phoenix-oss/llama-stack-mirror
0
0
Fork 1
mirror of https://github.com/meta-llama/llama-stack.git synced 2025-12-17 16:12:46 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'main' into add-nvidia-inference-adapter

Browse source
This commit is contained in:
Matthew Farrellee 2024-11-19 10:25:50 -05:00
commit 2a25ace2fa
131 changed files with 3927 additions and 1286 deletions
Download patch file Download diff file Expand all files Collapse all files

15
llama_stack/providers/inline/agents/meta_reference/config.py
View file

@ -4,11 +4,22 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from pydantic import BaseModel, Field
from typing import Any, Dict
from pydantic import BaseModel
from llama_stack.providers.utils.kvstore import KVStoreConfig
from llama_stack.providers.utils.kvstore.config import SqliteKVStoreConfig
class MetaReferenceAgentsImplConfig(BaseModel):
persistence_store: KVStoreConfig = Field(default=SqliteKVStoreConfig())
persistence_store: KVStoreConfig
@classmethod
def sample_run_config(cls, __distro_dir__: str) -> Dict[str, Any]:
return {
"persistence_store": SqliteKVStoreConfig.sample_run_config(
__distro_dir__=__distro_dir__,
db_name="agents_store.db",
)
}

1
llama_stack/providers/inline/eval/meta_reference/__init__.py
View file

@ -22,6 +22,7 @@ async def get_provider_impl(
deps[Api.datasets],
deps[Api.scoring],
deps[Api.inference],
deps[Api.agents],
)
await impl.initialize()
return impl

72
llama_stack/providers/inline/eval/meta_reference/eval.py
View file

@ -9,6 +9,7 @@ from llama_models.llama3.api.datatypes import * # noqa: F403
from .....apis.common.job_types import Job
from .....apis.eval.eval import Eval, EvalTaskConfig, EvaluateResponse, JobStatus
from llama_stack.apis.common.type_system import * # noqa: F403
from llama_stack.apis.agents import Agents
from llama_stack.apis.datasetio import DatasetIO
from llama_stack.apis.datasets import Datasets
from llama_stack.apis.eval_tasks import EvalTask
@ -39,12 +40,14 @@ class MetaReferenceEvalImpl(Eval, EvalTasksProtocolPrivate):
datasets_api: Datasets,
scoring_api: Scoring,
inference_api: Inference,
agents_api: Agents,
) -> None:
self.config = config
self.datasetio_api = datasetio_api
self.datasets_api = datasets_api
self.scoring_api = scoring_api
self.inference_api = inference_api
self.agents_api = agents_api
# TODO: assume sync job, will need jobs API for async scheduling
self.jobs = {}
@ -126,18 +129,50 @@ class MetaReferenceEvalImpl(Eval, EvalTasksProtocolPrivate):
self.jobs[job_id] = res
return Job(job_id=job_id)
async def evaluate_rows(
self,
task_id: str,
input_rows: List[Dict[str, Any]],
scoring_functions: List[str],
task_config: EvalTaskConfig,
) -> EvaluateResponse:
async def _run_agent_generation(
self, input_rows: List[Dict[str, Any]], task_config: EvalTaskConfig
) -> List[Dict[str, Any]]:
candidate = task_config.eval_candidate
if candidate.type == "agent":
raise NotImplementedError(
"Evaluation with generation has not been implemented for agents"
create_response = await self.agents_api.create_agent(candidate.config)
agent_id = create_response.agent_id
generations = []
for i, x in tqdm(enumerate(input_rows)):
assert ColumnName.chat_completion_input.value in x, "Invalid input row"
input_messages = eval(str(x[ColumnName.chat_completion_input.value]))
input_messages = [UserMessage(**x) for x in input_messages]
# NOTE: only single-turn agent generation is supported. Create a new session for each input row
session_create_response = await self.agents_api.create_agent_session(
agent_id, f"session-{i}"
)
session_id = session_create_response.session_id
turn_request = dict(
agent_id=agent_id,
session_id=session_id,
messages=input_messages,
stream=True,
)
turn_response = [
chunk
async for chunk in await self.agents_api.create_agent_turn(
**turn_request
)
]
final_event = turn_response[-1].event.payload
generations.append(
{
ColumnName.generated_answer.value: final_event.turn.output_message.content
}
)
return generations
async def _run_model_generation(
self, input_rows: List[Dict[str, Any]], task_config: EvalTaskConfig
) -> List[Dict[str, Any]]:
candidate = task_config.eval_candidate
assert (
candidate.sampling_params.max_tokens is not None
), "SamplingParams.max_tokens must be provided"
@ -179,6 +214,23 @@ class MetaReferenceEvalImpl(Eval, EvalTasksProtocolPrivate):
else:
raise ValueError("Invalid input row")
return generations
async def evaluate_rows(
self,
task_id: str,
input_rows: List[Dict[str, Any]],
scoring_functions: List[str],
task_config: EvalTaskConfig,
) -> EvaluateResponse:
candidate = task_config.eval_candidate
if candidate.type == "agent":
generations = await self._run_agent_generation(input_rows, task_config)
elif candidate.type == "model":
generations = await self._run_model_generation(input_rows, task_config)
else:
raise ValueError(f"Invalid candidate type: {candidate.type}")
# scoring with generated_answer
score_input_rows = [
input_r | generated_r

