paraphrase fast consistent model device (NVIDIA#898)

When running tests on a platform with a `cuda` capable device, errors
can be reported due to how the model is instantiated.

Recent changes to support a configuration of plugins and huggingface
models as used in this class lead to further identification of possible
exceptions when processing a `HFCompatible` class that does not have
`hf_args` or when providing options for huggingface model constructors
that favor the newer `device_map` argument to `device`.

garak/buffs/paraphrase.py

@@ -7,6 +7,7 @@
 
 import garak.attempt
 from garak import _config
+from garak.generators.huggingface import HFCompatible
 from garak.buffs.base import Buff
 
 
@@ -69,14 +70,17 @@ def transform(
             yield paraphrased_attempt
 
 
-class Fast(Buff):
+class Fast(Buff, HFCompatible):
     """CPU-friendly paraphrase buff based on Humarin's T5 paraphraser"""
 
+    DEFAULT_PARAMS = Buff.DEFAULT_PARAMS | {
+        "para_model_name": "garak-llm/chatgpt_paraphraser_on_T5_base",
+        "hf_args": {"device": "cpu", "torch_dtype": "float32"},
+    }
     bcp47 = "en"
     doc_uri = "https://huggingface.co/humarin/chatgpt_paraphraser_on_T5_base"
 
     def __init__(self, config_root=_config) -> None:
-        self.para_model_name = "garak-llm/chatgpt_paraphraser_on_T5_base"
         self.num_beams = 5
         self.num_beam_groups = 5
         self.num_return_sequences = 5
@@ -85,20 +89,23 @@ def __init__(self, config_root=_config) -> None:
         self.no_repeat_ngram_size = 2
         # self.temperature = 0.7
         self.max_length = 128
-        self.torch_device = None
+        self.device = None
         self.tokenizer = None
         self.para_model = None
         super().__init__(config_root=config_root)
 
     def _load_model(self):
-        import torch
         from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
 
-        self.torch_device = "cuda" if torch.cuda.is_available() else "cpu"
-        self.tokenizer = AutoTokenizer.from_pretrained(self.para_model_name)
+        self.device = self._select_hf_device()
+        model_kwargs = self._gather_hf_params(
+            hf_constructor=AutoModelForSeq2SeqLM.from_pretrained
+        )  # will defer to device_map if device map was `auto` may not match self.device
+
         self.para_model = AutoModelForSeq2SeqLM.from_pretrained(
-            self.para_model_name
-        ).to(self.torch_device)
+            self.para_model_name, **model_kwargs
+        ).to(self.device)
+        self.tokenizer = AutoTokenizer.from_pretrained(self.para_model_name)
 
     def _get_response(self, input_text):
         if self.para_model is None:

garak/generators/huggingface.py

@@ -53,10 +53,17 @@ def _set_hf_context_len(self, config):
                 self.context_len = config.n_ctx
 
     def _gather_hf_params(self, hf_constructor: Callable):
-        # this may be a bit too naive as it will pass any parameter valid for the pipeline signature
-        # this falls over when passed `from_pretrained` methods as the callable model params are not explicit
-        params = self.hf_args
-        if params["device"] is None:
+        """ "Identify arguments that impact huggingface transformers resources and behavior"""
+
+        # this may be a bit too naive as it will pass any parameter valid for the hf_constructor signature
+        # this falls over when passed some `from_pretrained` methods as the callable model params are not always explicit
+        params = (
+            self.hf_args
+            if hasattr(self, "hf_args") and isinstance(self.hf_args, dict)
+            else {}
+        )
+        if params is not None and not "device" in params and hasattr(self, "device"):
+            # consider setting self.device in all cases or if self.device is not found raise error `_select_hf_device` must be called
             params["device"] = self.device
 
         args = {}
@@ -96,6 +103,15 @@ def _gather_hf_params(self, hf_constructor: Callable):
                     continue
                 args[k] = params[k]
 
+        if (
+            not "device_map" in args
+            and "device_map" in params_to_process
+            and "device" in params_to_process
+            and "device" in args
+        ):
+            del args["device"]
+            args["device_map"] = self.device
+
         return args
 
     def _select_hf_device(self):