894 lines
30 KiB
Python
894 lines
30 KiB
Python
# Copyright (c) 2021 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 os
|
|
import pathlib
|
|
import shutil
|
|
import unittest
|
|
|
|
import numpy as np
|
|
from op_test import is_custom_device
|
|
|
|
import paddle
|
|
from paddle.device.cuda.graphs import CUDAGraph
|
|
|
|
|
|
def can_use_cuda_graph():
|
|
return (
|
|
paddle.is_compiled_with_cuda() or is_custom_device()
|
|
) and not paddle.is_compiled_with_rocm()
|
|
|
|
|
|
@unittest.skipIf(
|
|
not (paddle.is_compiled_with_cuda() or is_custom_device())
|
|
or float(paddle.version.cuda()) < 11.0,
|
|
"only support cuda >= 11.0",
|
|
)
|
|
class TestCUDAGraphInDygraphMode(unittest.TestCase):
|
|
def setUp(self):
|
|
if can_use_cuda_graph():
|
|
paddle.set_flags(
|
|
{
|
|
'FLAGS_allocator_strategy': 'auto_growth',
|
|
'FLAGS_sync_nccl_allreduce': False,
|
|
'FLAGS_cudnn_deterministic': True,
|
|
'FLAGS_use_stream_safe_cuda_allocator': False,
|
|
}
|
|
)
|
|
|
|
def random_tensor(self, shape):
|
|
return paddle.to_tensor(
|
|
np.random.randint(low=0, high=10, size=shape).astype("float32")
|
|
)
|
|
|
|
def test_cuda_graph_dynamic_graph(self):
|
|
if not can_use_cuda_graph():
|
|
return
|
|
|
|
shape = [2, 3]
|
|
x = self.random_tensor(shape)
|
|
z = self.random_tensor(shape)
|
|
|
|
g = CUDAGraph()
|
|
g.capture_begin()
|
|
y = x + 10
|
|
b = x.numel()
|
|
z.add_(x)
|
|
g.capture_end()
|
|
|
|
for _ in range(10):
|
|
z_np_init = z.numpy()
|
|
x_new = self.random_tensor(shape)
|
|
x.copy_(x_new, False)
|
|
g.replay()
|
|
x_np = x_new.numpy()
|
|
y_np = y.numpy()
|
|
z_np = z.numpy()
|
|
self.assertTrue((y_np - x_np == 10).all())
|
|
self.assertTrue((z_np - z_np_init == x_np).all())
|
|
|
|
g.reset()
|
|
|
|
def test_concat_and_split(self):
|
|
if not can_use_cuda_graph():
|
|
return
|
|
|
|
concat_num = 100
|
|
xs = []
|
|
xs_np = []
|
|
|
|
for i in range(concat_num):
|
|
x_np = np.random.random(size=[1]).astype(np.float32)
|
|
xs.append(paddle.to_tensor(x_np))
|
|
xs_np.append(x_np)
|
|
|
|
graph = CUDAGraph()
|
|
graph.capture_begin()
|
|
y = paddle.concat(xs)
|
|
zs = paddle.split(y, len(xs))
|
|
graph.capture_end()
|
|
graph.replay()
|
|
|
|
y_np = y.numpy()
|
|
y_np_expected = np.concatenate(xs_np)
|
|
np.testing.assert_array_equal(y_np, y_np_expected)
|
|
self.assertEqual(len(zs), len(xs_np))
|
|
for i, z in enumerate(zs):
|
|
np.testing.assert_array_equal(z.numpy(), xs_np[i])
|
|
|
|
output_dir = f'cuda_graph_dot_{os.getpid()}'
|
|
try:
|
|
graph.print_to_dot_files(pathlib.Path(output_dir))
|
|
graph.reset()
|
|
shutil.rmtree(output_dir)
|
|
except Exception as e:
|
|
msg = str(e)
|
|
sub_msg = "The print_to_dot_files() method is only supported when CUDA version >= 11.3"
|
|
self.assertTrue(sub_msg in msg)
|
|
finally:
|
|
graph.reset()
|
|
|
|
def test_dataloader(self):
|
|
if not can_use_cuda_graph():
|
|
return
|
|
|
|
class AutoIncDataset(paddle.io.Dataset):
|
|
def __init__(self, n, dtype):
|
|
self.n = n
|
|
self.dtype = dtype
|
|
|
|
def __len__(self):
|
|
return self.n
|
|
|
|
def __getitem__(self, idx):
|
|
return np.array([idx]).astype(self.dtype)
|
|
|
|
n = 100
|
|
dtype = 'int64'
|
|
dataset = AutoIncDataset(n, dtype)
|
|
data_loader = paddle.io.DataLoader(
|
|
dataset, batch_size=1, num_workers=2, use_buffer_reader=True
|
|
)
|
|
x = None
|
|
y = None
|
|
|
|
graph = None
|
|
for i, data in enumerate(data_loader):
|
|
if graph is None:
|
|
x = data
|
|
x = x.cuda()
|
|
graph = CUDAGraph()
|
|
graph.capture_begin()
|
|
y = x * x
|
|
graph.capture_end()
|
|
else:
|
|
x.copy_(data, False)
|
|
x = x.cuda()
|
|
|
|
graph.replay()
|
|
actual_x = np.array([[i]]).astype(dtype)
|
|
actual_y = np.array([[i * i]]).astype(dtype)
|
|
np.testing.assert_array_equal(actual_x, x.numpy())
|
|
np.testing.assert_array_equal(actual_y, y.numpy())
|
|
|
|
def test_dev_ctx_alloc(self):
|
|
if not can_use_cuda_graph():
|
|
return
|
|
|
|
x = paddle.to_tensor([2], dtype='float32')
|
|
graph = CUDAGraph()
|
|
graph.capture_begin()
|
|
y = paddle.cast(x, dtype='float16')
|
|
graph.capture_end()
|
|
|
|
def test_cuda_graph_with_enable_replace(self):
|
|
"""Test that CUDAGraph created with enable_replace=True captures and replays correctly."""
