fix: Safety in starter (#2731)

- fireworks, together do not support Llama-guard 3 8b model anymore 
- Need to default to ollama 
- current safety shields logic was not correct since the shield_id was
the provider ( which had duplicates )
- Followed similar logic to models 

Note: Seems a bit over-engineered but this can now be extended to other
providers and fits in the overall mechanism of how env_vars are used to
manage starter.

### How to test 
```
ENABLE_OLLAMA=ollama ENABLE_FIREWORKS=fireworks SAFETY_MODEL=llama-guard3:1b pytest -s -v tests/integration/ --stack-config starter -k 'not(supervised_fine_tune or builtin_tool_code or safety_with_image or code_interpreter_for or rag_and_code or truncation or register_and_unregister)' --text-model fireworks/meta-llama/Llama-3.3-70B-Instruct --vision-model fireworks/meta-llama/Llama-4-Scout-17B-16E-Instruct --safety-shield llama-guard3:1b --embedding-model all-MiniLM-L6-v2
```

### Related but not obvious in this PR 
In the llama-stack-ops repo, we run tests before publishing packages and
docker containers.
The actions in that repo were using the fireworks / together distros (
which are non-existent )

So need to update that to run with `starter` and use `ollama`
specifically for safety.

llama_stack/templates/nvidia/nvidia.py

@@ -68,7 +68,7 @@ def get_distribution_template() -> DistributionTemplate:
         ),
     ]
 
-    default_models = get_model_registry(available_models)
+    default_models, _ = get_model_registry(available_models)
     return DistributionTemplate(
         name="nvidia",
         distro_type="self_hosted",

llama_stack/templates/open-benchmark/open_benchmark.py

@@ -146,7 +146,8 @@ def get_distribution_template() -> DistributionTemplate:
         ),
     ]
 
-    default_models = get_model_registry(available_models) + [
+    models, _ = get_model_registry(available_models)
+    default_models = models + [
         ModelInput(
             model_id="meta-llama/Llama-3.3-70B-Instruct",
             provider_id="groq",

llama_stack/templates/starter/run.yaml

@@ -1171,24 +1171,8 @@ models:
   provider_id: ${env.ENABLE_SENTENCE_TRANSFORMERS:=sentence-transformers}
   model_type: embedding
 shields:
-- shield_id: ${env.ENABLE_OLLAMA:=__disabled__}
-  provider_id: llama-guard
-  provider_shield_id: ${env.ENABLE_OLLAMA:=__disabled__}/${env.SAFETY_MODEL:=llama-guard3:1b}
-- shield_id: ${env.ENABLE_FIREWORKS:=__disabled__}
-  provider_id: llama-guard
-  provider_shield_id: ${env.ENABLE_FIREWORKS:=__disabled__}/${env.SAFETY_MODEL:=accounts/fireworks/models/llama-guard-3-8b}
-- shield_id: ${env.ENABLE_FIREWORKS:=__disabled__}
-  provider_id: llama-guard
-  provider_shield_id: ${env.ENABLE_FIREWORKS:=__disabled__}/${env.SAFETY_MODEL:=accounts/fireworks/models/llama-guard-3-11b-vision}
-- shield_id: ${env.ENABLE_TOGETHER:=__disabled__}
-  provider_id: llama-guard
-  provider_shield_id: ${env.ENABLE_TOGETHER:=__disabled__}/${env.SAFETY_MODEL:=meta-llama/Llama-Guard-3-8B}
-- shield_id: ${env.ENABLE_TOGETHER:=__disabled__}
-  provider_id: llama-guard
-  provider_shield_id: ${env.ENABLE_TOGETHER:=__disabled__}/${env.SAFETY_MODEL:=meta-llama/Llama-Guard-3-11B-Vision-Turbo}
-- shield_id: ${env.ENABLE_SAMBANOVA:=__disabled__}
-  provider_id: llama-guard
-  provider_shield_id: ${env.ENABLE_SAMBANOVA:=__disabled__}/${env.SAFETY_MODEL:=sambanova/Meta-Llama-Guard-3-8B}
+- shield_id: ${env.SAFETY_MODEL:=__disabled__}
+  provider_shield_id: ${env.ENABLE_OLLAMA:=__disabled__}/${env.SAFETY_MODEL:=__disabled__}
 vector_dbs: []
 datasets: []
 scoring_fns: []

