@qvac/sdk 0.14.1 → 0.16.0

package/dist/schemas/llamacpp-config.d.ts

@@ -1,4 +1,4 @@
-import { z } from "zod";
+import { z } from 'zod';
 /**
  * Upper bound for `reasoning_budget`. Mirrors the llm-llamacpp addon, which
  * stores the budget as a 32-bit `int` and rejects values above
@@ -23,29 +23,35 @@ export declare const llmConfigBaseSchema: z.ZodObject<{
     predict: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<-1>, z.ZodLiteral<-2>, z.ZodNumber]>>;
     system_prompt: z.ZodOptional<z.ZodString>;
     no_mmap: z.ZodOptional<z.ZodBoolean>;
-    verbosity: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<0>, z.ZodLiteral<1>, z.ZodLiteral<2>, z.ZodLiteral<3>]>>;
+    verbosity: z.ZodOptional<z.ZodEnum<{
+        readonly ERROR: 0;
+        readonly WARN: 1;
+        readonly INFO: 2;
+        readonly DEBUG: 3;
+    }>>;
     presence_penalty: z.ZodOptional<z.ZodNumber>;
     frequency_penalty: z.ZodOptional<z.ZodNumber>;
     repeat_penalty: z.ZodOptional<z.ZodNumber>;
     stop_sequences: z.ZodOptional<z.ZodArray<z.ZodString>>;
     n_discarded: z.ZodOptional<z.ZodNumber>;
+    parallel: z.ZodOptional<z.ZodNumber>;
     tools: z.ZodOptional<z.ZodBoolean>;
     toolsMode: z.ZodOptional<z.ZodEnum<{
         static: "static";
         dynamic: "dynamic";
     }>>;
-    "cache-type-k": z.ZodOptional<z.ZodString>;
-    "cache-type-v": z.ZodOptional<z.ZodString>;
-    "main-gpu": z.ZodOptional<z.ZodUnion<readonly [z.ZodNumber, z.ZodEnum<{
+    'cache-type-k': z.ZodOptional<z.ZodString>;
+    'cache-type-v': z.ZodOptional<z.ZodString>;
+    'main-gpu': z.ZodOptional<z.ZodUnion<readonly [z.ZodNumber, z.ZodEnum<{
         integrated: "integrated";
         dedicated: "dedicated";
     }>]>>;
-    "split-mode": z.ZodOptional<z.ZodEnum<{
+    'split-mode': z.ZodOptional<z.ZodEnum<{
         none: "none";
         layer: "layer";
         row: "row";
     }>>;
-    "tensor-split": z.ZodOptional<z.ZodString>;
+    'tensor-split': z.ZodOptional<z.ZodString>;
     openclCacheDir: z.ZodOptional<z.ZodString>;
     reasoning_budget: z.ZodOptional<z.ZodNumber>;
     projectionModelSrc: z.ZodOptional<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
@@ -90,6 +96,7 @@ export declare const llmConfigBaseSchema: z.ZodObject<{
         batched: "batched";
         sequential: "sequential";
     }>>;
+    'mmproj-use-gpu': z.ZodOptional<z.ZodBoolean>;
 }, z.core.$strip>;
 export type LlmConfigInput = z.infer<typeof llmConfigBaseSchema>;
 export declare const LLM_CONFIG_DEFAULTS: {
@@ -111,29 +118,35 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
     predict: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<-1>, z.ZodLiteral<-2>, z.ZodNumber]>>;
     system_prompt: z.ZodOptional<z.ZodString>;
     no_mmap: z.ZodOptional<z.ZodBoolean>;
-    verbosity: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<0>, z.ZodLiteral<1>, z.ZodLiteral<2>, z.ZodLiteral<3>]>>;
+    verbosity: z.ZodOptional<z.ZodEnum<{
+        readonly ERROR: 0;
+        readonly WARN: 1;
+        readonly INFO: 2;
+        readonly DEBUG: 3;
+    }>>;
     presence_penalty: z.ZodOptional<z.ZodNumber>;
     frequency_penalty: z.ZodOptional<z.ZodNumber>;
     repeat_penalty: z.ZodOptional<z.ZodNumber>;
     stop_sequences: z.ZodOptional<z.ZodArray<z.ZodString>>;
     n_discarded: z.ZodOptional<z.ZodNumber>;
+    parallel: z.ZodOptional<z.ZodNumber>;
