together 2.0.0a8__py3-none-any.whl → 2.0.0a10__py3-none-any.whl

together/_streaming.py

@@ -55,30 +55,31 @@ class Stream(Generic[_T]):
         process_data = self._client._process_response_data
         iterator = self._iter_events()
 
-        for sse in iterator:
-            if sse.data.startswith("[DONE]"):
-                break
-
-            if sse.event is None:
-                data = sse.json()
-                if is_mapping(data) and data.get("error"):
-                    message = None
-                    error = data.get("error")
-                    if is_mapping(error):
-                        message = error.get("message")
-                    if not message or not isinstance(message, str):
-                        message = "An error occurred during streaming"
-
-                    raise APIError(
-                        message=message,
-                        request=self.response.request,
-                        body=data["error"],
-                    )
-
-                yield process_data(data=data, cast_to=cast_to, response=response)
-
-        # As we might not fully consume the response stream, we need to close it explicitly
-        response.close()
+        try:
+            for sse in iterator:
+                if sse.data.startswith("[DONE]"):
+                    break
+
+                if sse.event is None:
+                    data = sse.json()
+                    if is_mapping(data) and data.get("error"):
+                        message = None
+                        error = data.get("error")
+                        if is_mapping(error):
+                            message = error.get("message")
+                        if not message or not isinstance(message, str):
+                            message = "An error occurred during streaming"
+
+                        raise APIError(
+                            message=message,
+                            request=self.response.request,
+                            body=data["error"],
+                        )
+
+                    yield process_data(data=data, cast_to=cast_to, response=response)
+        finally:
+            # Ensure the response is closed even if the consumer doesn't read all data
+            response.close()
 
     def __enter__(self) -> Self:
         return self
@@ -137,30 +138,31 @@ class AsyncStream(Generic[_T]):
         process_data = self._client._process_response_data
         iterator = self._iter_events()
 
-        async for sse in iterator:
-            if sse.data.startswith("[DONE]"):
-                break
-
-            if sse.event is None:
-                data = sse.json()
-                if is_mapping(data) and data.get("error"):
-                    message = None
-                    error = data.get("error")
-                    if is_mapping(error):
-                        message = error.get("message")
-                    if not message or not isinstance(message, str):
-                        message = "An error occurred during streaming"
-
-                    raise APIError(
-                        message=message,
-                        request=self.response.request,
-                        body=data["error"],
-                    )
-
-                yield process_data(data=data, cast_to=cast_to, response=response)
-
-        # As we might not fully consume the response stream, we need to close it explicitly
-        await response.aclose()
+        try:
+            async for sse in iterator:
+                if sse.data.startswith("[DONE]"):
+                    break
+
+                if sse.event is None:
+                    data = sse.json()
+                    if is_mapping(data) and data.get("error"):
+                        message = None
+                        error = data.get("error")
+                        if is_mapping(error):
+                            message = error.get("message")
+                        if not message or not isinstance(message, str):
+                            message = "An error occurred during streaming"
+
+                        raise APIError(
+                            message=message,
+                            request=self.response.request,
+                            body=data["error"],
+                        )
+
+                    yield process_data(data=data, cast_to=cast_to, response=response)
+        finally:
+            # Ensure the response is closed even if the consumer doesn't read all data
+            await response.aclose()
 
     async def __aenter__(self) -> Self:
         return self

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/_utils/_utils.py

@@ -373,7 +373,7 @@ def get_required_header(headers: HeadersLike, header: str) -> str:
     lower_header = header.lower()
     if is_mapping_t(headers):
         # mypy doesn't understand the type narrowing here
-        for k, v in headers.items():  # type: ignore[misc, has-type]
+        for k, v in headers.items():  # type: ignore[misc, has-type, attr-defined]
             if k.lower() == lower_header and isinstance(v, str):  # type: ignore[has-type]
                 return v  # type: ignore[has-type]
 

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.8"  # x-release-please-version
+__version__ = "2.0.0-alpha.10"  # x-release-please-version

together/lib/cli/api/fine_tuning.py

@@ -13,6 +13,7 @@ from tabulate import tabulate
 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
@@ -24,13 +25,21 @@ 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 +332,60 @@ 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,

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/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/resources/audio/transcriptions.py

@@ -50,6 +50,8 @@ class TranscriptionsResource(SyncAPIResource):
         file: FileTypes,
         diarize: bool | Omit = omit,
         language: str | Omit = omit,
+        max_speakers: int | Omit = omit,
+        min_speakers: int | Omit = omit,
         model: Literal["openai/whisper-large-v3"] | Omit = omit,
         prompt: str | Omit = omit,
         response_format: Literal["json", "verbose_json"] | Omit = omit,
@@ -82,6 +84,12 @@ class TranscriptionsResource(SyncAPIResource):
           language: Optional ISO 639-1 language code. If `auto` is provided, language is
               auto-detected.
 
+          max_speakers: Maximum number of speakers expected in the audio. Used to improve diarization
+              accuracy when the approximate number of speakers is known.
+
+          min_speakers: Minimum number of speakers expected in the audio. Used to improve diarization
+              accuracy when the approximate number of speakers is known.
+
           model: Model to use for transcription
 
           prompt: Optional text to bias decoding.
@@ -107,6 +115,8 @@ class TranscriptionsResource(SyncAPIResource):
                 "file": file,
                 "diarize": diarize,
                 "language": language,
+                "max_speakers": max_speakers,
+                "min_speakers": min_speakers,
                 "model": model,
                 "prompt": prompt,
                 "response_format": response_format,
@@ -161,6 +171,8 @@ class AsyncTranscriptionsResource(AsyncAPIResource):
         file: FileTypes,
         diarize: bool | Omit = omit,
         language: str | Omit = omit,
+        max_speakers: int | Omit = omit,
+        min_speakers: int | Omit = omit,
         model: Literal["openai/whisper-large-v3"] | Omit = omit,
         prompt: str | Omit = omit,
         response_format: Literal["json", "verbose_json"] | Omit = omit,
@@ -193,6 +205,12 @@ class AsyncTranscriptionsResource(AsyncAPIResource):
           language: Optional ISO 639-1 language code. If `auto` is provided, language is
               auto-detected.
 
