generator: NIM

[leondz]

Add NIM generator support. This PR only covers NVIDIA-hosted NIMs

[leondz]

See https://build.nvidia.com/explore/reasoning for items to test on

[jmartin-tech]

Based on having #645 landed we might refactor this initial support as OpenAICompatible and later extend to add a NemoLLMCompatible implementation.

There is a little more work needed beyond my comments here to enable, context_len for unknown models to fallback to None when not known to the tool since OpenAICompatible currently just rejects any model without a known context length in the garak.generators.openai.context_lengths dictionary.

See: https://github.com/leondz/garak/blob/075796f044c2877977383dcf561b5a54f865c009/garak/generators/openai.py#L135-L138

[jmartin-tech]

Update look good, some nice to have items noted.

[jmartin-tech]

In a future update I would like to see these tests converted to sub in a mock response fixture when the ENV_VAR is not set.

Consider this tests, could assert the correct type is created by injecting a mock key similar to the test in test_ggml.py

https://github.com/leondz/garak/blob/77ebd8ab242b07dd9ec0d551154933831dbb8092/tests/generators/test_ggml.py#L12-L28

Integration level tests that actually call out to a service can provide a mock when no key was in the ENV at launch and intercept the request call via request-mock in this class otherwise allowing the code to still validate changes even when a true endpoint is not available.

[leondz]

That's a great idea, yes. I would also like to see this play better with generator internal api_keys, after the generator has validated / loaded its config, rather than going straight for the environment variable or _config. Tagging #589 as a note that this can be considered & addressed when handling that issue.

[jmartin-tech]

Not something to address in this PR, however I am noticing this file is used for testing, I wonder if it is possible to move it into the tests path as a fixture only loaded during testing.

Is there added value in having this available all the time?

[leondz]

I was thinking the same. No, there's no added value - the primary purpose of the plugin package test modules is to help integrators and users, not to ensure code integrity.

[jmartin-tech]

Functional testing passes however when executing the tests with a valid NIM_API_KEY in place tests/test_attempt.py reports an attempt to write to a file that no longer is closed.

======================================================================== FAILURES =========================================================================
_______________________________________________________________ test_attempt_sticky_params ________________________________________________________________

capsys = <_pytest.capture.CaptureFixture object at 0x33939e570>

    def test_attempt_sticky_params(capsys):
>       garak.cli.main(
            "-m test.Blank -g 1 -p atkgen,dan.Dan_6_0 --report_prefix _garak_test_attempt_sticky_params".split()
        )

tests/test_attempt.py:13:
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
garak/cli.py:486: in main
    command.probewise_run(generator, probe_names, evaluator, buff_names)
garak/command.py:212: in probewise_run
    probewise_h.run(generator, probe_names, evaluator, buffs)
garak/harnesses/probewise.py:106: in run
    h.run(model, [probe], detectors, evaluator, announce_probe=False)
garak/harnesses/base.py:124: in run
    evaluator.evaluate(attempt_results)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

self = <garak.evaluators.base.ThresholdEvaluator object at 0x3389ad370>, attempts = [<garak.attempt.Attempt object at 0x3394dbad0>]

    def evaluate(self, attempts: List[garak.attempt.Attempt]) -> None:
        """
        evaluate feedback from detectors
        expects a list of attempts that correspond to one probe
        outputs results once per detector
        """

        if len(attempts) == 0:
            logging.debug(
                "evaluators.base.Evaluator.evaluate called with 0 attempts, expected 1+"
            )
            return

        self.probename = attempts[0].probe_classname
        detector_names = attempts[0].detector_results.keys()

        for detector in detector_names:
            all_passes = []
            all_outputs = []
            for attempt in attempts:
                passes = [
                    1 if self.test(r) else 0
                    for r in map(float, attempt.detector_results[detector])
                ]
                all_passes += passes
                all_outputs += attempt.outputs
                for idx, score in enumerate(attempt.detector_results[detector]):
                    if not self.test(score):  # if we don't pass
                        if not _config.transient.hitlogfile:
                            if not _config.reporting.report_prefix:
                                hitlog_filename = f"{_config.reporting.report_dir}/garak.{_config.transient.run_id}.hitlog.jsonl"
                            else:
                                hitlog_filename = (
                                    _config.reporting.report_prefix + ".hitlog.jsonl"
                                )
                            logging.info("hit log in %s", hitlog_filename)
                            _config.transient.hitlogfile = open(
                                hitlog_filename,
                                "w",
                                buffering=1,
                                encoding="utf-8",
                            )

