File size: 7,256 Bytes
ea35292
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
67
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
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
from collections.abc import Iterable, Iterator
import logging
import os
from pathlib import Path

import torch

from bitsandbytes.cextension import HIP_ENVIRONMENT, get_cuda_bnb_library_path
from bitsandbytes.cuda_specs import CUDASpecs
from bitsandbytes.diagnostics.utils import print_dedented

CUDART_PATH_PREFERRED_ENVVARS = ("CONDA_PREFIX", "LD_LIBRARY_PATH")

CUDART_PATH_IGNORED_ENVVARS = {
    "DBUS_SESSION_BUS_ADDRESS",  # hardware related
    "GOOGLE_VM_CONFIG_LOCK_FILE",  # GCP: requires elevated permissions, causing problems in VMs and Jupyter notebooks
    "HOME",  # Linux shell default
    "LESSCLOSE",
    "LESSOPEN",  # related to the `less` command
    "MAIL",  # something related to emails
    "OLDPWD",
    "PATH",  # this is for finding binaries, not libraries
    "PWD",  # PWD: this is how the shell keeps track of the current working dir
    "SHELL",  # binary for currently invoked shell
    "SSH_AUTH_SOCK",  # SSH stuff, therefore unrelated
    "SSH_TTY",
    "TMUX",  # Terminal Multiplexer
    "XDG_DATA_DIRS",  # XDG: Desktop environment stuff
    "XDG_GREETER_DATA_DIR",  # XDG: Desktop environment stuff
    "XDG_RUNTIME_DIR",
    "_",  # current Python interpreter
}

CUDA_RUNTIME_LIB_PATTERNS = (
    (
        "libamdhip64.so*",  # Linux
        "amdhip64*.dll",  # Windows
    )
    if HIP_ENVIRONMENT
    else (
        "cudart64*.dll",  # Windows
        "libcudart*.so*",  # libcudart.so, libcudart.so.11.0, libcudart.so.12.0, libcudart.so.12.1, libcudart.so.12.2 etc.
        "nvcuda*.dll",  # Windows
    )
)

logger = logging.getLogger(__name__)


def find_cuda_libraries_in_path_list(paths_list_candidate: str) -> Iterable[Path]:
    for dir_string in paths_list_candidate.split(os.pathsep):
        if not dir_string:
            continue
        if os.sep not in dir_string:
            continue
        try:
            dir = Path(dir_string)
            try:
                if not dir.exists():
                    logger.warning(f"The directory listed in your path is found to be non-existent: {dir}")
                    continue
            except OSError:  # Assume an esoteric error trying to poke at the directory
                pass
            for lib_pattern in CUDA_RUNTIME_LIB_PATTERNS:
                for pth in dir.glob(lib_pattern):
                    if pth.is_file() and not pth.is_symlink():
                        yield pth
        except (OSError, PermissionError):
            pass


def is_relevant_candidate_env_var(env_var: str, value: str) -> bool:
    return (
        env_var in CUDART_PATH_PREFERRED_ENVVARS  # is a preferred location
        or (
            os.sep in value  # might contain a path
            and env_var not in CUDART_PATH_IGNORED_ENVVARS  # not ignored
            and "CONDA" not in env_var  # not another conda envvar
            and "BASH_FUNC" not in env_var  # not a bash function defined via envvar
            and "\n" not in value  # likely e.g. a script or something?
        )
    )


def get_potentially_lib_path_containing_env_vars() -> dict[str, str]:
    return {env_var: value for env_var, value in os.environ.items() if is_relevant_candidate_env_var(env_var, value)}


def find_cudart_libraries() -> Iterator[Path]:
    """
    Searches for a cuda installations, in the following order of priority:
        1. active conda env
        2. LD_LIBRARY_PATH
        3. any other env vars, while ignoring those that
            - are known to be unrelated
            - don't contain the path separator `/`

    If multiple libraries are found in part 3, we optimistically try one,
    while giving a warning message.
    """
    candidate_env_vars = get_potentially_lib_path_containing_env_vars()

    for envvar in CUDART_PATH_PREFERRED_ENVVARS:
        if envvar in candidate_env_vars:
            directory = candidate_env_vars[envvar]
            yield from find_cuda_libraries_in_path_list(directory)
            candidate_env_vars.pop(envvar)

    for env_var, value in candidate_env_vars.items():
        yield from find_cuda_libraries_in_path_list(value)


def _print_cuda_diagnostics(cuda_specs: CUDASpecs) -> None:
    print(
        f"PyTorch settings found: CUDA_VERSION={cuda_specs.cuda_version_string}, "
        f"Highest Compute Capability: {cuda_specs.highest_compute_capability}.",
    )

    binary_path = get_cuda_bnb_library_path(cuda_specs)
    if not binary_path.exists():
        print_dedented(
            f"""
            No compatible CUDA library found (tried: {binary_path.name}). You may need to compile from source:
            https://huggingface.co/docs/bitsandbytes/main/en/installation#cuda-compile
            """,
        )

    # 7.5 is the minimum CC for int8 tensor cores
    if not cuda_specs.has_imma:
        print_dedented(
            """
            WARNING: Compute capability < 7.5 detected! Only slow 8-bit matmul is supported for your GPU!
            If you run into issues with 8-bit matmul, you can try 4-bit quantization:
            https://huggingface.co/blog/4bit-transformers-bitsandbytes
            """,
        )


def _print_hip_diagnostics(cuda_specs: CUDASpecs) -> None:
    print(f"PyTorch settings found: ROCM_VERSION={cuda_specs.cuda_version_string}")

    rocm_override = os.environ.get("BNB_ROCM_VERSION")
    if rocm_override:
        print(f"BNB_ROCM_VERSION override: {rocm_override}")

    binary_path = get_cuda_bnb_library_path(cuda_specs)
    if not binary_path.exists():
        print_dedented(
            f"""
            No compatible ROCm library found (tried: {binary_path.name}). You may need to compile from source:
            https://huggingface.co/docs/bitsandbytes/main/en/installation#rocm-compile
            Use BNB_ROCM_VERSION to force a specific version if needed.
            """,
        )

    hip_major, hip_minor = cuda_specs.cuda_version_tuple
    if (hip_major, hip_minor) < (6, 1):
        print_dedented(
            """
            WARNING: bitsandbytes is fully supported only from ROCm 6.1.
            """,
        )


def print_diagnostics(cuda_specs: CUDASpecs) -> None:
    if HIP_ENVIRONMENT:
        _print_hip_diagnostics(cuda_specs)
    else:
        _print_cuda_diagnostics(cuda_specs)


def print_runtime_diagnostics() -> None:
    backend = "ROCm" if HIP_ENVIRONMENT else "CUDA"
    runtime_version = torch.version.hip if HIP_ENVIRONMENT else torch.version.cuda
    override_var = "BNB_ROCM_VERSION" if HIP_ENVIRONMENT else "BNB_CUDA_VERSION"
    override_example = "72" if HIP_ENVIRONMENT else "122"

    cudart_paths = list(find_cudart_libraries())
    if not cudart_paths:
        print(f"{backend} SETUP: WARNING! {backend} runtime files not found in any environmental path.")
    elif len(cudart_paths) > 1:
        print_dedented(
            f"""
            Found duplicate {backend} runtime files (see below).

            bitsandbytes will use PyTorch's {backend} runtime ({runtime_version}) and auto-select
            the closest available library version. If you need to force a specific version,
            set {override_var}, e.g.:
                export {override_var}={override_example}
            """,
        )
        for pth in cudart_paths:
            print(f"* Found {backend} runtime at: {pth}")