ommlds 0.0.0.dev466__py3-none-any.whl → 0.0.0.dev468__py3-none-any.whl

ommlds/minichain/backends/impls/mlx/chat.py

@@ -1,3 +1,4 @@
+import contextlib
 import typing as ta
 
 from omlish import check
@@ -5,6 +6,7 @@ from omlish import lang
 from omlish import typedvalues as tv
 
 from .....backends import mlx as mlxu
+from ....chat.choices.services import ChatChoicesOutputs
 from ....chat.choices.services import ChatChoicesRequest
 from ....chat.choices.services import ChatChoicesResponse
 from ....chat.choices.services import static_check_is_chat_choices_service
@@ -14,19 +16,28 @@ from ....chat.messages import AiMessage
 from ....chat.messages import Message
 from ....chat.messages import SystemMessage
 from ....chat.messages import UserMessage
+from ....chat.stream.services import ChatChoicesStreamRequest
+from ....chat.stream.services import ChatChoicesStreamResponse
+from ....chat.stream.services import static_check_is_chat_choices_stream_service
+from ....chat.stream.types import AiChoiceDeltas
+from ....chat.stream.types import AiChoicesDeltas
+from ....chat.stream.types import ContentAiChoiceDelta
 from ....configs import Config
 from ....llms.types import MaxTokens
 from ....models.configs import ModelPath
 from ....models.configs import ModelRepo
 from ....models.configs import ModelSpecifier
+from ....resources import UseResources
 from ....standard import DefaultOptions
+from ....stream.services import StreamResponseSink
+from ....stream.services import new_stream_response
 
 
 ##
 
 
 # @omlish-manifest $.minichain.backends.strings.manifests.BackendStringsManifest(
-#     ['ChatChoicesService'],
+#     ['ChatChoicesService', 'ChatChoicesStreamService'],
 #     'mlx',
 # )
 
@@ -34,12 +45,7 @@ from ....standard import DefaultOptions
 ##
 
 
-# @omlish-manifest $.minichain.registries.manifests.RegistryManifest(
-#     name='mlx',
-#     type='ChatChoicesService',
-# )
-@static_check_is_chat_choices_service
-class MlxChatChoicesService(lang.ExitStacked):
+class BaseMlxChatChoicesService(lang.ExitStacked):
     DEFAULT_MODEL: ta.ClassVar[ModelSpecifier] = (
         # 'mlx-community/DeepSeek-Coder-V2-Lite-Instruct-8bit'
         # 'mlx-community/Llama-3.3-70B-Instruct-4bit'
@@ -52,8 +58,8 @@ class MlxChatChoicesService(lang.ExitStacked):
         # 'mlx-community/Qwen2.5-0.5B-4bit'
         # 'mlx-community/Qwen2.5-32B-Instruct-8bit'
         # 'mlx-community/Qwen2.5-Coder-32B-Instruct-8bit'
-        # 'mlx-community/mamba-2.8b-hf-f16'
         # 'mlx-community/Qwen3-30B-A3B-6bit'
+        # 'mlx-community/mamba-2.8b-hf-f16'
     )
 
     def __init__(self, *configs: Config) -> None:
@@ -70,10 +76,7 @@ class MlxChatChoicesService(lang.ExitStacked):
     }
 
     def _get_msg_content(self, m: Message) -> str | None:
-        if isinstance(m, AiMessage):
-            return check.isinstance(m.c, str)
-
-        elif isinstance(m, (SystemMessage, UserMessage)):
+        if isinstance(m, (AiMessage, SystemMessage, UserMessage)):
             return check.isinstance(m.c, str)
 
         else:
@@ -96,10 +99,9 @@ class MlxChatChoicesService(lang.ExitStacked):
         max_tokens=MaxTokens,
     )
 
-    async def invoke(self, request: ChatChoicesRequest) -> ChatChoicesResponse:
-        loaded_model = self._load_model()
-
-        tokenizer = loaded_model.tokenization.tokenizer
+    @lang.cached_function(transient=True)
+    def _get_tokenizer(self) -> mlxu.tokenization.Tokenizer:
+        tokenizer = self._load_model().tokenization.tokenizer
 
         if not (
                 hasattr(tokenizer, 'apply_chat_template') and
@@ -107,26 +109,44 @@ class MlxChatChoicesService(lang.ExitStacked):
         ):
             raise RuntimeError(tokenizer)
 