12
llama_stack/providers/inline/inference/meta_reference/config.py
View file

@ -49,6 +49,18 @@ class MetaReferenceInferenceConfig(BaseModel):
resolved = resolve_model(self.model)
return resolved.pth_file_count
@classmethod
def sample_run_config(
cls,
model: str = "Llama3.2-3B-Instruct",
checkpoint_dir: str = "${env.CHECKPOINT_DIR:null}",
) -> Dict[str, Any]:
return {
"model": model,
"max_seq_len": 4096,
"checkpoint_dir": checkpoint_dir,
}
class MetaReferenceQuantizedInferenceConfig(MetaReferenceInferenceConfig):
quantization: QuantizationConfig

3
llama_stack/providers/inline/inference/meta_reference/generation.py
View file

@ -107,7 +107,7 @@ class Llama:
sys.stdout = open(os.devnull, "w")
start_time = time.time()
if config.checkpoint_dir:
if config.checkpoint_dir and config.checkpoint_dir != "null":
ckpt_dir = config.checkpoint_dir
else:
ckpt_dir = model_checkpoint_dir(model)
@ -137,7 +137,6 @@ class Llama:
), f"model_args vocab = {model_args.vocab_size} but tokenizer vocab = {tokenizer.n_words}"
if isinstance(config, MetaReferenceQuantizedInferenceConfig):
if isinstance(config.quantization, Fp8QuantizationConfig):
from .quantization.loader import convert_to_fp8_quantized_model

10
llama_stack/providers/inline/inference/vllm/config.py
View file

@ -34,6 +34,16 @@ class VLLMConfig(BaseModel):
default=0.3,
)
@classmethod
def sample_run_config(cls):
return {
"model": "${env.VLLM_INFERENCE_MODEL:Llama3.2-3B-Instruct}",
"tensor_parallel_size": "${env.VLLM_TENSOR_PARALLEL_SIZE:1}",
"max_tokens": "${env.VLLM_MAX_TOKENS:4096}",
"enforce_eager": "${env.VLLM_ENFORCE_EAGER:False}",
"gpu_memory_utilization": "${env.VLLM_GPU_MEMORY_UTILIZATION:0.3}",
}
@field_validator("model")
@classmethod
def validate_model(cls, model: str) -> str:

16
llama_stack/providers/inline/memory/faiss/config.py
View file

@ -4,10 +4,11 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Dict
from llama_models.schema_utils import json_schema_type
from pydantic import BaseModel
from llama_stack.distribution.utils.config_dirs import RUNTIME_BASE_DIR
from llama_stack.providers.utils.kvstore.config import (
KVStoreConfig,
SqliteKVStoreConfig,
@ -16,6 +17,13 @@ from llama_stack.providers.utils.kvstore.config import (
@json_schema_type
class FaissImplConfig(BaseModel):
kvstore: KVStoreConfig = SqliteKVStoreConfig(
db_path=(RUNTIME_BASE_DIR / "faiss_store.db").as_posix()
) # Uses SQLite config specific to FAISS storage
kvstore: KVStoreConfig
@classmethod
def sample_run_config(cls, __distro_dir__: str) -> Dict[str, Any]:
return {
"kvstore": SqliteKVStoreConfig.sample_run_config(
__distro_dir__=__distro_dir__,
db_name="faiss_store.db",
)
}

