together 2.0.0a9__py3-none-any.whl → 2.0.0a11__py3-none-any.whl

together/_base_client.py

@@ -1247,9 +1247,12 @@ class SyncAPIClient(BaseClient[httpx.Client, Stream[Any]]):
         *,
         cast_to: Type[ResponseT],
         body: Body | None = None,
+        files: RequestFiles | None = None,
         options: RequestOptions = {},
     ) -> ResponseT:
-        opts = FinalRequestOptions.construct(method="patch", url=path, json_data=body, **options)
+        opts = FinalRequestOptions.construct(
+            method="patch", url=path, json_data=body, files=to_httpx_files(files), **options
+        )
         return self.request(cast_to, opts)
 
     def put(
@@ -1767,9 +1770,12 @@ class AsyncAPIClient(BaseClient[httpx.AsyncClient, AsyncStream[Any]]):
         *,
         cast_to: Type[ResponseT],
         body: Body | None = None,
+        files: RequestFiles | None = None,
         options: RequestOptions = {},
     ) -> ResponseT:
-        opts = FinalRequestOptions.construct(method="patch", url=path, json_data=body, **options)
+        opts = FinalRequestOptions.construct(
+            method="patch", url=path, json_data=body, files=to_httpx_files(files), **options
+        )
         return await self.request(cast_to, opts)
 
     async def put(

together/_types.py

@@ -243,6 +243,9 @@ _T_co = TypeVar("_T_co", covariant=True)
 if TYPE_CHECKING:
     # This works because str.__contains__ does not accept object (either in typeshed or at runtime)
     # https://github.com/hauntsaninja/useful_types/blob/5e9710f3875107d068e7679fd7fec9cfab0eff3b/useful_types/__init__.py#L285
+    #
+    # Note: index() and count() methods are intentionally omitted to allow pyright to properly
+    # infer TypedDict types when dict literals are used in lists assigned to SequenceNotStr.
     class SequenceNotStr(Protocol[_T_co]):
         @overload
         def __getitem__(self, index: SupportsIndex, /) -> _T_co: ...
@@ -251,8 +254,6 @@ if TYPE_CHECKING:
         def __contains__(self, value: object, /) -> bool: ...
         def __len__(self) -> int: ...
         def __iter__(self) -> Iterator[_T_co]: ...
-        def index(self, value: Any, start: int = 0, stop: int = ..., /) -> int: ...
-        def count(self, value: Any, /) -> int: ...
         def __reversed__(self) -> Iterator[_T_co]: ...
 else:
     # just point this to a normal `Sequence` at runtime to avoid having to special case

together/_version.py

@@ -1,4 +1,4 @@
 # File generated from our OpenAPI spec by Stainless. See CONTRIBUTING.md for details.
 
 __title__ = "together"
-__version__ = "2.0.0-alpha.9"  # x-release-please-version
+__version__ = "2.0.0-alpha.11"  # x-release-please-version

together/lib/cli/api/fine_tuning.py

@@ -10,13 +10,15 @@ from textwrap import wrap
 import click
 from rich import print as rprint
 from tabulate import tabulate
+from rich.json import JSON
 from click.core import ParameterSource  # type: ignore[attr-defined]
 
 from together import Together
+from together.types import fine_tuning_estimate_price_params as pe_params
 from together._types import NOT_GIVEN, NotGiven
 from together.lib.utils import log_warn
 from together.lib.utils.tools import format_timestamp, finetune_price_to_dollars
-from together.lib.cli.api.utils import INT_WITH_MAX, BOOL_WITH_AUTO
+from together.lib.cli.api.utils import INT_WITH_MAX, BOOL_WITH_AUTO, generate_progress_bar
 from together.lib.resources.files import DownloadManager
 from together.lib.utils.serializer import datetime_serializer
 from together.types.finetune_response import TrainingTypeFullTrainingType, TrainingTypeLoRaTrainingType
@@ -24,13 +26,21 @@ from together.lib.resources.fine_tuning import get_model_limits
 
 _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 "
+    "The estimated price of this job is {price}. "
+    "The actual 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"
+    "the number of evaluations. Visit https://www.together.ai/pricing to learn more about pricing.\n"
+    "{warning}"
     "You can pass `-y` or `--confirm` to your command to skip this message.\n\n"
     "Do you want to proceed?"
 )
 
+_WARNING_MESSAGE_INSUFFICIENT_FUNDS = (
+    "The estimated price of this job is significantly greater than your current credit limit and balance combined. "
+    "It will likely get cancelled due to insufficient funds. "
+    "Consider increasing your credit limit at https://api.together.xyz/settings/profile\n"
+)
+
 _FT_JOB_WITH_STEP_REGEX = r"^ft-[\dabcdef-]+:\d+$"
 
 
@@ -323,7 +333,60 @@ def create(
     elif n_evals > 0 and not validation_file:
         raise click.BadParameter("You have specified a number of evaluation loops but no validation file.")
 
