temp commit

llama_stack/apis/post_training/post_training.py

@@ -16,7 +16,8 @@ from pydantic import BaseModel, Field
 from llama_models.llama3.api.datatypes import *  # noqa: F403
 from llama_stack.apis.datasets import *  # noqa: F403
 from llama_stack.apis.common.training_types import *  # noqa: F403
-import torch
+
+# import torch
 
 
 class OptimizerType(Enum):
@@ -36,7 +37,7 @@ class OptimizerConfig(BaseModel):
 
 @json_schema_type
 class TrainingConfig(BaseModel):
-    dtype: torch.dtype
+    dtype: str
     n_epochs: int
     max_steps_per_epoch: int
     gradient_accumulation_steps: int
@@ -116,9 +117,7 @@ class PostTrainingSFTRequest(BaseModel):
     validation_dataset_id: str
 
     algorithm: FinetuningAlgorithm
-    algorithm_config: Union[
+    algorithm_config: LoraFinetuningConfig
-        LoraFinetuningConfig, QLoraFinetuningConfig, DoraFinetuningConfig
-    ]
 
     optimizer_config: OptimizerConfig
     training_config: TrainingConfig
@@ -189,9 +188,7 @@ class PostTraining(Protocol):
         dataset_id: str,
         validation_dataset_id: str,
         algorithm: FinetuningAlgorithm,
-        algorithm_config: Union[
+        algorithm_config: LoraFinetuningConfig,
-            LoraFinetuningConfig, QLoraFinetuningConfig, DoraFinetuningConfig
-        ],
         optimizer_config: OptimizerConfig,
         training_config: TrainingConfig,
         hyperparam_search_config: Dict[str, Any],
@@ -206,7 +203,7 @@ class PostTraining(Protocol):
         dataset_id: str,
         validation_dataset_id: str,
         algorithm: RLHFAlgorithm,
-        algorithm_config: Union[DPOAlignmentConfig],
+        algorithm_config: DPOAlignmentConfig,
         optimizer_config: OptimizerConfig,
         training_config: TrainingConfig,
         hyperparam_search_config: Dict[str, Any],

llama_stack/distribution/resolver.py

@@ -24,6 +24,7 @@ from llama_stack.apis.inspect import Inspect
 from llama_stack.apis.memory import Memory
 from llama_stack.apis.memory_banks import MemoryBanks
 from llama_stack.apis.models import Models
+from llama_stack.apis.post_training import PostTraining
 from llama_stack.apis.safety import Safety
 from llama_stack.apis.scoring import Scoring
 from llama_stack.apis.scoring_functions import ScoringFunctions
@@ -58,6 +59,7 @@ def api_protocol_map() -> Dict[Api, Any]:
         Api.scoring_functions: ScoringFunctions,
         Api.eval: Eval,
         Api.eval_tasks: EvalTasks,
+        Api.post_training: PostTraining,
     }
 
 

llama_stack/providers/inline/post_training/meta_reference/post_training.py

@@ -8,6 +8,7 @@ from llama_stack.providers.inline.post_training.meta_reference.config import (
     MetaReferencePostTrainingConfig,
 )
 from llama_stack.apis.post_training import *  # noqa
+
 from llama_stack.providers.inline.post_training.meta_reference.recipes.lora_finetuning_single_device import (
     LoraFinetuningSingleDevice,
 )
@@ -20,17 +21,45 @@ class MetaReferencePostTrainingImpl:
         self.config = config
         self.datasetio_api = datasetio_api
 
