together 1.2.11__py3-none-any.whl → 2.0.0a8__py3-none-any.whl

together/cli/api/chat.py

@@ -1,276 +0,0 @@
-from __future__ import annotations
-
-import cmd
-import json
-from typing import List, Tuple
-
-import click
-
-from together import Together
-from together.types.chat_completions import (
-    ChatCompletionChoicesChunk,
-    ChatCompletionChunk,
-    ChatCompletionResponse,
-)
-
-
-class ChatShell(cmd.Cmd):
-    intro = "Type /exit to exit, /help, or /? to list commands.\n"
-    prompt = ">>> "
-
-    def __init__(
-        self,
-        client: Together,
-        model: str,
-        max_tokens: int | None = None,
-        stop: List[str] | None = None,
-        temperature: float | None = None,
-        top_p: float | None = None,
-        top_k: int | None = None,
-        repetition_penalty: float | None = None,
-        presence_penalty: float | None = None,
-        frequency_penalty: float | None = None,
-        min_p: float | None = None,
-        safety_model: str | None = None,
-        system_message: str | None = None,
-    ) -> None:
-        super().__init__()
-        self.client = client
-        self.model = model
-        self.max_tokens = max_tokens
-        self.stop = stop
-        self.temperature = temperature
-        self.top_p = top_p
-        self.top_k = top_k
-        self.repetition_penalty = repetition_penalty
-        self.presence_penalty = presence_penalty
-        self.frequency_penalty = frequency_penalty
-        self.min_p = min_p
-        self.safety_model = safety_model
-        self.system_message = system_message
-
-        self.messages = (
-            [{"role": "system", "content": self.system_message}]
-            if self.system_message
-            else []
-        )
-
-    def precmd(self, line: str) -> str:
-        if line.startswith("/"):
-            return line[1:]
-        else:
-            return "say " + line
-
-    def do_say(self, arg: str) -> None:
-        self.messages.append({"role": "user", "content": arg})
-
-        output = ""
-
-        for chunk in self.client.chat.completions.create(
-            messages=self.messages,
-            model=self.model,
-            max_tokens=self.max_tokens,
-            stop=self.stop,
-            temperature=self.temperature,
-            top_p=self.top_p,
-            top_k=self.top_k,
-            repetition_penalty=self.repetition_penalty,
-            presence_penalty=self.presence_penalty,
-            frequency_penalty=self.frequency_penalty,
-            min_p=self.min_p,
-            safety_model=self.safety_model,
-            stream=True,
-        ):
-            # assertions for type checking
-            assert isinstance(chunk, ChatCompletionChunk)
-            assert chunk.choices
-            assert chunk.choices[0].delta
-
-            token = chunk.choices[0].delta.content
-
-            click.echo(token, nl=False)
-
-            output += token or ""
-
-        click.echo("\n")
-
-        self.messages.append({"role": "assistant", "content": output})
-
-    def do_reset(self, arg: str) -> None:
-        self.messages = (
-            [{"role": "system", "content": self.system_message}]
-            if self.system_message
-            else []
-        )
-
-    def do_exit(self, arg: str) -> bool:
-        return True
-
-
-@click.command(name="chat.interactive")
-@click.pass_context
-@click.option("--model", type=str, required=True, help="Model name")
-@click.option("--max-tokens", type=int, help="Max tokens to generate")
-@click.option(
-    "--stop", type=str, multiple=True, help="List of strings to stop generation"
-)
-@click.option("--temperature", type=float, help="Sampling temperature")
-@click.option("--top-p", type=int, help="Top p sampling")
-@click.option("--top-k", type=float, help="Top k sampling")
-@click.option("--repetition-penalty", type=float, help="Repetition penalty")
-@click.option("--presence-penalty", type=float, help="Presence penalty")
-@click.option("--frequency-penalty", type=float, help="Frequency penalty")
-@click.option("--min-p", type=float, help="Minimum p")
-@click.option("--safety-model", type=str, help="Moderation model")
-@click.option("--system-message", type=str, help="System message to use for the chat")
-def interactive(
-    ctx: click.Context,
-    model: str,
-    max_tokens: int | None = None,
