wehub-resource-sync
184 lines
6.7 KiB
Python
184 lines
6.7 KiB
Python
# Copyright (c) 2025 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.
|
|
|
|
import time
|
|
import unittest
|
|
|
|
import numpy as np
|
|
|
|
import paddle
|
|
|
|
# Disable static mode so that .numpy() can be called.
|
|
paddle.disable_static()
|
|
|
|
from paddle.incubate.tensor.manipulation import (
|
|
async_offload,
|
|
async_offload_with_offset,
|
|
async_reload,
|
|
create_xpu_async_load,
|
|
)
|
|
|
|
|
|
def print_debug_info(tensor, name):
|
|
"""Prints debug information for a tensor."""
|
|
# print(f"{name} is on device: {tensor.place}")
|
|
# print(f"{name} shape: {tensor.shape}, dtype: {tensor.dtype}")
|
|
try:
|
|
arr = tensor.numpy()
|
|
flat = arr.flatten()
|
|
# print(f"{name} first 5 elements: {flat[:5]}")
|
|
# Uncomment the next line if you need to print the full array.
|
|
# print(f"{name} full array:\n{arr}")
|
|
except Exception as e:
|
|
# print(f"{name} cannot be converted to numpy array: {e}")
|
|
raise
|
|
|
|
|
|
class TestSaveLoadLargeParameters(unittest.TestCase):
|
|
def offload_and_reload(self, data0):
|
|
print_debug_info(data0, "data0 (original)")
|
|
|
|
# Create a fixed compute tensor for matmul.
|
|
data1 = paddle.arange(0, 100, dtype="float32").reshape([10, 10])
|
|
print_debug_info(data1, "data1 (for compute)")
|
|
|
|
loader = create_xpu_async_load()
|
|
|
|
# Offload data0 -> pinned memory.
|
|
cpu_data, task = async_offload(data0, loader)
|
|
print_debug_info(cpu_data, "cpu_data (after offload)")
|
|
|
|
# Do a compute on XPU.
|
|
res = paddle.matmul(data1, data1)
|
|
print_debug_info(res, "res (after first compute)")
|
|
|
|
# Wait for the offload task to complete (CPU side).
|
|
task.cpu_wait()
|
|
|
|
# Reload from pinned memory back to XPU.
|
|
xpu_data, task = async_reload(cpu_data, loader)
|
|
print_debug_info(xpu_data, "xpu_data (after reload)")
|
|
|
|
# Do another compute on XPU.
|
|
res = paddle.matmul(data1, data1)
|
|
print_debug_info(res, "res (after second compute)")
|
|
|
|
# Wait on both the device (XPU) and CPU sides.
|
|
task.xpu_wait()
|
|
task.cpu_wait()
|
|
|
|
# Extract numpy arrays and print max differences.
|
|
a = data0.numpy()
|
|
b = cpu_data.numpy()
|
|
c = xpu_data.numpy()
|
|
# print("Max diff (data0 - cpu_data):", np.max(np.abs(a - b)))
|
|
# print("Max diff (data0 - xpu_data):", np.max(np.abs(a - c)))
|
|
np.testing.assert_array_equal(a, b)
|
|
np.testing.assert_array_equal(a, c)
|
|
|
|
def test_large_parameters_paddle_save_tensor(self):
|
|
# Create a fixed tensor with known values using linspace.
|
|
arr = np.linspace(0, 1, 50).reshape([10, 5]).astype("float32")
|
|
data0 = paddle.to_tensor(arr, place=paddle.XPUPlace(0))
|
|
print_debug_info(
|
|
data0, "data0 in test_large_parameters_paddle_save_tensor"
|
|
)
|
|
self.offload_and_reload(data0)
|
|
|
|
def test_large_parameters_paddle_save_model_weight(self):
|
|
model = paddle.nn.Linear(10, 5)
|
|
data0 = model.weight
|
|
print_debug_info(
|
|
data0,
|
|
"model.weight in test_large_parameters_paddle_save_model_weight",
|
|
)
|
|
self.offload_and_reload(data0)
|
|
|
|
def test_offload_with_offset(self):
|
|
loader = create_xpu_async_load()
