fold in meta-reference-quantized

llama_stack/providers/inline/inference/meta_reference/__init__.py

@@ -4,13 +4,13 @@
 # This source code is licensed under the terms described in the LICENSE file in
 # the root directory of this source tree.
 
-from typing import Any, Dict, Union
+from typing import Any, Dict
 
-from .config import MetaReferenceInferenceConfig, MetaReferenceQuantizedInferenceConfig
+from .config import MetaReferenceInferenceConfig
 
 
 async def get_provider_impl(
-    config: Union[MetaReferenceInferenceConfig, MetaReferenceQuantizedInferenceConfig],
+    config: MetaReferenceInferenceConfig,
     _deps: Dict[str, Any],
 ):
     from .inference import MetaReferenceInferenceImpl

llama_stack/providers/inline/inference/meta_reference/config.py

@@ -31,6 +31,8 @@ class MetaReferenceInferenceConfig(BaseModel):
     # can override by specifying the directory explicitly
     checkpoint_dir: Optional[str] = None
 
+    quantization: Optional[QuantizationConfig] = None
+
     @field_validator("model")
     @classmethod
     def validate_model(cls, model: str) -> str:
@@ -47,27 +49,14 @@ class MetaReferenceInferenceConfig(BaseModel):
         cls,
         model: str = "Llama3.2-3B-Instruct",
         checkpoint_dir: str = "${env.CHECKPOINT_DIR:null}",
+        quantization_type: str = "${env.QUANTIZATION_TYPE:bf16}",
         **kwargs,
     ) -> Dict[str, Any]:
         return {
             "model": model,
             "max_seq_len": 4096,
             "checkpoint_dir": checkpoint_dir,
+            "quantization": {
+                "type": quantization_type,
+            },
         }
-
-
-class MetaReferenceQuantizedInferenceConfig(MetaReferenceInferenceConfig):
-    quantization: QuantizationConfig
-
-    @classmethod
-    def sample_run_config(
-        cls,
-        model: str = "Llama3.2-3B-Instruct",
-        checkpoint_dir: str = "${env.CHECKPOINT_DIR:null}",
-        **kwargs,
-    ) -> Dict[str, Any]:
-        config = super().sample_run_config(model, checkpoint_dir, **kwargs)
-        config["quantization"] = {
-            "type": "fp8",
-        }
-        return config

llama_stack/providers/inline/inference/meta_reference/generators.py

@@ -11,9 +11,7 @@ import torch
 from lmformatenforcer import JsonSchemaParser, TokenEnforcer, TokenEnforcerTokenizerData
 
 from llama_stack.apis.inference import (
-    Fp8QuantizationConfig,
     GreedySamplingStrategy,
-    Int4QuantizationConfig,
     JsonSchemaResponseFormat,
     ResponseFormat,
     SamplingParams,
@@ -32,7 +30,7 @@ from llama_stack.providers.utils.inference.prompt_adapter import (
 )
 
 from .common import model_checkpoint_dir
-from .config import MetaReferenceInferenceConfig, MetaReferenceQuantizedInferenceConfig
+from .config import MetaReferenceInferenceConfig
 from .inference import resolve_model
 
 Tokenizer = Llama4Tokenizer | Llama3Tokenizer
@@ -118,7 +116,7 @@ def _infer_tool_prompt_format(request: ChatCompletionRequestWithRawContent):
 class Llama4Generator:
     def __init__(
         self,
-        config: MetaReferenceInferenceConfig | MetaReferenceQuantizedInferenceConfig,
+        config: MetaReferenceInferenceConfig,
         model_id: str,
         llama_model: Model,
     ):
@@ -133,11 +131,13 @@ class Llama4Generator:
                 # if the model is a native llama model, get the default checkpoint_dir based on model core_model_id value
                 ckpt_dir = model_checkpoint_dir(resolved_model.descriptor())
 
