phoenix-oss/llama-stack-mirror
0
0
Fork 1
mirror of https://github.com/meta-llama/llama-stack.git synced 2025-12-29 07:22:22 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Merge branch 'main' into feat/litellm_sambanova_usage

Browse source
This commit is contained in:
Jorge Piedrahita Ortiz 2025-03-19 16:51:59 -05:00 • committed by GitHub
commit 02a4f9ac59
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
69 changed files with 1128 additions and 445 deletions
Download patch file Download diff file Expand all files Collapse all files

2
llama_stack/apis/benchmarks/benchmarks.py
View file

@ -52,7 +52,7 @@ class Benchmarks(Protocol):
async def get_benchmark(
self,
benchmark_id: str,
) -> Optional[Benchmark]: ...
) -> Benchmark: ...
@webmethod(route="/eval/benchmarks", method="POST")
async def register_benchmark(

4
llama_stack/apis/datasets/datasets.py
View file

@ -121,8 +121,6 @@ class Dataset(CommonDatasetFields, Resource):
class DatasetInput(CommonDatasetFields, BaseModel):
dataset_id: str
provider_id: Optional[str] = None
provider_dataset_id: Optional[str] = None
class ListDatasetsResponse(BaseModel):
@ -201,7 +199,7 @@ class Datasets(Protocol):
async def get_dataset(
self,
dataset_id: str,
) -> Optional[Dataset]: ...
) -> Dataset: ...
@webmethod(route="/datasets", method="GET")
async def list_datasets(self) -> ListDatasetsResponse: ...

2
llama_stack/apis/eval/eval.py
View file

@ -117,7 +117,7 @@ class Eval(Protocol):
"""
@webmethod(route="/eval/benchmarks/{benchmark_id}/jobs/{job_id}", method="GET")
async def job_status(self, benchmark_id: str, job_id: str) -> Optional[JobStatus]:
async def job_status(self, benchmark_id: str, job_id: str) -> JobStatus:
"""Get the status of a job.
:param benchmark_id: The ID of the benchmark to run the evaluation on.

2
llama_stack/apis/files/files.py
View file

@ -115,7 +115,7 @@ class Files(Protocol):
async def get_upload_session_info(
self,
upload_id: str,
) -> Optional[FileUploadResponse]:
) -> FileUploadResponse:
"""
Returns information about an existsing upload session

2
llama_stack/apis/models/models.py
View file

@ -66,7 +66,7 @@ class Models(Protocol):
async def get_model(
self,
model_id: str,
) -> Optional[Model]: ...
) -> Model: ...
@webmethod(route="/models", method="POST")
async def register_model(

10
llama_stack/apis/post_training/post_training.py
View file

@ -6,7 +6,7 @@
from datetime import datetime
from enum import Enum
from typing import Any, Dict, List, Literal, Optional, Protocol, Union
from typing import Any, Dict, List, Literal, Optional, Protocol
from pydantic import BaseModel, Field
from typing_extensions import Annotated
@ -89,7 +89,7 @@ class QATFinetuningConfig(BaseModel):
AlgorithmConfig = register_schema(
Annotated[Union[LoraFinetuningConfig, QATFinetuningConfig], Field(discriminator="type")],
Annotated[LoraFinetuningConfig | QATFinetuningConfig, Field(discriminator="type")],
name="AlgorithmConfig",
)
@ -184,7 +184,7 @@ class PostTraining(Protocol):
description="Model descriptor from `llama model list`",
),
checkpoint_dir: Optional[str] = None,
algorithm_config: Optional[AlgorithmConfig] = None,
algorithm_config: Optional[LoraFinetuningConfig | QATFinetuningConfig] = None,
) -> PostTrainingJob: ...
@webmethod(route="/post-training/preference-optimize", method="POST")
@ -202,10 +202,10 @@ class PostTraining(Protocol):
async def get_training_jobs(self) -> ListPostTrainingJobsResponse: ...
@webmethod(route="/post-training/job/status", method="GET")
async def get_training_job_status(self, job_uuid: str) -> Optional[PostTrainingJobStatusResponse]: ...
async def get_training_job_status(self, job_uuid: str) -> PostTrainingJobStatusResponse: ...
@webmethod(route="/post-training/job/cancel", method="POST")
async def cancel_training_job(self, job_uuid: str) -> None: ...
@webmethod(route="/post-training/job/artifacts", method="GET")
async def get_training_job_artifacts(self, job_uuid: str) -> Optional[PostTrainingJobArtifactsResponse]: ...
async def get_training_job_artifacts(self, job_uuid: str) -> PostTrainingJobArtifactsResponse: ...

2
llama_stack/apis/scoring_functions/scoring_functions.py
View file

@ -135,7 +135,7 @@ class ScoringFunctions(Protocol):
async def list_scoring_functions(self) -> ListScoringFunctionsResponse: ...
@webmethod(route="/scoring-functions/{scoring_fn_id:path}", method="GET")
async def get_scoring_function(self, scoring_fn_id: str, /) -> Optional[ScoringFn]: ...
async def get_scoring_function(self, scoring_fn_id: str, /) -> ScoringFn: ...
@webmethod(route="/scoring-functions", method="POST")
async def register_scoring_function(

2
llama_stack/apis/shields/shields.py
View file

@ -49,7 +49,7 @@ class Shields(Protocol):
async def list_shields(self) -> ListShieldsResponse: ...
@webmethod(route="/shields/{identifier:path}", method="GET")
async def get_shield(self, identifier: str) -> Optional[Shield]: ...
async def get_shield(self, identifier: str) -> Shield: ...
@webmethod(route="/shields", method="POST")
async def register_shield(

9
llama_stack/apis/tools/rag_tool.py
View file

@ -17,6 +17,15 @@ from llama_stack.schema_utils import json_schema_type, register_schema, webmetho
@json_schema_type
class RAGDocument(BaseModel):
"""
A document to be used for document ingestion in the RAG Tool.
:param document_id: The unique identifier for the document.
:param content: The content of the document.
:param mime_type: The MIME type of the document.
:param metadata: Additional metadata for the document.
"""
document_id: str
content: InterleavedContent | URL
mime_type: str | None = None