llama_stack/templates/starter/starter.py

@@ -12,7 +12,6 @@ from llama_stack.distribution.datatypes import (
     ModelInput,
     Provider,
     ProviderSpec,
-    ShieldInput,
     ToolGroupInput,
 )
 from llama_stack.distribution.utils.dynamic import instantiate_class_type
@@ -32,75 +31,39 @@ from llama_stack.providers.registry.inference import available_providers
 from llama_stack.providers.remote.inference.anthropic.models import (
     MODEL_ENTRIES as ANTHROPIC_MODEL_ENTRIES,
 )
-from llama_stack.providers.remote.inference.anthropic.models import (
-    SAFETY_MODELS_ENTRIES as ANTHROPIC_SAFETY_MODELS_ENTRIES,
-)
 from llama_stack.providers.remote.inference.bedrock.models import (
     MODEL_ENTRIES as BEDROCK_MODEL_ENTRIES,
 )
-from llama_stack.providers.remote.inference.bedrock.models import (
-    SAFETY_MODELS_ENTRIES as BEDROCK_SAFETY_MODELS_ENTRIES,
-)
 from llama_stack.providers.remote.inference.cerebras.models import (
     MODEL_ENTRIES as CEREBRAS_MODEL_ENTRIES,
 )
-from llama_stack.providers.remote.inference.cerebras.models import (
-    SAFETY_MODELS_ENTRIES as CEREBRAS_SAFETY_MODELS_ENTRIES,
-)
 from llama_stack.providers.remote.inference.databricks.databricks import (
     MODEL_ENTRIES as DATABRICKS_MODEL_ENTRIES,
 )
-from llama_stack.providers.remote.inference.databricks.databricks import (
-    SAFETY_MODELS_ENTRIES as DATABRICKS_SAFETY_MODELS_ENTRIES,
-)
 from llama_stack.providers.remote.inference.fireworks.models import (
     MODEL_ENTRIES as FIREWORKS_MODEL_ENTRIES,
 )
-from llama_stack.providers.remote.inference.fireworks.models import (
-    SAFETY_MODELS_ENTRIES as FIREWORKS_SAFETY_MODELS_ENTRIES,
-)
 from llama_stack.providers.remote.inference.gemini.models import (
     MODEL_ENTRIES as GEMINI_MODEL_ENTRIES,
 )
-from llama_stack.providers.remote.inference.gemini.models import (
-    SAFETY_MODELS_ENTRIES as GEMINI_SAFETY_MODELS_ENTRIES,
-)
 from llama_stack.providers.remote.inference.groq.models import (
     MODEL_ENTRIES as GROQ_MODEL_ENTRIES,
 )
-from llama_stack.providers.remote.inference.groq.models import (
-    SAFETY_MODELS_ENTRIES as GROQ_SAFETY_MODELS_ENTRIES,
-)
 from llama_stack.providers.remote.inference.nvidia.models import (
     MODEL_ENTRIES as NVIDIA_MODEL_ENTRIES,
 )
-from llama_stack.providers.remote.inference.nvidia.models import (
-    SAFETY_MODELS_ENTRIES as NVIDIA_SAFETY_MODELS_ENTRIES,
-)
 from llama_stack.providers.remote.inference.openai.models import (
     MODEL_ENTRIES as OPENAI_MODEL_ENTRIES,
 )
-from llama_stack.providers.remote.inference.openai.models import (
-    SAFETY_MODELS_ENTRIES as OPENAI_SAFETY_MODELS_ENTRIES,
-)
 from llama_stack.providers.remote.inference.runpod.runpod import (
     MODEL_ENTRIES as RUNPOD_MODEL_ENTRIES,
 )
-from llama_stack.providers.remote.inference.runpod.runpod import (
-    SAFETY_MODELS_ENTRIES as RUNPOD_SAFETY_MODELS_ENTRIES,
-)
 from llama_stack.providers.remote.inference.sambanova.models import (
     MODEL_ENTRIES as SAMBANOVA_MODEL_ENTRIES,
 )
-from llama_stack.providers.remote.inference.sambanova.models import (
-    SAFETY_MODELS_ENTRIES as SAMBANOVA_SAFETY_MODELS_ENTRIES,
-)
 from llama_stack.providers.remote.inference.together.models import (
     MODEL_ENTRIES as TOGETHER_MODEL_ENTRIES,
 )
-from llama_stack.providers.remote.inference.together.models import (
-    SAFETY_MODELS_ENTRIES as TOGETHER_SAFETY_MODELS_ENTRIES,
-)
 from llama_stack.providers.remote.vector_io.chroma.config import ChromaVectorIOConfig
 from llama_stack.providers.remote.vector_io.pgvector.config import (
     PGVectorVectorIOConfig,
@@ -111,6 +74,7 @@ from llama_stack.templates.template import (
     DistributionTemplate,
     RunConfigSettings,
     get_model_registry,
+    get_shield_registry,
 )
 
