Safety Request Processing

class SafetyRequestProcessing

Family-specific artifact parsing, chat formatting, and request shaping selected by the model recipe.

Subclassed by trt_edgellm::safety::InternVL3RequestProcessing

Public Functions

virtual ~SafetyRequestProcessing() = default

virtual SafetyStatus configure(

SafetyRuntimeArtifactBundle const &artifacts,

nlohmann::json const &llmConfig,

nlohmann::json const &visualConfig

) noexcept = 0

Load tokenizer/template/embedding configuration and validate artifact compatibility before allocation.

virtual ProcessingConfiguration const &configuration(

) const noexcept = 0

virtual ProcessingCapabilitySet capabilities() noexcept = 0

virtual SafetyStatus prepareRequest(

LLMGenerationRequest const &request,

ProcessingRequest &prepared

) noexcept = 0

Validate all requested features before SDK work; write into prepared, bounded input storage. The returned views remain valid until the next request.

virtual SafetyStatus validateResponse(

LLMGenerationRequest const &request,

LLMGenerationResponse const &response

) const noexcept = 0

Validate caller-owned output capacities before any request execution.

virtual SafetyStatus formatResponse(

ModelInferenceResult const &result,

LLMGenerationResponse &response

) noexcept = 0

Decode text into caller-prepared storage without allocation. Called only after successful generation.

std::unique_ptr<SafetyRequestProcessing> trt_edgellm::safety::createSafetyRequestProcessing(

ModelInferenceDefinition const &definition

)

Internal Initialization factory for the family adapter and recipe-selected capabilities. Returns nullptr when the selected implementations are unavailable; allocation failures may throw.