Wrapped LLM as a garak generator #1382

Nakul-Rajpal · 2025-09-25T22:54:54Z

Wrapped the Python Library LLM (for OpenAI, Anthropic’s Claude, Google’s Gemini etc) as a garak generator.

Fixes Issue #463

Wrapped the Python Library LLM (for OpenAI, Anthropic’s Claude, Google’s Gemini etc) as a garak generator.

Nakul-Rajpal · 2025-09-25T22:55:04Z

@leondz Please check

leondz · 2025-09-26T07:12:18Z

thanks, will take a look!

leondz · 2025-09-26T08:17:50Z

@Nakul-Rajpal This isn't passing tests - can you amend?

Nakul-Rajpal · 2025-09-26T18:24:56Z

@leondz It should be good now? The tests were failing due to me not adding the llm library to the requirements so it ran without the module.

Nakul-Rajpal · 2025-09-29T16:09:36Z

@leondz Should be ready to go now; really sorry about the errors before I request to do another issue I should familiarize myself with the repo further.

leondz · 2025-09-29T16:10:51Z

Yeah, it's in the review queue, thank you

jmartin-tech

This looks like a great start. I noted a few edge case and a mismatch in how a system_prompt is handled.

Please take a look, happy to offer further detail or answer question about how things flow.

jmartin-tech · 2025-09-29T13:54:52Z

garak/generators/llm.py

+        if self.system:
+            prompt_kwargs["system"] = self.system


Current system prompt support in garak is tied to the conversation passed as part of prompt. The DEFAULT_PARAMS entry here should likely be removed in favor of extracting the system_prompt from the prompt via prompt.last_message("system"). That is if passing a conversation that includes the system message would not apply it.

jmartin-tech · 2025-09-30T13:57:56Z

garak/generators/llm.py

+        "max_tokens": None,
+        "top_p": None,
+        "stop": [],
+        "system": None,


Remove, the system prompt is set via the run configuration and pass to generators as part of the prompt conversation.

Suggested change

"system": None,

jmartin-tech · 2025-09-30T14:03:13Z

garak/generators/llm.py

+        if self.max_tokens is not None:
+            prompt_kwargs["max_tokens"] = self.max_tokens
+        if self.temperature is not None:
+            prompt_kwargs["temperature"] = self.temperature
+        if self.top_p is not None:
+            prompt_kwargs["top_p"] = self.top_p
+        if self.stop:
+            prompt_kwargs["stop"] = self.stop


None == False and all keys defined in DEFAULT_PARAMS will exist on self.

Suggested change

if self.max_tokens is not None:

prompt_kwargs["max_tokens"] = self.max_tokens

if self.temperature is not None:

prompt_kwargs["temperature"] = self.temperature

if self.top_p is not None:

prompt_kwargs["top_p"] = self.top_p

if self.stop:

prompt_kwargs["stop"] = self.stop

if self.max_tokens:

prompt_kwargs["max_tokens"] = self.max_tokens

if self.temperature:

prompt_kwargs["temperature"] = self.temperature

if self.top_p:

prompt_kwargs["top_p"] = self.top_p

if self.stop:

prompt_kwargs["stop"] = self.stop

jmartin-tech · 2025-09-30T14:16:30Z

garak/generators/llm.py

+
+        This calls model.prompt() once per generation and materializes the text().
+        """
+        text_prompt = prompt.last_message().text


This does not grab out the full conversation. There is an existing helper function in the base class Generator._conversation_to_list() that will format the garak Conversation object as a list of dictionaries meeting the HuggingFace and OpenAI conversation list. Looking at how the llm library handles what it considers to be conversation I don't know if there is a way to load a prefilled history in a similar pattern to how chat completions APIs for other generators are working.

For best adoption, this generator should at least validate the conversation has at most one user and one system message to know if the prompt passed will be fully processed during inference.

jmartin-tech · 2025-09-30T14:18:46Z

garak/generators/llm.py

+        "temperature": None,
+        "max_tokens": None,


temperature and max_tokens are already in Generator.DEFAULT_PARAMS is there a reason to include here?

