together 1.5.17__py3-none-any.whl → 2.0.0a8__py3-none-any.whl

together/lib/resources/fine_tuning.py

@@ -0,0 +1,280 @@
+from __future__ import annotations
+
+from typing import TYPE_CHECKING, Literal
+
+from rich import print as rprint
+
+from together.lib.utils import log_warn_once
+
+if TYPE_CHECKING:
+    from together import Together, AsyncTogether
+from together.lib.types.fine_tuning import (
+    TrainingType,
+    FinetuneRequest,
+    FullTrainingType,
+    LoRATrainingType,
+    CosineLRScheduler,
+    LinearLRScheduler,
+    TrainingMethodDPO,
+    TrainingMethodSFT,
+    FinetuneLRScheduler,
+    CosineLRSchedulerArgs,
+    LinearLRSchedulerArgs,
+    FinetuneTrainingLimits,
+)
+
+AVAILABLE_TRAINING_METHODS = {
+    "sft",
+    "dpo",
+}
+
+
+def create_finetune_request(
+    model_limits: FinetuneTrainingLimits,
+    training_file: str,
+    model: str | None = None,
+    n_epochs: int = 1,
+    validation_file: str | None = "",
+    n_evals: int | None = 0,
+    n_checkpoints: int | None = 1,
+    batch_size: int | Literal["max"] = "max",
+    learning_rate: float | None = 0.00001,
+    lr_scheduler_type: Literal["linear", "cosine"] = "cosine",
+    min_lr_ratio: float | None = 0.0,
+    scheduler_num_cycles: float = 0.5,
+    warmup_ratio: float | None = None,
+    max_grad_norm: float = 1.0,
+    weight_decay: float | None = 0.0,
+    lora: bool = False,
+    lora_r: int | None = None,
+    lora_dropout: float | None = 0,
+    lora_alpha: float | None = None,
+    lora_trainable_modules: str | None = "all-linear",
+    suffix: str | None = None,
+    wandb_api_key: str | None = None,
+    wandb_base_url: str | None = None,
+    wandb_project_name: str | None = None,
+    wandb_name: str | None = None,
+    train_on_inputs: bool | Literal["auto"] | None = None,
+    training_method: str = "sft",
+    dpo_beta: float | None = None,
+    dpo_normalize_logratios_by_length: bool = False,
+    rpo_alpha: float | None = None,
+    simpo_gamma: float | None = None,
+    from_checkpoint: str | None = None,
+    from_hf_model: str | None = None,
+    hf_model_revision: str | None = None,
+    hf_api_token: str | None = None,
+    hf_output_repo_name: str | None = None,
+) -> FinetuneRequest:
+    if model is not None and from_checkpoint is not None:
+        raise ValueError("You must specify either a model or a checkpoint to start a job from, not both")
+
+    if model is None and from_checkpoint is None:
+        raise ValueError("You must specify either a model or a checkpoint")
+
+    if from_checkpoint is not None and from_hf_model is not None:
+        raise ValueError(
+            "You must specify either a Hugging Face Hub model or a previous checkpoint from "
+            "Together to start a job from, not both"
+        )
+
+    if from_hf_model is not None and model is None:
+        raise ValueError("You must specify the base model to fine-tune a model from the Hugging Face Hub")
+
+    model_or_checkpoint = model or from_checkpoint
+
+    if warmup_ratio is None:
+        warmup_ratio = 0.0
+
+    training_type: TrainingType = FullTrainingType()
+    if lora:
+        if model_limits.lora_training is None:
+            raise ValueError(f"LoRA adapters are not supported for the selected model ({model_or_checkpoint}).")
+
+        if lora_dropout is not None:
+            if not 0 <= lora_dropout < 1.0:
+                raise ValueError("LoRA dropout must be in [0, 1) range.")
+
+        lora_r = lora_r if lora_r is not None else model_limits.lora_training.max_rank
+        lora_alpha = lora_alpha if lora_alpha is not None else lora_r * 2
+        training_type = LoRATrainingType(
+            lora_r=lora_r,
+            lora_alpha=int(lora_alpha),
+            lora_dropout=lora_dropout or 0.0,
+            lora_trainable_modules=lora_trainable_modules or "all-linear",
+        )
+
+        max_batch_size = model_limits.lora_training.max_batch_size
+        min_batch_size = model_limits.lora_training.min_batch_size
+        max_batch_size_dpo = model_limits.lora_training.max_batch_size_dpo
+    else:
+        if model_limits.full_training is None:
+            raise ValueError(f"Full training is not supported for the selected model ({model_or_checkpoint}).")
+
+        max_batch_size = model_limits.full_training.max_batch_size
+        min_batch_size = model_limits.full_training.min_batch_size
+        max_batch_size_dpo = model_limits.full_training.max_batch_size_dpo
+
+    if batch_size != "max":
+        if training_method == "sft":
+            if batch_size > max_batch_size:
+                raise ValueError(
+                    f"Requested batch size of {batch_size} is higher that the maximum allowed value of {max_batch_size}."
+                )
+        elif training_method == "dpo":
+            if batch_size > max_batch_size_dpo:
+                raise ValueError(
+                    f"Requested batch size of {batch_size} is higher that the maximum allowed value of {max_batch_size_dpo}."
+                )
+
+        if batch_size < min_batch_size:
+            raise ValueError(
+                f"Requested batch size of {batch_size} is lower that the minimum allowed value of {min_batch_size}."
+            )
+
+    if warmup_ratio > 1 or warmup_ratio < 0:
+        raise ValueError(f"Warmup ratio should be between 0 and 1 (got {warmup_ratio})")
+
