together 1.5.5__tar.gz → 1.5.7__tar.gz

{together-1.5.5 → together-1.5.7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.3
 Name: together
-Version: 1.5.5
+Version: 1.5.7
 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-1.5.5 → together-1.5.7}/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-1.5.5 → together-1.5.7}/pyproject.toml

@@ -12,7 +12,7 @@ build-backend = "poetry.masonry.api"
 
 [tool.poetry]
 name = "together"
-version = "1.5.5"
+version = "1.5.7"
 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-1.5.5 → together-1.5.7}/src/together/cli/api/finetune.py

@@ -258,10 +258,13 @@ def create(
             raise click.BadParameter(
                 f"LoRA fine-tuning is not supported for the model `{model}`"
             )
-
+        if training_method == "dpo":
+            default_batch_size = model_limits.lora_training.max_batch_size_dpo
+        else:
+            default_batch_size = model_limits.lora_training.max_batch_size
         default_values = {
             "lora_r": model_limits.lora_training.max_rank,
-            "batch_size": model_limits.lora_training.max_batch_size,
+            "batch_size": default_batch_size,
             "learning_rate": 1e-3,
         }
 
@@ -288,7 +291,12 @@ def create(
 
         batch_size_source = ctx.get_parameter_source("batch_size")  # type: ignore[attr-defined]
         if batch_size_source == ParameterSource.DEFAULT:
-            training_args["batch_size"] = model_limits.full_training.max_batch_size
+            if training_method == "dpo":
+                training_args["batch_size"] = (
+                    model_limits.full_training.max_batch_size_dpo
+                )
+            else:
+                training_args["batch_size"] = model_limits.full_training.max_batch_size
 
     if n_evals <= 0 and validation_file:
         log_warn(

{together-1.5.5 → together-1.5.7}/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-1.5.5 → together-1.5.7}/src/together/resources/finetune.py

@@ -102,6 +102,7 @@ def create_finetune_request(
 
     training_type: TrainingType = FullTrainingType()
     max_batch_size: int = 0
+    max_batch_size_dpo: int = 0
     min_batch_size: int = 0
     if lora:
         if model_limits.lora_training is None:
@@ -119,7 +120,7 @@ def create_finetune_request(
 
         max_batch_size = model_limits.lora_training.max_batch_size
         min_batch_size = model_limits.lora_training.min_batch_size
-
+        max_batch_size_dpo = model_limits.lora_training.max_batch_size_dpo
     else:
         if model_limits.full_training is None:
             raise ValueError(
@@ -128,13 +129,24 @@ def create_finetune_request(
 
         max_batch_size = model_limits.full_training.max_batch_size
         min_batch_size = model_limits.full_training.min_batch_size
+        max_batch_size_dpo = model_limits.full_training.max_batch_size_dpo
 
-    batch_size = batch_size if batch_size != "max" else max_batch_size
+    if batch_size == "max":
+        if training_method == "dpo":
+            batch_size = max_batch_size_dpo
+        else:
+            batch_size = max_batch_size
 
-    if batch_size > max_batch_size:
-        raise ValueError(
-            f"Requested batch size of {batch_size} is higher that the maximum allowed value of {max_batch_size}."
-        )
+    if training_method == "sft":
+        if batch_size > max_batch_size:
+            raise ValueError(
+                f"Requested batch size of {batch_size} is higher that the maximum allowed value of {max_batch_size}."
+            )
+    elif training_method == "dpo":
+        if batch_size > max_batch_size_dpo:
+            raise ValueError(
+                f"Requested batch size of {batch_size} is higher that the maximum allowed value of {max_batch_size_dpo}."
+            )
 
     if batch_size < min_batch_size:
         raise ValueError(
@@ -600,7 +612,7 @@ 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

{together-1.5.5 → together-1.5.7}/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-1.5.5 → together-1.5.7}/src/together/types/finetune.py

@@ -1,7 +1,7 @@
 from __future__ import annotations
 
 from enum import Enum
-from typing import List, Literal
+from typing import List, Literal, Any
 
 from pydantic import StrictBool, Field, field_validator
 
@@ -329,8 +329,16 @@ class FinetuneDownloadResult(BaseModel):
 
 class FinetuneFullTrainingLimits(BaseModel):
     max_batch_size: int
+    max_batch_size_dpo: int = -1
     min_batch_size: int
 
+    def __init__(self, **data: Any) -> None:
+        super().__init__(**data)
+        if self.max_batch_size_dpo == -1:
+            half_max = self.max_batch_size // 2
+            rounded_half_max = (half_max // 8) * 8
+            self.max_batch_size_dpo = max(self.min_batch_size, rounded_half_max)
+
 
 class FinetuneLoraTrainingLimits(FinetuneFullTrainingLimits):
     max_rank: int