-        prompt = tokenizer.apply_chat_template(
+        return tokenizer
+
+    def _build_prompt(self, messages: ta.Sequence[Message]) -> str:
+        return check.isinstance(self._get_tokenizer().apply_chat_template(
             [  # type: ignore[arg-type]
                 dict(
                     role=self.ROLES_MAP[type(m)],
                     content=self._get_msg_content(m),
                 )
-                for m in request.v
+                for m in messages
             ],
             tokenize=False,
             add_generation_prompt=True,
-        )
+        ), str)
 
-        kwargs = dict()
+    def _build_kwargs(self, oc: tv.TypedValuesConsumer) -> dict[str, ta.Any]:
+        kwargs: dict[str, ta.Any] = {}
+        kwargs.update(oc.pop_scalar_kwargs(**self._OPTION_KWARG_NAMES_MAP))
+        return kwargs
+
+
+# @omlish-manifest $.minichain.registries.manifests.RegistryManifest(
+#     name='mlx',
+#     type='ChatChoicesService',
+# )
+@static_check_is_chat_choices_service
+class MlxChatChoicesService(BaseMlxChatChoicesService):
+    async def invoke(self, request: ChatChoicesRequest) -> ChatChoicesResponse:
+        loaded_model = self._load_model()
+
+        prompt = self._build_prompt(request.v)
 
         with tv.consume(
                 *self._default_options,
                 *request.options,
                 override=True,
         ) as oc:
-            kwargs.update(oc.pop_scalar_kwargs(**self._OPTION_KWARG_NAMES_MAP))
+            kwargs = self._build_kwargs(oc)
 
         response = mlxu.generate(
             loaded_model.model,
@@ -139,3 +159,57 @@ class MlxChatChoicesService(lang.ExitStacked):
         return ChatChoicesResponse([
             AiChoice([AiMessage(response)])  # noqa
         ])
+
+
+# @omlish-manifest $.minichain.registries.manifests.RegistryManifest(
+#     name='mlx',
+#     type='ChatChoicesStreamService',
+# )
+@static_check_is_chat_choices_stream_service
+class MlxChatChoicesStreamService(BaseMlxChatChoicesService):
+    def __init__(self, *configs: Config) -> None:
+        super().__init__()
+
+        with tv.consume(*configs) as cc:
+            self._model = cc.pop(MlxChatChoicesService.DEFAULT_MODEL)
+            self._default_options: tv.TypedValues = DefaultOptions.pop(cc)
+
+    READ_CHUNK_SIZE = 64 * 1024
+
+    async def invoke(
+            self,
+            request: ChatChoicesStreamRequest,
+            *,
+            max_tokens: int = 4096,  # FIXME: ChatOption
+    ) -> ChatChoicesStreamResponse:
+        loaded_model = self._load_model()
+
+        prompt = self._build_prompt(request.v)
+
+        with tv.consume(
+                *self._default_options,
+                *request.options,
+                override=True,
+        ) as oc:
+            oc.pop(UseResources, None)
+            kwargs = self._build_kwargs(oc)
+
+        async with UseResources.or_new(request.options) as rs:
+            gen: ta.Iterator[mlxu.GenerationOutput] = rs.enter_context(contextlib.closing(mlxu.stream_generate(
+                loaded_model.model,
+                loaded_model.tokenization,
+                check.isinstance(prompt, str),
+                mlxu.GenerationParams(**kwargs),
+                # verbose=True,
+            )))
+
+            async def inner(sink: StreamResponseSink[AiChoicesDeltas]) -> ta.Sequence[ChatChoicesOutputs]:
+                for go in gen:
+                    if go.text:
+                        await sink.emit(AiChoicesDeltas([AiChoiceDeltas([
+                            ContentAiChoiceDelta(go.text),
+                        ])]))
+
+                return []
+
+            return await new_stream_response(rs, inner)