2
llama_stack/apis/vector_dbs/vector_dbs.py
View file

@ -50,7 +50,7 @@ class VectorDBs(Protocol):
async def get_vector_db(
self,
vector_db_id: str,
) -> Optional[VectorDB]: ...
) -> VectorDB: ...
@webmethod(route="/vector-dbs", method="POST")
async def register_vector_db(

11
llama_stack/distribution/datatypes.py
View file

@ -125,6 +125,13 @@ class LoggingConfig(BaseModel):
)
class AuthenticationConfig(BaseModel):
endpoint: str = Field(
...,
description="Endpoint URL to validate authentication tokens",
)
class ServerConfig(BaseModel):
port: int = Field(
default=8321,
@ -140,6 +147,10 @@ class ServerConfig(BaseModel):
default=None,
description="Path to TLS key file for HTTPS",
)
auth: Optional[AuthenticationConfig] = Field(
default=None,
description="Authentication configuration for the server",
)
class StackRunConfig(BaseModel):

55
llama_stack/distribution/routers/routing_tables.py
View file

@ -20,6 +20,8 @@ from llama_stack.apis.datasets import (
DatasetType,
DataSource,
ListDatasetsResponse,
RowsDataSource,
URIDataSource,
)
from llama_stack.apis.models import ListModelsResponse, Model, Models, ModelType
from llama_stack.apis.resource import ResourceType
@ -219,8 +221,11 @@ class ModelsRoutingTable(CommonRoutingTableImpl, Models):
async def list_models(self) -> ListModelsResponse:
return ListModelsResponse(data=await self.get_all_with_type("model"))
async def get_model(self, model_id: str) -> Optional[Model]:
return await self.get_object_by_identifier("model", model_id)
async def get_model(self, model_id: str) -> Model:
model = await self.get_object_by_identifier("model", model_id)
if model is None:
raise ValueError(f"Model '{model_id}' not found")
return model
async def register_model(
self,
@ -267,8 +272,11 @@ class ShieldsRoutingTable(CommonRoutingTableImpl, Shields):
async def list_shields(self) -> ListShieldsResponse:
return ListShieldsResponse(data=await self.get_all_with_type(ResourceType.shield.value))
async def get_shield(self, identifier: str) -> Optional[Shield]:
return await self.get_object_by_identifier("shield", identifier)
async def get_shield(self, identifier: str) -> Shield:
shield = await self.get_object_by_identifier("shield", identifier)
if shield is None:
raise ValueError(f"Shield '{identifier}' not found")
return shield
async def register_shield(
self,
@ -303,8 +311,11 @@ class VectorDBsRoutingTable(CommonRoutingTableImpl, VectorDBs):
async def list_vector_dbs(self) -> ListVectorDBsResponse:
return ListVectorDBsResponse(data=await self.get_all_with_type("vector_db"))
async def get_vector_db(self, vector_db_id: str) -> Optional[VectorDB]:
return await self.get_object_by_identifier("vector_db", vector_db_id)
async def get_vector_db(self, vector_db_id: str) -> VectorDB:
vector_db = await self.get_object_by_identifier("vector_db", vector_db_id)
if vector_db is None:
raise ValueError(f"Vector DB '{vector_db_id}' not found")
return vector_db
async def register_vector_db(
self,
@ -355,8 +366,11 @@ class DatasetsRoutingTable(CommonRoutingTableImpl, Datasets):
async def list_datasets(self) -> ListDatasetsResponse:
return ListDatasetsResponse(data=await self.get_all_with_type(ResourceType.dataset.value))
async def get_dataset(self, dataset_id: str) -> Optional[Dataset]:
return await self.get_object_by_identifier("dataset", dataset_id)
async def get_dataset(self, dataset_id: str) -> Dataset:
dataset = await self.get_object_by_identifier("dataset", dataset_id)
if dataset is None:
raise ValueError(f"Dataset '{dataset_id}' not found")
return dataset
async def register_dataset(
self,
@ -365,6 +379,12 @@ class DatasetsRoutingTable(CommonRoutingTableImpl, Datasets):
metadata: Optional[Dict[str, Any]] = None,
dataset_id: Optional[str] = None,
) -> Dataset:
if isinstance(source, dict):
if source["type"] == "uri":
source = URIDataSource.parse_obj(source)
elif source["type"] == "rows":
source = RowsDataSource.parse_obj(source)
if not dataset_id:
dataset_id = f"dataset-{str(uuid.uuid4())}"
@ -408,8 +428,11 @@ class ScoringFunctionsRoutingTable(CommonRoutingTableImpl, ScoringFunctions):
async def list_scoring_functions(self) -> ListScoringFunctionsResponse:
return ListScoringFunctionsResponse(data=await self.get_all_with_type(ResourceType.scoring_function.value))
async def get_scoring_function(self, scoring_fn_id: str) -> Optional[ScoringFn]:
return await self.get_object_by_identifier("scoring_function", scoring_fn_id)
async def get_scoring_function(self, scoring_fn_id: str) -> ScoringFn:
scoring_fn = await self.get_object_by_identifier("scoring_function", scoring_fn_id)
if scoring_fn is None:
raise ValueError(f"Scoring function '{scoring_fn_id}' not found")
return scoring_fn
async def register_scoring_function(
self,
@ -445,8 +468,11 @@ class BenchmarksRoutingTable(CommonRoutingTableImpl, Benchmarks):
async def list_benchmarks(self) -> ListBenchmarksResponse:
return ListBenchmarksResponse(data=await self.get_all_with_type("benchmark"))
async def get_benchmark(self, benchmark_id: str) -> Optional[Benchmark]:
return await self.get_object_by_identifier("benchmark", benchmark_id)
async def get_benchmark(self, benchmark_id: str) -> Benchmark:
benchmark = await self.get_object_by_identifier("benchmark", benchmark_id)
if benchmark is None:
raise ValueError(f"Benchmark '{benchmark_id}' not found")
return benchmark
async def register_benchmark(
self,
@ -490,7 +516,10 @@ class ToolGroupsRoutingTable(CommonRoutingTableImpl, ToolGroups):
return ListToolGroupsResponse(data=await self.get_all_with_type("tool_group"))
async def get_tool_group(self, toolgroup_id: str) -> ToolGroup:
return await self.get_object_by_identifier("tool_group", toolgroup_id)
tool_group = await self.get_object_by_identifier("tool_group", toolgroup_id)
if tool_group is None:
raise ValueError(f"Tool group '{toolgroup_id}' not found")
return tool_group
async def get_tool(self, tool_name: str) -> Tool:
return await self.get_object_by_identifier("tool", tool_name)

