llama-stack-mirror/llama_stack/providers/remote/inference/bedrock/bedrock.py

# Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the terms described in the LICENSE file in
# the root directory of this source tree.

import json
from collections.abc import AsyncGenerator, AsyncIterator

from botocore.client import BaseClient

from llama_stack.apis.common.content_types import (
    InterleavedContent,
)
from llama_stack.apis.inference import (
    ChatCompletionRequest,
    ChatCompletionResponse,
    ChatCompletionResponseStreamChunk,
    Inference,
    LogProbConfig,
    Message,
    OpenAIEmbeddingsResponse,
    ResponseFormat,
    SamplingParams,
    ToolChoice,
    ToolConfig,
    ToolDefinition,
    ToolPromptFormat,
)
from llama_stack.providers.remote.inference.bedrock.config import BedrockConfig
from llama_stack.providers.utils.bedrock.client import create_bedrock_client
from llama_stack.providers.utils.inference.model_registry import (
    ModelRegistryHelper,
)
from llama_stack.providers.utils.inference.openai_compat import (
    OpenAIChatCompletionToLlamaStackMixin,
    OpenAICompatCompletionChoice,
    OpenAICompatCompletionResponse,
    OpenAICompletionToLlamaStackMixin,
    get_sampling_strategy_options,
    process_chat_completion_response,
    process_chat_completion_stream_response,
)
from llama_stack.providers.utils.inference.prompt_adapter import (
    chat_completion_request_to_prompt,
)

from .models import MODEL_ENTRIES

REGION_PREFIX_MAP = {
    "us": "us.",
    "eu": "eu.",
    "ap": "ap.",
}


def _get_region_prefix(region: str | None) -> str:
    # AWS requires region prefixes for inference profiles
    if region is None:
        return "us."  # default to US when we don't know

    # Handle case insensitive region matching
    region_lower = region.lower()
    for prefix in REGION_PREFIX_MAP:
        if region_lower.startswith(f"{prefix}-"):
            return REGION_PREFIX_MAP[prefix]

    # Fallback to US for anything we don't recognize
    return "us."


def _to_inference_profile_id(model_id: str, region: str = None) -> str:
    # Return ARNs unchanged
    if model_id.startswith("arn:"):
        return model_id

    # Return inference profile IDs that already have regional prefixes
    if any(model_id.startswith(p) for p in REGION_PREFIX_MAP.values()):
        return model_id

    # Default to US East when no region is provided
    if region is None:
        region = "us-east-1"

    return _get_region_prefix(region) + model_id


class BedrockInferenceAdapter(
    ModelRegistryHelper,
    OpenAIChatCompletionToLlamaStackMixin,
    OpenAICompletionToLlamaStackMixin,
    Inference,
):
    def __init__(self, config: BedrockConfig) -> None:
        ModelRegistryHelper.__init__(self, model_entries=MODEL_ENTRIES)
        self._config = config
        self._client = None

    @property
    def client(self) -> BaseClient:
        if self._client is None:
            self._client = create_bedrock_client(self._config)
        return self._client

    async def initialize(self) -> None:
        pass

    async def shutdown(self) -> None:
        if self._client is not None:
            self._client.close()

    async def completion(
        self,
        model_id: str,
        content: InterleavedContent,
        sampling_params: SamplingParams | None = None,
        response_format: ResponseFormat | None = None,
        stream: bool | None = False,
        logprobs: LogProbConfig | None = None,
    ) -> AsyncGenerator:
        raise NotImplementedError()

    async def chat_completion(
        self,
        model_id: str,
        messages: list[Message],
        sampling_params: SamplingParams | None = None,
        response_format: ResponseFormat | None = None,
        tools: list[ToolDefinition] | None = None,
        tool_choice: ToolChoice | None = ToolChoice.auto,
        tool_prompt_format: ToolPromptFormat | None = None,
        stream: bool | None = False,
        logprobs: LogProbConfig | None = None,
        tool_config: ToolConfig | None = None,
    ) -> ChatCompletionResponse | AsyncIterator[ChatCompletionResponseStreamChunk]:
        if sampling_params is None:
            sampling_params = SamplingParams()
        model = await self.model_store.get_model(model_id)
        request = ChatCompletionRequest(
            model=model.provider_resource_id,
            messages=messages,
            sampling_params=sampling_params,
            tools=tools or [],
            response_format=response_format,
            stream=stream,
            logprobs=logprobs,
            tool_config=tool_config,
        )

        if stream:
            return self._stream_chat_completion(request)
        else:
            return await self._nonstream_chat_completion(request)

    async def _nonstream_chat_completion(self, request: ChatCompletionRequest) -> ChatCompletionResponse:
        params = await self._get_params_for_chat_completion(request)
        res = self.client.invoke_model(**params)
        chunk = res["body"].read()
        result = json.loads(chunk.decode("utf-8"))

        choice = OpenAICompatCompletionChoice(
            finish_reason=result["stop_reason"],
            text=result["generation"],
        )

        response = OpenAICompatCompletionResponse(choices=[choice])
        return process_chat_completion_response(response, request)

    async def _stream_chat_completion(self, request: ChatCompletionRequest) -> AsyncGenerator:
        params = await self._get_params_for_chat_completion(request)
        res = self.client.invoke_model_with_response_stream(**params)
        event_stream = res["body"]

        async def _generate_and_convert_to_openai_compat():
            for event in event_stream:
                if "chunk" in event:
                    chunk_data = event["chunk"]["bytes"]
                    result = json.loads(chunk_data.decode("utf-8"))
                    if "generation" in result:
                        choice = OpenAICompatCompletionChoice(
                            finish_reason=result.get("stop_reason"),
                            text=result["generation"],
                        )
                        yield OpenAICompatCompletionResponse(choices=[choice])

        stream = _generate_and_convert_to_openai_compat()
        async for chunk in process_chat_completion_stream_response(stream, request):
            yield chunk

    async def _get_params_for_chat_completion(self, request: ChatCompletionRequest) -> dict:
        bedrock_model = request.model

        sampling_params = request.sampling_params
        options = get_sampling_strategy_options(sampling_params)

        if sampling_params.max_tokens:
            options["max_gen_len"] = sampling_params.max_tokens
        # Note: repetition_penalty is not supported by AWS Bedrock Llama models
        # if sampling_params.repetition_penalty > 0:
        #     options["repetition_penalty"] = sampling_params.repetition_penalty

        prompt = await chat_completion_request_to_prompt(request, self.get_llama_model(request.model))

        # Convert foundation model ID to inference profile ID
        region_name = self.client.meta.region_name
        inference_profile_id = _to_inference_profile_id(bedrock_model, region_name)

        return {
            "modelId": inference_profile_id,
            "body": json.dumps(
                {
                    "prompt": prompt,
                    **options,
                }
            ),
        }

    async def openai_embeddings(
        self,
        model: str,
        input: str | list[str],
        encoding_format: str | None = "float",
        dimensions: int | None = None,
        user: str | None = None,
    ) -> OpenAIEmbeddingsResponse:
        raise NotImplementedError()