 
@@ -164,28 +128,13 @@ def _get_model_entries_for_provider(provider_type: str) -> list[ProviderModelEnt
 def _get_model_safety_entries_for_provider(provider_type: str) -> list[ProviderModelEntry]:
     """Get model entries for a specific provider type."""
     safety_model_entries_map = {
-        "openai": OPENAI_SAFETY_MODELS_ENTRIES,
-        "fireworks": FIREWORKS_SAFETY_MODELS_ENTRIES,
-        "together": TOGETHER_SAFETY_MODELS_ENTRIES,
-        "anthropic": ANTHROPIC_SAFETY_MODELS_ENTRIES,
-        "gemini": GEMINI_SAFETY_MODELS_ENTRIES,
-        "groq": GROQ_SAFETY_MODELS_ENTRIES,
-        "sambanova": SAMBANOVA_SAFETY_MODELS_ENTRIES,
-        "cerebras": CEREBRAS_SAFETY_MODELS_ENTRIES,
-        "bedrock": BEDROCK_SAFETY_MODELS_ENTRIES,
-        "databricks": DATABRICKS_SAFETY_MODELS_ENTRIES,
-        "nvidia": NVIDIA_SAFETY_MODELS_ENTRIES,
-        "runpod": RUNPOD_SAFETY_MODELS_ENTRIES,
-    }
-
-    # Special handling for providers with dynamic model entries
-    if provider_type == "ollama":
-        return [
+        "ollama": [
             ProviderModelEntry(
-                provider_model_id="llama-guard3:1b",
+                provider_model_id="${env.SAFETY_MODEL:=__disabled__}",
                 model_type=ModelType.llm,
             ),
-        ]
+        ],
+    }
 
     return safety_model_entries_map.get(provider_type, [])
 
@@ -246,28 +195,20 @@ def get_remote_inference_providers() -> tuple[list[Provider], dict[str, list[Pro
 
