together 1.5.8__tar.gz → 1.5.11__tar.gz

{together-1.5.8 → together-1.5.11}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: together
-Version: 1.5.8
+Version: 1.5.11
 Summary: Python client for Together's Cloud Platform!
 License: Apache-2.0
 Author: Together AI
@@ -13,6 +13,7 @@ Classifier: Programming Language :: Python :: 3.10
 Classifier: Programming Language :: Python :: 3.11
 Classifier: Programming Language :: Python :: 3.12
 Classifier: Programming Language :: Python :: 3.13
+Provides-Extra: pyarrow
 Requires-Dist: aiohttp (>=3.9.3,<4.0.0)
 Requires-Dist: click (>=8.1.7,<9.0.0)
 Requires-Dist: eval-type-backport (>=0.1.3,<0.3.0)
@@ -20,7 +21,7 @@ Requires-Dist: filelock (>=3.13.1,<4.0.0)
 Requires-Dist: numpy (>=1.23.5) ; python_version < "3.12"
 Requires-Dist: numpy (>=1.26.0) ; python_version >= "3.12"
 Requires-Dist: pillow (>=11.1.0,<12.0.0)
-Requires-Dist: pyarrow (>=10.0.1)
+Requires-Dist: pyarrow (>=10.0.1) ; extra == "pyarrow"
 Requires-Dist: pydantic (>=2.6.3,<3.0.0)
 Requires-Dist: requests (>=2.31.0,<3.0.0)
 Requires-Dist: rich (>=13.8.1,<15.0.0)
@@ -92,7 +93,7 @@ client = Together()
 
 # Simple text message
 response = client.chat.completions.create(
-    model="mistralai/Mixtral-8x7B-Instruct-v0.1",
+    model="meta-llama/Llama-4-Scout-17B-16E-Instruct",
     messages=[{"role": "user", "content": "tell me about new york"}],
 )
 print(response.choices[0].message.content)
@@ -182,7 +183,7 @@ from together import Together
 
 client = Together()
 stream = client.chat.completions.create(
-    model="mistralai/Mixtral-8x7B-Instruct-v0.1",
+    model="meta-llama/Llama-4-Scout-17B-16E-Instruct",
     messages=[{"role": "user", "content": "tell me about new york"}],
     stream=True,
 )
@@ -207,7 +208,7 @@ async def async_chat_completion(messages):
     async_client = AsyncTogether()
     tasks = [
         async_client.chat.completions.create(
-            model="mistralai/Mixtral-8x7B-Instruct-v0.1",
+            model="meta-llama/Llama-4-Scout-17B-16E-Instruct",
             messages=[{"role": "user", "content": message}],
         )
         for message in messages
@@ -230,7 +231,7 @@ from together import Together
 client = Together()
 
 response = client.chat.completions.create(
-    model="mistralai/Mixtral-8x7B-Instruct-v0.1",
+    model="meta-llama/Llama-3.2-3B-Instruct-Turbo",
     messages=[{"role": "user", "content": "tell me about new york"}],
     logprobs=1
 )
@@ -381,7 +382,7 @@ client.files.delete(id="file-d0d318cb-b7d9-493a-bd70-1cfe089d3815") # deletes a
 
 ### Fine-tunes
 
-The finetune API is used for fine-tuning and allows developers to create finetuning jobs. It also has several methods to list all jobs, retrive statuses and get checkpoints. Please refer to our fine-tuning docs [here](https://docs.together.ai/docs/fine-tuning-python).
+The finetune API is used for fine-tuning and allows developers to create finetuning jobs. It also has several methods to list all jobs, retrive statuses and get checkpoints. Please refer to our fine-tuning docs [here](https://docs.together.ai/docs/fine-tuning-quickstart).
 
 ```python
 from together import Together
@@ -390,7 +391,7 @@ client = Together()
 
 client.fine_tuning.create(
   training_file = 'file-d0d318cb-b7d9-493a-bd70-1cfe089d3815',
-  model = 'mistralai/Mixtral-8x7B-Instruct-v0.1',
+  model = 'meta-llama/Llama-3.2-3B-Instruct',
   n_epochs = 3,
   n_checkpoints = 1,
   batch_size = "max",
@@ -428,7 +429,7 @@ for model in models:
 together chat.completions \
   --message "system" "You are a helpful assistant named Together" \
   --message "user" "What is your name?" \
-  --model mistralai/Mixtral-8x7B-Instruct-v0.1
+  --model meta-llama/Llama-4-Scout-17B-16E-Instruct
 ```
 
