together 1.5.4__tar.gz → 1.5.6__tar.gz

{together-1.5.4 → together-1.5.6}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: together
-Version: 1.5.4
+Version: 1.5.6
 Summary: Python client for Together's Cloud Platform!
 License: Apache-2.0
 Author: Together AI
@@ -23,7 +23,7 @@ Requires-Dist: pillow (>=11.1.0,<12.0.0)
 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: rich (>=13.8.1,<15.0.0)
 Requires-Dist: tabulate (>=0.9.0,<0.10.0)
 Requires-Dist: tqdm (>=4.66.2,<5.0.0)
 Requires-Dist: typer (>=0.9,<0.16)
@@ -220,6 +220,30 @@ async def async_chat_completion(messages):
 asyncio.run(async_chat_completion(messages))
 ```
 
+#### Fetching logprobs
+
+Logprobs are logarithms of token-level generation probabilities that indicate the likelihood of the generated token based on the previous tokens in the context. Logprobs allow us to estimate the model's confidence in its outputs, which can be used to decide how to optimally consume the model's output (e.g. rejecting low confidence outputs, retrying or ensembling model outputs etc).
+
+```python
+from together import Together
+
+client = Together()
+
+response = client.chat.completions.create(
+    model="mistralai/Mixtral-8x7B-Instruct-v0.1",
+    messages=[{"role": "user", "content": "tell me about new york"}],
+    logprobs=1
+)
+
+response_lobprobs = response.choices[0].logprobs
+
+print(dict(zip(response_lobprobs.tokens, response_lobprobs.token_logprobs)))
+# {'New': -2.384e-07, ' York': 0.0, ',': 0.0, ' also': -0.20703125, ' known': -0.20214844, ' as': -8.34465e-07, ... }
+```
+
+More details about using logprobs in Together's API can be found [here](https://docs.together.ai/docs/logprobs).
+
+
 ### Completions
 
 Completions are for code and language models shown [here](https://docs.together.ai/docs/inference-models). Below, a code model example is shown.

{together-1.5.4 → together-1.5.6}/README.md

@@ -186,6 +186,30 @@ async def async_chat_completion(messages):
 asyncio.run(async_chat_completion(messages))
 ```
 
+#### Fetching logprobs
+
+Logprobs are logarithms of token-level generation probabilities that indicate the likelihood of the generated token based on the previous tokens in the context. Logprobs allow us to estimate the model's confidence in its outputs, which can be used to decide how to optimally consume the model's output (e.g. rejecting low confidence outputs, retrying or ensembling model outputs etc).
+
+```python
+from together import Together
+
+client = Together()
+
+response = client.chat.completions.create(
+    model="mistralai/Mixtral-8x7B-Instruct-v0.1",
+    messages=[{"role": "user", "content": "tell me about new york"}],
+    logprobs=1
+)
+
+response_lobprobs = response.choices[0].logprobs
+
+print(dict(zip(response_lobprobs.tokens, response_lobprobs.token_logprobs)))
+# {'New': -2.384e-07, ' York': 0.0, ',': 0.0, ' also': -0.20703125, ' known': -0.20214844, ' as': -8.34465e-07, ... }
+```
+
+More details about using logprobs in Together's API can be found [here](https://docs.together.ai/docs/logprobs).
+
+
 ### Completions
 
 Completions are for code and language models shown [here](https://docs.together.ai/docs/inference-models). Below, a code model example is shown.

{together-1.5.4 → together-1.5.6}/pyproject.toml

@@ -12,7 +12,7 @@ build-backend = "poetry.masonry.api"
 
 [tool.poetry]
 name = "together"
-version = "1.5.4"
+version = "1.5.6"
 authors = ["Together AI <support@together.ai>"]
 description = "Python client for Together's Cloud Platform!"
 readme = "README.md"
@@ -29,7 +29,7 @@ homepage = "https://github.com/togethercomputer/together-python"
 python = "^3.10"
 typer = ">=0.9,<0.16"
 requests = "^2.31.0"
-rich = "^13.8.1"
+rich = ">=13.8.1,<15.0.0"
 tqdm = "^4.66.2"
 tabulate = "^0.9.0"
 pydantic = "^2.6.3"
@@ -49,10 +49,10 @@ optional = true
 
