@fugood/bricks-ctor 2.24.0-beta.40

package/types/generators/LlmMediaTekNeuroPilot.ts

@@ -0,0 +1,235 @@
+/* Auto generated by build script
+ *
+ * On-device LLM inference using MediaTek NeuroPilot native SDK integration on Android
+ */
+import type { SwitchCondInnerStateCurrentCanvas, SwitchCondData, SwitchDef } from '../switch'
+import type { Data, DataLink } from '../data'
+import type {
+  Brick,
+  Generator,
+  EventAction,
+  ActionWithDataParams,
+  ActionWithParams,
+  Action,
+  EventProperty,
+} from '../common'
+import type { TemplateEventPropsMap } from '../../utils/event-props'
+
+/* Load or validate the NeuroPilot model context */
+export type GeneratorNeuropilotLlmActionLoadModel = ActionWithParams & {
+  __actionName: 'GENERATOR_NEUROPILOT_LLM_LOAD_MODEL'
+  params?: Array<
+    | {
+        input: 'runnerPath'
+        value?: string | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'configPath'
+        value?: string | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'workingDirectory'
+        value?: string | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'libraryPaths'
+        value?: Array<any> | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'daemonHost'
+        value?: string | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'daemonPort'
+        value?: number | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'daemonSocketName'
+        value?: string | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'daemonSocketNamespace'
+        value?: string | DataLink | EventProperty
+        mapping?: string
+      }
+  >
+}
+
+/* Run text generation with the current NeuroPilot context */
+export type GeneratorNeuropilotLlmActionGenerate = ActionWithParams & {
+  __actionName: 'GENERATOR_NEUROPILOT_LLM_GENERATE'
+  params?: Array<
+    | {
+        input: 'prompt'
+        value?: string | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'messages'
+        value?: Array<any> | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'maxNewTokens'
+        value?: number | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'preformatter'
+        value?: string | DataLink | EventProperty
+        mapping?: string
+      }
+  >
+}
+
+/* Abort an in-flight NeuroPilot generation request */
+export type GeneratorNeuropilotLlmActionAbortGeneration = Action & {
+  __actionName: 'GENERATOR_NEUROPILOT_LLM_ABORT_GENERATION'
+}
+
+/* Release the current NeuroPilot context */
+export type GeneratorNeuropilotLlmActionReleaseContext = Action & {
+  __actionName: 'GENERATOR_NEUROPILOT_LLM_RELEASE_CONTEXT'
+}
+
+interface GeneratorNeuropilotLlmDef {
+  /*
+Default property:
+{
+  "runnerPath": "/data/local/tmp/llm_sdk/main",
+  "configPath": "/data/local/tmp/llm_sdk/config_gemma2_2b_instruct.yaml",
+  "workingDirectory": "/data/local/tmp/llm_sdk",
+  "libraryPaths": [
+    "/vendor/lib64",
+    "/system_ext/lib64",
+    "/vendor/lib",
+    "/system_ext/lib"
+  ],
+  "runtimeMode": "auto",
+  "preformatter": "GemmaNoInput",
+  "maxNewTokens": 128
+}
+  */
+  property?: {
+    /* Preloadable NeuroPilot model bundle preset */
+    modelBundle?:
+      | 'Gemma 2 2B Instruct'
+      | 'Gemma 2 2B Instruct (Tailpatched)'
+      | 'Gemma 2 2B Instruct (Tailpatched MDLA53)'
+      | 'Qwen 2.5 0.5B Instruct'
+      | 'Qwen 2.5 1.5B Instruct'
+      | 'Llama 3 8B Instruct'
+      | DataLink
+    /* Override base URL for NeuroPilot model bundle downloads */
+    modelBaseUrl?: string | DataLink
+    /* Validate runner/config paths on generator initialization */
+    init?: boolean | DataLink
+    /* Runner binary path on device */
+    runnerPath?: string | DataLink
+    /* YAML config path on device */
+    configPath?: string | DataLink
+    /* Working directory for the runner process */
+    workingDirectory?: string | DataLink
+    /* Extra library search paths for the runner process */