-    if confirm or click.confirm(_CONFIRMATION_MESSAGE, default=True, show_default=True):
+    training_type_cls: pe_params.TrainingType
+    if lora:
+        training_type_cls = pe_params.TrainingTypeLoRaTrainingType(
+            lora_alpha=int(lora_alpha or 0),
+            lora_r=lora_r or 0,
+            lora_dropout=lora_dropout or 0,
+            lora_trainable_modules=lora_trainable_modules or "all-linear",
+            type="Lora",
+        )
+    else:
+        training_type_cls = pe_params.TrainingTypeFullTrainingType(
+            type="Full",
+        )
+
+    training_method_cls: pe_params.TrainingMethod
+    if training_method == "sft":
+        training_method_cls = pe_params.TrainingMethodTrainingMethodSft(
+            method="sft",
+            train_on_inputs=train_on_inputs or "auto",
+        )
+    else:
+        training_method_cls = pe_params.TrainingMethodTrainingMethodDpo(
+            method="dpo",
+            dpo_beta=dpo_beta or 0,
+            dpo_normalize_logratios_by_length=dpo_normalize_logratios_by_length or False,
+            dpo_reference_free=False,
+            rpo_alpha=rpo_alpha or 0,
+            simpo_gamma=simpo_gamma or 0,
+        )
+
+    finetune_price_estimation_result = client.fine_tuning.estimate_price(
+        training_file=training_file,
+        validation_file=validation_file,
+        model=model or "",
+        n_epochs=n_epochs,
+        n_evals=n_evals,
+        training_type=training_type_cls,
+        training_method=training_method_cls,
+    )
+    price = click.style(
+        f"${finetune_price_estimation_result.estimated_total_price:.2f}",
+        bold=True,
+    )
+    if not finetune_price_estimation_result.allowed_to_proceed:
+        warning = click.style(_WARNING_MESSAGE_INSUFFICIENT_FUNDS, fg="red", bold=True)
+    else:
+        warning = ""
+
+    confirmation_message = _CONFIRMATION_MESSAGE.format(
+        price=price,
+        warning=warning,
+    )
+
+    if confirm or click.confirm(confirmation_message, default=True, show_default=True):
         response = client.fine_tuning.create(
             **training_args,
             verbose=True,
@@ -363,6 +426,9 @@ def list(ctx: click.Context) -> None:
                 "Price": f"""${
                     finetune_price_to_dollars(float(str(i.total_price)))
                 }""",  # convert to string for mypy typing
+                "Progress": generate_progress_bar(
+                    i, datetime.now().astimezone(), use_rich=False
+                ),
             }
         )
     table = tabulate(display_list, headers="keys", tablefmt="grid", showindex=True)
@@ -382,7 +448,12 @@ def retrieve(ctx: click.Context, fine_tune_id: str) -> None:
     # remove events from response for cleaner output
     response.events = None
 
-    click.echo(json.dumps(response.model_dump(exclude_none=True), indent=4))
+    rprint(JSON.from_data(response.model_json_schema()))
+    progress_text = generate_progress_bar(
+        response, datetime.now().astimezone(), use_rich=True
+    )
+    prefix = f"Status: [bold]{response.status}[/bold],"
+    rprint(f"{prefix} {progress_text}")
 
 
 @fine_tuning.command()

together/lib/cli/api/models.py

@@ -7,7 +7,6 @@ from tabulate import tabulate
 from together import Together, omit
 from together._models import BaseModel
 from together._response import APIResponse as APIResponse
-from together.lib.resources.models import filter_by_dedicated_models
 from together.types.model_upload_response import ModelUploadResponse
 
 
@@ -34,11 +33,7 @@ def list(ctx: click.Context, type: Optional[str], json: bool) -> None:
     """List models"""
     client: Together = ctx.obj
 