 [tool.poetry.group.quality.dependencies]
 black = ">=23.1,<26.0"
-ruff = ">=0.3.2,<0.10.0"
+ruff = ">=0.3.2,<0.12.0"
 types-tqdm = "^4.65.0.0"
 types-tabulate = "^0.9.0.3"
-pre-commit = "4.1.0"
+pre-commit = "4.2.0"
 types-requests = "^2.31.0.20240218"
 pyarrow-stubs = ">=10.0.1.7,<20240831.0.0.0"
 mypy = "^1.9.0"

{together-1.5.4 → together-1.5.6}/src/together/filemanager.py

@@ -378,7 +378,8 @@ class UploadManager:
 
             if not callback_response.status_code == 200:
                 raise APIError(
-                    f"Error code: {callback_response.status_code} - Failed to process uploaded file"
+                    f"Error during file upload: {callback_response.content.decode()}, headers: {callback_response.headers}",
+                    http_status=callback_response.status_code,
                 )
 
             response = self.callback(f"{url}/{file_id}/preprocess")

{together-1.5.4 → together-1.5.6}/src/together/resources/code_interpreter.py

@@ -1,11 +1,12 @@
 from __future__ import annotations
 
-from typing import Dict, Literal, Optional
+from typing import Any, Dict, List, Literal, Optional
+from pydantic import ValidationError
 
 from together.abstract import api_requestor
 from together.together_response import TogetherResponse
 from together.types import TogetherClient, TogetherRequest
-from together.types.code_interpreter import ExecuteResponse
+from together.types.code_interpreter import ExecuteResponse, FileInput
 
 
 class CodeInterpreter:
@@ -19,22 +20,28 @@ class CodeInterpreter:
         code: str,
         language: Literal["python"],
         session_id: Optional[str] = None,
+        files: Optional[List[Dict[str, Any]]] = None,
     ) -> ExecuteResponse:
-        """Execute a code snippet.
+        """Execute a code snippet, optionally with files.
 
         Args:
             code (str): Code snippet to execute
             language (str): Programming language for the code to execute. Currently only supports Python.
             session_id (str, optional): Identifier of the current session. Used to make follow-up calls.
+            files (List[Dict], optional): Files to upload to the session before executing the code.
 
         Returns:
             ExecuteResponse: Object containing execution results and outputs
+
+        Raises:
+            ValidationError: If any dictionary in the `files` list does not conform to the
+                required structure or types.
         """
         requestor = api_requestor.APIRequestor(
             client=self._client,
         )
 
-        data: Dict[str, str] = {
+        data: Dict[str, Any] = {
             "code": code,
             "language": language,
         }
@@ -42,6 +49,23 @@ class CodeInterpreter:
         if session_id is not None:
             data["session_id"] = session_id
 
+        if files is not None:
+            serialized_files = []
+            try:
+                for file_dict in files:
+                    # Validate the dictionary by creating a FileInput instance
+                    validated_file = FileInput(**file_dict)
+                    # Serialize the validated model back to a dict for the API call
+                    serialized_files.append(validated_file.model_dump())
+            except ValidationError as e:
+                raise ValueError(f"Invalid file input format: {e}") from e
+            except TypeError as e:
+                raise ValueError(
+                    f"Invalid file input: Each item in 'files' must be a dictionary. Error: {e}"
+                ) from e
+
+            data["files"] = serialized_files
+
         # Use absolute URL to bypass the /v1 prefix
         response, _, _ = requestor.request(
             options=TogetherRequest(

{together-1.5.4 → together-1.5.6}/src/together/resources/finetune.py

@@ -2,7 +2,7 @@ from __future__ import annotations
 
 import re
 from pathlib import Path
-from typing import Literal, List
+from typing import List, Literal
 