+    libraryPaths?: Array<string | DataLink> | DataLink
+    /* Daemon host for a preloaded NeuroPilot service */
+    daemonHost?: string | DataLink
+    /* Daemon TCP port for a preloaded NeuroPilot service */
+    daemonPort?: number | DataLink
+    /* Android local socket name for a privileged NeuroPilot service */
+    daemonSocketName?: string | DataLink
+    /* Android local socket namespace used by the privileged NeuroPilot service */
+    daemonSocketNamespace?: 'abstract' | 'reserved' | 'filesystem' | DataLink
+    /* Runtime selection strategy for MediaTek execution */
+    runtimeMode?: 'auto' | 'daemon' | 'root-runner' | 'native' | DataLink
+    /* Prompt preformatter used by the vendor runner */
+    preformatter?:
+      | 'AlpacaNoInput'
+      | 'OneShotConversation'
+      | 'VicunaNoInput'
+      | 'QwenNoInput'
+      | 'Qwen3NoInput'
+      | 'Qwen3NoInputNoThink'
+      | 'Llama3NoInput'
+      | 'Phi3NoInput'
+      | 'MinicpmNoInput'
+      | 'MinicpmNoInputZh'
+      | 'InternLM2'
+      | 'GemmaNoInput'
+      | DataLink
+    /* Prompt to run */
+    prompt?: string | DataLink
+    /* Chat messages to flatten into a prompt */
+    messages?: Array<DataLink | {}> | DataLink
+    /* Maximum tokens requested from the runner */
+    maxNewTokens?: number | DataLink
+  }
+  events?: {
+    /* Event triggered when the NeuroPilot context state changes */
+    onContextStateChange?: Array<
+      EventAction<string & keyof TemplateEventPropsMap['NeuropilotLlm']['onContextStateChange']>
+    >
+    /* Event triggered when a completion token or partial result is emitted */
+    onCompletion?: Array<
+      EventAction<string & keyof TemplateEventPropsMap['NeuropilotLlm']['onCompletion']>
+    >
+    /* Event triggered when generation finishes */
+    onCompletionFinished?: Array<
+      EventAction<string & keyof TemplateEventPropsMap['NeuropilotLlm']['onCompletionFinished']>
+    >
+    /* Event triggered when a NeuroPilot error occurs */
+    onError?: Array<EventAction<string & keyof TemplateEventPropsMap['NeuropilotLlm']['onError']>>
+  }
+  outlets?: {
+    /* Current NeuroPilot context state */
+    contextState?: () => Data<string>
+    /* Final generated result text */
+    result?: () => Data<string>
+    /* Full context returned by the runner */
+    fullContext?: () => Data<string>
+    /* Last emitted token or chunk */
+    lastToken?: () => Data<string>
+    /* Raw output captured from the NeuroPilot runner */
+    rawOutput?: () => Data<string>
+    /* Prompt-phase performance in tokens per second */
+    promptTokensPerSec?: () => Data<string>
+    /* Generation-phase performance in tokens per second */
+    generationTokensPerSec?: () => Data<string>
+  }
+}
+
+/* On-device LLM inference using MediaTek NeuroPilot native SDK integration on Android */
+export type GeneratorNeuropilotLlm = Generator &
+  GeneratorNeuropilotLlmDef & {
+    templateKey: 'GENERATOR_NEUROPILOT_LLM'
+    switches?: Array<
+      SwitchDef &
+        GeneratorNeuropilotLlmDef & {
+          conds?: Array<{
+            method: '==' | '!=' | '>' | '<' | '>=' | '<='
+            cond:
+              | SwitchCondInnerStateCurrentCanvas
+              | SwitchCondData
+              | {
+                  __typename: 'SwitchCondInnerStateOutlet'
+                  outlet:
+                    | 'contextState'
+                    | 'result'
+                    | 'fullContext'
+                    | 'lastToken'
+                    | 'rawOutput'
+                    | 'promptTokensPerSec'
+                    | 'generationTokensPerSec'
+                  value: any
+                }
+          }>
+        }
+    >
+  }

package/types/generators/LlmMlx.ts

@@ -0,0 +1,227 @@
+/* Auto generated by build script
+ *
+ * On-device LLM inference using Apple MLX framework on iOS/tvOS
+ *
+ *  ## Features
+ *  - Powered by MLX (Apple's ML framework optimized for Apple Silicon)
