together 1.4.6__tar.gz → 1.5.3__tar.gz

{together-1.4.6 → together-1.5.3}/PKG-INFO

@@ -1,15 +1,14 @@
 Metadata-Version: 2.3
 Name: together
-Version: 1.4.6
+Version: 1.5.3
 Summary: Python client for Together's Cloud Platform!
 License: Apache-2.0
 Author: Together AI
 Author-email: support@together.ai
-Requires-Python: >=3.9,<4.0
+Requires-Python: >=3.10,<4.0
 Classifier: License :: OSI Approved :: Apache Software License
 Classifier: Operating System :: POSIX :: Linux
 Classifier: Programming Language :: Python :: 3
-Classifier: Programming Language :: Python :: 3.9
 Classifier: Programming Language :: Python :: 3.10
 Classifier: Programming Language :: Python :: 3.11
 Classifier: Programming Language :: Python :: 3.12
@@ -45,7 +44,7 @@ Description-Content-Type: text/markdown
 [![Discord](https://dcbadge.vercel.app/api/server/9Rk6sSeWEG?style=flat&compact=true)](https://discord.com/invite/9Rk6sSeWEG)
 [![Twitter](https://img.shields.io/twitter/url/https/twitter.com/togethercompute.svg?style=social&label=Follow%20%40togethercompute)](https://twitter.com/togethercompute)
 
-The [Together Python API Library](https://pypi.org/project/together/) is the official Python client for Together's API platform, providing a convenient way for interacting with the REST APIs and enables easy integrations with Python 3.8+ applications with easy to use synchronous and asynchronous clients.
+The [Together Python API Library](https://pypi.org/project/together/) is the official Python client for Together's API platform, providing a convenient way for interacting with the REST APIs and enables easy integrations with Python 3.10+ applications with easy to use synchronous and asynchronous clients.
 
 
 

{together-1.4.6 → together-1.5.3}/README.md

@@ -10,7 +10,7 @@
 [![Discord](https://dcbadge.vercel.app/api/server/9Rk6sSeWEG?style=flat&compact=true)](https://discord.com/invite/9Rk6sSeWEG)
 [![Twitter](https://img.shields.io/twitter/url/https/twitter.com/togethercompute.svg?style=social&label=Follow%20%40togethercompute)](https://twitter.com/togethercompute)
 
-The [Together Python API Library](https://pypi.org/project/together/) is the official Python client for Together's API platform, providing a convenient way for interacting with the REST APIs and enables easy integrations with Python 3.8+ applications with easy to use synchronous and asynchronous clients.
+The [Together Python API Library](https://pypi.org/project/together/) is the official Python client for Together's API platform, providing a convenient way for interacting with the REST APIs and enables easy integrations with Python 3.10+ applications with easy to use synchronous and asynchronous clients.
 
 
 

{together-1.4.6 → together-1.5.3}/pyproject.toml

@@ -12,7 +12,7 @@ build-backend = "poetry.masonry.api"
 
 [tool.poetry]
 name = "together"
-version = "1.4.6"
+version = "1.5.3"
 authors = [
     "Together AI <support@together.ai>"
 ]
@@ -28,7 +28,7 @@ repository = "https://github.com/togethercomputer/together-python"
 homepage = "https://github.com/togethercomputer/together-python"
 
 [tool.poetry.dependencies]
-python = "^3.9"
+python = "^3.10"
 typer = ">=0.9,<0.16"
 requests = "^2.31.0"
 rich = "^13.8.1"

