@fugood/bricks-project 2.24.0-beta.4 → 2.24.0-beta.41

package/types/generators/LlmGgml.ts

@@ -1,4 +1,17 @@
-/* Auto generated by build script */
+/* Auto generated by build script
+ *
+ * Local Large Language Model (LLM) inference based on GGML and [llama.cpp](https://github.com/ggerganov/llama.cpp)
+ *
+ *  ## Notice
+ *  - The device RAM must be larger than 8GB
+ *  - iOS: Recommended use M1+ / A17+ chip device. Supported GPU acceleration by Metal.
+ *  - macOS: Recommended use M1+ chip device. Supported GPU acceleration by Metal.
+ *  - Android: Recommended use Android 13+ system.
+ *    - Supported GPU acceleration by OpenCL, currently only for Qualcomm Adreno 700+ GPUs, other GPUs are not supported.
+ *    - Supported Hexagon NPU for Qualcomm Snapdragon 8 Gen 1+ GPUs.
+ *  - Linux / Windows  [@nextline    - Supported GPU acceleration, you can choose `vulkan` or `cuda` backend in Accel Variant property
+ *    - Supported Hexagon NPU for Qualcomm Dragonwing IQ9 series+ (Linux)
+ */
 import type { SwitchCondInnerStateCurrentCanvas, SwitchCondData, SwitchDef } from '../switch'
 import type { Data, DataLink } from '../data'
 import type {
@@ -10,6 +23,7 @@ import type {
   Action,
   EventProperty,
 } from '../common'
+import type { TemplateEventPropsMap } from '../../utils/event-props'
 
 /* Load the model */
 export type GeneratorLLMActionLoadModel = Action & {
@@ -97,6 +111,16 @@ export type GeneratorLLMActionProcessPrompt = ActionWithParams & {
         value?: boolean | DataLink | EventProperty
         mapping?: string
       }
+    | {
+        input: 'thinkingBudgetTokens'
+        value?: number | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'thinkingBudgetMessage'
+        value?: string | DataLink | EventProperty
+        mapping?: string
+      }
     | {
         input: 'prompt'
         value?: string | DataLink | EventProperty
@@ -137,6 +161,11 @@ export type GeneratorLLMActionProcessPrompt = ActionWithParams & {
         value?: string | DataLink | EventProperty
         mapping?: string
       }
+    | {
+        input: 'forcePureContent'
+        value?: boolean | DataLink | EventProperty
+        mapping?: string
+      }
   >
 }
 
@@ -179,6 +208,16 @@ export type GeneratorLLMActionCompletion = ActionWithParams & {
         value?: boolean | DataLink | EventProperty
         mapping?: string
       }
+    | {
+        input: 'thinkingBudgetTokens'
+        value?: number | DataLink | EventProperty
+        mapping?: string
+      }
+    | {
+        input: 'thinkingBudgetMessage'
+        value?: string | DataLink | EventProperty
+        mapping?: string
+      }
     | {
         input: 'useReasoningFormat'
         value?: string | DataLink | EventProperty
@@ -224,6 +263,11 @@ export type GeneratorLLMActionCompletion = ActionWithParams & {
         value?: string | DataLink | EventProperty
         mapping?: string
       }
+    | {
+        input: 'forcePureContent'
+        value?: boolean | DataLink | EventProperty
+        mapping?: string
+      }
     | {
         input: 'grammar'
         value?: string | DataLink | EventProperty
@@ -432,6 +476,7 @@ Default property:
   "completionEnableThinking": true,
   "completionAddGenerationPrompt": true,
   "completionChatTemplateKwargs": {},
+  "completionForcePureContent": false,
   "completionUseReasoningFormat": "auto",
   "completionStopWords": [],
   "completionPredict": 400,
@@ -459,10 +504,10 @@ Default property:
 }
   */
   property?: {
-    /* Initialize the Llama context on generator initialization  
+    /* Initialize the Llama context on generator initialization
  Please note that it will take some RAM depending on the model size */
     init?: boolean | DataLink
-    /* The URL or path of model  
+    /* The URL or path of model
  We used GGUF format model, please refer to https://github.com/ggerganov/llama.cpp/tree/master#description */
     modelUrl?: string | DataLink
     /* Hash type of model */
@@ -477,10 +522,10 @@ Default property:
     mmprojHashType?: 'md5' | 'sha256' | 'sha1' | DataLink
     /* Hash of mmproj file (PREVIEW FEATURE) */
     mmprojHash?: string | DataLink
-    /* Minimum tokens for image encoding in multimodal (PREVIEW FEATURE)  
+    /* Minimum tokens for image encoding in multimodal (PREVIEW FEATURE)
  Useful for dynamic resolution models (e.g. Qwen-VL). Default: -1 (auto) */
     imageMinTokens?: number | DataLink
-    /* Maximum tokens for image encoding in multimodal (PREVIEW FEATURE)  
+    /* Maximum tokens for image encoding in multimodal (PREVIEW FEATURE)
  Limit tokens for dynamic resolution models to balance speed vs. detail. Default: -1 (auto) */
     imageMaxTokens?: number | DataLink
     /* Chat Template (Jinja format) to override the default template from model */
@@ -493,13 +538,13 @@ Default property:
     uBatchSize?: number | DataLink
     /* Number of threads */
     maxThreads?: number | DataLink
-    /* Accelerator variant (Only for desktop)  
+    /* Accelerator variant (Only for desktop)
  `default` - CPU / Metal (macOS)  
  `vulkan` - Use Vulkan  
  `cuda` - Use CUDA  
  `snapdragon` - Use OpenCL/Hexagon of Qualcomm Snapdragon */
     accelVariant?: 'default' | 'vulkan' | 'cuda' | 'snapdragon' | DataLink
-    /* Devices. For example:  
+    /* Devices. For example:
   
