Proxy fails to start on Windows with AttributeError: module 'os' has no attribute 'WNOHANG'.

[shivamrawat1]

Cause: PR #21899 added a Prisma engine watchdog that uses os.waitpid() and os.WNOHANG, which are Unix-only. On Windows these are missing and cause the crash.
Fix: Add sys.platform == "win32" checks so that on Windows _try_waitpid_watch returns False (falling back to os.kill polling) and _reap_all_zombies returns an empty set.

Relevant issues

Pre-Submission checklist

Please complete all items before asking a LiteLLM maintainer to review your PR

• I have Added testing in the tests/test_litellm/ directory, Adding at least 1 test is a hard requirement - see details
• My PR passes all unit tests on make test-unit
• My PR's scope is as isolated as possible, it only solves 1 specific problem
• I have requested a Greptile review by commenting @greptileai and received a Confidence Score of at least 4/5 before requesting a maintainer review

CI (LiteLLM team)

CI status guideline:

• 50-55 passing tests: main is stable with minor issues.
• 45-49 passing tests: acceptable but needs attention
• <= 40 passing tests: unstable; be careful with your merges and assess the risk.

• Branch creation CI run
  Link:

• CI run for the last commit
  Link:

• Merge / cherry-pick CI run
  Links:

Type

🐛 Bug Fix
✅ Test

Changes

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated (UTC)
litellm	Ready	Preview, Comment	Mar 12, 2026 10:51pm

[greptile-apps]

Greptile Summary

This PR fixes a Windows-specific startup crash introduced in PR #21899, where the Prisma engine watchdog called os.waitpid() and os.WNOHANG — both Unix-only APIs — causing an AttributeError on Windows. The fix adds sys.platform == "win32" early-return guards to _reap_all_zombies and _try_waitpid_watch, ensuring Windows falls through gracefully to the existing os.kill(pid, 0) polling fallback (which works cross-platform).

• litellm/proxy/utils.py: adds import sys and two if sys.platform == "win32": return ... guards. The _try_pidfd_watch method was already safe (guarded by hasattr(os, "pidfd_open")), and the os.kill polling path (_poll_engine_proc) is cross-platform.
• tests/litellm/proxy/test_prisma_engine_watchdog.py: adds test_try_waitpid_watch_returns_false_on_windows and test_reap_all_zombies_returns_empty_on_windows using patch("sys.platform", "win32") to exercise the new guards. All new tests are mocked and make no real network calls.
• Minor note: existing Unix-specific tests such as test_try_waitpid_watch_returns_false_when_not_child do not patch sys.platform, so on an actual Windows CI runner they would return False via the new Windows guard instead of the ChildProcessError path — the assertions still pass, but the test exercises a different code path than intended.

Confidence Score: 5/5

• This PR is safe to merge — it is a narrow, well-tested defensive fix for a Windows crash with no impact on existing Unix behaviour.
• The change is minimal (two early-return guards + import sys), the logic is straightforward, and the fallback to os.kill polling on Windows was already implemented and tested. No existing Unix code paths are altered. Two new mock-only tests cover the new branches. No backwards-incompatible changes, no new database queries, and no provider-specific code introduced.
• No files require special attention.

Important Files Changed

Filename	Overview
litellm/proxy/utils.py	Adds sys.platform == "win32" guards to _reap_all_zombies and _try_waitpid_watch to prevent AttributeError on Windows where os.waitpid/os.WNOHANG are unavailable. The fix is minimal and correct; _try_pidfd_watch already had a hasattr(os, "pidfd_open") guard. The fallback to os.kill polling remains fully functional on Windows.
tests/litellm/proxy/test_prisma_engine_watchdog.py	Adds two new unit tests for the Windows early-return paths. Tests use patch("sys.platform", "win32") which correctly patches the global sys module. One subtle issue: existing Unix-only tests (e.g. test_try_waitpid_watch_returns_false_when_not_child) do not guard against running on Windows, where they would return False via the new Windows check rather than the ChildProcessError path — still passing, but for the wrong reason.

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A[_start_engine_watcher] --> B[_try_waitpid_watch]
    B -- "Windows: sys.platform == 'win32'" --> C[return False]
    B -- "Unix: call os.waitpid with WNOHANG" --> D{is child?}
    D -- No --> E[return False]
    D -- Yes --> F[Start waitpid thread]
    F --> G[return True]

    C --> H[_try_pidfd_watch]
    E --> H
    H -- "no hasattr os.pidfd_open" --> I[return False]
    H -- "Linux: pidfd_open succeeds" --> J[return True]

    I --> K[os.kill polling fallback\n_poll_engine_proc every 1s]

    subgraph "_reap_all_zombies"
        L{sys.platform == 'win32'?}
        L -- Yes --> M[return empty set]
        L -- No --> N[os.waitpid -1 WNOHANG loop]
    end

Loading

_{Last reviewed commit: 4e87fca}