phoenix-oss/llama-stack
1
0
Fork 0
forked from phoenix-oss/llama-stack-mirror
Code Pull requests Releases Packages 2 Activity Actions

Introduce Llama stack distributions (#22)

Browse source 
* Add distribution CLI scaffolding

* More progress towards `llama distribution install`

* getting closer to a distro definition, distro install + configure works

* Distribution server now functioning

* read existing configuration, save enums properly

* Remove inference uvicorn server entrypoint and llama inference CLI command

* updated dependency and client model name

* Improved exception handling

* local imports for faster cli

* undo a typo, add a passthrough distribution

* implement full-passthrough in the server

* add safety adapters, configuration handling, server + clients

* cleanup, moving stuff to common, nuke utils

* Add a Path() wrapper at the earliest place

* fixes

* Bring agentic system api to toolchain

Add adapter dependencies and resolve adapters using a topological sort

* refactor to reduce size of `agentic_system`

* move straggler files and fix some important existing bugs

* ApiSurface -> Api

* refactor a method out

* Adapter -> Provider

* Make each inference provider into its own subdirectory

* installation fixes

* Rename Distribution -> DistributionSpec, simplify RemoteProviders

* dict key instead of attr

* update inference config to take model and not model_dir

* Fix passthrough streaming, send headers properly not part of body :facepalm

* update safety to use model sku ids and not model dirs

* Update cli_reference.md

* minor fixes

* add DistributionConfig, fix a bug in model download

* Make install + start scripts do proper configuration automatically

* Update CLI_reference

* Nuke fp8_requirements, fold fbgemm into common requirements

* Update README, add newline between API surface configurations

* Refactor download functionality out of the Command so can be reused

* Add `llama model download` alias for `llama download`

* Show message about checksum file so users can check themselves

* Simpler intro statements

* get ollama working

* Reduce a bunch of dependencies from toolchain

Some improvements to the distribution install script

* Avoid using `conda run` since it buffers everything

* update dependencies and rely on LLAMA_TOOLCHAIN_DIR for dev purposes

* add validation for configuration input

* resort imports

* make optional subclasses default to yes for configuration

* Remove additional_pip_packages; move deps to providers

* for inline make 8b model the default

* Add scripts to MANIFEST

* allow installing from test.pypi.org

* Fix #2 to help with testing packages

* Must install llama-models at that same version first

* fix PIP_ARGS

---------

Co-authored-by: Hardik Shah <hjshah@fb.com>
Co-authored-by: Hardik Shah <hjshah@meta.com>
This commit is contained in:
Ashwin Bharambe 2024-08-08 13:38:41 -07:00 • committed by GitHub
parent da4645a27a
commit e830814399
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
115 changed files with 5839 additions and 1120 deletions
Download patch file Download diff file Expand all files Collapse all files

156
llama_toolchain/safety/meta_reference/shields/prompt_guard.py Normal file
View file

@ -0,0 +1,156 @@
# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from enum import auto, Enum
from typing import List
import torch
from llama_models.llama3_1.api.datatypes import Message
from termcolor import cprint
from transformers import AutoModelForSequenceClassification, AutoTokenizer
from .base import message_content_as_str, OnViolationAction, ShieldResponse, TextShield
from llama_toolchain.safety.api.datatypes import * # noqa: F403
class PromptGuardShield(TextShield):
class Mode(Enum):
INJECTION = auto()
JAILBREAK = auto()
_instances = {}
_model_cache = None
@staticmethod
def instance(
model_dir: str,
threshold: float = 0.9,
temperature: float = 1.0,
mode: "PromptGuardShield.Mode" = Mode.JAILBREAK,
on_violation_action=OnViolationAction.RAISE,
) -> "PromptGuardShield":
action_value = on_violation_action.value
key = (model_dir, threshold, temperature, mode, action_value)
if key not in PromptGuardShield._instances:
PromptGuardShield._instances[key] = PromptGuardShield(
model_dir=model_dir,
threshold=threshold,
temperature=temperature,
mode=mode,
on_violation_action=on_violation_action,
)
return PromptGuardShield._instances[key]
def __init__(
self,
model_dir: str,
threshold: float = 0.9,
temperature: float = 1.0,
mode: "PromptGuardShield.Mode" = Mode.JAILBREAK,
on_violation_action: OnViolationAction = OnViolationAction.RAISE,
):
super().__init__(on_violation_action)
assert (
model_dir is not None
), "Must provide a model directory for prompt injection shield"
if temperature <= 0:
raise ValueError("Temperature must be greater than 0")
self.device = "cuda"
if PromptGuardShield._model_cache is None:
# load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained(model_dir)
model = AutoModelForSequenceClassification.from_pretrained(
model_dir, device_map=self.device
)
PromptGuardShield._model_cache = (tokenizer, model)
self.tokenizer, self.model = PromptGuardShield._model_cache
self.temperature = temperature
self.threshold = threshold
self.mode = mode
def get_shield_type(self) -> ShieldType:
return (
BuiltinShield.jailbreak_shield
if self.mode == self.Mode.JAILBREAK
else BuiltinShield.injection_shield
)
def convert_messages_to_text(self, messages: List[Message]) -> str:
return message_content_as_str(messages[-1])
async def run_impl(self, text: str) -> ShieldResponse:
# run model on messages and return response
inputs = self.tokenizer(text, return_tensors="pt")
inputs = {name: tensor.to(self.model.device) for name, tensor in inputs.items()}
with torch.no_grad():
outputs = self.model(**inputs)
logits = outputs[0]
probabilities = torch.softmax(logits / self.temperature, dim=-1)
score_embedded = probabilities[0, 1].item()
score_malicious = probabilities[0, 2].item()
cprint(
f"Ran PromptGuardShield and got Scores: Embedded: {score_embedded}, Malicious: {score_malicious}",
color="magenta",
)
if self.mode == self.Mode.INJECTION and (
score_embedded + score_malicious > self.threshold
):
return ShieldResponse(
shield_type=self.get_shield_type(),
is_violation=True,
violation_type=f"prompt_injection:embedded={score_embedded},malicious={score_malicious}",
violation_return_message="Sorry, I cannot do this.",
)
elif self.mode == self.Mode.JAILBREAK and score_malicious > self.threshold:
return ShieldResponse(
shield_type=self.get_shield_type(),
is_violation=True,
violation_type=f"prompt_injection:malicious={score_malicious}",
violation_return_message="Sorry, I cannot do this.",
)
return ShieldResponse(
shield_type=self.get_shield_type(),
is_violation=False,
)
class JailbreakShield(PromptGuardShield):
def __init__(
self,
model_dir: str,
threshold: float = 0.9,
temperature: float = 1.0,
on_violation_action: OnViolationAction = OnViolationAction.RAISE,
):
super().__init__(
model_dir=model_dir,
threshold=threshold,
temperature=temperature,
mode=PromptGuardShield.Mode.JAILBREAK,
on_violation_action=on_violation_action,
)
class InjectionShield(PromptGuardShield):
def __init__(
self,
model_dir: str,
threshold: float = 0.9,
temperature: float = 1.0,
on_violation_action: OnViolationAction = OnViolationAction.RAISE,
):
super().__init__(
model_dir=model_dir,
threshold=threshold,
temperature=temperature,
mode=PromptGuardShield.Mode.INJECTION,
on_violation_action=on_violation_action,
)