26
llama_stack/providers/inline/safety/llama_guard/llama_guard.py
View file

@ -73,18 +73,21 @@ DEFAULT_LG_V3_SAFETY_CATEGORIES = [
CAT_ELECTIONS,
]
LLAMA_GUARD_MODEL_IDS = [
CoreModelId.llama_guard_3_8b.value,
CoreModelId.llama_guard_3_1b.value,
CoreModelId.llama_guard_3_11b_vision.value,
]
# accept both CoreModelId and huggingface repo id
LLAMA_GUARD_MODEL_IDS = {
CoreModelId.llama_guard_3_8b.value: "meta-llama/Llama-Guard-3-8B",
"meta-llama/Llama-Guard-3-8B": "meta-llama/Llama-Guard-3-8B",
CoreModelId.llama_guard_3_1b.value: "meta-llama/Llama-Guard-3-1B",
"meta-llama/Llama-Guard-3-1B": "meta-llama/Llama-Guard-3-1B",
CoreModelId.llama_guard_3_11b_vision.value: "meta-llama/Llama-Guard-3-11B-Vision",
"meta-llama/Llama-Guard-3-11B-Vision": "meta-llama/Llama-Guard-3-11B-Vision",
}
MODEL_TO_SAFETY_CATEGORIES_MAP = {
CoreModelId.llama_guard_3_8b.value: (
DEFAULT_LG_V3_SAFETY_CATEGORIES + [CAT_CODE_INTERPRETER_ABUSE]
),
CoreModelId.llama_guard_3_1b.value: DEFAULT_LG_V3_SAFETY_CATEGORIES,
CoreModelId.llama_guard_3_11b_vision.value: DEFAULT_LG_V3_SAFETY_CATEGORIES,
"meta-llama/Llama-Guard-3-8B": DEFAULT_LG_V3_SAFETY_CATEGORIES
+ [CAT_CODE_INTERPRETER_ABUSE],
"meta-llama/Llama-Guard-3-1B": DEFAULT_LG_V3_SAFETY_CATEGORIES,
"meta-llama/Llama-Guard-3-11B-Vision": DEFAULT_LG_V3_SAFETY_CATEGORIES,
}
@ -150,8 +153,9 @@ class LlamaGuardSafetyImpl(Safety, ShieldsProtocolPrivate):
if len(messages) > 0 and messages[0].role != Role.user.value:
messages[0] = UserMessage(content=messages[0].content)
model = LLAMA_GUARD_MODEL_IDS[shield.provider_resource_id]
impl = LlamaGuardShield(
model=shield.provider_resource_id,
model=model,
inference_api=self.inference_api,
excluded_categories=self.config.excluded_categories,
)

91
llama_stack/providers/inline/scoring/llm_as_judge/scoring_fn/fn_defs/llm_as_judge_405b_simpleqa.py Normal file
View file

@ -0,0 +1,91 @@
# 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 llama_stack.apis.common.type_system import NumberType
from llama_stack.apis.scoring_functions import LLMAsJudgeScoringFnParams, ScoringFn
GRADER_TEMPLATE = """
Your job is to look at a question, a gold target, and a predicted answer, and then assign a grade of either ["CORRECT", "INCORRECT", "NOT_ATTEMPTED"].
First, I will give examples of each grade, and then you will grade a new example.
The following are examples of CORRECT predicted answers.
```
Question: What are the names of Barack Obama's children?
Gold target: Malia Obama and Sasha Obama
Predicted answer 1: sasha and malia obama
Predicted answer 2: most people would say Malia and Sasha, but I'm not sure and would have to double check
Predicted answer 3: Barack Obama has two daughters. Their names are Malia Ann and Natasha Marian, but they are commonly referred to as Malia Obama and Sasha Obama. Malia was born on July 4, 1998, and Sasha was born on June 10, 2001.
```
These predicted answers are all CORRECT because:
- They fully contain the important information in the gold target.
- They do not contain any information that contradicts the gold target.
- Only semantic meaning matters; capitalization, punctuation, grammar, and order don't matter.
- Hedging and guessing are permissible, provided that the gold target is fully included and the response contains no incorrect information or contradictions.
The following are examples of INCORRECT predicted answers.
```
Question: What are the names of Barack Obama's children?
Gold target: Malia and Sasha
Predicted answer 1: Malia.
Predicted answer 2: Malia, Sasha, and Susan.
Predicted answer 3: Barack Obama does not have any children.
Predicted answer 4: I think it's either Malia and Sasha. Or it could be Malia and Jackie. Or it could be Joey and Malia.
Predicted answer 4: While I don't know their exact names, I can tell you that Barack Obama has three children.
Predicted answer 5: It's possible you may mean Betsy and Olivia. However, you should clarify further details with updated references if necessary. Is that the correct answer?
Predicted answer 6: It may be the case that Obama's child is named James. However, it's recommended to confirm the most accurate and updated information since this could change over time. This model may not always reflect the most current information.
```
These predicted answers are all INCORRECT because:
- A factual statement in the answer contradicts the gold target. Incorrect statements that have some hedging (e.g., "it is possible that", "although i'm not sure, i think") are also considered incorrect.
The following are examples of NOT_ATTEMPTED predicted answers.
```
Question: What are the names of Barack Obama's children?
Gold target: Malia and Sasha
Predicted answer 1: I don't know.
Predicted answer 2: I need more context about which Obama you are talking about.
Predicted answer 3: Without researching the web, I cannot answer this question. However, I can tell you that Barack Obama has two children.
Predicted answer 4: Barack Obama has two children. I know that one of them is Malia, but I'm not sure about the other one.
```
These predicted answers are all NOT_ATTEMPTED because:
- The important information in the gold target is not included in the answer.
- No statements in the answer contradict the gold target.
Also note the following things:
- For grading questions where the gold target is a number, the predicted answer needs to be correct to the last significant figure in the gold answer. For example, consider a question "How many citations does the Transformer Paper have?" with gold target "120k".
- Predicted answers "120k", "124k", and 115k" are all CORRECT.
- Predicted answers "100k" and "113k" are INCORRECT.
- Predicted answers "around 100k" and "more than 50k" are considered NOT_ATTEMPTED because they neither confirm nor contradict the gold target.
- The gold target may contain more information than the question. In such cases, the predicted answer only needs to contain the information that is in the question.
- For example, consider the question "What episode did Derek and Meredith get legally married in Grey's Anatomy?" with gold target "Season 7, Episode 20: White Wedding". Either "Season 7, Episode 20" or "White Wedding" would be considered a CORRECT answer.
- Do not punish predicted answers if they omit information that would be clearly inferred from the question.
- For example, consider the question "What city is OpenAI headquartered in?" and the gold target "San Francisco, California". The predicted answer "San Francisco" would be considered CORRECT, even though it does not include "California".
- Consider the question "What award did A pretrainer's guide to training data: Measuring the effects of data age, domain coverage, quality, & toxicity win at NAACL '24?", the gold target is "Outstanding Paper Award". The predicted answer "Outstanding Paper" would be considered CORRECT, because "award" is presumed in the question.
- For the question "What is the height of Jason Wei in meters?", the gold target is "1.73 m". The predicted answer "1.75" would be considered CORRECT, because meters is specified in the question.
- For the question "What is the name of Barack Obama's wife?", the gold target is "Michelle Obama". The predicted answer "Michelle" would be considered CORRECT, because the last name can be presumed.
- Do not punish for typos in people's name if it's clearly the same name.
- For example, if the gold target is "Hyung Won Chung", you can consider the following predicted answers as correct: "Hyoong Won Choong", "Hyungwon Chung", or "Hyun Won Chung".
Here is a new example. Simply reply with either CORRECT, INCORRECT, NOT ATTEMPTED. Don't apologize or correct yourself if there was a mistake; we are just trying to grade the answer.
```
Question: {input_query}
Gold target: {expected_answer}
Predicted answer: {generated_answer}
```
Grade the predicted answer of this new question as one of:
A: CORRECT
B: INCORRECT
C: NOT_ATTEMPTED
Just return the letters "A", "B", or "C", with no text around it.
""".strip()
llm_as_judge_405b_simpleqa = ScoringFn(
identifier="llm-as-judge::405b-simpleqa",
description="Llm As Judge Scoring Function for SimpleQA Benchmark (https://github.com/openai/simple-evals/blob/main/simpleqa_eval.py)",
return_type=NumberType(),
provider_id="llm-as-judge",
provider_resource_id="llm-as-judge-405b-simpleqa",
params=LLMAsJudgeScoringFnParams(
judge_model="Llama3.1-405B-Instruct",
prompt_template=GRADER_TEMPLATE,
judge_score_regexes=[r"(A|B|C)"],
),
)

