together 1.2.11__tar.gz → 1.2.13__tar.gz

{together-1.2.11 → together-1.2.13}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: together
-Version: 1.2.11
+Version: 1.2.13
 Summary: Python client for Together's Cloud Platform!
 Home-page: https://github.com/togethercomputer/together-python
 License: Apache-2.0
@@ -25,6 +25,7 @@ Requires-Dist: pillow (>=10.3.0,<11.0.0)
 Requires-Dist: pyarrow (>=10.0.1)
 Requires-Dist: pydantic (>=2.6.3,<3.0.0)
 Requires-Dist: requests (>=2.31.0,<3.0.0)
+Requires-Dist: rich (>=13.8.1,<14.0.0)
 Requires-Dist: tabulate (>=0.9.0,<0.10.0)
 Requires-Dist: tqdm (>=4.66.2,<5.0.0)
 Requires-Dist: typer (>=0.9,<0.13)

{together-1.2.11 → together-1.2.13}/pyproject.toml

@@ -12,7 +12,7 @@ build-backend = "poetry.masonry.api"
 
 [tool.poetry]
 name = "together"
-version = "1.2.11"
+version = "1.2.13"
 authors = [
     "Together AI <support@together.ai>"
 ]
@@ -31,6 +31,7 @@ homepage = "https://github.com/togethercomputer/together-python"
 python = "^3.8"
 typer = ">=0.9,<0.13"
 requests = "^2.31.0"
+rich = "^13.8.1"
 tqdm = "^4.66.2"
 tabulate = "^0.9.0"
 pydantic = "^2.6.3"
@@ -70,7 +71,7 @@ tox = "^4.14.1"
 optional = true
 
 [tool.poetry.group.examples.dependencies]
-datasets = "^2.18.0"
+datasets = ">=2.18,<4.0"
 transformers = "^4.39.3"
 
 

{together-1.2.11 → together-1.2.13}/src/together/cli/api/finetune.py

@@ -1,15 +1,27 @@
 from __future__ import annotations
 
 import json
+from datetime import datetime
 from textwrap import wrap
 
 import click
 from click.core import ParameterSource  # type: ignore[attr-defined]
+from rich import print as rprint
 from tabulate import tabulate
 
 from together import Together
-from together.utils import finetune_price_to_dollars, log_warn, parse_timestamp
 from together.types.finetune import DownloadCheckpointType
+from together.utils import finetune_price_to_dollars, log_warn, parse_timestamp
+
+
+_CONFIRMATION_MESSAGE = (
+    "You are about to create a fine-tuning job. "
+    "The cost of your job will be determined by the model size, the number of tokens "
+    "in the training file, the number of tokens in the validation file, the number of epochs, and "
+    "the number of evaluations. Visit https://www.together.ai/pricing to get a price estimate.\n"
+    "You can pass `-y` or `--confirm` to your command to skip this message.\n\n"
+    "Do you want to proceed?"
+)
 
 
 class DownloadCheckpointTypeChoice(click.Choice):
@@ -65,6 +77,14 @@ def fine_tuning(ctx: click.Context) -> None:
     "--suffix", 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(
+    "--confirm",
+    "-y",
+    type=bool,
+    is_flag=True,
+    default=False,
+    help="Whether to skip the launch confirmation message",
+)
 def create(
     ctx: click.Context,
     training_file: str,
@@ -82,6 +102,7 @@ def create(
     lora_trainable_modules: str,
     suffix: str,
     wandb_api_key: str,
+    confirm: bool,
 ) -> None:
     """Start fine-tuning"""
     client: Together = ctx.obj
@@ -109,30 +130,37 @@ def create(
             "You have specified a number of evaluation loops but no validation file."
         )
 
