Flags for varying NIM seed and temperature every call

garak/generators/nim.py

@@ -3,10 +3,12 @@
 
 """NVIDIA Inference Microservice LLM interface"""
 
+import random
 from typing import List, Union
 
 import openai
 
+from garak import _config
 from garak.generators.openai import OpenAICompatible
 
 
@@ -38,6 +40,8 @@ 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/",
+        "vary_seed_each_call": True,  # encourage variation when generations>1. not respected by all NIMs
+        "vary_temp_each_call": True,  # encourage variation when generations>1. not respected by all NIMs
         "suppressed_params": {"n", "frequency_penalty", "presence_penalty"},
     }
     active = True
@@ -67,6 +71,13 @@ def _call_model(
         assert (
             generations_this_call == 1
         ), "generations_per_call / n > 1 is not supported"
+
+        if self.vary_seed_each_call:
+            _config.run.seed = random.randint(0, 65535)
+
+        if self.vary_temp_each_call:
+            self.temperature = random.random()
+
         return super()._call_model(prompt, generations_this_call)
 
 

garak/generators/openai.py

@@ -25,9 +25,9 @@
 # lists derived from https://platform.openai.com/docs/models
 chat_models = (
     "gpt-4",  # links to latest version
-    "gpt-4-turbo", # links to latest version
-    "gpt-4o", # links to latest version
-    "gpt-4o-mini", # links to latest version
+    "gpt-4-turbo",  # links to latest version
+    "gpt-4o",  # links to latest version
+    "gpt-4o-mini",  # links to latest version
     "gpt-4-turbo-preview",
     "gpt-3.5-turbo",  # links to latest version
     "gpt-4-32k",
@@ -185,6 +185,9 @@ def _call_model(
             "stop": self.stop,
         }
 
+        if _config.run.seed is not None:
+            create_args["seed"] = _config.run.seed
+
         create_args = {
             k: v
             for k, v in create_args.items()