69
llama_stack/distribution/server/auth.py Normal file
View file

@ -0,0 +1,69 @@
# 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.
import json
from urllib.parse import parse_qs
import httpx
from llama_stack.log import get_logger
logger = get_logger(name=__name__, category="auth")
class AuthenticationMiddleware:
def __init__(self, app, auth_endpoint):
self.app = app
self.auth_endpoint = auth_endpoint
async def __call__(self, scope, receive, send):
if scope["type"] == "http":
headers = dict(scope.get("headers", []))
auth_header = headers.get(b"authorization", b"").decode()
if not auth_header or not auth_header.startswith("Bearer "):
return await self._send_auth_error(send, "Missing or invalid Authorization header")
api_key = auth_header.split("Bearer ", 1)[1]
path = scope.get("path", "")
request_headers = {k.decode(): v.decode() for k, v in headers.items()}
query_string = scope.get("query_string", b"").decode()
params = parse_qs(query_string)
auth_data = {
"api_key": api_key,
"request": {
"path": path,
"headers": request_headers,
"params": params,
},
}
# Validate with authentication endpoint
try:
async with httpx.AsyncClient() as client:
response = await client.post(self.auth_endpoint, json=auth_data)
if response.status_code != 200:
logger.warning(f"Authentication failed: {response.status_code}")
return await self._send_auth_error(send, "Authentication failed")
except Exception:
logger.exception("Error during authentication")
return await self._send_auth_error(send, "Authentication service error")
return await self.app(scope, receive, send)
async def _send_auth_error(self, send, message):
await send(
{
"type": "http.response.start",
"status": 401,
"headers": [[b"content-type", b"application/json"]],
}
)
error_msg = json.dumps({"error": {"message": message}}).encode()
await send({"type": "http.response.body", "body": error_msg})

6
llama_stack/distribution/server/server.py
View file

@ -52,6 +52,7 @@ from llama_stack.providers.utils.telemetry.tracing import (
start_trace,
)
from .auth import AuthenticationMiddleware
from .endpoints import get_all_api_endpoints
REPO_ROOT = Path(__file__).parent.parent.parent.parent
@ -351,6 +352,11 @@ def main():
if not os.environ.get("LLAMA_STACK_DISABLE_VERSION_CHECK"):
app.add_middleware(ClientVersionMiddleware)
# Add authentication middleware if configured
if config.server.auth and config.server.auth.endpoint:
logger.info(f"Enabling authentication with endpoint: {config.server.auth.endpoint}")
app.add_middleware(AuthenticationMiddleware, auth_endpoint=config.server.auth.endpoint)
try:
impls = asyncio.run(construct_stack(config))
except InvalidProviderError as e:

3
llama_stack/distribution/ui/page/distribution/datasets.py
View file

@ -5,7 +5,8 @@
# the root directory of this source tree.
import streamlit as st
from modules.api import llama_stack_api
from llama_stack.distribution.ui.modules.api import llama_stack_api
def datasets():

3
llama_stack/distribution/ui/page/distribution/eval_tasks.py
View file

@ -5,7 +5,8 @@
# the root directory of this source tree.
import streamlit as st
from modules.api import llama_stack_api
from llama_stack.distribution.ui.modules.api import llama_stack_api
def benchmarks():

3
llama_stack/distribution/ui/page/distribution/models.py
View file

@ -5,7 +5,8 @@
# the root directory of this source tree.
import streamlit as st
from modules.api import llama_stack_api
from llama_stack.distribution.ui.modules.api import llama_stack_api
def models():

3
llama_stack/distribution/ui/page/distribution/providers.py
View file

@ -5,7 +5,8 @@
# the root directory of this source tree.
import streamlit as st
from modules.api import llama_stack_api
from llama_stack.distribution.ui.modules.api import llama_stack_api
def providers():

13
llama_stack/distribution/ui/page/distribution/resources.py
View file

@ -4,14 +4,15 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from page.distribution.benchmarks import benchmarks
from page.distribution.datasets import datasets
from page.distribution.models import models
from page.distribution.scoring_functions import scoring_functions
from page.distribution.shields import shields
from page.distribution.vector_dbs import vector_dbs
from streamlit_option_menu import option_menu
from llama_stack.distribution.ui.page.distribution.datasets import datasets
from llama_stack.distribution.ui.page.distribution.eval_tasks import benchmarks
from llama_stack.distribution.ui.page.distribution.models import models
from llama_stack.distribution.ui.page.distribution.scoring_functions import scoring_functions
from llama_stack.distribution.ui.page.distribution.shields import shields
from llama_stack.distribution.ui.page.distribution.vector_dbs import vector_dbs
def resources_page():
options = [

3
llama_stack/distribution/ui/page/distribution/scoring_functions.py
View file

@ -5,7 +5,8 @@
# the root directory of this source tree.
import streamlit as st
from modules.api import llama_stack_api
from llama_stack.distribution.ui.modules.api import llama_stack_api
def scoring_functions():

3
llama_stack/distribution/ui/page/distribution/shields.py
View file

@ -5,7 +5,8 @@
# the root directory of this source tree.
import streamlit as st
from modules.api import llama_stack_api
from llama_stack.distribution.ui.modules.api import llama_stack_api
def shields():

3
llama_stack/distribution/ui/page/distribution/vector_dbs.py
View file

@ -5,7 +5,8 @@
# the root directory of this source tree.
import streamlit as st
from modules.api import llama_stack_api
from llama_stack.distribution.ui.modules.api import llama_stack_api
def vector_dbs():

5
llama_stack/distribution/ui/page/evaluations/app_eval.py
View file

@ -8,8 +8,9 @@ import json
import pandas as pd
import streamlit as st
from modules.api import llama_stack_api
from modules.utils import process_dataset
from llama_stack.distribution.ui.modules.api import llama_stack_api
from llama_stack.distribution.ui.modules.utils import process_dataset
def application_evaluation_page():