ommlds/minichain/backends/impls/ollama/chat.py

@@ -0,0 +1,196 @@
+import typing as ta
+
+from omlish import check
+from omlish import lang
+from omlish import marshal as msh
+from omlish import typedvalues as tv
+from omlish.formats import json
+from omlish.http import all as http
+from omlish.io.buffers import DelimitingBuffer
+
+from .....backends.ollama import protocol as pt
+from ....chat.choices.services import ChatChoicesOutputs
+from ....chat.choices.services import ChatChoicesRequest
+from ....chat.choices.services import ChatChoicesResponse
+from ....chat.choices.services import static_check_is_chat_choices_service
+from ....chat.choices.types import AiChoice
+from ....chat.messages import AiMessage
+from ....chat.messages import AnyAiMessage
+from ....chat.messages import Message
+from ....chat.messages import SystemMessage
+from ....chat.messages import UserMessage
+from ....chat.stream.services import ChatChoicesStreamRequest
+from ....chat.stream.services import ChatChoicesStreamResponse
+from ....chat.stream.services import static_check_is_chat_choices_stream_service
+from ....chat.stream.types import AiChoiceDeltas
+from ....chat.stream.types import AiChoicesDeltas
+from ....chat.stream.types import ContentAiChoiceDelta
+from ....models.configs import ModelName
+from ....resources import UseResources
+from ....standard import ApiUrl
+from ....stream.services import StreamResponseSink
+from ....stream.services import new_stream_response
+
+
+##
+
+
+# @omlish-manifest $.minichain.backends.strings.manifests.BackendStringsManifest(
+#     [
+#         'ChatChoicesService',
+#         'ChatChoicesStreamService',
+#     ],
+#     'ollama',
+# )
+
+
+##
+
+
+class BaseOllamaChatChoicesService(lang.Abstract):
+    DEFAULT_API_URL: ta.ClassVar[ApiUrl] = ApiUrl('http://localhost:11434/api')
+    DEFAULT_MODEL_NAME: ta.ClassVar[ModelName] = ModelName('llama3.2')
+
+    def __init__(
+            self,
+            *configs: ApiUrl | ModelName,
+    ) -> None:
+        super().__init__()
+
+        with tv.consume(*configs) as cc:
+            self._api_url = cc.pop(self.DEFAULT_API_URL)
+            self._model_name = cc.pop(self.DEFAULT_MODEL_NAME)
+
+    #
+
+    ROLE_MAP: ta.ClassVar[ta.Mapping[type[Message], pt.Role]] = {
+        SystemMessage: 'system',
+        UserMessage: 'user',
+        AiMessage: 'assistant',
+    }
+
+    @classmethod
+    def _get_message_content(cls, m: Message) -> str | None:
+        if isinstance(m, (AiMessage, UserMessage, SystemMessage)):
+            return check.isinstance(m.c, str)
+        else:
+            raise TypeError(m)
+
+    @classmethod
+    def _build_request_messages(cls, mc_msgs: ta.Iterable[Message]) -> ta.Sequence[pt.Message]:
+        messages: list[pt.Message] = []
+        for m in mc_msgs:
+            messages.append(pt.Message(
+                role=cls.ROLE_MAP[type(m)],
+                content=cls._get_message_content(m),
+            ))
+        return messages
+
+
+##
+
+
+# @omlish-manifest $.minichain.registries.manifests.RegistryManifest(
+#     name='ollama',
+#     type='ChatChoicesService',
+# )
+@static_check_is_chat_choices_service
+class OllamaChatChoicesService(BaseOllamaChatChoicesService):
+    async def invoke(
+            self,
+            request: ChatChoicesRequest,
+    ) -> ChatChoicesResponse:
+        messages = self._build_request_messages(request.v)
+
+        a_req = pt.ChatRequest(
+            model=self._model_name.v,
+            messages=messages,
+            # tools=tools or None,
+            stream=False,
+        )
+
+        raw_request = msh.marshal(a_req)
+
+        raw_response = http.request(
+            self._api_url.v.removesuffix('/') + '/chat',
+            data=json.dumps(raw_request).encode('utf-8'),
+        )
+
+        json_response = json.loads(check.not_none(raw_response.data).decode('utf-8'))
+
+        resp = msh.unmarshal(json_response, pt.ChatResponse)
+
+        out: list[AnyAiMessage] = []
+        if resp.message.role == 'assistant':
+            out.append(AiMessage(
+                check.not_none(resp.message.content),
+            ))
+        else:
+            raise TypeError(resp.message.role)
+
+        return ChatChoicesResponse([
+            AiChoice(out),
+        ])
+
+
+##
+
+
+# @omlish-manifest $.minichain.registries.manifests.RegistryManifest(
+#     name='ollama',
+#     type='ChatChoicesStreamService',
+# )
+@static_check_is_chat_choices_stream_service
+class OllamaChatChoicesStreamService(BaseOllamaChatChoicesService):
+    READ_CHUNK_SIZE = 64 * 1024
+
+    async def invoke(
+            self,
+            request: ChatChoicesStreamRequest,
+    ) -> ChatChoicesStreamResponse:
+        messages = self._build_request_messages(request.v)
+
+        a_req = pt.ChatRequest(
+            model=self._model_name.v,
+            messages=messages,
+            # tools=tools or None,
+            stream=True,
+        )
+
+        raw_request = msh.marshal(a_req)
+
+        http_request = http.HttpRequest(
+            self._api_url.v.removesuffix('/') + '/chat',
+            data=json.dumps(raw_request).encode('utf-8'),
+        )
+
+        async with UseResources.or_new(request.options) as rs:
+            http_client = rs.enter_context(http.client())
+            http_response = rs.enter_context(http_client.stream_request(http_request))
+
+            async def inner(sink: StreamResponseSink[AiChoicesDeltas]) -> ta.Sequence[ChatChoicesOutputs] | None:
+                db = DelimitingBuffer([b'\r', b'\n', b'\r\n'])
+                while True:
+                    # FIXME: read1 not on response stream protocol
+                    b = http_response.stream.read1(self.READ_CHUNK_SIZE)  # type: ignore[attr-defined]
+                    for l in db.feed(b):
+                        if isinstance(l, DelimitingBuffer.Incomplete):
+                            # FIXME: handle
+                            return []
+
+                        lj = json.loads(l.decode('utf-8'))
+                        lp: pt.ChatResponse = msh.unmarshal(lj, pt.ChatResponse)
+
+                        check.state(lp.message.role == 'assistant')
+                        check.none(lp.message.tool_name)
+                        check.state(not lp.message.tool_calls)
+
+                        if (c := lp.message.content):
+                            await sink.emit(AiChoicesDeltas([AiChoiceDeltas([ContentAiChoiceDelta(
+                                c,
+                            )])]))
+
+                    if not b:
+                        return []
+
+            return await new_stream_response(rs, inner)