 from rich import print as rprint
 
@@ -10,37 +10,38 @@ from together.abstract import api_requestor
 from together.filemanager import DownloadManager
 from together.together_response import TogetherResponse
 from together.types import (
+    CosineLRScheduler,
+    CosineLRSchedulerArgs,
+    FinetuneCheckpoint,
     FinetuneDownloadResult,
     FinetuneList,
     FinetuneListEvents,
+    FinetuneLRScheduler,
     FinetuneRequest,
     FinetuneResponse,
     FinetuneTrainingLimits,
     FullTrainingType,
+    LinearLRScheduler,
+    LinearLRSchedulerArgs,
     LoRATrainingType,
     TogetherClient,
     TogetherRequest,
-    TrainingType,
-    FinetuneLRScheduler,
-    FinetuneLinearLRScheduler,
-    FinetuneCosineLRScheduler,
-    FinetuneLinearLRSchedulerArgs,
-    FinetuneCosineLRSchedulerArgs,
     TrainingMethodDPO,
     TrainingMethodSFT,
-    FinetuneCheckpoint,
+    TrainingType,
 )
 from together.types.finetune import (
     DownloadCheckpointType,
-    FinetuneEventType,
     FinetuneEvent,
+    FinetuneEventType,
 )
 from together.utils import (
+    get_event_step,
     log_warn_once,
     normalize_key,
-    get_event_step,
 )
 
+
 _FT_JOB_WITH_STEP_REGEX = r"^ft-[\dabcdef-]+:\d+$"
 
 
@@ -50,7 +51,7 @@ AVAILABLE_TRAINING_METHODS = {
 }
 
 
-def createFinetuneRequest(
+def create_finetune_request(
     model_limits: FinetuneTrainingLimits,
     training_file: str,
     model: str | None = None,
@@ -63,7 +64,7 @@ def createFinetuneRequest(
     lr_scheduler_type: Literal["linear", "cosine"] = "linear",
     min_lr_ratio: float = 0.0,
     scheduler_num_cycles: float = 0.5,
-    warmup_ratio: float = 0.0,
+    warmup_ratio: float | None = None,
     max_grad_norm: float = 1.0,
     weight_decay: float = 0.0,
     lora: bool = False,
@@ -81,7 +82,6 @@ def createFinetuneRequest(
     dpo_beta: float | None = None,
     from_checkpoint: str | None = None,
 ) -> FinetuneRequest:
-
     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"
@@ -90,6 +90,8 @@ def createFinetuneRequest(
     if model is None and from_checkpoint is None:
         raise ValueError("You must specify either a model or a checkpoint")
 
+    model_or_checkpoint = model or from_checkpoint
+
     if batch_size == "max":
         log_warn_once(
             "Starting from together>=1.3.0, "
@@ -103,7 +105,9 @@ def createFinetuneRequest(
     min_batch_size: int = 0
     if lora:
         if model_limits.lora_training is None:
-            raise ValueError("LoRA adapters are not supported for the selected model.")
+            raise ValueError(
+                f"LoRA adapters are not supported for the selected model ({model_or_checkpoint})."
+            )
         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(
@@ -118,7 +122,9 @@ def createFinetuneRequest(
 
     else:
         if model_limits.full_training is None:
-            raise ValueError("Full training is not supported for the selected model.")
+            raise ValueError(
+                f"Full training is not supported for the selected model ({model_or_checkpoint})."
+            )
 
         max_batch_size = model_limits.full_training.max_batch_size
         min_batch_size = model_limits.full_training.min_batch_size
@@ -127,46 +133,50 @@ def createFinetuneRequest(
 
     if batch_size > max_batch_size:
         raise ValueError(
-            "Requested batch size is higher that the maximum allowed value."
+            f"Requested batch size of {batch_size} is higher that the maximum allowed value of {max_batch_size}."
         )
 
     if batch_size < min_batch_size:
         raise ValueError(
-            "Requested batch size is lower that the minimum allowed value."
+            f"Requested batch size of {batch_size} is lower that the minimum allowed value of {min_batch_size}."
         )
 
     if warmup_ratio > 1 or warmup_ratio < 0:
-        raise ValueError("Warmup ratio should be between 0 and 1")
+        raise ValueError(f"Warmup ratio should be between 0 and 1 (got {warmup_ratio})")
 
     if min_lr_ratio is not None and (min_lr_ratio > 1 or min_lr_ratio < 0):
-        raise ValueError("Min learning rate ratio should be between 0 and 1")
+        raise ValueError(
+            f"Min learning rate ratio should be between 0 and 1 (got {min_lr_ratio})"
+        )
 
     if max_grad_norm < 0:
-        raise ValueError("Max gradient norm should be non-negative")
+        raise ValueError(
+            f"Max gradient norm should be non-negative (got {max_grad_norm})"
+        )
 
     if weight_decay is not None and (weight_decay < 0):
-        raise ValueError("Weight decay should be non-negative")
+        raise ValueError(f"Weight decay should be non-negative (got {weight_decay})")
 
     if training_method not in AVAILABLE_TRAINING_METHODS:
         raise ValueError(
             f"training_method must be one of {', '.join(AVAILABLE_TRAINING_METHODS)}"
         )
 
