ommlds 0.0.0.dev448__py3-none-any.whl → 0.0.0.dev450__py3-none-any.whl

ommlds/.omlish-manifests.json

@@ -96,7 +96,7 @@
     "module": ".minichain.backends.impls.google.chat",
     "attr": null,
     "file": "ommlds/minichain/backends/impls/google/chat.py",
-    "line": 29,
+    "line": 33,
     "value": {
       "!.minichain.registries.manifests.RegistryManifest": {
         "module": "ommlds.minichain.backends.impls.google.chat",

ommlds/backends/google/protocol/__init__.py

@@ -0,0 +1,3 @@
+from omlish import marshal as _msh  # noqa
+
+_msh.register_global_module_import('._marshal', __package__)

ommlds/backends/google/protocol/_marshal.py

@@ -0,0 +1,16 @@
+from omlish import lang
+from omlish import marshal as msh
+
+from .types import Value
+
+
+##
+
+
+@lang.static_init
+def _install_standard_marshalling() -> None:
+    msh.install_standard_factories(
+        *msh.standard_polymorphism_factories(
+            msh.polymorphism_from_subclasses(Value),
+        ),
+    )

ommlds/backends/google/protocol/types.py

@@ -1,3 +1,6 @@
+"""
+https://ai.google.dev/api/generate-content
+"""
 import typing as ta
 
 from omlish import dataclasses as dc
@@ -8,19 +11,36 @@ from omlish import marshal as msh
 ##
 
 
+def _set_class_marshal_options(cls):
+    msh.update_object_metadata(
+        cls,
+        field_naming=msh.Naming.LOW_CAMEL,
+        field_defaults=msh.FieldMetadata(
+            options=msh.FieldOptions(
+                omit_if=lang.is_none,
+            ),
+        ),
+    )
+
+    return cls
+
+
 @dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
-@msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+@_set_class_marshal_options
+@msh.update_fields_metadata(
+    ['data'],
+    marshaler=msh.Base64MarshalerUnmarshaler(bytes),
+    unmarshaler=msh.Base64MarshalerUnmarshaler(bytes),
+)
 class Blob(lang.Final):
     mine_type: str
     data: bytes
 
 
 @dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
-@msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+@_set_class_marshal_options
 class FunctionCall(lang.Final):
-    id: str
+    id: str | None = None
     name: str
     args: ta.Mapping[str, ta.Any] | None = None
 
@@ -42,10 +62,9 @@ Scheduling: ta.TypeAlias = ta.Literal[
 
 
 @dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
-@msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+@_set_class_marshal_options
 class FunctionResponse(lang.Final):
-    id: str
+    id: str | None = None
     name: str
     response: ta.Mapping[str, ta.Any] | None = None
     will_continue: bool | None = None
@@ -53,8 +72,7 @@ class FunctionResponse(lang.Final):
 
 
 @dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
-@msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+@_set_class_marshal_options
 class FileData(lang.Final):
     mime_type: str
     file_uri: str
@@ -70,8 +88,7 @@ Language: ta.TypeAlias = ta.Literal[
 
 
 @dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
-@msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+@_set_class_marshal_options
 class ExecutableCode(lang.Final):
     language: Language
     code: str
@@ -93,16 +110,14 @@ Outcome: ta.TypeAlias = ta.Literal[
 
 
 @dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
-@msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+@_set_class_marshal_options
 class CodeExecutionResult(lang.Final):
     outcome: Outcome
     output: str
 
 
 @dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
-@msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+@_set_class_marshal_options
 class VideoMetadata(lang.Final):
     start_offset: str  # Duration
     end_offset: str  # Duration
@@ -110,8 +125,12 @@ class VideoMetadata(lang.Final):
 
 
 @dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
-@msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+@msh.update_fields_metadata(
+    ['thought_signature'],
+    marshaler=msh.OptionalMarshaler(msh.Base64MarshalerUnmarshaler(bytes)),
+    unmarshaler=msh.OptionalUnmarshaler(msh.Base64MarshalerUnmarshaler(bytes)),
+)
+@_set_class_marshal_options
 class Part(lang.Final):
     # TODO: data: msh.oneof ...
     text: str | None = None
@@ -130,8 +149,7 @@ class Part(lang.Final):
 
 
 @dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
-@msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+@_set_class_marshal_options
 class Content(lang.Final):
     parts: ta.Sequence[Part] | None = None
     role: ta.Literal['user', 'model'] | None = None
@@ -170,60 +188,51 @@ Type: ta.TypeAlias = ta.Literal[
 Struct: ta.TypeAlias = ta.Mapping[str, 'Value']
 