     tools: z.ZodOptional<z.ZodBoolean>;
     toolsMode: z.ZodOptional<z.ZodEnum<{
         static: "static";
         dynamic: "dynamic";
     }>>;
-    "cache-type-k": z.ZodOptional<z.ZodString>;
-    "cache-type-v": z.ZodOptional<z.ZodString>;
-    "main-gpu": z.ZodOptional<z.ZodUnion<readonly [z.ZodNumber, z.ZodEnum<{
+    'cache-type-k': z.ZodOptional<z.ZodString>;
+    'cache-type-v': z.ZodOptional<z.ZodString>;
+    'main-gpu': z.ZodOptional<z.ZodUnion<readonly [z.ZodNumber, z.ZodEnum<{
         integrated: "integrated";
         dedicated: "dedicated";
     }>]>>;
-    "split-mode": z.ZodOptional<z.ZodEnum<{
+    'split-mode': z.ZodOptional<z.ZodEnum<{
         none: "none";
         layer: "layer";
         row: "row";
     }>>;
-    "tensor-split": z.ZodOptional<z.ZodString>;
+    'tensor-split': z.ZodOptional<z.ZodString>;
     openclCacheDir: z.ZodOptional<z.ZodString>;
     reasoning_budget: z.ZodOptional<z.ZodNumber>;
     projectionModelSrc: z.ZodOptional<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
@@ -178,6 +191,7 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
         batched: "batched";
         sequential: "sequential";
     }>>;
+    'mmproj-use-gpu': z.ZodOptional<z.ZodBoolean>;
 }, z.core.$strip>, z.ZodTransform<{
     ctx_size: number | undefined;
     temp?: number | undefined;
@@ -196,13 +210,14 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
     repeat_penalty?: number | undefined;
     stop_sequences?: string[] | undefined;
     n_discarded?: number | undefined;
+    parallel?: number | undefined;
     tools?: boolean | undefined;
     toolsMode?: "static" | "dynamic" | undefined;
-    "cache-type-k"?: string | undefined;
-    "cache-type-v"?: string | undefined;
-    "main-gpu"?: number | "integrated" | "dedicated" | undefined;
-    "split-mode"?: "none" | "layer" | "row" | undefined;
-    "tensor-split"?: string | undefined;
+    'cache-type-k'?: string | undefined;
+    'cache-type-v'?: string | undefined;
+    'main-gpu'?: number | "integrated" | "dedicated" | undefined;
+    'split-mode'?: "none" | "layer" | "row" | undefined;
+    'tensor-split'?: string | undefined;
     openclCacheDir?: string | undefined;
     reasoning_budget?: number | undefined;
     projectionModelSrc?: string | {
@@ -219,6 +234,7 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
         addon?: "embeddings" | "diffusion" | "tts" | "ocr" | "vla" | "llamacpp-completion" | "whispercpp-transcription" | "bci-whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "tts-ggml" | "parakeet-transcription" | "ggml-ocr" | "sdcpp-generation" | "ggml-vla" | "ggml-classification" | "llm" | "whisper" | "bci" | "nmt" | "parakeet" | "classification" | "vad" | undefined;
     } | undefined;
     image_tile_mode: "disabled" | "batched" | "sequential" | undefined;
+    'mmproj-use-gpu'?: boolean | undefined;
 }, {
     ctx_size?: number | undefined;
     temp?: number | undefined;
@@ -237,13 +253,14 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
     repeat_penalty?: number | undefined;
     stop_sequences?: string[] | undefined;
     n_discarded?: number | undefined;
+    parallel?: number | undefined;
     tools?: boolean | undefined;
     toolsMode?: "static" | "dynamic" | undefined;
-    "cache-type-k"?: string | undefined;
-    "cache-type-v"?: string | undefined;
