# What does this PR do?
This PR adds support for NVIDIA's NeMo Customizer API to the Llama Stack
post-training module. The integration enables users to fine-tune models
using NVIDIA's cloud-based customization service through a consistent
Llama Stack interface.
[//]: # (If resolving an issue, uncomment and update the line below)
[//]: # (Closes #[issue-number])
## Test Plan
[Describe the tests you ran to verify your changes with result
summaries. *Provide clear instructions so the plan can be easily
re-executed.*]
Yet to be done
Things pending under this PR:
- [x] Integration of fine-tuned model(new checkpoint) for inference with
nvidia llm distribution
- [x] distribution integration of API
- [x] Add test cases for customizer(In Progress)
- [x] Documentation
```
LLAMA_STACK_BASE_URL=http://localhost:5002 pytest -v tests/client-sdk/post_training/test_supervised_fine_tuning.py
============================================================================================================================================================================ test session starts =============================================================================================================================================================================
platform linux -- Python 3.10.0, pytest-8.3.4, pluggy-1.5.0 -- /home/ubuntu/llama-stack/.venv/bin/python
cachedir: .pytest_cache
metadata: {'Python': '3.10.0', 'Platform': 'Linux-6.8.0-1021-gcp-x86_64-with-glibc2.35', 'Packages': {'pytest': '8.3.4', 'pluggy': '1.5.0'}, 'Plugins': {'nbval': '0.11.0', 'metadata': '3.1.1', 'anyio': '4.8.0', 'html': '4.1.1', 'asyncio': '0.25.3'}}
rootdir: /home/ubuntu/llama-stack
configfile: pyproject.toml
plugins: nbval-0.11.0, metadata-3.1.1, anyio-4.8.0, html-4.1.1, asyncio-0.25.3
asyncio: mode=strict, asyncio_default_fixture_loop_scope=None
collected 2 items
tests/client-sdk/post_training/test_supervised_fine_tuning.py::test_post_training_provider_registration[txt=8B] PASSED [ 50%]
tests/client-sdk/post_training/test_supervised_fine_tuning.py::test_list_training_jobs[txt=8B] PASSED [100%]
======================================================================================================================================================================== 2 passed, 1 warning in 0.10s ========================================================================================================================================================================
```
cc: @mattf @dglogo @sumitb
---------
Co-authored-by: Ubuntu <ubuntu@llama-stack-customizer-dev-inst-2tx95fyisatvlic4we8hidx5tfj.us-central1-a.c.brevdevprod.internal>
Each model known to the system has two identifiers:
- the `provider_resource_id` (what the provider calls it) -- e.g.,
`accounts/fireworks/models/llama-v3p1-8b-instruct`
- the `identifier` (`model_id`) under which it is registered and gets
routed to the appropriate provider.
We have so far used the HuggingFace repo alias as the standardized
identifier you can use to refer to the model. So in the above example,
we'd use `meta-llama/Llama-3.1-8B-Instruct` as the name under which it
gets registered. This makes it convenient for users to refer to these
models across providers.
However, we forgot to register the _actual_ provider model ID also. You
should be able to route via `provider_resource_id` also, of course.
This change fixes this (somewhat grave) omission.
*Note*: this change is additive -- more aliases work now compared to
before.
## Test Plan
Run the following for distro=(ollama fireworks together)
```
LLAMA_STACK_CONFIG=$distro \
pytest -s -v tests/client-sdk/inference/test_text_inference.py \
--inference-model=meta-llama/Llama-3.1-8B-Instruct --vision-inference-model=""
```
# What does this PR do?
add /v1/inference/embeddings implementation to NVIDIA provider
**open topics** -
- *asymmetric models*. NeMo Retriever includes asymmetric models, which
are models that embed differently depending on if the input is destined
for storage or lookup against storage. the /v1/inference/embeddings api
does not allow the user to indicate the type of embedding to perform.
see https://github.com/meta-llama/llama-stack/issues/934
- *truncation*. embedding models typically have a limited context
window, e.g. 1024 tokens is common though newer models have 8k windows.