3
llama_stack/distribution/ui/page/evaluations/native_eval.py
View file

@ -8,7 +8,8 @@ import json
import pandas as pd
import streamlit as st
from modules.api import llama_stack_api
from llama_stack.distribution.ui.modules.api import llama_stack_api
def select_benchmark_1():

3
llama_stack/distribution/ui/page/playground/chat.py
View file

@ -5,7 +5,8 @@
# the root directory of this source tree.
import streamlit as st
from modules.api import llama_stack_api
from llama_stack.distribution.ui.modules.api import llama_stack_api
# Sidebar configurations
with st.sidebar:

7
llama_stack/distribution/ui/page/playground/rag.py
View file

@ -7,9 +7,10 @@
import streamlit as st
from llama_stack_client.lib.agents.agent import Agent
from llama_stack_client.lib.agents.event_logger import EventLogger
from llama_stack_client.types.memory_insert_params import Document
from modules.api import llama_stack_api
from modules.utils import data_url_from_file
from llama_stack_client.types.shared.document import Document
from llama_stack.distribution.ui.modules.api import llama_stack_api
from llama_stack.distribution.ui.modules.utils import data_url_from_file
def rag_chat_page():

9
llama_stack/models/llama/datatypes.py
View file

@ -47,7 +47,14 @@ RecursiveType = Union[Primitive, List[Primitive], Dict[str, Primitive]]
class ToolCall(BaseModel):
call_id: str
tool_name: Union[BuiltinTool, str]
arguments: Dict[str, RecursiveType]
# Plan is to deprecate the Dict in favor of a JSON string
# that is parsed on the client side instead of trying to manage
# the recursive type here.
# Making this a union so that client side can start prepping for this change.
# Eventually, we will remove both the Dict and arguments_json field,
# and arguments will just be a str
arguments: Union[str, Dict[str, RecursiveType]]
arguments_json: Optional[str] = None
@field_validator("tool_name", mode="before")
@classmethod

9
llama_stack/models/llama/llama3/chat_format.py
View file

@ -12,6 +12,7 @@
# the top-level of this source tree.
import io
import json
import uuid
from dataclasses import dataclass
from typing import Dict, List, Optional, Tuple
@ -203,9 +204,10 @@ class ChatFormat:
# This code tries to handle that case
if tool_name in BuiltinTool.__members__:
tool_name = BuiltinTool[tool_name]
tool_arguments = {
"query": list(tool_arguments.values())[0],
}
if isinstance(tool_arguments, dict):
tool_arguments = {
"query": list(tool_arguments.values())[0],
}
else:
builtin_tool_info = ToolUtils.maybe_extract_builtin_tool_call(content)
if builtin_tool_info is not None:
@ -229,6 +231,7 @@ class ChatFormat:
call_id=call_id,
tool_name=tool_name,
arguments=tool_arguments,
arguments_json=json.dumps(tool_arguments),
)
)
content = ""

7
llama_stack/models/llama/llama3/template_data.py
View file

@ -11,11 +11,8 @@
# top-level folder for each specific model found within the models/ directory at
# the top-level of this source tree.
from llama_stack.models.llama.datatypes import (
BuiltinTool,
StopReason,
ToolCall,
)
from llama_stack.models.llama.datatypes import BuiltinTool, StopReason, ToolCall
from .prompt_templates import (
BuiltinToolGenerator,

8
llama_stack/providers/inline/datasetio/localfs/datasetio.py
View file

@ -35,12 +35,12 @@ class PandasDataframeDataset:
else:
return self.df.iloc[idx].to_dict()
def load(self) -> None:
async def load(self) -> None:
if self.df is not None:
return
if self.dataset_def.source.type == "uri":
self.df = get_dataframe_from_uri(self.dataset_def.source.uri)
self.df = await get_dataframe_from_uri(self.dataset_def.source.uri)
elif self.dataset_def.source.type == "rows":
self.df = pandas.DataFrame(self.dataset_def.source.rows)
else:
@ -95,7 +95,7 @@ class LocalFSDatasetIOImpl(DatasetIO, DatasetsProtocolPrivate):
) -> IterrowsResponse:
dataset_def = self.dataset_infos[dataset_id]
dataset_impl = PandasDataframeDataset(dataset_def)
dataset_impl.load()
await dataset_impl.load()
start_index = start_index or 0
@ -114,7 +114,7 @@ class LocalFSDatasetIOImpl(DatasetIO, DatasetsProtocolPrivate):
async def append_rows(self, dataset_id: str, rows: List[Dict[str, Any]]) -> None:
dataset_def = self.dataset_infos[dataset_id]
dataset_impl = PandasDataframeDataset(dataset_def)
dataset_impl.load()
await dataset_impl.load()
new_rows_df = pandas.DataFrame(rows)
dataset_impl.df = pandas.concat([dataset_impl.df, new_rows_df], ignore_index=True)

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

@ -10,6 +10,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
import copy
import json
import logging
import multiprocessing
@ -213,7 +214,7 @@ def maybe_parse_message(maybe_json: Optional[str]) -> Optional[ProcessingMessage
def parse_message(json_str: str) -> ProcessingMessage:
data = json.loads(json_str)
return ProcessingMessageWrapper(**data).payload
return copy.deepcopy(ProcessingMessageWrapper(**data).payload)
def worker_process_entrypoint(

1
llama_stack/providers/inline/inference/vllm/vllm.py
View file

@ -582,6 +582,7 @@ class VLLMInferenceImpl(Inference, ModelsProtocolPrivate):
tool_name=t.function.name,
# vLLM function args come back as a string. Llama Stack expects JSON.
arguments=json.loads(t.function.arguments),
arguments_json=t.function.arguments,
)
for t in vllm_message.tool_calls
],

8
llama_stack/providers/inline/post_training/common/validator.py
View file

@ -9,6 +9,9 @@
#
# 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
from llama_stack.apis.common.type_system import (
ChatCompletionInputType,
DialogType,
@ -20,7 +23,7 @@ from llama_stack.providers.utils.common.data_schema_validator import (
validate_dataset_schema,
)
EXPECTED_DATASET_SCHEMA = {
EXPECTED_DATASET_SCHEMA: dict[str, list[dict[str, Any]]] = {
"instruct": [
{
ColumnName.chat_completion_input.value: ChatCompletionInputType(),
@ -41,6 +44,9 @@ async def validate_input_dataset_schema(
dataset_type: str,
) -> None:
dataset_def = await datasets_api.get_dataset(dataset_id=dataset_id)
if not dataset_def:
raise ValueError(f"Dataset {dataset_id} does not exist.")
if not dataset_def.dataset_schema or len(dataset_def.dataset_schema) == 0:
raise ValueError(f"Dataset {dataset_id} does not have a schema defined.")

