wehub-resource-sync
335 lines
14 KiB
Python
335 lines
14 KiB
Python
# Copyright (c) 2024 PaddlePaddle Authors. All Rights Reserved.
|
|
#
|
|
# Licensed under the Apache License, Version 2.0 (the "License");
|
|
# you may not use this file except in compliance with the License.
|
|
# You may obtain a copy of the License at
|
|
#
|
|
# http://www.apache.org/licenses/LICENSE-2.0
|
|
#
|
|
# Unless required by applicable law or agreed to in writing, software
|
|
# distributed under the License is distributed on an "AS IS" BASIS,
|
|
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
# See the License for the specific language governing permissions and
|
|
# limitations under the License.
|
|
|
|
# [AUTO-GENERATED]
|
|
# Target file: python/paddle/optimizer/fusion_utils.py
|
|
# Coverage target: get_current_device_type, get_align, FusionStorage class
|
|
# 未覆盖行: FusionStorageHelper class (requires IPC metadata)
|
|
|
|
import unittest
|
|
|
|
import numpy as np
|
|
|
|
import paddle
|
|
from paddle.optimizer.fusion_utils import (
|
|
FusionStorage,
|
|
get_align,
|
|
get_current_device_type,
|
|
)
|
|
|
|
|
|
class TestGetCurrentDeviceType(unittest.TestCase):
|
|
"""Test get_current_device_type function.
|
|
测试 get_current_device_type 函数。"""
|
|
|
|
def test_returns_string(self):
|
|
"""get_current_device_type should return a string.
|
|
get_current_device_type 应该返回字符串。"""
|
|
result = get_current_device_type()
|
|
self.assertIsInstance(result, str)
|
|
|
|
def test_returns_valid_device(self):
|
|
"""get_current_device_type should return 'gpu', 'xpu', or 'unknown'.
|
|
get_current_device_type 应该返回 'gpu'、'xpu' 或 'unknown'。"""
|
|
result = get_current_device_type()
|
|
self.assertIn(result, ["gpu", "xpu", "unknown"])
|
|
|
|
def test_consistent_result(self):
|
|
"""Multiple calls should return the same result (cached).
|
|
多次调用应返回相同结果（缓存）。"""
|
|
result1 = get_current_device_type()
|
|
result2 = get_current_device_type()
|
|
self.assertEqual(result1, result2)
|
|
|
|
|
|
class TestGetAlign(unittest.TestCase):
|
|
"""Test get_align function.
|
|
测试 get_align 函数。"""
|
|
|
|
def setUp(self):
|
|
paddle.disable_static()
|
|
|
|
def test_get_align_float32(self):
|
|
"""Test alignment for float32 tensor.
|
|
测试 float32 张量的对齐。"""
|
|
t = paddle.zeros([10], dtype=paddle.float32)
|
|
# Should return an integer alignment value
|
|
result = get_align(t)
|
|
self.assertIsInstance(result, (int, np.integer))
|
|
|
|
def test_get_align_float16(self):
|
|
"""Test alignment for float16 tensor.
|
|
测试 float16 张量的对齐。"""
|
|
t = paddle.zeros([10], dtype=paddle.float16)
|
|
result = get_align(t)
|
|
self.assertIsInstance(result, (int, np.integer))
|
|
|
|
def test_get_align_bfloat16(self):
|
|
"""Test alignment for bfloat16 tensor.
|
|
测试 bfloat16 张量的对齐。"""
|
|
t = paddle.zeros([10], dtype=paddle.bfloat16)
|
|
result = get_align(t)
|
|
self.assertIsInstance(result, (int, np.integer))
|
|
|
|
def test_get_align_large_tensor(self):
|
|
"""Test alignment for a large tensor that may already be aligned.
|
|
测试已经对齐的大张量的对齐值。"""
|
|
# 256 bytes / 4 bytes per float32 = 64 elements -> should be aligned
|
|
t = paddle.zeros([64], dtype=paddle.float32)
|
|
result = get_align(t)
|
|
self.assertEqual(result, 0)
|
|
|
|
def test_get_align_unaligned_tensor(self):
|
|
"""Test alignment for a small unaligned tensor.
