### Context
This is the 1st of series PRs that integrate torchtune with llama-stack
as meta reference post-training implementation. For MVP, we will focus
on single device LoRA SFT.
Though this PR is still WIP, we want to get early feedback on the high
level design of this skeleton while still working on several details
### Scope
To limit the scope of this PR, we focus on the skeleton of the
implementation.
**What are included?**
- refine the post-training SFT apis
- skeleton of supervised_fine_tune implementation. We verified that we
can call the supervised_fine_tune API successfully from llama stack
client SDK (client side PR:
https://github.com/meta-llama/llama-stack-client-python/pull/51)
- a very basic single device LoRA training recipe based on torchtune
core components
- parity check with torchtune library and post training api unit test
**What are not includes?**
- implementation of other job management, get training artifacts apis
(separate PR)
- refactor the meta reference inference logic to support eval on
finetuned model (separate PR)
- several necessary functionality in the training recipe such as
logging, validation etc (separate PR)
- interop with telemetry for tracing and metrics logging, currently
temporarily log to local disk (separate PR)
### Testing
**e2e test**
Although we haven't added detailed testing and numerical parity check
with torchtune yet, we did a simple E2E test from client to server
1. setup server with` llama stack build --template
experimental-post-training --image-type conda` and `llama stack run
experimental-post-training `
2. On client, run `llama-stack-client --endpoint
http://devgpu018.nha2.facebook.com:5000 post_training
supervised_fine_tune`
3. Training finishes successfully. On server side, get the finetune
checkpoints under output dir. On client side, get the job uuid
server
<img width="1110" alt="Screenshot 2024-12-02 at 5 52 32 PM"
src="https://github.com/user-attachments/assets/b548eb90-7a9b-4edc-a858-ee237cc4361d">
client
<img width="807" alt="Screenshot 2024-12-02 at 5 52 37 PM"
src="https://github.com/user-attachments/assets/1138ffa8-4698-40fa-b190-3d7b99646838">
**parity check**
torchtune dataloader output and llama-stack post training dataloader
output are same
<img width="1116" alt="Screenshot 2024-12-04 at 8 18 46 PM"
src="https://github.com/user-attachments/assets/5e295cdc-4c24-4ea6-82c0-ca96ef1bd6ee">
torchtune LoRA SFT and llama-stack post training LoRA SFT on alpaca
dataset with llama3.2 3B instruct model are numerical match
<img width="860" alt="Screenshot 2024-12-04 at 8 17 01 PM"
src="https://github.com/user-attachments/assets/c05cf0a8-c674-4d2e-9f0a-c5d01b2dca99">
<img width="1049" alt="Screenshot 2024-12-04 at 8 17 06 PM"
src="https://github.com/user-attachments/assets/b911d4e2-e7b1-41a9-b62c-d75529b6d443">
**unit test **
![Uploading Screenshot 2024-12-09 at 1.35.10 PM.png…]()
# What does this PR do?
This PR moves all print statements to use logging. Things changed:
- Had to add `await start_trace("sse_generator")` to server.py to
actually get tracing working. else was not seeing any logs
- If no telemetry provider is provided in the run.yaml, we will write to
stdout
- by default, the logs are going to be in JSON, but we expose an option
to configure to output in a human readable way.
# What does this PR do?
This PR kills the notion of "pure passthrough" remote providers. You
cannot specify a single provider you must specify a whole distribution
(stack) as remote.
This PR also significantly fixes / upgrades testing infrastructure so
you can now test against a remotely hosted stack server by just doing
```bash
pytest -s -v -m remote test_agents.py \
--inference-model=Llama3.1-8B-Instruct --safety-shield=Llama-Guard-3-1B \
--env REMOTE_STACK_URL=http://localhost:5001
```
Also fixed `test_agents_persistence.py` (which was broken) and killed
some deprecated testing functions.
## Test Plan
All the tests.
Splits the meta-reference safety implementation into three distinct providers:
- inline::llama-guard
- inline::prompt-guard
- inline::code-scanner
Note that this PR is a backward incompatible change to the llama stack server. I have added deprecation_error field to ProviderSpec -- the server reads it and immediately barfs. This is used to direct the user with a specific message on what action to perform. An automagical "config upgrade" is a bit too much work to implement right now :/
(Note that we will be gradually prefixing all inline providers with inline:: -- I am only doing this for this set of new providers because otherwise existing configuration files will break even more badly.)
* persist registered objects with distribution
* linter fixes
* comment
* use annotate and field discriminator
* workign tests
* donot use global state
* precommit failures fixed
* add back Any
* fix imports
* remove unnecessary changes in ollama
* precommit failures fixed
* make kvstore configurable for dist and rename registry
* add comment about registry list return
* fix linter errors
* use registry to hydrate
* remove debug print
* linter fixes
* remove kvstore.db
* rename distribution_registry_store
---------
Co-authored-by: Dinesh Yeduguru <dineshyv@fb.com>
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.
