mirror of
https://github.com/zebrajr/pytorch.git
synced 2025-12-07 12:21:27 +01:00
f32593630b
On attention_is_all_you_need_pytorch: Perf: 1.526x -> 1.544x Memory: 1.00 -> 1.05x Fix for https://github.com/pytorch/pytorch/issues/102319, although I'm not sure all the perf is recovered. Pull Request resolved: https://github.com/pytorch/pytorch/pull/103330 Approved by: https://github.com/jansel
43 lines
1.2 KiB
Python
43 lines
1.2 KiB
Python
# Copyright (c) Facebook, Inc. and its affiliates.
|
|
# All rights reserved.
|
|
#
|
|
# This source code is licensed under the BSD-style license found in the
|
|
# LICENSE file in the root directory of this source tree.
|
|
|
|
"""
|
|
Global flags for aot autograd
|
|
"""
|
|
import os
|
|
import sys
|
|
|
|
# Converts torch rng ops to their functional philox rng equivalents. Note that
|
|
# we functionalize only CUDA rng ops today.
|
|
functionalize_rng_ops = False
|
|
|
|
# can be useful for debugging if we are incorrectly creating meta fake tensors
|
|
fake_tensor_allow_meta = os.environ.get("FAKE_ALLOW_META", True)
|
|
|
|
# Enables optional asserts in hotpath code to check for errors. If
|
|
# you are seeing weird accuracy problems, try turning this on.
|
|
# This is currently off by default as it will harm tracing time,
|
|
# but it is on by default for aot_eager.
|
|
debug_assert = False
|
|
|
|
debug_partitioner = os.environ.get("AOT_PARTITIONER_DEBUG", False)
|
|
|
|
partitioner_aggressive_fusion = False
|
|
|
|
static_weight_shapes = True
|
|
|
|
# Applies CSE to the graph before partitioning
|
|
cse = True
|
|
|
|
# Restricts the amount of computation AOTAutograd can do.
|
|
max_dist_from_bw = 3
|
|
|
|
|
|
from .._dynamo.config_utils import install_config_module
|
|
|
|
# adds patch, save_config, invalid config checks, etc
|
|
install_config_module(sys.modules[__name__])
|