+ *  - Download models directly from HuggingFace Hub
+ *  - Streaming token generation
+ *  - Supports LLM and VLM (Vision Language Models)
+ *  - Requires iOS 17+ or tvOS 17+
+ */
+import type { SwitchCondInnerStateCurrentCanvas, SwitchCondData, SwitchDef } from '../switch'
+import type { Data, DataLink } from '../data'
+import type {
+  Brick,
+  Generator,
+  EventAction,
+  ActionWithDataParams,
+  ActionWithParams,
+  Action,
+  EventProperty,
+} from '../common'
+import type { TemplateEventPropsMap } from '../../utils/event-props'
+
+/* Load model */
+export type GeneratorMlxLLMActionLoadModel = ActionWithParams & {
+  __actionName: 'GENERATOR_MLX_LLM_LOAD_MODEL'
+  params?: Array<
+    | {
+        input: 'modelId'
+        value?: string | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'vlm'
+        value?: boolean | DataLink | EventProperty
+        mapping?: string
+      }
+  >
+}
+
+/* Run text completion */
+export type GeneratorMlxLLMActionCompletion = ActionWithParams & {
+  __actionName: 'GENERATOR_MLX_LLM_COMPLETION'
+  params?: Array<
+    | {
+        input: 'messages'
+        value?: Array<any> | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'tools'
+        value?: Array<any> | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'maxTokens'
+        value?: number | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'temperature'
+        value?: number | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'topP'
+        value?: number | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'repetitionPenalty'
+        value?: number | DataLink | EventProperty
+        mapping?: string
+      }
+  >
+}
+
+/* Stop text completion */
+export type GeneratorMlxLLMActionStopCompletion = Action & {
+  __actionName: 'GENERATOR_MLX_LLM_STOP_COMPLETION'
+}
+
+/* Release model context */
+export type GeneratorMlxLLMActionReleaseContext = Action & {
+  __actionName: 'GENERATOR_MLX_LLM_RELEASE_CONTEXT'
+}
+
+interface GeneratorMlxLLMDef {
+  /*
+Default property:
+{
+  "init": false,
+  "modelId": "mlx-community/Qwen3-4B-4bit",
+  "vlm": false,
+  "completionMessages": [
+    {
+      "role": "system",
+      "content": "You are a helpful assistant."
+    },
+    {
+      "role": "user",
+      "content": "Hello"
+    }
+  ],
+  "completionMaxTokens": 1024,
+  "completionTemperature": 0.6,
+  "completionTopP": 1,
+  "completionRepetitionContextSize": 20
+}
+  */
+  property?: {
+    /* Initialize model on generator init */
+    init?: boolean | DataLink
+    /* HuggingFace model ID or local path to model directory
+ e.g. "mlx-community/Qwen3-4B-4bit" */
+    modelId?: string | DataLink
+    /* Enable Vision Language Model (VLM) mode */
+    vlm?: boolean | DataLink
+    /* Chat messages (if first message has role 'system', it will be used as system prompt) */
+    completionMessages?:
+      | Array<
+          | DataLink
+          | {
+              role?: string | DataLink
+              content?: string | DataLink
+            }
+        >
+      | DataLink
+    /* Maximum tokens to generate */
+    completionMaxTokens?: number | DataLink
+    /* Temperature (0.0 to 2.0) */
+    completionTemperature?: number | DataLink
+    /* Top P sampling */
+    completionTopP?: number | DataLink
+    /* Repetition penalty factor */
+    completionRepetitionPenalty?: number | DataLink
+    /* Number of tokens to consider for repetition penalty */
+    completionRepetitionContextSize?: number | DataLink
+    /* Maximum KV cache size (context window). Uses rotating cache to limit memory.
+ Smaller values (e.g. 512) use less memory but lower quality.  
+ Larger values (e.g. 4096) use more memory but better quality.  
+ Leave empty for unlimited (model default). */
+    contextSize?: number | DataLink
+    /* Enable thinking mode (model-dependent, e.g. Qwen3).
+ When enabled, passes enable_thinking=true to the chat template.  