 
 # build a list of shields for all possible providers
-def get_shields_for_providers(providers: list[Provider]) -> list[ShieldInput]:
-    shields = []
+def get_safety_models_for_providers(providers: list[Provider]) -> dict[str, list[ProviderModelEntry]]:
+    available_models = {}
     for provider in providers:
         provider_type = provider.provider_type.split("::")[1]
         safety_model_entries = _get_model_safety_entries_for_provider(provider_type)
         if len(safety_model_entries) == 0:
             continue
-        if provider.provider_id:
-            shield_id = provider.provider_id
-        else:
-            raise ValueError(f"Provider {provider.provider_type} has no provider_id")
-        for safety_model_entry in safety_model_entries:
-            print(f"provider.provider_id: {provider.provider_id}")
-            print(f"safety_model_entry.provider_model_id: {safety_model_entry.provider_model_id}")
-            shields.append(
-                ShieldInput(
-                    provider_id="llama-guard",
-                    shield_id=shield_id,
-                    provider_shield_id=f"{provider.provider_id}/${{env.SAFETY_MODEL:={safety_model_entry.provider_model_id}}}",
-                )
-            )
-    return shields
+
+        env_var = f"ENABLE_{provider_type.upper().replace('-', '_').replace('::', '_')}"
+        provider_id = f"${{env.{env_var}:=__disabled__}}"
+
+        available_models[provider_id] = safety_model_entries
+
+    return available_models
 
 
 def get_distribution_template() -> DistributionTemplate:
@@ -307,8 +248,6 @@ def get_distribution_template() -> DistributionTemplate:
         ),
     ]
 
-    shields = get_shields_for_providers(remote_inference_providers)
-
     providers = {
         "inference": ([p.provider_type for p in remote_inference_providers] + ["inline::sentence-transformers"]),
         "vector_io": ([p.provider_type for p in vector_io_providers]),
@@ -361,7 +300,10 @@ def get_distribution_template() -> DistributionTemplate:
         },
     )
 
-    default_models = get_model_registry(available_models)
+    default_models, ids_conflict_in_models = get_model_registry(available_models)
+
+    available_safety_models = get_safety_models_for_providers(remote_inference_providers)
+    shields = get_shield_registry(available_safety_models, ids_conflict_in_models)
 
     return DistributionTemplate(
         name=name,

llama_stack/templates/template.py

@@ -37,7 +37,7 @@ from llama_stack.providers.utils.sqlstore.sqlstore import get_pip_packages as ge
 
 def get_model_registry(
     available_models: dict[str, list[ProviderModelEntry]],
-) -> list[ModelInput]:
+) -> tuple[list[ModelInput], bool]:
     models = []
 
     # check for conflicts in model ids
@@ -74,7 +74,50 @@ def get_model_registry(
                         metadata=entry.metadata,
                     )
                 )
-    return models
+    return models, ids_conflict
+
+
+def get_shield_registry(
+    available_safety_models: dict[str, list[ProviderModelEntry]],
+    ids_conflict_in_models: bool,
+) -> list[ShieldInput]:
+    shields = []
+
+    # check for conflicts in shield ids
+    all_ids = set()
+    ids_conflict = False
+
+    for _, entries in available_safety_models.items():
+        for entry in entries:
+            ids = [entry.provider_model_id] + entry.aliases
+            for model_id in ids:
+                if model_id in all_ids:
+                    ids_conflict = True
+                    rich.print(
+                        f"[yellow]Shield id {model_id} conflicts; all shield ids will be prefixed with provider id[/yellow]"
+                    )
+                    break
+            all_ids.update(ids)
+            if ids_conflict:
+                break
+        if ids_conflict:
+            break
+
+    for provider_id, entries in available_safety_models.items():
+        for entry in entries:
+            ids = [entry.provider_model_id] + entry.aliases
+            for model_id in ids:
+                identifier = f"{provider_id}/{model_id}" if ids_conflict and provider_id not in model_id else model_id
+                shields.append(
+                    ShieldInput(
+                        shield_id=identifier,
+                        provider_shield_id=f"{provider_id}/{entry.provider_model_id}"
+                        if ids_conflict_in_models
+                        else entry.provider_model_id,
+                    )
+                )
+
+    return shields
 
 
 class DefaultModel(BaseModel):

llama_stack/templates/watsonx/watsonx.py

@@ -69,7 +69,7 @@ def get_distribution_template() -> DistributionTemplate:
         },
     )
 
-    default_models = get_model_registry(available_models)
+    default_models, _ = get_model_registry(available_models)
     return DistributionTemplate(
         name="watsonx",
         distro_type="remote_hosted",