|
|
测试小的未对齐张量的对齐值。"""
|
|
t = paddle.zeros([3], dtype=paddle.float32)
|
|
result = get_align(t)
|
|
# 3 * 4 = 12 bytes, 256 - 12 = 244 remaining, 244 / 4 = 61
|
|
self.assertGreaterEqual(result, 0)
|
|
|
|
def test_get_align_2d_tensor(self):
|
|
"""Test alignment for 2D tensor.
|
|
测试二维张量的对齐值。"""
|
|
t = paddle.zeros([5, 10], dtype=paddle.float32)
|
|
result = get_align(t)
|
|
self.assertIsInstance(result, (int, np.integer))
|
|
|
|
def test_get_align_result_non_negative(self):
|
|
"""Alignment result should always be non-negative.
|
|
对齐结果应始终为非负。"""
|
|
t = paddle.zeros([7], dtype=paddle.float32)
|
|
result = get_align(t)
|
|
self.assertGreaterEqual(result, 0)
|
|
|
|
|
|
class TestFusionStorage(unittest.TestCase):
|
|
"""Test FusionStorage class.
|
|
测试 FusionStorage 类。"""
|
|
|
|
def setUp(self):
|
|
paddle.disable_static()
|
|
|
|
def test_basic_creation(self):
|
|
"""Test basic FusionStorage creation with accumulators and master_weights.
|
|
测试使用 accumulators 和 master_weights 创建基本 FusionStorage。"""
|
|
acc = paddle.randn([10], dtype=paddle.float32)
|
|
mw = paddle.randn([10], dtype=paddle.float32)
|
|
accumulators = {"momentum": {"weight": acc}}
|
|
master_weights = {"weight": mw}
|
|
storage = FusionStorage(accumulators, master_weights)
|
|
self.assertIsNotNone(storage.buffer)
|
|
self.assertEqual(storage.buffer.dtype, paddle.float32)
|
|
|
|
def test_with_merged_model_params(self):
|
|
"""Test FusionStorage with merged_model_params.
|
|
测试带有 merged_model_params 的 FusionStorage。"""
|
|
acc = paddle.randn([8], dtype=paddle.float32)
|
|
mw = paddle.randn([8], dtype=paddle.float32)
|
|
mp = paddle.randn([8], dtype=paddle.float32)
|
|
accumulators = {"momentum": {"w": acc}}
|
|
master_weights = {"w": mw}
|
|
merged_model_params = {"w": mp}
|
|
storage = FusionStorage(
|
|
accumulators,
|
|
master_weights,
|
|
merged_model_params=merged_model_params,
|
|
)
|
|
self.assertIsNotNone(storage.buffer)
|
|
self.assertIsNotNone(storage.merged_model_params_meta)
|
|
|
|
def test_buffer_shape(self):
|
|
"""Test that buffer size accounts for alignment padding.
|
|
测试 buffer 大小考虑了对齐填充。"""
|
|
acc = paddle.randn([10], dtype=paddle.float32)
|
|
mw = paddle.randn([10], dtype=paddle.float32)
|
|
accumulators = {"momentum": {"w": acc}}
|
|
master_weights = {"w": mw}
|
|
storage = FusionStorage(accumulators, master_weights)
|
|
# Buffer should be at least as large as raw data
|
|
raw_size = 10 + 10
|
|
self.assertGreaterEqual(storage.buffer.shape[0], raw_size)
|
|
|
|
def test_mapping_tensor_preserves_values(self):
|
|
"""Test that mapping_tensor copies values into the buffer correctly.
|
|
测试 mapping_tensor 正确地将值复制到 buffer 中。"""
|
|
acc_val = paddle.ones([5], dtype=paddle.float32) * 3.14
|
|
mw_val = paddle.ones([5], dtype=paddle.float32) * 2.71
|
|
accumulators = {"momentum": {"w": acc_val.clone()}}
|
|
master_weights = {"w": mw_val.clone()}
|
|
storage = FusionStorage(accumulators, master_weights)
|
|
# Check that accumulator values are in buffer
|
|
acc_meta = storage.accumulators_meta["momentum"]["w"]
|
|
buf_slice = storage.buffer._slice(acc_meta["start"], acc_meta["end"])
|
|
# The first 5 elements should be 3.14
|
|
np.testing.assert_array_almost_equal(
|
|
buf_slice.numpy()[:5], np.full(5, 3.14), decimal=5
|
|
)
|
|
# Check that master weight values are in buffer
|
|
mw_meta = storage.master_weights_meta["w"]
|
|
mw_slice = storage.buffer._slice(mw_meta["start"], mw_meta["end"])
|
|
np.testing.assert_array_almost_equal(
|
|
mw_slice.numpy()[:5], np.full(5, 2.71), decimal=5
|
|
)
|
|
|
|
def test_accumulators_meta_structure(self):
|
|
"""Test that accumulators_meta has correct keys.