+    if min_lr_ratio is not None and (min_lr_ratio > 1 or min_lr_ratio < 0):
+        raise ValueError(f"Min learning rate ratio should be between 0 and 1 (got {min_lr_ratio})")
+
+    if max_grad_norm < 0:
+        raise ValueError(f"Max gradient norm should be non-negative (got {max_grad_norm})")
+
+    if weight_decay is not None and (weight_decay < 0):
+        raise ValueError(f"Weight decay should be non-negative (got {weight_decay})")
+
+    if training_method not in AVAILABLE_TRAINING_METHODS:
+        raise ValueError(f"training_method must be one of {', '.join(AVAILABLE_TRAINING_METHODS)}")
+
+    if train_on_inputs is not None and training_method != "sft":
+        raise ValueError("train_on_inputs is only supported for SFT training")
+
+    if dpo_beta is not None and training_method != "dpo":
+        raise ValueError("dpo_beta is only supported for DPO training")
+    if dpo_normalize_logratios_by_length and training_method != "dpo":
+        raise ValueError("dpo_normalize_logratios_by_length=True is only supported for DPO training")
+    if rpo_alpha is not None:
+        if training_method != "dpo":
+            raise ValueError("rpo_alpha is only supported for DPO training")
+        if not rpo_alpha >= 0.0:
+            raise ValueError(f"rpo_alpha should be non-negative (got {rpo_alpha})")
+
+    if simpo_gamma is not None:
+        if training_method != "dpo":
+            raise ValueError("simpo_gamma is only supported for DPO training")
+        if not simpo_gamma >= 0.0:
+            raise ValueError(f"simpo_gamma should be non-negative (got {simpo_gamma})")
+
+    lr_scheduler: FinetuneLRScheduler
+    if lr_scheduler_type == "cosine":
+        if scheduler_num_cycles <= 0.0:
+            raise ValueError(f"Number of cycles should be greater than 0 (got {scheduler_num_cycles})")
+
+        lr_scheduler = CosineLRScheduler(
+            lr_scheduler_args=CosineLRSchedulerArgs(min_lr_ratio=min_lr_ratio, num_cycles=scheduler_num_cycles),
+        )
+    else:
+        lr_scheduler = LinearLRScheduler(
+            lr_scheduler_args=LinearLRSchedulerArgs(min_lr_ratio=min_lr_ratio),
+        )
+
+    training_method_cls: TrainingMethodSFT | TrainingMethodDPO
+    if training_method == "sft":
+        if train_on_inputs is None:
+            log_warn_once("train_on_inputs is not set for SFT training, it will be set to 'auto'")
+            train_on_inputs = "auto"
+        training_method_cls = TrainingMethodSFT(train_on_inputs=train_on_inputs)
+    elif training_method == "dpo":
+        if simpo_gamma is not None and simpo_gamma > 0:
+            dpo_reference_free = True
+            dpo_normalize_logratios_by_length = True
+            rprint(
+                f"Parameter simpo_gamma was set to {simpo_gamma}. "
+                "SimPO training detected. Reference logits will not be used "
+                "and length normalization of log-probabilities will be enabled."
+            )
+        else:
+            dpo_reference_free = False
+
+        training_method_cls = TrainingMethodDPO(
+            dpo_beta=dpo_beta,
+            dpo_normalize_logratios_by_length=dpo_normalize_logratios_by_length,
+            dpo_reference_free=dpo_reference_free,
+            rpo_alpha=rpo_alpha,
+            simpo_gamma=simpo_gamma,
+        )
+
+    finetune_request = FinetuneRequest(
+        model=model,
+        training_file=training_file,
+        validation_file=validation_file,
+        n_epochs=n_epochs,
+        n_evals=n_evals,
+        n_checkpoints=n_checkpoints,
+        batch_size=batch_size,
+        learning_rate=learning_rate or 0.00001,
+        lr_scheduler=lr_scheduler,
+        warmup_ratio=warmup_ratio,
+        max_grad_norm=max_grad_norm,
+        weight_decay=weight_decay or 0.0,
+        training_type=training_type,
+        suffix=suffix,
+        wandb_key=wandb_api_key,
+        wandb_base_url=wandb_base_url,
+        wandb_project_name=wandb_project_name,
+        wandb_name=wandb_name,
+        training_method=training_method_cls,  # pyright: ignore[reportPossiblyUnboundVariable]
+        from_checkpoint=from_checkpoint,
+        from_hf_model=from_hf_model,
+        hf_model_revision=hf_model_revision,
+        hf_api_token=hf_api_token,
+        hf_output_repo_name=hf_output_repo_name,
+    )
+
+    return finetune_request
+
+
+def get_model_limits(client: Together, model: str) -> FinetuneTrainingLimits:
+    """
+    Requests training limits for a specific model
+
+    Args:
+        model_name (str): Name of the model to get limits for
+
+    Returns:
+    FinetuneTrainingLimits: Object containing training limits for the model
+    """
+
+    response = client.get(
+        "/fine-tunes/models/limits",
+        cast_to=FinetuneTrainingLimits,
+        options={
+            "params": {"model_name": model},
+        },
+    )
+
+    return response
+
+
+async def async_get_model_limits(client: AsyncTogether, model: str) -> FinetuneTrainingLimits:
+    """
+    Requests training limits for a specific model
+
+    Args:
+        model_name (str): Name of the model to get limits for
+
+    Returns:
+    FinetuneTrainingLimits: Object containing training limits for the model
+    """
+
+    response = await client.get(
+        "/fine-tunes/models/limits",
+        cast_to=FinetuneTrainingLimits,
+        options={
+            "params": {"model_name": model},
+        },
+    )
+
+    return response