 The Chat Completions CLI enables streaming tokens to stdout by default. To disable streaming, use `--no-stream`.
@@ -438,7 +439,7 @@ The Chat Completions CLI enables streaming tokens to stdout by default. To disab
 ```bash
 together completions \
   "Large language models are " \
-  --model mistralai/Mixtral-8x7B-v0.1 \
+  --model meta-llama/Llama-4-Scout-17B-16E-Instruct \
   --max-tokens 512 \
   --stop "."
 ```

{together-1.5.8 → together-1.5.11}/README.md

@@ -58,7 +58,7 @@ client = Together()
 
 # Simple text message
 response = client.chat.completions.create(
-    model="mistralai/Mixtral-8x7B-Instruct-v0.1",
+    model="meta-llama/Llama-4-Scout-17B-16E-Instruct",
     messages=[{"role": "user", "content": "tell me about new york"}],
 )
 print(response.choices[0].message.content)
@@ -148,7 +148,7 @@ from together import Together
 
 client = Together()
 stream = client.chat.completions.create(
-    model="mistralai/Mixtral-8x7B-Instruct-v0.1",
+    model="meta-llama/Llama-4-Scout-17B-16E-Instruct",
     messages=[{"role": "user", "content": "tell me about new york"}],
     stream=True,
 )
@@ -173,7 +173,7 @@ async def async_chat_completion(messages):
     async_client = AsyncTogether()
     tasks = [
         async_client.chat.completions.create(
-            model="mistralai/Mixtral-8x7B-Instruct-v0.1",
+            model="meta-llama/Llama-4-Scout-17B-16E-Instruct",
             messages=[{"role": "user", "content": message}],
         )
         for message in messages
@@ -196,7 +196,7 @@ from together import Together
 client = Together()
 
 response = client.chat.completions.create(
-    model="mistralai/Mixtral-8x7B-Instruct-v0.1",
+    model="meta-llama/Llama-3.2-3B-Instruct-Turbo",
     messages=[{"role": "user", "content": "tell me about new york"}],
     logprobs=1
 )
@@ -347,7 +347,7 @@ client.files.delete(id="file-d0d318cb-b7d9-493a-bd70-1cfe089d3815") # deletes a
 
 ### Fine-tunes
 
-The finetune API is used for fine-tuning and allows developers to create finetuning jobs. It also has several methods to list all jobs, retrive statuses and get checkpoints. Please refer to our fine-tuning docs [here](https://docs.together.ai/docs/fine-tuning-python).
+The finetune API is used for fine-tuning and allows developers to create finetuning jobs. It also has several methods to list all jobs, retrive statuses and get checkpoints. Please refer to our fine-tuning docs [here](https://docs.together.ai/docs/fine-tuning-quickstart).
 
 ```python
 from together import Together
@@ -356,7 +356,7 @@ client = Together()
 
 client.fine_tuning.create(
   training_file = 'file-d0d318cb-b7d9-493a-bd70-1cfe089d3815',
-  model = 'mistralai/Mixtral-8x7B-Instruct-v0.1',
+  model = 'meta-llama/Llama-3.2-3B-Instruct',
   n_epochs = 3,
   n_checkpoints = 1,
   batch_size = "max",
@@ -394,7 +394,7 @@ for model in models:
 together chat.completions \
   --message "system" "You are a helpful assistant named Together" \
   --message "user" "What is your name?" \
-  --model mistralai/Mixtral-8x7B-Instruct-v0.1
+  --model meta-llama/Llama-4-Scout-17B-16E-Instruct
 ```
 
 The Chat Completions CLI enables streaming tokens to stdout by default. To disable streaming, use `--no-stream`.
@@ -404,7 +404,7 @@ The Chat Completions CLI enables streaming tokens to stdout by default. To disab
 ```bash
 together completions \
   "Large language models are " \
-  --model mistralai/Mixtral-8x7B-v0.1 \
+  --model meta-llama/Llama-4-Scout-17B-16E-Instruct \
   --max-tokens 512 \
   --stop "."
 ```