-    response = client.fine_tuning.create(
-        training_file=training_file,
-        model=model,
-        n_epochs=n_epochs,
-        validation_file=validation_file,
-        n_evals=n_evals,
-        n_checkpoints=n_checkpoints,
-        batch_size=batch_size,
-        learning_rate=learning_rate,
-        lora=lora,
-        lora_r=lora_r,
-        lora_dropout=lora_dropout,
-        lora_alpha=lora_alpha,
-        lora_trainable_modules=lora_trainable_modules,
-        suffix=suffix,
-        wandb_api_key=wandb_api_key,
-    )
-
-    click.echo(json.dumps(response.model_dump(exclude_none=True), indent=4))
+    if confirm or click.confirm(_CONFIRMATION_MESSAGE, default=True, show_default=True):
+        response = client.fine_tuning.create(
+            training_file=training_file,
+            model=model,
+            n_epochs=n_epochs,
+            validation_file=validation_file,
+            n_evals=n_evals,
+            n_checkpoints=n_checkpoints,
+            batch_size=batch_size,
+            learning_rate=learning_rate,
+            lora=lora,
+            lora_r=lora_r,
+            lora_dropout=lora_dropout,
+            lora_alpha=lora_alpha,
+            lora_trainable_modules=lora_trainable_modules,
+            suffix=suffix,
+            wandb_api_key=wandb_api_key,
+            verbose=True,
+        )
 
-    # TODO: Remove it after the 21st of August
-    log_warn(
-        "The default value of batch size has been changed from 32 to 16 since together version >= 1.2.6"
-    )
+        report_string = f"Successfully submitted a fine-tuning job {response.id}"
+        if response.created_at is not None:
+            created_time = datetime.strptime(
+                response.created_at, "%Y-%m-%dT%H:%M:%S.%f%z"
+            )
+            # created_at reports UTC time, we use .astimezone() to convert to local time
+            formatted_time = created_time.astimezone().strftime("%m/%d/%Y, %H:%M:%S")
+            report_string += f" at {formatted_time}"
+        rprint(report_string)
+    else:
+        click.echo("No confirmation received, stopping job launch")
 
 
 @fine_tuning.command()

{together-1.2.11 → together-1.2.13}/src/together/constants.py

@@ -19,7 +19,7 @@ Please set it as an environment variable or set it as together.api_key
 Find your TOGETHER_API_KEY at https://api.together.xyz/settings/api-keys"""
 
 # Minimum number of samples required for fine-tuning file
-MIN_SAMPLES = 100
+MIN_SAMPLES = 1
 
 # the number of bytes in a gigabyte, used to convert bytes to GB for readable comparison
 NUM_BYTES_IN_GB = 2**30

{together-1.2.11 → together-1.2.13}/src/together/resources/chat/completions.py

@@ -20,7 +20,7 @@ class ChatCompletions:
     def create(
         self,
         *,
-        messages: List[Dict[str, str]],
+        messages: List[Dict[str, Any]],
         model: str,
         max_tokens: int | None = None,
         stop: List[str] | None = None,
@@ -32,13 +32,14 @@ class ChatCompletions:
         frequency_penalty: float | None = None,
         min_p: float | None = None,
         logit_bias: Dict[str, float] | None = None,
+        seed: int | None = None,
         stream: bool = False,
         logprobs: int | None = None,
         echo: bool | None = None,
         n: int | None = None,
         safety_model: str | None = None,
         response_format: Dict[str, str | Dict[str, Any]] | None = None,
-        tools: Dict[str, str | Dict[str, Any]] | None = None,
+        tools: List[Dict[str, Any]] | None = None,
         tool_choice: str | Dict[str, str | Dict[str, str]] | None = None,
         **kwargs: Any,
     ) -> ChatCompletionResponse | Iterator[ChatCompletionChunk]:
@@ -79,6 +80,7 @@ class ChatCompletions:
             logit_bias (Dict[str, float], optional): A dictionary of tokens and their bias values that modify the
                 likelihood of specific tokens being sampled. Bias values must be in the range [-100, 100].
                 Defaults to None.
+            seed (int, optional): A seed value to use for reproducibility.
             stream (bool, optional): Flag indicating whether to stream the generated completions.
                 Defaults to False.
             logprobs (int, optional): Number of top-k logprobs to return
@@ -124,6 +126,7 @@ class ChatCompletions:
             frequency_penalty=frequency_penalty,
             min_p=min_p,
             logit_bias=logit_bias,
+            seed=seed,
             stream=stream,
             logprobs=logprobs,
             echo=echo,
@@ -171,6 +174,7 @@ class AsyncChatCompletions:
         frequency_penalty: float | None = None,
         min_p: float | None = None,
         logit_bias: Dict[str, float] | None = None,
+        seed: int | None = None,
         stream: bool = False,
         logprobs: int | None = None,
         echo: bool | None = None,
@@ -218,6 +222,7 @@ class AsyncChatCompletions:
             logit_bias (Dict[str, float], optional): A dictionary of tokens and their bias values that modify the
                 likelihood of specific tokens being sampled. Bias values must be in the range [-100, 100].
                 Defaults to None.
+            seed (int, optional): A seed value to use for reproducibility.
             stream (bool, optional): Flag indicating whether to stream the generated completions.
                 Defaults to False.
             logprobs (int, optional): Number of top-k logprobs to return
@@ -263,6 +268,7 @@ class AsyncChatCompletions:
             frequency_penalty=frequency_penalty,
             min_p=min_p,
             logit_bias=logit_bias,
+            seed=seed,
             stream=stream,
             logprobs=logprobs,
             echo=echo,

