Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

feature: reinforce trust_remote_code=False #979

Merged
merged 4 commits into from
Nov 12, 2024
Merged

feature: reinforce trust_remote_code=False #979

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
fc27f68
strengthen protections around trust_remote_code
leondz Nov 11, 2024
191ccc9
inject paraphrase.PegasusT5 trust_remote_code
leondz Nov 12, 2024
c86a39f
handle case where trust_remote_code val not prepared for Pegasus model
leondz Nov 12, 2024
dd2258c
move trust_remote_code to pegasus configurable params
leondz Nov 12, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
7 changes: 5 additions & 2 deletions garak/buffs/paraphrase.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,8 @@ class PegasusT5(Buff, HFCompatible):
DEFAULT_PARAMS = Buff.DEFAULT_PARAMS | {
"para_model_name": "garak-llm/pegasus_paraphrase",
"hf_args": {
"device": "cpu"
"device": "cpu",
"trust_remote_code": False,
}, # torch_dtype doesn't have standard support in Pegasus
"max_length": 60,
"temperature": 1.5,
Expand All @@ -39,7 +40,9 @@ def _load_model(self):
self.para_model = PegasusForConditionalGeneration.from_pretrained(
self.para_model_name
).to(self.device)
self.tokenizer = PegasusTokenizer.from_pretrained(self.para_model_name)
self.tokenizer = PegasusTokenizer.from_pretrained(
self.para_model_name, trust_remote_code=self.hf_args["trust_remote_code"]
)

def _get_response(self, input_text):
if self.para_model is None:
Expand Down
6 changes: 1 addition & 5 deletions garak/generators/huggingface.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -436,15 +436,11 @@ def _load_client(self):
if _config.run.seed is not None:
transformers.set_seed(_config.run.seed)

trust_remote_code = self.name.startswith("mosaicml/mpt-")

model_kwargs = self._gather_hf_params(
hf_constructor=transformers.AutoConfig.from_pretrained
) # will defer to device_map if device map was `auto` may not match self.device

self.config = transformers.AutoConfig.from_pretrained(
self.name, trust_remote_code=trust_remote_code, **model_kwargs
)
self.config = transformers.AutoConfig.from_pretrained(self.name, **model_kwargs)

self._set_hf_context_len(self.config)
self.config.init_device = self.device # determined by Pipeline `__init__``
Expand Down
8 changes: 7 additions & 1 deletion garak/resources/api/huggingface.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,7 +9,6 @@


class HFCompatible:

"""Mixin class providing private utility methods for using Huggingface
transformers within garak"""

Expand Down Expand Up @@ -79,6 +78,13 @@ def _gather_hf_params(self, hf_constructor: Callable):
del args["device"]
args["device_map"] = self.device

# trust_remote_code reset to default disabled unless unlocked in garak HF item config
if (
"trust_remote_code" in params_to_process
and "trust_remote_code" not in params
):
args["trust_remote_code"] = False

return args

def _select_hf_device(self):
Expand Down
Loading