llama-stack-api 0.4.3__py3-none-any.whl → 0.5.0rc1__py3-none-any.whl

llama_stack_api/post_training/__init__.py

@@ -0,0 +1,73 @@
+# 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.
+
+"""Post-Training API protocol and models.
+
+This module contains the Post-Training protocol definition.
+Pydantic models are defined in llama_stack_api.post_training.models.
+The FastAPI router is defined in llama_stack_api.post_training.fastapi_routes.
+"""
+
+# Import fastapi_routes for router factory access
+from . import fastapi_routes
+
+# Import protocol for re-export
+from .api import PostTraining
+
+# Import models for re-export
+from .models import (
+    AlgorithmConfig,
+    CancelTrainingJobRequest,
+    DataConfig,
+    DatasetFormat,
+    DPOAlignmentConfig,
+    DPOLossType,
+    EfficiencyConfig,
+    GetTrainingJobArtifactsRequest,
+    GetTrainingJobStatusRequest,
+    ListPostTrainingJobsResponse,
+    LoraFinetuningConfig,
+    OptimizerConfig,
+    OptimizerType,
+    PostTrainingJob,
+    PostTrainingJobArtifactsResponse,
+    PostTrainingJobLogStream,
+    PostTrainingJobStatusResponse,
+    PostTrainingRLHFRequest,
+    PreferenceOptimizeRequest,
+    QATFinetuningConfig,
+    RLHFAlgorithm,
+    SupervisedFineTuneRequest,
+    TrainingConfig,
+)
+
+__all__ = [
+    "PostTraining",
+    "AlgorithmConfig",
+    "CancelTrainingJobRequest",
+    "DataConfig",
+    "DatasetFormat",
+    "DPOAlignmentConfig",
+    "DPOLossType",
+    "EfficiencyConfig",
+    "GetTrainingJobArtifactsRequest",
+    "GetTrainingJobStatusRequest",
+    "ListPostTrainingJobsResponse",
+    "LoraFinetuningConfig",
+    "OptimizerConfig",
+    "OptimizerType",
+    "PostTrainingJob",
+    "PostTrainingJobArtifactsResponse",
+    "PostTrainingJobLogStream",
+    "PostTrainingJobStatusResponse",
+    "PostTrainingRLHFRequest",
+    "PreferenceOptimizeRequest",
+    "QATFinetuningConfig",
+    "RLHFAlgorithm",
+    "SupervisedFineTuneRequest",
+    "TrainingConfig",
+    "fastapi_routes",
+]

llama_stack_api/post_training/api.py

@@ -0,0 +1,36 @@
+# 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 typing import Protocol, runtime_checkable
+
+from .models import (
+    CancelTrainingJobRequest,
+    GetTrainingJobArtifactsRequest,
+    GetTrainingJobStatusRequest,
+    ListPostTrainingJobsResponse,
+    PostTrainingJob,
+    PostTrainingJobArtifactsResponse,
+    PostTrainingJobStatusResponse,
+    PreferenceOptimizeRequest,
+    SupervisedFineTuneRequest,
+)
+
+
+@runtime_checkable
+class PostTraining(Protocol):
+    async def supervised_fine_tune(self, request: SupervisedFineTuneRequest) -> PostTrainingJob: ...
+
+    async def preference_optimize(self, request: PreferenceOptimizeRequest) -> PostTrainingJob: ...
+
+    async def get_training_jobs(self) -> ListPostTrainingJobsResponse: ...
+
+    async def get_training_job_status(self, request: GetTrainingJobStatusRequest) -> PostTrainingJobStatusResponse: ...
+
+    async def cancel_training_job(self, request: CancelTrainingJobRequest) -> None: ...
+
+    async def get_training_job_artifacts(
+        self, request: GetTrainingJobArtifactsRequest
+    ) -> PostTrainingJobArtifactsResponse: ...

llama_stack_api/post_training/fastapi_routes.py

@@ -0,0 +1,116 @@
+# 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.
+
+"""FastAPI router for the Post-Training API.
+
+This module defines the FastAPI router for the Post-Training API using standard
+FastAPI route decorators.
+"""
+
+from typing import Annotated
+
+from fastapi import APIRouter, Body, Depends
+
+from llama_stack_api.router_utils import create_path_dependency, standard_responses
+from llama_stack_api.version import LLAMA_STACK_API_V1ALPHA
+
+from .api import PostTraining
+from .models import (
+    CancelTrainingJobRequest,
+    GetTrainingJobArtifactsRequest,
+    GetTrainingJobStatusRequest,
+    ListPostTrainingJobsResponse,
+    PostTrainingJob,
+    PostTrainingJobArtifactsResponse,
+    PostTrainingJobStatusResponse,
+    PreferenceOptimizeRequest,
+    SupervisedFineTuneRequest,
+)
+
+# Path parameter dependencies for single-field models
+get_training_job_status_request = create_path_dependency(GetTrainingJobStatusRequest)
+cancel_training_job_request = create_path_dependency(CancelTrainingJobRequest)
+get_training_job_artifacts_request = create_path_dependency(GetTrainingJobArtifactsRequest)
+
+
+def create_router(impl: PostTraining) -> APIRouter:
+    """Create a FastAPI router for the Post-Training API."""
+    router = APIRouter(
+        prefix=f"/{LLAMA_STACK_API_V1ALPHA}",
+        tags=["Post Training"],
+        responses=standard_responses,
+    )
+
+    @router.post(
+        "/post-training/supervised-fine-tune",
+        response_model=PostTrainingJob,
+        summary="Run supervised fine-tuning of a model.",
+        description="Run supervised fine-tuning of a model.",
+        responses={200: {"description": "A PostTrainingJob."}},
+    )
+    async def supervised_fine_tune(
+        request: Annotated[SupervisedFineTuneRequest, Body(...)],
+    ) -> PostTrainingJob:
+        return await impl.supervised_fine_tune(request)
+
+    @router.post(
+        "/post-training/preference-optimize",
+        response_model=PostTrainingJob,
+        summary="Run preference optimization of a model.",
+        description="Run preference optimization of a model.",
+        responses={200: {"description": "A PostTrainingJob."}},
+    )
+    async def preference_optimize(
+        request: Annotated[PreferenceOptimizeRequest, Body(...)],
+    ) -> PostTrainingJob:
+        return await impl.preference_optimize(request)
+
+    @router.get(
+        "/post-training/jobs",
+        response_model=ListPostTrainingJobsResponse,
+        summary="Get all training jobs.",
+        description="Get all training jobs.",
+        responses={200: {"description": "A ListPostTrainingJobsResponse."}},
+    )
+    async def get_training_jobs() -> ListPostTrainingJobsResponse:
+        return await impl.get_training_jobs()
+
+    @router.get(
+        "/post-training/job/status",
+        response_model=PostTrainingJobStatusResponse,
+        summary="Get the status of a training job.",
+        description="Get the status of a training job.",
+        responses={200: {"description": "A PostTrainingJobStatusResponse."}},
+    )
+    async def get_training_job_status(
+        request: Annotated[GetTrainingJobStatusRequest, Depends(get_training_job_status_request)],
+    ) -> PostTrainingJobStatusResponse:
+        return await impl.get_training_job_status(request)
+
+    @router.post(
+        "/post-training/job/cancel",
+        summary="Cancel a training job.",
+        description="Cancel a training job.",
+        responses={200: {"description": "Successfully cancelled the training job."}},
+    )
+    async def cancel_training_job(
+        request: Annotated[CancelTrainingJobRequest, Depends(cancel_training_job_request)],
+    ) -> None:
+        return await impl.cancel_training_job(request)
+
+    @router.get(
+        "/post-training/job/artifacts",
+        response_model=PostTrainingJobArtifactsResponse,
+        summary="Get the artifacts of a training job.",
+        description="Get the artifacts of a training job.",
+        responses={200: {"description": "A PostTrainingJobArtifactsResponse."}},
+    )
+    async def get_training_job_artifacts(
+        request: Annotated[GetTrainingJobArtifactsRequest, Depends(get_training_job_artifacts_request)],
+    ) -> PostTrainingJobArtifactsResponse:
+        return await impl.get_training_job_artifacts(request)
+
+    return router

llama_stack_api/post_training/models.py

@@ -0,0 +1,339 @@
+# 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.
+
+"""Pydantic models for Post-Training API requests and responses.
+
+This module defines the request and response models for the Post-Training API
+using Pydantic with Field descriptions for OpenAPI schema generation.
+"""
+
+from datetime import datetime
+from enum import Enum
+from typing import Annotated, Any, Literal
+
+from pydantic import BaseModel, Field
+
+from llama_stack_api.common.content_types import URL
+from llama_stack_api.common.job_types import JobStatus
+from llama_stack_api.common.training_types import Checkpoint
+from llama_stack_api.schema_utils import json_schema_type, register_schema
+
+
+@json_schema_type
+class OptimizerType(Enum):
+    """Available optimizer algorithms for training.
+    :cvar adam: Adaptive Moment Estimation optimizer
+    :cvar adamw: AdamW optimizer with weight decay
+    :cvar sgd: Stochastic Gradient Descent optimizer
+    """
+
+    adam = "adam"
+    adamw = "adamw"
+    sgd = "sgd"
+
+
+@json_schema_type
+class DatasetFormat(Enum):
+    """Format of the training dataset.
+    :cvar instruct: Instruction-following format with prompt and completion
+    :cvar dialog: Multi-turn conversation format with messages
+    """
+
+    instruct = "instruct"
+    dialog = "dialog"
+
+
+@json_schema_type
+class DataConfig(BaseModel):
+    """Configuration for training data and data loading.
+
+    :param dataset_id: Unique identifier for the training dataset
+    :param batch_size: Number of samples per training batch
+    :param shuffle: Whether to shuffle the dataset during training
+    :param data_format: Format of the dataset (instruct or dialog)
+    :param validation_dataset_id: (Optional) Unique identifier for the validation dataset
+    :param packed: (Optional) Whether to pack multiple samples into a single sequence for efficiency
+    :param train_on_input: (Optional) Whether to compute loss on input tokens as well as output tokens
+    """
+
+    dataset_id: str
+    batch_size: int
+    shuffle: bool
+    data_format: DatasetFormat
+    validation_dataset_id: str | None = None
+    packed: bool | None = False
+    train_on_input: bool | None = False
+
+
+@json_schema_type
+class OptimizerConfig(BaseModel):
+    """Configuration parameters for the optimization algorithm.
+
+    :param optimizer_type: Type of optimizer to use (adam, adamw, or sgd)
+    :param lr: Learning rate for the optimizer
+    :param weight_decay: Weight decay coefficient for regularization
+    :param num_warmup_steps: Number of steps for learning rate warmup
+    """
+
+    optimizer_type: OptimizerType
+    lr: float
+    weight_decay: float
+    num_warmup_steps: int
+
+
+@json_schema_type
+class EfficiencyConfig(BaseModel):
+    """Configuration for memory and compute efficiency optimizations.
+
+    :param enable_activation_checkpointing: (Optional) Whether to use activation checkpointing to reduce memory usage
+    :param enable_activation_offloading: (Optional) Whether to offload activations to CPU to save GPU memory
+    :param memory_efficient_fsdp_wrap: (Optional) Whether to use memory-efficient FSDP wrapping
+    :param fsdp_cpu_offload: (Optional) Whether to offload FSDP parameters to CPU
+    """
+
+    enable_activation_checkpointing: bool | None = False
+    enable_activation_offloading: bool | None = False
+    memory_efficient_fsdp_wrap: bool | None = False
+    fsdp_cpu_offload: bool | None = False
+
+
+@json_schema_type
+class TrainingConfig(BaseModel):
+    """Comprehensive configuration for the training process.
+
+    :param n_epochs: Number of training epochs to run
+    :param max_steps_per_epoch: Maximum number of steps to run per epoch
+    :param gradient_accumulation_steps: Number of steps to accumulate gradients before updating
+    :param max_validation_steps: (Optional) Maximum number of validation steps per epoch
+    :param data_config: (Optional) Configuration for data loading and formatting
+    :param optimizer_config: (Optional) Configuration for the optimization algorithm
+    :param efficiency_config: (Optional) Configuration for memory and compute optimizations
+    :param dtype: (Optional) Data type for model parameters (bf16, fp16, fp32)
+    """
+
+    n_epochs: int
+    max_steps_per_epoch: int = 1
+    gradient_accumulation_steps: int = 1
+    max_validation_steps: int | None = 1
+    data_config: DataConfig | None = None
+    optimizer_config: OptimizerConfig | None = None
+    efficiency_config: EfficiencyConfig | None = None
+    dtype: str | None = "bf16"
+
+
+@json_schema_type
+class LoraFinetuningConfig(BaseModel):
+    """Configuration for Low-Rank Adaptation (LoRA) fine-tuning.
+
+    :param type: Algorithm type identifier, always "LoRA"
+    :param lora_attn_modules: List of attention module names to apply LoRA to
+    :param apply_lora_to_mlp: Whether to apply LoRA to MLP layers
+    :param apply_lora_to_output: Whether to apply LoRA to output projection layers
+    :param rank: Rank of the LoRA adaptation (lower rank = fewer parameters)
+    :param alpha: LoRA scaling parameter that controls adaptation strength
+    :param use_dora: (Optional) Whether to use DoRA (Weight-Decomposed Low-Rank Adaptation)
+    :param quantize_base: (Optional) Whether to quantize the base model weights
+    """
+
+    type: Literal["LoRA"] = "LoRA"
+    lora_attn_modules: list[str]
+    apply_lora_to_mlp: bool
+    apply_lora_to_output: bool
+    rank: int
+    alpha: int
+    use_dora: bool | None = False
+    quantize_base: bool | None = False
+
+
+@json_schema_type
+class QATFinetuningConfig(BaseModel):
+    """Configuration for Quantization-Aware Training (QAT) fine-tuning.
+
+    :param type: Algorithm type identifier, always "QAT"
+    :param quantizer_name: Name of the quantization algorithm to use
+    :param group_size: Size of groups for grouped quantization
+    """
+
+    type: Literal["QAT"] = "QAT"
+    quantizer_name: str
+    group_size: int
+
+
+AlgorithmConfig = Annotated[LoraFinetuningConfig | QATFinetuningConfig, Field(discriminator="type")]
+register_schema(AlgorithmConfig, name="AlgorithmConfig")
+
+
+@json_schema_type
+class PostTrainingJobLogStream(BaseModel):
+    """Stream of logs from a finetuning job.
+
+    :param job_uuid: Unique identifier for the training job
+    :param log_lines: List of log message strings from the training process
+    """
+
+    job_uuid: str
+    log_lines: list[str]
+
+
+@json_schema_type
+class RLHFAlgorithm(Enum):
+    """Available reinforcement learning from human feedback algorithms.
+    :cvar dpo: Direct Preference Optimization algorithm
+    """
+
+    dpo = "dpo"
+
+
+@json_schema_type
+class DPOLossType(Enum):
+    sigmoid = "sigmoid"
+    hinge = "hinge"
+    ipo = "ipo"
+    kto_pair = "kto_pair"
+
+
+@json_schema_type
+class DPOAlignmentConfig(BaseModel):
+    """Configuration for Direct Preference Optimization (DPO) alignment.
+
+    :param beta: Temperature parameter for the DPO loss
+    :param loss_type: The type of loss function to use for DPO
+    """
+
+    beta: float
+    loss_type: DPOLossType = DPOLossType.sigmoid
+
+
+@json_schema_type
+class PostTrainingRLHFRequest(BaseModel):
+    """Request to finetune a model using reinforcement learning from human feedback.
+
+    :param job_uuid: Unique identifier for the training job
+    :param finetuned_model: URL or path to the base model to fine-tune
+    :param dataset_id: Unique identifier for the training dataset
+    :param validation_dataset_id: Unique identifier for the validation dataset
+    :param algorithm: RLHF algorithm to use for training
+    :param algorithm_config: Configuration parameters for the RLHF algorithm
+    :param optimizer_config: Configuration parameters for the optimization algorithm
+    :param training_config: Configuration parameters for the training process
+    :param hyperparam_search_config: Configuration for hyperparameter search
+    :param logger_config: Configuration for training logging
+    """
+
+    job_uuid: str
+
+    finetuned_model: URL
+
+    dataset_id: str
+    validation_dataset_id: str
+
+    algorithm: RLHFAlgorithm
+    algorithm_config: DPOAlignmentConfig
+
+    optimizer_config: OptimizerConfig
+    training_config: TrainingConfig
+
+    # TODO: define these
+    hyperparam_search_config: dict[str, Any]
+    logger_config: dict[str, Any]
+
+
+@json_schema_type
+class PostTrainingJob(BaseModel):
+    job_uuid: str
+
+
+@json_schema_type
+class PostTrainingJobStatusResponse(BaseModel):
+    """Status of a finetuning job.
+
+    :param job_uuid: Unique identifier for the training job
+    :param status: Current status of the training job
+    :param scheduled_at: (Optional) Timestamp when the job was scheduled
+    :param started_at: (Optional) Timestamp when the job execution began
+    :param completed_at: (Optional) Timestamp when the job finished, if completed
+    :param resources_allocated: (Optional) Information about computational resources allocated to the job
+    :param checkpoints: List of model checkpoints created during training
+    """
+
+    job_uuid: str
+    status: JobStatus
+
+    scheduled_at: datetime | None = None
+    started_at: datetime | None = None
+    completed_at: datetime | None = None
+
+    resources_allocated: dict[str, Any] | None = None
+
+    checkpoints: list[Checkpoint] = Field(default_factory=list)
+
+
+@json_schema_type
+class ListPostTrainingJobsResponse(BaseModel):
+    data: list[PostTrainingJob]
+
+
+@json_schema_type
+class PostTrainingJobArtifactsResponse(BaseModel):
+    """Artifacts of a finetuning job.
+
+    :param job_uuid: Unique identifier for the training job
+    :param checkpoints: List of model checkpoints created during training
+    """
+
+    job_uuid: str
+    checkpoints: list[Checkpoint] = Field(default_factory=list)
+
+    # TODO(ashwin): metrics, evals
+
+
+@json_schema_type
+class SupervisedFineTuneRequest(BaseModel):
+    """Request to run supervised fine-tuning of a model."""
+
+    job_uuid: str = Field(..., description="The UUID of the job to create.")
+    training_config: TrainingConfig = Field(..., description="The training configuration.")
+    hyperparam_search_config: dict[str, Any] = Field(..., description="The hyperparam search configuration.")
+    logger_config: dict[str, Any] = Field(..., description="The logger configuration.")
+    model: str | None = Field(
+        default=None,
+        description="Model descriptor for training if not in provider config",
+    )
+    checkpoint_dir: str | None = Field(default=None, description="The directory to save checkpoint(s) to.")
+    algorithm_config: AlgorithmConfig | None = Field(default=None, description="The algorithm configuration.")
+
+
+@json_schema_type
+class PreferenceOptimizeRequest(BaseModel):
+    """Request to run preference optimization of a model."""
+
+    job_uuid: str = Field(..., description="The UUID of the job to create.")
+    finetuned_model: str = Field(..., description="The model to fine-tune.")
+    algorithm_config: DPOAlignmentConfig = Field(..., description="The algorithm configuration.")
+    training_config: TrainingConfig = Field(..., description="The training configuration.")
+    hyperparam_search_config: dict[str, Any] = Field(..., description="The hyperparam search configuration.")
+    logger_config: dict[str, Any] = Field(..., description="The logger configuration.")
+
+
+@json_schema_type
+class GetTrainingJobStatusRequest(BaseModel):
+    """Request to get the status of a training job."""
+
+    job_uuid: str = Field(..., description="The UUID of the job to get the status of.")
+
+
+@json_schema_type
+class CancelTrainingJobRequest(BaseModel):
+    """Request to cancel a training job."""
+
+    job_uuid: str = Field(..., description="The UUID of the job to cancel.")
+
+
+@json_schema_type
+class GetTrainingJobArtifactsRequest(BaseModel):
+    """Request to get the artifacts of a training job."""
+
+    job_uuid: str = Field(..., description="The UUID of the job to get the artifacts of.")

llama_stack_api/prompts/__init__.py

@@ -0,0 +1,47 @@
+# 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.
+
+"""Prompts API protocol and models.
+
+This module contains the Prompts protocol definition.
+Pydantic models are defined in llama_stack_api.prompts.models.
+The FastAPI router is defined in llama_stack_api.prompts.fastapi_routes.
+"""
+
+# Import fastapi_routes for router factory access
+from . import fastapi_routes
+
+# Import protocol for FastAPI router
+from .api import Prompts
+
+# Import models for re-export
+from .models import (
+    CreatePromptRequest,
+    DeletePromptRequest,
+    GetPromptRequest,
+    ListPromptsResponse,
+    ListPromptVersionsRequest,
+    Prompt,
+    SetDefaultVersionBodyRequest,
+    SetDefaultVersionRequest,
+    UpdatePromptBodyRequest,
+    UpdatePromptRequest,
+)
+
+__all__ = [
+    "CreatePromptRequest",
+    "DeletePromptRequest",
+    "GetPromptRequest",
+    "ListPromptVersionsRequest",
+    "ListPromptsResponse",
+    "Prompt",
+    "Prompts",
+    "SetDefaultVersionBodyRequest",
+    "SetDefaultVersionRequest",
+    "UpdatePromptBodyRequest",
+    "UpdatePromptRequest",
+    "fastapi_routes",
+]

llama_stack_api/prompts/api.py

@@ -0,0 +1,44 @@
+# 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.
+
+"""Prompts API protocol definition.
+
+This module contains the Prompts protocol definition.
+Pydantic models are defined in llama_stack_api.prompts.models.
+The FastAPI router is defined in llama_stack_api.prompts.fastapi_routes.
+"""
+
+from typing import Protocol, runtime_checkable
+
+from .models import (
+    CreatePromptRequest,
+    DeletePromptRequest,
+    GetPromptRequest,
+    ListPromptsResponse,
+    ListPromptVersionsRequest,
+    Prompt,
+    SetDefaultVersionRequest,
+    UpdatePromptRequest,
+)
+
+
+@runtime_checkable
+class Prompts(Protocol):
+    """Protocol for prompt management operations."""
+
+    async def list_prompts(self) -> ListPromptsResponse: ...
+
+    async def list_prompt_versions(self, request: ListPromptVersionsRequest) -> ListPromptsResponse: ...
+
+    async def get_prompt(self, request: GetPromptRequest) -> Prompt: ...
+
+    async def create_prompt(self, request: CreatePromptRequest) -> Prompt: ...
+
+    async def update_prompt(self, request: UpdatePromptRequest) -> Prompt: ...
+
+    async def delete_prompt(self, request: DeletePromptRequest) -> None: ...
+
+    async def set_default_version(self, request: SetDefaultVersionRequest) -> Prompt: ...