|
|
# Create a fixed source tensor with all elements equal to 3.14.
|
|
src_arr = np.full([100], 3.14, dtype="float32")
|
|
data1 = paddle.to_tensor(src_arr, place=paddle.XPUPlace(0))
|
|
print_debug_info(data1, "data1 in test_offload_with_offset")
|
|
# Create a destination tensor on CPU (pinned memory) initialized to zeros.
|
|
dst_arr = np.zeros([100], dtype="float32")
|
|
data2 = paddle.to_tensor(dst_arr, place=paddle.XPUPinnedPlace())
|
|
print_debug_info(data2, "data2 in test_offload_with_offset (CPU)")
|
|
|
|
# Offload in two segments.
|
|
task1 = async_offload_with_offset(
|
|
src_tensor=data1,
|
|
dst_tensor=data2,
|
|
src_offset=0,
|
|
dst_offset=0,
|
|
offload_size=50,
|
|
async_loader=loader,
|
|
)
|
|
task2 = async_offload_with_offset(
|
|
src_tensor=data1,
|
|
dst_tensor=data2,
|
|
src_offset=50,
|
|
dst_offset=50,
|
|
offload_size=50,
|
|
async_loader=loader,
|
|
)
|
|
|
|
# Wait for both tasks.
|
|
task1.xpu_wait()
|
|
task2.cpu_wait()
|
|
|
|
print_debug_info(data1, "data1 after offload_with_offset")
|
|
print_debug_info(data2, "data2 after offload_with_offset")
|
|
diff = np.max(np.abs(data1.numpy() - data2.numpy()))
|
|
# print("Max diff (data1 - data2):", diff)
|
|
np.testing.assert_array_equal(data1.numpy(), data2.numpy())
|
|
|
|
def test_large_data_performance(self):
|
|
# Create a large tensor (~100MB) on XPU.
|
|
# For float32 (4 bytes), 100MB = 104857600 bytes => 104857600 / 4 = 26214400 elements.
|
|
# We use shape [512, 512, 100] since 512*512*100 = 26214400.
|
|
# print("Starting large data performance test...")
|
|
large_arr = np.random.rand(512, 512, 1000).astype("float32")
|
|
large_tensor = paddle.to_tensor(large_arr, place=paddle.XPUPlace(0))
|
|
print_debug_info(large_tensor, "large_tensor (original)")
|
|
loader = create_xpu_async_load()
|
|
|
|
# Measure offload time.
|
|
t0 = time.time()
|
|
cpu_large, task_offload = async_offload(large_tensor, loader)
|
|
task_offload.cpu_wait() # Wait for offload completion.
|
|
t1 = time.time()
|
|
offload_time = t1 - t0
|
|
# print(f"Offload time for 100MB tensor: {offload_time:.4f} seconds")
|
|
|
|
# Measure reload time.
|
|
t2 = time.time()
|
|
xpu_large, task_reload = async_reload(cpu_large, loader)
|
|
task_reload.cpu_wait() # Wait for reload completion.
|
|
t3 = time.time()
|
|
reload_time = t3 - t2
|
|
# print(f"Reload time for 100MB tensor: {reload_time:.4f} seconds")
|
|
|
|
# Verify that the reloaded tensor matches the original.
|
|
a = large_tensor.numpy()
|
|
c = xpu_large.numpy()
|
|
max_diff = np.max(np.abs(a - c))
|
|
# print("Max diff (large_tensor - xpu_large):", max_diff)
|
|
np.testing.assert_array_equal(a, c)
|
|
|
|
|
|
if __name__ == '__main__':
|
|
# print("Default Paddle device:", paddle.get_device())
|
|
unittest.main()
|