8
llama_stack/providers/inline/post_training/torchtune/common/checkpointer.py
View file

@ -37,7 +37,7 @@ class TorchtuneCheckpointer:
checkpoint_files: List[str],
output_dir: str,
model_type: str,
) -> None:
):
# Fail fast if ``checkpoint_files`` is invalid
# TODO: support loading more than one file
if len(checkpoint_files) != 1:
@ -58,7 +58,7 @@ class TorchtuneCheckpointer:
"""
Load Meta checkpoint from file. Currently only loading from a single file is supported.
"""
state_dict: Dict[str:Any] = {}
state_dict: Dict[str, Any] = {}
model_state_dict = safe_torch_load(self._checkpoint_path)
if self._model_type == ModelType.LLAMA3_VISION:
from torchtune.models.llama3_2_vision._convert_weights import (
@ -85,10 +85,10 @@ class TorchtuneCheckpointer:
state_dict: Dict[str, Any],
epoch: int,
adapter_only: bool = False,
checkpoint_format: str = "meta",
checkpoint_format: str | None = None,
) -> str:
model_file_path = Path(self._output_dir) / f"{self._model_id}-{self._training_algorithm}-{epoch}"
if checkpoint_format == "meta":
if checkpoint_format == "meta" or checkpoint_format is None:
self._save_meta_format_checkpoint(model_file_path, state_dict, adapter_only)
elif checkpoint_format == "huggingface":
# Note: for saving hugging face format checkpoints, we only suppport saving adapter weights now

13
llama_stack/providers/inline/post_training/torchtune/common/utils.py
View file

@ -10,7 +10,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
from typing import Any, Callable, Dict
from typing import Callable, Dict
import torch
from pydantic import BaseModel
@ -25,10 +25,13 @@ from llama_stack.apis.post_training import DatasetFormat
from llama_stack.models.llama.datatypes import Model
from llama_stack.models.llama.sku_list import resolve_model
BuildLoraModelCallable = Callable[..., torch.nn.Module]
BuildTokenizerCallable = Callable[..., Llama3Tokenizer]
class ModelConfig(BaseModel):
model_definition: Any
tokenizer_type: Any
model_definition: BuildLoraModelCallable
tokenizer_type: BuildTokenizerCallable
checkpoint_type: str
@ -51,10 +54,6 @@ DATA_FORMATS: Dict[str, Transform] = {
}
BuildLoraModelCallable = Callable[..., torch.nn.Module]
BuildTokenizerCallable = Callable[..., Llama3Tokenizer]
def _validate_model_id(model_id: str) -> Model:
model = resolve_model(model_id)
if model is None or model.core_model_id.value not in MODEL_CONFIGS:

2
llama_stack/providers/inline/post_training/torchtune/datasets/sft.py
View file

@ -55,7 +55,7 @@ class SFTDataset(Dataset):
if "messages" in transformed_sample:
validate_messages(transformed_sample["messages"])
tokenized_dict = self._model_transform(transformed_sample)
tokenized_dict: dict[str, Any] = self._model_transform(transformed_sample)
if not ("tokens" in tokenized_dict and "mask" in tokenized_dict):
keys_str = ", ".join(tokenized_dict.keys())

45
llama_stack/providers/inline/post_training/torchtune/recipes/lora_finetuning_single_device.py
View file

@ -37,10 +37,10 @@ from llama_stack.apis.common.training_types import PostTrainingMetric
from llama_stack.apis.datasetio import DatasetIO
from llama_stack.apis.datasets import Datasets
from llama_stack.apis.post_training import (
AlgorithmConfig,
Checkpoint,
LoraFinetuningConfig,
OptimizerConfig,
QATFinetuningConfig,
TrainingConfig,
)
from llama_stack.distribution.utils.config_dirs import DEFAULT_CHECKPOINT_DIR
@ -73,6 +73,9 @@ class LoraFinetuningSingleDevice:
# Currently logging only logs limited training metrics to local disk
# will figure out more loggings and how it works with telemetry in future PRs
_checkpointer: TorchtuneCheckpointer
def __init__(
self,
config: TorchtunePostTrainingConfig,
@ -82,7 +85,7 @@ class LoraFinetuningSingleDevice:
logger_config: Dict[str, Any],
model: str,
checkpoint_dir: Optional[str],
algorithm_config: Optional[AlgorithmConfig],
algorithm_config: LoraFinetuningConfig | QATFinetuningConfig | None,
datasetio_api: DatasetIO,
datasets_api: Datasets,
) -> None:
@ -109,12 +112,12 @@ class LoraFinetuningSingleDevice:
return str(checkpoint_dir)
if checkpoint_dir and checkpoint_dir != "null":
self.checkpoint_dir = config.checkpoint_dir
self.checkpoint_dir = checkpoint_dir
else:
model = resolve_model(self.model_id)
if model is None:
model_obj = resolve_model(self.model_id)
if model_obj is None:
raise ValueError(f"{self.model_id} not found. Your model id should be in the llama models SKU list")
self.checkpoint_dir = model_checkpoint_dir(model)
self.checkpoint_dir = model_checkpoint_dir(model_obj)
self._output_dir = str(DEFAULT_CHECKPOINT_DIR)
self._checkpoint_format = config.checkpoint_format
@ -135,16 +138,16 @@ class LoraFinetuningSingleDevice:
self.max_validation_steps = training_config.max_validation_steps
self._clip_grad_norm = 1.0
self._enable_activation_checkpointing = (
(training_config.efficiency_config.enable_activation_checkpointing)
if training_config.efficiency_config
else False
)
self._enable_activation_offloading = (
(training_config.efficiency_config.enable_activation_offloading)
if training_config.efficiency_config
else False
)
self._enable_activation_checkpointing = False
self._enable_activation_offloading = False
if training_config.efficiency_config:
if training_config.efficiency_config.enable_activation_checkpointing:
self._enable_activation_checkpointing = (
training_config.efficiency_config.enable_activation_checkpointing
)
if training_config.efficiency_config.enable_activation_offloading:
self._enable_activation_offloading = training_config.efficiency_config.enable_activation_offloading
self.datasetio_api = datasetio_api
self.datasets_api = datasets_api
@ -451,12 +454,12 @@ class LoraFinetuningSingleDevice:
"""
# Initialize tokens count and running loss (for grad accumulation)
t0 = time.perf_counter()
running_loss = 0
running_loss: float = 0.0
num_tokens = 0
# training artifacts
checkpoints = []
memory_stats = {}
memory_stats: Dict[str, Any] = {}
# self.epochs_run should be non-zero when we're resuming from a checkpoint
for curr_epoch in range(self.epochs_run, self.total_epochs):
@ -484,7 +487,7 @@ class LoraFinetuningSingleDevice:
# Loss is normalized by default so we multiply by the number of tokens
# This way we can normalize by the total number of tokens if we're accumulating gradients
current_loss = await self._loss_step(batch) * current_num_tokens
running_loss += current_loss
running_loss += current_loss.detach().item()
current_loss.backward()
# Step with optimizer
@ -500,7 +503,7 @@ class LoraFinetuningSingleDevice:
# Update the number of steps when the weights are updated
self.global_step += 1
loss_to_log = running_loss.item() / num_tokens
loss_to_log = running_loss / num_tokens
pbar.update(1)
pbar.set_description(f"{curr_epoch + 1}|{self.global_step}|Loss: {loss_to_log}")
@ -523,7 +526,7 @@ class LoraFinetuningSingleDevice:
)
# Reset running stats for the next step
running_loss = 0
running_loss = 0.0
num_tokens = 0
t0 = time.perf_counter()

