Add Openai compatible Completion class (#60)

* Add openai compatible Completion class

* Add pydantic dependency

Author: orangetin

poetry.lock

@@ -26,6 +26,7 @@ requests = "^2.31.0"
 tqdm = "^4.66.1"
 sseclient-py = "^1.7.2"
 tabulate = "^0.9.0"
+pydantic = "^2.5.0"
 
 [tool.poetry.group.quality]
 optional = true

pyproject.toml

@@ -32,7 +32,7 @@
 
 min_samples = 100
 
-from .complete import Complete
+from .complete import Complete, Completion
 from .embeddings import Embeddings
 from .error import *
 from .files import Files
@@ -54,6 +54,7 @@
     "default_embedding_model",
     "Models",
     "Complete",
+    "Completion",
     "Files",
     "Finetune",
     "Image",

src/together/__init__.py

@@ -114,6 +114,7 @@ def do_say(self, arg: str) -> None:
                 top_k=self.args.top_k,
                 repetition_penalty=self.args.repetition_penalty,
             ):
+                assert isinstance(token, str)
                 print(token, end="", flush=True)
                 output += token
         except together.AuthenticationError:

src/together/commands/chat.py

@@ -146,6 +146,7 @@ def _run_complete(args: argparse.Namespace) -> None:
         except together.AuthenticationError:
             logger.critical(together.MISSING_API_KEY_MESSAGE)
             exit(0)
+        assert isinstance(response, dict)
         no_streamer(args, response)
     else:
         try:

src/together/commands/complete.py

@@ -1,7 +1,8 @@
 import json
-from typing import Any, Dict, Iterator, List, Optional
+from typing import Any, Dict, Iterator, List, Optional, Union
 
 import together
+from together.types import TogetherResponse
 from together.utils import create_post_request, get_logger, sse_client
 
 
@@ -21,7 +22,9 @@ def create(
         top_k: Optional[int] = 50,
         repetition_penalty: Optional[float] = None,
         logprobs: Optional[int] = None,
-    ) -> Dict[str, Any]:
+        api_key: Optional[str] = None,
+        cast: bool = False,
+    ) -> Union[Dict[str, Any], TogetherResponse]:
         if model == "":
             model = together.default_text_model
 
@@ -39,14 +42,18 @@ def create(
 
         # send request
         response = create_post_request(
-            url=together.api_base_complete, json=parameter_payload
+            url=together.api_base_complete, json=parameter_payload, api_key=api_key
         )
 
         try:
             response_json = dict(response.json())
 
         except Exception as e:
             raise together.JSONError(e, http_status=response.status_code)
+
+        if cast:
+            return TogetherResponse(**response_json)
+
         return response_json
 
     @classmethod
@@ -55,13 +62,15 @@ def create_streaming(
         prompt: str,
         model: Optional[str] = "",
         max_tokens: Optional[int] = 128,
-        stop: Optional[str] = None,
+        stop: Optional[List[str]] = None,
         temperature: Optional[float] = 0.7,
         top_p: Optional[float] = 0.7,
         top_k: Optional[int] = 50,
         repetition_penalty: Optional[float] = None,
         raw: Optional[bool] = False,
-    ) -> Iterator[str]:
+        api_key: Optional[str] = None,
+        cast: Optional[bool] = False,
+    ) -> Union[Iterator[str], Iterator[TogetherResponse]]:
         """
         Prints streaming responses and returns the completed text.
         """
@@ -83,19 +92,25 @@ def create_streaming(
 
         # send request
         response = create_post_request(
-            url=together.api_base_complete, json=parameter_payload, stream=True
+            url=together.api_base_complete,
+            json=parameter_payload,
+            api_key=api_key,
+            stream=True,
         )
 
         output = ""
         client = sse_client(response)
         for event in client.events():
-            if raw:
+            if cast:
+                if event.data != "[DONE]":
+                    yield TogetherResponse(**json.loads(event.data))
+            elif raw:
                 yield str(event.data)
             elif event.data != "[DONE]":
                 json_response = dict(json.loads(event.data))
                 if "error" in json_response.keys():
                     raise together.ResponseError(
-                        json_response["error"]["error"],
+                        json_response["error"],
                         request_id=json_response["error"]["request_id"],
                     )
                 elif "choices" in json_response.keys():
@@ -106,3 +121,50 @@ def create_streaming(
                     raise together.ResponseError(
                         f"Unknown error occured. Received unhandled response: {event.data}"
                     )
+
+
+class Completion:
+    @classmethod
+    def create(
+        self,
+        prompt: str,
+        model: Optional[str] = "",
+        max_tokens: Optional[int] = 128,
+        stop: Optional[List[str]] = [],
+        temperature: Optional[float] = 0.7,
+        top_p: Optional[float] = 0.7,
+        top_k: Optional[int] = 50,
+        repetition_penalty: Optional[float] = None,
+        logprobs: Optional[int] = None,
+        api_key: Optional[str] = None,
+        stream: bool = False,
+    ) -> Union[
+        TogetherResponse, Iterator[TogetherResponse], Iterator[str], Dict[str, Any]
+    ]:
+        if stream:
+            return Complete.create_streaming(
+                prompt=prompt,
+                model=model,
+                max_tokens=max_tokens,
+                stop=stop,
+                temperature=temperature,
+                top_p=top_p,
+                top_k=top_k,
+                repetition_penalty=repetition_penalty,
+                api_key=api_key,
+                cast=True,
+            )
+        else:
+            return Complete.create(
+                prompt=prompt,
+                model=model,
+                max_tokens=max_tokens,
+                stop=stop,
+                temperature=temperature,
+                top_p=top_p,
+                top_k=top_k,
+                repetition_penalty=repetition_penalty,
+                logprobs=logprobs,
+                api_key=api_key,
+                cast=True,
+            )