-    "main-gpu"?: number | "integrated" | "dedicated" | undefined;
-    "split-mode"?: "none" | "layer" | "row" | undefined;
-    "tensor-split"?: string | undefined;
+    'cache-type-k'?: string | undefined;
+    'cache-type-v'?: string | undefined;
+    'main-gpu'?: number | "integrated" | "dedicated" | undefined;
+    'split-mode'?: "none" | "layer" | "row" | undefined;
+    'tensor-split'?: string | undefined;
     openclCacheDir?: string | undefined;
     reasoning_budget?: number | undefined;
     projectionModelSrc?: string | {
@@ -260,13 +277,14 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
         addon?: "embeddings" | "diffusion" | "tts" | "ocr" | "vla" | "llamacpp-completion" | "whispercpp-transcription" | "bci-whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "tts-ggml" | "parakeet-transcription" | "ggml-ocr" | "sdcpp-generation" | "ggml-vla" | "ggml-classification" | "llm" | "whisper" | "bci" | "nmt" | "parakeet" | "classification" | "vad" | undefined;
     } | undefined;
     image_tile_mode?: "disabled" | "batched" | "sequential" | undefined;
+    'mmproj-use-gpu'?: boolean | undefined;
 }>>;
 export type LlmConfig = z.infer<typeof llmConfigSchema>;
 export declare const embedConfigBaseSchema: z.ZodObject<{
     gpuLayers: z.ZodOptional<z.ZodNumber>;
     device: z.ZodOptional<z.ZodEnum<{
-        cpu: "cpu";
         gpu: "gpu";
+        cpu: "cpu";
     }>>;
     batchSize: z.ZodOptional<z.ZodNumber>;
     pooling: z.ZodOptional<z.ZodEnum<{
@@ -296,7 +314,12 @@ export declare const embedConfigBaseSchema: z.ZodObject<{
         row: "row";
     }>>;
     tensorSplit: z.ZodOptional<z.ZodString>;
-    verbosity: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<0>, z.ZodLiteral<1>, z.ZodLiteral<2>, z.ZodLiteral<3>]>>;
+    verbosity: z.ZodOptional<z.ZodEnum<{
+        readonly ERROR: 0;
+        readonly WARN: 1;
+        readonly INFO: 2;
+        readonly DEBUG: 3;
+    }>>;
     openclCacheDir: z.ZodOptional<z.ZodString>;
 }, z.core.$strip>;
 export type EmbedConfigInput = z.infer<typeof embedConfigBaseSchema>;
@@ -308,8 +331,8 @@ export declare const EMBED_CONFIG_DEFAULTS: {
 export declare const embedConfigSchema: z.ZodPipe<z.ZodObject<{
     gpuLayers: z.ZodOptional<z.ZodNumber>;
     device: z.ZodOptional<z.ZodEnum<{
-        cpu: "cpu";
         gpu: "gpu";
+        cpu: "cpu";
     }>>;
     batchSize: z.ZodOptional<z.ZodNumber>;
     pooling: z.ZodOptional<z.ZodEnum<{
@@ -339,11 +362,16 @@ export declare const embedConfigSchema: z.ZodPipe<z.ZodObject<{
         row: "row";
     }>>;
     tensorSplit: z.ZodOptional<z.ZodString>;
-    verbosity: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<0>, z.ZodLiteral<1>, z.ZodLiteral<2>, z.ZodLiteral<3>]>>;
+    verbosity: z.ZodOptional<z.ZodEnum<{
+        readonly ERROR: 0;
+        readonly WARN: 1;
+        readonly INFO: 2;
+        readonly DEBUG: 3;
+    }>>;
     openclCacheDir: z.ZodOptional<z.ZodString>;
 }, z.core.$strip>, z.ZodTransform<{
     gpuLayers: number | undefined;
-    device: "cpu" | "gpu" | undefined;
+    device: "gpu" | "cpu" | undefined;
     batchSize: number | undefined;
     pooling?: "none" | "mean" | "cls" | "last" | "rank" | undefined;
     attention?: "causal" | "non-causal" | undefined;
@@ -356,7 +384,7 @@ export declare const embedConfigSchema: z.ZodPipe<z.ZodObject<{