|
|
if not can_use_cuda_graph():
|
|
return
|
|
|
|
shape = [4, 4]
|
|
x = self.random_tensor(shape)
|
|
x_val = x.numpy().copy()
|
|
|
|
g = CUDAGraph(enable_replace=True)
|
|
g.capture_begin()
|
|
y = x * 2.0
|
|
g.capture_end()
|
|
|
|
g.replay()
|
|
np.testing.assert_allclose(y.numpy(), x_val * 2.0, rtol=1e-5)
|
|
g.reset()
|
|
|
|
def test_replace_input_ptrs(self):
|
|
"""Test replace_input_ptrs exercises CacheKernelNodeInfos and ReplaceInputPtrs code paths."""
|
|
if not can_use_cuda_graph():
|
|
return
|
|
|
|
shape = [4, 4]
|
|
x = self.random_tensor(shape)
|
|
x_val = x.numpy().copy()
|
|
|
|
g = CUDAGraph(enable_replace=True)
|
|
g.capture_begin()
|
|
y = x * 2.0
|
|
g.capture_end()
|
|
|
|
# First replay: result should match x * 2
|
|
g.replay()
|
|
np.testing.assert_allclose(y.numpy(), x_val * 2.0, rtol=1e-5)
|
|
|
|
# Create a new input buffer and replace the pointer
|
|
x_new = self.random_tensor(shape)
|
|
x_new_val = x_new.numpy().copy()
|
|
old_ptr = x.data_ptr()
|
|
new_ptr = x_new.data_ptr()
|
|
|
|
# Should not raise; exercises ReplaceInputPtrs
|
|
g.replace_input_ptrs([old_ptr], [new_ptr])
|
|
g.replay()
|
|
|
|
# On CUDA >= 12.4 the replacement is effective; on older CUDA
|
|
# GetKernelParamInfos returns empty so replacement is a no-op.
|
|
# Either way we validate no crash and replay succeeds.
|
|
result = y.numpy()
|
|
cuda_ver = (
|
|
float(paddle.version.cuda())
|
|
if paddle.version.cuda() != 'False'
|
|
else 0.0
|
|
)
|
|
if cuda_ver >= 12.4:
|
|
np.testing.assert_allclose(result, x_new_val * 2.0, rtol=1e-5)
|
|
|
|
g.reset()
|
|
|
|
def test_replace_input_ptrs_requires_enable_replace(self):
|
|
"""Test that replace_input_ptrs raises an error when enable_replace=False."""
|
|
if not can_use_cuda_graph():
|
|
return
|
|
|
|
shape = [2, 3]
|
|
x = self.random_tensor(shape)
|
|
|
|
g = CUDAGraph() # enable_replace defaults to False
|
|
g.capture_begin()
|
|
y = x + 1.0
|
|
g.capture_end()
|
|
|
|
raised = False
|
|
try:
|
|
g.replace_input_ptrs([x.data_ptr()], [x.data_ptr()])
|
|
except Exception:
|
|
raised = True
|
|
finally:
|
|
g.reset()
|
|
|
|
self.assertTrue(raised, "Expected exception when enable_replace=False")
|
|
|
|
def test_replace_input_ptrs_empty(self):
|
|
"""Test replace_input_ptrs with empty pointer lists (no-op path)."""
