lionagi 0.4.0__py3-none-any.whl → 0.5.0__py3-none-any.whl

lionagi/integrations/ollama_/api_endpoints/model/list_model.py

@@ -0,0 +1,60 @@
+# Copyright (c) 2023 - 2024, HaiyangLi <quantocean.li at gmail dot com>
+#
+# SPDX-License-Identifier: Apache-2.0
+
+
+from pydantic import BaseModel, Field
+
+from ..data_models import OllamaEndpointResponseBody
+
+
+class Detail(BaseModel):
+    parent_model: str = Field(None)
+
+    format: str = Field(None)
+
+    family: str = Field(None)
+
+    families: list[str] | None = Field(None)
+
+    parameter_size: str = Field(None)
+
+    quantization_level: str = Field(None)
+
+
+class LocalModel(BaseModel):
+    name: str = Field(None)
+
+    model: str = Field(None)
+
+    modified_at: str = Field(None)
+
+    size: int = Field(None)
+
+    digest: str = Field(None)
+
+    details: Detail = Field(None)
+
+
+class OllamaListLocalModelsResponseBody(OllamaEndpointResponseBody):
+    models: list[LocalModel] = Field(None)
+
+
+class RunningModel(BaseModel):
+    name: str = Field(None)
+
+    model: str = Field(None)
+
+    size: int = Field(None)
+
+    digest: str = Field(None)
+
+    details: Detail = Field(None)
+
+    expires_at: str = Field(None)
+
+    size_vram: int = Field(None)
+
+
+class OllamaListRunningModelsResponseBody(OllamaEndpointResponseBody):
+    models: list[RunningModel] = Field(None)

lionagi/integrations/ollama_/api_endpoints/model/pull_model.py

@@ -0,0 +1,34 @@
+# Copyright (c) 2023 - 2024, HaiyangLi <quantocean.li at gmail dot com>
+#
+# SPDX-License-Identifier: Apache-2.0
+
+
+from pydantic import Field
+
+from ..data_models import OllamaEndpointRequestBody, OllamaEndpointResponseBody
+
+
+class OllamaPullModelRequestBody(OllamaEndpointRequestBody):
+    name: str = Field(description="Name of the model to pull")
+
+    insecure: bool = Field(
+        False,
+        description="Allow insecure connections to the library. "
+        "Only use this if you are pulling from your own library during development.",
+    )
+
+    stream: bool = Field(
+        True,
+        description="If 'false' the response will be returned as a single response object, "
+        "rather than a stream of objects",
+    )
+
+
+class OllamaPullModelResponseBody(OllamaEndpointResponseBody):
+    status: str = Field(None)
+
+    digest: str = Field(None)
+
+    total: int = Field(None)
+
+    completed: int = Field(None)

lionagi/integrations/ollama_/api_endpoints/model/push_model.py

@@ -0,0 +1,35 @@
+# Copyright (c) 2023 - 2024, HaiyangLi <quantocean.li at gmail dot com>
+#
+# SPDX-License-Identifier: Apache-2.0
+
+from typing import Any
+
+from pydantic import Field
+
+from ..data_models import OllamaEndpointRequestBody, OllamaEndpointResponseBody
+
+
+class OllamaPushModelRequestBody(OllamaEndpointRequestBody):
+    name: str = Field(
+        description="Name of the model to push in the form of '<namespace>/<model>:<tag>'"
+    )
+
+    insecure: Any = Field(
+        None,
+        description="Allow insecure connections to the library. "
+        "Only use this if you are pushing to your library during development.",
+    )
+
+    stream: bool = Field(
+        True,
+        description="If 'false' the response will be returned as a single response object, "
+        "rather than a stream of objects",
+    )
+
+
+class OllamaPushModelResponseBody(OllamaEndpointResponseBody):
+    status: str = Field(None)
+
+    digest: str = Field(None)
+
+    total: int = Field(None)