{together-1.4.6 → together-1.5.3}/src/together/cli/api/finetune.py

@@ -58,30 +58,50 @@ def fine_tuning(ctx: click.Context) -> None:
 @fine_tuning.command()
 @click.pass_context
 @click.option(
-    "--training-file", type=str, required=True, help="Training file ID from Files API"
+    "--training-file",
+    "-t",
+    type=str,
+    required=True,
+    help="Training file ID from Files API",
+)
+@click.option("--model", "-m", type=str, help="Base model name")
+@click.option(
+    "--n-epochs", "-ne", type=int, default=1, help="Number of epochs to train for"
 )
-@click.option("--model", type=str, required=True, help="Base model name")
-@click.option("--n-epochs", type=int, default=1, help="Number of epochs to train for")
 @click.option(
     "--validation-file", type=str, default="", help="Validation file ID from Files API"
 )
 @click.option("--n-evals", type=int, default=0, help="Number of evaluation loops")
 @click.option(
-    "--n-checkpoints", type=int, default=1, help="Number of checkpoints to save"
+    "--n-checkpoints", "-c", type=int, default=1, help="Number of checkpoints to save"
+)
+@click.option(
+    "--batch-size", "-b", type=INT_WITH_MAX, default="max", help="Train batch size"
+)
+@click.option("--learning-rate", "-lr", type=float, default=1e-5, help="Learning rate")
+@click.option(
+    "--lr-scheduler-type",
+    type=click.Choice(["linear", "cosine"]),
+    default="linear",
+    help="Learning rate scheduler type",
 )
-@click.option("--batch-size", type=INT_WITH_MAX, default="max", help="Train batch size")
-@click.option("--learning-rate", type=float, default=1e-5, help="Learning rate")
 @click.option(
     "--min-lr-ratio",
     type=float,
     default=0.0,
     help="The ratio of the final learning rate to the peak learning rate",
 )
+@click.option(
+    "--scheduler-num-cycles",
+    type=float,
+    default=0.5,
+    help="Number or fraction of cycles for the cosine learning rate scheduler.",
+)
 @click.option(
     "--warmup-ratio",
     type=float,
     default=0.0,
-    help="Warmup ratio for learning rate scheduler.",
+    help="Warmup ratio for the learning rate scheduler.",
 )
 @click.option(
     "--max-grad-norm",
@@ -123,7 +143,11 @@ def fine_tuning(ctx: click.Context) -> None:
     help="Beta parameter for DPO training (only used when '--training-method' is 'dpo')",
 )
 @click.option(
-    "--suffix", type=str, default=None, help="Suffix for the fine-tuned model name"
+    "--suffix",
+    "-s",
+    type=str,
+    default=None,
+    help="Suffix for the fine-tuned model name",
 )
 @click.option("--wandb-api-key", type=str, default=None, help="Wandb API key")
 @click.option("--wandb-base-url", type=str, default=None, help="Wandb base URL")
@@ -162,7 +186,9 @@ def create(
     n_checkpoints: int,
     batch_size: int | Literal["max"],
     learning_rate: float,
+    lr_scheduler_type: Literal["linear", "cosine"],
     min_lr_ratio: float,
+    scheduler_num_cycles: float,
     warmup_ratio: float,
     max_grad_norm: float,
     weight_decay: float,
@@ -194,7 +220,9 @@ def create(
         n_checkpoints=n_checkpoints,
         batch_size=batch_size,
         learning_rate=learning_rate,
+        lr_scheduler_type=lr_scheduler_type,
         min_lr_ratio=min_lr_ratio,
+        scheduler_num_cycles=scheduler_num_cycles,
         warmup_ratio=warmup_ratio,
         max_grad_norm=max_grad_norm,
         weight_decay=weight_decay,
@@ -214,8 +242,15 @@ def create(
         from_checkpoint=from_checkpoint,
     )
 
+    if model is None and from_checkpoint is None:
+        raise click.BadParameter("You must specify either a model or a checkpoint")
+
+    model_name = model
+    if from_checkpoint is not None:
+        model_name = from_checkpoint.split(":")[0]
+
     model_limits: FinetuneTrainingLimits = client.fine_tuning.get_model_limits(
-        model=model
+        model=model_name
     )
 
     if lora:
@@ -411,6 +446,7 @@ def list_checkpoints(ctx: click.Context, fine_tune_id: str) -> None:
 @click.argument("fine_tune_id", type=str, required=True)
 @click.option(
     "--output_dir",
+    "-o",
     type=click.Path(exists=True, file_okay=False, resolve_path=True),
     required=False,
     default=None,
@@ -418,6 +454,7 @@ def list_checkpoints(ctx: click.Context, fine_tune_id: str) -> None:
 )
 @click.option(
     "--checkpoint-step",
+    "-s",
     type=int,
     required=False,
     default=None,