-    stop: List[str] | None = None,
-    temperature: float | None = None,
-    top_p: float | None = None,
-    top_k: int | None = None,
-    repetition_penalty: float | None = None,
-    presence_penalty: float | None = None,
-    frequency_penalty: float | None = None,
-    min_p: float | None = None,
-    safety_model: str | None = None,
-    system_message: str | None = None,
-) -> None:
-    """Interactive chat shell"""
-    client: Together = ctx.obj
-
-    ChatShell(
-        client=client,
-        model=model,
-        max_tokens=max_tokens,
-        stop=stop,
-        temperature=temperature,
-        top_p=top_p,
-        top_k=top_k,
-        repetition_penalty=repetition_penalty,
-        presence_penalty=presence_penalty,
-        frequency_penalty=frequency_penalty,
-        min_p=min_p,
-        safety_model=safety_model,
-        system_message=system_message,
-    ).cmdloop()
-
-
-@click.command(name="chat.completions")
-@click.pass_context
-@click.option(
-    "--message",
-    type=(str, str),
-    multiple=True,
-    required=True,
-    help="Message to generate chat completions from",
-)
-@click.option("--model", type=str, required=True, help="Model name")
-@click.option("--max-tokens", type=int, help="Max tokens to generate")
-@click.option(
-    "--stop", type=str, multiple=True, help="List of strings to stop generation"
-)
-@click.option("--temperature", type=float, help="Sampling temperature")
-@click.option("--top-p", type=int, help="Top p sampling")
-@click.option("--top-k", type=float, help="Top k sampling")
-@click.option("--repetition-penalty", type=float, help="Repetition penalty")
-@click.option("--presence-penalty", type=float, help="Presence penalty sampling method")
-@click.option(
-    "--frequency-penalty", type=float, help="Frequency penalty sampling method"
-)
-@click.option("--min-p", type=float, help="Min p sampling")
-@click.option("--no-stream", is_flag=True, help="Disable streaming")
-@click.option("--logprobs", type=int, help="Return logprobs. Only works with --raw.")
-@click.option("--echo", is_flag=True, help="Echo prompt. Only works with --raw.")
-@click.option("--n", type=int, help="Number of output generations")
-@click.option("--safety-model", type=str, help="Moderation model")
-@click.option("--raw", is_flag=True, help="Output raw JSON")
-def chat(
-    ctx: click.Context,
-    message: List[Tuple[str, str]],
-    model: str,
-    max_tokens: int | None = None,
-    stop: List[str] | None = None,
-    temperature: float | None = None,
-    top_p: float | None = None,
-    top_k: int | None = None,
-    repetition_penalty: float | None = None,
-    presence_penalty: float | None = None,
-    frequency_penalty: float | None = None,
-    min_p: float | None = None,
-    no_stream: bool = False,
-    logprobs: int | None = None,
-    echo: bool | None = None,
-    n: int | None = None,
-    safety_model: str | None = None,
-    raw: bool = False,
-) -> None:
-    """Generate chat completions from messages"""
-    client: Together = ctx.obj
-
-    messages = [{"role": msg[0], "content": msg[1]} for msg in message]
-
-    response = client.chat.completions.create(
-        model=model,
-        messages=messages,
-        top_p=top_p,
-        top_k=top_k,
-        temperature=temperature,
-        max_tokens=max_tokens,
-        stop=stop,
-        repetition_penalty=repetition_penalty,
-        presence_penalty=presence_penalty,
-        frequency_penalty=frequency_penalty,
-        min_p=min_p,
-        stream=not no_stream,
-        logprobs=logprobs,
-        echo=echo,
-        n=n,
-        safety_model=safety_model,
-    )
-
-    if not no_stream:
-        for chunk in response:
-            # assertions for type checking
-            assert isinstance(chunk, ChatCompletionChunk)
-            assert chunk.choices
-
-            if raw:
-                click.echo(f"{json.dumps(chunk.model_dump(exclude_none=True))}")
-                continue
-
-            should_print_header = len(chunk.choices) > 1
-            for stream_choice in sorted(chunk.choices, key=lambda c: c.index):  # type: ignore
-                assert isinstance(stream_choice, ChatCompletionChoicesChunk)
-                assert stream_choice.delta
-
-                if should_print_header:
-                    click.echo(f"\n===== Completion {stream_choice.index} =====\n")
-                click.echo(f"{stream_choice.delta.content}", nl=False)
-
-                if should_print_header:
-                    click.echo("\n")
-
-        # new line after stream ends
-        click.echo("\n")
-    else:
-        # assertions for type checking
-        assert isinstance(response, ChatCompletionResponse)
-        assert isinstance(response.choices, list)
-
-        if raw:
-            click.echo(
-                f"{json.dumps(response.model_dump(exclude_none=True), indent=4)}"
-            )
-            return
-
-        should_print_header = len(response.choices) > 1
-        for i, choice in enumerate(response.choices):
-            if should_print_header:
-                click.echo(f"===== Completion {i} =====")
-            click.echo(choice.message.content)  # type: ignore
-
-            if should_print_header:
-                click.echo("\n")

together/cli/api/completions.py

@@ -1,119 +0,0 @@
-from __future__ import annotations
-
-import json
-from typing import List
-
-import click
-
-from together import Together
-from together.types import CompletionChunk
-from together.types.completions import CompletionChoicesChunk, CompletionResponse
-
-
-@click.command()
-@click.pass_context
-@click.argument("prompt", type=str, required=True)
-@click.option("--model", type=str, required=True, help="Model name")
-@click.option("--max-tokens", type=int, help="Max tokens to generate")
-@click.option(
-    "--stop", type=str, multiple=True, help="List of strings to stop generation"
-)
-@click.option("--temperature", type=float, help="Sampling temperature")
-@click.option("--top-p", type=int, help="Top p sampling")
-@click.option("--top-k", type=float, help="Top k sampling")
-@click.option("--repetition-penalty", type=float, help="Repetition penalty")
-@click.option("--presence-penalty", type=float, help="Presence penalty")
-@click.option("--frequency-penalty", type=float, help="Frequency penalty")
-@click.option("--min-p", type=float, help="Minimum p")
-@click.option("--no-stream", is_flag=True, help="Disable streaming")
-@click.option("--logprobs", type=int, help="Return logprobs. Only works with --raw.")
-@click.option("--echo", is_flag=True, help="Echo prompt. Only works with --raw.")
-@click.option("--n", type=int, help="Number of output generations")
-@click.option("--safety-model", type=str, help="Moderation model")
-@click.option("--raw", is_flag=True, help="Return raw JSON response")
-def completions(
-    ctx: click.Context,
-    prompt: str,
-    model: str,
-    max_tokens: int | None = 512,
-    stop: List[str] | None = None,
-    temperature: float | None = None,
-    top_p: float | None = None,
-    top_k: int | None = None,
-    repetition_penalty: float | None = None,
-    presence_penalty: float | None = None,
-    frequency_penalty: float | None = None,
-    min_p: float | None = None,
-    no_stream: bool = False,
-    logprobs: int | None = None,
-    echo: bool | None = None,
-    n: int | None = None,
-    safety_model: str | None = None,
-    raw: bool = False,
-) -> None:
-    """Generate text completions"""
-    client: Together = ctx.obj
-
-    response = client.completions.create(
-        model=model,
-        prompt=prompt,
-        top_p=top_p,
-        top_k=top_k,
-        temperature=temperature,
-        max_tokens=max_tokens,
-        stop=stop,
-        repetition_penalty=repetition_penalty,
-        presence_penalty=presence_penalty,
-        frequency_penalty=frequency_penalty,
-        min_p=min_p,
-        stream=not no_stream,
-        logprobs=logprobs,
-        echo=echo,
-        n=n,
-        safety_model=safety_model,
-    )
-
-    if not no_stream:
-        for chunk in response:
-            # assertions for type checking
-            assert isinstance(chunk, CompletionChunk)
-            assert chunk.choices
-
-            if raw:
-                click.echo(f"{json.dumps(chunk.model_dump(exclude_none=True))}")
-                continue
-
-            should_print_header = len(chunk.choices) > 1
-            for stream_choice in sorted(chunk.choices, key=lambda c: c.index):  # type: ignore
-                # assertions for type checking
-                assert isinstance(stream_choice, CompletionChoicesChunk)