 
-@dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
-@msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+@dc.dataclass(frozen=True)
 class Value(lang.Abstract, lang.Sealed):
-    pass
+    """https://protobuf.dev/reference/protobuf/google.protobuf/#value"""
 
 
-@dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
+@dc.dataclass(frozen=True)
 @msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
 class NullValue(Value, lang.Final):
     null_value: None = None
 
 
-@dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
-@msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+@dc.dataclass(frozen=True)
+@_set_class_marshal_options
 class NumberValue(Value, lang.Final):
     number_value: float
 
 
-@dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
-@msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+@dc.dataclass(frozen=True)
+@_set_class_marshal_options
 class StringValue(Value, lang.Final):
     string_value: str
 
 
-@dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
-@msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+@dc.dataclass(frozen=True)
+@_set_class_marshal_options
 class BoolValue(Value, lang.Final):
     bool_value: bool
 
 
-@dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
-@msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+@dc.dataclass(frozen=True)
+@_set_class_marshal_options
 class StructValue(Value, lang.Final):
     struct_value: Struct
 
 
-@dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
-@msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+@dc.dataclass(frozen=True)
+@_set_class_marshal_options
 class ListValue(Value, lang.Final):
     list_value: ta.Sequence[Value]
 
 
 @dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
-@msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+@_set_class_marshal_options
 class Schema(lang.Final):
-    type: Type | None = None
+    type: Type | None = None  # FIXME: required
     format: str | None = None
     title: str | None = None
     description: str | None = None
@@ -261,19 +270,18 @@ FunctionBehavior: ta.TypeAlias = ta.Literal[
 
 
 @dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
-@msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+@_set_class_marshal_options
 class FunctionDeclaration(lang.Final):
     name: str
     description: str
 
-    behavior: FunctionBehavior
+    behavior: FunctionBehavior | None = None
 
-    parameters: Schema
-    parameters_json_schema: Value
+    parameters: Schema | None = None
+    parameters_json_schema: Value | None = None
 
-    response: Schema
-    response_json_schema: Value
+    response: Schema | None = None
+    response_json_schema: Value | None = None
 
 
 DynamicRetrievalMode: ta.TypeAlias = ta.Literal[
@@ -286,8 +294,7 @@ DynamicRetrievalMode: ta.TypeAlias = ta.Literal[
 
 
 @dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
-@msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+@_set_class_marshal_options
 class DynamicRetrievalConfig(lang.Final):
     mode: DynamicRetrievalMode | None = None
 
@@ -295,44 +302,38 @@ class DynamicRetrievalConfig(lang.Final):
 
 
 @dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
-@msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+@_set_class_marshal_options
 class GoogleSearchRetrieval(lang.Final):
     dynamic_retrieval_config: DynamicRetrievalConfig
 
 
 @dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
-@msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+@_set_class_marshal_options
 class CodeExecution(lang.Final):
     pass
 
 
 @dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
-@msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+@_set_class_marshal_options
 class Interval(lang.Final):
     start_time: str  # Timestamp
     end_time: str  # Timestamp
 
 
 @dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
-@msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+@_set_class_marshal_options
 class GoogleSearch(lang.Final):
     time_range_filter: Interval | None = None
 
 
 @dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
-@msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+@_set_class_marshal_options
 class UrlContext(lang.Final):
     pass
 
 
 @dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
-@msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+@_set_class_marshal_options
 class Tool(lang.Final):
     function_declarations: ta.Sequence[FunctionDeclaration] | None = None
     google_search_retrieval: GoogleSearchRetrieval | None = None
@@ -341,34 +342,261 @@ class Tool(lang.Final):
     url_context: UrlContext | None = None
 