{together-1.4.6 → together-1.5.3}/src/together/resources/finetune.py

@@ -22,7 +22,10 @@ from together.types import (
     TogetherRequest,
     TrainingType,
     FinetuneLRScheduler,
+    FinetuneLinearLRScheduler,
+    FinetuneCosineLRScheduler,
     FinetuneLinearLRSchedulerArgs,
+    FinetuneCosineLRSchedulerArgs,
     TrainingMethodDPO,
     TrainingMethodSFT,
     FinetuneCheckpoint,
@@ -50,14 +53,16 @@ AVAILABLE_TRAINING_METHODS = {
 def createFinetuneRequest(
     model_limits: FinetuneTrainingLimits,
     training_file: str,
-    model: 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"] = "linear",
     min_lr_ratio: float = 0.0,
+    scheduler_num_cycles: float = 0.5,
     warmup_ratio: float = 0.0,
     max_grad_norm: float = 1.0,
     weight_decay: float = 0.0,
@@ -77,6 +82,14 @@ def createFinetuneRequest(
     from_checkpoint: 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 batch_size == "max":
         log_warn_once(
             "Starting from together>=1.3.0, "
@@ -86,6 +99,8 @@ def createFinetuneRequest(
         warmup_ratio = 0.0
 
     training_type: TrainingType = FullTrainingType()
+    max_batch_size: int = 0
+    min_batch_size: int = 0
     if lora:
         if model_limits.lora_training is None:
             raise ValueError("LoRA adapters are not supported for the selected model.")
@@ -98,18 +113,26 @@ def createFinetuneRequest(
             lora_trainable_modules=lora_trainable_modules,
         )
 
-        batch_size = (
-            batch_size
-            if batch_size != "max"
-            else model_limits.lora_training.max_batch_size
-        )
+        max_batch_size = model_limits.lora_training.max_batch_size
+        min_batch_size = model_limits.lora_training.min_batch_size
+
     else:
         if model_limits.full_training is None:
             raise ValueError("Full training is not supported for the selected model.")
-        batch_size = (
-            batch_size
-            if batch_size != "max"
-            else model_limits.full_training.max_batch_size
+
+        max_batch_size = model_limits.full_training.max_batch_size
+        min_batch_size = model_limits.full_training.min_batch_size
+
+    batch_size = batch_size if batch_size != "max" else max_batch_size
+
+    if batch_size > max_batch_size:
+        raise ValueError(
+            "Requested batch size is higher that the maximum allowed value."
+        )
+
+    if batch_size < min_batch_size:
+        raise ValueError(
+            "Requested batch size is lower that the minimum allowed value."
         )
 
     if warmup_ratio > 1 or warmup_ratio < 0:
@@ -129,10 +152,22 @@ def createFinetuneRequest(
             f"training_method must be one of {', '.join(AVAILABLE_TRAINING_METHODS)}"
         )
 
-    lrScheduler = FinetuneLRScheduler(
-        lr_scheduler_type="linear",
-        lr_scheduler_args=FinetuneLinearLRSchedulerArgs(min_lr_ratio=min_lr_ratio),
-    )
+    # Default to generic lr scheduler
+    lrScheduler: FinetuneLRScheduler = FinetuneLRScheduler(lr_scheduler_type="linear")
+
+    if lr_scheduler_type == "cosine":
+        if scheduler_num_cycles <= 0.0:
+            raise ValueError("Number of cycles should be greater than 0")
+
+        lrScheduler = FinetuneCosineLRScheduler(
+            lr_scheduler_args=FinetuneCosineLRSchedulerArgs(
+                min_lr_ratio=min_lr_ratio, num_cycles=scheduler_num_cycles
+            ),
+        )
+    else:
+        lrScheduler = FinetuneLinearLRScheduler(
+            lr_scheduler_args=FinetuneLinearLRSchedulerArgs(min_lr_ratio=min_lr_ratio),
+        )
 
     training_method_cls: TrainingMethodSFT | TrainingMethodDPO = TrainingMethodSFT()
     if training_method == "dpo":
@@ -237,14 +272,16 @@ class FineTuning:
         self,
         *,
         training_file: str,
-        model: 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"] = "linear",
         min_lr_ratio: float = 0.0,
+        scheduler_num_cycles: float = 0.5,
         warmup_ratio: float = 0.0,
         max_grad_norm: float = 1.0,
         weight_decay: float = 0.0,
@@ -270,7 +307,7 @@ class FineTuning:
 