when the input is larger than this window the endpoint cannot perform
its designed function. two options: 0. return an error so the user can
reduce the input size and retry; 1. perform truncation for the user and
proceed (common strategies are left or right truncation). many users
encounter context window size limits and will struggle to write reliable
programs. this struggle is especially acute without access to the
model's tokenizer. the /v1/inference/embeddings api does not allow the
user to delegate truncation policy. see
https://github.com/meta-llama/llama-stack/issues/933
- *dimensions*. "Matryoshka" embedding models are available. they allow
users to control the number of embedding dimensions the model produces.
this is a critical feature for managing storage constraints. embeddings
of 1024 dimensions what achieve 95% recall for an application may not be
worth the storage cost if a 512 dimensions can achieve 93% recall.
controlling embedding dimensions allows applications to determine their
recall and storage tradeoffs. the /v1/inference/embeddings api does not
allow the user to control the output dimensions. see
https://github.com/meta-llama/llama-stack/issues/932
## Test Plan
- `llama stack run llama_stack/templates/nvidia/run.yaml`
- `LLAMA_STACK_BASE_URL=http://localhost:8321 pytest -v
tests/client-sdk/inference/test_embedding.py --embedding-model
baai/bge-m3`
## Sources
Please link relevant resources if necessary.
## Before submitting
- [ ] This PR fixes a typo or improves the docs (you can dismiss the
other checks if that's the case).
- [x] Ran pre-commit to handle lint / formatting issues.
- [x] Read the [contributor
guideline](https://github.com/meta-llama/llama-stack/blob/main/CONTRIBUTING.md),
Pull Request section?
- [ ] Updated relevant documentation.
- [x] Wrote necessary unit or integration tests.
---------
Co-authored-by: Ashwin Bharambe <ashwin.bharambe@gmail.com>
# What does this PR do?
Before this change, `distro_codegen.py` would only work if the user
manually installed multiple provider-specific dependencies (see #1122).
Now, users can run `distro_codegen.py` without any provider-specific
dependencies because we avoid importing the entire provider
implementations just to get the config needed to build the provider
template.
Concretely, this mostly means moving the
MODEL_ALIASES (and related variants) definitions to a new models.py
class within the provider implementation for those providers that
require additional dependencies. It also meant moving a couple of
imports from top-level imports to inside `get_adapter_impl` for some
providers, which follows the pattern used by multiple existing
providers.
To ensure we don't regress and accidentally add new imports that cause
distro_codegen.py to fail, the stubbed-in pre-commit hook for
distro_codegen.py was uncommented and slightly tweaked to run via `uv
run python ...` to ensure it runs with only the project's default
dependencies and to run automatically instead of manually.
Lastly, this updates distro_codegen.py itself to keep track of paths it
might have changed and to only `git diff` those specific paths when
checking for changed files instead of doing a diff on the entire working
tree. The latter was overly broad and would require a user have no other
unstaged changes in their working tree, even if those unstaged changes
were unrelated to generated code. Now it only flags uncommitted changes
for paths distro_codegen.py actually writes to.
Our generated code was also out-of-date, presumably because of these
issues, so this commit also has some updates to the generated code
purely because it was out of sync, and the pre-commit hook now enforces
things to be updated.
(Closes#1122)
## Test Plan
I manually tested distro_codegen.py and the pre-commit hook to verify
those work as expected, flagging any uncommited changes and catching any
imports that attempt to pull in provider-specific dependencies.
However, I do not have valid api keys to the impacted provider
implementations, and am unable to easily run the inference tests against
each changed provider. There are no functional changes to the provider
implementations here, but I'd appreciate a second set of eyes on the
changed import statements and moving of MODEL_ALIASES type code to a
separate models.py to ensure I didn't make any obvious errors.
---------
Signed-off-by: Ben Browning <bbrownin@redhat.com>
Co-authored-by: Ashwin Bharambe <ashwin.bharambe@gmail.com>
llama-models should have extremely minimal cruft. Its sole purpose
should be didactic -- show the simplest implementation of the llama
models and document the prompt formats, etc.
This PR is the complement to
https://github.com/meta-llama/llama-models/pull/279
## Test Plan
Ensure all `llama` CLI `model` sub-commands work:
```bash
llama model list
llama model download --model-id ...
llama model prompt-format -m ...
