forked from phoenix-oss/llama-stack-mirror

Composable building blocks to build Llama Apps

Find a file

Ihar Hrachyshka aca82df7ed fix: Multiple fixes for server shutdown (fix lifespan handling; fix handling CancelledError when raised by provider; let uvicorn handle signals) (#1495 ) # What does this PR do? If implementation raises CancelledError (e.g. when it runs its own async loop for jobs), the main server shutdown handler gets confused and doesn't attempt to shut down the main loop tasks. While at it, also fixing the following failure when this happens: ``` UnboundLocalError: cannot access local variable 'loop' where it is not associated with a value ``` Shutdown handlers were not running because lifespan logic was broken since ~Oct 2024. Fixed that too and enforcing `lifespan` now (making sure server will crash when it fails to interact with app through middleware). [//]: # (If resolving an issue, uncomment and update the line below) [//]: # (Closes #[issue-number]) ## Test Plan Spotted while working on https://github.com/meta-llama/llama-stack/pull/1437 One way to trigger it without the PR above is to add `raise CancelledError` in any of the running providers' `shutdown` methods; then `kill -INT <pid>` the server process. Validated this with the following test patch: ``` diff --git a/llama_stack/distribution/server/server.py b/llama_stack/distribution/server/server.py index b85c463a..10dad83e 100644 --- a/llama_stack/distribution/server/server.py +++ b/llama_stack/distribution/server/server.py @@ -174,6 +174,7 @@ def handle_signal(app, signum, _) -> None: except asyncio.CancelledError: pass finally: + logger.info("Stopping event loop") loop.stop() loop = asyncio.get_running_loop() diff --git a/llama_stack/providers/inline/post_training/torchtune/post_training.py b/llama_stack/providers/inline/post_training/torchtune/post_training.py index b837362d..163f43d8 100644 --- a/llama_stack/providers/inline/post_training/torchtune/post_training.py +++ b/llama_stack/providers/inline/post_training/torchtune/post_training.py @@ -3,6 +3,7 @@ # # This source code is licensed under the terms described in the LICENSE file in # the root directory of this source tree. +import asyncio from datetime import datetime from typing import Any, Dict, Optional @@ -43,6 +44,9 @@ class TorchtunePostTrainingImpl: self.jobs = {} self.checkpoints_dict = {} + async def shutdown(self) -> None: + raise asyncio.CancelledError("Shutdown") + async def supervised_fine_tune( self, job_uuid: str, ``` Without the fix: ``` INFO: Uvicorn running on http://['::', '0.0.0.0']:8321 (Press CTRL+C to quit) INFO: Shutting down INFO: Finished server process [52099] INFO 2025-03-07 23:25:33,548 __main__:143 server: Received signal SIGINT (2). Exiting gracefully... INFO 2025-03-07 23:25:33,550 __main__:150 server: Shutting down DatasetsRoutingTable INFO 2025-03-07 23:25:33,551 __main__:177 server: Stopping event loop ERROR 2025-03-07 23:25:33,552 asyncio:1785 uncategorized: unhandled exception during asyncio.run() shutdown task: <Task finished name='Task-12' coro=<handle_signal.<locals>.shutdown() done, defined at /home/ec2-user/src/llama-stack/schedule/llama_stack/distribution/server/server.py:145> exception=UnboundLocalError("cannot access local variable 'loop' where it is not associated with a value")> ╭───────────────────────────────────── Traceback (most recent call last) ─────────────────────────────────────╮ │ /home/ec2-user/src/llama-stack/schedule/llama_stack/distribution/server/server.py:178 in shutdown │ │ │ │ 175 │ │ │ pass │ │ 176 │ │ finally: │ │ 177 │ │ │ logger.info("Stopping event loop") │ │ ❱ 178 │ │ │ loop.stop() │ │ 179 │ │ │ 180 │ loop = asyncio.get_running_loop() │ │ 181 │ loop.create_task(shutdown()) │ ╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────╯ UnboundLocalError: cannot access local variable 'loop' where it is not associated with a value ``` With the fix, now seeing the following messages when the server is killed: ``` INFO: Uvicorn running on http://['::', '0.0.0.0']:8321 (Press CTRL+C to quit) INFO: Shutting down INFO: Finished server process [50836] INFO 2025-03-07 23:20:35,182 __main__:143 server: Received signal SIGINT (2). Exiting gracefully... INFO 2025-03-07 23:20:35,184 __main__:149 server: Shutting down DatasetsRoutingTable ERROR 2025-03-07 23:20:35,185 __main__:158 server: Failed to shutdown DatasetsRoutingTable: {CancelledError()} ╭───────────────────────────────────── Traceback (most recent call last) ─────────────────────────────────────╮ │ /usr/lib64/python3.11/asyncio/tasks.py:476 in wait_for │ │ │ │ 473 │ try: │ │ 474 │ │ # wait until the future completes or the timeout │ │ 475 │ │ try: │ │ ❱ 476 │ │ │ await waiter │ │ 477 │ │ except exceptions.CancelledError: │ │ 478 │ │ │ if fut.done(): │ │ 479 │ │ │ │ return fut.result() │ ╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────╯ CancelledError During handling of the above exception, another exception occurred: ╭───────────────────────────────────── Traceback (most recent call last) ─────────────────────────────────────╮ │ /home/ec2-user/src/llama-stack/schedule/llama_stack/distribution/server/server.py:152 in shutdown │ │ │ │ 149 │ │ │ logger.info("Shutting down %s", impl_name) │ │ 150 │ │ │ try: │ │ 151 │ │ │ │ if hasattr(impl, "shutdown"): │ │ ❱ 152 │ │ │ │ │ await asyncio.wait_for(impl.shutdown(), timeout=5) │ │ 153 │ │ │ │ else: │ │ 154 │ │ │ │ │ logger.warning("No shutdown method for %s", impl_name) │ │ 155 │ │ │ except asyncio.TimeoutError: │ │ │ │ /usr/lib64/python3.11/asyncio/tasks.py:479 in wait_for │ │ │ │ 476 │ │ │ await waiter │ │ 477 │ │ except exceptions.CancelledError: │ │ 478 │ │ │ if fut.done(): │ │ ❱ 479 │ │ │ │ return fut.result() │ │ 480 │ │ │ else: │ │ 481 │ │ │ │ fut.remove_done_callback(cb) │ │ 482 │ │ │ │ # We must ensure that the task is not running │ │ │ │ /home/ec2-user/src/llama-stack/schedule/llama_stack/distribution/routers/routing_tables.py:131 in shutdown │ │ │ │ 128 │ │ │ elif api == Api.tool_runtime: │ │ 129 │ │ │ │ p.tool_store = self │ │ 130 │ │ │ ❱ 131 │ async def shutdown(self) -> None: │ │ 132 │ │ for p in self.impls_by_provider_id.values(): │ │ 133 │ │ │ await p.shutdown() │ │ 134 │ ╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────╯ CancelledError INFO 2025-03-07 23:20:35,295 __main__:149 server: Shutting down DatasetIORouter INFO 2025-03-07 23:20:35,296 __main__:149 server: Shutting down ScoringFunctionsRoutingTable INFO 2025-03-07 23:20:35,297 __main__:149 server: Shutting down ScoringRouter INFO 2025-03-07 23:20:35,298 __main__:149 server: Shutting down ModelsRoutingTable INFO 2025-03-07 23:20:35,299 __main__:149 server: Shutting down InferenceRouter INFO 2025-03-07 23:20:35,300 __main__:149 server: Shutting down ShieldsRoutingTable INFO 2025-03-07 23:20:35,300 __main__:149 server: Shutting down SafetyRouter INFO 2025-03-07 23:20:35,301 __main__:149 server: Shutting down VectorDBsRoutingTable INFO 2025-03-07 23:20:35,302 __main__:149 server: Shutting down VectorIORouter INFO 2025-03-07 23:20:35,303 __main__:149 server: Shutting down ToolGroupsRoutingTable INFO 2025-03-07 23:20:35,304 __main__:149 server: Shutting down ToolRuntimeRouter INFO 2025-03-07 23:20:35,304 __main__:149 server: Shutting down MetaReferenceAgentsImpl INFO 2025-03-07 23:20:35,305 __main__:149 server: Shutting down TelemetryAdapter INFO 2025-03-07 23:20:35,306 __main__:149 server: Shutting down TorchtunePostTrainingImpl ERROR 2025-03-07 23:20:35,307 __main__:158 server: Failed