2
llama_stack/providers/inline/scoring/llm_as_judge/scoring_fn/fn_defs/llm_as_judge_base.py
View file

@ -9,7 +9,7 @@ from llama_stack.apis.scoring_functions import ScoringFn
llm_as_judge_base = ScoringFn(
identifier="llm-as-judge::llm_as_judge_base",
identifier="llm-as-judge::base",
description="Llm As Judge Scoring Function",
return_type=NumberType(),
provider_id="llm-as-judge",

3
llama_stack/providers/inline/scoring/llm_as_judge/scoring_fn/llm_as_judge_scoring_fn.py
View file

@ -11,6 +11,8 @@ from llama_stack.apis.scoring import * # noqa: F401, F403
from llama_stack.apis.common.type_system import * # noqa: F403
import re
from .fn_defs.llm_as_judge_405b_simpleqa import llm_as_judge_405b_simpleqa
from .fn_defs.llm_as_judge_base import llm_as_judge_base
@ -24,6 +26,7 @@ class LlmAsJudgeScoringFn(BaseScoringFn):
self.inference_api = inference_api
self.supported_fn_defs_registry = {
llm_as_judge_base.identifier: llm_as_judge_base,
llm_as_judge_405b_simpleqa.identifier: llm_as_judge_405b_simpleqa,
}
async def score_row(

1
llama_stack/providers/registry/eval.py
View file

@ -22,6 +22,7 @@ def available_providers() -> List[ProviderSpec]:
Api.datasets,
Api.scoring,
Api.inference,
Api.agents,
],
),
]

9
llama_stack/providers/remote/inference/fireworks/config.py
View file

@ -4,7 +4,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Optional
from typing import Any, Dict, Optional
from llama_models.schema_utils import json_schema_type
from pydantic import BaseModel, Field
@ -20,3 +20,10 @@ class FireworksImplConfig(BaseModel):
default=None,
description="The Fireworks.ai API Key",
)
@classmethod
def sample_run_config(cls) -> Dict[str, Any]:
return {
"url": "https://api.fireworks.ai/inference",
"api_key": "${env.FIREWORKS_API_KEY}",
}