|
|
测试 accumulators_meta 包含正确的键。"""
|
|
acc = paddle.randn([4], dtype=paddle.float32)
|
|
mw = paddle.randn([4], dtype=paddle.float32)
|
|
accumulators = {"sgd": {"param1": acc}}
|
|
master_weights = {"param1": mw}
|
|
storage = FusionStorage(accumulators, master_weights)
|
|
self.assertIn("sgd", storage.accumulators_meta)
|
|
self.assertIn("param1", storage.accumulators_meta["sgd"])
|
|
meta = storage.accumulators_meta["sgd"]["param1"]
|
|
self.assertIn("start", meta)
|
|
self.assertIn("end", meta)
|
|
self.assertIn("name", meta)
|
|
self.assertIn("shape", meta)
|
|
|
|
def test_master_weights_meta_structure(self):
|
|
"""Test that master_weights_meta has correct keys.
|
|
测试 master_weights_meta 包含正确的键。"""
|
|
acc = paddle.randn([6], dtype=paddle.float32)
|
|
mw = paddle.randn([6], dtype=paddle.float32)
|
|
accumulators = {"momentum": {"p": acc}}
|
|
master_weights = {"p": mw}
|
|
storage = FusionStorage(accumulators, master_weights)
|
|
self.assertIn("p", storage.master_weights_meta)
|
|
meta = storage.master_weights_meta["p"]
|
|
self.assertIn("start", meta)
|
|
self.assertIn("end", meta)
|
|
self.assertIn("name", meta)
|
|
self.assertIn("shape", meta)
|
|
|
|
def test_merged_model_params_meta_structure(self):
|
|
"""Test that merged_model_params_meta is populated correctly.
|
|
测试 merged_model_params_meta 正确填充。"""
|
|
acc = paddle.randn([4], dtype=paddle.float32)
|
|
mw = paddle.randn([4], dtype=paddle.float32)
|
|
mp = paddle.randn([4], dtype=paddle.float32)
|
|
accumulators = {"adam": {"p": acc}}
|
|
master_weights = {"p": mw}
|
|
merged = {"p": mp}
|
|
storage = FusionStorage(
|
|
accumulators, master_weights, merged_model_params=merged
|
|
)
|
|
self.assertIn("p", storage.merged_model_params_meta)
|
|
meta = storage.merged_model_params_meta["p"]
|
|
self.assertIn("start", meta)
|
|
self.assertIn("end", meta)
|
|
self.assertIn("shape", meta)
|
|
|
|
def test_multiple_accumulator_groups(self):
|
|
"""Test with multiple accumulator groups.
|
|
测试多个累加器组的情况。"""
|
|
acc1 = paddle.randn([3], dtype=paddle.float32)
|
|
acc2 = paddle.randn([3], dtype=paddle.float32)
|
|
mw = paddle.randn([3], dtype=paddle.float32)
|
|
accumulators = {
|
|
"momentum": {"w": acc1},
|
|
"variance": {"w": acc2},
|
|
}
|
|
master_weights = {"w": mw}
|
|
storage = FusionStorage(accumulators, master_weights)
|
|
self.assertIn("momentum", storage.accumulators_meta)
|
|
self.assertIn("variance", storage.accumulators_meta)
|
|
|
|
def test_none_merged_model_params(self):
|
|
"""Test FusionStorage with merged_model_params=None (default).