+    LoraFinetuningConfig(
+        lora_attn_modules=["q_proj", "v_proj", "output_proj"],
+        apply_lora_to_mlp=True,
+        apply_lora_to_output=False,
+        rank=8,
+        alpha=16,
+    )
+
+    OptimizerConfig(
+        lr=3e-4,
+        lr_min=3e-5,
+        weight_decay=0.1,
+        num_warmup_steps=100,
+    )
+
+    TrainingConfig(
+        dtype="bf16",
+        n_epochs=1,
+        max_steps_per_epoch=10,
+        gradient_accumulation_steps=1,
+        batch_size=1,
+        shuffle=1,
+        enable_activation_checkpointing=False,
+        memory_efficient_fsdp_wrap=False,
+        fsdp_cpu_offload=False,
+    )
+
     def supervised_fine_tune(
         self,
-        job_uuid: str,
+        job_uuid: str = "1234",
-        model: str,
+        model: str = " meta-llama/Llama-3.2-3B-Instruct",
-        dataset_id: str,
+        dataset_id: str = "alpaca",
-        validation_dataset_id: str,
+        validation_dataset_id: str = "alpaca",
-        algorithm: FinetuningAlgorithm,
+        algorithm: FinetuningAlgorithm = FinetuningAlgorithm.lora,
-        algorithm_config: LoraFinetuningConfig,
+        algorithm_config: LoraFinetuningConfig = LoraFinetuningConfig,
-        optimizer_config: OptimizerConfig,
+        optimizer_config: OptimizerConfig = OptimizerConfig,
-        training_config: TrainingConfig,
+        training_config: TrainingConfig = TrainingConfig,
-        logger_config: Dict[str, Any],
+        hyperparam_search_config: Dict[str, Any] = {},
+        logger_config: Dict[str, Any] = {},
     ) -> PostTrainingJob:
         # wrapper request to make it easier to pass around (internal only, not exposed to API)
         request = PostTrainingSFTRequest(
@@ -54,3 +83,36 @@ class MetaReferencePostTrainingImpl:
             raise NotImplementedError()
 
         return PostTrainingJob(job_uuid=job_uuid)
+
+    def preference_optimize(
+        self,
+        job_uuid: str,
+        finetuned_model: URL,
+        dataset_id: str,
+        validation_dataset_id: str,
+        algorithm: RLHFAlgorithm,
+        algorithm_config: DPOAlignmentConfig,
+        optimizer_config: OptimizerConfig,
+        training_config: TrainingConfig,
+        hyperparam_search_config: Dict[str, Any],
+        logger_config: Dict[str, Any],
+    ) -> PostTrainingJob: ...
+
+    def get_training_jobs(self) -> List[PostTrainingJob]: ...
+
+    # sends SSE stream of logs
+    @webmethod(route="/post-training/job/logs")
+    def get_training_job_logstream(self, job_uuid: str) -> PostTrainingJobLogStream: ...
+
+    @webmethod(route="/post-training/job/status")
+    def get_training_job_status(
+        self, job_uuid: str
+    ) -> PostTrainingJobStatusResponse: ...
+
+    @webmethod(route="/post-training/job/cancel")
+    def cancel_training_job(self, job_uuid: str) -> None: ...
+
+    @webmethod(route="/post-training/job/artifacts")
+    def get_training_job_artifacts(
+        self, job_uuid: str
+    ) -> PostTrainingJobArtifactsResponse: ...

llama_stack/providers/inline/post_training/meta_reference/recipes/lora_finetuning_single_device.py

@@ -38,7 +38,7 @@ from torchtune.modules.peft import (
     set_trainable_params,
     validate_missing_and_unexpected_for_lora,
 )
-from torchtune.training.lr_scheduler import get_cosine_schedule_with_warmup
+from torchtune.training.lr_schedulers import get_cosine_schedule_with_warmup
 
 log = logging.getLogger(__name__)
 

llama_stack/providers/registry/post_training.py

@@ -12,6 +12,7 @@ from llama_stack.distribution.datatypes import *  # noqa: F403
 META_REFERENCE_DEPS = [
     "torch",
     "torchtune",
+    "torchao",
     "numpy",
 ]
 
@@ -24,5 +25,8 @@ def available_providers() -> List[ProviderSpec]:
             pip_packages=META_REFERENCE_DEPS,
             module="llama_stack.providers.inline.post_training.meta_reference",
             config_class="llama_stack.providers.inline.post_training.meta_reference.MetaReferencePostTrainingConfig",
+            api_dependencies=[
+                Api.datasetio,
+            ],
         ),
     ]

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

@@ -6,6 +6,8 @@ distribution_spec:
   providers:
     post_training:
     - inline::meta-reference
+    datasetio:
+    - remote::huggingface
     inference:
     - inline::meta-reference
     memory:

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

@@ -8,6 +8,8 @@ apis:
 - memory
 - safety
 - telemetry
+- datasetio
+- post_training
 providers:
   inference:
   - provider_id: meta-reference-inference
@@ -16,6 +18,10 @@ providers:
       model: ${env.INFERENCE_MODEL}
       max_seq_len: 4096
       checkpoint_dir: ${env.INFERENCE_CHECKPOINT_DIR:null}
+  datasetio:
+  - provider_id: huggingface-0
+    provider_type: remote::huggingface
+    config: {}
   memory:
   - provider_id: faiss
     provider_type: inline::faiss