Suggested change

"temperature": None,

"max_tokens": None,

leondz

Looks pretty good. Requests to add a pattern supporting parallelisation, some renaming, and vars ensuring test consistency

leondz · 2025-10-09T04:34:13Z

garak/generators/llm.py

+        "system": None,
+    }
+
+    generator_family_name = "LLM"


this might be better as lower case - that's how the tool is described

leondz · 2025-10-09T04:39:30Z

garak/generators/llm.py

+
+        try:
+            # Resolve the llm model; fall back to llm's default if no name given
+            self.model = llm.get_model(self.name) if self.name else llm.get_model()


Can we rename self.model to self.target to be consistent with overall garak nomenclature? llm supports both systems and models so the change fits this use-case fine too

leondz · 2025-10-09T04:41:13Z

garak/generators/llm.py

It's worth implementing the _load_client() / _clear_client() pattern here to support parallelisation - see openai.OpenAICompatible for an example

leondz · 2025-10-09T04:42:06Z

tests/generators/test_llm.py

+# SPDX-FileCopyrightText: Portions Copyright (c) 2025 NVIDIA CORPORATION &
+#                         AFFILIATES. All rights reserved.


Suggested change

# SPDX-FileCopyrightText: Portions Copyright (c) 2025 NVIDIA CORPORATION &

# AFFILIATES. All rights reserved.

# SPDX-FileCopyrightText: Portions Copyright (c) 2025 NVIDIA CORPORATION & AFFILIATES. All rights reserved.

leondz · 2025-10-09T04:44:58Z

pyproject.toml

we should land #1199 before landing this, and then move this PR to the deferred loading pattern

leondz · 2025-10-09T04:52:50Z

tests/generators/test_llm.py

+def test_generate_returns_message(cfg, fake_llm):
+    gen = LLMGenerator(name="alias", config_root=cfg)
+
+    conv = Conversation([Turn("user", Message(text="ping"))])


Suggested change

conv = Conversation([Turn("user", Message(text="ping"))])

test_txt = "ping"

conv = Conversation([Turn("user", Message(text=test_txt))])

leondz · 2025-10-09T04:53:00Z

tests/generators/test_llm.py

+    assert out[0].text == "OK_FAKE"
+
+    prompt_text, kwargs = fake_llm.calls[0]
+    assert prompt_text == "ping"


Suggested change

assert prompt_text == "ping"

assert prompt_text == test_txt

leondz · 2025-10-09T04:53:47Z

tests/generators/test_llm.py

+    gen.temperature = 0.2
+    gen.max_tokens = 64
+    gen.top_p = 0.9
+    gen.stop = ["\n\n"]
+    gen.system = "you are testy"


use vars for these values (and the checks later)

leondz · 2025-10-09T04:54:40Z

tests/generators/test_llm.py

+    assert kwargs["temperature"] == 0.2
+    assert kwargs["max_tokens"] == 64
+    assert kwargs["top_p"] == 0.9
+    assert kwargs["stop"] == ["\n\n"]
+    assert kwargs["system"] == "you are testy"


vars here. could do a list assignment / check likex,y = 1,2 for brevity

leondz · 2025-10-09T04:59:13Z

tests/generators/test_llm.py

+    class BoomModel:
+        def prompt(self, *a, **k):
+            raise RuntimeError("boom")
+    monkeypatch.setattr(llm, "get_model", lambda *a, **k: BoomModel())


jmartin-tech

Some tweaks to ensure consistent behaviour.

Code suggestions are untested.

jmartin-tech · 2025-10-16T14:36:16Z

garak/generators/llm.py

+        self._load_client()
+
+        super().__init__(self.name, config_root=config_root)
+
+        self._clear_client()


It may be best to call super().__init__() before calling _load_client() to ensure the end object state is not impacted. Also no need to call _clear_client() during __init__().