|
|
if not can_use_cuda_graph():
|
|
return
|
|
|
|
shape = [2, 3]
|
|
x = self.random_tensor(shape)
|
|
|
|
g = CUDAGraph(enable_replace=True)
|
|
g.capture_begin()
|
|
y = x + 1.0
|
|
g.capture_end()
|
|
|
|
# Calling with empty lists should be a no-op and not raise
|
|
g.replace_input_ptrs([], [])
|
|
g.replay()
|
|
np.testing.assert_allclose(y.numpy(), x.numpy() + 1.0, rtol=1e-5)
|
|
g.reset()
|
|
|
|
def test_replace_input_ptrs_no_match(self):
|
|
"""Test replace_input_ptrs when old_ptr does not match any kernel param (no modification)."""
|
|
if not can_use_cuda_graph():
|
|
return
|
|
|
|
shape = [2, 3]
|
|
x = self.random_tensor(shape)
|
|
x_val = x.numpy().copy()
|
|
|
|
g = CUDAGraph(enable_replace=True)
|
|
g.capture_begin()
|
|
y = x + 1.0
|
|
g.capture_end()
|
|
|
|
g.replay()
|
|
|
|
# Pass a dummy pointer that won't match anything
|
|
dummy_ptr = 0
|
|
g.replace_input_ptrs([dummy_ptr], [dummy_ptr])
|
|
g.replay()
|
|
|
|
# Output should remain unchanged from original x
|
|
np.testing.assert_allclose(y.numpy(), x_val + 1.0, rtol=1e-5)
|
|
g.reset()
|
|
|
|
def test_replace_multiple_input_ptrs(self):
|
|
"""Test replacing multiple input tensor pointers simultaneously.
|
|
Exercises the j-loop over multiple old_ptrs entries in ReplaceInputPtrs.
|
|
"""
|
|
if not can_use_cuda_graph():
|
|
return
|
|
|
|
shape = [4, 4]
|
|
x = self.random_tensor(shape)
|
|
w = self.random_tensor(shape)
|
|
x_val = x.numpy().copy()
|
|
w_val = w.numpy().copy()
|
|
|
|
g = CUDAGraph(enable_replace=True)
|
|
g.capture_begin()
|
|
y = x + w
|
|
g.capture_end()
|
|
|
|
g.replay()
|
|
np.testing.assert_allclose(y.numpy(), x_val + w_val, rtol=1e-5)
|
|
|
|
x_new = self.random_tensor(shape)
|
|
w_new = self.random_tensor(shape)
|
|
x_new_val = x_new.numpy().copy()
|
|
w_new_val = w_new.numpy().copy()
|
|
|
|
# Replace both input pointers at once
|
|
g.replace_input_ptrs(
|
|
[x.data_ptr(), w.data_ptr()],
|
|
[x_new.data_ptr(), w_new.data_ptr()],
|
|
)
|
|
g.replay()
|
|
|
|
cuda_ver = (
|
|
float(paddle.version.cuda())
|
|
if paddle.version.cuda() != 'False'
|
|
else 0.0
|
|
)
|
|
if cuda_ver >= 12.4:
|
|
np.testing.assert_allclose(
|
|
y.numpy(), x_new_val + w_new_val, rtol=1e-5
|
|
)
|
|
g.reset()
|
|
|
|
def test_replace_input_ptrs_repeated(self):
|
|
"""Test calling replace_input_ptrs multiple times in a row.
|
|
Verifies that each replacement overwrites the previous one.
|
|
"""
|
|
if not can_use_cuda_graph():
|
|
return
|
|
|
|
shape = [4, 4]
|
|
x = self.random_tensor(shape)
|
|
|
|
g = CUDAGraph(enable_replace=True)
|
|
g.capture_begin()
|
|
y = x * 3.0
|
|
g.capture_end()
|
|
|
|
g.replay()
|
|
|
|
x_new1 = self.random_tensor(shape)
|
|
x_new2 = self.random_tensor(shape)
|
|
x_new2_val = x_new2.numpy().copy()
|
|
|
|
# First replacement: x -> x_new1
|
|
g.replace_input_ptrs([x.data_ptr()], [x_new1.data_ptr()])
|
|
# Second replacement: x_new1 -> x_new2 (chain replacement)
|
|
g.replace_input_ptrs([x_new1.data_ptr()], [x_new2.data_ptr()])
|
|
g.replay()
|
|
|
|
cuda_ver = (
|
|
float(paddle.version.cuda())
|
|
if paddle.version.cuda() != 'False'
|
|
else 0.0
|
|
)
|
|
if cuda_ver >= 12.4:
|
|
np.testing.assert_allclose(y.numpy(), x_new2_val * 3.0, rtol=1e-5)
|
|
g.reset()
|
|
|
|
def test_replace_input_ptrs_after_multiple_replays(self):
|
|
"""Test replace_input_ptrs interleaved with multiple replays.