5
llama_stack/providers/inline/tool_runtime/code_interpreter/code_interpreter.py
View file

@ -5,6 +5,7 @@
# the root directory of this source tree.
import asyncio
import logging
import os
import tempfile
@ -37,7 +38,7 @@ class CodeInterpreterToolRuntimeImpl(ToolsProtocolPrivate, ToolRuntime):
async def initialize(self):
pass
async def register_tool(self, tool: Tool):
async def register_tool(self, tool: Tool) -> None:
pass
async def unregister_tool(self, tool_id: str) -> None:
@ -65,7 +66,7 @@ class CodeInterpreterToolRuntimeImpl(ToolsProtocolPrivate, ToolRuntime):
# Use environment variable to control bwrap usage
force_disable_bwrap = os.environ.get("DISABLE_CODE_SANDBOX", "").lower() in ("1", "true", "yes")
req = CodeExecutionRequest(scripts=[script], use_bwrap=not force_disable_bwrap)
res = self.code_executor.execute(req)
res = await asyncio.to_thread(self.code_executor.execute, req)
pieces = [res["process_status"]]
for out_type in ["stdout", "stderr"]:
res_out = res[out_type]

19
llama_stack/providers/inline/vector_io/qdrant/__init__.py Normal file
View file

@ -0,0 +1,19 @@
# 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 Dict
from llama_stack.providers.datatypes import Api, ProviderSpec
from .config import QdrantVectorIOConfig
async def get_adapter_impl(config: QdrantVectorIOConfig, deps: Dict[Api, ProviderSpec]):
from llama_stack.providers.remote.vector_io.qdrant.qdrant import QdrantVectorIOAdapter
impl = QdrantVectorIOAdapter(config, deps[Api.inference])
await impl.initialize()
return impl

23
llama_stack/providers/inline/vector_io/qdrant/config.py Normal file
View file

@ -0,0 +1,23 @@
# 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
from llama_stack.schema_utils import json_schema_type
@json_schema_type
class QdrantVectorIOConfig(BaseModel):
path: str
@classmethod
def sample_run_config(cls, __distro_dir__: str) -> Dict[str, Any]:
return {
"path": "${env.QDRANT_PATH:~/.llama/" + __distro_dir__ + "}/" + "qdrant.db",
}

8
llama_stack/providers/registry/vector_io.py
View file

@ -92,6 +92,14 @@ def available_providers() -> List[ProviderSpec]:
),
api_dependencies=[Api.inference],
),
InlineProviderSpec(
api=Api.vector_io,
provider_type="inline::qdrant",
pip_packages=["qdrant-client"],
module="llama_stack.providers.inline.vector_io.qdrant",
config_class="llama_stack.providers.inline.vector_io.qdrant.QdrantVectorIOConfig",
api_dependencies=[Api.inference],
),
remote_provider_spec(
Api.vector_io,
AdapterSpec(

