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

375 lines
13 KiB
Python

# 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 datetime import datetime
from enum import Enum
from typing import Annotated, Any, Literal, Protocol
from pydantic import BaseModel, Field
from llama_stack.apis.common.content_types import URL
from llama_stack.apis.common.job_types import JobStatus
from llama_stack.apis.common.training_types import Checkpoint
from llama_stack.schema_utils import json_schema_type, register_schema, webmethod
@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 reward_scale: Scaling factor for the reward signal
:param reward_clip: Maximum absolute value for reward clipping
:param epsilon: Small value added for numerical stability
:param gamma: Discount factor for future rewards
:param beta: Temperature parameter for the DPO loss
:param loss_type: The type of loss function to use for DPO
"""
reward_scale: float
reward_clip: float
epsilon: float
gamma: float
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]
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)
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
class PostTraining(Protocol):
@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: str | None = Field(
default=None,
description="Model descriptor for training if not in provider config`",
),
checkpoint_dir: str | None = None,
algorithm_config: AlgorithmConfig | None = None,
) -> PostTrainingJob:
"""Run supervised fine-tuning of a model.
:param job_uuid: The UUID of the job to create.
:param training_config: The training configuration.
:param hyperparam_search_config: The hyperparam search configuration.
:param logger_config: The logger configuration.
:param model: The model to fine-tune.
:param checkpoint_dir: The directory to save checkpoint(s) to.
:param algorithm_config: The algorithm configuration.
:returns: A 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:
"""Run preference optimization of a model.
:param job_uuid: The UUID of the job to create.
:param finetuned_model: The model to fine-tune.
:param algorithm_config: The algorithm configuration.
:param training_config: The training configuration.
:param hyperparam_search_config: The hyperparam search configuration.
:param logger_config: The logger configuration.
:returns: A PostTrainingJob.
"""
...
@webmethod(route="/post-training/jobs", method="GET")
async def get_training_jobs(self) -> ListPostTrainingJobsResponse:
"""Get all training jobs.
:returns: A ListPostTrainingJobsResponse.
"""
...
@webmethod(route="/post-training/job/status", method="GET")
async def get_training_job_status(self, job_uuid: str) -> PostTrainingJobStatusResponse:
"""Get the status of a training job.
:param job_uuid: The UUID of the job to get the status of.
:returns: A PostTrainingJobStatusResponse.
"""
...
@webmethod(route="/post-training/job/cancel", method="POST")
async def cancel_training_job(self, job_uuid: str) -> None:
"""Cancel a training job.
:param job_uuid: The UUID of the job to cancel.
"""
...
@webmethod(route="/post-training/job/artifacts", method="GET")
async def get_training_job_artifacts(self, job_uuid: str) -> PostTrainingJobArtifactsResponse:
"""Get the artifacts of a training job.
:param job_uuid: The UUID of the job to get the artifacts of.
:returns: A PostTrainingJobArtifactsResponse.
"""
...