-    response = client.models.list()
-    models_list = response
-
-    if type == "dedicated":
-        models_list = filter_by_dedicated_models(client, models_list)
+    models_list = client.models.list(dedicated=type == "dedicated" if type else omit)
 
     display_list: List[Dict[str, Any]] = []
     model: BaseModel

together/lib/cli/api/utils.py

@@ -1,18 +1,25 @@
 from __future__ import annotations
 
-from typing import Literal
+import re
+import math
+from typing import List, Union, Literal
 from gettext import gettext as _
-from typing_extensions import override
+from datetime import datetime
 
 import click
 
+from together.lib.types.fine_tuning import COMPLETED_STATUSES, FinetuneResponse
+from together.types.finetune_response import FinetuneResponse as _FinetuneResponse
+from together.types.fine_tuning_list_response import Data
+
+_PROGRESS_BAR_WIDTH = 40
+
 
 class AutoIntParamType(click.ParamType):
     name = "integer_or_max"
     _number_class = int
 
-    @override
-    def convert(
+    def convert(  # pyright: ignore[reportImplicitOverride]
         self, value: str, param: click.Parameter | None, ctx: click.Context | None
     ) -> int | Literal["max"] | None:
         if value == "max":
@@ -21,7 +28,9 @@ class AutoIntParamType(click.ParamType):
             return int(value)
         except ValueError:
             self.fail(
-                _("{value!r} is not a valid {number_type}.").format(value=value, number_type=self.name),
+                _("{value!r} is not a valid {number_type}.").format(
+                    value=value, number_type=self.name
+                ),
                 param,
                 ctx,
             )
@@ -30,8 +39,7 @@ class AutoIntParamType(click.ParamType):
 class BooleanWithAutoParamType(click.ParamType):
     name = "boolean_or_auto"
 
-    @override
-    def convert(
+    def convert( # pyright: ignore[reportImplicitOverride]
         self, value: str, param: click.Parameter | None, ctx: click.Context | None
     ) -> bool | Literal["auto"] | None:
         if value == "auto":
@@ -40,7 +48,9 @@ class BooleanWithAutoParamType(click.ParamType):
             return bool(value)
         except ValueError:
             self.fail(
-                _("{value!r} is not a valid {type}.").format(value=value, type=self.name),
+                _("{value!r} is not a valid {type}.").format(
+                    value=value, type=self.name
+                ),
                 param,
                 ctx,
             )
@@ -48,3 +58,82 @@ class BooleanWithAutoParamType(click.ParamType):
 
 INT_WITH_MAX = AutoIntParamType()
 BOOL_WITH_AUTO = BooleanWithAutoParamType()
+
+
+def _human_readable_time(timedelta: float) -> str:
+    """Convert a timedelta to a compact human-readble string
+    Examples:
+        00:00:10 -> 10s
+        01:23:45 -> 1h 23min 45s
+        1 Month 23 days 04:56:07 -> 1month 23d 4h 56min 7s
+    Args:
+        timedelta (float): The timedelta in seconds to convert.
+    Returns:
+        A string representing the timedelta in a human-readable format.
+    """
+    units = [
+        (30 * 24 * 60 * 60, "month"),  # 30 days
+        (24 * 60 * 60, "d"),
+        (60 * 60, "h"),
+        (60, "min"),
+        (1, "s"),
+    ]
+
+    total_seconds = int(timedelta)
+    parts: List[str] = []
+
+    for unit_seconds, unit_name in units:
+        if total_seconds >= unit_seconds:
+            value = total_seconds // unit_seconds
+            total_seconds %= unit_seconds
+            parts.append(f"{value}{unit_name}")
+
+    return " ".join(parts) if parts else "0s"
+
+
+def generate_progress_bar(
+    finetune_job: Union[Data, FinetuneResponse, _FinetuneResponse], current_time: datetime, use_rich: bool = False
+) -> str:
+    """Generate a progress bar for a finetune job.
+    Args:
+        finetune_job: The finetune job to generate a progress bar for.
+        current_time: The current time.
+        use_rich: Whether to use rich formatting.
+    Returns:
+        A string representing the progress bar.
+    """
+    progress = "Progress: [bold red]unavailable[/bold red]"
+    if finetune_job.status in COMPLETED_STATUSES:
+        progress = "Progress: [bold green]completed[/bold green]"
+    elif finetune_job.updated_at is not None:
+        update_at = finetune_job.updated_at.astimezone()
+
+        if finetune_job.progress is not None:
+            if current_time < update_at:
+                return progress
+
+            if not finetune_job.progress.estimate_available:
+                return progress
+
+            if finetune_job.progress.seconds_remaining <= 0:
+                return progress
+
+            elapsed_time = (current_time - update_at).total_seconds()
+            ratio_filled = min(
+                elapsed_time / finetune_job.progress.seconds_remaining, 1.0
+            )
+            percentage = ratio_filled * 100
+            filled = math.ceil(ratio_filled * _PROGRESS_BAR_WIDTH)
+            bar = "█" * filled + "░" * (_PROGRESS_BAR_WIDTH - filled)
+            time_left = "N/A"
+            if finetune_job.progress.seconds_remaining > elapsed_time:
+                time_left = _human_readable_time(
+                    finetune_job.progress.seconds_remaining - elapsed_time
+                )
+            time_text = f"{time_left} left"
+            progress = f"Progress: {bar} [bold]{percentage:>3.0f}%[/bold] [yellow]{time_text}[/yellow]"
+
+    if use_rich:
+        return progress
+
+    return re.sub(r"\[/?[^\]]+\]", "", progress)