     openclCacheDir?: string | undefined;
 }, {
     gpuLayers?: number | undefined;
-    device?: "cpu" | "gpu" | undefined;
+    device?: "gpu" | "cpu" | undefined;
     batchSize?: number | undefined;
     pooling?: "none" | "mean" | "cls" | "last" | "rank" | undefined;
     attention?: "causal" | "non-causal" | undefined;

package/dist/schemas/llamacpp-config.js

@@ -1,6 +1,6 @@
-import { z } from "zod";
-import { modelSrcInputSchema } from "./model-src-utils.js";
-import { TOOLS_MODE } from "./tools.js";
+import { z } from 'zod';
+import { modelSrcInputSchema } from './model-src-utils.js';
+import { TOOLS_MODE } from './tools.js';
 /**
  * Upper bound for `reasoning_budget`. Mirrors the llm-llamacpp addon, which
  * stores the budget as a 32-bit `int` and rejects values above
@@ -11,14 +11,9 @@ export const VERBOSITY = {
     ERROR: 0,
     WARN: 1,
     INFO: 2,
-    DEBUG: 3,
+    DEBUG: 3
 };
-const verbositySchema = z.union([
-    z.literal(VERBOSITY.ERROR),
-    z.literal(VERBOSITY.WARN),
-    z.literal(VERBOSITY.INFO),
-    z.literal(VERBOSITY.DEBUG),
-]);
+const verbositySchema = z.enum(VERBOSITY);
 // Base schema - validates types, all fields optional (for client-side validation)
 export const llmConfigBaseSchema = z.object({
     ctx_size: z.number().optional(),
@@ -33,7 +28,7 @@ export const llmConfigBaseSchema = z.object({
         .union([
         z.literal(-1), // special: until stop token
         z.literal(-2), // special: until context filled
-        z.number().int().min(1), // positive integer: fixed token count
+        z.number().int().min(1) // positive integer: fixed token count
     ])
         .optional(),
     /** JS-side only: seeds conversation history. Never forwarded to the C++ addon. */
@@ -45,18 +40,22 @@ export const llmConfigBaseSchema = z.object({
     repeat_penalty: z.number().optional(),
     stop_sequences: z.array(z.string()).optional(),
     n_discarded: z.number().optional(),
+    /**
+     * Concurrent sequence slots for continuous batching. Defaults to 1
+     * (sequential, batching disabled); values >= 2 enable batched decoding in
+     * @qvac/llm-llamacpp.
+     */
+    parallel: z.number().int().min(1).optional(),
     tools: z.boolean().optional(),
     toolsMode: z
         .enum([TOOLS_MODE.static, TOOLS_MODE.dynamic])
         .describe('Controls tool placement in the prompt. "static" (default) prepends the tool set once and reuses it across the session. "dynamic" anchors tools after the last user message and trims them from the kv-cache after the chain resolves so each user prompt can carry its own tools.')
         .optional(),
-    "cache-type-k": z.string().optional(),
-    "cache-type-v": z.string().optional(),
-    "main-gpu": z
-        .union([z.number().int().min(0), z.enum(["integrated", "dedicated"])])
-        .optional(),
-    "split-mode": z.enum(["none", "layer", "row"]).optional(),
-    "tensor-split": z.string().optional(),
+    'cache-type-k': z.string().optional(),
+    'cache-type-v': z.string().optional(),
+    'main-gpu': z.union([z.number().int().min(0), z.enum(['integrated', 'dedicated'])]).optional(),
+    'split-mode': z.enum(['none', 'layer', 'row']).optional(),
+    'tensor-split': z.string().optional(),
     /**
      * Writable directory for OpenCL kernel binary cache. Required on Android
      * for fast GPU startup.