{together-1.5.8 → together-1.5.11}/pyproject.toml

@@ -12,7 +12,7 @@ build-backend = "poetry.masonry.api"
 
 [tool.poetry]
 name = "together"
-version = "1.5.8"
+version = "1.5.11"
 authors = ["Together AI <support@together.ai>"]
 description = "Python client for Together's Cloud Platform!"
 readme = "README.md"
@@ -37,13 +37,16 @@ aiohttp = "^3.9.3"
 filelock = "^3.13.1"
 eval-type-backport = ">=0.1.3,<0.3.0"
 click = "^8.1.7"
-pyarrow = ">=10.0.1"
+pyarrow = { version = ">=10.0.1", optional = true }
 numpy = [
     { version = ">=1.23.5", python = "<3.12" },
     { version = ">=1.26.0", python = ">=3.12" },
 ]
 pillow = "^11.1.0"
 
+[tool.poetry.extras]
+pyarrow = ["pyarrow"]
+
 [tool.poetry.group.quality]
 optional = true
 

{together-1.5.8 → together-1.5.11}/src/together/cli/api/endpoints.py

@@ -82,7 +82,7 @@ def endpoints(ctx: click.Context) -> None:
 @click.option(
     "--model",
     required=True,
-    help="The model to deploy (e.g. mistralai/Mixtral-8x7B-Instruct-v0.1)",
+    help="The model to deploy (e.g. meta-llama/Llama-4-Scout-17B-16E-Instruct)",
 )
 @click.option(
     "--min-replicas",

{together-1.5.8 → together-1.5.11}/src/together/cli/api/finetune.py

@@ -1,10 +1,10 @@
 from __future__ import annotations
 
 import json
+import re
 from datetime import datetime, timezone
 from textwrap import wrap
 from typing import Any, Literal
-import re
 
 import click
 from click.core import ParameterSource  # type: ignore[attr-defined]
@@ -13,17 +13,17 @@ from tabulate import tabulate
 
 from together import Together
 from together.cli.api.utils import BOOL_WITH_AUTO, INT_WITH_MAX
+from together.types.finetune import (
+    DownloadCheckpointType,
+    FinetuneEventType,
+    FinetuneTrainingLimits,
+)
 from together.utils import (
     finetune_price_to_dollars,
+    format_timestamp,
     log_warn,
     log_warn_once,
     parse_timestamp,
-    format_timestamp,
-)
-from together.types.finetune import (
-    DownloadCheckpointType,
-    FinetuneTrainingLimits,
-    FinetuneEventType,
 )
 
 
@@ -348,9 +348,9 @@ def list(ctx: click.Context) -> None:
                 "Model Output Name": "\n".join(wrap(i.output_name or "", width=30)),
                 "Status": i.status,
                 "Created At": i.created_at,
-                "Price": f"""${finetune_price_to_dollars(
-                    float(str(i.total_price))
-                )}""",  # convert to string for mypy typing
+                "Price": f"""${
+                    finetune_price_to_dollars(float(str(i.total_price)))
+                }""",  # convert to string for mypy typing
             }
         )
     table = tabulate(display_list, headers="keys", tablefmt="grid", showindex=True)

{together-1.5.8 → together-1.5.11}/src/together/resources/finetune.py

@@ -2,7 +2,7 @@ from __future__ import annotations
 
 import re
 from pathlib import Path
-from typing import Dict, List, Literal
+from typing import List, Dict, Literal
 