         Args:
             training_file (str): File-ID of a file uploaded to the Together API
-            model (str): Name of the base model to run fine-tune job on
+            model (str, optional): Name of the base model to run fine-tune job on
             n_epochs (int, optional): Number of epochs for fine-tuning. Defaults to 1.
             validation file (str, optional): File ID of a file uploaded to the Together API for validation.
             n_evals (int, optional): Number of evaluation loops to run. Defaults to 0.
@@ -279,9 +316,11 @@ class FineTuning:
             batch_size (int or "max"): Batch size for fine-tuning. Defaults to max.
             learning_rate (float, optional): Learning rate multiplier to use for training
                 Defaults to 0.00001.
+            lr_scheduler_type (Literal["linear", "cosine"]): Learning rate scheduler type. Defaults to "linear".
             min_lr_ratio (float, optional): Min learning rate ratio of the initial learning rate for
                 the learning rate scheduler. Defaults to 0.0.
-            warmup_ratio (float, optional): Warmup ratio for learning rate scheduler.
+            scheduler_num_cycles (float, optional): Number or fraction of cycles for the cosine learning rate scheduler. Defaults to 0.5.
+            warmup_ratio (float, optional): Warmup ratio for the learning rate scheduler.
             max_grad_norm (float, optional): Max gradient norm. Defaults to 1.0, set to 0 to disable.
             weight_decay (float, optional): Weight decay. Defaults to 0.0.
             lora (bool, optional): Whether to use LoRA adapters. Defaults to True.
@@ -325,7 +364,16 @@ class FineTuning:
         )
 
         if model_limits is None:
-            model_limits = self.get_model_limits(model=model)
+            # mypy doesn't understand that model or from_checkpoint is not None
+            if model is not None:
+                model_name = model
+            elif from_checkpoint is not None:
+                model_name = from_checkpoint.split(":")[0]
+            else:
+                # this branch is unreachable, but mypy doesn't know that
+                pass
+            model_limits = self.get_model_limits(model=model_name)
+
         finetune_request = createFinetuneRequest(
             model_limits=model_limits,
             training_file=training_file,
@@ -336,7 +384,9 @@ class FineTuning:
             n_checkpoints=n_checkpoints,
             batch_size=batch_size,
             learning_rate=learning_rate,
+            lr_scheduler_type=lr_scheduler_type,
             min_lr_ratio=min_lr_ratio,
+            scheduler_num_cycles=scheduler_num_cycles,
             warmup_ratio=warmup_ratio,
             max_grad_norm=max_grad_norm,
             weight_decay=weight_decay,
@@ -610,14 +660,16 @@ class AsyncFineTuning:
         self,
         *,
         training_file: str,
-        model: 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"] = "linear",
         min_lr_ratio: float = 0.0,
+        scheduler_num_cycles: float = 0.5,
         warmup_ratio: float = 0.0,
         max_grad_norm: float = 1.0,
         weight_decay: float = 0.0,
@@ -643,7 +695,7 @@ class AsyncFineTuning:
 
         Args:
             training_file (str): File-ID of a file uploaded to the Together API
-            model (str): Name of the base model to run fine-tune job on
+            model (str, optional): Name of the base model to run fine-tune job on
             n_epochs (int, optional): Number of epochs for fine-tuning. Defaults to 1.
             validation file (str, optional): File ID of a file uploaded to the Together API for validation.
             n_evals (int, optional): Number of evaluation loops to run. Defaults to 0.
@@ -652,9 +704,11 @@ class AsyncFineTuning:
             batch_size (int, optional): Batch size for fine-tuning. Defaults to max.
             learning_rate (float, optional): Learning rate multiplier to use for training
                 Defaults to 0.00001.
+            lr_scheduler_type (Literal["linear", "cosine"]): Learning rate scheduler type. Defaults to "linear".
             min_lr_ratio (float, optional): Min learning rate ratio of the initial learning rate for
                 the learning rate scheduler. Defaults to 0.0.
-            warmup_ratio (float, optional): Warmup ratio for learning rate scheduler.
+            scheduler_num_cycles (float, optional): Number or fraction of cycles for the cosine learning rate scheduler. Defaults to 0.5.
+            warmup_ratio (float, optional): Warmup ratio for the learning rate scheduler.
             max_grad_norm (float, optional): Max gradient norm. Defaults to 1.0, set to 0 to disable.
             weight_decay (float, optional): Weight decay. Defaults to 0.0.
             lora (bool, optional): Whether to use LoRA adapters. Defaults to True.
@@ -698,7 +752,15 @@ class AsyncFineTuning:
         )
 
         if model_limits is None:
-            model_limits = await self.get_model_limits(model=model)
+            # mypy doesn't understand that model or from_checkpoint is not None
+            if model is not None:
+                model_name = model
+            elif from_checkpoint is not None:
+                model_name = from_checkpoint.split(":")[0]
+            else:
+                # this branch is unreachable, but mypy doesn't know that
+                pass
+            model_limits = await self.get_model_limits(model=model_name)
 
         finetune_request = createFinetuneRequest(
             model_limits=model_limits,
@@ -710,7 +772,9 @@ class AsyncFineTuning:
             n_checkpoints=n_checkpoints,
             batch_size=batch_size,
             learning_rate=learning_rate,
+            lr_scheduler_type=lr_scheduler_type,
             min_lr_ratio=min_lr_ratio,
+            scheduler_num_cycles=scheduler_num_cycles,
             warmup_ratio=warmup_ratio,
             max_grad_norm=max_grad_norm,
             weight_decay=weight_decay,

