Merge branch 'meta-llama:main' into main

llama_stack/providers/registry/memory.py

@@ -8,11 +8,25 @@ from typing import List
 
 from llama_stack.distribution.datatypes import *  # noqa: F403
 
+
 EMBEDDING_DEPS = [
     "blobfile",
     "chardet",
     "pypdf",
-    "sentence-transformers",
+    "tqdm",
+    "numpy",
+    "scikit-learn",
+    "scipy",
+    "nltk",
+    "sentencepiece",
+    "transformers",
+    # this happens to work because special dependencies are always installed last
+    # so if there was a regular torch installed first, this would be ignored
+    # we need a better way to do this to identify potential conflicts, etc.
+    # for now, this lets us significantly reduce the size of the container which
+    # does not have any "local" inference code (and hence does not need GPU-enabled torch)
+    "torch --index-url https://download.pytorch.org/whl/cpu",
+    "sentence-transformers --no-deps",
 ]
 
 

llama_stack/providers/registry/safety.py

@@ -15,13 +15,15 @@ def available_providers() -> List[ProviderSpec]:
             api=Api.safety,
             provider_id="meta-reference",
             pip_packages=[
-                "accelerate",
                 "codeshield",
-                "torch",
                 "transformers",
+                "torch --index-url https://download.pytorch.org/whl/cpu",
             ],
             module="llama_stack.providers.impls.meta_reference.safety",
             config_class="llama_stack.providers.impls.meta_reference.safety.SafetyConfig",
+            api_dependencies=[
+                Api.inference,
+            ],
         ),
         remote_provider_spec(
             api=Api.safety,