4
llama_stack/providers/remote/inference/fireworks/fireworks.py
View file

@ -35,7 +35,7 @@ from llama_stack.providers.utils.inference.prompt_adapter import (
from .config import FireworksImplConfig
model_aliases = [
MODEL_ALIASES = [
build_model_alias(
"fireworks/llama-v3p1-8b-instruct",
CoreModelId.llama3_1_8b_instruct.value,
@ -79,7 +79,7 @@ class FireworksInferenceAdapter(
ModelRegistryHelper, Inference, NeedsRequestProviderData
):
def __init__(self, config: FireworksImplConfig) -> None:
ModelRegistryHelper.__init__(self, model_aliases)
ModelRegistryHelper.__init__(self, MODEL_ALIASES)
self.config = config
self.formatter = ChatFormat(Tokenizer.get_instance())

20
llama_stack/providers/remote/inference/nvidia/_nvidia.py
View file

@ -30,7 +30,7 @@ from llama_stack.apis.inference import (
ResponseFormat,
)
from llama_stack.providers.utils.inference.model_registry import (
build_model_alias,
build_model_alias_with_just_provider_model_id,
ModelRegistryHelper,
)
@ -43,39 +43,39 @@ from ._openai_utils import (
from ._utils import check_health
_MODEL_ALIASES = [
build_model_alias(
build_model_alias_with_just_provider_model_id(
"meta/llama3-8b-instruct",
CoreModelId.llama3_8b_instruct.value,
),
build_model_alias(
build_model_alias_with_just_provider_model_id(
"meta/llama3-70b-instruct",
CoreModelId.llama3_70b_instruct.value,
),
build_model_alias(
build_model_alias_with_just_provider_model_id(
"meta/llama-3.1-8b-instruct",
CoreModelId.llama3_1_8b_instruct.value,
),
build_model_alias(
build_model_alias_with_just_provider_model_id(
"meta/llama-3.1-70b-instruct",
CoreModelId.llama3_1_70b_instruct.value,
),
build_model_alias(
build_model_alias_with_just_provider_model_id(
"meta/llama-3.1-405b-instruct",
CoreModelId.llama3_1_405b_instruct.value,
),
build_model_alias(
build_model_alias_with_just_provider_model_id(
"meta/llama-3.2-1b-instruct",
CoreModelId.llama3_2_1b_instruct.value,
),
build_model_alias(
build_model_alias_with_just_provider_model_id(
"meta/llama-3.2-3b-instruct",
CoreModelId.llama3_2_3b_instruct.value,
),
build_model_alias(
build_model_alias_with_just_provider_model_id(
"meta/llama-3.2-11b-vision-instruct",
CoreModelId.llama3_2_11b_vision_instruct.value,
),
build_model_alias(
build_model_alias_with_just_provider_model_id(
"meta/llama-3.2-90b-vision-instruct",
CoreModelId.llama3_2_90b_vision_instruct.value,
),

8
llama_stack/providers/remote/inference/ollama/__init__.py
View file

@ -4,14 +4,10 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from llama_stack.distribution.datatypes import RemoteProviderConfig
from .config import OllamaImplConfig
class OllamaImplConfig(RemoteProviderConfig):
port: int = 11434
async def get_adapter_impl(config: RemoteProviderConfig, _deps):
async def get_adapter_impl(config: OllamaImplConfig, _deps):
from .ollama import OllamaInferenceAdapter
impl = OllamaInferenceAdapter(config.url)

22
llama_stack/providers/remote/inference/ollama/config.py Normal file
View file

@ -0,0 +1,22 @@
# 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 typing import Any, Dict
from pydantic import BaseModel
DEFAULT_OLLAMA_URL = "http://localhost:11434"
class OllamaImplConfig(BaseModel):
url: str = DEFAULT_OLLAMA_URL
@classmethod
def sample_run_config(
cls, url: str = "${env.OLLAMA_URL:http://localhost:11434}", **kwargs
) -> Dict[str, Any]:
return {"url": url}

33
llama_stack/providers/remote/inference/ollama/ollama.py
View file

@ -16,6 +16,7 @@ from ollama import AsyncClient
from llama_stack.providers.utils.inference.model_registry import (
build_model_alias,
build_model_alias_with_just_provider_model_id,
ModelRegistryHelper,
)
@ -44,10 +45,18 @@ model_aliases = [
"llama3.1:8b-instruct-fp16",
CoreModelId.llama3_1_8b_instruct.value,
),
build_model_alias_with_just_provider_model_id(
"llama3.1:8b",
CoreModelId.llama3_1_8b_instruct.value,
),
build_model_alias(
"llama3.1:70b-instruct-fp16",
CoreModelId.llama3_1_70b_instruct.value,
),
build_model_alias_with_just_provider_model_id(
"llama3.1:70b",
CoreModelId.llama3_1_70b_instruct.value,
),
build_model_alias(
"llama3.2:1b-instruct-fp16",
CoreModelId.llama3_2_1b_instruct.value,
@ -56,6 +65,24 @@ model_aliases = [
"llama3.2:3b-instruct-fp16",
CoreModelId.llama3_2_3b_instruct.value,
),
build_model_alias_with_just_provider_model_id(
"llama3.2:1b",
CoreModelId.llama3_2_1b_instruct.value,
),
build_model_alias_with_just_provider_model_id(
"llama3.2:3b",
CoreModelId.llama3_2_3b_instruct.value,
),
build_model_alias(
"llama3.2-vision:11b-instruct-fp16",
CoreModelId.llama3_2_11b_vision_instruct.value,
),
build_model_alias_with_just_provider_model_id(
"llama3.2-vision",
CoreModelId.llama3_2_11b_vision_instruct.value,
),
# The Llama Guard models don't have their full fp16 versions
# so we are going to alias their default version to the canonical SKU
build_model_alias(
"llama-guard3:8b",
CoreModelId.llama_guard_3_8b.value,
@ -64,10 +91,6 @@ model_aliases = [
"llama-guard3:1b",
CoreModelId.llama_guard_3_1b.value,
),
build_model_alias(
"x/llama3.2-vision:11b-instruct-fp16",
CoreModelId.llama3_2_11b_vision_instruct.value,
),
]
@ -82,7 +105,7 @@ class OllamaInferenceAdapter(Inference, ModelsProtocolPrivate):
return AsyncClient(host=self.url)
async def initialize(self) -> None:
print("Initializing Ollama, checking connectivity to server...")
print(f"checking connectivity to Ollama at `{self.url}`...")
try:
await self.client.ps()
except httpx.ConnectError as e:

17
llama_stack/providers/remote/inference/tgi/config.py
View file

@ -12,19 +12,20 @@ from pydantic import BaseModel, Field
@json_schema_type
class TGIImplConfig(BaseModel):
host: str = "localhost"
port: int = 8080
protocol: str = "http"
@property
def url(self) -> str:
return f"{self.protocol}://{self.host}:{self.port}"
url: str = Field(
description="The URL for the TGI serving endpoint",
)
api_token: Optional[str] = Field(
default=None,
description="A bearer token if your TGI endpoint is protected.",
)
@classmethod
def sample_run_config(cls, url: str = "${env.TGI_URL}", **kwargs):
return {
"url": url,
}
@json_schema_type
class InferenceEndpointImplConfig(BaseModel):

9
llama_stack/providers/remote/inference/together/config.py
View file

@ -4,7 +4,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Optional
from typing import Any, Dict, Optional
from llama_models.schema_utils import json_schema_type
from pydantic import BaseModel, Field
@ -20,3 +20,10 @@ class TogetherImplConfig(BaseModel):
default=None,
description="The Together AI API Key",
)
@classmethod
def sample_run_config(cls, **kwargs) -> Dict[str, Any]:
return {
"url": "https://api.together.xyz/v1",
"api_key": "${env.TOGETHER_API_KEY}",
}

4
llama_stack/providers/remote/inference/together/together.py
View file

@ -38,7 +38,7 @@ from llama_stack.providers.utils.inference.prompt_adapter import (
from .config import TogetherImplConfig
model_aliases = [
MODEL_ALIASES = [
build_model_alias(
"meta-llama/Meta-Llama-3.1-8B-Instruct-Turbo",
CoreModelId.llama3_1_8b_instruct.value,
@ -78,7 +78,7 @@ class TogetherInferenceAdapter(
ModelRegistryHelper, Inference, NeedsRequestProviderData
):
def __init__(self, config: TogetherImplConfig) -> None:
ModelRegistryHelper.__init__(self, model_aliases)
ModelRegistryHelper.__init__(self, MODEL_ALIASES)
self.config = config
self.formatter = ChatFormat(Tokenizer.get_instance())

12
llama_stack/providers/remote/inference/vllm/config.py
View file

@ -24,3 +24,15 @@ class VLLMInferenceAdapterConfig(BaseModel):
default="fake",
description="The API token",
)
@classmethod
def sample_run_config(
cls,
url: str = "${env.VLLM_URL}",
**kwargs,
):
return {
"url": url,
"max_tokens": "${env.VLLM_MAX_TOKENS:4096}",
"api_token": "${env.VLLM_API_TOKEN:fake}",
}

2
llama_stack/providers/tests/README.md
View file

@ -44,7 +44,7 @@ Finally, you can override the model completely by doing:
```bash
pytest -s -v llama_stack/providers/tests/inference/test_text_inference.py \
-m fireworks \
--inference-model "Llama3.1-70B-Instruct" \
--inference-model "meta-llama/Llama3.1-70B-Instruct" \
--env FIREWORKS_API_KEY=<...>
```

4
llama_stack/providers/tests/agents/conftest.py
View file

@ -81,13 +81,13 @@ def pytest_addoption(parser):
parser.addoption(
"--inference-model",
action="store",
default="Llama3.1-8B-Instruct",
default="meta-llama/Llama-3.1-8B-Instruct",
help="Specify the inference model to use for testing",
)
parser.addoption(
"--safety-shield",
action="store",
default="Llama-Guard-3-8B",
default="meta-llama/Llama-Guard-3-8B",
help="Specify the safety shield to use for testing",
)

2
llama_stack/providers/tests/agents/fixtures.py
View file

@ -83,6 +83,6 @@ async def agents_stack(request, inference_model, safety_shield):
)
for model in inference_models
],
shields=[safety_shield],
shields=[safety_shield] if safety_shield else [],
)
return test_stack

