Add support for specifying seed (#139)

* Add support for specifying seed

* update test models

* fix model name

* update model list

* bump version to 1.2.12

Author: orangetin

pyproject.toml

@@ -12,7 +12,7 @@ build-backend = "poetry.masonry.api"
 
 [tool.poetry]
 name = "together"
-version = "1.2.11"
+version = "1.2.12"
 authors = [
     "Together AI <[email protected]>"
 ]

src/together/resources/chat/completions.py

@@ -32,6 +32,7 @@ def create(
         frequency_penalty: float | None = None,
         min_p: float | None = None,
         logit_bias: Dict[str, float] | None = None,
+        seed: int | None = None,
         stream: bool = False,
         logprobs: int | None = None,
         echo: bool | None = None,
@@ -79,6 +80,7 @@ def create(
             logit_bias (Dict[str, float], optional): A dictionary of tokens and their bias values that modify the
                 likelihood of specific tokens being sampled. Bias values must be in the range [-100, 100].
                 Defaults to None.
+            seed (int, optional): A seed value to use for reproducibility.
             stream (bool, optional): Flag indicating whether to stream the generated completions.
                 Defaults to False.
             logprobs (int, optional): Number of top-k logprobs to return
@@ -124,6 +126,7 @@ def create(
             frequency_penalty=frequency_penalty,
             min_p=min_p,
             logit_bias=logit_bias,
+            seed=seed,
             stream=stream,
             logprobs=logprobs,
             echo=echo,
@@ -171,6 +174,7 @@ async def create(
         frequency_penalty: float | None = None,
         min_p: float | None = None,
         logit_bias: Dict[str, float] | None = None,
+        seed: int | None = None,
         stream: bool = False,
         logprobs: int | None = None,
         echo: bool | None = None,
@@ -218,6 +222,7 @@ async def create(
             logit_bias (Dict[str, float], optional): A dictionary of tokens and their bias values that modify the
                 likelihood of specific tokens being sampled. Bias values must be in the range [-100, 100].
                 Defaults to None.
+            seed (int, optional): A seed value to use for reproducibility.
             stream (bool, optional): Flag indicating whether to stream the generated completions.
                 Defaults to False.
             logprobs (int, optional): Number of top-k logprobs to return
@@ -263,6 +268,7 @@ async def create(
             frequency_penalty=frequency_penalty,
             min_p=min_p,
             logit_bias=logit_bias,
+            seed=seed,
             stream=stream,
             logprobs=logprobs,
             echo=echo,

src/together/resources/completions.py

@@ -32,6 +32,7 @@ def create(
         frequency_penalty: float | None = None,
         min_p: float | None = None,
         logit_bias: Dict[str, float] | None = None,
+        seed: int | None = None,
         stream: bool = False,
         logprobs: int | None = None,
         echo: bool | None = None,
@@ -75,6 +76,7 @@ def create(
             logit_bias (Dict[str, float], optional): A dictionary of tokens and their bias values that modify the
                 likelihood of specific tokens being sampled. Bias values must be in the range [-100, 100].
                 Defaults to None.
+            seed (int, optional): Seed value for reproducibility.
             stream (bool, optional): Flag indicating whether to stream the generated completions.
                 Defaults to False.
             logprobs (int, optional): Number of top-k logprobs to return
@@ -107,6 +109,7 @@ def create(
             repetition_penalty=repetition_penalty,
             presence_penalty=presence_penalty,
             frequency_penalty=frequency_penalty,
+            seed=seed,
             min_p=min_p,
             logit_bias=logit_bias,
             stream=stream,
@@ -153,6 +156,7 @@ async def create(
         frequency_penalty: float | None = None,
         min_p: float | None = None,
         logit_bias: Dict[str, float] | None = None,
+        seed: int | None = None,
         stream: bool = False,
         logprobs: int | None = None,
         echo: bool | None = None,
@@ -196,6 +200,7 @@ async def create(
             logit_bias (Dict[str, float], optional): A dictionary of tokens and their bias values that modify the
                 likelihood of specific tokens being sampled. Bias values must be in the range [-100, 100].
                 Defaults to None.
+            seed (int, optional): Seed value for reproducibility.
             stream (bool, optional): Flag indicating whether to stream the generated completions.
                 Defaults to False.
             logprobs (int, optional): Number of top-k logprobs to return
@@ -230,6 +235,7 @@ async def create(
             frequency_penalty=frequency_penalty,
             min_p=min_p,
             logit_bias=logit_bias,
+            seed=seed,
             stream=stream,
             logprobs=logprobs,
             echo=echo,

src/together/types/chat_completions.py

@@ -96,6 +96,7 @@ class ChatCompletionRequest(BaseModel):
     frequency_penalty: float | None = None
     min_p: float | None = None
     logit_bias: Dict[str, float] | None = None
+    seed: int | None = None
     # stream SSE token chunks
     stream: bool = False
     # return logprobs
@@ -126,6 +127,7 @@ def verify_parameters(self) -> Self:
 class ChatCompletionChoicesData(BaseModel):
     index: int | None = None
     logprobs: LogprobsPart | None = None
+    seed: int | None = None
     finish_reason: FinishReason | None = None
     message: ChatCompletionMessage | None = None
 
@@ -150,6 +152,7 @@ class ChatCompletionResponse(BaseModel):
 class ChatCompletionChoicesChunk(BaseModel):
     index: int | None = None
     logprobs: float | None = None
+    seed: int | None = None
     finish_reason: FinishReason | None = None
     delta: DeltaContent | None = None
 

src/together/types/completions.py

@@ -35,6 +35,7 @@ class CompletionRequest(BaseModel):
     frequency_penalty: float | None = None
     min_p: float | None = None
     logit_bias: Dict[str, float] | None = None
+    seed: int | None = None
     # stream SSE token chunks
     stream: bool = False
     # return logprobs
@@ -61,13 +62,15 @@ def verify_parameters(self) -> Self:
 class CompletionChoicesData(BaseModel):
     index: int
     logprobs: LogprobsPart | None = None
-    finish_reason: FinishReason | None = None
+    seed: int | None = None
+    finish_reason: FinishReason
     text: str
 
 
 class CompletionChoicesChunk(BaseModel):
     index: int | None = None
     logprobs: float | None = None
+    seed: int | None = None
     finish_reason: FinishReason | None = None
     delta: DeltaContent | None = None
 

tests/integration/resources/test_completion.py

@@ -530,3 +530,27 @@ def test_logit_bias(
         )
 
         assert isinstance(response, CompletionResponse)
+
+    @pytest.mark.parametrize(
+        "model,prompt",
+        product(
+            completion_test_model_list,
+            completion_prompt_list,
+        ),
+    )
+    def test_seed(
+        self,
+        model,
+        prompt,
+        sync_together_client,
+    ):
+        response = sync_together_client.completions.create(
+            prompt=prompt,
+            model=model,
+            stop=STOP,
+            max_tokens=1,
+            seed=4242,
+        )
+
+        assert isinstance(response, CompletionResponse)
+        assert response.choices[0].seed == 4242