-        if isinstance(config, MetaReferenceQuantizedInferenceConfig):
+        if config.quantization:
-            if isinstance(config.quantization, Fp8QuantizationConfig):
+            if config.quantization.type == "fp8":
                 quantization_mode = QuantizationMode.fp8_mixed
-            elif isinstance(config.quantization, Int4QuantizationConfig):
+            elif config.quantization.type == "int4":
                 quantization_mode = QuantizationMode.int4_mixed
+            elif config.quantization.type == "bf16":
+                quantization_mode = None
             else:
                 raise ValueError(f"Unsupported quantization mode {config.quantization}")
         else:
@@ -207,7 +207,7 @@ class Llama4Generator:
 class Llama3Generator:
     def __init__(
         self,
-        config: MetaReferenceInferenceConfig | MetaReferenceQuantizedInferenceConfig,
+        config: MetaReferenceInferenceConfig,
         model_id: str,
         llama_model: Model,
     ):
@@ -222,11 +222,13 @@ class Llama3Generator:
                 # if the model is a native llama model, get the default checkpoint_dir based on model core_model_id value
                 ckpt_dir = model_checkpoint_dir(resolved_model.descriptor())
 
-        if isinstance(config, MetaReferenceQuantizedInferenceConfig):
+        if config.quantization:
-            if isinstance(config.quantization, Fp8QuantizationConfig):
+            if config.quantization.type == "fp8":
                 quantization_mode = QuantizationMode.fp8_mixed
-            elif isinstance(config.quantization, Int4QuantizationConfig):
+            elif config.quantization.type == "int4":
                 quantization_mode = QuantizationMode.int4_mixed
+            elif config.quantization.type == "bf16":
+                quantization_mode = None
             else:
                 raise ValueError(f"Unsupported quantization mode {config.quantization}")
         else:

llama_stack/providers/registry/inference.py

@@ -24,6 +24,8 @@ META_REFERENCE_DEPS = [
     "zmq",
     "lm-format-enforcer",
     "sentence-transformers",
+    "torchao==0.5.0",
+    "fbgemm-gpu-genai==1.1.2",
 ]
 
 
@@ -36,13 +38,6 @@ def available_providers() -> List[ProviderSpec]:
             module="llama_stack.providers.inline.inference.meta_reference",
             config_class="llama_stack.providers.inline.inference.meta_reference.MetaReferenceInferenceConfig",
         ),
-        InlineProviderSpec(
-            api=Api.inference,
-            provider_type="inline::meta-reference-quantized",
-            pip_packages=META_REFERENCE_DEPS + ["fbgemm-gpu", "torchao==0.5.0"],
-            module="llama_stack.providers.inline.inference.meta_reference",
-            config_class="llama_stack.providers.inline.inference.meta_reference.MetaReferenceQuantizedInferenceConfig",
-        ),
         InlineProviderSpec(
             api=Api.inference,
             provider_type="inline::vllm",

llama_stack/templates/meta-reference-quantized-gpu/__init__.py

@@ -1,7 +0,0 @@
-# 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.
-
-from .meta_reference import get_distribution_template  # noqa: F401

llama_stack/templates/meta-reference-quantized-gpu/build.yaml

@@ -1,32 +0,0 @@
-version: '2'
-distribution_spec:
-  description: Use Meta Reference with fp8, int4 quantization for running LLM inference
-  providers:
-    inference:
-    - inline::meta-reference-quantized
-    vector_io:
-    - inline::faiss
-    - remote::chromadb
-    - remote::pgvector
-    safety:
-    - inline::llama-guard
-    agents:
-    - inline::meta-reference
-    telemetry:
-    - inline::meta-reference
-    eval:
-    - inline::meta-reference
-    datasetio:
-    - remote::huggingface
-    - inline::localfs
-    scoring:
-    - inline::basic
-    - inline::llm-as-judge
-    - inline::braintrust
-    tool_runtime:
-    - remote::brave-search
-    - remote::tavily-search
-    - inline::code-interpreter
-    - inline::rag-runtime
-    - remote::model-context-protocol
-image_type: conda

