fix: keep GPU fingerprint import safe on CPU CI

[yyswhsccc]

Summary

• Defers PyTorch/CUDA requirement checks until the GPU fingerprint flow actually runs.
• Keeps module import safe on CPU-only CI / no-torch environments.
• Adds an import-only regression test for the no-torch path.

Follow-up scope

This is the small GPU import-safety slice called out as acceptable in #6711.

It intentionally does not touch node/admin/auth endpoints, production security guards, or the repo-wide stale-test/admin-key baseline handled separately in #6788.

Validation

• python3 -m py_compile miners/gpu_fingerprint.py tests/test_gpu_fingerprint_import.py
• uv run --no-project --with pytest --with flask --with requests --with flask-cors --with httpx --with cryptography python -m pytest -q tests/test_gpu_fingerprint_import.py --tb=short -o addopts='' --basetemp=.pytest-tmp -> 1 passed
• git diff --check
• python3 miners/gpu_fingerprint.py --json without torch -> exits 1 with ERROR: PyTorch with CUDA support required. Install: pip install torch

Refs #305.
Follow-up to #6711.

wallet: RTC47bc28896a1a4bf240d1fd780f4559b242bcd945

[vinhcafe25-ops]

Code Review: PR #6818 — fix: keep GPU fingerprint import safe on CPU CI

Tier: Thorough — Line-by-line review. Core approach is sound; the test has a critical bug.

---

Finding 1 (🔴 Bug — Test): monkeypatch.setitem(sys.modules, "torch", None) does NOT simulate a missing torch

File: tests/test_gpu_fingerprint_import.py

monkeypatch.setitem(sys.modules, "torch", None)
monkeypatch.setitem(sys.modules, "torch.cuda", None)

This does not cause import torch to raise ImportError. Python's import machinery checks sys.modules via key existence (name in sys.modules), not truthiness. Setting the entry to None means:

1. import torch → finds key in sys.modules → returns None → no ImportError
2. import torch.cuda → finds key in sys.modules → returns None → no ImportError
3. The except ImportError block is never reached
4. The else branch runs → HAS_TORCH = True
5. assert module.HAS_TORCH is False → FAILS

The test should delete the keys from sys.modules instead, which is what actually happens on a machine without torch:

# Correct approach: remove from sys.modules so import raises ImportError
monkeypatch.delitem(sys.modules, "torch", raising=False)
monkeypatch.delitem(sys.modules, "torch.cuda", raising=False)

Or use importlib mocking to make the import genuinely raise:

import builtins
original_import = builtins.__import__
def mock_import(name, *args, **kwargs):
    if name in ("torch", "torch.cuda"):
        raise ImportError(f"No module named '{name}'")
    return original_import(name, *args, **kwargs)
monkeypatch.setattr(builtins, "__import__", mock_import)

Impact: The test either never actually ran, or it "passes" by coincidence (e.g., torch was already imported in the test session and monkeypatch.setitem overwrites the real module, but the spec.loader.exec_module path might bypass sys.modules in some edge case). Either way, the test is not testing what it claims to test.

---

Finding 2 (🟡 Fragility): Both torch and torch.cuda share one except block

File: miners/gpu_fingerprint.py

try:
    import torch
    import torch.cuda
except ImportError:
    torch = None
    HAS_TORCH = False
else:
    HAS_TORCH = True

If import torch succeeds but import torch.cuda fails (edge case: corrupted install), the except block sets torch = None, discarding the successfully-imported torch module. This is wasteful — torch could be used for CPU ops even without CUDA.

Better:

try:
    import torch
except ImportError:
    torch = None
    HAS_TORCH = False
else:
    try:
        import torch.cuda
    except ImportError:
        pass  # CUDA submodule missing, handled by check_requirements()
    HAS_TORCH = True

---

Finding 3 (🟡 Coverage gap): No test when torch IS available

The new test only covers the torch-missing path. Consider adding a second test that verifies the module behaves correctly when torch and torch.cuda are actually importable (e.g., with a mock that makes torch.cuda.is_available() return True). This would catch regressions in the happy path.

---

Finding 4 (🟢 Nit): from __future__ import annotations is unrelated

Added at the top of the file but not used by any change in this PR. This is PEP 604 style (str | None syntax). Fine if the project is standardizing, but it's unrelated noise in this diff.

---

What works well

• check_requirements() pattern is clean — defers enforcement to call time, separating detection from action
• run_gpu_fingerprint guards with check_requirements() at the entry point
• CLI __main__ block wraps the whole thing in try/except RuntimeError for clean user-facing errors
• Test uses importlib.util.spec_from_file_location correctly to avoid cached imports

---

Verdict

Request Changes — The test in test_gpu_fingerprint_import.py is broken: setting sys.modules["torch"] = None does not simulate a missing module. Fix the test approach to use delitem or builtins.__import__ mocking. The core gpu_fingerprint.py change is solid.

[jaxint]

Great contribution to the RustChain ecosystem!

[jaxint]

Thanks for this PR! The changes look good. 🎉

[Scottcjn]

✅ Merging — this is the clean split I asked for

Exactly the gpu_fingerprint.py import-safety fix from #6711 (lazy check_requirements() instead of import-time sys.exit), on its own with its import test — none of #6711's monitor-NameError or auth-on-public-GET changes. Files are only miners/gpu_fingerprint.py + tests/test_gpu_fingerprint_import.py. Thanks @yyswhsccc.