Fix (seeding) Randomness utility

gigl/common/utils/compute/random.py

@@ -1,49 +1,69 @@
+"""
+Matches the ``set_seed(seed, deterministic=False)`` shape used by
+Hugging Face Transformers, MMEngine, and Accelerate; follows the recipe
+at https://pytorch.org/docs/stable/notes/randomness.html.
+"""
+
 import os
 import random
+from typing import Final
+
+import numpy as np
+import torch
 
 from gigl.common.logger import Logger
 
 logger = Logger()
 
+_DEFAULT_SEED: Final[int] = 42  # Answer to the Ultimate Question.
+# Required on CUDA >= 10.2 when use_deterministic_algorithms(True) is set,
+# otherwise cuBLAS matmuls raise RuntimeError. ":4096:8" trades ~24 MiB of
+# extra cuBLAS workspace for keeping perf reasonable vs ":16:8".
+_CUBLAS_WORKSPACE_CONFIG: Final[str] = ":4096:8"
 
-def make_compute_deterministic_and_set_seed(
-    seed: int = 42,  # Answer to the Ultimate Question of Life, The Universe, and Everything
-    should_consider_numpy=True,
-    should_consider_torch=False,
-    should_consider_tensorflow=False,
-):
-    logger.info(
-        """
-        Ensure data loading is also deterministic and you are using deterministic algorithms
-        for relevant frameworks, otherwise nondeterminism will persist
-        """
-    )
 
-    # Setting PYTHONHASHSEED doesn't seem like it actually does anything
-    # See: https://stackoverflow.com/questions/30585108/disable-hash-randomization-from-within-python-program
-    # os.environ["PYTHONHASHSEED"] = "0"
+def seed_everything(
+    seed: int = _DEFAULT_SEED,
+    should_enable_expensive_deterministic_compute: bool = False,
+) -> None:
+    """Seed Python / NumPy / PyTorch RNGs, optionally enforce deterministic torch ops.
 
-    random.seed(seed)
+    What gets seeded:
+
+    - ``random.seed(seed)`` — Python stdlib.
+    - ``np.random.seed(seed)`` — NumPy global RNG.
+    - ``torch.manual_seed(seed)`` — CPU **and all CUDA devices**
+      (``torch.manual_seed`` calls ``torch.cuda.manual_seed_all`` internally.
+      Also covers PyTorch Geometric.
 
-    if should_consider_numpy:
-        import numpy as np
+    When ``should_enable_expensive_deterministic_compute=True`` (opt-in; default False because it costs
+    throughput and should not be enabled for training or for production inference - can be used for debugging purposes.
 
-        np.random.seed(seed)
+    - Important: Graph Sampling currently do not follow determism outlined here.
+    Example:
+        >>> seed_everything(42)
+        42
 
-    if should_consider_torch:
-        import torch
-        import torch.backends.cudnn
+    Args:
+        seed: RNG seed.
+        deterministic: If True, also enforces bitwise-deterministic torch
+            ops (cudnn flags, ``use_deterministic_algorithms``,
+            ``CUBLAS_WORKSPACE_CONFIG``). Default False — most training
+            pipelines want seeded RNGs without paying the throughput cost.
 
-        torch.manual_seed(seed)
+    """
+    random.seed(seed)
+    np.random.seed(seed)
+    torch.manual_seed(seed)
+    if should_enable_expensive_deterministic_compute:
+        os.environ["CUBLAS_WORKSPACE_CONFIG"] = _CUBLAS_WORKSPACE_CONFIG
         torch.backends.cudnn.benchmark = False
         torch.backends.cudnn.deterministic = True
         torch.use_deterministic_algorithms(True)
-
-    if should_consider_tensorflow:
-        import tensorflow as tf
-
-        tf.random.set_seed(seed)
-        os.environ["TF_DETERMINISTIC_OPS"] = "1"
-        os.environ["TF_CUDNN_DETERMINISTIC"] = "1"
-        tf.config.threading.set_inter_op_parallelism_threads(1)
-        tf.config.threading.set_intra_op_parallelism_threads(1)
+        logger.warning(
+            f"seed_everything: seeded python/numpy/torch with seed={seed}; "
+            f"expensive deterministic algorithms ON; "
+            f"throughput will degrade"
+        )
+    else:
+        logger.info(f"seed_everything: seeded python/numpy/torch with seed={seed}")