```
Ran tests:
```bash
cd tests/client-sdk
LLAMA_STACK_CONFIG=fireworks pytest -s -v inference/
LLAMA_STACK_CONFIG=fireworks pytest -s -v vector_io/
LLAMA_STACK_CONFIG=fireworks pytest -s -v agents/
```
Create a fresh venv `uv venv && source .venv/bin/activate` and run
`llama stack build --template fireworks --image-type venv` followed by
`llama stack run together --image-type venv` <-- the server runs
Also checked that the OpenAPI generator can run and there is no change
in the generated files as a result.
```bash
cd docs/openapi_generator
sh run_openapi_generator.sh
```
Lint check in main branch is failing. This fixes the lint check after we
moved to ruff in https://github.com/meta-llama/llama-stack/pull/921. We
need to move to a `ruff.toml` file as well as fixing and ignoring some
additional checks.
Signed-off-by: Yuan Tang <terrytangyuan@gmail.com>
It's a more generic term and applicable to alternatives of Docker, such
as Podman or other OCI-compliant technologies.
---------
Signed-off-by: Yuan Tang <terrytangyuan@gmail.com>
# What does this PR do?
- fixes to nvidia inference provider to account for strategy update
- update nvidia templates
## Test Plan
```
llama stack run ./llama_stack/templates/nvidia/run.yaml --port 5000
LLAMA_STACK_BASE_URL="http://localhost:5000" pytest -v tests/client-sdk/inference/test_inference.py --html=report.html --self-contained-html
```
<img width="1288" alt="image"
src="https://github.com/user-attachments/assets/d20f9aea-525e-47de-a5be-586e022e0d55"
/>
**NOTE**
- vision inference broken
- tool calling broken
- /completion broken
cc @mattf @cdgamarose-nv for improving NVIDIA inference adapter
## Sources
Please link relevant resources if necessary.
## Before submitting
- [ ] This PR fixes a typo or improves the docs (you can dismiss the
other checks if that's the case).
- [ ] Ran pre-commit to handle lint / formatting issues.
- [ ] Read the [contributor
guideline](https://github.com/meta-llama/llama-stack/blob/main/CONTRIBUTING.md),
Pull Request section?
- [ ] Updated relevant documentation.
- [ ] Wrote necessary unit or integration tests.
# What does this PR do?
adds nvidia template for creating a distribution using inference adapter
for NVIDIA NIMs.
## Test Plan
Please describe:
Build llama stack distribution for nvidia using the template, docker and
conda.
```bash
(.venv) local-cdgamarose@a4u8g-0006:~/llama-stack$ llama-stack-client configure --endpoint http://localhost:5000
Done! You can now use the Llama Stack Client CLI with endpoint http://localhost:5000
(.venv) local-cdgamarose@a4u8g-0006:~/llama-stack$ llama-stack-client models list
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━┓
┃ identifier ┃ provider_id ┃ provider_resource_id ┃ metadata ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━┩
│ Llama3.1-8B-Instruct │ nvidia │ meta/llama-3.1-8b-instruct │ {} │
│ meta-llama/Llama-3.2-3B-Instruct │ nvidia │ meta/llama-3.2-3b-instruct │ {} │
└──────────────────────────────────┴─────────────┴────────────────────────────┴──────────┘
(.venv) local-cdgamarose@a4u8g-0006:~/llama-stack$ llama-stack-client inference chat-completion --message "hello, write me a 2 sentence poem"
ChatCompletionResponse(
completion_message=CompletionMessage(
content='Here is a 2 sentence poem:\n\nThe sun sets slow and paints the sky, \nA gentle hue of pink that makes me sigh.',
role='assistant',
stop_reason='end_of_turn',
tool_calls=[]
),
logprobs=None
)
```
## Before submitting
- [ ] This PR fixes a typo or improves the docs (you can dismiss the
other checks if that's the case).
- [x] Ran pre-commit to handle lint / formatting issues.
- [x] Read the [contributor
guideline](https://github.com/meta-llama/llama-stack/blob/main/CONTRIBUTING.md),
Pull Request section?
- [x] Updated relevant documentation.
- [ ] Wrote necessary unit or integration tests.
---------
Co-authored-by: Matthew Farrellee <matt@cs.wisc.edu>