|
|
Ensures ReplaceInputPtrs works correctly on non-first-run graphs.
|
|
"""
|
|
if not can_use_cuda_graph():
|
|
return
|
|
|
|
shape = [4, 4]
|
|
x = self.random_tensor(shape)
|
|
x_val = x.numpy().copy()
|
|
|
|
g = CUDAGraph(enable_replace=True)
|
|
g.capture_begin()
|
|
y = x + 5.0
|
|
g.capture_end()
|
|
|
|
# Multiple replays before replacing
|
|
for _ in range(3):
|
|
g.replay()
|
|
np.testing.assert_allclose(y.numpy(), x_val + 5.0, rtol=1e-5)
|
|
|
|
x_new = self.random_tensor(shape)
|
|
x_new_val = x_new.numpy().copy()
|
|
g.replace_input_ptrs([x.data_ptr()], [x_new.data_ptr()])
|
|
g.replay()
|
|
|
|
cuda_ver = (
|
|
float(paddle.version.cuda())
|
|
if paddle.version.cuda() != 'False'
|
|
else 0.0
|
|
)
|
|
if cuda_ver >= 12.4:
|
|
np.testing.assert_allclose(y.numpy(), x_new_val + 5.0, rtol=1e-5)
|
|
g.reset()
|
|
|
|
def test_replace_input_ptrs_multiple_kernel_nodes(self):
|
|
if not can_use_cuda_graph():
|
|
return
|
|
|
|
shape = [4, 4]
|
|
x = self.random_tensor(shape)
|
|
x_val = x.numpy().copy()
|
|
|
|
g = CUDAGraph(enable_replace=True)
|
|
g.capture_begin()
|
|
y = x * 2.0
|
|
z = y + x
|
|
g.capture_end()
|
|
|
|
g.replay()
|
|
np.testing.assert_allclose(z.numpy(), x_val * 3.0, rtol=1e-5)
|
|
|
|
x_new = self.random_tensor(shape)
|
|
x_new_val = x_new.numpy().copy()
|
|
g.replace_input_ptrs([x.data_ptr()], [x_new.data_ptr()])
|
|
g.replay()
|
|
|
|
cuda_ver = (
|
|
float(paddle.version.cuda())
|
|
if paddle.version.cuda() != 'False'
|
|
else 0.0
|
|
)
|
|
if cuda_ver >= 12.4:
|
|
np.testing.assert_allclose(y.numpy(), x_new_val * 2.0, rtol=1e-5)
|
|
np.testing.assert_allclose(z.numpy(), x_new_val * 3.0, rtol=1e-5)
|
|
g.reset()
|
|
|
|
def test_replace_input_ptrs_with_pool_id(self):
|
|
if not can_use_cuda_graph():
|
|
return
|
|
|
|
shape = [4, 4]
|
|
x = self.random_tensor(shape)
|
|
x_val = x.numpy().copy()
|
|
|
|
g = CUDAGraph(pool_id=0, enable_replace=True)
|
|
g.capture_begin()
|
|
y = x - 1.0
|
|
g.capture_end()
|
|
|
|
g.replay()
|
|
np.testing.assert_allclose(y.numpy(), x_val - 1.0, rtol=1e-5)
|
|
|
|
x_new = self.random_tensor(shape)
|
|
x_new_val = x_new.numpy().copy()
|
|
g.replace_input_ptrs([x.data_ptr()], [x_new.data_ptr()])
|
|
g.replay()
|
|
|
|
cuda_ver = (
|
|
float(paddle.version.cuda())
|
|
if paddle.version.cuda() != 'False'
|
|
else 0.0
|
|
)
|
|
if cuda_ver >= 12.4:
|
|
np.testing.assert_allclose(y.numpy(), x_new_val - 1.0, rtol=1e-5)
|
|
g.reset()
|
|
|
|
|
|
@unittest.skipIf(
|
|
not (paddle.is_compiled_with_cuda() or is_custom_device())
|
|
or float(paddle.version.cuda()) < 11.0,
|
|
"only support cuda >= 11.0",
|
|
)
|
|
class TestPaddleCudaCUDAGraphCompat(unittest.TestCase):
|
|
"""Tests for the PyTorch-compat surface of ``paddle.cuda.CUDAGraph``.
|
|
|
|
``paddle.cuda.CUDAGraph`` is the same class as
|
|
``paddle.device.cuda.graphs.CUDAGraph`` (re-exported via
|
|
``paddle.cuda.graphs``). These tests cover the PyTorch-compat keyword
|
|
arguments and helper methods (``keep_graph``, ``pool``, ``instantiate``,
|
|
``enable_debug_mode``/``debug_dump``, ``raw_cuda_graph[_exec]``) plus an
|
|
end-to-end capture/replay round-trip.