2
llama_stack/providers/tests/eval/conftest.py
View file

@ -63,7 +63,7 @@ def pytest_addoption(parser):
parser.addoption(
"--inference-model",
action="store",
default="Llama3.2-3B-Instruct",
default="meta-llama/Llama-3.2-3B-Instruct",
help="Specify the inference model to use for testing",
)

8
llama_stack/providers/tests/inference/conftest.py
View file

@ -32,8 +32,12 @@ def pytest_configure(config):
MODEL_PARAMS = [
pytest.param("Llama3.1-8B-Instruct", marks=pytest.mark.llama_8b, id="llama_8b"),
pytest.param("Llama3.2-3B-Instruct", marks=pytest.mark.llama_3b, id="llama_3b"),
pytest.param(
"meta-llama/Llama-3.1-8B-Instruct", marks=pytest.mark.llama_8b, id="llama_8b"
),
pytest.param(
"meta-llama/Llama-3.2-3B-Instruct", marks=pytest.mark.llama_3b, id="llama_3b"
),
]
VISION_MODEL_PARAMS = [

50
llama_stack/providers/tests/inference/test_model_registration.py
View file

@ -6,7 +6,6 @@
import pytest
from llama_models.datatypes import CoreModelId
# How to run this test:
#
@ -17,11 +16,22 @@ from llama_models.datatypes import CoreModelId
class TestModelRegistration:
@pytest.mark.asyncio
async def test_register_unsupported_model(self, inference_stack):
_, models_impl = inference_stack
async def test_register_unsupported_model(self, inference_stack, inference_model):
inference_impl, models_impl = inference_stack
provider = inference_impl.routing_table.get_provider_impl(inference_model)
if provider.__provider_spec__.provider_type not in (
"meta-reference",
"remote::ollama",
"remote::vllm",
"remote::tgi",
):
pytest.skip(
"Skipping test for remote inference providers since they can handle large models like 70B instruct"
)
# Try to register a model that's too large for local inference
with pytest.raises(Exception) as exc_info:
with pytest.raises(ValueError) as exc_info:
await models_impl.register_model(
model_id="Llama3.1-70B-Instruct",
)
@ -37,21 +47,27 @@ class TestModelRegistration:
)
@pytest.mark.asyncio
async def test_update_model(self, inference_stack):
async def test_register_with_llama_model(self, inference_stack):
_, models_impl = inference_stack
# Register a model to update
model_id = CoreModelId.llama3_1_8b_instruct.value
old_model = await models_impl.register_model(model_id=model_id)
# Update the model
new_model_id = CoreModelId.llama3_2_3b_instruct.value
updated_model = await models_impl.update_model(
model_id=model_id, provider_model_id=new_model_id
_ = await models_impl.register_model(
model_id="custom-model",
metadata={"llama_model": "meta-llama/Llama-2-7b"},
)
# Retrieve the updated model to verify changes
assert updated_model.provider_resource_id != old_model.provider_resource_id
with pytest.raises(ValueError) as exc_info:
await models_impl.register_model(
model_id="custom-model-2",
metadata={"llama_model": "meta-llama/Llama-2-7b"},
provider_model_id="custom-model",
)
# Cleanup
await models_impl.unregister_model(model_id=model_id)
@pytest.mark.asyncio
async def test_register_with_invalid_llama_model(self, inference_stack):
_, models_impl = inference_stack
with pytest.raises(ValueError) as exc_info:
await models_impl.register_model(
model_id="custom-model-2",
metadata={"llama_model": "invalid-llama-model"},
)

2
llama_stack/providers/tests/resolver.py
View file

@ -6,7 +6,6 @@
import json
import tempfile
from datetime import datetime
from typing import Any, Dict, List, Optional
from llama_stack.distribution.datatypes import * # noqa: F403
@ -37,7 +36,6 @@ async def construct_stack_for_test(
) -> TestStack:
sqlite_file = tempfile.NamedTemporaryFile(delete=False, suffix=".db")
run_config = dict(
built_at=datetime.now(),
image_name="test-fixture",
apis=apis,
providers=providers,

3
llama_stack/providers/tests/safety/fixtures.py
View file

@ -47,6 +47,9 @@ def safety_shield(request):
else:
params = {}
if not shield_id:
return None
return ShieldInput(
shield_id=shield_id,
params=params,

2
llama_stack/providers/tests/scoring/conftest.py
View file

@ -58,7 +58,7 @@ def pytest_addoption(parser):
parser.addoption(
"--inference-model",
action="store",
default="Llama3.2-3B-Instruct",
default="meta-llama/Llama-3.2-3B-Instruct",
help="Specify the inference model to use for testing",
)

5
llama_stack/providers/utils/inference/__init__.py
View file

@ -31,3 +31,8 @@ def supported_inference_models() -> List[str]:
or is_supported_safety_model(m)
)
]
ALL_HUGGINGFACE_REPOS_TO_MODEL_DESCRIPTOR = {
m.huggingface_repo: m.descriptor() for m in all_registered_models()
}