|
|
测试 merged_model_params=None（默认值）的 FusionStorage。"""
|
|
acc = paddle.randn([4], dtype=paddle.float32)
|
|
mw = paddle.randn([4], dtype=paddle.float32)
|
|
accumulators = {"momentum": {"w": acc}}
|
|
master_weights = {"w": mw}
|
|
storage = FusionStorage(
|
|
accumulators, master_weights, merged_model_params=None
|
|
)
|
|
self.assertIsNone(storage.merged_model_params)
|
|
self.assertEqual(storage.merged_model_params_meta, {})
|
|
|
|
def test_assert_accumulators_dict(self):
|
|
"""Test that accumulators must be a dict.
|
|
测试 accumulators 必须是字典。"""
|
|
mw = paddle.randn([4], dtype=paddle.float32)
|
|
with self.assertRaises(AssertionError):
|
|
FusionStorage("not_a_dict", {"w": mw})
|
|
|
|
def test_assert_master_weights_dict(self):
|
|
"""Test that master_weights must be a dict.
|
|
测试 master_weights 必须是字典。"""
|
|
acc = paddle.randn([4], dtype=paddle.float32)
|
|
with self.assertRaises(AssertionError):
|
|
FusionStorage({"w": acc}, "not_a_dict")
|
|
|
|
def test_assert_merged_model_params_type(self):
|
|
"""Test that merged_model_params must be dict or None.
|
|
测试 merged_model_params 必须是字典或 None。"""
|
|
acc = paddle.randn([4], dtype=paddle.float32)
|
|
mw = paddle.randn([4], dtype=paddle.float32)
|
|
with self.assertRaises(AssertionError):
|
|
FusionStorage({"w": acc}, {"w": mw}, merged_model_params="invalid")
|
|
|
|
@unittest.skipIf(
|
|
not paddle.is_compiled_with_cuda(), "GPU required for IPC metadata"
|
|
)
|
|
def test_buffer_ipc_meta_on_gpu(self):
|
|
"""Test buffer_ipc_meta property on GPU.
|
|
测试 GPU 上的 buffer_ipc_meta 属性。"""
|
|
acc = paddle.randn([4], dtype=paddle.float32)
|
|
mw = paddle.randn([4], dtype=paddle.float32)
|
|
accumulators = {"momentum": {"w": acc.cuda()}}
|
|
master_weights = {"w": mw.cuda()}
|
|
storage = FusionStorage(accumulators, master_weights)
|
|
# On CUDA (non-ROCm), buffer_ipc_meta should return IPC metadata
|
|
meta = storage.buffer_ipc_meta
|
|
# meta could be None on ROCm or could be a tuple on CUDA
|
|
if not paddle.is_compiled_with_rocm():
|
|
self.assertIsNotNone(meta)
|
|
|
|
def test_dtype_float16_storage(self):
|
|
"""Test FusionStorage with float16 dtype.
|
|
测试 float16 类型的 FusionStorage。"""
|
|
acc = paddle.randn([8], dtype=paddle.float16)
|
|
mw = paddle.randn([8], dtype=paddle.float16)
|
|
# Note: FusionStorage requires matching dtype, but we test float16
|
|
# which uses align=2 vs float32 align=4
|
|
accumulators = {"momentum": {"w": acc}}
|
|
master_weights = {"w": mw}
|
|
storage = FusionStorage(
|
|
accumulators, master_weights, dtype=paddle.float16
|
|
)
|
|
self.assertEqual(storage.dtype, paddle.float16)
|
|
self.assertEqual(storage.buffer.dtype, paddle.float16)
|
|
|
|
def test_meta_shape_matches(self):
|
|
"""Test that stored shape in metadata matches original tensor shape.
|
|
测试元数据中存储的形状与原始张量形状匹配。"""
|
|
acc = paddle.randn([3, 4], dtype=paddle.float32)
|
|
mw = paddle.randn([3, 4], dtype=paddle.float32)
|
|
accumulators = {"momentum": {"w": acc}}
|
|
master_weights = {"w": mw}
|
|
storage = FusionStorage(accumulators, master_weights)
|
|
self.assertEqual(
|
|
tuple(storage.accumulators_meta["momentum"]["w"]["shape"]),
|
|
(3, 4),
|
|
)
|
|
self.assertEqual(
|
|
tuple(storage.master_weights_meta["w"]["shape"]), (3, 4)
|
|
)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
unittest.main()
|