|
|
"""
|
|
|
|
def setUp(self):
|
|
if can_use_cuda_graph():
|
|
paddle.set_flags(
|
|
{
|
|
'FLAGS_allocator_strategy': 'auto_growth',
|
|
'FLAGS_sync_nccl_allreduce': False,
|
|
'FLAGS_cudnn_deterministic': True,
|
|
'FLAGS_use_stream_safe_cuda_allocator': False,
|
|
}
|
|
)
|
|
|
|
def test_paddle_cuda_class_is_consolidated(self):
|
|
# The PyTorch-facing name and the native name must be the same class
|
|
# to avoid duplicated implementations.
|
|
from paddle.device.cuda.graphs import CUDAGraph as DeviceCUDAGraph
|
|
|
|
self.assertIs(paddle.cuda.CUDAGraph, DeviceCUDAGraph)
|
|
self.assertIs(paddle.cuda.graphs.CUDAGraph, DeviceCUDAGraph)
|
|
|
|
def test_constructor_keep_graph(self):
|
|
g = paddle.cuda.CUDAGraph()
|
|
self.assertFalse(g._keep_graph)
|
|
g2 = paddle.cuda.CUDAGraph(keep_graph=True)
|
|
self.assertTrue(g2._keep_graph)
|
|
|
|
def test_constructor_positional_bool_is_keep_graph(self):
|
|
# ``torch.cuda.CUDAGraph(True)`` passes the bool positionally; our
|
|
# signature would otherwise bind it to ``place`` and surface only as
|
|
# an obscure pybind error inside ``capture_begin``. The bool-shim in
|
|
# ``__init__`` re-routes it to ``keep_graph``.
|
|
g_true = paddle.cuda.CUDAGraph(True)
|
|
self.assertTrue(g_true._keep_graph)
|
|
g_false = paddle.cuda.CUDAGraph(False)
|
|
self.assertFalse(g_false._keep_graph)
|
|
|
|
# Conflict: ``CUDAGraph(True, keep_graph=True)`` is ambiguous and
|
|
# must raise.
|
|
with self.assertRaises(TypeError):
|
|
paddle.cuda.CUDAGraph(True, keep_graph=True)
|
|
|
|
def test_pool_returns_int_idempotent(self):
|
|
g = paddle.cuda.CUDAGraph()
|
|
token = g.pool()
|
|
self.assertIsInstance(token, int)
|
|
self.assertEqual(g.pool(), token)
|
|
|
|
def test_graph_pool_handle_returns_int(self):
|
|
token = paddle.cuda.graph_pool_handle()
|
|
self.assertIsInstance(token, int)
|
|
# Each call should yield a fresh token.
|
|
self.assertNotEqual(paddle.cuda.graph_pool_handle(), token)
|
|
|
|
def test_capture_begin_pool_propagation(self):
|
|
g = paddle.cuda.CUDAGraph()
|
|
# Stub the C++ binding so we don't enter an actual capture; we only
|
|
# care that the pool argument is propagated into ``_pool_id``.
|
|
from paddle.device.cuda import graphs as graphs_module
|
|
|
|
original = graphs_module.CoreCUDAGraph.begin_capture_with_pool_id
|
|
graphs_module.CoreCUDAGraph.begin_capture_with_pool_id = (
|
|
lambda *args, **kwargs: None
|
|
)
|
|
try:
|
|
g.capture_begin(pool=42)
|
|
self.assertEqual(g._pool_id, 42)
|
|
finally:
|
|
graphs_module.CoreCUDAGraph.begin_capture_with_pool_id = original
|
|
|
|
def test_capture_error_mode_propagated_and_validated(self):
|
|
# capture_error_mode must be validated against ALL_MODES and the
|
|
# resolved mode index must be forwarded to the C++ binding rather
|
|
# than ignored. Construct with the matching mode so the precedence
|
|
# warning (covered separately) does not fire here.
|
|
from paddle.device.cuda import graphs as graphs_module
|
|
|
|
captured = {}
|
|
original = graphs_module.CoreCUDAGraph.begin_capture_with_pool_id
|
|
|
|
def fake_begin(place, mode, pool_id, enable_replace):
|
|
captured["mode"] = mode
|
|
|
|
graphs_module.CoreCUDAGraph.begin_capture_with_pool_id = fake_begin
|
|
try:
|
|
for name, expected in [
|
|
("global", 0),
|
|
("thread_local", 1),
|
|
("relaxed", 2),
|
|
]:
|
|
g = paddle.cuda.CUDAGraph(mode=name)
|
|
g.capture_begin(capture_error_mode=name)
|
|
self.assertEqual(captured["mode"], expected)
|
|
|
|
g = paddle.cuda.CUDAGraph()
|
|
with self.assertRaises(ValueError):
|
|
g.capture_begin(capture_error_mode="not-a-mode")
|
|
finally:
|
|
graphs_module.CoreCUDAGraph.begin_capture_with_pool_id = original
|
|
|
|
def test_capture_error_mode_override_warns(self):
