remove extraneous default params for nims that expect conservative pa…

garak/generators/nim.py

@@ -38,6 +38,7 @@ class NVOpenAIChat(OpenAICompatible):
         "top_p": 0.7,
         "top_k": 0,  # top_k is hard set to zero as of 24.04.30
         "uri": "https://integrate.api.nvidia.com/v1/",
+        "suppressed_params": {"n", "frequency_penalty", "presence_penalty"},
     }
     active = True
     supports_multiple_generations = False

garak/generators/openai.py

@@ -98,14 +98,9 @@ class OpenAICompatible(Generator):
         "frequency_penalty": 0.0,
         "presence_penalty": 0.0,
         "stop": ["#", ";"],
+        "suppressed_params": set(),
     }
 
-    temperature = 0.7
-    top_p = 1.0
-    frequency_penalty = 0.0
-    presence_penalty = 0.0
-    stop = ["#", ";"]
-
     # avoid attempt to pickle the client attribute
     def __getstate__(self) -> object:
         self._clear_client()
@@ -162,6 +157,20 @@ def _call_model(
         if self.client is None:
             # reload client once when consuming the generator
             self._load_client()
+
+        create_args = {
+            "model": self.name,
+            "temperature": self.temperature,
+            "max_tokens": self.max_tokens,
+            "n": generations_this_call,
+            "top_p": self.top_p,
+            "frequency_penalty": self.frequency_penalty,
+            "presence_penalty": self.presence_penalty,
+            "stop": self.stop,
+        }
+
+        create_args = {k: v for k, v in create_args.items() if v is not None and k not in self.suppressed_params}
+
         if self.generator == self.client.completions:
             if not isinstance(prompt, str):
                 msg = (
@@ -172,17 +181,9 @@ def _call_model(
                 print(msg)
                 return list()
 
-            response = self.generator.create(
-                model=self.name,
-                prompt=prompt,
-                temperature=self.temperature,
-                max_tokens=self.max_tokens,
-                n=generations_this_call,
-                top_p=self.top_p,
-                frequency_penalty=self.frequency_penalty,
-                presence_penalty=self.presence_penalty,
-                stop=self.stop,
-            )
+            create_args["prompt"] = prompt
+
+            response = self.generator.create(**create_args)
             return [c.text for c in response.choices]
 
         elif self.generator == self.client.chat.completions:
@@ -199,17 +200,9 @@ def _call_model(
                 print(msg)
                 return list()
             try:
-                response = self.generator.create(
-                    model=self.name,
-                    messages=messages,
-                    temperature=self.temperature,
-                    top_p=self.top_p,
-                    n=generations_this_call,
-                    stop=self.stop,
-                    max_tokens=self.max_tokens,
-                    presence_penalty=self.presence_penalty,
-                    frequency_penalty=self.frequency_penalty,
-                )
+                create_args["messages"] = messages
+                response = self.generator.create(**create_args)
+
                 return [c.message.content for c in response.choices]
             except openai.BadRequestError:
                 msg = "Bad request: " + str(repr(prompt))

tests/generators/test_nim.py

@@ -52,3 +52,21 @@ def test_nim_parallel_attempts():
 def test_nim_hf_detector():
     garak.cli.main("-m nim -p lmrc.Bullying -g 1 -n google/gemma-2b".split())
     assert True
+
+
+@pytest.mark.skipif(
+    os.getenv(NVOpenAIChat.ENV_VAR, None) is None,
+    reason=f"NIM API key is not set in {NVOpenAIChat.ENV_VAR}",
+)
+def test_nim_conservative_api():  # extraneous params can throw 422
+    g = NVOpenAIChat(name="nvidia/nemotron-4-340b-instruct")
+    result = g._call_model("this is a test", generations_this_call=1)
+    assert isinstance(result, list), "NIM _call_model should return a list"
+    assert len(result) == 1, "NIM _call_model result list should have one item"
+    assert isinstance(result[0], str), "NIM _call_model should return a list"
+    result = g.generate("this is a test", generations_this_call=1)
+    assert isinstance(result, list), "NIM generate() should return a list"
+    assert (
+        len(result) == 1
+    ), "NIM generate() result list should have one item when generations_this_call=1"
+    assert isinstance(result[0], str), "NIM generate() should return a list"