ommlds/minichain/backends/impls/openai/chat.py

@@ -26,8 +26,8 @@ from ....chat.choices.services import static_check_is_chat_choices_service
 from ....models.configs import ModelName
 from ....standard import ApiKey
 from ....standard import DefaultOptions
-from .format2 import OpenaiChatRequestHandler
-from .format2 import build_mc_choices_response
+from .format import OpenaiChatRequestHandler
+from .format import build_mc_choices_response
 from .names import MODEL_NAMES
 
 

ommlds/minichain/backends/impls/openai/format.py

@@ -2,18 +2,17 @@ import typing as ta
 
 from omlish import cached
 from omlish import check
-from omlish import lang
 from omlish import typedvalues as tv
 from omlish.formats import json
 
+from .....backends.openai import protocol as pt
 from ....chat.choices.services import ChatChoicesResponse
 from ....chat.choices.types import AiChoice
+from ....chat.choices.types import AiChoices
 from ....chat.choices.types import ChatChoicesOptions
-from ....chat.messages import AiChat
 from ....chat.messages import AiMessage
 from ....chat.messages import AnyAiMessage
 from ....chat.messages import Chat
-from ....chat.messages import Message
 from ....chat.messages import SystemMessage
 from ....chat.messages import ToolUseMessage
 from ....chat.messages import ToolUseResultMessage
@@ -28,7 +27,7 @@ from ....llms.types import MaxTokens
 from ....llms.types import Temperature
 from ....llms.types import TokenUsage
 from ....llms.types import TokenUsageOutput
-from ....tools.jsonschema import build_tool_spec_json_schema
+from ....tools.jsonschema import build_tool_spec_params_json_schema
 from ....tools.types import ToolSpec
 from ....tools.types import ToolUse
 from ....types import Option
@@ -37,61 +36,115 @@ from ....types import Option
 ##
 
 
-def build_request_messages(chat: Chat) -> ta.Sequence[ta.Mapping[str, ta.Any]]:
-    out: list[dict[str, ta.Any]] = []
+def build_oai_request_msgs(mc_chat: Chat) -> ta.Sequence[pt.ChatCompletionMessage]:
+    oai_msgs: list[pt.ChatCompletionMessage] = []
 