Suggested change

self._load_client()

super().__init__(self.name, config_root=config_root)

self._clear_client()

super().__init__(self.name, config_root=config_root)

self._load_client()

jmartin-tech · 2025-10-16T14:37:10Z

garak/generators/llm.py

+    Calls model.prompt() with the prompt text and relays the response. Per-provider
+    options and API keys are all handled by `llm` (e.g., `llm keys set openai`).
+
+    Set --model_name to the `llm` model id or alias (e.g., "gpt-4o-mini",


Suggested change

Set --model_name to the `llm` model id or alias (e.g., "gpt-4o-mini",

Set --target_name to the `llm` model id or alias (e.g., "gpt-4o-mini",

jmartin-tech · 2025-10-17T13:34:05Z

garak/generators/llm.py

+        if assistant_turns:
+            raise ValueError("llm generator does not accept assistant turns")
+        if len(system_turns) > 1:
+            raise ValueError("llm generator supports at most one system turn")
+        if len(user_turns) != 1:
+            raise ValueError("llm generator requires exactly one user turn")


Unsupported prompts in _call_model() are currently expected to return None to avoid early termination of the test run. While I understand the thought process of using raise here, currently logging the reason for skipping the prompt would align better.

import logging

Suggested change

if assistant_turns:

raise ValueError("llm generator does not accept assistant turns")

if len(system_turns) > 1:

raise ValueError("llm generator supports at most one system turn")

if len(user_turns) != 1:

raise ValueError("llm generator requires exactly one user turn")

if assistant_turns:

logging.debug("llm generator does not accept assistant turns")

return [None] * generations_this_call

if len(system_turns) > 1:

logging.debug("llm generator supports at most one system turn")

return [None] * generations_this_call

if len(user_turns) != 1:

logging.debug("llm generator requires exactly one user turn")

return [None] * generations_this_call

jmartin-tech · 2025-10-17T13:40:20Z

garak/generators/llm.py

+        prompt_kwargs = {
+            key: getattr(self, key)
+            for key in ("max_tokens", "temperature", "top_p")
+            if getattr(self, key) is not None
+        }
+        if self.stop:
+            prompt_kwargs["stop"] = self.stop


Could this inspect the accepted arguments to self.target.prompt() vs a hard coded list here? Something similar exists in the OpenAICompatible class, where we collect all options set on the generator that the target API accepts.

Wrapped LLM as a garak generator

5f0b1af

Wrapped the Python Library LLM (for OpenAI, Anthropic’s Claude, Google’s Gemini etc) as a garak generator.

Nakul-Rajpal added 2 commits September 26, 2025 10:01

Added LLM library Dependency

a7d109b

Added LLM library to requirements.

6a54502

Nakul-Rajpal added 2 commits September 26, 2025 15:30

Add LLM Docs

65645dc

Added LLM library to generators.rst

094141d

jmartin-tech requested changes Oct 1, 2025

View reviewed changes

leondz requested changes Oct 9, 2025

View reviewed changes

leondz added the generators Interfaces with LLMs label Oct 12, 2025

Added Recommended Changes

7f20877

Nakul-Rajpal requested review from jmartin-tech and leondz October 15, 2025 22:57

Merge branch 'main' into generator-wrapLLM

f2d730e

jmartin-tech requested changes Oct 23, 2025

View reviewed changes

		# SPDX-FileCopyrightText: Portions Copyright (c) 2025 NVIDIA CORPORATION &
		# AFFILIATES. All rights reserved.

	conv = Conversation([Turn("user", Message(text="ping"))])
	test_txt = "ping"
	conv = Conversation([Turn("user", Message(text=test_txt))])

	Set --model_name to the `llm` model id or alias (e.g., "gpt-4o-mini",
	Set --target_name to the `llm` model id or alias (e.g., "gpt-4o-mini",

Wrapped LLM as a garak generator #1382

Are you sure you want to change the base?

Wrapped LLM as a garak generator #1382

Uh oh!

Conversation

Nakul-Rajpal commented Sep 25, 2025

Uh oh!

Nakul-Rajpal commented Sep 25, 2025

Uh oh!

leondz commented Sep 26, 2025

Uh oh!

leondz commented Sep 26, 2025

Uh oh!

Nakul-Rajpal commented Sep 26, 2025

Uh oh!

Nakul-Rajpal commented Sep 29, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

leondz commented Sep 29, 2025

Uh oh!

jmartin-tech left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

leondz left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

jmartin-tech left a comment

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

Nakul-Rajpal commented Sep 29, 2025 •

edited

Loading