-    # Default to generic lr scheduler
-    lrScheduler: FinetuneLRScheduler = FinetuneLRScheduler(lr_scheduler_type="linear")
-
+    lr_scheduler: FinetuneLRScheduler
     if lr_scheduler_type == "cosine":
         if scheduler_num_cycles <= 0.0:
-            raise ValueError("Number of cycles should be greater than 0")
+            raise ValueError(
+                f"Number of cycles should be greater than 0 (got {scheduler_num_cycles})"
+            )
 
-        lrScheduler = FinetuneCosineLRScheduler(
-            lr_scheduler_args=FinetuneCosineLRSchedulerArgs(
+        lr_scheduler = CosineLRScheduler(
+            lr_scheduler_args=CosineLRSchedulerArgs(
                 min_lr_ratio=min_lr_ratio, num_cycles=scheduler_num_cycles
             ),
         )
     else:
-        lrScheduler = FinetuneLinearLRScheduler(
-            lr_scheduler_args=FinetuneLinearLRSchedulerArgs(min_lr_ratio=min_lr_ratio),
+        lr_scheduler = LinearLRScheduler(
+            lr_scheduler_args=LinearLRSchedulerArgs(min_lr_ratio=min_lr_ratio),
         )
 
     training_method_cls: TrainingMethodSFT | TrainingMethodDPO = TrainingMethodSFT()
@@ -182,7 +192,7 @@ def createFinetuneRequest(
         n_checkpoints=n_checkpoints,
         batch_size=batch_size,
         learning_rate=learning_rate,
-        lr_scheduler=lrScheduler,
+        lr_scheduler=lr_scheduler,
         warmup_ratio=warmup_ratio,
         max_grad_norm=max_grad_norm,
         weight_decay=weight_decay,
@@ -374,7 +384,7 @@ class FineTuning:
                 pass
             model_limits = self.get_model_limits(model=model_name)
 
-        finetune_request = createFinetuneRequest(
+        finetune_request = create_finetune_request(
             model_limits=model_limits,
             training_file=training_file,
             model=model,
@@ -590,7 +600,7 @@ class FineTuning:
                 raise ValueError(
                     "Only DEFAULT checkpoint type is allowed for FullTrainingType"
                 )
-            url += "&checkpoint=modelOutputPath"
+            url += "&checkpoint=model_output_path"
         elif isinstance(ft_job.training_type, LoRATrainingType):
             if checkpoint_type == DownloadCheckpointType.DEFAULT:
                 checkpoint_type = DownloadCheckpointType.MERGED
@@ -762,7 +772,7 @@ class AsyncFineTuning:
                 pass
             model_limits = await self.get_model_limits(model=model_name)
 
-        finetune_request = createFinetuneRequest(
+        finetune_request = create_finetune_request(
             model_limits=model_limits,
             training_file=training_file,
             model=model,

{together-1.5.4 → together-1.5.6}/src/together/types/__init__.py

@@ -34,11 +34,11 @@ from together.types.finetune import (
     TrainingMethodDPO,
     TrainingMethodSFT,
     FinetuneCheckpoint,
-    FinetuneCosineLRScheduler,
-    FinetuneCosineLRSchedulerArgs,
+    CosineLRScheduler,
+    CosineLRSchedulerArgs,
     FinetuneDownloadResult,
-    FinetuneLinearLRScheduler,
-    FinetuneLinearLRSchedulerArgs,
+    LinearLRScheduler,
+    LinearLRSchedulerArgs,
     FinetuneLRScheduler,
     FinetuneList,
     FinetuneListEvents,
@@ -72,10 +72,10 @@ __all__ = [
     "FinetuneListEvents",
     "FinetuneDownloadResult",
     "FinetuneLRScheduler",
-    "FinetuneLinearLRScheduler",
-    "FinetuneLinearLRSchedulerArgs",
-    "FinetuneCosineLRScheduler",
-    "FinetuneCosineLRSchedulerArgs",
+    "LinearLRScheduler",
+    "LinearLRSchedulerArgs",
+    "CosineLRScheduler",
+    "CosineLRSchedulerArgs",
     "FileRequest",
     "FileResponse",
     "FileList",