lionagi/integrations/ollama_/api_endpoints/model/show_model.py

@@ -0,0 +1,36 @@
+# Copyright (c) 2023 - 2024, HaiyangLi <quantocean.li at gmail dot com>
+#
+# SPDX-License-Identifier: Apache-2.0
+
+
+from pydantic import Field
+
+from ..data_models import OllamaEndpointRequestBody, OllamaEndpointResponseBody
+from .list_model import Detail
+
+
+class OllamaShowModelRequestBody(OllamaEndpointRequestBody):
+    name: str = Field(description="Name of the model to show")
+
+    verbose: bool | None = Field(
+        None,
+        description="If set to true, returns full data for verbose response fields",
+    )
+
+
+class OllamaShowModelResponseBody(OllamaEndpointResponseBody):
+    license: str = Field(None)
+
+    modelfile: str = Field(None)
+
+    parameters: str = Field(None)
+
+    template: str = Field(None)
+
+    details: Detail = Field(None)
+
+    model_info: dict = Field(None)
+
+    projector_info: dict = Field(None)
+
+    modified_at: str = Field(None)

lionagi/integrations/ollama_/api_endpoints/option_models.py

@@ -0,0 +1,68 @@
+# Copyright (c) 2023 - 2024, HaiyangLi <quantocean.li at gmail dot com>
+#
+# SPDX-License-Identifier: Apache-2.0
+
+from pydantic import BaseModel, Field
+
+
+class Option(BaseModel):
+    mirostat: int = Field(
+        0, description="Enable Mirostat sampling for controlling perplexity."
+    )
+
+    mirostat_eta: float = Field(
+        0.1,
+        description="Influences how quickly the algorithm responds to feedback from the generated text. ",
+    )
+
+    mirostat_tau: float = Field(
+        5.0,
+        description="Controls the balance between coherence and diversity of the output.",
+    )
+
+    num_ctx: int = Field(
+        2048,
+        description="Sets the size of the context window used to generate the next token.",
+    )
+
+    repeat_last_n: int = Field(
+        64,
+        description="Sets how far back for the model to look back to prevent repetition.",
+    )
+
+    repeat_penalty: float = Field(
+        1.1, description="Sets how strongly to penalize repetitions."
+    )
+
+    temperature: float = Field(
+        0.8, description="The temperature of the model."
+    )
+
+    seed: int = Field(
+        0, description="Sets the random number seed to use for generation."
+    )
+
+    stop: str = Field(None, description="Sets the stop sequences to use.")
+
+    tfs_z: float = Field(
+        1,
+        description="Tail free sampling is used to reduce the impact of less probable tokens from the output.",
+    )
+
+    num_predict: int = Field(
+        128,
+        description="Maximum number of tokens to predict when generating text.",
+    )
+
+    top_k: int = Field(
+        40, description="Reduces the probability of generating nonsense."
+    )
+
+    top_p: float = Field(0.9, description="Works together with top-k.")
+
+    min_p: float = Field(
+        0.0,
+        description="Alternative to the top_p, and aims to ensure a balance of quality and variety. "
+        "The parameter p represents the minimum probability for a token to be considered, "
+        "relative to the probability of the most likely token. ",
+    )