                        trigger = None
                        if "trigger" in attempt.notes:
                            trigger = attempt.notes["trigger"]
                        elif "triggers" in attempt.notes:
                            if (
                                isinstance(attempt.notes["triggers"], list)
                                and len(attempt.notes["triggers"]) == 1
                            ):  # a list of one can be reported just as a string
                                trigger = attempt.notes["triggers"][0]
                            else:
                                trigger = attempt.notes["triggers"]
>                       _config.transient.hitlogfile.write(
                            json.dumps(
                                {
                                    "goal": attempt.goal,
                                    "prompt": attempt.prompt,
                                    "output": attempt.outputs[idx],
                                    "trigger": trigger,
                                    "score": score,
                                    "run_id": str(_config.transient.run_id),
                                    "attempt_id": str(attempt.uuid),
                                    "attempt_seq": attempt.seq,
                                    "attempt_idx": idx,
                                    "generator": f"{_config.plugins.model_type} {_config.plugins.model_name}",
                                    "probe": self.probename,
                                    "detector": detector,
                                    "generations_per_prompt": _config.run.generations,
                                }
                            )
                            + "\n"  # generator,probe,prompt,trigger,result,detector,score,run id,attemptid,
                        )
E                       ValueError: I/O operation on closed file.

garak/evaluators/base.py:91: ValueError
------------------------------------------------------------------ Captured stdout call -------------------------------------------------------------------
garak LLM security probe v0.9.0.12.post1 ( https://github.com/leondz/garak ) at 2024-05-07T15:06:41.549252
📜 reporting to _garak_test_attempt_sticky_params.report.jsonl
🦜 loading generator: Test: Blank
🕵️  queue of probes: atkgen.Tox, dan.Dan_6_0
🔴🪖  🦜 loading generator: Hugging Face 🤗 pipeline: leondz/artgpt2tox
atkgen.Tox                                                                toxicity.ToxicCommentModel: PASS  ok on    2/   2
dan.Dan_6_0                                                                                  dan.DAN: PASS  ok on    1/   1
------------------------------------------------------------------ Captured stderr call -------------------------------------------------------------------

================================================================= short test summary info =================================================================
FAILED tests/test_attempt.py::test_attempt_sticky_params - ValueError: I/O operation on closed file.
================================================== 1 failed, 908 passed, 5 skipped in 191.22s (0:03:11) ===================================================

This consistently fails in local testing on my end. However executing the single test file in isolation passes:

% python -m pytest tests/test_attempt.py
=================================================================== test session starts ===================================================================
platform darwin -- Python 3.12.2, pytest-8.1.1, pluggy-1.4.0
rootdir: /Users/jemartin/Projects/nvidia/garak
configfile: pyproject.toml
plugins: anyio-4.3.0, requests-mock-1.12.1
collected 1 item

tests/test_attempt.py .                                                                                                                             [100%]

==================================================================== 1 passed in 4.77s ====================================================================

And fails in the same manor when skipped nim tests execute first:

python -m pytest tests/generators/test_nim.py tests/test_attempt.py

This suggests there is some sort of data leakage or stale state in global config introduced by the skipped tests.

It seems reasonable to file a known issue with tests when NIM_API_KEY is in the env to let this move forward. Other efforts are planned to adjust _config scoping that may address this issue in #646.

[leondz]

Interesting. Thanks for writing this up. Yes, agree with your hypothesis.

The failure may be related to these - tagging for reference:

• test order affects test success/speed. it shouldn't test order affects test success/speed. it shouldn't #473
• _config object contains state not suitable for multiprocessing _config object contains state not suitable for multiprocessing #646
• Hitlogfile write exception Hitlogfile write exception #599

If we're lucky, addressing #646 may clear up a bunch of these.

[leondz]

Now getting the exact same test fail error, also in test_attempt_sticky_params. Will prioritise fixing this for release milestone.