{together-1.2.11 → together-1.2.13}/src/together/resources/completions.py

@@ -32,6 +32,7 @@ class Completions:
         frequency_penalty: float | None = None,
         min_p: float | None = None,
         logit_bias: Dict[str, float] | None = None,
+        seed: int | None = None,
         stream: bool = False,
         logprobs: int | None = None,
         echo: bool | None = None,
@@ -75,6 +76,7 @@ class Completions:
             logit_bias (Dict[str, float], optional): A dictionary of tokens and their bias values that modify the
                 likelihood of specific tokens being sampled. Bias values must be in the range [-100, 100].
                 Defaults to None.
+            seed (int, optional): Seed value for reproducibility.
             stream (bool, optional): Flag indicating whether to stream the generated completions.
                 Defaults to False.
             logprobs (int, optional): Number of top-k logprobs to return
@@ -107,6 +109,7 @@ class Completions:
             repetition_penalty=repetition_penalty,
             presence_penalty=presence_penalty,
             frequency_penalty=frequency_penalty,
+            seed=seed,
             min_p=min_p,
             logit_bias=logit_bias,
             stream=stream,
@@ -153,6 +156,7 @@ class AsyncCompletions:
         frequency_penalty: float | None = None,
         min_p: float | None = None,
         logit_bias: Dict[str, float] | None = None,
+        seed: int | None = None,
         stream: bool = False,
         logprobs: int | None = None,
         echo: bool | None = None,
@@ -196,6 +200,7 @@ class AsyncCompletions:
             logit_bias (Dict[str, float], optional): A dictionary of tokens and their bias values that modify the
                 likelihood of specific tokens being sampled. Bias values must be in the range [-100, 100].
                 Defaults to None.
+            seed (int, optional): Seed value for reproducibility.
             stream (bool, optional): Flag indicating whether to stream the generated completions.
                 Defaults to False.
             logprobs (int, optional): Number of top-k logprobs to return
@@ -230,6 +235,7 @@ class AsyncCompletions:
             frequency_penalty=frequency_penalty,
             min_p=min_p,
             logit_bias=logit_bias,
+            seed=seed,
             stream=stream,
             logprobs=logprobs,
             echo=echo,

{together-1.2.11 → together-1.2.13}/src/together/resources/finetune.py

@@ -2,6 +2,8 @@ from __future__ import annotations
 
 from pathlib import Path
 
+from rich import print as rprint
+
 from together.abstract import api_requestor
 from together.filemanager import DownloadManager
 from together.together_response import TogetherResponse
@@ -43,6 +45,7 @@ class FineTuning:
         lora_trainable_modules: str | None = "all-linear",
         suffix: str | None = None,
         wandb_api_key: str | None = None,
+        verbose: bool = False,
     ) -> FinetuneResponse:
         """
         Method to initiate a fine-tuning job
@@ -67,6 +70,8 @@ class FineTuning:
                 Defaults to None.
             wandb_api_key (str, optional): API key for Weights & Biases integration.
                 Defaults to None.
+            verbose (bool, optional): whether to print the job parameters before submitting a request.
+                Defaults to False.
 