src/together/complete.py

@@ -0,0 +1,111 @@
+import typing
+from enum import Enum
+from typing import Any, Dict, List, Optional
+
+from pydantic import BaseModel
+
+
+# Decoder input tokens
+class InputToken(BaseModel):
+    # Token ID from the model tokenizer
+    id: int
+    # Token text
+    text: str
+    # Logprob
+    # Optional since the logprob of the first token cannot be computed
+    logprob: Optional[float]
+
+
+# Generated tokens
+class Token(BaseModel):
+    # Token ID
+    id: int
+    # Logprob
+    logprob: Optional[float]
+    # Is the token a special token
+    # Can be used to ignore tokens when concatenating
+    special: bool
+
+
+# Generation finish reason
+class FinishReason(str, Enum):
+    # number of generated tokens == `max_new_tokens`
+    Length = "length"
+    # the model generated its end of sequence token
+    EndOfSequenceToken = "eos_token"
+    # the model generated a text included in `stop_sequences`
+    StopSequence = "stop_sequence"
+
+
+# Additional sequences when using the `best_of` parameter
+class BestOfSequence(BaseModel):
+    # Generated text
+    generated_text: str
+    # Generation finish reason
+    finish_reason: FinishReason
+    # Number of generated tokens
+    generated_tokens: int
+    # Sampling seed if sampling was activated
+    seed: Optional[int]
+    # Decoder input tokens, empty if decoder_input_details is False
+    prefill: List[InputToken]
+    # Generated tokens
+    tokens: List[Token]
+
+
+# `generate` details
+class Details(BaseModel):
+    # Generation finish reason
+    finish_reason: FinishReason
+    # Number of generated tokens
+    generated_tokens: int
+    # Sampling seed if sampling was activated
+    seed: Optional[int]
+    # Decoder input tokens, empty if decoder_input_details is False
+    prefill: List[InputToken]
+    # Generated tokens
+    tokens: List[Token]
+    # Additional sequences when using the `best_of` parameter
+    best_of_sequences: Optional[List[BestOfSequence]]
+
+
+# `generate` return value
+class Response(BaseModel):
+    # Generated text
+    generated_text: str
+    # Generation details
+    details: Details
+
+
+class Choice(BaseModel):
+    # Generated text
+    text: str
+    finish_reason: Optional[FinishReason] = None
+    logprobs: Optional[List[float]] = None
+
+
+# `generate_stream` details
+class StreamDetails(BaseModel):
+    # Number of generated tokens
+    generated_tokens: int
+    # Sampling seed if sampling was activated
+    seed: Optional[int]
+
+
+# `generate_stream` return value
+class TogetherResponse(BaseModel):
+    choices: Optional[List[Choice]] = None
+    id: Optional[str] = None
+    token: Optional[Token] = None
+    error: Optional[str] = None
+    error_type: Optional[str] = None
+    generated_text: Optional[str] = None
+    # Generation details
+    # Only available when the generation is finished
+    details: Optional[StreamDetails] = None
+
+    def __init__(self, **kwargs: Optional[Dict[str, Any]]) -> None:
+        if kwargs.get("output"):
+            kwargs["choices"] = typing.cast(Dict[str, Any], kwargs["output"])["choices"]
+            kwargs["details"] = kwargs.get("details")
+        super().__init__(**kwargs)

src/together/types.py

@@ -81,13 +81,14 @@ def create_post_request(
     json: Optional[Dict[Any, Any]] = None,
     stream: Optional[bool] = False,
     check_auth: Optional[bool] = True,
+    api_key: Optional[str] = None,
 ) -> requests.Response:
-    if check_auth:
+    if check_auth and api_key is None:
         verify_api_key()
 
     if not headers:
         headers = {
-            "Authorization": f"Bearer {together.api_key}",
+            "Authorization": f"Bearer {api_key or together.api_key}",
             "Content-Type": "application/json",
             "User-Agent": together.user_agent,
         }