|
|
# An explicit ``capture_error_mode`` overrides the constructor's
|
|
# ``mode`` for this capture; surface a ``UserWarning`` so callers
|
|
# know which one wins. Matching values must not warn.
|
|
import warnings as _w
|
|
|
|
from paddle.device.cuda import graphs as graphs_module
|
|
|
|
original = graphs_module.CoreCUDAGraph.begin_capture_with_pool_id
|
|
graphs_module.CoreCUDAGraph.begin_capture_with_pool_id = (
|
|
lambda *args, **kwargs: None
|
|
)
|
|
try:
|
|
g = paddle.cuda.CUDAGraph(mode='thread_local')
|
|
with _w.catch_warnings(record=True) as ws:
|
|
_w.simplefilter('always')
|
|
g.capture_begin(capture_error_mode='global')
|
|
msgs = [str(w.message) for w in ws]
|
|
self.assertTrue(
|
|
any('takes precedence' in m for m in msgs),
|
|
f'expected precedence warning, got: {msgs}',
|
|
)
|
|
|
|
g = paddle.cuda.CUDAGraph(mode='thread_local')
|
|
with _w.catch_warnings(record=True) as ws:
|
|
_w.simplefilter('always')
|
|
g.capture_begin(capture_error_mode='thread_local')
|
|
msgs = [str(w.message) for w in ws]
|
|
self.assertFalse(
|
|
any('takes precedence' in m for m in msgs),
|
|
f'unexpected precedence warning: {msgs}',
|
|
)
|
|
finally:
|
|
graphs_module.CoreCUDAGraph.begin_capture_with_pool_id = original
|
|
|
|
def test_capture_begin_uses_constructor_mode_when_unset(self):
|
|
# When ``capture_error_mode`` is omitted, ``capture_begin`` must
|
|
# fall back to the constructor's ``mode`` rather than silently
|
|
# forcing 'global'. Existing Paddle callers like
|
|
# ``CUDAGraph(mode='relaxed').capture_begin()`` rely on this.
|
|
from paddle.device.cuda import graphs as graphs_module
|
|
|
|
captured = {}
|
|
original = graphs_module.CoreCUDAGraph.begin_capture_with_pool_id
|
|
|
|
def fake_begin(place, mode, pool_id, enable_replace):
|
|
captured["mode"] = mode
|
|
|
|
graphs_module.CoreCUDAGraph.begin_capture_with_pool_id = fake_begin
|
|
try:
|
|
for name, expected in [
|
|
("global", 0),
|
|
("thread_local", 1),
|
|
("relaxed", 2),
|
|
]:
|
|
g = paddle.cuda.CUDAGraph(mode=name)
|
|
g.capture_begin()
|
|
self.assertEqual(captured["mode"], expected)
|
|
finally:
|
|
graphs_module.CoreCUDAGraph.begin_capture_with_pool_id = original
|
|
|
|
def test_capture_begin_materializes_pool_id(self):
|
|
# ``g.pool()`` must return the same id that ``capture_begin`` passed
|
|
# to the C++ side, so that ``g2.capture_begin(pool=g.pool())`` shares
|
|
# the same memory pool. Before materialization, capture_begin would
|
|
# forward ``None`` and pool() would later mint a *different* id.
|
|
from paddle.device.cuda import graphs as graphs_module
|
|
|
|
captured = {}
|
|
original = graphs_module.CoreCUDAGraph.begin_capture_with_pool_id
|
|
|
|
def fake_begin(place, mode, pool_id, enable_replace):
|
|
captured["pool_id"] = pool_id
|
|
|
|
graphs_module.CoreCUDAGraph.begin_capture_with_pool_id = fake_begin
|
|
try:
|
|
g = paddle.cuda.CUDAGraph()
|
|
g.capture_begin()
|
|
self.assertIsNotNone(captured["pool_id"])
|
|
self.assertEqual(g.pool(), captured["pool_id"])
|
|
finally:
|
|
graphs_module.CoreCUDAGraph.begin_capture_with_pool_id = original
|
|
|
|
def test_instantiate_requires_capture(self):