  Metal or CPU for iOS/tvOS/MacOS  
  OpenCL or CPU for Android  
@@ -514,6 +559,8 @@ Default property:
     useMlock?: boolean | DataLink
     /* Use mmap */
     useMmap?: boolean | DataLink
+    /* Disable extra buffer types for weight repacking. Reduces memory usage at the cost of slower prompt processing. */
+    noExtraBuffs?: boolean | DataLink
     /* Use Flash Attention for inference (Recommended with GPU enabled) */
     useFlashAttn?: 'auto' | 'on' | 'off' | DataLink
     /* KV cache data type for the K (Default: f16) */
@@ -534,7 +581,7 @@ Default property:
     transformScriptCode?: string | DataLink
     /* Variables used in Transform Script (object) */
     transformScriptVariables?: {} | DataLink
-    /* Session save mode  
+    /* Session save mode
  `none` - No session saving  
  `prompt` - Save session when prompt processed  
  `completion` - Save session when completion finished  
@@ -546,7 +593,7 @@ Default property:
     sessionRemain?: number | DataLink
     /* TODO:loran_gqarms_norm_epsrope_freq_baserope_freq_scale */
     completionMode?: 'auto' | 'chat' | 'text' | DataLink
-    /* Tools for chat mode using OpenAI-compatible function calling format  
+    /* 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
@@ -566,7 +613,7 @@ Default property:
       | DataLink
     /* Prompt (text mode) */
     completionPrompt?: string | DataLink
-    /* Media paths to be used in the prompt template (PREVIEW FEATURE)  
+    /* Media paths to be used in the prompt template (PREVIEW FEATURE)
  In prompt, use `<__media__>` for position of media content */
     completionPromptMediaPaths?: Array<string | DataLink> | DataLink
     /* Data to be used in the prompt template (e.g. `Hello ${name}`). Supports nested data, such as `Hello ${user.name}`. */
@@ -588,13 +635,19 @@ Default property:
         }
     /* Enable thinking */
     completionEnableThinking?: boolean | DataLink
+    /* Maximum tokens allowed inside the model's thinking block before forcing it closed. Only applies when chat formatting exposes thinking tags. */
+    completionThinkingBudgetTokens?: number | DataLink
+    /* Message injected before the thinking end tag when the thinking budget is exhausted. */
+    completionThinkingBudgetMessage?: string | DataLink
     /* Add generation prompt */
     completionAddGenerationPrompt?: boolean | DataLink
     /* Now (for fill current date in chat template if supported) */
     completionNow?: string | DataLink
     /* Additional keyword arguments for chat template (object) */
     completionChatTemplateKwargs?: {} | DataLink
-    /* Use reasoning format for enhanced response structure  
+    /* Force pure content (accept any model that has a chat_template without requiring template validation) */
+    completionForcePureContent?: boolean | DataLink
+    /* Use reasoning format for enhanced response structure
  `auto` - Auto-determine the reasoning format of the model  
  `none` - Disable reasoning format */
     completionUseReasoningFormat?: 'auto' | 'none' | DataLink
@@ -666,47 +719,103 @@ Default property:
   }
   events?: {
     /* Event triggered when context state changes */
-    onContextStateChange?: Array<EventAction>
+    onContextStateChange?: Array<
+      EventAction<string & keyof TemplateEventPropsMap['Llm']['onContextStateChange']>
+    >
     /* Event triggered when error occurs */
-    onError?: Array<EventAction>
+    onError?: Array<EventAction<string & keyof TemplateEventPropsMap['Llm']['onError']>>
     /* Event triggered when completion */
-    onCompletion?: Array<EventAction>
+    onCompletion?: Array<EventAction<string & keyof TemplateEventPropsMap['Llm']['onCompletion']>>
     /* Event triggered when completion finished */
-    onCompletionFinished?: Array<EventAction>