to shutdown TorchtunePostTrainingImpl: {CancelledError('Shutdown')} ╭───────────────────────────────────── Traceback (most recent call last) ─────────────────────────────────────╮ │ /home/ec2-user/src/llama-stack/schedule/llama_stack/distribution/server/server.py:152 in shutdown │ │ │ │ 149 │ │ │ logger.info("Shutting down %s", impl_name) │ │ 150 │ │ │ try: │ │ 151 │ │ │ │ if hasattr(impl, "shutdown"): │ │ ❱ 152 │ │ │ │ │ await asyncio.wait_for(impl.shutdown(), timeout=5) │ │ 153 │ │ │ │ else: │ │ 154 │ │ │ │ │ logger.warning("No shutdown method for %s", impl_name) │ │ 155 │ │ │ except asyncio.TimeoutError: │ │ │ │ /usr/lib64/python3.11/asyncio/tasks.py:489 in wait_for │ │ │ │ 486 │ │ │ │ raise │ │ 487 │ │ │ │ 488 │ │ if fut.done(): │ │ ❱ 489 │ │ │ return fut.result() │ │ 490 │ │ else: │ │ 491 │ │ │ fut.remove_done_callback(cb) │ │ 492 │ │ │ # We must ensure that the task is not running │ │ │ │ /home/ec2-user/src/llama-stack/schedule/llama_stack/providers/inline/post_training/torchtune/post_training. │ │ py:48 in shutdown │ │ │ │ 45 │ │ self.checkpoints_dict = {} │ │ 46 │ │ │ 47 │ async def shutdown(self) -> None: │ │ ❱ 48 │ │ raise asyncio.CancelledError("Shutdown") │ │ 49 │ │ │ 50 │ async def supervised_fine_tune( │ │ 51 │ │ self, │ ╰─────────────────────────────────────────────────────────────────────────────────────────────────────────────╯ CancelledError: Shutdown INFO 2025-03-07 23:20:35,352 __main__:149 server: Shutting down BenchmarksRoutingTable INFO 2025-03-07 23:20:35,353 __main__:149 server: Shutting down EvalRouter INFO 2025-03-07 23:20:35,354 __main__:149 server: Shutting down DistributionInspectImpl INFO 2025-03-07 23:20:35,355 __main__:177 server: Stopping event loop Traceback (most recent call last): File "<frozen runpy>", line 198, in _run_module_as_main File "<frozen runpy>", line 88, in _run_code File "/home/ec2-user/src/llama-stack/schedule/llama_stack/distribution/server/server.py", line 488, in <module> main() File "/home/ec2-user/src/llama-stack/schedule/llama_stack/distribution/server/server.py", line 476, in main uvicorn.run(*uvicorn_config) File "/home/ec2-user/src/llama-stack/schedule/venv/lib64/python3.11/site-packages/uvicorn/main.py", line 579, in run server.run() File "/home/ec2-user/src/llama-stack/schedule/venv/lib64/python3.11/site-packages/uvicorn/server.py", line 66, in run return asyncio.run(self.serve(sockets=sockets)) ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^ File "/usr/lib64/python3.11/asyncio/runners.py", line 189, in run with Runner(debug=debug) as runner: File "/usr/lib64/python3.11/asyncio/runners.py", line 63, in __exit__ self.close() File "/usr/lib64/python3.11/asyncio/runners.py", line 71, in close _cancel_all_tasks(loop) File "/usr/lib64/python3.11/asyncio/runners.py", line 201, in _cancel_all_tasks loop.run_until_complete(tasks.gather(to_cancel, return_exceptions=True)) File "/usr/lib64/python3.11/asyncio/base_events.py", line 652, in run_until_complete raise RuntimeError('Event loop stopped before Future completed.') RuntimeError: Event loop stopped before Future completed. ++ error_handler 104 ++ echo 'Error occurred in script at line: 104' Error occurred in script at line: 104 ++ exit 1 ``` With all patches included, the shutdown now looks as follows: ``` $ kill -INT $(ps ax \| grep llama_stack.distribution.server.server \| grep -v nvim \| awk -e '{print $1}' \| sort \| head -n 1) ``` ``` 20:56:09.308 [START] INFO: Uvicorn running on http://['::', '0.0.0.0']:8321 (Press CTRL+C to quit) INFO: Shutting down INFO: Waiting for application shutdown. INFO 2025-03-10 20:56:43,961 __main__:140 server: Shutting down INFO 2025-03-10 20:56:43,962 __main__:124 server: Shutting down DatasetsRoutingTable INFO 2025-03-10 20:56:43,964 __main__:124 server: Shutting down DatasetIORouter INFO 2025-03-10 20:56:43,965 __main__:124 server: Shutting down ScoringFunctionsRoutingTable INFO 2025-03-10 20:56:43,966 __main__:124 server: Shutting down ScoringRouter INFO 2025-03-10 20:56:43,967 __main__:124 server: Shutting down ModelsRoutingTable INFO 2025-03-10 20:56:43,968 __main__:124 server: Shutting down InferenceRouter INFO 2025-03-10 20:56:43,969 __main__:124 server: Shutting down ShieldsRoutingTable INFO 2025-03-10 20:56:43,971 __main__:124 server: Shutting down SafetyRouter INFO 2025-03-10 20:56:43,972 __main__:124 server: Shutting down VectorDBsRoutingTable INFO 2025-03-10 20:56:43,973 __main__:124 server: Shutting down VectorIORouter INFO 2025-03-10 20:56:43,974 __main__:124 server: Shutting down ToolGroupsRoutingTable INFO 2025-03-10 20:56:43,975 __main__:124 server: Shutting down ToolRuntimeRouter INFO 2025-03-10 20:56:43,976 __main__:124 server: Shutting down MetaReferenceAgentsImpl INFO 2025-03-10 20:56:43,977 __main__:124 server: Shutting down TelemetryAdapter INFO 2025-03-10 20:56:43,978 __main__:124 server: Shutting down TorchtunePostTrainingImpl WARNING 2025-03-10 20:56:43,979 __main__:129 server: No shutdown method for TorchtunePostTrainingImpl INFO 2025-03-10 20:56:43,979 __main__:124 server: Shutting down BenchmarksRoutingTable INFO 2025-03-10 20:56:43,980 __main__:124 server: Shutting down EvalRouter INFO 2025-03-10 20:56:43,981 __main__:124 server: Shutting down DistributionInspectImpl INFO: Application shutdown complete. INFO: Finished server process [33862] ``` [//]: # (## Documentation) --------- Signed-off-by: Ihar Hrachyshka <ihar.hrachyshka@gmail.com>		2025-03-11 10:30:55 -07:00
.cursor/rules	test: first unit test for resolver (#1475 )	2025-03-07 10:20:51 -08:00
.github	fix: Use `--with-editable` to capture accurate code coverage reporting (#1532 )	2025-03-10 19:30:28 -04:00
distributions	fix: revert to using faiss for ollama distro (#1530 )	2025-03-10 16:15:17 -07:00
docs	docs: Small nits in llama CLI reference (#1542 )	2025-03-11 10:12:18 -07:00
llama_stack	fix: Multiple fixes for server shutdown (fix lifespan handling; fix handling CancelledError when raised by provider; let uvicorn handle signals) (#1495 )	2025-03-11 10:30:55 -07:00
rfcs	chore: remove straggler references to llama-models (#1345 )	2025-03-01 14:26:03 -08:00
scripts	fix: clean up detailed history for CHANGELOG (#1494 )	2025-03-07 14:03:54 -08:00
tests	docs: improve integration test doc (#1502 )	2025-03-10 15:50:46 -07:00
.gitignore	chore: Display code coverage for unit tests in PR builds (#1512 )	2025-03-10 16:27:33 -04:00
.pre-commit-config.yaml	fix: make sure readthedocs is triggered if pyproject.toml is updated	2025-03-08 23:05:10 -08:00
.python-version	build: hint on Python version for uv venv (#1172 )	2025-02-25 10:37:45 -05:00
.readthedocs.yaml	first version of readthedocs (#278 )	2024-10-22 10:15:58 +05:30
CHANGELOG.md	docs: Add v0.1.6 release notes to changelog (#1506 )	2025-03-08 16:20:08 -08:00
CODE_OF_CONDUCT.md	Initial commit	2024-07-23 08:32:33 -07:00
CONTRIBUTING.md	docs: Remove duplicate docs on api docs generator (#1534 )	2025-03-11 10:01:46 -07:00
LICENSE	Update LICENSE (#47 )	2024-08-29 07:39:50 -07:00
MANIFEST.in	build: include .md (#1482 )	2025-03-07 12:10:52 -08:00
pyproject.toml	build: revamp "test" dependencies from pyproject (#1468 )	2025-03-10 15:43:16 -07:00
README.md	chore: remove dependency on llama_models completely (#1344 )	2025-03-01 12:48:08 -08:00
requirements.txt	fix: include jinja2 as a core llama-stack dependency (#1529 )	2025-03-10 14:59:11 -07:00
SECURITY.md	Create SECURITY.md	2024-10-08 13:30:40 -04:00
uv.lock	build: revamp "test" dependencies from pyproject (#1468 )	2025-03-10 15:43:16 -07:00