51
llama_stack/providers/remote/inference/nvidia/nvidia.py
View file

@ -6,6 +6,7 @@
import logging
import warnings
from functools import lru_cache
from typing import AsyncIterator, List, Optional, Union
from openai import APIConnectionError, AsyncOpenAI, BadRequestError
@ -82,12 +83,42 @@ class NVIDIAInferenceAdapter(Inference, ModelRegistryHelper):
# )
self._config = config
# make sure the client lives longer than any async calls
self._client = AsyncOpenAI(
base_url=f"{self._config.url}/v1",
api_key=(self._config.api_key.get_secret_value() if self._config.api_key else "NO KEY"),
timeout=self._config.timeout,
)
@lru_cache # noqa: B019
def _get_client(self, provider_model_id: str) -> AsyncOpenAI:
"""
For hosted models, https://integrate.api.nvidia.com/v1 is the primary base_url. However,
some models are hosted on different URLs. This function returns the appropriate client
for the given provider_model_id.
This relies on lru_cache and self._default_client to avoid creating a new client for each request
or for each model that is hosted on https://integrate.api.nvidia.com/v1.
:param provider_model_id: The provider model ID
:return: An OpenAI client
"""
@lru_cache # noqa: B019
def _get_client_for_base_url(base_url: str) -> AsyncOpenAI:
"""
Maintain a single OpenAI client per base_url.
"""
return AsyncOpenAI(
base_url=base_url,
api_key=(self._config.api_key.get_secret_value() if self._config.api_key else "NO KEY"),
timeout=self._config.timeout,
)
special_model_urls = {
"meta/llama-3.2-11b-vision-instruct": "https://ai.api.nvidia.com/v1/gr/meta/llama-3.2-11b-vision-instruct",
"meta/llama-3.2-90b-vision-instruct": "https://ai.api.nvidia.com/v1/gr/meta/llama-3.2-90b-vision-instruct",
}
base_url = f"{self._config.url}/v1"
if _is_nvidia_hosted(self._config) and provider_model_id in special_model_urls:
base_url = special_model_urls[provider_model_id]
return _get_client_for_base_url(base_url)
async def completion(
self,
@ -105,9 +136,10 @@ class NVIDIAInferenceAdapter(Inference, ModelRegistryHelper):
await check_health(self._config) # this raises errors
provider_model_id = self.get_provider_model_id(model_id)
request = convert_completion_request(
request=CompletionRequest(
model=self.get_provider_model_id(model_id),
model=provider_model_id,
content=content,
sampling_params=sampling_params,
response_format=response_format,
@ -118,7 +150,7 @@ class NVIDIAInferenceAdapter(Inference, ModelRegistryHelper):
)
try:
response = await self._client.completions.create(**request)
response = await self._get_client(provider_model_id).completions.create(**request)
except APIConnectionError as e:
raise ConnectionError(f"Failed to connect to NVIDIA NIM at {self._config.url}: {e}") from e
@ -206,6 +238,7 @@ class NVIDIAInferenceAdapter(Inference, ModelRegistryHelper):
await check_health(self._config) # this raises errors
provider_model_id = self.get_provider_model_id(model_id)
request = await convert_chat_completion_request(
request=ChatCompletionRequest(
model=self.get_provider_model_id(model_id),
@ -221,7 +254,7 @@ class NVIDIAInferenceAdapter(Inference, ModelRegistryHelper):
)
try:
response = await self._client.chat.completions.create(**request)
response = await self._get_client(provider_model_id).chat.completions.create(**request)
except APIConnectionError as e:
raise ConnectionError(f"Failed to connect to NVIDIA NIM at {self._config.url}: {e}") from e

8
llama_stack/providers/remote/inference/vllm/vllm.py
View file

@ -90,15 +90,12 @@ def _convert_to_vllm_tool_calls_in_response(
if not tool_calls:
return []
call_function_arguments = None
for call in tool_calls:
call_function_arguments = json.loads(call.function.arguments)
return [
ToolCall(
call_id=call.id,
tool_name=call.function.name,
arguments=call_function_arguments,
arguments=json.loads(call.function.arguments),
arguments_json=call.function.arguments,
)
for call in tool_calls
]
@ -183,6 +180,7 @@ async def _process_vllm_chat_completion_stream_response(
call_id=tool_call_buf.call_id,
tool_name=tool_call_buf.tool_name,
arguments=args,
arguments_json=args_str,
),
parse_status=ToolCallParseStatus.succeeded,
),

1
llama_stack/providers/remote/vector_io/qdrant/config.py
View file

@ -23,7 +23,6 @@ class QdrantVectorIOConfig(BaseModel):
prefix: Optional[str] = None
timeout: Optional[int] = None
host: Optional[str] = None
path: Optional[str] = None
@classmethod
def sample_run_config(cls, **kwargs: Any) -> Dict[str, Any]:

20
llama_stack/providers/remote/vector_io/qdrant/qdrant.py
View file

@ -6,7 +6,7 @@
import logging
import uuid
from typing import Any, Dict, List, Optional
from typing import Any, Dict, List, Optional, Union
from numpy.typing import NDArray
from qdrant_client import AsyncQdrantClient, models
@ -16,12 +16,13 @@ from llama_stack.apis.inference import InterleavedContent
from llama_stack.apis.vector_dbs import VectorDB
from llama_stack.apis.vector_io import Chunk, QueryChunksResponse, VectorIO
from llama_stack.providers.datatypes import Api, VectorDBsProtocolPrivate
from llama_stack.providers.inline.vector_io.qdrant import QdrantVectorIOConfig as InlineQdrantVectorIOConfig
from llama_stack.providers.utils.memory.vector_store import (
EmbeddingIndex,
VectorDBWithIndex,
)
from .config import QdrantVectorIOConfig
from .config import QdrantVectorIOConfig as RemoteQdrantVectorIOConfig
log = logging.getLogger(__name__)
CHUNK_ID_KEY = "_chunk_id"
@ -99,17 +100,19 @@ class QdrantIndex(EmbeddingIndex):
class QdrantVectorIOAdapter(VectorIO, VectorDBsProtocolPrivate):
def __init__(self, config: QdrantVectorIOConfig, inference_api: Api.inference) -> None:
def __init__(
self, config: Union[RemoteQdrantVectorIOConfig, InlineQdrantVectorIOConfig], inference_api: Api.inference
) -> None:
self.config = config
self.client = AsyncQdrantClient(**self.config.model_dump(exclude_none=True))
self.client: AsyncQdrantClient = None
self.cache = {}
self.inference_api = inference_api
async def initialize(self) -> None:
pass
self.client = AsyncQdrantClient(**self.config.model_dump(exclude_none=True))
async def shutdown(self) -> None:
self.client.close()
await self.client.close()
async def register_vector_db(
self,
@ -123,6 +126,11 @@ class QdrantVectorIOAdapter(VectorIO, VectorDBsProtocolPrivate):
self.cache[vector_db.identifier] = index
async def unregister_vector_db(self, vector_db_id: str) -> None:
if vector_db_id in self.cache:
await self.cache[vector_db_id].index.delete()
del self.cache[vector_db_id]
async def _get_and_cache_vector_db_index(self, vector_db_id: str) -> Optional[VectorDBWithIndex]:
if vector_db_id in self.cache:
return self.cache[vector_db_id]