@@ -76,50 +75,59 @@ export const llmConfigBaseSchema = z.object({
      *   - `"disabled"`: no multi-tile encoding (single tile).
      * Ignored by text-only models. Default is `"sequential"`.
      */
-    image_tile_mode: z.enum(["disabled", "batched", "sequential"]).optional(),
+    image_tile_mode: z.enum(['disabled', 'batched', 'sequential']).optional(),
+    /**
+     * Run the multimodal projector (mmproj / vision encoder) on the GPU
+     * (multimodal models only). `true` forces GPU, `false` forces CPU. When
+     * unset, the llm-llamacpp addon auto-selects the backend per device class:
+     * GPU on desktop/iOS and positively-detected Android Adreno 800+ GPUs; CPU on
+     * every other Android GPU — Mali, Adreno <800, and any tier that can't be
+     * detected (only Adreno 800+ is benchmarked to encode the projector faster on
+     * the GPU than the CPU). Only honoured when the model itself runs on a GPU
+     * backend — ignored with a warning on CPU.
+     */
+    'mmproj-use-gpu': z.boolean().optional()
 });
 // Default values - typed as partial of the config
 export const LLM_CONFIG_DEFAULTS = {
     ctx_size: 1024,
     gpu_layers: 99,
-    device: "gpu",
-    system_prompt: "You are a helpful assistant.",
-    image_tile_mode: "sequential",
+    device: 'gpu',
+    system_prompt: 'You are a helpful assistant.',
+    image_tile_mode: 'sequential'
 };
 // Full schema - applies defaults via transform (no duplication)
 export const llmConfigSchema = llmConfigBaseSchema.transform((data) => ({
     ...LLM_CONFIG_DEFAULTS,
-    ...data,
+    ...data
 }));
 // Base schema - validates types, all fields optional (for client-side validation)
 export const embedConfigBaseSchema = z.object({
     gpuLayers: z.number().int().optional(),
-    device: z.enum(["gpu", "cpu"]).optional(),
+    device: z.enum(['gpu', 'cpu']).optional(),
     batchSize: z.number().int().min(1).optional(),
-    pooling: z.enum(["none", "mean", "cls", "last", "rank"]).optional(),
-    attention: z.enum(["causal", "non-causal"]).optional(),
+    pooling: z.enum(['none', 'mean', 'cls', 'last', 'rank']).optional(),
+    attention: z.enum(['causal', 'non-causal']).optional(),
     embdNormalize: z.number().int().optional(),
-    flashAttention: z.enum(["on", "off", "auto"]).optional(),
-    mainGpu: z
-        .union([z.number().int().min(0), z.enum(["integrated", "dedicated"])])
-        .optional(),
-    splitMode: z.enum(["none", "layer", "row"]).optional(),
+    flashAttention: z.enum(['on', 'off', 'auto']).optional(),
+    mainGpu: z.union([z.number().int().min(0), z.enum(['integrated', 'dedicated'])]).optional(),
+    splitMode: z.enum(['none', 'layer', 'row']).optional(),
     tensorSplit: z.string().optional(),
     verbosity: verbositySchema.optional(),
     /**
      * Writable directory for OpenCL kernel binary cache. Required on Android
      * for fast GPU startup.
      */
-    openclCacheDir: z.string().optional(),
+    openclCacheDir: z.string().optional()
 });
 // Default values - typed as partial of the config
 export const EMBED_CONFIG_DEFAULTS = {
     gpuLayers: 99,
-    device: "gpu",
-    batchSize: 1024,
+    device: 'gpu',
+    batchSize: 1024
 };
 // Full schema - validates then applies defaults via transform
 export const embedConfigSchema = embedConfigBaseSchema.transform((data) => ({
     ...EMBED_CONFIG_DEFAULTS,
-    ...data,
+    ...data
 }));