lionagi/integrations/openai_/OpenAIModel.py

@@ -0,0 +1,414 @@
+import warnings
+from pathlib import Path
+
+import yaml
+from dotenv import load_dotenv
+from pydantic import (
+    BaseModel,
+    ConfigDict,
+    Field,
+    field_serializer,
+    model_validator,
+)
+
+from lionagi.service.rate_limiter import RateLimiter, RateLimitError
+from lionagi.service.service_util import invoke_retry
+from lionagi.service.token_calculator import TiktokenCalculator
+
+from .api_endpoints.api_request import OpenAIRequest
+from .api_endpoints.chat_completions.request.request_body import (
+    OpenAIChatCompletionRequestBody,
+    StreamOptions,
+)
+from .api_endpoints.chat_completions.util import get_images, get_text_messages
+from .api_endpoints.data_models import OpenAIEndpointRequestBody
+from .api_endpoints.embeddings.request_body import OpenAIEmbeddingRequestBody
+from .api_endpoints.match_response import match_response
+from .image_token_calculator.image_token_calculator import (
+    OpenAIImageTokenCalculator,
+)
+
+load_dotenv()
+path = Path(__file__).parent
+
+price_config_file_name = path / "openai_max_output_token_data.yaml"
+max_output_token_file_name = path / "openai_price_data.yaml"
+
+
+class OpenAIModel(BaseModel):
+    model: str = Field(description="ID of the model to use.")
+
+    request_model: OpenAIRequest = Field(description="Making requests")
+
+    rate_limiter: RateLimiter = Field(
+        description="Rate Limiter to track usage"
+    )
+
+    text_token_calculator: TiktokenCalculator = Field(
+        default=None, description="Token Calculator"
+    )
+
+    image_token_calculator: OpenAIImageTokenCalculator = Field(
+        default=None, description="Image Token Calculator"
+    )
+
+    estimated_output_len: int = Field(
+        default=0, description="Expected output len before making request"
+    )
+
+    model_config = ConfigDict(extra="forbid")
+
+    @model_validator(mode="before")
+    @classmethod
+    def parse_input(cls, data: dict):
+        if not isinstance(data, dict):
+            raise ValueError("Invalid init param")
+
+        # parse request model
+        request_model_params = {
+            "api_key": data.pop("api_key", None),
+            "endpoint": data.pop("endpoint", None),
+            "method": data.pop("method", None),
+        }
+        if org := data.pop("openai_organization", None):
+            request_model_params["openai_organization"] = org
+        if proj := data.pop("openai_project", None):
+            request_model_params["openai_project"] = proj
+        if content_type := data.pop("content_type", None):
+            request_model_params["content_type"] = content_type
+
+        data["request_model"] = OpenAIRequest(**request_model_params)
+
+        # parse rate limiter
+        if "rate_limiter" not in data:
+            rate_limiter_params = {}
+            if limit_tokens := data.pop("limit_tokens", None):
+                rate_limiter_params["limit_tokens"] = limit_tokens
+            if limit_requests := data.pop("limit_requests", None):
+                rate_limiter_params["limit_requests"] = limit_requests
+
+            data["rate_limiter"] = RateLimiter(**rate_limiter_params)
+
+        # parse token calculator
+        try:
+            text_calc = TiktokenCalculator(encoding_name=data.get("model"))
+            data["text_token_calculator"] = text_calc
+        except Exception:
+            pass
+
+        # set image calcultor
+        try:
+            data["image_token_calculator"] = OpenAIImageTokenCalculator(
+                model=data.get("model")
+            )
+        except Exception:
+            pass
+
+        return data
+
+    @field_serializer("request_model")
+    def serialize_request_model(self, value: OpenAIRequest):
+        return value.model_dump(exclude_unset=True)
+
+    @invoke_retry(max_retries=3, base_delay=1, max_delay=60)
+    async def invoke(
+        self,
+        request_body: OpenAIEndpointRequestBody,
+        estimated_output_len: int = 0,
+        output_file=None,
+        parse_response=True,
+    ):
+        if request_model := getattr(request_body, "model"):
+            if request_model != self.model:
+                raise ValueError(
+                    f"Request model does not match. Model is {self.model}, but request is made for {request_model}."
+                )
+
+        if getattr(request_body, "stream", None) and not isinstance(
+            request_body, OpenAIChatCompletionRequestBody
+        ):
+            raise ValueError("Stream is only supported for chat completions")
+
+        # check remaining rate limit
+        input_token_len = await self.get_input_token_len(request_body)
+
+        # chat completion request body attribute
+        if getattr(request_body, "max_completion_tokens", None):
+            estimated_output_len = request_body.max_completion_tokens
+
+        invoke_viability_result = self.verify_invoke_viability(
+            input_tokens_len=input_token_len,
+            estimated_output_len=estimated_output_len,
+        )
+        if not invoke_viability_result:
+            raise RateLimitError("Rate limit reached for requests")
+
+        try:
+            if getattr(request_body, "stream", None):
+                return await self.stream(
+                    request_body,
+                    output_file=output_file,