-    for m in chat:
-        if isinstance(m, SystemMessage):
-            out.append(dict(
-                role='system',
-                content=m.c,
+    for mc_msg in mc_chat:
+        if isinstance(mc_msg, SystemMessage):
+            oai_msgs.append(pt.SystemChatCompletionMessage(
+                content=check.isinstance(mc_msg.c, str),
             ))
 
-        elif isinstance(m, AiMessage):
-            out.append(dict(
-                role='assistant',
-                content=check.isinstance(m.c, (str, None)),
+        elif isinstance(mc_msg, AiMessage):
+            oai_msgs.append(pt.AssistantChatCompletionMessage(
+                content=check.isinstance(mc_msg.c, (str, None)),
             ))
 
-        elif isinstance(m, ToolUseMessage):
-            out.append(dict(
-                role='assistant',
-                tool_calls=[
-                    dict(
-                        id=m.tu.id,
-                        function=dict(
-                            arguments=check.not_none(m.tu.raw_args),
-                            name=m.tu.name,
-                        ),
-                        type='function',
+        elif isinstance(mc_msg, ToolUseMessage):
+            oai_msgs.append(pt.AssistantChatCompletionMessage(
+                tool_calls=[pt.AssistantChatCompletionMessage.ToolCall(
+                    id=check.not_none(mc_msg.tu.id),
+                    function=pt.AssistantChatCompletionMessage.ToolCall.Function(
+                        arguments=check.not_none(mc_msg.tu.raw_args),
+                        name=mc_msg.tu.name,
                     ),
-                ],
+                )],
             ))
 
-        elif isinstance(m, UserMessage):
-            out.append(dict(
-                role='user',
-                content=prepare_content_str(m.c),
+        elif isinstance(mc_msg, UserMessage):
+            oai_msgs.append(pt.UserChatCompletionMessage(
+                content=prepare_content_str(mc_msg.c),
             ))
 
-        elif isinstance(m, ToolUseResultMessage):
+        elif isinstance(mc_msg, ToolUseResultMessage):
             tc: str
-            if isinstance(m.tur.c, str):
-                tc = m.tur.c
-            elif isinstance(m.tur.c, JsonContent):
-                tc = json.dumps_compact(m.tur.c)
+            if isinstance(mc_msg.tur.c, str):
+                tc = mc_msg.tur.c
+            elif isinstance(mc_msg.tur.c, JsonContent):
+                tc = json.dumps_compact(mc_msg.tur.c)
             else:
-                raise TypeError(m.tur.c)
-            out.append(dict(
-                role='tool',
-                tool_call_id=m.tur.id,
+                raise TypeError(mc_msg.tur.c)
+            oai_msgs.append(pt.ToolChatCompletionMessage(
+                tool_call_id=check.not_none(mc_msg.tur.id),
                 content=tc,
             ))
 
         else:
-            raise TypeError(m)
+            raise TypeError(mc_msg)
 
-    return out
+    return oai_msgs
+
+
+#
+
+
+def build_mc_ai_choice(oai_choice: pt.ChatCompletionResponseChoice) -> AiChoice:
+    cur: list[AnyAiMessage] = []
+
+    oai_msg = oai_choice.message
+
+    if (oai_c := oai_msg.content) is not None:
+        cur.append(AiMessage(check.isinstance(oai_c, str)))
+
+    for oai_tc in oai_msg.tool_calls or []:
+        cur.append(ToolUseMessage(ToolUse(
+            id=oai_tc.id,
+            name=oai_tc.function.name,
+            args=json.loads(oai_tc.function.arguments or '{}'),
+            raw_args=oai_tc.function.arguments,
+        )))
+
+    return AiChoice(cur)
+
+
+def build_mc_ai_choices(oai_resp: pt.ChatCompletionResponse) -> AiChoices:
+    return [
+        build_mc_ai_choice(oai_choice)
+        for oai_choice in oai_resp.choices
+    ]
+
+
+def build_mc_choices_response(oai_resp: pt.ChatCompletionResponse) -> ChatChoicesResponse:
+    return ChatChoicesResponse(
+        build_mc_ai_choices(oai_resp),
+
+        tv.TypedValues(
+            *([TokenUsageOutput(TokenUsage(
+                input=tu.prompt_tokens,
+                output=tu.completion_tokens,
+                total=tu.total_tokens,
+            ))] if (tu := oai_resp.usage) is not None else []),
+        ),
+    )
+
+
+def build_mc_ai_choice_delta(delta: pt.ChatCompletionChunkChoiceDelta) -> AiChoiceDelta:
+    if delta.content is not None:
+        check.state(not delta.tool_calls)
+        return ContentAiChoiceDelta(delta.content)
+
+    elif delta.tool_calls is not None:
+        check.state(delta.content is None)
+        tc = check.single(delta.tool_calls)
+        tc_fn = check.not_none(tc.function)
+        return PartialToolUseAiChoiceDelta(
+            id=tc.id,
+            name=tc_fn.name,
+            raw_args=tc_fn.arguments,
+        )
+
+    else:
+        raise ValueError(delta)
 