+          max_speakers: Maximum number of speakers expected in the audio. Used to improve diarization
+              accuracy when the approximate number of speakers is known.
+
+          min_speakers: Minimum number of speakers expected in the audio. Used to improve diarization
+              accuracy when the approximate number of speakers is known.
+
           model: Model to use for transcription
 
           prompt: Optional text to bias decoding.
@@ -218,6 +236,8 @@ class AsyncTranscriptionsResource(AsyncAPIResource):
                 "file": file,
                 "diarize": diarize,
                 "language": language,
+                "max_speakers": max_speakers,
+                "min_speakers": min_speakers,
                 "model": model,
                 "prompt": prompt,
                 "response_format": response_format,

together/resources/chat/completions.py

@@ -136,6 +136,14 @@ class CompletionsResource(SyncAPIResource):
 
           response_format: An object specifying the format that the model must output.
 
+              Setting to `{ "type": "json_schema", "json_schema": {...} }` enables Structured
+              Outputs which ensures the model will match your supplied JSON schema. Learn more
+              in the [Structured Outputs guide](https://docs.together.ai/docs/json-mode).
+
+              Setting to `{ "type": "json_object" }` enables the older JSON mode, which
+              ensures the message the model generates is valid JSON. Using `json_schema` is
+              preferred for models that support it.
+
           safety_model: The name of the moderation model used to validate tokens. Choose from the
               available moderation models found
               [here](https://docs.together.ai/docs/inference-models#moderation-models).
@@ -277,6 +285,14 @@ class CompletionsResource(SyncAPIResource):
 
           response_format: An object specifying the format that the model must output.
 
+              Setting to `{ "type": "json_schema", "json_schema": {...} }` enables Structured
+              Outputs which ensures the model will match your supplied JSON schema. Learn more
+              in the [Structured Outputs guide](https://docs.together.ai/docs/json-mode).
+
+              Setting to `{ "type": "json_object" }` enables the older JSON mode, which
+              ensures the message the model generates is valid JSON. Using `json_schema` is
+              preferred for models that support it.
+
           safety_model: The name of the moderation model used to validate tokens. Choose from the
               available moderation models found
               [here](https://docs.together.ai/docs/inference-models#moderation-models).
@@ -414,6 +430,14 @@ class CompletionsResource(SyncAPIResource):
 
           response_format: An object specifying the format that the model must output.
 
+              Setting to `{ "type": "json_schema", "json_schema": {...} }` enables Structured
+              Outputs which ensures the model will match your supplied JSON schema. Learn more
+              in the [Structured Outputs guide](https://docs.together.ai/docs/json-mode).
+
+              Setting to `{ "type": "json_object" }` enables the older JSON mode, which
+              ensures the message the model generates is valid JSON. Using `json_schema` is
+              preferred for models that support it.
+
           safety_model: The name of the moderation model used to validate tokens. Choose from the
               available moderation models found
               [here](https://docs.together.ai/docs/inference-models#moderation-models).
@@ -653,6 +677,14 @@ class AsyncCompletionsResource(AsyncAPIResource):
 
           response_format: An object specifying the format that the model must output.
 
+              Setting to `{ "type": "json_schema", "json_schema": {...} }` enables Structured
+              Outputs which ensures the model will match your supplied JSON schema. Learn more
+              in the [Structured Outputs guide](https://docs.together.ai/docs/json-mode).
+
+              Setting to `{ "type": "json_object" }` enables the older JSON mode, which
+              ensures the message the model generates is valid JSON. Using `json_schema` is
+              preferred for models that support it.
+
           safety_model: The name of the moderation model used to validate tokens. Choose from the
               available moderation models found
               [here](https://docs.together.ai/docs/inference-models#moderation-models).
@@ -794,6 +826,14 @@ class AsyncCompletionsResource(AsyncAPIResource):
 
           response_format: An object specifying the format that the model must output.
 
+              Setting to `{ "type": "json_schema", "json_schema": {...} }` enables Structured
+              Outputs which ensures the model will match your supplied JSON schema. Learn more
+              in the [Structured Outputs guide](https://docs.together.ai/docs/json-mode).
+
+              Setting to `{ "type": "json_object" }` enables the older JSON mode, which
+              ensures the message the model generates is valid JSON. Using `json_schema` is
+              preferred for models that support it.
+
           safety_model: The name of the moderation model used to validate tokens. Choose from the
               available moderation models found
               [here](https://docs.together.ai/docs/inference-models#moderation-models).
@@ -931,6 +971,14 @@ class AsyncCompletionsResource(AsyncAPIResource):
 
           response_format: An object specifying the format that the model must output.
 
+              Setting to `{ "type": "json_schema", "json_schema": {...} }` enables Structured
+              Outputs which ensures the model will match your supplied JSON schema. Learn more
+              in the [Structured Outputs guide](https://docs.together.ai/docs/json-mode).
+
+              Setting to `{ "type": "json_object" }` enables the older JSON mode, which
+              ensures the message the model generates is valid JSON. Using `json_schema` is
+              preferred for models that support it.
+
           safety_model: The name of the moderation model used to validate tokens. Choose from the
               available moderation models found
               [here](https://docs.together.ai/docs/inference-models#moderation-models).