{together-1.5.4 → together-1.5.6}/src/together/types/code_interpreter.py

@@ -7,6 +7,16 @@ from pydantic import Field
 from together.types.endpoints import TogetherJSONModel
 
 
+class FileInput(TogetherJSONModel):
+    """File input to be uploaded to the code interpreter session."""
+
+    name: str = Field(description="The name of the file.")
+    encoding: Literal["string", "base64"] = Field(
+        description="Encoding of the file content. Use 'string' for text files and 'base64' for binary files."
+    )
+    content: str = Field(description="The content of the file, encoded as specified.")
+
+
 class InterpreterOutput(TogetherJSONModel):
     """Base class for interpreter output types."""
 
@@ -40,6 +50,7 @@ class ExecuteResponse(TogetherJSONModel):
 
 
 __all__ = [
+    "FileInput",
     "InterpreterOutput",
     "ExecuteResponseData",
     "ExecuteResponse",

{together-1.5.4 → together-1.5.6}/src/together/types/finetune.py

@@ -1,9 +1,9 @@
 from __future__ import annotations
 
 from enum import Enum
-from typing import List, Literal, Union
+from typing import List, Literal
 
-from pydantic import StrictBool, Field, validator, field_validator, ValidationInfo
+from pydantic import StrictBool, Field, field_validator
 
 from together.types.abstract import BaseModel
 from together.types.common import (
@@ -176,7 +176,7 @@ class FinetuneRequest(BaseModel):
     # training learning rate
     learning_rate: float
     # learning rate scheduler type and args
-    lr_scheduler: FinetuneLinearLRScheduler | FinetuneCosineLRScheduler | None = None
+    lr_scheduler: LinearLRScheduler | CosineLRScheduler | None = None
     # learning rate warmup ratio
     warmup_ratio: float
     # max gradient norm
@@ -239,7 +239,7 @@ class FinetuneResponse(BaseModel):
     # training learning rate
     learning_rate: float | None = None
     # learning rate scheduler type and args
-    lr_scheduler: FinetuneLinearLRScheduler | FinetuneCosineLRScheduler | None = None
+    lr_scheduler: LinearLRScheduler | CosineLRScheduler | EmptyLRScheduler | None = None
     # learning rate warmup ratio
     warmup_ratio: float | None = None
     # max gradient norm
@@ -345,11 +345,11 @@ class FinetuneTrainingLimits(BaseModel):
     lora_training: FinetuneLoraTrainingLimits | None = None
 
 
-class FinetuneLinearLRSchedulerArgs(BaseModel):
+class LinearLRSchedulerArgs(BaseModel):
     min_lr_ratio: float | None = 0.0
 
 
-class FinetuneCosineLRSchedulerArgs(BaseModel):
+class CosineLRSchedulerArgs(BaseModel):
     min_lr_ratio: float | None = 0.0
     num_cycles: float | None = 0.5
 
@@ -358,14 +358,20 @@ class FinetuneLRScheduler(BaseModel):
     lr_scheduler_type: str
 
 
-class FinetuneLinearLRScheduler(FinetuneLRScheduler):
+class LinearLRScheduler(FinetuneLRScheduler):
     lr_scheduler_type: Literal["linear"] = "linear"
-    lr_scheduler: FinetuneLinearLRSchedulerArgs | None = None
+    lr_scheduler_args: LinearLRSchedulerArgs | None = None
 
 
-class FinetuneCosineLRScheduler(FinetuneLRScheduler):
+class CosineLRScheduler(FinetuneLRScheduler):
     lr_scheduler_type: Literal["cosine"] = "cosine"
-    lr_scheduler: FinetuneCosineLRSchedulerArgs | None = None
+    lr_scheduler_args: CosineLRSchedulerArgs | None = None
+
+
+# placeholder for old fine-tuning jobs with no lr_scheduler_type specified
+class EmptyLRScheduler(FinetuneLRScheduler):
+    lr_scheduler_type: Literal[""]
+    lr_scheduler_args: None = None
 
 
 class FinetuneCheckpoint(BaseModel):