exists
+    if throw_on_invalid_target and not file_path.exists():
+        raise FileNotFoundError(f"File not found: {file_path}")
+
     # If the link already exists, remove it
     if remove_existing:
         try:

{nshtrainer-1.0.0b42.dist-info → nshtrainer-1.0.0b43.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: nshtrainer
-Version: 1.0.0b42
+Version: 1.0.0b43
 Summary: 
 Author: Nima Shoghi
 Author-email: nimashoghi@gmail.com

{nshtrainer-1.0.0b42.dist-info → nshtrainer-1.0.0b43.dist-info}/RECORD

@@ -1,8 +1,8 @@
 nshtrainer/.nshconfig.generated.json,sha256=yZd6cn1RhvNNJUgiUTRYut8ofZYvbulnpPG-rZIRhi4,106
 nshtrainer/__init__.py,sha256=g_moPnfQxSxFZX5NB9ILQQOJrt4RTRuiFt9N0STIpxM,874
 nshtrainer/_callback.py,sha256=tXQCDzS6CvMTuTY5lQSH5qZs1pXUi-gt9bQdpXMVdEs,12715
-nshtrainer/_checkpoint/metadata.py,sha256=Kp1BK1iHCsOzjjy_B0Rbs0yIyd_yV9gZl4uym7rNQ_E,4796
-nshtrainer/_checkpoint/saver.py,sha256=qG8dEetyqwLJS0MkdWDz4DCAInaSyafW8-Oe8x0UVnE,1353
+nshtrainer/_checkpoint/metadata.py,sha256=LQZ8g50rKxQQx-FqiW3n8EWmal9qSWRouOpIIn6NJJY,4758
+nshtrainer/_checkpoint/saver.py,sha256=rWl4d2lCTMU4_wt8yZFL2pFQaP9hj5sPgqHMPQ4zuyI,1584
 nshtrainer/_directory.py,sha256=TJR9ccyuzRlAVfVjGyeQ3E2AFAcz-XbBCxWfiXo2SlY,3191
 nshtrainer/_experimental/__init__.py,sha256=U4S_2y3zgLZVfMenHRaJFBW8yqh2mUBuI291LGQVOJ8,35
 nshtrainer/_hf_hub.py,sha256=4OsCbIITnZk_YLyoMrVyZ0SIN04FBxlC0ig2Et8UAdo,14287
@@ -122,7 +122,6 @@ nshtrainer/nn/mlp.py,sha256=_a8rJJniSCvM08gyQGO-5MUoO18U9_FSGGn3tZL2_U4,7101
 nshtrainer/nn/module_dict.py,sha256=9plb8aQUx5TUEPhX5jI9u8LrpTeKe7jZAHi8iIqcN8w,2365
 nshtrainer/nn/module_list.py,sha256=UB43pcwD_3nUke_DyLQt-iXKhWdKM6Zjm84lRC1hPYA,1755
 nshtrainer/nn/nonlinearity.py,sha256=xmaL4QCRvCxqmaGIOwetJeKK-6IK4m2OV7D3SjxSwJQ,6322
-nshtrainer/nn/tests/test_mlp.py,sha256=COw88Oc4FMvj_mGdQf6F2UtgZ39FE2lbQjTmMAwtCWE,4031
 nshtrainer/optimizer.py,sha256=u968GRNPUNn3f_9BEY2RBNuJq5O3wJWams3NG0dkrOA,1738
 nshtrainer/profiler/__init__.py,sha256=RjaNBoVcTFu8lF0dNlFp-2LaPYdonoIbDy2_KhgF0Ek,594
 nshtrainer/profiler/_base.py,sha256=kFcSVn9gJuMwgDxbfyHh46CmEAIPZjxw3yjPbKgzvwA,950