|
|
# ``instantiate`` returns the core graph built by ``capture_end``; on a
|
|
# freshly constructed graph there is nothing captured yet, so it raises.
|
|
g = paddle.cuda.CUDAGraph()
|
|
with self.assertRaises(RuntimeError):
|
|
g.instantiate()
|
|
|
|
def test_debug_dump_requires_enable_debug_mode(self):
|
|
g = paddle.cuda.CUDAGraph()
|
|
with self.assertRaises(RuntimeError):
|
|
g.debug_dump("/tmp/no_such_path")
|
|
g.enable_debug_mode()
|
|
self.assertTrue(g._debug_mode)
|
|
|
|
def test_raw_handles_not_implemented(self):
|
|
g = paddle.cuda.CUDAGraph()
|
|
with self.assertRaises(NotImplementedError):
|
|
g.raw_cuda_graph()
|
|
with self.assertRaises(NotImplementedError):
|
|
g.raw_cuda_graph_exec()
|
|
|
|
def test_capture_replay_round_trip(self):
|
|
if not can_use_cuda_graph():
|
|
return
|
|
|
|
shape = [2, 3]
|
|
x = paddle.to_tensor(
|
|
np.random.randint(0, 10, size=shape).astype("float32")
|
|
)
|
|
z = paddle.to_tensor(
|
|
np.random.randint(0, 10, size=shape).astype("float32")
|
|
)
|
|
|
|
g = paddle.cuda.CUDAGraph()
|
|
g.capture_begin()
|
|
y = x + 10
|
|
z.add_(x)
|
|
g.capture_end()
|
|
# ``instantiate`` returns the core CUDA graph held after capture.
|
|
self.assertIs(g.instantiate(), g._graph)
|
|
|
|
z_before = z.numpy().copy()
|
|
x_new = paddle.to_tensor(
|
|
np.random.randint(0, 10, size=shape).astype("float32")
|
|
)
|
|
x.copy_(x_new, False)
|
|
g.replay()
|
|
self.assertTrue((y.numpy() - x.numpy() == 10).all())
|
|
self.assertTrue((z.numpy() - z_before == x.numpy()).all())
|
|
g.reset()
|
|
|
|
|
|
@unittest.skipIf(
|
|
not (paddle.is_compiled_with_cuda() or is_custom_device())
|
|
or float(paddle.version.cuda()) < 11.0,
|
|
"only support cuda >= 11.0",
|
|
)
|
|
class TestPaddleCudaGraphContextManager(unittest.TestCase):
|
|
"""Tests for ``paddle.cuda.graph``, the PyTorch-compat context manager."""
|
|
|
|
def setUp(self):
|
|
if can_use_cuda_graph():
|
|
paddle.set_flags(
|
|
{
|
|
'FLAGS_allocator_strategy': 'auto_growth',
|
|
'FLAGS_sync_nccl_allreduce': False,
|
|
'FLAGS_cudnn_deterministic': True,
|
|
'FLAGS_use_stream_safe_cuda_allocator': False,
|
|
}
|
|
)
|
|
|
|
def test_graph_context_manager_propagates_args(self):
|
|
# Verify ``with paddle.cuda.graph(g, pool=...)`` invokes
|
|
# ``capture_begin``/``capture_end`` and threads ``pool`` through.
|
|
g = paddle.cuda.CUDAGraph()
|
|
from paddle.device.cuda import graphs as graphs_module
|
|
|
|
called = {"begin": 0, "end": 0}
|
|
original_begin = graphs_module.CoreCUDAGraph.begin_capture_with_pool_id
|
|
original_end = graphs_module.CoreCUDAGraph.end_capture
|
|
graphs_module.CoreCUDAGraph.begin_capture_with_pool_id = (
|
|
lambda *args, **kwargs: called.__setitem__(
|
|
"begin", called["begin"] + 1
|
|
)
|
|
)
|
|
graphs_module.CoreCUDAGraph.end_capture = lambda: called.__setitem__(
|
|
"end", called["end"] + 1
|
|
)
|
|
try:
|
|
with paddle.cuda.graph(g, pool=7):
|
|
pass
|
|
self.assertEqual(called["begin"], 1)
|
|
self.assertEqual(called["end"], 1)
|
|
self.assertEqual(g._pool_id, 7)
|
|
finally:
|
|
graphs_module.CoreCUDAGraph.begin_capture_with_pool_id = (
|
|
original_begin
|
|
)
|
|
graphs_module.CoreCUDAGraph.end_capture = original_end
|
|
|
|
def test_graph_context_manager_capture_replay(self):
|
|
if not can_use_cuda_graph():
|
|
return
|
|
|
|
shape = [2, 3]
|
|
x = paddle.to_tensor(
|
|
np.random.randint(0, 10, size=shape).astype("float32")
|
|
)
|
|
|
|
g = paddle.cuda.CUDAGraph()
|
|
with paddle.cuda.graph(g):
|
|
y = x + 1
|
|
g.replay()
|
|
np.testing.assert_allclose(y.numpy(), x.numpy() + 1, rtol=1e-5)
|
|
g.reset()
|
|
|
|
def test_graph_context_manager_switches_stream(self):