+ The model may output reasoning in &lt;think&gt; tags which will be extracted as reasoning_content. */
+    completionEnableThinking?: boolean | DataLink
+    /* Additional keyword arguments for chat template (object) */
+    completionChatTemplateKwargs?: {} | DataLink
+    /* Tools for chat mode using OpenAI-compatible function calling format
+ Format: Array of objects with {type, function: {name, description, parameters}} structure  
+ See: https://platform.openai.com/docs/guides/function-calling */
+    completionTools?: Array<{} | DataLink> | DataLink
+    /* Buttress connection settings for remote inference */
+    buttressConnectionSettings?:
+      | DataLink
+      | {
+          enabled?: boolean | DataLink
+          url?: string | DataLink
+          fallbackType?: 'use-local' | 'no-op' | DataLink
+          strategy?: 'prefer-local' | 'prefer-buttress' | 'prefer-best' | DataLink
+        }
+  }
+  events?: {
+    /* Event triggered when context state changes */
+    onContextStateChange?: Array<
+      EventAction<string & keyof TemplateEventPropsMap['MlxLlm']['onContextStateChange']>
+    >
+    /* Error event */
+    onError?: Array<EventAction<string & keyof TemplateEventPropsMap['MlxLlm']['onError']>>
+    /* Completion streaming event (emitted for each token) */
+    onCompletion?: Array<
+      EventAction<string & keyof TemplateEventPropsMap['MlxLlm']['onCompletion']>
+    >
+    /* Completion finished event */
+    onCompletionFinished?: Array<
+      EventAction<string & keyof TemplateEventPropsMap['MlxLlm']['onCompletionFinished']>
+    >
+  }
+  outlets?: {
+    /* Context state */
+    contextState?: () => Data<string>
+    /* Model load progress (0-1) */
+    loadProgress?: () => Data<number>
+    /* Whether the model is evaluating */
+    isEvaluating?: () => Data<boolean>
+    /* Completion result */
+    completionResult?: () => Data<string>
+    /* Last token */
+    completionLastToken?: () => Data<string>
+  }
+}
+
+/* On-device LLM inference using Apple MLX framework on iOS/tvOS
+
+ ## Features  
+ - Powered by MLX (Apple's ML framework optimized for Apple Silicon)  
+ - Download models directly from HuggingFace Hub  
+ - Streaming token generation  
+ - Supports LLM and VLM (Vision Language Models)  
+ - Requires iOS 17+ or tvOS 17+ */
+export type GeneratorMlxLLM = Generator &
+  GeneratorMlxLLMDef & {
+    templateKey: 'GENERATOR_MLX_LLM'
+    switches?: Array<
+      SwitchDef &
+        GeneratorMlxLLMDef & {
+          conds?: Array<{
+            method: '==' | '!=' | '>' | '<' | '>=' | '<='
+            cond:
+              | SwitchCondInnerStateCurrentCanvas
+              | SwitchCondData
+              | {
+                  __typename: 'SwitchCondInnerStateOutlet'
+                  outlet:
+                    | 'contextState'
+                    | 'loadProgress'
+                    | 'isEvaluating'
+                    | 'completionResult'
+                    | 'completionLastToken'
+                  value: any
+                }
+          }>
+        }
+    >
+  }

package/types/generators/LlmOnnx.ts

@@ -0,0 +1,213 @@
+/* Auto generated by build script
+ *
+ * Local LLM inference based on [transformers.js](https://huggingface.co/docs/transformers.js)
+ *  You can use any converted model on HuggingFace.
+ */
+import type { SwitchCondInnerStateCurrentCanvas, SwitchCondData, SwitchDef } from '../switch'
+import type { Data, DataLink } from '../data'
+import type {
+  Brick,
+  Generator,
+  EventAction,
+  ActionWithDataParams,
+  ActionWithParams,
+  Action,
+  EventProperty,
+} from '../common'
+import type { TemplateEventPropsMap } from '../../utils/event-props'
+
+/* Load the model */
+export type GeneratorOnnxLLMActionLoadModel = Action & {
+  __actionName: 'GENERATOR_ONNX_LLM_LOAD_MODEL'
+}
+
+/* Inference */
+export type GeneratorOnnxLLMActionInfer = ActionWithParams & {
+  __actionName: 'GENERATOR_ONNX_LLM_INFER'
+  params?: Array<
+    | {
+        input: 'prompt'
+        value?: string | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'chat'
+        value?: Array<any> | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'images'
+        value?: Array<any> | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'audios'
+        value?: Array<any> | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'tools'
+        value?: Array<any> | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'toolChoice'
+        value?: string | DataLink | EventProperty
+        mapping?: string
+      }
+  >
+}
+
+/* Clean cache */
+export type GeneratorOnnxLLMActionCleanCache = Action & {
+  __actionName: 'GENERATOR_ONNX_LLM_CLEAN_CACHE'
+}
+
+/* Release context */
+export type GeneratorOnnxLLMActionReleaseContext = Action & {