 
+FunctionCallingMode: ta.TypeAlias = ta.Literal[
+    # Unspecified function calling mode. This value should not be used.
+    'MODE_UNSPECIFIED',
+
+    # Default model behavior, model decides to predict either a function call or a natural language response.
+    'AUTO',
+
+    # Model is constrained to always predicting a function call only. If "allowedFunctionNames" are set, the predicted
+    # function call will be limited to any one of "allowedFunctionNames", else the predicted function call will be any
+    # one of the provided "functionDeclarations".
+    'ANY',
+
+    # Model will not predict any function call. Model behavior is same as when not passing any function declarations.
+    'NONE',
+
+    # Model decides to predict either a function call or a natural language response, but will validate function calls
+    # with constrained decoding. If "allowedFunctionNames" are set, the predicted function call will be limited to any
+    # one of "allowedFunctionNames", else the predicted function call will be any one of the provided
+    # "functionDeclarations".
+    'VALIDATED',
+]
+
+
 @dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
-@msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+@_set_class_marshal_options
+class FunctionCallingConfig(lang.Final):
+    mode: FunctionCallingMode | None = None
+    allowed_function_names: ta.Sequence[str] | None = None
+
+
+@dc.dataclass(frozen=True, kw_only=True)
+@_set_class_marshal_options
+class ToolConfig(lang.Final):
+    function_calling_config: FunctionCallingConfig | None = None
+
+
+HarmCategory: ta.TypeAlias = ta.Literal[
+    # Category is unspecified.
+    'HARM_CATEGORY_UNSPECIFIED',
+
+    # PaLM - Negative or harmful comments targeting identity and/or protected attribute.
+    'HARM_CATEGORY_DEROGATORY',
+
+    # PaLM - Content that is rude, disrespectful, or profane.
+    'HARM_CATEGORY_TOXICITY',
+
+    # PaLM - Describes scenarios depicting violence against an individual or group, or general descriptions of gore.
+    'HARM_CATEGORY_VIOLENCE',
+
+    # PaLM - Contains references to sexual acts or other lewd content.
+    'HARM_CATEGORY_SEXUAL',
+
+    # PaLM - Promotes unchecked medical advice.
+    'HARM_CATEGORY_MEDICAL',
+
+    # PaLM - Dangerous content that promotes, facilitates, or encourages harmful acts.
+    'HARM_CATEGORY_DANGEROUS',
+
+    # Gemini - Harassment content.
+    'HARM_CATEGORY_HARASSMENT',
+
+    # Gemini - Hate speech and content.
+    'HARM_CATEGORY_HATE_SPEECH',
+
+    # Gemini - Sexually explicit content.
+    'HARM_CATEGORY_SEXUALLY_EXPLICIT',
+
+    # Gemini - Dangerous content.
+    'HARM_CATEGORY_DANGEROUS_CONTENT',
+
+    # Gemini - Content that may be used to harm civic integrity. DEPRECATED: use enableEnhancedCivicAnswers instead.
+    'HARM_CATEGORY_CIVIC_INTEGRITY',
+]
+
+
+HarmBlockThreshold: ta.TypeAlias = ta.Literal[
+    # Threshold is unspecified.
+    'HARM_BLOCK_THRESHOLD_UNSPECIFIED',
+
+    # Content with NEGLIGIBLE will be allowed.
+    'BLOCK_LOW_AND_ABOVE',
+
+    # Content with NEGLIGIBLE and LOW will be allowed.
+    'BLOCK_MEDIUM_AND_ABOVE',
+
+    # Content with NEGLIGIBLE, LOW, and MEDIUM will be allowed.
+    'BLOCK_ONLY_HIGH',
+
+    # All content will be allowed.
+    'BLOCK_NONE',
+
+    # Turn off the safety filter.
+    'OFF',
+]
+
+
+@dc.dataclass(frozen=True, kw_only=True)
+@_set_class_marshal_options
+class SafetySetting(lang.Final):
+    category: HarmCategory
+    threshold: HarmBlockThreshold
+
+
+@dc.dataclass(frozen=True, kw_only=True)
+@_set_class_marshal_options
+class ThinkingConfig(lang.Final):
+    include_thoughts: bool | None = None
+    thinking_budget: int | None = None
+
+
+Modality: ta.TypeAlias = ta.Literal[
+    # Default value.
+    'MODALITY_UNSPECIFIED',
+
+    # Indicates the model should return text.
+    'TEXT',
+
+    # Indicates the model should return images.
+    'IMAGE',
+
+    # Indicates the model should return audio.
+    'AUDIO',
+]
+
+
+MediaResolution: ta.TypeAlias = ta.Literal[
+    # Media resolution has not been set.
+    'MEDIA_RESOLUTION_UNSPECIFIED',
+
+    # Media resolution set to low (64 tokens).
+    'MEDIA_RESOLUTION_LOW',
+
+    # Media resolution set to medium (256 tokens).
+    'MEDIA_RESOLUTION_MEDIUM',
+
+    # Media resolution set to high (zoomed reframing with 256 tokens).
+    'MEDIA_RESOLUTION_HIGH',
+]
+
+
+@dc.dataclass(frozen=True, kw_only=True)
+@_set_class_marshal_options
+class GenerationConfig(lang.Final):
+    stop_sequences: ta.Sequence[str] | None = None
+
+    response_mime_type: str | None = None
+    response_schema: Schema | None = None
+    response_json_schema: Value | None = None
+    response_modalities: ta.Sequence[Modality] | None = None
+
+    candidate_count: int | None = None
+    max_output_tokens: int | None = None
+    temperature: float | None = None
+    top_p: float | None = None
+    top_k: int | None = None
+    seed: int | None = None
+    presence_penalty: float | None = None
+    frequency_penalty: float | None = None
+
+    response_logprobs: bool | None = None
+    logprobs: int | None = None
+
+    enable_enhanced_civic_answers: bool | None = None
+
+    # speech_config: SpeechConfig | None = None
+
+    thinking_config: ThinkingConfig | None = None
+
+    media_resolution: MediaResolution | None = None
+
+
+@dc.dataclass(frozen=True, kw_only=True)
+@_set_class_marshal_options
 class GenerateContentRequest(lang.Final):
     """https://ai.google.dev/api/generate-content#request-body"""
 
     contents: ta.Sequence[Content] | None = None
+    tools: ta.Sequence[Tool] | None = None
+    tool_config: ToolConfig | None = None
+    safety_settings: ta.Sequence[SafetySetting] | None = None
+    system_instruction: Content | None = None
+    generation_config: GenerationConfig | None = None
+    cached_content: str | None = None
+
+
+FinishReason: ta.TypeAlias = ta.Literal[
+    # Default value. This value is unused.
+    'FINISH_REASON_UNSPECIFIED',
+
+    # Natural stop point of the model or provided stop sequence.
+    'STOP',
+
+    # The maximum number of tokens as specified in the request was reached.
+    'MAX_TOKENS',
+
+    # The response candidate content was flagged for safety reasons.
+    'SAFETY',
+
+    # The response candidate content was flagged for recitation reasons.
+    'RECITATION',
+
+    # The response candidate content was flagged for using an unsupported language.
+    'LANGUAGE',
+
+    # Unknown reason.
+    'OTHER',
+
+    # Token generation stopped because the content contains forbidden terms.
+    'BLOCKLIST',
+
+    # Token generation stopped for potentially containing prohibited content.
+    'PROHIBITED_CONTENT',
+
+    # Token generation stopped because the content potentially contains Sensitive Personally Identifiable Information
+    # (SPII).
+    'SPII',
+
+    # The function call generated by the model is invalid.
+    'MALFORMED_FUNCTION_CALL',
+
+    # Token generation stopped because generated images contain safety violations.
+    'IMAGE_SAFETY',
+
+    # Model generated a tool call but no tools were enabled in the request.
+    'UNEXPECTED_TOOL_CALL',
+
+    # Model called too many tools consecutively, thus the system exited execution.
+    'TOO_MANY_TOOL_CALLS',
+]
 
 
 @dc.dataclass(frozen=True, kw_only=True)
-@msh.update_fields_metadata(omit_if=lang.is_none)
-@msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+@_set_class_marshal_options
 class GenerateContentResponse(lang.Final):
     """https://ai.google.dev/api/generate-content#v1beta.GenerateContentResponse"""
 
     @dc.dataclass(frozen=True, kw_only=True)
-    @msh.update_fields_metadata(omit_if=lang.is_none)
-    @msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+    @_set_class_marshal_options
     class Candidate(lang.Final):
         content: Content | None = None
-        finish_reason: ta.Literal['STOP'] | None = None
+        finish_reason: FinishReason | None = None
+        # safety_ratings: ta.Sequence[SafetyRating] | None = None
+        # citation_metadata: CitationMetadata | None = None
+        token_count: int | None = None
+        # grounding_attributions: ta.Sequence[GroundingAttribution] | None = None
+        # grounding_metadata: GroundingMetadata | None = None
+        avg_logprobs: float | None = None
+        # logprobs_result: LogprobsResult | None = None
+        # url_context_metadata: UrlContextMetadata | None = None
         index: int | None = None
 
     candidates: ta.Sequence[Candidate] | None = None
 
     @dc.dataclass(frozen=True, kw_only=True)
-    @msh.update_fields_metadata(omit_if=lang.is_none)
-    @msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+    @_set_class_marshal_options
     class UsageMetadata(lang.Final):
         prompt_token_count: int | None = None
         cached_content_token_count: int | None = None
@@ -377,8 +605,7 @@ class GenerateContentResponse(lang.Final):
         thoughts_token_count: int | None = None
 
         @dc.dataclass(frozen=True, kw_only=True)
-        @msh.update_fields_metadata(omit_if=lang.is_none)
-        @msh.update_object_metadata(field_naming=msh.Naming.LOW_CAMEL)
+        @_set_class_marshal_options
         class ModalityTokenCount:
             modality: str | None = None
             token_count: int | None = None

ommlds/backends/mlx/generation.py

@@ -254,7 +254,7 @@ def stream_generate(
                 add_special_tokens=add_special_tokens,
             )
 
-        prompt = mx.array(prompt)  # type: ignore[arg-type]
+        prompt = mx.array(ta.cast(ta.Any, prompt))
 
     detokenizer = tokenization.detokenizer
     detokenizer.reset()