@@ -141,18 +140,18 @@ nshtrainer/trainer/plugin/layer_sync.py,sha256=h-ydZwXepnsw5-paLgiDatqPyQ_8C0QEv
 nshtrainer/trainer/plugin/precision.py,sha256=I0QsB1bVxmsFmBOkgrAfGONsuYae_lD9Bz0PfJEQvH4,5598
 nshtrainer/trainer/signal_connector.py,sha256=GhfGcSzfaTNhnj2QFkBDq5aT7FqbLMA7eC8SYQs8_8w,10828
 nshtrainer/trainer/strategy.py,sha256=VPTn5z3zvXTydY8IJchjhjcOfpvtoejnvUkq5E4WTus,1368
-nshtrainer/trainer/trainer.py,sha256=l2kJs27v4IHZnzxExr0zX0sVex0wukgiD2Wn_0wiGJg,20836
+nshtrainer/trainer/trainer.py,sha256=ed_Pn-yQCb9BqaHXo2wVhkt2CSfGNEzMAM6RsDoTo-I,20834
 nshtrainer/util/_environment_info.py,sha256=MT8mBe6ZolRfKiwU-les1P-lPNPqXpHQcfADrh_A3uY,24629
 nshtrainer/util/bf16.py,sha256=9QhHZCkYSfYpIcxwAMoXyuh2yTSHBzT-EdLQB297jEs,762
 nshtrainer/util/config/__init__.py,sha256=Z39JJufSb61Lhn2GfVcv3eFW_eorOrN9-9llDWlnZZM,272
 nshtrainer/util/config/dtype.py,sha256=Fn_MhhQoHPyFAnFPSwvcvLiGR3yWFIszMba02CJiC4g,2213
 nshtrainer/util/config/duration.py,sha256=mM-UfU_HvhXwW33TYEDg0x58n80tnle2e6VaWtxZTjk,764
 nshtrainer/util/environment.py,sha256=s-B5nY0cKYXdFMdNYumvC_xxacMATiI4DvV2gUDu20k,4195
-nshtrainer/util/path.py,sha256=IL8MS29MzP2aBdQhqMl2Eh4RvyuZbzaIbAVZzsdLDSk,3754
+nshtrainer/util/path.py,sha256=9fIjE3S78pPL6wjAgEJUYfIJQAPdKOQqIYvTS9lWTUk,3959
 nshtrainer/util/seed.py,sha256=diMV8iwBKN7Xxt5pELmui-gyqyT80_CZzomrWhNss0k,316
 nshtrainer/util/slurm.py,sha256=HflkP5iI_r4UHMyPjw9R4dD5AHsJUpcfJw5PLvGYBRM,1603
 nshtrainer/util/typed.py,sha256=Xt5fUU6zwLKSTLUdenovnKK0N8qUq89Kddz2_XeykVQ,164
 nshtrainer/util/typing_utils.py,sha256=MjY-CUX9R5Tzat-BlFnQjwl1PQ_W2yZQoXhkYHlJ_VA,442
-nshtrainer-1.0.0b42.dist-info/METADATA,sha256=m8vNeWow5AfufFEd6uCjm6w-0NOCTTzCFjXDed0eq6U,988
-nshtrainer-1.0.0b42.dist-info/WHEEL,sha256=XbeZDeTWKc1w7CSIyre5aMDU_-PohRwTQceYnisIYYY,88
-nshtrainer-1.0.0b42.dist-info/RECORD,,
+nshtrainer-1.0.0b43.dist-info/METADATA,sha256=ZE3l6CN34ptFgx3SDPfKIgjdV2s3J8qdP729eb58vzo,988
+nshtrainer-1.0.0b43.dist-info/WHEEL,sha256=XbeZDeTWKc1w7CSIyre5aMDU_-PohRwTQceYnisIYYY,88
+nshtrainer-1.0.0b43.dist-info/RECORD,,

