Repository navigation
Expand file tree
/
Copy path__init__.py
More file actions
135 lines (105 loc) · 5.56 KB
/
Copy path__init__.py
File metadata and controls
135 lines (105 loc) · 5.56 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
# Copyright 2024 Bytedance Ltd. and/or its affiliates
#
# 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 importlib
import logging
import os
from packaging.version import parse as parse_version
# `verl.protocol` (imported just below) is the first module that pulls in Ray, so any
# Ray flag read at import time must be set here first. When the training driver runs
# under `uv run`, Ray >=2.47 registers an automatic "uv" runtime-env hook that fires
# during `ray.init()`. That hook rewrites `runtime_env["working_dir"]`, and because
# verl deliberately passes an explicit `working_dir=None` (see
# verl/trainer/constants_ppo.py) it crashes in `_is_path(None)` with
# "TypeError: path_or_uri must be a string, got NoneType". verl already forwards the
# uv environment to Ray workers via `runtime_env.py_executable`, so the hook is
# redundant; disable it. `setdefault` keeps it overridable (set "1" to re-enable).
os.environ.setdefault("RAY_ENABLE_UV_RUN_RUNTIME_ENV", "0")
from .protocol import DataProto # noqa: E402
from .utils.device import is_npu_available # noqa: E402
from .utils.import_utils import import_external_libs # noqa: E402
from .utils.logging_utils import set_basic_config # noqa: E402
version_folder = os.path.dirname(os.path.join(os.path.abspath(__file__)))
with open(os.path.join(version_folder, "version/version")) as f:
__version__ = f.read().strip()
set_basic_config(level=logging.WARNING)
__all__ = ["DataProto", "__version__"]
modules = os.getenv("VERL_USE_EXTERNAL_MODULES", "")
if modules:
modules = modules.split(",")
import_external_libs(modules)
# Auto-discover plugins via setuptools entry_points.
# Controlled by VERL_USE_EXTERNAL_PLUGINS:
# "auto" — load all entry_points in the "verl.plugins" group (default)
# "none" — disable entry_point discovery entirely
# "pkg1,pkg2" — only load the named entry_points
_plugins_policy = os.getenv("VERL_USE_EXTERNAL_PLUGINS", "auto").strip().lower()
if _plugins_policy != "none":
from importlib.metadata import entry_points as _entry_points
_discovered = _entry_points(group="verl.plugins")
if _plugins_policy == "auto":
_allowed = None
else:
_allowed = {name.strip() for name in _plugins_policy.split(",") if name.strip()}
for _ep in _discovered:
if _allowed is not None and _ep.name not in _allowed:
continue
try:
_ep.load()
except Exception as _e:
logging.getLogger(__name__).debug("Failed to load plugin '%s': %s", _ep.name, _e)
if os.getenv("VERL_USE_MODELSCOPE", "False").lower() == "true":
if importlib.util.find_spec("modelscope") is None:
raise ImportError("You are using the modelscope hub, please install modelscope by `pip install modelscope -U`")
# Patch hub to download models from modelscope to speed up.
from modelscope.utils.hf_util import patch_hub
patch_hub()
if is_npu_available:
# Workaround for torch-npu's lack of support for creating nested tensors from NPU tensors.
#
# ```
# >>> a, b = torch.arange(3).npu(), torch.arange(5).npu() + 3
# >>> nt = torch.nested.nested_tensor([a, b], layout=torch.jagged)
# ```
# throws "not supported in npu" on Ascend NPU.
# See https://github.com/Ascend/pytorch/blob/294cdf5335439b359991cecc042957458a8d38ae/torch_npu/utils/npu_intercept.py#L109
# for details.
import torch
try:
if hasattr(torch.nested.nested_tensor, "__wrapped__"):
torch.nested.nested_tensor = torch.nested.nested_tensor.__wrapped__
if hasattr(torch.nested.as_nested_tensor, "__wrapped__"):
torch.nested.as_nested_tensor = torch.nested.as_nested_tensor.__wrapped__
except AttributeError:
pass
# In verl, the driver process aggregates the computation results of workers via Ray.
# Therefore, after a worker completes its computation job, it will package the output
# using tensordict and transfer it to the CPU. Since the `to` operation of tensordict
# is non-blocking, when transferring data from a device to the CPU, it is necessary to
# ensure that a batch of data has been completely transferred before being used on the
# host; otherwise, unexpected precision issues may arise. Tensordict has already noticed
# this problem and fixed it. Ref: https://github.com/pytorch/tensordict/issues/725
# However, the relevant modifications only cover CUDA and MPS devices and do not take effect
# for third-party devices such as NPUs. This patch fixes this issue, and the relevant
# modifications can be removed once the fix is merged into tensordict.
import tensordict
if parse_version(tensordict.__version__) < parse_version("0.10.0"):
from tensordict.base import TensorDictBase
def _sync_all_patch(self):
from torch._utils import _get_available_device_type, _get_device_module
device_type = _get_available_device_type()
if device_type is None:
return
device_module = _get_device_module(device_type)
device_module.synchronize()
TensorDictBase._sync_all = _sync_all_patch