10
llama_stack/providers/utils/datasetio/url_utils.py
View file

@ -4,6 +4,7 @@
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.
import asyncio
import base64
import io
from urllib.parse import unquote
@ -13,12 +14,15 @@ import pandas
from llama_stack.providers.utils.memory.vector_store import parse_data_url
def get_dataframe_from_uri(uri: str):
async def get_dataframe_from_uri(uri: str):
df = None
if uri.endswith(".csv"):
df = pandas.read_csv(uri)
# Moving to its own thread to avoid io from blocking the eventloop
# This isn't ideal as it moves more then just the IO to a new thread
# but it is as close as we can easly get
df = await asyncio.to_thread(pandas.read_csv, uri)
elif uri.endswith(".xlsx"):
df = pandas.read_excel(uri)
df = await asyncio.to_thread(pandas.read_excel, uri)
elif uri.startswith("data:"):
parts = parse_data_url(uri)
data = parts["data"]

14
llama_stack/providers/utils/inference/openai_compat.py
View file

@ -529,7 +529,11 @@ async def convert_message_to_openai_dict_new(
) -> Union[str, Iterable[OpenAIChatCompletionContentPartParam]]:
async def impl(
content_: InterleavedContent,
) -> Union[str, OpenAIChatCompletionContentPartParam, List[OpenAIChatCompletionContentPartParam]]:
) -> Union[
str,
OpenAIChatCompletionContentPartParam,
List[OpenAIChatCompletionContentPartParam],
]:
# Llama Stack and OpenAI spec match for str and text input
if isinstance(content_, str):
return content_
@ -570,7 +574,7 @@ async def convert_message_to_openai_dict_new(
OpenAIChatCompletionMessageToolCall(
id=tool.call_id,
function=OpenAIFunction(
name=tool.tool_name if not isinstance(tool.tool_name, BuiltinTool) else tool.tool_name.value,
name=(tool.tool_name if not isinstance(tool.tool_name, BuiltinTool) else tool.tool_name.value),
arguments=json.dumps(tool.arguments),
),
type="function",
@ -609,6 +613,7 @@ def convert_tool_call(
call_id=tool_call.id,
tool_name=tool_call.function.name,
arguments=json.loads(tool_call.function.arguments),
arguments_json=tool_call.function.arguments,
)
except Exception:
return UnparseableToolCall(
@ -759,6 +764,7 @@ def _convert_openai_tool_calls(
call_id=call.id,
tool_name=call.function.name,
arguments=json.loads(call.function.arguments),
arguments_json=call.function.arguments,
)
for call in tool_calls
]
@ -890,7 +896,8 @@ async def convert_openai_chat_completion_stream(
# ChatCompletionResponseEvent only supports one per stream
if len(choice.delta.tool_calls) > 1:
warnings.warn(
"multiple tool calls found in a single delta, using the first, ignoring the rest", stacklevel=2
"multiple tool calls found in a single delta, using the first, ignoring the rest",
stacklevel=2,
)
if not enable_incremental_tool_calls:
@ -971,6 +978,7 @@ async def convert_openai_chat_completion_stream(
call_id=buffer["call_id"],
tool_name=buffer["name"],
arguments=arguments,
arguments_json=buffer["arguments"],
)
yield ChatCompletionResponseStreamChunk(
event=ChatCompletionResponseEvent(

5
llama_stack/templates/open-benchmark/open_benchmark.py
View file

@ -170,7 +170,6 @@ def get_distribution_template() -> DistributionTemplate:
default_datasets = [
DatasetInput(
dataset_id="simpleqa",
provider_id="huggingface",
purpose=DatasetPurpose.eval_messages_answer,
source=URIDataSource(
uri="huggingface://datasets/llamastack/simpleqa?split=train",
@ -178,7 +177,6 @@ def get_distribution_template() -> DistributionTemplate:
),
DatasetInput(
dataset_id="mmlu_cot",
provider_id="huggingface",
purpose=DatasetPurpose.eval_messages_answer,
source=URIDataSource(
uri="huggingface://datasets/llamastack/mmlu_cot?split=test&name=all",
@ -186,7 +184,6 @@ def get_distribution_template() -> DistributionTemplate:
),
DatasetInput(
dataset_id="gpqa_cot",
provider_id="huggingface",
purpose=DatasetPurpose.eval_messages_answer,
source=URIDataSource(
uri="huggingface://datasets/llamastack/gpqa_0shot_cot?split=test&name=gpqa_main",
@ -194,7 +191,6 @@ def get_distribution_template() -> DistributionTemplate:
),
DatasetInput(
dataset_id="math_500",
provider_id="huggingface",
purpose=DatasetPurpose.eval_messages_answer,
source=URIDataSource(
uri="huggingface://datasets/llamastack/math_500?split=test",
@ -202,7 +198,6 @@ def get_distribution_template() -> DistributionTemplate:
),
DatasetInput(
dataset_id="bfcl",
provider_id="huggingface",
purpose=DatasetPurpose.eval_messages_answer,
source=URIDataSource(
uri="huggingface://datasets/llamastack/bfcl_v3?split=train",

5
llama_stack/templates/open-benchmark/run.yaml
View file

@ -164,35 +164,30 @@ datasets:
uri: huggingface://datasets/llamastack/simpleqa?split=train
metadata: {}
dataset_id: simpleqa
provider_id: huggingface
- purpose: eval/messages-answer
source:
type: uri
uri: huggingface://datasets/llamastack/mmlu_cot?split=test&name=all
metadata: {}
dataset_id: mmlu_cot
provider_id: huggingface
- purpose: eval/messages-answer
source:
type: uri
uri: huggingface://datasets/llamastack/gpqa_0shot_cot?split=test&name=gpqa_main
metadata: {}
dataset_id: gpqa_cot
provider_id: huggingface
- purpose: eval/messages-answer
source:
type: uri
uri: huggingface://datasets/llamastack/math_500?split=test
metadata: {}
dataset_id: math_500
provider_id: huggingface
- purpose: eval/messages-answer
source:
type: uri
uri: huggingface://datasets/llamastack/bfcl_v3?split=train
metadata: {}
dataset_id: bfcl
provider_id: huggingface
scoring_fns: []
benchmarks:
- dataset_id: simpleqa