+    onCompletionFinished?: Array<
+      EventAction<string & keyof TemplateEventPropsMap['Llm']['onCompletionFinished']>
+    >
     /* Event triggered on get function call request */
-    onCompletionFunctionCall?: Array<EventAction>
+    onCompletionFunctionCall?: Array<
+      EventAction<string & keyof TemplateEventPropsMap['Llm']['onCompletionFunctionCall']>
+    >
   }
   outlets?: {
     /* Context state */
-    contextState?: () => Data
+    contextState?: () => Data<string>
     /* Context load progress (0-100) */
-    contextLoadProgress?: () => Data
+    contextLoadProgress?: () => Data<number>
     /* Context details */
-    contextDetails?: () => Data
+    contextDetails?: () => Data<{
+      state?: string
+      contextId?: string
+      gpu?: string
+      reasonNoGPU?: string
+      model?: { [key: string]: any }
+      isMultimodalEnabled?: boolean
+      [key: string]: any
+    }>
     /* Session details */
-    sessions?: () => Data
+    sessions?: () => Data<{
+      last_session_id?: string
+      sessions?: Array<{
+        id?: string
+        type?: string
+        prompt?: string
+        sessionKey?: string
+        model_instance_id?: string
+        tokens_evaluated?: number
+        t?: number
+        [key: string]: any
+      }>
+      last_custom_session_id?: string
+      custom_sessions?: Array<{
+        id?: string
+        type?: string
+        prompt?: string
+        sessionKey?: string
+        model_instance_id?: string
+        tokens_evaluated?: number
+        t?: number
+        [key: string]: any
+      }>
+      [key: string]: any
+    }>
     /* Is evaluating */
-    isEvaluating?: () => Data
+    isEvaluating?: () => Data<boolean>
     /* Tokenize result */
-    tokenizeResult?: () => Data
+    tokenizeResult?: () => Data<Array<number>>
     /* Detokenize result */
-    detokenizeResult?: () => Data
+    detokenizeResult?: () => Data<string>
     /* Last formatted prompt (messages or prompt) */
-    completionLastFormattedPrompt?: () => Data
+    completionLastFormattedPrompt?: () => Data<string>
     /* Last completion token */
-    completionLastToken?: () => Data
+    completionLastToken?: () => Data<string>
     /* Completion result */
-    completionResult?: () => Data
+    completionResult?: () => Data<string>
     /* Reasoning content from model responses */
-    completionReasoningContent?: () => Data
+    completionReasoningContent?: () => Data<string>
     /* Full context (Prompt + Completion) */
-    completionFullContext?: () => Data
+    completionFullContext?: () => Data<string>
     /* Inference result details */
-    completionResultDetails?: () => Data
+    completionResultDetails?: () => Data<{
+      prompt?: string
+      full_context?: string
+      text?: string
+      content?: string
+      reasoning_content?: string
+      token?: string
+      tool_calls?: Array<{
+        id?: string
+        type?: string
+        function?: {
+          name?: string
+          arguments?: string
+          [key: string]: any
+        }
+        [key: string]: any
+      }>
+      [key: string]: any
+    }>
   }
 }
 
-/* Local Large Language Model (LLM) inference based on GGML and [llama.cpp](https://github.com/ggerganov/llama.cpp)  
+/* Local Large Language Model (LLM) inference based on GGML and [llama.cpp](https://github.com/ggerganov/llama.cpp)
 
  ## Notice  
  - The device RAM must be larger than 8GB  
@@ -720,7 +829,7 @@ Default property:
 export type GeneratorLLM = Generator &
   GeneratorLLMDef & {
     templateKey: 'GENERATOR_LLM'
-    switches: Array<
+    switches?: Array<
       SwitchDef &
         GeneratorLLMDef & {
           conds?: Array<{

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
+                }
+          }>
+        }
+    >
+  }