+                    parse_response=parse_response,
+                )
+
+            if getattr(request_body, "file", None):
+                response_body, response_headers = (
+                    await self.request_model.invoke(
+                        form_data=request_body,
+                        output_file=output_file,
+                        with_response_header=True,
+                        parse_response=False,
+                    )
+                )
+            else:
+                response_body, response_headers = (
+                    await self.request_model.invoke(
+                        json_data=request_body,
+                        output_file=output_file,
+                        with_response_header=True,
+                        parse_response=False,
+                    )
+                )
+
+            self.check_limits_info(response_headers)
+
+            if response_body:
+                # mainly for chat/completions and embedding endpoints
+                # update rate limit
+                if response_body.get("usage"):
+                    total_token_usage = response_body["usage"]["total_tokens"]
+                    self.rate_limiter.update_rate_limit(
+                        response_headers.get("Date"), total_token_usage
+                    )
+                else:  # No Token limits condition (request limit only)
+                    self.rate_limiter.update_rate_limit(
+                        response_headers.get("Date")
+                    )
+            else:
+                # for audio/speech endpoint (without response body object)
+                self.rate_limiter.update_rate_limit(
+                    response_headers.get("Date")
+                )
+
+            self.check_remaining_info(response_headers)
+
+            if parse_response:
+                return match_response(self.request_model, response_body)
+            else:
+                return response_body
+
+        except Exception as e:  # TODO: example
+            raise e
+
+    async def stream(
+        self,
+        request_body: OpenAIEndpointRequestBody,
+        output_file=None,
+        parse_response=True,
+    ):
+        if not isinstance(request_body, OpenAIChatCompletionRequestBody):
+            raise ValueError("Stream is only supported for chat completions")
+
+        stream_options_included = bool(getattr(request_body, "stream_options"))
+        if not stream_options_included:
+            setattr(
+                request_body,
+                "stream_options",
+                StreamOptions(include_usage=True),
+            )
+
+        response_list = []
+        async for chunk in self.request_model.stream(
+            json_data=request_body,
+            output_file=output_file,
+            with_response_header=True,
+        ):
+            response_list.append(chunk)
+
+        response_headers = response_list.pop()
+        if stream_options_included:
+            usage_chunk = response_list[-1]
+        else:
+            usage_chunk = response_list.pop()
+
+        self.check_limits_info(response_headers)
+        total_token_usage = usage_chunk["usage"]["total_tokens"]
+        self.rate_limiter.update_rate_limit(
+            response_headers.get("Date"), total_token_usage
+        )
+
+        if parse_response:
+            return match_response(self.request_model, response_list)
+        else:
+            return response_list
+
+    async def get_input_token_len(
+        self, request_body: OpenAIEndpointRequestBody
+    ):
+        if request_model := getattr(request_body, "model"):
+            if request_model != self.model:
+                raise ValueError(
+                    f"Request model does not match. Model is {self.model}, but request is made for {request_model}."
+                )
+        # TODO: match with new Request Body format
+        if isinstance(request_body, OpenAIChatCompletionRequestBody):
+            messages_text = get_text_messages(request_body)
+            # text_token_calculator should always be available for chat completions
+            text_tokens = self.text_token_calculator.calculate(messages_text)
+
+            image_urls = get_images(request_body)
+            image_tokens = 0
+            for url, detail in image_urls:
+                if self.image_token_calculator:
+                    image_tokens += (
+                        await self.image_token_calculator.calculate(
+                            url, detail
+                        )
+                    )
+                else:
+                    raise ValueError(
+                        "The model does not have vision capabilities."
+                    )
+
+            return text_tokens + image_tokens
+        elif isinstance(request_body, OpenAIEmbeddingRequestBody):
+            text = request_body.input
+            if isinstance(text, str):  # str
+                return self.text_token_calculator.calculate(text)
+            elif isinstance(text[0], int):  # List[int]
+                return len(text)
+            elif isinstance(text[0], str):  # List[str]
+                total_totkens = 0
+                for t in text:
+                    total_totkens += self.text_token_calculator.calculate(t)
+                return total_totkens
+            else:  # List[List[int]]
+                total_tokens = 0
+                for t in text:
