[Chore] Separate out system utilities from vllm.utils

[dongbo910220]

Part of #26900

This PR continues the effort to clean up vllm.utils by extracting various system-level helper functions into a new, dedicated module vllm/utils/system_utils.py.

The following functions have been moved:

• vllm.utils.update_environment_variables -> vllm.utils.system_utils.update_environment_variables
• vllm.utils.set_env_var -> vllm.utils.system_utils.set_env_var
• vllm.utils.unique_filepath -> vllm.utils.system_utils.unique_filepath
• vllm.utils.set_process_title -> vllm.utils.system_utils.set_process_title
• vllm.utils.decorate_logs -> vllm.utils.system_utils.decorate_logs

CC: @DarkLight1337

Test Plan

Test Result

[gemini-code-assist]

Code Review

This pull request is a good step in refactoring vllm.utils by moving system-related utilities to a new system_utils.py module. The changes are clean and preserve the public API. I've found one instance of redundant code that should be removed.

[mergify]

Documentation preview: https://vllm--27201.org.readthedocs.build/en/27201/

[DarkLight1337]

/gemini review

[gemini-code-assist]

Code Review

This pull request is a good step in refactoring vllm.utils by extracting system-level utilities into a new vllm.utils.system_utils module. The moved functions are now better organized and include some nice improvements, such as safer environment variable handling and checking for optional dependencies.

I've identified one area for improvement in the newly created system_utils.py file. The log decoration logic relies on monkey-patching sys.stdout and sys.stderr, which is a fragile approach. I've suggested a more robust implementation using a wrapper class. This change will make the logging functionality more reliable, especially in diverse environments like testing or when using custom loggers.

[gemini-code-assist]

The current implementation of decorate_logs and _add_prefix relies on monkey-patching sys.stdout.write and sys.stderr.write, and adding a custom attribute start_new_line to the stream objects. This approach is fragile and can lead to unexpected errors if sys.stdout or sys.stderr are redirected to objects that do not support attribute or method assignment (e.g., in certain testing environments or with custom logging handlers). Additionally, the monkey-patched write method does not return the number of characters written, which violates the TextIO.write interface.

A more robust and cleaner approach is to use a wrapper class for the stream objects. This avoids modifying global objects directly, encapsulates the prefixing logic cleanly, and correctly implements the stream interface. The suggested change refactors this logic into a _LogPrefixStream class, making the log decoration more reliable and maintainable.

class _LogPrefixStream:
    """A stream wrapper that adds a prefix to each line of output."""

    def __init__(self, stream: TextIO, prefix: str):
        self._stream = stream
        self._prefix = prefix
        self.start_new_line = True

    def write(self, s: str) -> int:
        if not s:
            return 0

        if self.start_new_line:
            self._stream.write(self._prefix)

        idx = 0
        while (next_idx := s.find("\n", idx)) != -1:
            next_idx += 1
            self._stream.write(s[idx:next_idx])
            if next_idx == len(s):
                self.start_new_line = True
                return len(s)
            self._stream.write(self._prefix)
            idx = next_idx

        self._stream.write(s[idx:])
        self.start_new_line = False
        return len(s)

    def __getattr__(self, name: str):
        return getattr(self._stream, name)


def decorate_logs(process_name: str | None = None) -> None:
    """Decorate stdout/stderr with process name and PID prefix."""
    from vllm.utils import get_mp_context

    if process_name is None:
        process_name = get_mp_context().current_process().name
    pid = os.getpid()
    prefix = f"{CYAN}({process_name} pid={pid}){RESET} "
    sys.stdout = _LogPrefixStream(sys.stdout, prefix)
    sys.stderr = _LogPrefixStream(sys.stderr, prefix)

[DarkLight1337]

Gemini didn't flag any missing imports so LGTM, see if the tests pass now

[dongbo910220]

Hi @DarkLight1337 ,It looks like the Docker build CI is failing. I took a look at the logs and found this error: Error: No solution found when resolving dependencies: apache-tvm-ffi==0.1.0b15

Since this PR only refactors some Python utility functions and doesn't touch any dependency or build configurations, I'm wondering if this might be an unrelated issue from the base branch .

[DarkLight1337]

It should be fixed on main now, let me rebase

[dongbo910220]

Hi @DarkLight1337 , After merging the latest main branch, the CI failures are still the same error we saw in the previous run.
The persistent error:
All failing tests show the identical error:
ImportError while loading conftest '/vllm-workspace/tests/conftest.py'.
ImportError: /usr/local/lib/python3.12/dist-packages/vllm/_C.abi3.so:
undefined symbol: ZNK3c106SymInt6sym_neERKS0

This PR only modifies Python code - function locations and import paths. No C++ code or
compilation configuration was touched.
Maybe we could wait until this issue is resolved

[DarkLight1337]

PTAL at the failing tests

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @dongbo910220.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[DarkLight1337]

Please fix merge conflicts

[DarkLight1337]

You seem to have accidentally added a file called @@ at the root