feat(providers): add NVIDIA Inference embedding provider and tests (#935)

# 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>

docs/source/distributions/remote_hosted_distro/nvidia.md

@@ -36,6 +36,7 @@ The following models are available by default:
 - `meta-llama/Llama-3.2-3B-Instruct (meta/llama-3.2-3b-instruct)`
 - `meta-llama/Llama-3.2-11B-Vision-Instruct (meta/llama-3.2-11b-vision-instruct)`
 - `meta-llama/Llama-3.2-90B-Vision-Instruct (meta/llama-3.2-90b-vision-instruct)`
+- `baai/bge-m3 (baai/bge-m3)`
 
 
 ### Prerequisite: API Keys

llama_stack/providers/remote/inference/nvidia/models.py

@@ -4,8 +4,10 @@
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
 
+from llama_stack.apis.models import ModelType
 from llama_stack.models.llama.datatypes import CoreModelId
 from llama_stack.providers.utils.inference.model_registry import (
+    ProviderModelEntry,
     build_hf_repo_model_entry,
 )
 
@@ -46,6 +48,14 @@ _MODEL_ENTRIES = [
         "meta/llama-3.2-90b-vision-instruct",
         CoreModelId.llama3_2_90b_vision_instruct.value,
     ),
+    ProviderModelEntry(
+        provider_model_id="baai/bge-m3",
+        model_type=ModelType.embedding,
+        metadata={
+            "embedding_dimensions": 1024,
+            "context_length": 8192,
+        },
+    ),
     # TODO(mf): how do we handle Nemotron models?
     # "Llama3.1-Nemotron-51B-Instruct" -> "meta/llama-3.1-nemotron-51b-instruct",
 ]

llama_stack/providers/remote/inference/nvidia/nvidia.py

@@ -10,6 +10,11 @@ from typing import AsyncIterator, List, Optional, Union
 
 from openai import APIConnectionError, AsyncOpenAI
 
+from llama_stack.apis.common.content_types import (
+    InterleavedContent,
+    InterleavedContentItem,
+    TextContentItem,
+)
 from llama_stack.apis.inference import (
     ChatCompletionRequest,
     ChatCompletionResponse,
@@ -19,7 +24,6 @@ from llama_stack.apis.inference import (
     CompletionResponseStreamChunk,
     EmbeddingsResponse,
     Inference,
-    InterleavedContent,
     LogProbConfig,
     Message,
     ResponseFormat,
@@ -117,9 +121,38 @@ class NVIDIAInferenceAdapter(Inference, ModelRegistryHelper):
     async def embeddings(
         self,
         model_id: str,
-        contents: List[InterleavedContent],
+        contents: List[str] | List[InterleavedContentItem],
     ) -> EmbeddingsResponse:
-        raise NotImplementedError()
+        if any(content_has_media(content) for content in contents):
+            raise NotImplementedError("Media is not supported")
+
+        #
+        # Llama Stack: contents = List[str] | List[InterleavedContentItem]
+        #  ->
+        # OpenAI: input = str | List[str]
+        #
+        # we can ignore str and always pass List[str] to OpenAI
+        #
+        flat_contents = [
+            item.text if isinstance(item, TextContentItem) else item
+            for content in contents
+            for item in (content if isinstance(content, list) else [content])
+        ]
+        input = [content.text if isinstance(content, TextContentItem) else content for content in flat_contents]
+        model = self.get_provider_model_id(model_id)
+
+        response = await self._client.embeddings.create(
+            model=model,
+            input=input,
+            # extra_body={"input_type": "passage"|"query"},  # TODO(mf): how to tell caller's intent?
+        )
+
+        #
+        # OpenAI: CreateEmbeddingResponse(data=[Embedding(embedding=List[float], ...)], ...)
+        #  ->
+        # Llama Stack: EmbeddingsResponse(embeddings=List[List[float]])
+        #
+        return EmbeddingsResponse(embeddings=[embedding.embedding for embedding in response.data])
 
     async def chat_completion(
         self,

llama_stack/templates/nvidia/nvidia.py

@@ -41,9 +41,11 @@ def get_distribution_template() -> DistributionTemplate:
     core_model_to_hf_repo = {m.descriptor(): m.huggingface_repo for m in all_registered_models()}
     default_models = [
         ModelInput(
-            model_id=core_model_to_hf_repo[m.llama_model],
+            model_id=core_model_to_hf_repo[m.llama_model] if m.llama_model else m.provider_model_id,
             provider_model_id=m.provider_model_id,
             provider_id="nvidia",
+            model_type=m.model_type,
+            metadata=m.metadata,
         )
         for m in _MODEL_ENTRIES
     ]