         Returns:
             FinetuneResponse: Object containing information about fine-tuning job.
@@ -85,7 +90,7 @@ class FineTuning:
                 lora_trainable_modules=lora_trainable_modules,
             )
 
-        parameter_payload = FinetuneRequest(
+        finetune_request = FinetuneRequest(
             model=model,
             training_file=training_file,
             validation_file=validation_file,
@@ -97,7 +102,13 @@ class FineTuning:
             training_type=training_type,
             suffix=suffix,
             wandb_key=wandb_api_key,
-        ).model_dump(exclude_none=True)
+        )
+        if verbose:
+            rprint(
+                "Submitting a fine-tuning job with the following parameters:",
+                finetune_request,
+            )
+        parameter_payload = finetune_request.model_dump(exclude_none=True)
 
         response, _, _ = requestor.request(
             options=TogetherRequest(
@@ -110,11 +121,6 @@ class FineTuning:
 
         assert isinstance(response, TogetherResponse)
 
-        # TODO: Remove it after the 21st of August
-        log_warn(
-            "The default value of batch size has been changed from 32 to 16 since together version >= 1.2.6"
-        )
-
         # TODO: Remove after next LoRA default change
         log_warn(
             "Some of the jobs run _directly_ from the together-python library might be trained using LoRA adapters. "
@@ -266,7 +272,7 @@ class FineTuning:
                 raise ValueError(
                     "Only DEFAULT checkpoint type is allowed for FullTrainingType"
                 )
-            url += f"&checkpoint=modelOutputPath"
+            url += "&checkpoint=modelOutputPath"
         elif isinstance(ft_job.training_type, LoRATrainingType):
             if checkpoint_type == DownloadCheckpointType.DEFAULT:
                 checkpoint_type = DownloadCheckpointType.MERGED

{together-1.2.11 → together-1.2.13}/src/together/types/chat_completions.py

@@ -40,9 +40,24 @@ class ToolCalls(BaseModel):
     function: FunctionCall | None = None
 
 
+class ChatCompletionMessageContentType(str, Enum):
+    TEXT = "text"
+    IMAGE_URL = "image_url"
+
+
+class ChatCompletionMessageContentImageURL(BaseModel):
+    url: str
+
+
+class ChatCompletionMessageContent(BaseModel):
+    type: ChatCompletionMessageContentType
+    text: str | None = None
+    image_url: ChatCompletionMessageContentImageURL | None = None
+
+
 class ChatCompletionMessage(BaseModel):
     role: MessageRole
-    content: str | None = None
+    content: str | List[ChatCompletionMessageContent] | None = None
     tool_calls: List[ToolCalls] | None = None
 
 
@@ -96,6 +111,7 @@ class ChatCompletionRequest(BaseModel):
     frequency_penalty: float | None = None
     min_p: float | None = None
     logit_bias: Dict[str, float] | None = None
+    seed: int | None = None
     # stream SSE token chunks
     stream: bool = False
     # return logprobs
@@ -126,6 +142,7 @@ class ChatCompletionRequest(BaseModel):
 class ChatCompletionChoicesData(BaseModel):
     index: int | None = None
     logprobs: LogprobsPart | None = None
+    seed: int | None = None
     finish_reason: FinishReason | None = None
     message: ChatCompletionMessage | None = None
 
@@ -150,6 +167,7 @@ class ChatCompletionResponse(BaseModel):
 class ChatCompletionChoicesChunk(BaseModel):
     index: int | None = None
     logprobs: float | None = None
+    seed: int | None = None
     finish_reason: FinishReason | None = None
     delta: DeltaContent | None = None
 

{together-1.2.11 → together-1.2.13}/src/together/types/completions.py

@@ -35,6 +35,7 @@ class CompletionRequest(BaseModel):
     frequency_penalty: float | None = None
     min_p: float | None = None
     logit_bias: Dict[str, float] | None = None
+    seed: int | None = None
     # stream SSE token chunks
     stream: bool = False
     # return logprobs
@@ -61,13 +62,15 @@ class CompletionRequest(BaseModel):
 class CompletionChoicesData(BaseModel):
     index: int
     logprobs: LogprobsPart | None = None
-    finish_reason: FinishReason | None = None
+    seed: int | None = None
+    finish_reason: FinishReason
     text: str
 
 
 class CompletionChoicesChunk(BaseModel):
     index: int | None = None
     logprobs: float | None = None
+    seed: int | None = None
     finish_reason: FinishReason | None = None
     delta: DeltaContent | None = None