llama_stack/templates/meta-reference-quantized-gpu/doc_template.md

@@ -1,113 +0,0 @@
----
-orphan: true
----
-# Meta Reference Quantized Distribution
-
-```{toctree}
-:maxdepth: 2
-:hidden:
-
-self
-```
-
-The `llamastack/distribution-{{ name }}` distribution consists of the following provider configurations:
-
-{{ providers_table }}
-
-The only difference vs. the `meta-reference-gpu` distribution is that it has support for more efficient inference -- with fp8, int4 quantization, etc.
-
-Note that you need access to nvidia GPUs to run this distribution. This distribution is not compatible with CPU-only machines or machines with AMD GPUs.
-
-{% if run_config_env_vars %}
-### Environment Variables
-
-The following environment variables can be configured:
-
-{% for var, (default_value, description) in run_config_env_vars.items() %}
-- `{{ var }}`: {{ description }} (default: `{{ default_value }}`)
-{% endfor %}
-{% endif %}
-
-
-## Prerequisite: Downloading Models
-
-Please use `llama model list --downloaded` to check that you have llama model checkpoints downloaded in `~/.llama` before proceeding. See [installation guide](https://llama-stack.readthedocs.io/en/latest/references/llama_cli_reference/download_models.html) here to download the models. Run `llama model list` to see the available models to download, and `llama model download` to download the checkpoints.
-
-```
-$ llama model list --downloaded
-┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━┓
-┃ Model                                   ┃ Size     ┃ Modified Time       ┃
-┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━┩
-│ Llama3.2-1B-Instruct:int4-qlora-eo8     │ 1.53 GB  │ 2025-02-26 11:22:28 │
-├─────────────────────────────────────────┼──────────┼─────────────────────┤
-│ Llama3.2-1B                             │ 2.31 GB  │ 2025-02-18 21:48:52 │
-├─────────────────────────────────────────┼──────────┼─────────────────────┤
-│ Prompt-Guard-86M                        │ 0.02 GB  │ 2025-02-26 11:29:28 │
-├─────────────────────────────────────────┼──────────┼─────────────────────┤
-│ Llama3.2-3B-Instruct:int4-spinquant-eo8 │ 3.69 GB  │ 2025-02-26 11:37:41 │
-├─────────────────────────────────────────┼──────────┼─────────────────────┤
-│ Llama3.2-3B                             │ 5.99 GB  │ 2025-02-18 21:51:26 │
-├─────────────────────────────────────────┼──────────┼─────────────────────┤
-│ Llama3.1-8B                             │ 14.97 GB │ 2025-02-16 10:36:37 │
-├─────────────────────────────────────────┼──────────┼─────────────────────┤
-│ Llama3.2-1B-Instruct:int4-spinquant-eo8 │ 1.51 GB  │ 2025-02-26 11:35:02 │
-├─────────────────────────────────────────┼──────────┼─────────────────────┤
-│ Llama-Guard-3-1B                        │ 2.80 GB  │ 2025-02-26 11:20:46 │
-├─────────────────────────────────────────┼──────────┼─────────────────────┤
-│ Llama-Guard-3-1B:int4                   │ 0.43 GB  │ 2025-02-26 11:33:33 │
-└─────────────────────────────────────────┴──────────┴─────────────────────┘
-```
-
-## Running the Distribution
-
-You can do this via Conda (build code) or Docker which has a pre-built image.
-
-### Via Docker
-
-This method allows you to get started quickly without having to build the distribution code.
-
-```bash
-LLAMA_STACK_PORT=8321
-docker run \
-  -it \
-  --pull always \
-  -p $LLAMA_STACK_PORT:$LLAMA_STACK_PORT \
-  -v ~/.llama:/root/.llama \
-  llamastack/distribution-{{ name }} \
-  --port $LLAMA_STACK_PORT \
-  --env INFERENCE_MODEL=meta-llama/Llama-3.2-3B-Instruct
-```
-
-If you are using Llama Stack Safety / Shield APIs, use:
-
-```bash
-docker run \
-  -it \
-  --pull always \
-  -p $LLAMA_STACK_PORT:$LLAMA_STACK_PORT \
-  -v ~/.llama:/root/.llama \
-  llamastack/distribution-{{ name }} \
-  --port $LLAMA_STACK_PORT \
-  --env INFERENCE_MODEL=meta-llama/Llama-3.2-3B-Instruct \
-  --env SAFETY_MODEL=meta-llama/Llama-Guard-3-1B
-```
-
-### Via Conda
-
-Make sure you have done `uv pip install llama-stack` and have the Llama Stack CLI available.
-
-```bash
-llama stack build --template {{ name }} --image-type conda
-llama stack run distributions/{{ name }}/run.yaml \
-  --port $LLAMA_STACK_PORT \
-  --env INFERENCE_MODEL=meta-llama/Llama-3.2-3B-Instruct
-```
-
-If you are using Llama Stack Safety / Shield APIs, use:
-
-```bash
-llama stack run distributions/{{ name }}/run-with-safety.yaml \
-  --port $LLAMA_STACK_PORT \
-  --env INFERENCE_MODEL=meta-llama/Llama-3.2-3B-Instruct \
-  --env SAFETY_MODEL=meta-llama/Llama-Guard-3-1B
-```