together/lib/resources/models.py

@@ -0,0 +1,35 @@
+from typing import List
+from typing_extensions import TypeAlias
+
+from together import Together
+from together._models import BaseModel
+from together.types.model_list_response import ModelListResponse
+
+
+class DedicatedModel(BaseModel):
+    name: str
+    id: str
+
+
+ModelList: TypeAlias = List[DedicatedModel]
+
+
+def filter_by_dedicated_models(client: Together, models: ModelListResponse) -> ModelListResponse:
+    """
+    Filter models based on dedicated model response.
+
+    Args:
+    models (List[BaseModel]): List of all models
+    dedicated_response (APIResponse): Response from autoscale models endpoint
+
+    Returns:
+    List[BaseModel]: Filtered list of models
+    """
+    dedicated_models = client.get("/autoscale/models", cast_to=ModelList)
+
+    # Create a set of dedicated model names for efficient lookup
+    dedicated_model_names = {model.name for model in dedicated_models}
+
+    # Filter models to only include those in dedicated_model_names
+    # Note: The model.id from ModelObject matches the name field in the autoscale response
+    return [model for model in models if model.id in dedicated_model_names]

together/lib/types/__init__.py

@@ -0,0 +1,13 @@
+from .error import (
+    DownloadError,
+    FileTypeError,
+)
+from .fine_tuning import (
+    FinetuneTrainingLimits,
+)
+
+__all__ = [
+    "FinetuneTrainingLimits",
+    "DownloadError",
+    "FileTypeError",
+]

together/lib/types/error.py

@@ -0,0 +1,9 @@
+from ..._exceptions import TogetherError
+
+
+class DownloadError(TogetherError):
+    pass
+
+
+class FileTypeError(TogetherError):
+    pass