{together-1.4.6 → together-1.5.3}/src/together/types/__init__.py

@@ -34,11 +34,14 @@ from together.types.finetune import (
     TrainingMethodDPO,
     TrainingMethodSFT,
     FinetuneCheckpoint,
+    FinetuneCosineLRScheduler,
+    FinetuneCosineLRSchedulerArgs,
     FinetuneDownloadResult,
+    FinetuneLinearLRScheduler,
     FinetuneLinearLRSchedulerArgs,
+    FinetuneLRScheduler,
     FinetuneList,
     FinetuneListEvents,
-    FinetuneLRScheduler,
     FinetuneRequest,
     FinetuneResponse,
     FinetuneTrainingLimits,
@@ -69,7 +72,10 @@ __all__ = [
     "FinetuneListEvents",
     "FinetuneDownloadResult",
     "FinetuneLRScheduler",
+    "FinetuneLinearLRScheduler",
     "FinetuneLinearLRSchedulerArgs",
+    "FinetuneCosineLRScheduler",
+    "FinetuneCosineLRSchedulerArgs",
     "FileRequest",
     "FileResponse",
     "FileList",

{together-1.4.6 → together-1.5.3}/src/together/types/finetune.py

@@ -1,9 +1,9 @@
 from __future__ import annotations
 
 from enum import Enum
-from typing import List, Literal
+from typing import List, Literal, Union
 
-from pydantic import StrictBool, Field, validator, field_validator
+from pydantic import StrictBool, Field, validator, field_validator, ValidationInfo
 
 from together.types.abstract import BaseModel
 from together.types.common import (
@@ -170,13 +170,13 @@ class FinetuneRequest(BaseModel):
     # validation file id
     validation_file: str | None = None
     # base model string
-    model: str
+    model: str | None = None
     # number of epochs to train for
     n_epochs: int
     # training learning rate
     learning_rate: float
     # learning rate scheduler type and args
-    lr_scheduler: FinetuneLRScheduler | None = None
+    lr_scheduler: FinetuneLinearLRScheduler | FinetuneCosineLRScheduler | None = None
     # learning rate warmup ratio
     warmup_ratio: float
     # max gradient norm
@@ -239,7 +239,7 @@ class FinetuneResponse(BaseModel):
     # training learning rate
     learning_rate: float | None = None
     # learning rate scheduler type and args
-    lr_scheduler: FinetuneLRScheduler | None = None
+    lr_scheduler: FinetuneLinearLRScheduler | FinetuneCosineLRScheduler | None = None
     # learning rate warmup ratio
     warmup_ratio: float | None = None
     # max gradient norm
@@ -345,13 +345,27 @@ class FinetuneTrainingLimits(BaseModel):
     lora_training: FinetuneLoraTrainingLimits | None = None
 
 
+class FinetuneLinearLRSchedulerArgs(BaseModel):
+    min_lr_ratio: float | None = 0.0
+
+
+class FinetuneCosineLRSchedulerArgs(BaseModel):
+    min_lr_ratio: float | None = 0.0
+    num_cycles: float | None = 0.5
+
+
 class FinetuneLRScheduler(BaseModel):
     lr_scheduler_type: str
-    lr_scheduler_args: FinetuneLinearLRSchedulerArgs | None = None
 
 
-class FinetuneLinearLRSchedulerArgs(BaseModel):
-    min_lr_ratio: float | None = 0.0
+class FinetuneLinearLRScheduler(FinetuneLRScheduler):
+    lr_scheduler_type: Literal["linear"] = "linear"
+    lr_scheduler: FinetuneLinearLRSchedulerArgs | None = None
+
+
+class FinetuneCosineLRScheduler(FinetuneLRScheduler):
+    lr_scheduler_type: Literal["cosine"] = "cosine"
+    lr_scheduler: FinetuneCosineLRSchedulerArgs | None = None
 
 
 class FinetuneCheckpoint(BaseModel):