mineru-rocm/docker/scripts/patch_vllm_platform.py

#!/usr/bin/env python3
"""vllm 平台检测补丁

问题 1：amdsmi 不可用时 platform 回退到 torch.version.hip
问题 2：rocm.py 中 logger.warning_once() 导致循环导入（默认不再改写；仅保留为显式开关）
"""

import os
import re
import sys
import sysconfig
import traceback

VLLM_DIR = '/opt/vllm/vllm'
VLLM_SRC_DIR = '/opt/vllm'


def patch6_init_platform_fallback():
    """补丁 6：platforms/__init__.py —— torch.version.hip 兜底"""
    f = os.path.join(VLLM_DIR, 'platforms', '__init__.py')
    c = open(f).read()
    old = "    return 'vllm.platforms.rocm.RocmPlatform' if is_rocm else None"
    new = ("    # amdsmi fallback: also check torch.version.hip\n"
           "    if not is_rocm:\n"
           "        try:\n"
           "            import torch\n"
           "            if torch.version.hip is not None:\n"
           "                is_rocm = True\n"
           "        except Exception:\n"
           "            pass\n"
           "    return 'vllm.platforms.rocm.RocmPlatform' if is_rocm else None")
    c2 = c.replace(old, new)
    if c2 != c:
        open(f, 'w').write(c2)
        print('Patch 6: __init__.py platform fallback applied.')
    else:
        print('Patch 6: already applied or pattern not found.')


def patch7_rocm_break_import_cycle():
    """补丁 7：platforms/rocm.py —— logger.warning_once → sys.stderr.write

    默认启用安全的“整段 except 块”替换，避免 logger.warning_once() 在平台检测
    期间触发循环导入，导致 current_platform 落到 UnspecifiedPlatform。
    如需禁用，设置 VLLM_PATCH_ROCM_WARNING_ONCE=0。
    """
    if os.environ.get('VLLM_PATCH_ROCM_WARNING_ONCE') == '0':
        print('Patch 7: skipped (VLLM_PATCH_ROCM_WARNING_ONCE=0).')
        return
    f = os.path.join(VLLM_DIR, 'platforms', 'rocm.py')
    c = open(f).read()
    if 'amdsmi unavailable, using torch.cuda fallback' in c:
        print('Patch 7: already applied.')
        return
    pattern = re.compile(
        r'(\n\s*except Exception as e:\n'
        r'\s*logger\.debug\("Failed to get GCN arch via amdsmi: %s", e\)\n'
        r'\s*logger\.warning_once\(\n'
        r'(?:\s*"[^"]*"\n)+'
        r'\s*\)\n)',
        re.MULTILINE,
    )

    def repl(match):
        indent = re.search(r'\n(\s*)except Exception as e:', match.group(1)).group(1)
        body_indent = indent + '    '
        return (
            f'\n{indent}except Exception as e:\n'
            f'{body_indent}import sys as _sys\n'
            f'{body_indent}_sys.stderr.write('
            '"vLLM ROCm: amdsmi unavailable, using torch.cuda fallback for GPU detection\\n")\n'
        )

    c2, n = pattern.subn(repl, c, count=1)
    if c2 != c:
        open(f, 'w').write(c2)
        print('Patch 7: rocm.py logger.warning_once circular import patch applied.')
    else:
        print('Patch 7: already applied or pattern not found.')


def ensure_vllm_dist_info():
    """为 /opt/vllm 源码导入创建最小 dist-info。

    vLLM 平台检测会通过 importlib.metadata 查询 vllm 分发元数据/entry points。
    如果只靠 PYTHONPATH 导入源码，没有 dist-info，就会出现：
    "The vLLM package was not found..."，并可能得到 UnspecifiedPlatform。
    """
    try:
        import vllm
    except Exception:
        return

    version = getattr(vllm, '__version__', '0.1.dev1') or '0.1.dev1'
    site_packages = sysconfig.get_paths().get('purelib')
    if not site_packages:
        return
    dist_info = os.path.join(site_packages, f'vllm-{version}.dist-info')
    os.makedirs(dist_info, exist_ok=True)

    metadata = os.path.join(dist_info, 'METADATA')
    if not os.path.exists(metadata):
        with open(metadata, 'w') as fp:
            fp.write(
                'Metadata-Version: 2.1\n'
                'Name: vllm\n'
                f'Version: {version}\n'
                'Summary: vLLM source tree mounted at /opt/vllm\n'
            )

    with open(os.path.join(dist_info, 'top_level.txt'), 'w') as fp:
        fp.write('vllm\n')
    with open(os.path.join(dist_info, 'INSTALLER'), 'w') as fp:
        fp.write('mineru-rocm-runtime\n')
    with open(os.path.join(dist_info, 'entry_points.txt'), 'w') as fp:
        fp.write(
            '[vllm.platform_plugins]\n'
            'rocm = vllm.platforms.rocm:rocm_platform_plugin\n'
        )
    with open(os.path.join(dist_info, 'RECORD'), 'w') as fp:
        fp.write('')
    print(f'vllm dist-info ensured: {dist_info}')


def ensure_vllm_installed():
    """确保 vLLM Python 包在当前虚拟环境中可导入。

    某些 MinerU 依赖安装流程后，site-packages 中可能缺少 vLLM 分发元数据，但
    /opt/vllm 源码仍在。此时直接把 /opt/vllm 加入 PYTHONPATH/sys.path 再导入，
    避免 editable 安装触发 pyproject 元数据校验失败。
    """
    try:
        import vllm  # noqa: F401
        return
    except ModuleNotFoundError as exc:
        if exc.name != 'vllm':
            raise

    if not os.path.isdir(VLLM_SRC_DIR):
        raise ModuleNotFoundError(
            'vllm is not installed and /opt/vllm source directory is missing'
        )

    print('vllm package missing; trying PYTHONPATH fallback with /opt/vllm...')
    current = os.environ.get('PYTHONPATH', '')
    paths = [p for p in current.split(':') if p]
    if VLLM_SRC_DIR not in paths:
        os.environ['PYTHONPATH'] = f"{VLLM_SRC_DIR}:{current}" if current else VLLM_SRC_DIR
    if VLLM_SRC_DIR not in sys.path:
        sys.path.insert(0, VLLM_SRC_DIR)

    import vllm  # noqa: F401
    print('vllm import recovered via PYTHONPATH fallback.')


def main():
    patch6_init_platform_fallback()
    patch7_rocm_break_import_cycle()
    try:
        ensure_vllm_installed()
        ensure_vllm_dist_info()
        import vllm
        from vllm.platforms import current_platform
        print(f'vllm runtime import OK: {vllm.__version__}, platform={type(current_platform).__name__}')
        if type(current_platform).__name__ == 'UnspecifiedPlatform':
            raise RuntimeError(
                'vLLM platform detection returned UnspecifiedPlatform; '
                'ROCm backend is not usable. Check amdsmi, /dev/kfd, /dev/dri, '
                'and vllm.platform_plugins metadata.'
            )
    except Exception:
        print('ERROR: vllm runtime import failed after platform patches:')
        traceback.print_exc()
        raise
    print('vllm platform patches done.')


if __name__ == '__main__':
    main()