together/lib/constants.py

@@ -14,6 +14,9 @@ import enum
 # Download defaults
 DOWNLOAD_BLOCK_SIZE = 10 * 1024 * 1024  # 10 MB
 DISABLE_TQDM = False
+MAX_DOWNLOAD_RETRIES = 5  # Maximum retries for download failures
+DOWNLOAD_INITIAL_RETRY_DELAY = 1.0  # Initial retry delay in seconds
+DOWNLOAD_MAX_RETRY_DELAY = 30.0  # Maximum retry delay in seconds
 
 # Upload defaults
 MAX_CONCURRENT_PARTS = 4  # Maximum concurrent parts for multipart upload

together/lib/resources/files.py

@@ -3,6 +3,7 @@ from __future__ import annotations
 import os
 import math
 import stat
+import time
 import uuid
 import shutil
 import asyncio
@@ -29,12 +30,15 @@ from ..constants import (
     MAX_MULTIPART_PARTS,
     TARGET_PART_SIZE_MB,
     MAX_CONCURRENT_PARTS,
+    MAX_DOWNLOAD_RETRIES,
     MULTIPART_THRESHOLD_GB,
+    DOWNLOAD_MAX_RETRY_DELAY,
     MULTIPART_UPLOAD_TIMEOUT,
+    DOWNLOAD_INITIAL_RETRY_DELAY,
 )
 from ..._resource import SyncAPIResource, AsyncAPIResource
 from ..types.error import DownloadError, FileTypeError
-from ..._exceptions import APIStatusError, AuthenticationError
+from ..._exceptions import APIStatusError, APIConnectionError, AuthenticationError
 
 log: logging.Logger = logging.getLogger(__name__)
 
@@ -198,6 +202,11 @@ class DownloadManager(SyncAPIResource):
 
                 assert file_size != 0, "Unable to retrieve remote file."
 
+                # Download with retry logic
+                bytes_downloaded = 0
+                retry_count = 0
+                retry_delay = DOWNLOAD_INITIAL_RETRY_DELAY
+
                 with tqdm(
                     total=file_size,
                     unit="B",
@@ -205,14 +214,64 @@ class DownloadManager(SyncAPIResource):
                     desc=f"Downloading file {file_path.name}",
                     disable=bool(DISABLE_TQDM),
                 ) as pbar:
-                    for chunk in response.iter_bytes(DOWNLOAD_BLOCK_SIZE):
-                        pbar.update(len(chunk))
-                        temp_file.write(chunk)  # type: ignore
+                    while bytes_downloaded < file_size:
+                        try:
+                            # If this is a retry, close the previous response and create a new one with Range header
+                            if bytes_downloaded > 0:
+                                response.close()
+
+                                log.info(f"Resuming download from byte {bytes_downloaded}")
+                                response = self._client.get(
+                                    path=url,
+                                    cast_to=httpx.Response,
+                                    stream=True,
+                                    options=RequestOptions(
+                                        headers={"Range": f"bytes={bytes_downloaded}-"},
+                                    ),
+                                )
+
+                            # Download chunks
+                            for chunk in response.iter_bytes(DOWNLOAD_BLOCK_SIZE):
+                                temp_file.write(chunk)  # type: ignore
+                                bytes_downloaded += len(chunk)
+                                pbar.update(len(chunk))
+
+                            # Successfully completed download
+                            break
+
+                        except (httpx.RequestError, httpx.StreamError, APIConnectionError) as e:
+                            if retry_count >= MAX_DOWNLOAD_RETRIES:
+                                log.error(f"Download failed after {retry_count} retries")
+                                raise DownloadError(
+                                    f"Download failed after {retry_count} retries. Last error: {str(e)}"
+                                ) from e
+
+                            retry_count += 1
+                            log.warning(
+                                f"Download interrupted at {bytes_downloaded}/{file_size} bytes. "
+                                f"Retry {retry_count}/{MAX_DOWNLOAD_RETRIES} in {retry_delay}s..."
+                            )
+                            time.sleep(retry_delay)
+
+                            # Exponential backoff with max delay cap
+                            retry_delay = min(retry_delay * 2, DOWNLOAD_MAX_RETRY_DELAY)
+
+                        except APIStatusError as e:
+                            # For API errors, don't retry
+                            log.error(f"API error during download: {e}")
+                            raise APIStatusError(
+                                "Error downloading file",
+                                response=e.response,
+                                body=e.response,
+                            ) from e
+
+                # Close the response
+                response.close()
 