|
|
# When ``stream`` is provided, the body must execute on that stream
|
|
# and the previous stream must be restored on exit.
|
|
if not can_use_cuda_graph():
|
|
return
|
|
|
|
side_stream = paddle.device.Stream()
|
|
prev_stream = paddle.device.current_stream()
|
|
observed = {}
|
|
|
|
g = paddle.cuda.CUDAGraph()
|
|
from paddle.device.cuda import graphs as graphs_module
|
|
|
|
original_begin = graphs_module.CoreCUDAGraph.begin_capture_with_pool_id
|
|
original_end = graphs_module.CoreCUDAGraph.end_capture
|
|
graphs_module.CoreCUDAGraph.begin_capture_with_pool_id = (
|
|
lambda *args, **kwargs: observed.__setitem__(
|
|
"inside", paddle.device.current_stream()
|
|
)
|
|
)
|
|
graphs_module.CoreCUDAGraph.end_capture = lambda: None
|
|
try:
|
|
with paddle.cuda.graph(g, stream=side_stream):
|
|
pass
|
|
finally:
|
|
graphs_module.CoreCUDAGraph.begin_capture_with_pool_id = (
|
|
original_begin
|
|
)
|
|
graphs_module.CoreCUDAGraph.end_capture = original_end
|
|
|
|
# ``Stream`` wrappers are fresh Python objects each call; compare by
|
|
# repr which embeds the underlying CUDA stream pointer.
|
|
self.assertEqual(repr(observed["inside"]), repr(side_stream))
|
|
self.assertEqual(
|
|
repr(paddle.device.current_stream()), repr(prev_stream)
|
|
)
|
|
|
|
def test_graph_context_manager_default_does_not_warn(self):
|
|
# ``with paddle.cuda.graph(CUDAGraph()):`` is the most common
|
|
# default usage; it must not surface the
|
|
# ``capture_error_mode-takes-precedence`` warning, since the user
|
|
# didn't pick ``capture_error_mode`` here.
|
|
import warnings as _w
|
|
|
|
from paddle.device.cuda import graphs as graphs_module
|
|
|
|
original_begin = graphs_module.CoreCUDAGraph.begin_capture_with_pool_id
|
|
original_end = graphs_module.CoreCUDAGraph.end_capture
|
|
graphs_module.CoreCUDAGraph.begin_capture_with_pool_id = (
|
|
lambda *args, **kwargs: None
|
|
)
|
|
graphs_module.CoreCUDAGraph.end_capture = lambda: None
|
|
try:
|
|
g = paddle.cuda.CUDAGraph()
|
|
with _w.catch_warnings(record=True) as ws:
|
|
_w.simplefilter('always')
|
|
with paddle.cuda.graph(g):
|
|
pass
|
|
msgs = [str(w.message) for w in ws]
|
|
self.assertFalse(
|
|
any('takes precedence' in m for m in msgs),
|
|
f'unexpected precedence warning: {msgs}',
|
|
)
|
|
finally:
|
|
graphs_module.CoreCUDAGraph.begin_capture_with_pool_id = (
|
|
original_begin
|
|
)
|
|
graphs_module.CoreCUDAGraph.end_capture = original_end
|
|
|
|
def test_graph_context_manager_stream_none_is_noop(self):
|
|
# When ``stream`` is None the current stream must be unchanged.
|
|
if not can_use_cuda_graph():
|
|
return
|
|
|
|
prev_stream = paddle.device.current_stream()
|
|
g = paddle.cuda.CUDAGraph()
|
|
from paddle.device.cuda import graphs as graphs_module
|
|
|
|
original_begin = graphs_module.CoreCUDAGraph.begin_capture_with_pool_id
|
|
original_end = graphs_module.CoreCUDAGraph.end_capture
|
|
graphs_module.CoreCUDAGraph.begin_capture_with_pool_id = (
|
|
lambda *args, **kwargs: None
|
|
)
|
|
graphs_module.CoreCUDAGraph.end_capture = lambda: None
|
|
try:
|
|
with paddle.cuda.graph(g):
|
|
self.assertEqual(
|
|
repr(paddle.device.current_stream()), repr(prev_stream)
|
|
)
|
|
finally:
|
|
graphs_module.CoreCUDAGraph.begin_capture_with_pool_id = (
|
|
original_begin
|
|
)
|
|
graphs_module.CoreCUDAGraph.end_capture = original_end
|
|
|
|
self.assertEqual(
|
|
repr(paddle.device.current_stream()), repr(prev_stream)
|
|
)
|
|
|
|
|
|
if __name__ == "__main__":
|
|
unittest.main()
|