add nvidia distribution

llama_stack/providers/registry/post_training.py

@@ -6,7 +6,7 @@
 
 from typing import List
 
-from llama_stack.providers.datatypes import Api, InlineProviderSpec, ProviderSpec
+from llama_stack.providers.datatypes import AdapterSpec, Api, InlineProviderSpec, ProviderSpec, remote_provider_spec
 
 
 def available_providers() -> List[ProviderSpec]:
@@ -22,15 +22,13 @@ def available_providers() -> List[ProviderSpec]:
                 Api.datasets,
             ],
         ),
-        InlineProviderSpec(
+        remote_provider_spec(
             api=Api.post_training,
-            provider_type="remote::nvidia",
+            adapter=AdapterSpec(
-            pip_packages=["torch", "numpy"],
+                adapter_type="nvidia",
-            module="llama_stack.providers.remote.post_training.nvidia",
+                pip_packages=["requests"],
-            config_class="llama_stack.providers.remote.post_training.nvidia.NvidiaPostTrainingConfig",
+                module="llama_stack.providers.remote.post_training.nvidia",
-            api_dependencies=[
+                config_class="llama_stack.providers.remote.post_training.nvidia.NvidiaPostTrainingConfig",
-                Api.datasetio,
+            ),
-                Api.datasets,
-            ],
         ),
     ]

llama_stack/providers/remote/post_training/__init__.py

@@ -0,0 +1,5 @@
+# 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.

llama_stack/providers/remote/post_training/nvidia/__init__.py

@@ -13,7 +13,7 @@ from .config import NvidiaPostTrainingConfig
 # post_training api and the torchtune provider is still experimental and under heavy development
 
 
-async def get_provider_impl(
+async def get_adapter_impl(
     config: NvidiaPostTrainingConfig,
     deps: Dict[Api, ProviderSpec],
 ):

llama_stack/providers/remote/post_training/nvidia/config.py

@@ -5,7 +5,7 @@
 # the root directory of this source tree.
 
 import os
-from typing import Optional
+from typing import Any, Dict, Optional
 
 from pydantic import BaseModel, Field
 
@@ -58,3 +58,15 @@ class NvidiaPostTrainingConfig(BaseModel):
         default_factory=lambda: os.getenv("NVIDIA_OUTPUT_MODEL_DIR", "test-example-model@v1"),
         description="Directory to save the output model",
     )
+
+    @classmethod
+    def sample_run_config(cls, **kwargs) -> Dict[str, Any]:
+        return {
+            "api_key": "${env.NVIDIA_API_KEY:}",
+            "user_id": "${env.NVIDIA_USER_ID:llama-stack-user}",
+            "dataset_namespace": "${env.NVIDIA_DATASET_NAMESPACE:default}",
+            "access_policies": "${env.NVIDIA_ACCESS_POLICIES:}",
+            "project_id": "${env.NVIDIA_PROJECT_ID:test-project}",
+            "customizer_url": "${env.NVIDIA_CUSTOMIZER_URL:}",
+            "output_model_dir": "${env.NVIDIA_OUTPUT_MODEL_DIR:test-example-model@v1}",
+        }

llama_stack/templates/nvidia/build.yaml

@@ -21,6 +21,8 @@ distribution_spec:
     - inline::basic
     - inline::llm-as-judge
     - inline::braintrust
+    post_training:
+    - remote::nvidia
     tool_runtime:
     - remote::brave-search
     - remote::tavily-search

llama_stack/templates/nvidia/nvidia.py

@@ -10,6 +10,7 @@ from llama_stack.distribution.datatypes import ModelInput, Provider, ToolGroupIn
 from llama_stack.models.llama.sku_list import all_registered_models
 from llama_stack.providers.remote.inference.nvidia import NVIDIAConfig
 from llama_stack.providers.remote.inference.nvidia.models import _MODEL_ENTRIES
+from llama_stack.providers.remote.post_training.nvidia import NvidiaPostTrainingConfig
 from llama_stack.templates.template import DistributionTemplate, RunConfigSettings
 
 
@@ -18,6 +19,7 @@ def get_distribution_template() -> DistributionTemplate:
         "inference": ["remote::nvidia"],
         "vector_io": ["inline::faiss"],
         "safety": ["inline::llama-guard"],
+        "post_training": ["remote::nvidia"],
         "agents": ["inline::meta-reference"],
         "telemetry": ["inline::meta-reference"],
         "eval": ["inline::meta-reference"],
@@ -38,6 +40,12 @@ def get_distribution_template() -> DistributionTemplate:
         config=NVIDIAConfig.sample_run_config(),
     )
 
+    post_training_provider = Provider(
+        provider_id="nvidia",
+        provider_type="remote::nvidia",
+        config=NvidiaPostTrainingConfig.sample_run_config(),
+    )
+
     core_model_to_hf_repo = {m.descriptor(): m.huggingface_repo for m in all_registered_models()}
     default_models = [
         ModelInput(
@@ -90,5 +98,30 @@ def get_distribution_template() -> DistributionTemplate:
                 "",
                 "NVIDIA API Key",
             ),
+            ## Nemo Customizer related variables
+            "NVIDIA_USER_ID": (
+                "llama-stack-user",
+                "NVIDIA User ID",
+            ),
+            "NVIDIA_DATASET_NAMESPACE": (
+                "default",
+                "NVIDIA Dataset Namespace",
+            ),
+            "NVIDIA_ACCESS_POLICIES": (
+                "{}",
+                "NVIDIA Access Policies",
+            ),
+            "NVIDIA_PROJECT_ID": (
+                "test-project",
+                "NVIDIA Project ID",
+            ),
+            "NVIDIA_CUSTOMIZER_URL": (
+                "https://customizer.api.nvidia.com",
+                "NVIDIA Customizer URL",
+            ),
+            "NVIDIA_OUTPUT_MODEL_DIR": (
+                "test-example-model@v1",
+                "NVIDIA Output Model Directory",
+            ),
         },
     )

llama_stack/templates/nvidia/run.yaml

@@ -8,6 +8,7 @@ apis:
 - safety
 - scoring
 - telemetry
+- post_training
 - tool_runtime
 - vector_io
 providers:
@@ -66,6 +67,10 @@ providers:
     provider_type: inline::braintrust
     config:
       openai_api_key: ${env.OPENAI_API_KEY:}
+  post_training:
+  - provider_id: nvidia-customizer
+    provider_type: remote::nvidia
+    config: {}
   tool_runtime:
   - provider_id: brave-search
     provider_type: remote::brave-search