-                assert stream_choice.delta
-
-                if should_print_header:
-                    click.echo(f"\n===== Completion {stream_choice.index} =====\n")
-                click.echo(f"{stream_choice.delta.content}", nl=False)
-
-                if should_print_header:
-                    click.echo("\n")
-
-        # new line after stream ends
-        click.echo("\n")
-    else:
-        # assertions for type checking
-        assert isinstance(response, CompletionResponse)
-        assert isinstance(response.choices, list)
-
-        if raw:
-            click.echo(
-                f"{json.dumps(response.model_dump(exclude_none=True), indent=4)}"
-            )
-            return
-
-        should_print_header = len(response.choices) > 1
-        for i, choice in enumerate(response.choices):
-            if should_print_header:
-                click.echo(f"===== Completion {i} =====")
-            click.echo(choice.text)
-
-            if should_print_header or not choice.text.endswith("\n"):
-                click.echo("\n")

together/cli/api/finetune.py

@@ -1,272 +0,0 @@
-from __future__ import annotations
-
-import json
-from textwrap import wrap
-
-import click
-from click.core import ParameterSource  # type: ignore[attr-defined]
-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
-
-
-class DownloadCheckpointTypeChoice(click.Choice):
-    def __init__(self) -> None:
-        super().__init__([ct.value for ct in DownloadCheckpointType])
-
-    def convert(
-        self, value: str, param: click.Parameter | None, ctx: click.Context | None
-    ) -> DownloadCheckpointType:
-        value = super().convert(value, param, ctx)
-        return DownloadCheckpointType(value)
-
-
-@click.group(name="fine-tuning")
-@click.pass_context
-def fine_tuning(ctx: click.Context) -> None:
-    """Fine-tunes API commands"""
-    pass
-
-
-@fine_tuning.command()
-@click.pass_context
-@click.option(
-    "--training-file", type=str, required=True, help="Training file ID from Files API"
-)
-@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"
-)
-@click.option("--batch-size", type=int, default=16, help="Train batch size")
-@click.option("--learning-rate", type=float, default=1e-5, help="Learning rate")
-@click.option(
-    "--lora/--no-lora",
-    type=bool,
-    default=False,
-    help="Whether to use LoRA adapters for fine-tuning",
-)
-@click.option("--lora-r", type=int, default=8, help="LoRA adapters' rank")
-@click.option("--lora-dropout", type=float, default=0, help="LoRA adapters' dropout")
-@click.option("--lora-alpha", type=float, default=8, help="LoRA adapters' alpha")
-@click.option(
-    "--lora-trainable-modules",
-    type=str,
-    default="all-linear",
-    help="Trainable modules for LoRA adapters. For example, 'all-linear', 'q_proj,v_proj'",
-)
-@click.option(
-    "--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")
-def create(
-    ctx: click.Context,
-    training_file: str,
-    validation_file: str,
-    model: str,
-    n_epochs: int,
-    n_evals: int,
-    n_checkpoints: int,
-    batch_size: int,
-    learning_rate: float,
-    lora: bool,
-    lora_r: int,
-    lora_dropout: float,
-    lora_alpha: float,
-    lora_trainable_modules: str,
-    suffix: str,
-    wandb_api_key: str,
-) -> None:
-    """Start fine-tuning"""
-    client: Together = ctx.obj
-
-    if lora:
-        learning_rate_source = click.get_current_context().get_parameter_source(  # type: ignore[attr-defined]
-            "learning_rate"
-        )
-        if learning_rate_source == ParameterSource.DEFAULT:
-            learning_rate = 1e-3
-    else:
-        for param in ["lora_r", "lora_dropout", "lora_alpha", "lora_trainable_modules"]:
-            param_source = click.get_current_context().get_parameter_source(param)  # type: ignore[attr-defined]
-            if param_source != ParameterSource.DEFAULT:
-                raise click.BadParameter(
-                    f"You set LoRA parameter `{param}` for a full fine-tuning job. "
-                    f"Please change the job type with --lora or remove `{param}` from the arguments"
-                )
-    if n_evals <= 0 and validation_file:
-        log_warn(
-            "Warning: You have specified a validation file but the number of evaluation loops is set to 0. No evaluations will be performed."
-        )
-    elif n_evals > 0 and not validation_file:
-        raise click.BadParameter(
-            "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))
-
-    # 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"
-    )
-
-
-@fine_tuning.command()
-@click.pass_context
-def list(ctx: click.Context) -> None:
-    """List fine-tuning jobs"""
-    client: Together = ctx.obj
-
-    response = client.fine_tuning.list()
-
-    response.data = response.data or []
-
-    response.data.sort(key=lambda x: parse_timestamp(x.created_at or ""))
-
-    display_list = []
-    for i in response.data:
-        display_list.append(
-            {
-                "Fine-tune ID": i.id,
-                "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
-            }
-        )
-    table = tabulate(display_list, headers="keys", tablefmt="grid", showindex=True)
-
-    click.echo(table)
-
-
-@fine_tuning.command()
-@click.pass_context
-@click.argument("fine_tune_id", type=str, required=True)
-def retrieve(ctx: click.Context, fine_tune_id: str) -> None:
-    """Retrieve fine-tuning job details"""
-    client: Together = ctx.obj
-
-    response = client.fine_tuning.retrieve(fine_tune_id)
-
-    # remove events from response for cleaner output
-    response.events = None
-
-    click.echo(json.dumps(response.model_dump(exclude_none=True), indent=4))
-
-
-@fine_tuning.command()
-@click.pass_context
-@click.argument("fine_tune_id", type=str, required=True)
-@click.option(
-    "--quiet", is_flag=True, help="Do not prompt for confirmation before cancelling job"
-)
-def cancel(ctx: click.Context, fine_tune_id: str, quiet: bool = False) -> None:
-    """Cancel fine-tuning job"""
-    client: Together = ctx.obj
-    if not quiet:
-        confirm_response = input(
-            "You will be billed for any completed training steps upon cancellation. "
-            f"Do you want to cancel job {fine_tune_id}? [y/N]"
-        )
-        if "y" not in confirm_response.lower():
-            click.echo({"status": "Cancel not submitted"})
-            return
-    response = client.fine_tuning.cancel(fine_tune_id)
-
-    click.echo(json.dumps(response.model_dump(exclude_none=True), indent=4))
-
-
-@fine_tuning.command()
-@click.pass_context
-@click.argument("fine_tune_id", type=str, required=True)
-def list_events(ctx: click.Context, fine_tune_id: str) -> None:
-    """List fine-tuning events"""
-    client: Together = ctx.obj
-
-    response = client.fine_tuning.list_events(fine_tune_id)
-
-    response.data = response.data or []
-
-    display_list = []
-    for i in response.data:
-        display_list.append(
-            {
-                "Message": "\n".join(wrap(i.message or "", width=50)),
-                "Type": i.type,
-                "Created At": parse_timestamp(i.created_at or ""),
-                "Hash": i.hash,
-            }
-        )
-    table = tabulate(display_list, headers="keys", tablefmt="grid", showindex=True)
-
-    click.echo(table)
-
-
-@fine_tuning.command()
-@click.pass_context
-@click.argument("fine_tune_id", type=str, required=True)
-@click.option(
-    "--output_dir",
-    type=click.Path(exists=True, file_okay=False, resolve_path=True),
-    required=False,
-    default=None,
-    help="Output directory",
-)
-@click.option(
-    "--checkpoint-step",
-    type=int,
-    required=False,
-    default=-1,
-    help="Download fine-tuning checkpoint. Defaults to latest.",
-)
-@click.option(
-    "--checkpoint-type",
-    type=DownloadCheckpointTypeChoice(),
-    required=False,
-    default=DownloadCheckpointType.DEFAULT.value,
-    help="Specifies checkpoint type. 'merged' and 'adapter' options work only for LoRA jobs.",
-)
-def download(
-    ctx: click.Context,
-    fine_tune_id: str,
-    output_dir: str,
-    checkpoint_step: int,
-    checkpoint_type: DownloadCheckpointType,
-) -> None:
-    """Download fine-tuning checkpoint"""
-    client: Together = ctx.obj
-
-    response = client.fine_tuning.download(
-        fine_tune_id,
-        output=output_dir,
-        checkpoint_step=checkpoint_step,
-        checkpoint_type=checkpoint_type,
-    )
-
-    click.echo(json.dumps(response.model_dump(exclude_none=True), indent=4))