48
llama_stack/providers/utils/inference/model_registry.py
View file

@ -11,6 +11,10 @@ from llama_models.sku_list import all_registered_models
from llama_stack.providers.datatypes import Model, ModelsProtocolPrivate
from llama_stack.providers.utils.inference import (
ALL_HUGGINGFACE_REPOS_TO_MODEL_DESCRIPTOR,
)
ModelAlias = namedtuple("ModelAlias", ["provider_model_id", "aliases", "llama_model"])
@ -32,6 +36,16 @@ def build_model_alias(provider_model_id: str, model_descriptor: str) -> ModelAli
)
def build_model_alias_with_just_provider_model_id(
provider_model_id: str, model_descriptor: str
) -> ModelAlias:
return ModelAlias(
provider_model_id=provider_model_id,
aliases=[],
llama_model=model_descriptor,
)
class ModelRegistryHelper(ModelsProtocolPrivate):
def __init__(self, model_aliases: List[ModelAlias]):
self.alias_to_provider_id_map = {}
@ -51,7 +65,7 @@ class ModelRegistryHelper(ModelsProtocolPrivate):
if identifier in self.alias_to_provider_id_map:
return self.alias_to_provider_id_map[identifier]
else:
raise ValueError(f"Unknown model: `{identifier}`")
return None
def get_llama_model(self, provider_model_id: str) -> str:
if provider_model_id in self.provider_id_to_llama_model_map:
@ -60,8 +74,34 @@ class ModelRegistryHelper(ModelsProtocolPrivate):
return None
async def register_model(self, model: Model) -> Model:
model.provider_resource_id = self.get_provider_model_id(
model.provider_resource_id
)
provider_resource_id = self.get_provider_model_id(model.provider_resource_id)
if provider_resource_id:
model.provider_resource_id = provider_resource_id
else:
if model.metadata.get("llama_model") is None:
raise ValueError(
f"Model '{model.provider_resource_id}' is not available and no llama_model was specified in metadata. "
"Please specify a llama_model in metadata or use a supported model identifier"
)
existing_llama_model = self.get_llama_model(model.provider_resource_id)
if existing_llama_model:
if existing_llama_model != model.metadata["llama_model"]:
raise ValueError(
f"Provider model id '{model.provider_resource_id}' is already registered to a different llama model: '{existing_llama_model}'"
)
else:
if (
model.metadata["llama_model"]
not in ALL_HUGGINGFACE_REPOS_TO_MODEL_DESCRIPTOR
):
raise ValueError(
f"Invalid llama_model '{model.metadata['llama_model']}' specified in metadata. "
f"Must be one of: {', '.join(ALL_HUGGINGFACE_REPOS_TO_MODEL_DESCRIPTOR.keys())}"
)
self.provider_id_to_llama_model_map[model.provider_resource_id] = (
ALL_HUGGINGFACE_REPOS_TO_MODEL_DESCRIPTOR[
model.metadata["llama_model"]
]
)
return model

35
llama_stack/providers/utils/kvstore/config.py
View file

@ -36,6 +36,15 @@ class RedisKVStoreConfig(CommonConfig):
def url(self) -> str:
return f"redis://{self.host}:{self.port}"
@classmethod
def sample_run_config(cls):
return {
"type": "redis",
"namespace": None,
"host": "${env.REDIS_HOST:localhost}",
"port": "${env.REDIS_PORT:6379}",
}
class SqliteKVStoreConfig(CommonConfig):
type: Literal[KVStoreType.sqlite.value] = KVStoreType.sqlite.value
@ -44,6 +53,19 @@ class SqliteKVStoreConfig(CommonConfig):
description="File path for the sqlite database",
)
@classmethod
def sample_run_config(
cls, __distro_dir__: str = "runtime", db_name: str = "kvstore.db"
):
return {
"type": "sqlite",
"namespace": None,
"db_path": "${env.SQLITE_STORE_DIR:~/.llama/"
+ __distro_dir__
+ "}/"
+ db_name,
}
class PostgresKVStoreConfig(CommonConfig):
type: Literal[KVStoreType.postgres.value] = KVStoreType.postgres.value
@ -54,6 +76,19 @@ class PostgresKVStoreConfig(CommonConfig):
password: Optional[str] = None
table_name: str = "llamastack_kvstore"
@classmethod
def sample_run_config(cls, table_name: str = "llamastack_kvstore"):
return {
"type": "postgres",
"namespace": None,
"host": "${env.POSTGRES_HOST:localhost}",
"port": "${env.POSTGRES_PORT:5432}",
"db": "${env.POSTGRES_DB}",
"user": "${env.POSTGRES_USER}",
"password": "${env.POSTGRES_PASSWORD}",
"table_name": "${env.POSTGRES_TABLE_NAME:" + table_name + "}",
}
@classmethod
@field_validator("table_name")
def validate_table_name(cls, v: str) -> str: