llama-stack-mirror/llama_stack/apis/post_training/post_training.py

# 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 enum import Enum
from typing import Any, Dict, List, Literal, Optional, Protocol, Union

from pydantic import BaseModel, Field
from typing_extensions import Annotated

from llama_stack.apis.common.content_types import URL
from llama_stack.apis.common.job_types import BaseJob
from llama_stack.schema_utils import json_schema_type, register_schema, webmethod


@json_schema_type
class OptimizerType(Enum):
    adam = "adam"
    adamw = "adamw"
    sgd = "sgd"


@json_schema_type
class DatasetFormat(Enum):
    instruct = "instruct"
    dialog = "dialog"


@json_schema_type
class DataConfig(BaseModel):
    dataset_id: str
    batch_size: int
    shuffle: bool
    data_format: DatasetFormat
    validation_dataset_id: Optional[str] = None
    packed: Optional[bool] = False
    train_on_input: Optional[bool] = False


@json_schema_type
class OptimizerConfig(BaseModel):
    optimizer_type: OptimizerType
    lr: float
    weight_decay: float
    num_warmup_steps: int


@json_schema_type
class EfficiencyConfig(BaseModel):
    enable_activation_checkpointing: Optional[bool] = False
    enable_activation_offloading: Optional[bool] = False
    memory_efficient_fsdp_wrap: Optional[bool] = False
    fsdp_cpu_offload: Optional[bool] = False


@json_schema_type
class TrainingConfig(BaseModel):
    n_epochs: int
    max_steps_per_epoch: int = 1
    gradient_accumulation_steps: int = 1
    max_validation_steps: Optional[int] = 1
    data_config: Optional[DataConfig] = None
    optimizer_config: Optional[OptimizerConfig] = None
    efficiency_config: Optional[EfficiencyConfig] = None
    dtype: Optional[str] = "bf16"


@json_schema_type
class LoraFinetuningConfig(BaseModel):
    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: Optional[bool] = False
    quantize_base: Optional[bool] = False


@json_schema_type
class QATFinetuningConfig(BaseModel):
    type: Literal["QAT"] = "QAT"
    quantizer_name: str
    group_size: int


AlgorithmConfig = Annotated[Union[LoraFinetuningConfig, QATFinetuningConfig], Field(discriminator="type")]
register_schema(AlgorithmConfig, name="AlgorithmConfig")


@json_schema_type
class RLHFAlgorithm(Enum):
    dpo = "dpo"


@json_schema_type
class DPOAlignmentConfig(BaseModel):
    reward_scale: float
    reward_clip: float
    epsilon: float
    gamma: float


@json_schema_type
class PostTrainingRLHFRequest(BaseModel):
    """Request to finetune a model."""

    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(BaseJob, BaseModel):
    type: Literal["post-training"] = "post-training"


class ListPostTrainingJobsResponse(BaseModel):
    data: list[PostTrainingJob]


class PostTraining(Protocol):
    # This is how you create a new job - POST against the root endpoint
    @webmethod(route="/post-training/supervised-fine-tune", method="POST")
    async def supervised_fine_tune(
        self,
        job_uuid: str,
        training_config: TrainingConfig,
        hyperparam_search_config: Dict[str, Any],
        logger_config: Dict[str, Any],
        model: Optional[str] = Field(
            default=None,
            description="Model descriptor for training if not in provider config`",
        ),
        checkpoint_dir: Optional[str] = None,
        algorithm_config: Optional[AlgorithmConfig] = None,
    ) -> PostTrainingJob: ...

    @webmethod(route="/post-training/preference-optimize", method="POST")
    async def preference_optimize(
        self,
        job_uuid: str,
        finetuned_model: str,
        algorithm_config: DPOAlignmentConfig,
        training_config: TrainingConfig,
        hyperparam_search_config: Dict[str, Any],
        logger_config: Dict[str, Any],
    ) -> PostTrainingJob: ...

    # CRUD operations on running jobs
    @webmethod(route="/post-training/jobs/{job_id:path}", method="GET")
    async def get_post_training_job(self, job_id: str) -> PostTrainingJob: ...

    @webmethod(route="/post-training/jobs", method="GET")
    async def list_post_training_jobs(self) -> ListPostTrainingJobsResponse: ...

    @webmethod(route="/post-training/jobs/{job_id:path}", method="POST")
    async def update_post_training_job(self, job: PostTrainingJob) -> PostTrainingJob: ...

    @webmethod(route="/post-training/job/{job_id:path}", method="DELETE")
    async def delete_post_training_job(self, job_id: str) -> None: ...

    # Note: pause/resume/cancel are achieved as follows:
    # - POST with status=paused
    # - POST with status=resuming
    # - POST with status=cancelled