README.md

Llama Stack

Quick Start | Documentation | Colab Notebook

Llama Stack standardizes the core building blocks that simplify AI application development. It codifies best practices across the Llama ecosystem. More specifically, it provides

Unified API layer for Inference, RAG, Agents, Tools, Safety, Evals, and Telemetry.
Plugin architecture to support the rich ecosystem of different API implementations in various environments, including local development, on-premises, cloud, and mobile.
Prepackaged verified distributions which offer a one-stop solution for developers to get started quickly and reliably in any environment.
Multiple developer interfaces like CLI and SDKs for Python, Typescript, iOS, and Android.
Standalone applications as examples for how to build production-grade AI applications with Llama Stack.

Llama Stack Benefits

Flexible Options: Developers can choose their preferred infrastructure without changing APIs and enjoy flexible deployment choices.
Consistent Experience: With its unified APIs, Llama Stack makes it easier to build, test, and deploy AI applications with consistent application behavior.
Robust Ecosystem: Llama Stack is already integrated with distribution partners (cloud providers, hardware vendors, and AI-focused companies) that offer tailored infrastructure, software, and services for deploying Llama models.

By reducing friction and complexity, Llama Stack empowers developers to focus on what they do best: building transformative generative AI applications.

API Providers

Here is a list of the various API providers and available distributions that can help developers get started easily with Llama Stack.

API Provider Builder	Environments	Agents	Inference	Memory	Safety	Telemetry
Meta Reference	Single Node	✅	✅	✅	✅	✅
SambaNova	Hosted		✅
Cerebras	Hosted		✅
Fireworks	Hosted	✅	✅	✅
AWS Bedrock	Hosted		✅		✅
Together	Hosted	✅	✅		✅
Groq	Hosted		✅
Ollama	Single Node		✅
TGI	Hosted and Single Node		✅
NVIDIA NIM	Hosted and Single Node		✅
Chroma	Single Node			✅
PG Vector	Single Node			✅
PyTorch ExecuTorch	On-device iOS	✅	✅
vLLM	Hosted and Single Node		✅

Distributions

A Llama Stack Distribution (or "distro") is a pre-configured bundle of provider implementations for each API component. Distributions make it easy to get started with a specific deployment scenario - you can begin with a local development setup (eg. ollama) and seamlessly transition to production (eg. Fireworks) without changing your application code. Here are some of the distributions we support:

Distribution	Llama Stack Docker	Start This Distribution
Meta Reference	llamastack/distribution-meta-reference-gpu	Guide
Meta Reference Quantized	llamastack/distribution-meta-reference-quantized-gpu	Guide
SambaNova	llamastack/distribution-sambanova	Guide
Cerebras	llamastack/distribution-cerebras	Guide
Ollama	llamastack/distribution-ollama	Guide
TGI	llamastack/distribution-tgi	Guide
Together	llamastack/distribution-together	Guide
Fireworks	llamastack/distribution-fireworks	Guide
vLLM	llamastack/distribution-remote-vllm	Guide

Installation

You have two ways to install this repository:

Install as a package: You can install the repository directly from PyPI by running the following command:
```
pip install llama-stack
```
Install from source: If you prefer to install from the source code, we recommend using uv. Then, run the following commands:
```
 git clone git@github.com:meta-llama/llama-stack.git
 cd llama-stack

 uv sync
 uv pip install -e .
```

Documentation

Please checkout our Documentation page for more details.

CLI references
- llama (server-side) CLI Reference: Guide for using the llama CLI to work with Llama models (download, study prompts), and building/starting a Llama Stack distribution.
- llama (client-side) CLI Reference: Guide for using the llama-stack-client CLI, which allows you to query information about the distribution.
Getting Started
- Quick guide to start a Llama Stack server.
- Jupyter notebook to walk-through how to use simple text and vision inference llama_stack_client APIs
- The complete Llama Stack lesson Colab notebook of the new Llama 3.2 course on Deeplearning.ai.
- A Zero-to-Hero Guide that guide you through all the key components of llama stack with code samples.
Contributing
- Adding a new API Provider to walk-through how to add a new API provider.

Llama Stack Client SDKs

Language	Client SDK	Package
Python	llama-stack-client-python
Swift	llama-stack-client-swift
Typescript	llama-stack-client-typescript
Kotlin	llama-stack-client-kotlin

Check out our client SDKs for connecting to a Llama Stack server in your preferred language, you can choose from python, typescript, swift, and kotlin programming languages to quickly build your applications.

You can find more example scripts with client SDKs to talk with the Llama Stack server in our llama-stack-apps repo.