 
 ##
@@ -112,14 +165,6 @@ class OpenaiChatRequestHandler:
         self._model = model
         self._mandatory_kwargs = mandatory_kwargs
 
-    ROLES_MAP: ta.ClassVar[ta.Mapping[type[Message], str]] = {
-        SystemMessage: 'system',
-        UserMessage: 'user',
-        AiMessage: 'assistant',
-        ToolUseMessage: 'assistant',
-        ToolUseResultMessage: 'tool',
-    }
-
     DEFAULT_OPTIONS: ta.ClassVar[tv.TypedValues[Option]] = tv.TypedValues[Option](
         Temperature(0.),
         MaxTokens(1024),
@@ -162,72 +207,26 @@ class OpenaiChatRequestHandler:
         )
 
     @cached.function
-    def raw_request(self) -> ta.Mapping[str, ta.Any]:
+    def oai_request(self) -> pt.ChatCompletionRequest:
         po = self._process_options()
 
-        tools = [
-            dict(
-                type='function',
-                function=build_tool_spec_json_schema(ts),
+        tools: list[pt.ChatCompletionRequestTool] = [
+            pt.ChatCompletionRequestTool(
+                function=pt.ChatCompletionRequestTool.Function(
+                    name=check.not_none(ts.name),
+                    description=prepare_content_str(ts.desc),
+                    parameters=build_tool_spec_params_json_schema(ts),
+                ),
             )
             for ts in po.tools_by_name.values()
         ]
 
-        return dict(
+        return pt.ChatCompletionRequest(
             model=self._model,
-            messages=build_request_messages(self._chat),
+            messages=build_oai_request_msgs(self._chat),
             top_p=1,
-            **lang.opt_kw(tools=tools),
+            tools=tools or None,
             frequency_penalty=0.0,
             presence_penalty=0.0,
             **po.kwargs,
         )
-
-    def build_ai_chat(self, message: ta.Mapping[str, ta.Any]) -> AiChat:
-        out: list[AnyAiMessage] = []
-        if (c := message.get('content')) is not None:
-            out.append(AiMessage(c))
-        for tc in message.get('tool_calls', []):
-            out.append(ToolUseMessage(
-                ToolUse(
-                    id=tc['id'],
-                    name=tc['function']['name'],
-                    args=json.loads(tc['function']['arguments'] or '{}'),
-                    raw_args=tc['function']['arguments'],
-                ),
-            ))
-        return out
-
-    def build_response(self, raw_response: ta.Mapping[str, ta.Any]) -> ChatChoicesResponse:
-        return ChatChoicesResponse(
-            [
-                AiChoice(self.build_ai_chat(choice['message']))
-                for choice in raw_response['choices']
-            ],
-
-            tv.TypedValues(
-                *([TokenUsageOutput(TokenUsage(
-                    input=tu['prompt_tokens'],
-                    output=tu['completion_tokens'],
-                    total=tu['total_tokens'],
-                ))] if (tu := raw_response.get('usage')) is not None else []),
-            ),
-        )
-
-    def build_ai_choice_delta(self, delta: ta.Mapping[str, ta.Any]) -> AiChoiceDelta:
-        if (c := delta.get('content')) is not None:
-            check.state(not delta.get('tool_calls'))
-            return ContentAiChoiceDelta(c)
-
-        elif (tcs := delta.get('tool_calls')) is not None:  # noqa
-            check.state(delta.get('content') is None)
-            tc = check.single(tcs)
-            tc_fn = tc['function']
-            return PartialToolUseAiChoiceDelta(
-                id=tc.get('id'),
-                name=tc_fn.get('name'),
-                raw_args=tc_fn.get('arguments'),
-            )
-
-        else:
-            raise ValueError(delta)