+  __actionName: 'GENERATOR_ONNX_LLM_RELEASE_CONTEXT'
+}
+
+interface GeneratorOnnxLLMDef {
+  /*
+Default property:
+{
+  "modelType": "auto",
+  "toolCallParser": "llama3_json",
+  "toolChoice": "auto",
+  "maxNewTokens": 256,
+  "temperature": 0.7,
+  "topK": 50,
+  "topP": 0.9,
+  "repetitionPenalty": 1,
+  "noRepeatNgramSize": 0,
+  "numBeams": 1,
+  "doSample": true,
+  "executionMode": "sequential"
+}
+  */
+  property?: {
+    /* Initialize the TTS context on generator initialization */
+    init?: boolean | DataLink
+    /* LLM model */
+    model?: string | DataLink
+    /* Model type */
+    modelType?: string | DataLink
+    /* Quantize type */
+    quantizeType?:
+      | 'auto'
+      | 'none'
+      | 'fp16'
+      | 'q8'
+      | 'int8'
+      | 'uint8'
+      | 'q4'
+      | 'bnb4'
+      | 'q4f16'
+      | DataLink
+    /* Prompt to inference */
+    prompt?: string | DataLink
+    /* Messages to inference */
+    messages?: Array<DataLink | {}> | DataLink
+    /* Images with message to inference */
+    images?: Array<string | DataLink> | DataLink
+    /* Audios with message to inference */
+    audios?: Array<string | DataLink> | DataLink
+    /* Tool call parser */
+    toolCallParser?: 'llama3_json' | 'mistral' | 'hermes' | 'internlm' | 'phi4' | DataLink
+    /* Tools for chat mode using OpenAI-compatible function calling format
+ Format: Array of objects with {type, function: {name, description, parameters}} structure  
+ See: https://platform.openai.com/docs/guides/function-calling */
+    tools?: Array<{} | DataLink> | DataLink
+    /* Tool choice for chat mode */
+    toolChoice?: 'none' | 'auto' | DataLink
+    /* Max new tokens to generate */
+    maxNewTokens?: number | DataLink
+    /* Temperature */
+    temperature?: number | DataLink
+    /* Top k */
+    topK?: number | DataLink
+    /* Top p */
+    topP?: number | DataLink
+    /* Repetition penalty */
+    repetitionPenalty?: number | DataLink
+    /* No repeat ngram size */
+    noRepeatNgramSize?: number | DataLink
+    /* Number of beams */
+    numBeams?: number | DataLink
+    /* Do sampling */
+    doSample?: boolean | DataLink
+    /* Executor candidates, descending order of priority
+  Default will be xnnpack, wasm, cpu */
+    executors?:
+      | Array<'qnn' | 'dml' | 'nnapi' | 'xnnpack' | 'coreml' | 'cpu' | 'wasm' | 'webgpu' | DataLink>
+      | DataLink
+    /* Execution mode
+  Usually when the model has many branches, setting this option to `parallel` will give you better performance. */
+    executionMode?: 'sequential' | 'parallel' | DataLink
+    /* QNN backend */
+    qnnBackend?: 'HTP' | 'HTA' | 'DSP' | 'GPU' | 'CPU' | DataLink
+    /* Enable FP16 for QNN HTP */
+    qnnHtpEnableFp16?: boolean | DataLink
+    /* Enable QNN debug */
+    qnnEnableDebug?: boolean | DataLink
+  }
+  events?: {
+    /* Event triggered when state change */
+    onContextStateChange?: Array<
+      EventAction<string & keyof TemplateEventPropsMap['OnnxLlm']['onContextStateChange']>
+    >
+    /* Event triggered on get function call request */
+    onFunctionCall?: Array<
+      EventAction<string & keyof TemplateEventPropsMap['OnnxLlm']['onFunctionCall']>
+    >
+    /* Event triggered on completion finished */
+    onCompletionFinished?: Array<
+      EventAction<string & keyof TemplateEventPropsMap['OnnxLlm']['onCompletionFinished']>
+    >
+    /* Event triggered when error occurs */
+    onError?: Array<EventAction<string & keyof TemplateEventPropsMap['OnnxLlm']['onError']>>
+  }
+  outlets?: {
+    /* Context state */
+    contextState?: () => Data<string>
+    /* Generated output */
+    generated?: () => Data<string>
+    /* Full result of generation */
+    fullResult?: () => Data<string>
+    /* Last function call */
+    lastFunctionCall?: () => Data<{
+      id?: string
+      type?: string
+      function?: {
+        name?: string
+        arguments?: string
+        [key: string]: any
+      }
+      [key: string]: any
+    }>
+  }
+}
+
+/* Local LLM inference based on [transformers.js](https://huggingface.co/docs/transformers.js)
+ You can use any converted model on HuggingFace. */
+export type GeneratorOnnxLLM = Generator &
+  GeneratorOnnxLLMDef & {
+    templateKey: 'GENERATOR_ONNX_LLM'
+    switches?: Array<
+      SwitchDef &
+        GeneratorOnnxLLMDef & {
+          conds?: Array<{
+            method: '==' | '!=' | '>' | '<' | '>=' | '<='
+            cond:
+              | SwitchCondInnerStateCurrentCanvas
+              | SwitchCondData
+              | {
+                  __typename: 'SwitchCondInnerStateOutlet'
+                  outlet: 'contextState' | 'generated' | 'fullResult' | 'lastFunctionCall'
+                  value: any
+                }
+          }>
+        }
+    >
+  }