llama_stack/templates/nvidia/run.yaml

@@ -135,6 +135,13 @@ models:
   provider_id: nvidia
   provider_model_id: meta/llama-3.2-90b-vision-instruct
   model_type: llm
+- metadata:
+    embedding_dimensions: 1024
+    context_length: 8192
+  model_id: baai/bge-m3
+  provider_id: nvidia
+  provider_model_id: baai/bge-m3
+  model_type: embedding
 shields: []
 vector_dbs: []
 datasets: []

tests/client-sdk/conftest.py

@@ -58,6 +58,12 @@ def pytest_addoption(parser):
         default="meta-llama/Llama-Guard-3-1B",
         help="Specify the safety shield model to use for testing",
     )
+    parser.addoption(
+        "--embedding-model",
+        action="store",
+        default=TEXT_MODEL,
+        help="Specify the embedding model to use for testing",
+    )
 
 
 @pytest.fixture(scope="session")
@@ -105,3 +111,9 @@ def pytest_generate_tests(metafunc):
             [metafunc.config.getoption("--vision-inference-model")],
             scope="session",
         )
+    if "embedding_model_id" in metafunc.fixturenames:
+        metafunc.parametrize(
+            "embedding_model_id",
+            [metafunc.config.getoption("--embedding-model")],
+            scope="session",
+        )

tests/client-sdk/inference/test_embedding.py

@@ -0,0 +1,103 @@
+# 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.
+
+
+#
+# Test plan:
+#
+#  Types of input:
+#   - array of a string
+#   - array of a image (ImageContentItem, either URL or base64 string)
+#   - array of a text (TextContentItem)
+#   - array of array of texts, images, or both
+#  Types of output:
+#   - list of list of floats
+#
+# Todo:
+#  - negative tests
+#    - empty
+#      - empty list
+#      - empty string
+#      - empty text
+#      - empty image
+#      - list of empty texts
+#      - list of empty images
+#      - list of empty texts and images
+#    - long
+#      - long string
+#      - long text
+#      - large image
+#      - appropriate combinations
+#    - batch size
+#      - many inputs
+#    - invalid
+#      - invalid URL
+#      - invalid base64
+#      - list of list of strings
+#
+# Notes:
+#  - use llama_stack_client fixture
+#  - use pytest.mark.parametrize when possible
+#  - no accuracy tests: only check the type of output, not the content
+#
+
+import pytest
+from llama_stack_client.types import EmbeddingsResponse
+from llama_stack_client.types.shared.interleaved_content import (
+    URL,
+    ImageContentItem,
+    ImageContentItemImage,
+    TextContentItem,
+)
+
+DUMMY_STRING = "hello"
+DUMMY_STRING2 = "world"
+DUMMY_TEXT = TextContentItem(text=DUMMY_STRING, type="text")
+DUMMY_TEXT2 = TextContentItem(text=DUMMY_STRING2, type="text")
+# TODO(mf): add a real image URL and base64 string
+DUMMY_IMAGE_URL = ImageContentItem(
+    image=ImageContentItemImage(url=URL(uri="https://example.com/image.jpg")), type="image"
+)
+DUMMY_IMAGE_BASE64 = ImageContentItem(image=ImageContentItemImage(data="base64string"), type="image")
+
+
+@pytest.mark.parametrize(
+    "contents",
+    [
+        [DUMMY_STRING, DUMMY_STRING2],
+        [DUMMY_TEXT, DUMMY_TEXT2],
+    ],
+    ids=[
+        "list[string]",
+        "list[text]",
+    ],
+)
+def test_embedding_text(llama_stack_client, embedding_model_id, contents):
+    response = llama_stack_client.inference.embeddings(model_id=embedding_model_id, contents=contents)
+    assert isinstance(response, EmbeddingsResponse)
+    assert len(response.embeddings) == sum(len(content) if isinstance(content, list) else 1 for content in contents)
+    assert isinstance(response.embeddings[0], list)
+    assert isinstance(response.embeddings[0][0], float)
+
+
+@pytest.mark.parametrize(
+    "contents",
+    [
+        [DUMMY_IMAGE_URL, DUMMY_IMAGE_BASE64],
+        [DUMMY_IMAGE_URL, DUMMY_STRING, DUMMY_IMAGE_BASE64, DUMMY_TEXT],
+    ],
+    ids=[
+        "list[url,base64]",
+        "list[url,string,base64,text]",
+    ],
+)
+@pytest.mark.skip(reason="Media is not supported")
+def test_embedding_image(llama_stack_client, embedding_model_id, contents):
+    response = llama_stack_client.inference.embeddings(model_id=embedding_model_id, contents=contents)
+    assert isinstance(response, EmbeddingsResponse)
+    assert len(response.embeddings) == sum(len(content) if isinstance(content, list) else 1 for content in contents)
+    assert isinstance(response.embeddings[0], list)
+    assert isinstance(response.embeddings[0][0], float)