 from rich import print as rprint
 
@@ -69,7 +69,7 @@ def create_finetune_request(
     wandb_base_url: str | None = None,
     wandb_project_name: str | None = None,
     wandb_name: str | None = None,
-    train_on_inputs: bool | Literal["auto"] = "auto",
+    train_on_inputs: bool | Literal["auto"] | None = None,
     training_method: str = "sft",
     dpo_beta: float | None = None,
     from_checkpoint: str | None = None,
@@ -101,6 +101,11 @@ def create_finetune_request(
             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(
@@ -166,6 +171,18 @@ def create_finetune_request(
             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 train_on_inputs is None and training_method == "sft":
+        log_warn_once(
+            "train_on_inputs is not set for SFT training, it will be set to 'auto'"
+        )
+        train_on_inputs = "auto"
+
+    if dpo_beta is not None and training_method != "dpo":
+        raise ValueError("dpo_beta is only supported for DPO training")
+
     lr_scheduler: FinetuneLRScheduler
     if lr_scheduler_type == "cosine":
         if scheduler_num_cycles <= 0.0:
@@ -183,8 +200,10 @@ def create_finetune_request(
             lr_scheduler_args=LinearLRSchedulerArgs(min_lr_ratio=min_lr_ratio),
         )
 
-    training_method_cls: TrainingMethodSFT | TrainingMethodDPO = TrainingMethodSFT()
-    if training_method == "dpo":
+    training_method_cls: TrainingMethodSFT | TrainingMethodDPO
+    if training_method == "sft":
+        training_method_cls = TrainingMethodSFT(train_on_inputs=train_on_inputs)
+    elif training_method == "dpo":
         training_method_cls = TrainingMethodDPO(dpo_beta=dpo_beta)
 
     finetune_request = FinetuneRequest(
@@ -206,7 +225,6 @@ def create_finetune_request(
         wandb_base_url=wandb_base_url,
         wandb_project_name=wandb_project_name,
         wandb_name=wandb_name,
-        train_on_inputs=train_on_inputs,
         training_method=training_method_cls,
         from_checkpoint=from_checkpoint,
     )
@@ -281,7 +299,7 @@ class FineTuning:
         wandb_name: str | None = None,
         verbose: bool = False,
         model_limits: FinetuneTrainingLimits | None = None,
-        train_on_inputs: bool | Literal["auto"] = "auto",
+        train_on_inputs: bool | Literal["auto"] | None = None,
         training_method: str = "sft",
         dpo_beta: float | None = None,
         from_checkpoint: str | None = None,
@@ -326,12 +344,12 @@ class FineTuning:
                 Defaults to False.
             model_limits (FinetuneTrainingLimits, optional): Limits for the hyperparameters the model in Fine-tuning.
                 Defaults to None.
-            train_on_inputs (bool or "auto"): Whether to mask the user messages in conversational data or prompts in instruction data.
+            train_on_inputs (bool or "auto", optional): Whether to mask the user messages in conversational data or prompts in instruction data.
                 "auto" will automatically determine whether to mask the inputs based on the data format.
                 For datasets with the "text" field (general format), inputs will not be masked.
                 For datasets with the "messages" field (conversational format) or "prompt" and "completion" fields
                 (Instruction format), inputs will be masked.
-                Defaults to "auto".
+                Defaults to None, or "auto" if training_method is "sft" (set in create_finetune_request).
             training_method (str, optional): Training method. Defaults to "sft".
                 Supported methods: "sft", "dpo".
             dpo_beta (float, optional): DPO beta parameter. Defaults to None.
@@ -545,7 +563,7 @@ class FineTuning:
         *,
         output: Path | str | None = None,
         checkpoint_step: int | None = None,
-        checkpoint_type: DownloadCheckpointType = DownloadCheckpointType.DEFAULT,
+        checkpoint_type: DownloadCheckpointType | str = DownloadCheckpointType.DEFAULT,
     ) -> FinetuneDownloadResult:
         """
         Downloads compressed fine-tuned model or checkpoint to local disk.
@@ -558,7 +576,7 @@ class FineTuning:
                 Defaults to None.
             checkpoint_step (int, optional): Specifies step number for checkpoint to download.
                 Defaults to -1 (download the final model)
-            checkpoint_type (CheckpointType, optional): Specifies which checkpoint to download.
+            checkpoint_type (CheckpointType | str, optional): Specifies which checkpoint to download.
                 Defaults to CheckpointType.DEFAULT.
 