             # Raise exception if remote file size does not match downloaded file size
             if os.stat(temp_file.name).st_size != file_size:
-                DownloadError(
-                    f"Downloaded file size `{pbar.n}` bytes does not match remote file size `{file_size}` bytes."
+                raise DownloadError(
+                    f"Downloaded file size `{bytes_downloaded}` bytes does not match remote file size `{file_size}` bytes."
                 )
 
             # Moves temp file to output file path

together/lib/resources/fine_tuning.py

@@ -4,6 +4,7 @@ from typing import TYPE_CHECKING, Literal
 
 from rich import print as rprint
 
+from together.types import fine_tuning_estimate_price_params as pe_params
 from together.lib.utils import log_warn_once
 
 if TYPE_CHECKING:
@@ -66,7 +67,7 @@ def create_finetune_request(
     hf_model_revision: str | None = None,
     hf_api_token: str | None = None,
     hf_output_repo_name: str | None = None,
-) -> FinetuneRequest:
+) -> tuple[FinetuneRequest, pe_params.TrainingType, pe_params.TrainingMethod]:
     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")
 
@@ -233,8 +234,46 @@ def create_finetune_request(
         hf_output_repo_name=hf_output_repo_name,
     )
 
-    return finetune_request
+    training_type_pe, training_method_pe = create_price_estimation_params(finetune_request)
 
+    return finetune_request, training_type_pe, training_method_pe
+
+def create_price_estimation_params(finetune_request: FinetuneRequest) -> tuple[pe_params.TrainingType, pe_params.TrainingMethod]:
+    training_type_cls: pe_params.TrainingType
+    if isinstance(finetune_request.training_type, FullTrainingType):
+        training_type_cls = pe_params.TrainingTypeFullTrainingType(
+            type="Full",
+        )
+    elif isinstance(finetune_request.training_type, LoRATrainingType):
+        training_type_cls = pe_params.TrainingTypeLoRaTrainingType(
+            lora_alpha=finetune_request.training_type.lora_alpha,
+            lora_r=finetune_request.training_type.lora_r,
+            lora_dropout=finetune_request.training_type.lora_dropout,
+            lora_trainable_modules=finetune_request.training_type.lora_trainable_modules,
+            type="Lora",
+        )
+    else:
+        raise ValueError(f"Unknown training type: {finetune_request.training_type}")
+
+    training_method_cls: pe_params.TrainingMethod
+    if isinstance(finetune_request.training_method, TrainingMethodSFT):
+        training_method_cls = pe_params.TrainingMethodTrainingMethodSft(
+            method="sft",
+            train_on_inputs=finetune_request.training_method.train_on_inputs,
+        )
+    elif isinstance(finetune_request.training_method, TrainingMethodDPO):
+        training_method_cls = pe_params.TrainingMethodTrainingMethodDpo(
+            method="dpo",
+            dpo_beta=finetune_request.training_method.dpo_beta or 0,
+            dpo_normalize_logratios_by_length=finetune_request.training_method.dpo_normalize_logratios_by_length,
+            dpo_reference_free=finetune_request.training_method.dpo_reference_free,
+            rpo_alpha=finetune_request.training_method.rpo_alpha or 0,
+            simpo_gamma=finetune_request.training_method.simpo_gamma or 0,
+        )
+    else:
+        raise ValueError(f"Unknown training method: {finetune_request.training_method}")
+
+    return training_type_cls, training_method_cls
 
 def get_model_limits(client: Together, model: str) -> FinetuneTrainingLimits:
     """

together/lib/types/fine_tuning.py

@@ -25,6 +25,14 @@ class FinetuneJobStatus(str, Enum):
     STATUS_COMPLETED = "completed"
 
 
+COMPLETED_STATUSES = [
+    FinetuneJobStatus.STATUS_ERROR,
+    FinetuneJobStatus.STATUS_USER_ERROR,
+    FinetuneJobStatus.STATUS_COMPLETED,
+    FinetuneJobStatus.STATUS_CANCELLED,
+]
+
+
 class FinetuneEventType(str, Enum):
     """
     Fine-tune job event types
@@ -260,6 +268,15 @@ FinetuneLRScheduler: TypeAlias = Union[
 ]
 
 
+class FinetuneProgress(BaseModel):
+    """
+    Fine-tune job progress
+    """
+
+    estimate_available: bool = False
+    seconds_remaining: float = 0
+
+
 class FinetuneResponse(BaseModel):
     """
     Fine-tune API response type
@@ -393,6 +410,8 @@ class FinetuneResponse(BaseModel):
     training_file_size: Optional[int] = Field(None, alias="TrainingFileSize")
     train_on_inputs: Union[StrictBool, Literal["auto"], None] = "auto"
 
+    progress: Union[FinetuneProgress, None] = None
+
     @classmethod
     def validate_training_type(cls, v: TrainingType) -> TrainingType:
         if v.type == "Full" or v.type == "":

together/resources/audio/transcriptions.py

@@ -47,7 +47,7 @@ class TranscriptionsResource(SyncAPIResource):
     def create(
         self,
         *,
-        file: FileTypes,
+        file: Union[FileTypes, str],
         diarize: bool | Omit = omit,
         language: str | Omit = omit,
         max_speakers: int | Omit = omit,
@@ -68,7 +68,8 @@ class TranscriptionsResource(SyncAPIResource):
         Transcribes audio into text
 
         Args:
-          file: Audio file to transcribe
+          file: Audio file upload or public HTTP/HTTPS URL. Supported formats .wav, .mp3, .m4a,
+              .webm, .flac.
 
           diarize: Whether to enable speaker diarization. When enabled, you will get the speaker id
               for each word in the transcription. In the response, in the words array, you
@@ -168,7 +169,7 @@ class AsyncTranscriptionsResource(AsyncAPIResource):
     async def create(
         self,
         *,
-        file: FileTypes,
+        file: Union[FileTypes, str],
         diarize: bool | Omit = omit,
         language: str | Omit = omit,
         max_speakers: int | Omit = omit,
@@ -189,7 +190,8 @@ class AsyncTranscriptionsResource(AsyncAPIResource):
         Transcribes audio into text
 
         Args:
-          file: Audio file to transcribe
+          file: Audio file upload or public HTTP/HTTPS URL. Supported formats .wav, .mp3, .m4a,
+              .webm, .flac.
 
           diarize: Whether to enable speaker diarization. When enabled, you will get the speaker id
               for each word in the transcription. In the response, in the words array, you

together/resources/audio/translations.py

@@ -47,7 +47,7 @@ class TranslationsResource(SyncAPIResource):
     def create(
         self,
         *,
-        file: FileTypes,
+        file: Union[FileTypes, str],
         language: str | Omit = omit,
         model: Literal["openai/whisper-large-v3"] | Omit = omit,
         prompt: str | Omit = omit,
@@ -65,7 +65,8 @@ class TranslationsResource(SyncAPIResource):
         Translates audio into English
 
         Args:
-          file: Audio file to translate
+          file: Audio file upload or public HTTP/HTTPS URL. Supported formats .wav, .mp3, .m4a,
+              .webm, .flac.
 
           language: Target output language. Optional ISO 639-1 language code. If omitted, language
               is set to English.
@@ -145,7 +146,7 @@ class AsyncTranslationsResource(AsyncAPIResource):
     async def create(
         self,
         *,
-        file: FileTypes,
+        file: Union[FileTypes, str],
         language: str | Omit = omit,
         model: Literal["openai/whisper-large-v3"] | Omit = omit,
         prompt: str | Omit = omit,
@@ -163,7 +164,8 @@ class AsyncTranslationsResource(AsyncAPIResource):
         Translates audio into English
 
         Args:
-          file: Audio file to translate
+          file: Audio file upload or public HTTP/HTTPS URL. Supported formats .wav, .mp3, .m4a,
+              .webm, .flac.
 
           language: Target output language. Optional ISO 639-1 language code. If omitted, language
               is set to English.