nshtrainer/nn/tests/test_mlp.py

@@ -1,120 +0,0 @@
-from __future__ import annotations
-
-from typing import cast
-
-import pytest
-import torch
-
-from nshtrainer.nn.mlp import MLP, custom_seed_context
-
-
-def test_mlp_seed_reproducibility():
-    """Test that the seed parameter in MLP ensures reproducible weights."""
-
-    # Test dimensions
-    dims = [10, 20, 5]
-
-    # Create two MLPs with the same seed
-    seed1 = 42
-    mlp1 = MLP(dims, activation=torch.nn.ReLU(), seed=seed1)
-    mlp2 = MLP(dims, activation=torch.nn.ReLU(), seed=seed1)
-
-    # Create an MLP with a different seed
-    seed2 = 123
-    mlp3 = MLP(dims, activation=torch.nn.ReLU(), seed=seed2)
-
-    # Check first layer weights
-    layer1_weights1 = cast(torch.Tensor, mlp1[0].weight)
-    layer1_weights2 = cast(torch.Tensor, mlp2[0].weight)
-    layer1_weights3 = cast(torch.Tensor, mlp3[0].weight)
-
-    # Same seed should produce identical weights
-    assert torch.allclose(layer1_weights1, layer1_weights2)
-
-    # Different seeds should produce different weights
-    assert not torch.allclose(layer1_weights1, layer1_weights3)
-
-    # Check second layer weights
-    layer2_weights1 = cast(torch.Tensor, mlp1[2].weight)
-    layer2_weights2 = cast(torch.Tensor, mlp2[2].weight)
-    layer2_weights3 = cast(torch.Tensor, mlp3[2].weight)
-
-    # Same seed should produce identical weights for all layers
-    assert torch.allclose(layer2_weights1, layer2_weights2)
-
-    # Different seeds should produce different weights for all layers
-    assert not torch.allclose(layer2_weights1, layer2_weights3)
-
-    # Test that not providing a seed gives different results each time
-    mlp4 = MLP(dims, activation=torch.nn.ReLU(), seed=None)
-    mlp5 = MLP(dims, activation=torch.nn.ReLU(), seed=None)
-
-    # Without seeds, weights should be different
-    assert not torch.allclose(
-        cast(torch.Tensor, mlp4[0].weight), cast(torch.Tensor, mlp5[0].weight)
-    )
-
-
-def test_custom_seed_context():
-    """Test that custom_seed_context properly controls random number generation."""
-
-    # Test that the same seed produces the same random numbers
-    with custom_seed_context(42):
-        tensor1 = torch.randn(10)
-
-    with custom_seed_context(42):
-        tensor2 = torch.randn(10)
-
-    # Same seed should produce identical random tensors
-    assert torch.allclose(tensor1, tensor2)
-
-    # Test that different seeds produce different random numbers
-    with custom_seed_context(123):
-        tensor3 = torch.randn(10)
-
-    # Different seeds should produce different random tensors
-    assert not torch.allclose(tensor1, tensor3)
-
-
-def test_custom_seed_context_preserves_state():
-    """Test that custom_seed_context preserves the original random state."""
-
-    # Set a known seed for the test
-    original_seed = 789
-    torch.manual_seed(original_seed)
-
-    # Generate a tensor with the original seed
-    original_tensor = torch.randn(10)
-
-    # Use a different seed in the context
-    with custom_seed_context(42):
-        # This should use the temporary seed
-        context_tensor = torch.randn(10)
-
-    # After exiting the context, we should be back to the original seed state
-    # Reset the generator to get the same sequence again
-    torch.manual_seed(original_seed)
-    expected_tensor = torch.randn(10)
-
-    # The tensor generated after the context should match what we would get
-    # if we had just set the original seed again
-    assert torch.allclose(original_tensor, expected_tensor)
-
-    # And it should be different from the tensor generated inside the context
-    assert not torch.allclose(original_tensor, context_tensor)
-
-
-def test_custom_seed_context_with_none():
-    """Test that custom_seed_context with None seed doesn't affect randomization."""
-
-    # Set a known seed
-    torch.manual_seed(555)
-    expected_tensor = torch.randn(10)
-
-    # Reset and use None seed in context
-    torch.manual_seed(555)
-    with custom_seed_context(None):
-        actual_tensor = torch.randn(10)
-
-    # With None seed, the context should not affect the random state
-    assert torch.allclose(expected_tensor, actual_tensor)