+                    total_tokens += len(t)
+                return total_tokens
+
+        # TODO: add other rules for other endpoints if input tokens should be calculated
+        return 0  # no tokens rate limit
+
+    def verify_invoke_viability(
+        self, input_tokens_len: int = 0, estimated_output_len: int = 0
+    ):
+        self.rate_limiter.release_tokens()
+
+        estimated_output_len = (
+            estimated_output_len
+            if estimated_output_len != 0
+            else self.estimated_output_len
+        )
+        if estimated_output_len == 0:
+            with open(max_output_token_file_name) as file:
+                output_token_config = yaml.safe_load(file)
+                estimated_output_len = output_token_config.get(self.model, 0)
+                self.estimated_output_len = (
+                    estimated_output_len  # update to default max output len
+                )
+
+        if self.rate_limiter.check_availability(
+            input_tokens_len, estimated_output_len
+        ):
+            return True
+        else:
+            return False
+
+    def check_limits_info(self, response_headers):
+        if response_headers.get("x-ratelimit-limit-requests"):
+            if self.rate_limiter.limit_requests is None:
+                self.rate_limiter.limit_requests = int(
+                    response_headers.get("x-ratelimit-limit-requests")
+                )
+            else:
+                if self.rate_limiter.limit_requests > int(
+                    response_headers.get("x-ratelimit-limit-requests")
+                ):
+                    warnings.warn(
+                        "The configured request limit exceeds the account's allowed request limit."
+                        "This may lead to unexpected throttling or rejection of requests.",
+                        UserWarning,
+                    )
+        if response_headers.get("x-ratelimit-limit-tokens"):
+            if self.rate_limiter.limit_tokens is None:
+                self.rate_limiter.limit_tokens = int(
+                    response_headers.get("x-ratelimit-limit-tokens")
+                )
+            else:
+                if self.rate_limiter.limit_tokens > int(
+                    response_headers.get("x-ratelimit-limit-tokens")
+                ):
+                    warnings.warn(
+                        "The configured token limit exceeds the account's allowed token limit."
+                        "This may lead to unexpected throttling or rejection of requests.",
+                        UserWarning,
+                    )
+
+    def check_remaining_info(self, response_headers):
+        if response_headers.get("x-ratelimit-remaining-tokens"):
+            if (
+                int(response_headers.get("x-ratelimit-remaining-tokens"))
+                < self.rate_limiter.remaining_tokens
+            ):
+                token_diff = self.rate_limiter.remaining_tokens - int(
+                    response_headers.get("x-ratelimit-remaining-tokens")
+                )
+                self.rate_limiter.update_rate_limit(
+                    response_headers.get("Date"), token_diff
+                )
+
+        if response_headers.get("x-ratelimit-remaining-requests"):
+            if (
+                int(response_headers.get("x-ratelimit-remaining-requests"))
+                < self.rate_limiter.remaining_requests
+            ):
+                request_diff = self.rate_limiter.remaining_requests - int(
+                    response_headers.get("x-ratelimit-remaining-requests")
+                )
+                for i in range(request_diff):
+                    self.rate_limiter.update_rate_limit(
+                        response_headers.get("Date")
+                    )
+
+    def estimate_text_price(
+        self,
+        input_text: str,
+        with_batch: bool = False,
+        estimated_num_of_output_tokens: int = 0,
+    ):
+        if self.text_token_calculator is None:
+            raise ValueError(
+                "Estimating price currently only supports chat/completions endpoint"
+            )
+
+        # only if text_token_calculator is available
+        num_of_input_tokens = self.text_token_calculator.calculate(input_text)
+
+        # read openai price info from config file
+        with open(price_config_file_name) as file:
+            price_config = yaml.safe_load(file)
+
+        if self.request_model.endpoint == "chat/completions":
+            model_price_info_dict = price_config["model"][self.model]
+            if with_batch:
+                estimated_price = (
+                    model_price_info_dict["input_tokens_with_batch"]
+                    * num_of_input_tokens
+                    + model_price_info_dict["output_tokens_with_batch"]
+                    * estimated_num_of_output_tokens
+                )
+            else:
+                estimated_price = (
+                    model_price_info_dict["input_tokens"] * num_of_input_tokens
+                    + model_price_info_dict["output_tokens"]
+                    * estimated_num_of_output_tokens
+                )
+
+        else:
+            # TODO: add price config for other endpoints
+            raise ValueError(
+                "Estimating price currently only supports chat/completions endpoint"
+            )
+        return estimated_price