* Significantly simpler and malleable test setup
* convert memory tests
* refactor fixtures and add support for composable fixtures
* Fix memory to use the newer fixture organization
* Get agents tests working
* Safety tests work
* yet another refactor to make this more general
now it accepts --inference-model, --safety-model options also
* get multiple providers working for meta-reference (for inference + safety)
* Add README.md
---------
Co-authored-by: Ashwin Bharambe <ashwin@meta.com>
* Added hadamard transform for spinquant
* Changed from config to model_args
* Added an assertion for model args
* Use enum.value to check against str
* pre-commit
---------
Co-authored-by: Sachin Mehta <sacmehta@fb.com>
Co-authored-by: Ashwin Bharambe <ashwin.bharambe@gmail.com>
Added support for structured output in the API and added a reference implementation for meta-reference.
A few notes:
* Two formats are specified in the API: Json schema and EBNF based grammar
* Implementation only supports Json for now
We use lm-format-enhancer to provide the implementation right now but may change this especially because BNF grammars aren't supported by that library.
Fireworks has support for structured output and Together has limited supported for it too. Subsequent PRs will add these changes. We would like all our inference providers to provide structured output for llama models since it is an extremely important and highly sought-after need by the developers.
PR #201 had made several changes while trying to fix issues with getting the stream=False branches of inference and agents API working. As part of this, it made a change which was slightly gratuitous. Namely, making chat_completion() and brethren "def" instead of "async def".
The rationale was that this allowed the user (within llama-stack) of this to use it as:
```
async for chunk in api.chat_completion(params)
```
However, it causes unnecessary confusion for several folks. Given that clients (e.g., llama-stack-apps) anyway use the SDK methods (which are completely isolated) this choice was not ideal. Let's revert back so the call now looks like:
```
async for chunk in await api.chat_completion(params)
```
Bonus: Added a completion() implementation for the meta-reference provider. Technically should have been another PR :)
I only tested with "on-the-fly" bf16 -> fp8 conversion, not the "load
from fp8" codepath.
YAML I tested with:
```
providers:
- provider_id: quantized
provider_type: meta-reference-quantized
config:
model: Llama3.1-8B-Instruct
quantization:
type: fp8
```
This PR makes several core changes to the developer experience surrounding Llama Stack.
Background: PR #92 introduced the notion of "routing" to the Llama Stack. It introduces three object types: (1) models, (2) shields and (3) memory banks. Each of these objects can be associated with a distinct provider. So you can get model A to be inferenced locally while model B, C can be inference remotely (e.g.)
However, this had a few drawbacks:
you could not address the provider instances -- i.e., if you configured "meta-reference" with a given model, you could not assign an identifier to this instance which you could re-use later.
the above meant that you could not register a "routing_key" (e.g. model) dynamically and say "please use this existing provider I have already configured" for a new model.
the terms "routing_table" and "routing_key" were exposed directly to the user. in my view, this is way too much overhead for a new user (which almost everyone is.) people come to the stack wanting to do ML and encounter a completely unexpected term.
What this PR does: This PR structures the run config with only a single prominent key:
- providers
Providers are instances of configured provider types. Here's an example which shows two instances of the remote::tgi provider which are serving two different models.
providers:
inference:
- provider_id: foo
provider_type: remote::tgi
config: { ... }
- provider_id: bar
provider_type: remote::tgi
config: { ... }
Secondly, the PR adds dynamic registration of { models | shields | memory_banks } to the API surface. The distribution still acts like a "routing table" (as previously) except that it asks the backing providers for a listing of these objects. For example it asks a TGI or Ollama inference adapter what models it is serving. Only the models that are being actually served can be requested by the user for inference. Otherwise, the Stack server will throw an error.
When dynamically registering these objects, you can use the provider IDs shown above. Info about providers can be obtained using the Api.inspect set of endpoints (/providers, /routes, etc.)
The above examples shows the correspondence between inference providers and models registry items. Things work similarly for the safety <=> shields and memory <=> memory_banks pairs.
Registry: This PR also makes it so that Providers need to implement additional methods for registering and listing objects. For example, each Inference provider is now expected to implement the ModelsProtocolPrivate protocol (naming is not great!) which consists of two methods
register_model
list_models
The goal is to inform the provider that a certain model needs to be supported so the provider can make any relevant backend changes if needed (or throw an error if the model cannot be supported.)
There are many other cleanups included some of which are detailed in a follow-up comment.
We should use Inference APIs to execute Llama Guard instead of directly needing to use HuggingFace modeling related code. The actual inference consideration is handled by Inference.
I got this error message and tried to the run the command presented
and it didn't work. The model needs to be give with `--model-id`
instead of as a positional argument.
Signed-off-by: Russell Bryant <rbryant@redhat.com>
Test Plan:
First, start a TGI container with `meta-llama/Llama-Guard-3-8B` model
serving on port 5099. See https://github.com/meta-llama/llama-stack/pull/53 and its
description for how.
Then run llama-stack with the following run config:
```
image_name: safety
docker_image: null
conda_env: safety
apis_to_serve:
- models
- inference
- shields
- safety
api_providers:
inference:
providers:
- remote::tgi
safety:
providers:
- meta-reference
telemetry:
provider_id: meta-reference
config: {}
routing_table:
inference:
- provider_id: remote::tgi
config:
url: http://localhost:5099
api_token: null
hf_endpoint_name: null
routing_key: Llama-Guard-3-8B
safety:
- provider_id: meta-reference
config:
llama_guard_shield:
model: Llama-Guard-3-8B
excluded_categories: []
disable_input_check: false
disable_output_check: false
prompt_guard_shield: null
routing_key: llama_guard
```
Now simply run `python -m llama_stack.apis.safety.client localhost
<port>` and check that the llama_guard shield calls run correctly. (The
injection_shield calls fail as expected since we have not set up a
router for them.)
This is yet another of those large PRs (hopefully we will have less and less of them as things mature fast). This one introduces substantial improvements and some simplifications to the stack.
Most important bits:
* Agents reference implementation now has support for session / turn persistence. The default implementation uses sqlite but there's also support for using Redis.
* We have re-architected the structure of the Stack APIs to allow for more flexible routing. The motivating use cases are:
- routing model A to ollama and model B to a remote provider like Together
- routing shield A to local impl while shield B to a remote provider like Bedrock
- routing a vector memory bank to Weaviate while routing a keyvalue memory bank to Redis
* Support for provider specific parameters to be passed from the clients. A client can pass data using `x_llamastack_provider_data` parameter which can be type-checked and provided to the Adapter implementations.