llama_stack/templates/meta-reference-quantized-gpu/meta_reference.py

@@ -1,115 +0,0 @@
-# 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.
-
-from pathlib import Path
-
-from llama_stack.apis.models.models import ModelType
-from llama_stack.distribution.datatypes import ModelInput, Provider, ToolGroupInput
-from llama_stack.providers.inline.inference.meta_reference import (
-    MetaReferenceQuantizedInferenceConfig,
-)
-from llama_stack.providers.inline.inference.sentence_transformers import (
-    SentenceTransformersInferenceConfig,
-)
-from llama_stack.providers.inline.vector_io.faiss.config import FaissVectorIOConfig
-from llama_stack.templates.template import DistributionTemplate, RunConfigSettings
-
-
-def get_distribution_template() -> DistributionTemplate:
-    providers = {
-        "inference": ["inline::meta-reference-quantized"],
-        "vector_io": ["inline::faiss", "remote::chromadb", "remote::pgvector"],
-        "safety": ["inline::llama-guard"],
-        "agents": ["inline::meta-reference"],
-        "telemetry": ["inline::meta-reference"],
-        "eval": ["inline::meta-reference"],
-        "datasetio": ["remote::huggingface", "inline::localfs"],
-        "scoring": ["inline::basic", "inline::llm-as-judge", "inline::braintrust"],
-        "tool_runtime": [
-            "remote::brave-search",
-            "remote::tavily-search",
-            "inline::code-interpreter",
-            "inline::rag-runtime",
-            "remote::model-context-protocol",
-        ],
-    }
-    default_tool_groups = [
-        ToolGroupInput(
-            toolgroup_id="builtin::websearch",
-            provider_id="tavily-search",
-        ),
-        ToolGroupInput(
-            toolgroup_id="builtin::rag",
-            provider_id="rag-runtime",
-        ),
-        ToolGroupInput(
-            toolgroup_id="builtin::code_interpreter",
-            provider_id="code-interpreter",
-        ),
-    ]
-    name = "meta-reference-quantized-gpu"
-    inference_provider = Provider(
-        provider_id="meta-reference-inference",
-        provider_type="inline::meta-reference-quantized",
-        config=MetaReferenceQuantizedInferenceConfig.sample_run_config(
-            model="${env.INFERENCE_MODEL}",
-            checkpoint_dir="${env.INFERENCE_CHECKPOINT_DIR:null}",
-        ),
-    )
-    embedding_provider = Provider(
-        provider_id="sentence-transformers",
-        provider_type="inline::sentence-transformers",
-        config=SentenceTransformersInferenceConfig.sample_run_config(),
-    )
-    vector_io_provider = Provider(
-        provider_id="faiss",
-        provider_type="inline::faiss",
-        config=FaissVectorIOConfig.sample_run_config(f"~/.llama/distributions/{name}"),
-    )
-
-    inference_model = ModelInput(
-        model_id="${env.INFERENCE_MODEL}",
-        provider_id="meta-reference-inference",
-    )
-    embedding_model = ModelInput(
-        model_id="all-MiniLM-L6-v2",
-        provider_id="sentence-transformers",
-        model_type=ModelType.embedding,
-        metadata={
-            "embedding_dimension": 384,
-        },
-    )
-    return DistributionTemplate(
-        name=name,
-        distro_type="self_hosted",
-        description="Use Meta Reference with fp8, int4 quantization for running LLM inference",
-        template_path=Path(__file__).parent / "doc_template.md",
-        providers=providers,
-        run_configs={
-            "run.yaml": RunConfigSettings(
-                provider_overrides={
-                    "inference": [inference_provider, embedding_provider],
-                    "vector_io": [vector_io_provider],
-                },
-                default_models=[inference_model, embedding_model],
-                default_tool_groups=default_tool_groups,
-            ),
-        },
-        run_config_env_vars={
-            "LLAMA_STACK_PORT": (
-                "8321",
-                "Port for the Llama Stack distribution server",
-            ),
-            "INFERENCE_MODEL": (
-                "meta-llama/Llama-3.2-3B-Instruct",
-                "Inference model loaded into the Meta Reference server",
-            ),
-            "INFERENCE_CHECKPOINT_DIR": (
-                "null",
-                "Directory containing the Meta Reference model checkpoint",
-            ),
-        },
-    )

llama_stack/templates/meta-reference-quantized-gpu/run.yaml

@@ -1,134 +0,0 @@
-version: '2'
-image_name: meta-reference-quantized-gpu
-apis:
-- agents
-- datasetio
-- eval
-- inference
-- safety
-- scoring
-- telemetry
-- tool_runtime
-- vector_io
-providers:
-  inference:
-  - provider_id: meta-reference-inference
-    provider_type: inline::meta-reference-quantized
-    config:
-      model: ${env.INFERENCE_MODEL}
-      max_seq_len: 4096
-      checkpoint_dir: ${env.INFERENCE_CHECKPOINT_DIR:null}
-      quantization:
-        type: fp8
-  - provider_id: sentence-transformers
-    provider_type: inline::sentence-transformers
-    config: {}
-  vector_io:
-  - provider_id: faiss
-    provider_type: inline::faiss
-    config:
-      kvstore:
-        type: sqlite
-        namespace: null
-        db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/meta-reference-quantized-gpu}/faiss_store.db
-  safety:
-  - provider_id: llama-guard
-    provider_type: inline::llama-guard
-    config:
-      excluded_categories: []
-  agents:
-  - provider_id: meta-reference
-    provider_type: inline::meta-reference
-    config:
-      persistence_store:
-        type: sqlite
-        namespace: null
-        db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/meta-reference-quantized-gpu}/agents_store.db
-  telemetry:
-  - provider_id: meta-reference
-    provider_type: inline::meta-reference
-    config:
-      service_name: "${env.OTEL_SERVICE_NAME:\u200B}"
-      sinks: ${env.TELEMETRY_SINKS:console,sqlite}
-      sqlite_db_path: ${env.SQLITE_DB_PATH:~/.llama/distributions/meta-reference-quantized-gpu/trace_store.db}
-  eval:
-  - provider_id: meta-reference
-    provider_type: inline::meta-reference
-    config:
-      kvstore:
-        type: sqlite
-        namespace: null
-        db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/meta-reference-quantized-gpu}/meta_reference_eval.db
-  datasetio:
-  - provider_id: huggingface
-    provider_type: remote::huggingface
-    config:
-      kvstore:
-        type: sqlite
-        namespace: null
-        db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/meta-reference-quantized-gpu}/huggingface_datasetio.db
-  - provider_id: localfs
-    provider_type: inline::localfs
-    config:
-      kvstore:
-        type: sqlite
-        namespace: null
-        db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/meta-reference-quantized-gpu}/localfs_datasetio.db
-  scoring:
-  - provider_id: basic
-    provider_type: inline::basic
-    config: {}
-  - provider_id: llm-as-judge
-    provider_type: inline::llm-as-judge
-    config: {}
-  - provider_id: braintrust
-    provider_type: inline::braintrust
-    config:
-      openai_api_key: ${env.OPENAI_API_KEY:}
-  tool_runtime:
-  - provider_id: brave-search
-    provider_type: remote::brave-search
-    config:
-      api_key: ${env.BRAVE_SEARCH_API_KEY:}
-      max_results: 3
-  - provider_id: tavily-search
-    provider_type: remote::tavily-search
-    config:
-      api_key: ${env.TAVILY_SEARCH_API_KEY:}
-      max_results: 3
-  - provider_id: code-interpreter
-    provider_type: inline::code-interpreter
-    config: {}
-  - provider_id: rag-runtime
-    provider_type: inline::rag-runtime
-    config: {}
-  - provider_id: model-context-protocol
-    provider_type: remote::model-context-protocol
-    config: {}
-metadata_store:
-  type: sqlite
-  db_path: ${env.SQLITE_STORE_DIR:~/.llama/distributions/meta-reference-quantized-gpu}/registry.db
-models:
-- metadata: {}
-  model_id: ${env.INFERENCE_MODEL}
-  provider_id: meta-reference-inference
-  model_type: llm
-- metadata:
-    embedding_dimension: 384
-  model_id: all-MiniLM-L6-v2
-  provider_id: sentence-transformers
-  model_type: embedding
-shields: []
-vector_dbs: []
-datasets: []
-scoring_fns: []
-benchmarks: []
-tool_groups:
-- toolgroup_id: builtin::websearch
-  provider_id: tavily-search
-- toolgroup_id: builtin::rag
-  provider_id: rag-runtime
-- toolgroup_id: builtin::code_interpreter
-  provider_id: code-interpreter
-server:
-  port: 8321