         Returns:
@@ -582,6 +600,16 @@ class FineTuning:
 
         ft_job = self.retrieve(id)
 
+        # convert str to DownloadCheckpointType
+        if isinstance(checkpoint_type, str):
+            try:
+                checkpoint_type = DownloadCheckpointType(checkpoint_type.lower())
+            except ValueError:
+                enum_strs = ", ".join(e.value for e in DownloadCheckpointType)
+                raise ValueError(
+                    f"Invalid checkpoint type: {checkpoint_type}. Choose one of {{{enum_strs}}}."
+                )
+
         if isinstance(ft_job.training_type, FullTrainingType):
             if checkpoint_type != DownloadCheckpointType.DEFAULT:
                 raise ValueError(
@@ -592,10 +620,11 @@ class FineTuning:
             if checkpoint_type == DownloadCheckpointType.DEFAULT:
                 checkpoint_type = DownloadCheckpointType.MERGED
 
-            if checkpoint_type == DownloadCheckpointType.MERGED:
-                url += f"&checkpoint={DownloadCheckpointType.MERGED.value}"
-            elif checkpoint_type == DownloadCheckpointType.ADAPTER:
-                url += f"&checkpoint={DownloadCheckpointType.ADAPTER.value}"
+            if checkpoint_type in {
+                DownloadCheckpointType.MERGED,
+                DownloadCheckpointType.ADAPTER,
+            }:
+                url += f"&checkpoint={checkpoint_type.value}"
             else:
                 raise ValueError(
                     f"Invalid checkpoint type for LoRATrainingType: {checkpoint_type}"
@@ -682,7 +711,7 @@ class AsyncFineTuning:
         wandb_name: str | None = None,
         verbose: bool = False,
         model_limits: FinetuneTrainingLimits | None = None,
-        train_on_inputs: bool | Literal["auto"] = "auto",
+        train_on_inputs: bool | Literal["auto"] | None = None,
         training_method: str = "sft",
         dpo_beta: float | None = None,
         from_checkpoint: str | None = None,
@@ -732,7 +761,7 @@ class AsyncFineTuning:
                 For datasets with the "text" field (general format), inputs will not be masked.
                 For datasets with the "messages" field (conversational format) or "prompt" and "completion" fields
                 (Instruction format), inputs will be masked.
-                Defaults to "auto".
+                Defaults to None, or "auto" if training_method is "sft" (set in create_finetune_request).
             training_method (str, optional): Training method. Defaults to "sft".
                 Supported methods: "sft", "dpo".
             dpo_beta (float, optional): DPO beta parameter. Defaults to None.

{together-1.5.8 → together-1.5.11}/src/together/types/finetune.py

@@ -3,7 +3,7 @@ from __future__ import annotations
 from enum import Enum
 from typing import List, Literal, Any
 
-from pydantic import StrictBool, Field, field_validator
+from pydantic import Field, StrictBool, field_validator
 
 from together.types.abstract import BaseModel
 from together.types.common import (
@@ -149,6 +149,7 @@ class TrainingMethodSFT(TrainingMethod):
     """
 
     method: Literal["sft"] = "sft"
+    train_on_inputs: StrictBool | Literal["auto"] = "auto"
 
 
 class TrainingMethodDPO(TrainingMethod):
@@ -201,8 +202,6 @@ class FinetuneRequest(BaseModel):
     wandb_name: str | None = None
     # training type
     training_type: FullTrainingType | LoRATrainingType | None = None
-    # train on inputs
-    train_on_inputs: StrictBool | Literal["auto"] = "auto"
     # training method
     training_method: TrainingMethodSFT | TrainingMethodDPO = Field(
         default_factory=TrainingMethodSFT

{together-1.5.8 → together-1.5.11}/src/together/utils/files.py

@@ -6,7 +6,6 @@ from pathlib import Path
 from traceback import format_exc
 from typing import Any, Dict, List
 
-from pyarrow import ArrowInvalid, parquet
 
 from together.constants import (
     MAX_FILE_SIZE_GB,
@@ -372,6 +371,14 @@ def _check_jsonl(file: Path) -> Dict[str, Any]:
 
 
 def _check_parquet(file: Path) -> Dict[str, Any]:
+    try:
+        # Pyarrow is optional as it's large (~80MB) and isn't compatible with older systems.
+        from pyarrow import ArrowInvalid, parquet
+    except ImportError:
+        raise ImportError(
+            "pyarrow is not installed and is required to use parquet files. Please install it via `pip install together[pyarrow]`"
+        )
+
     report_dict: Dict[str, Any] = {}
 
     try: