@qvac/sdk 0.18.2 → 0.19.0

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

@@ -1,404 +0,0 @@
-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
- * `std::numeric_limits<int>::max()`.
- */
-export declare const REASONING_BUDGET_MAX = 2147483647;
-export declare const VERBOSITY: {
-    readonly ERROR: 0;
-    readonly WARN: 1;
-    readonly INFO: 2;
-    readonly DEBUG: 3;
-};
-export declare const llmConfigBaseSchema: z.ZodObject<{
-    ctx_size: z.ZodOptional<z.ZodNumber>;
-    temp: z.ZodOptional<z.ZodNumber>;
-    top_p: z.ZodOptional<z.ZodNumber>;
-    top_k: z.ZodOptional<z.ZodNumber>;
-    seed: z.ZodOptional<z.ZodNumber>;
-    gpu_layers: z.ZodOptional<z.ZodNumber>;
-    lora: z.ZodOptional<z.ZodString>;
-    device: z.ZodOptional<z.ZodString>;
-    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.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>;
-    '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<{
-        none: "none";
-        layer: "layer";
-        row: "row";
-    }>>;
-    '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<{
-        src: z.ZodString;
-        name: z.ZodOptional<z.ZodString>;
-        modelId: z.ZodOptional<z.ZodString>;
-        registryPath: z.ZodOptional<z.ZodString>;
-        registrySource: z.ZodOptional<z.ZodString>;
-        blobCoreKey: z.ZodOptional<z.ZodString>;
-        blobIndex: z.ZodOptional<z.ZodNumber>;
-        engine: z.ZodOptional<z.ZodString>;
-        expectedSize: z.ZodOptional<z.ZodNumber>;
-        sha256Checksum: z.ZodOptional<z.ZodString>;
-        addon: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
-            embeddings: "embeddings";
-            diffusion: "diffusion";
-            audiogen: "audiogen";
-            tts: "tts";
-            ocr: "ocr";
-            vla: "vla";
-            "llamacpp-completion": "llamacpp-completion";
-            "whispercpp-transcription": "whispercpp-transcription";
-            "bci-whispercpp-transcription": "bci-whispercpp-transcription";
-            "llamacpp-embedding": "llamacpp-embedding";
-            "nmtcpp-translation": "nmtcpp-translation";
-            "onnx-tts": "onnx-tts";
-            "tts-ggml": "tts-ggml";
-            "parakeet-transcription": "parakeet-transcription";
-            "ggml-ocr": "ggml-ocr";
-            "sdcpp-generation": "sdcpp-generation";
-            "audiogen-ggml": "audiogen-ggml";
-            "ggml-vla": "ggml-vla";
-            "ggml-classification": "ggml-classification";
-            llm: "llm";
-            whisper: "whisper";
-            bci: "bci";
-            nmt: "nmt";
-            parakeet: "parakeet";
-            classification: "classification";
-        }>, z.ZodLiteral<"vad">]>>;
-    }, z.core.$strip>]>>;
-    image_tile_mode: z.ZodOptional<z.ZodEnum<{
-        disabled: "disabled";
-        batched: "batched";
-        sequential: "sequential";
-    }>>;
-    image_no_upscale: z.ZodOptional<z.ZodEnum<{
-        on: "on";
-        off: "off";
-    }>>;
-    'mmproj-use-gpu': z.ZodOptional<z.ZodBoolean>;
-}, z.core.$strip>;
-export type LlmConfigInput = z.infer<typeof llmConfigBaseSchema>;
-export declare const LLM_CONFIG_DEFAULTS: {
-    readonly ctx_size: 1024;
-    readonly gpu_layers: 99;
-    readonly device: "gpu";
-    readonly system_prompt: "You are a helpful assistant.";
-    readonly image_tile_mode: "sequential";
-};
-export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
-    ctx_size: z.ZodOptional<z.ZodNumber>;
-    temp: z.ZodOptional<z.ZodNumber>;
-    top_p: z.ZodOptional<z.ZodNumber>;
-    top_k: z.ZodOptional<z.ZodNumber>;
-    seed: z.ZodOptional<z.ZodNumber>;
-    gpu_layers: z.ZodOptional<z.ZodNumber>;
-    lora: z.ZodOptional<z.ZodString>;
-    device: z.ZodOptional<z.ZodString>;
-    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.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>;
-    '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<{
-        none: "none";
-        layer: "layer";
-        row: "row";
-    }>>;
-    '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<{
-        src: z.ZodString;
-        name: z.ZodOptional<z.ZodString>;
-        modelId: z.ZodOptional<z.ZodString>;
-        registryPath: z.ZodOptional<z.ZodString>;
-        registrySource: z.ZodOptional<z.ZodString>;
-        blobCoreKey: z.ZodOptional<z.ZodString>;
-        blobIndex: z.ZodOptional<z.ZodNumber>;
-        engine: z.ZodOptional<z.ZodString>;
-        expectedSize: z.ZodOptional<z.ZodNumber>;
-        sha256Checksum: z.ZodOptional<z.ZodString>;
-        addon: z.ZodOptional<z.ZodUnion<readonly [z.ZodEnum<{
-            embeddings: "embeddings";
-            diffusion: "diffusion";
-            audiogen: "audiogen";
-            tts: "tts";
-            ocr: "ocr";
-            vla: "vla";
-            "llamacpp-completion": "llamacpp-completion";
-            "whispercpp-transcription": "whispercpp-transcription";
-            "bci-whispercpp-transcription": "bci-whispercpp-transcription";
-            "llamacpp-embedding": "llamacpp-embedding";
-            "nmtcpp-translation": "nmtcpp-translation";
-            "onnx-tts": "onnx-tts";
-            "tts-ggml": "tts-ggml";
-            "parakeet-transcription": "parakeet-transcription";
-            "ggml-ocr": "ggml-ocr";
-            "sdcpp-generation": "sdcpp-generation";
-            "audiogen-ggml": "audiogen-ggml";
-            "ggml-vla": "ggml-vla";
-            "ggml-classification": "ggml-classification";
-            llm: "llm";
-            whisper: "whisper";
-            bci: "bci";
-            nmt: "nmt";
-            parakeet: "parakeet";
-            classification: "classification";
-        }>, z.ZodLiteral<"vad">]>>;
-    }, z.core.$strip>]>>;
-    image_tile_mode: z.ZodOptional<z.ZodEnum<{
-        disabled: "disabled";
-        batched: "batched";
-        sequential: "sequential";
-    }>>;
-    image_no_upscale: z.ZodOptional<z.ZodEnum<{
-        on: "on";
-        off: "off";
-    }>>;
-    'mmproj-use-gpu': z.ZodOptional<z.ZodBoolean>;
-}, z.core.$strip>, z.ZodTransform<{
-    ctx_size: number | undefined;
-    temp?: number | undefined;
-    top_p?: number | undefined;
-    top_k?: number | undefined;
-    seed?: number | undefined;
-    gpu_layers: number | undefined;
-    lora?: string | undefined;
-    device: string | undefined;
-    predict?: number | undefined;
-    system_prompt: string | undefined;
-    no_mmap?: boolean | undefined;
-    verbosity?: 0 | 3 | 2 | 1 | undefined;
-    presence_penalty?: number | undefined;
-    frequency_penalty?: number | undefined;
-    repeat_penalty?: number | undefined;
-    stop_sequences?: string[] | undefined;
-    n_discarded?: number | undefined;
-    parallel?: number | undefined;
-    tools?: boolean | 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 | {
-        src: string;
-        name?: string | undefined;
-        modelId?: string | undefined;
-        registryPath?: string | undefined;
-        registrySource?: string | undefined;
-        blobCoreKey?: string | undefined;
-        blobIndex?: number | undefined;
-        engine?: string | undefined;
-        expectedSize?: number | undefined;
-        sha256Checksum?: string | undefined;
-        addon?: "embeddings" | "diffusion" | "audiogen" | "tts" | "ocr" | "vla" | "llamacpp-completion" | "whispercpp-transcription" | "bci-whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "tts-ggml" | "parakeet-transcription" | "ggml-ocr" | "sdcpp-generation" | "audiogen-ggml" | "ggml-vla" | "ggml-classification" | "llm" | "whisper" | "bci" | "nmt" | "parakeet" | "classification" | "vad" | undefined;
-    } | undefined;
-    image_tile_mode: "disabled" | "batched" | "sequential" | undefined;
-    image_no_upscale?: "on" | "off" | undefined;
-    'mmproj-use-gpu'?: boolean | undefined;
-}, {
-    ctx_size?: number | undefined;
-    temp?: number | undefined;
-    top_p?: number | undefined;
-    top_k?: number | undefined;
-    seed?: number | undefined;
-    gpu_layers?: number | undefined;
-    lora?: string | undefined;
-    device?: string | undefined;
-    predict?: number | undefined;
-    system_prompt?: string | undefined;
-    no_mmap?: boolean | undefined;
-    verbosity?: 0 | 3 | 2 | 1 | undefined;
-    presence_penalty?: number | undefined;
-    frequency_penalty?: number | undefined;
-    repeat_penalty?: number | undefined;
-    stop_sequences?: string[] | undefined;
-    n_discarded?: number | undefined;
-    parallel?: number | undefined;
-    tools?: boolean | 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 | {
-        src: string;
-        name?: string | undefined;
-        modelId?: string | undefined;
-        registryPath?: string | undefined;
-        registrySource?: string | undefined;
-        blobCoreKey?: string | undefined;
-        blobIndex?: number | undefined;
-        engine?: string | undefined;
-        expectedSize?: number | undefined;
-        sha256Checksum?: string | undefined;
-        addon?: "embeddings" | "diffusion" | "audiogen" | "tts" | "ocr" | "vla" | "llamacpp-completion" | "whispercpp-transcription" | "bci-whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "tts-ggml" | "parakeet-transcription" | "ggml-ocr" | "sdcpp-generation" | "audiogen-ggml" | "ggml-vla" | "ggml-classification" | "llm" | "whisper" | "bci" | "nmt" | "parakeet" | "classification" | "vad" | undefined;
-    } | undefined;
-    image_tile_mode?: "disabled" | "batched" | "sequential" | undefined;
-    image_no_upscale?: "on" | "off" | 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<{
-        gpu: "gpu";
-        cpu: "cpu";
-    }>>;
-    batchSize: z.ZodOptional<z.ZodNumber>;
-    pooling: z.ZodOptional<z.ZodEnum<{
-        none: "none";
-        mean: "mean";
-        cls: "cls";
-        last: "last";
-        rank: "rank";
-    }>>;
-    attention: z.ZodOptional<z.ZodEnum<{
-        causal: "causal";
-        "non-causal": "non-causal";
-    }>>;
-    embdNormalize: z.ZodOptional<z.ZodNumber>;
-    flashAttention: z.ZodOptional<z.ZodEnum<{
-        on: "on";
-        off: "off";
-        auto: "auto";
-    }>>;
-    mainGpu: z.ZodOptional<z.ZodUnion<readonly [z.ZodNumber, z.ZodEnum<{
-        integrated: "integrated";
-        dedicated: "dedicated";
-    }>]>>;
-    splitMode: z.ZodOptional<z.ZodEnum<{
-        none: "none";
-        layer: "layer";
-        row: "row";
-    }>>;
-    tensorSplit: z.ZodOptional<z.ZodString>;
-    verbosity: z.ZodOptional<z.ZodEnum<{
-        readonly ERROR: 0;
-        readonly WARN: 1;
-        readonly INFO: 2;
-        readonly DEBUG: 3;
-    }>>;
-    openclCacheDir: z.ZodOptional<z.ZodString>;
-}, z.core.$strip>;
-export type EmbedConfigInput = z.infer<typeof embedConfigBaseSchema>;
-export declare const EMBED_CONFIG_DEFAULTS: {
-    readonly gpuLayers: 99;
-    readonly device: "gpu";
-    readonly batchSize: 1024;
-};
-export declare const embedConfigSchema: z.ZodPipe<z.ZodObject<{
-    gpuLayers: z.ZodOptional<z.ZodNumber>;
-    device: z.ZodOptional<z.ZodEnum<{
-        gpu: "gpu";
-        cpu: "cpu";
-    }>>;
-    batchSize: z.ZodOptional<z.ZodNumber>;
-    pooling: z.ZodOptional<z.ZodEnum<{
-        none: "none";
-        mean: "mean";
-        cls: "cls";
-        last: "last";
-        rank: "rank";
-    }>>;
-    attention: z.ZodOptional<z.ZodEnum<{
-        causal: "causal";
-        "non-causal": "non-causal";
-    }>>;
-    embdNormalize: z.ZodOptional<z.ZodNumber>;
-    flashAttention: z.ZodOptional<z.ZodEnum<{
-        on: "on";
-        off: "off";
-        auto: "auto";
-    }>>;
-    mainGpu: z.ZodOptional<z.ZodUnion<readonly [z.ZodNumber, z.ZodEnum<{
-        integrated: "integrated";
-        dedicated: "dedicated";
-    }>]>>;
-    splitMode: z.ZodOptional<z.ZodEnum<{
-        none: "none";
-        layer: "layer";
-        row: "row";
-    }>>;
-    tensorSplit: z.ZodOptional<z.ZodString>;
-    verbosity: z.ZodOptional<z.ZodEnum<{
-        readonly ERROR: 0;
-        readonly WARN: 1;
-        readonly INFO: 2;
-        readonly DEBUG: 3;
-    }>>;
-    openclCacheDir: z.ZodOptional<z.ZodString>;
-}, z.core.$strip>, z.ZodTransform<{
-    gpuLayers: number | undefined;
-    device: "gpu" | "cpu" | undefined;
-    batchSize: number | undefined;
-    pooling?: "none" | "mean" | "cls" | "last" | "rank" | undefined;
-    attention?: "causal" | "non-causal" | undefined;
-    embdNormalize?: number | undefined;
-    flashAttention?: "on" | "off" | "auto" | undefined;
-    mainGpu?: number | "integrated" | "dedicated" | undefined;
-    splitMode?: "none" | "layer" | "row" | undefined;
-    tensorSplit?: string | undefined;
-    verbosity?: 0 | 3 | 2 | 1 | undefined;
-    openclCacheDir?: string | undefined;
-}, {
-    gpuLayers?: number | undefined;
-    device?: "gpu" | "cpu" | undefined;
-    batchSize?: number | undefined;
-    pooling?: "none" | "mean" | "cls" | "last" | "rank" | undefined;
-    attention?: "causal" | "non-causal" | undefined;
-    embdNormalize?: number | undefined;
-    flashAttention?: "on" | "off" | "auto" | undefined;
-    mainGpu?: number | "integrated" | "dedicated" | undefined;
-    splitMode?: "none" | "layer" | "row" | undefined;
-    tensorSplit?: string | undefined;
-    verbosity?: 0 | 3 | 2 | 1 | undefined;
-    openclCacheDir?: string | undefined;
-}>>;
-export type EmbedConfig = z.infer<typeof embedConfigSchema>;
-//# sourceMappingURL=llamacpp-config.d.ts.map

package/dist/schemas/llamacpp-config.js

@@ -1,219 +0,0 @@
-import { z } from 'zod';
-import { modelSrcInputSchema } from './model-src-utils.js';
-/**
- * Upper bound for `reasoning_budget`. Mirrors the llm-llamacpp addon, which
- * stores the budget as a 32-bit `int` and rejects values above
- * `std::numeric_limits<int>::max()`.
- */
-export const REASONING_BUDGET_MAX = 2147483647;
-export const VERBOSITY = {
-    ERROR: 0,
-    WARN: 1,
-    INFO: 2,
-    DEBUG: 3
-};
-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()
-        .describe("Context window size in tokens; `0` uses the model's trained context length. Default 1024."),
-    temp: z.number().min(0).max(2).optional().describe('Sampling temperature (0–2). Default 0.8.'),
-    top_p: z
-        .number()
-        .min(0)
-        .max(1)
-        .optional()
-        .describe('Top-p (nucleus) sampling cutoff (0–1). Default 0.9.'),
-    top_k: z
-        .number()
-        .int()
-        .min(0)
-        .max(128)
-        .optional()
-        .describe('Top-k sampling — keep only the top K tokens (0–128). Default 40.'),
-    seed: z.number().optional().describe('Sampling RNG seed; `-1` (default) picks a random seed.'),
-    gpu_layers: z
-        .number()
-        .optional()
-        .describe('Number of model layers to offload to the GPU. Default 99 (offload all).'),
-    lora: z
-        .string()
-        .optional()
-        .describe('Path to a LoRA adapter file to apply on top of the base model.'),
-    device: z
-        .string()
-        .optional()
-        .describe("Device to run inference on: `'gpu'` or `'cpu'`. Default `'gpu'`."),
-    predict: z
-        .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
-    ])
-        .optional()
-        .describe('Max tokens to predict. `-1` = until stop token, `-2` = until context filled.'),
-    system_prompt: z
-        .string()
-        .optional()
-        .describe("Seeds conversation history on the JS side only; never forwarded to the addon. Default `'You are a helpful assistant.'`"),
-    no_mmap: z.boolean().optional().describe('Disable memory-mapped model loading. Default false.'),
-    verbosity: verbositySchema
-        .optional()
-        .describe('Native log verbosity: `0`=ERROR, `1`=WARN, `2`=INFO, `3`=DEBUG. Default 0.'),
-    presence_penalty: z
-        .number()
-        .optional()
-        .describe('Presence penalty applied to tokens that have already appeared. Default 0.'),
-    frequency_penalty: z
-        .number()
-        .optional()
-        .describe('Frequency penalty applied to tokens by their frequency so far. Default 0.'),
-    repeat_penalty: z
-        .number()
-        .optional()
-        .describe('Repetition penalty applied to repeated tokens. Default 1.1.'),
-    stop_sequences: z
-        .array(z.string())
-        .optional()
-        .describe('Strings that stop generation when produced (forwarded to the addon as `reverse_prompt`).'),
-    n_discarded: z
-        .number()
-        .optional()
-        .describe('Tokens to discard from the front of the context when it fills (sliding window); `0` (default) disables sliding. In batch mode clamped to the per-slot window (`ctx_size / parallel`).'),
-    parallel: z
-        .number()
-        .int()
-        .min(1)
-        .optional()
-        .describe('Concurrent sequence slots for continuous batching (1–256). Default 1 (sequential, batching off); `>= 2` enables batched decoding and splits the KV cache evenly across slots.'),
-    tools: z
-        .boolean()
-        .optional()
-        .describe("Enable tool calling via the model's jinja chat template. Default false."),
-    'cache-type-k': z
-        .string()
-        .optional()
-        .describe('KV-cache key quantization type (`f16`, `f32`, `bf16`, `q8_0`, `q4_0`, …). Unset selects a backend-safe default.'),
-    'cache-type-v': z
-        .string()
-        .optional()
-        .describe('KV-cache value quantization type; quantizing the value cache requires flash attention. Unset selects a backend-safe default.'),
-    'main-gpu': z
-        .union([z.number().int().min(0), z.enum(['integrated', 'dedicated'])])
-        .optional()
-        .describe("GPU to use on multi-GPU systems: a device index, or `'integrated'`/`'dedicated'` to restrict selection to that class."),
-    'split-mode': z
-        .enum(['none', 'layer', 'row'])
-        .optional()
-        .describe("How to split the model across GPUs: `'none'` (default, single GPU), `'layer'` (pipeline parallelism), or `'row'` (tensor parallelism)."),
-    'tensor-split': z
-        .string()
-        .optional()
-        .describe("Proportions for distributing layers/rows across GPUs, e.g. `'1,1'` (equal) or `'3,1'` (75/25)."),
-    openclCacheDir: z
-        .string()
-        .optional()
-        .describe('Writable directory for the OpenCL kernel binary cache; required on Android for fast GPU startup.'),
-    reasoning_budget: z
-        .number()
-        .int()
-        .min(-1)
-        .max(REASONING_BUDGET_MAX)
-        .optional()
-        .describe('Reasoning-channel token budget. `-1` (default) unrestricted, `0` disables it, any positive integer caps the reasoning channel at that many tokens (the closing think tag is force-emitted once the budget is spent).'),
-    projectionModelSrc: modelSrcInputSchema
-        .optional()
-        .describe('Multimodal projection (mmproj / vision encoder) model source; multimodal models only.'),
-    image_tile_mode: z
-        .enum(['disabled', 'batched', 'sequential'])
-        .optional()
-        .describe("Qwen3.5-VL multi-tile image encoding mode (multimodal models only): `'sequential'` (default) encodes image tiles one at a time, `'batched'` encodes all tiles in a single batched pass, `'disabled'` does no multi-tile encoding (single tile). Ignored by text-only models."),
-    image_no_upscale: z
-        .enum(['on', 'off'])
-        .optional()
-        .describe("idefics3-style image preprocessing rule (multimodal models only): `'on'` rounds the image's long side up to a whole number of slices and caps it, so an image smaller than the cap keeps its own resolution and becomes far fewer slices; `'off'` always stretches the long side to the cap. When unset, the model's own GGUF value is used; ignored with a warning by models that do not use idefics3-style preprocessing. Changes the number of image tokens (and therefore both accuracy and encode time), so a checkpoint whose GGUF omits the key needs this set to preprocess correctly."),
-    'mmproj-use-gpu': z
-        .boolean()
-        .optional()
-        .describe('Run the multimodal projector (mmproj / vision encoder) on the GPU (multimodal models only). `true` forces GPU, `false` forces CPU. When unset, the backend is auto-selected per device class: GPU on desktop/iOS and Android Adreno 800+; CPU on every other Android GPU (Arm Mali, Adreno <800, and undetectable Adreno tiers), with the LLM layers still on the GPU. Only honoured when the model itself runs on a GPU backend — ignored with a warning on CPU.')
-});
-// 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'
-};
-// Full schema - applies defaults via transform (no duplication)
-export const llmConfigSchema = llmConfigBaseSchema.transform((data) => ({
-    ...LLM_CONFIG_DEFAULTS,
-    ...data
-}));
-// Base schema - validates types, all fields optional (for client-side validation)
-export const embedConfigBaseSchema = z.object({
-    gpuLayers: z
-        .number()
-        .int()
-        .optional()
-        .describe('Number of model layers to offload to the GPU. Default 99 (offload all).'),
-    device: z
-        .enum(['gpu', 'cpu'])
-        .optional()
-        .describe("Device to run inference on: `'gpu'` or `'cpu'`. Default `'gpu'`."),
-    batchSize: z
-        .number()
-        .int()
-        .min(1)
-        .optional()
-        .describe('Tokens processed per batch (input throughput). Default 1024.'),
-    pooling: z
-        .enum(['none', 'mean', 'cls', 'last', 'rank'])
-        .optional()
-        .describe("Pooling strategy collapsing token embeddings into one sequence vector: `'none'`, `'mean'`, `'cls'`, `'last'`, or `'rank'`. Unset uses the model default."),
-    attention: z
-        .enum(['causal', 'non-causal'])
-        .optional()
-        .describe("Attention type: `'causal'` or `'non-causal'`. Unset uses the model default."),
-    embdNormalize: z
-        .number()
-        .int()
-        .optional()
-        .describe('Embedding normalization: `-1` none, `0` max-abs int16, `1` taxicab, `2` euclidean, `>2` p-norm. Default 2 (euclidean).'),
-    flashAttention: z
-        .enum(['on', 'off', 'auto'])
-        .optional()
-        .describe("Flash attention: `'on'`, `'off'`, or `'auto'`. Default `'auto'`."),
-    mainGpu: z
-        .union([z.number().int().min(0), z.enum(['integrated', 'dedicated'])])
-        .optional()
-        .describe("GPU to use on multi-GPU systems: a device index, or `'integrated'`/`'dedicated'` to restrict selection to that class."),
-    splitMode: z
-        .enum(['none', 'layer', 'row'])
-        .optional()
-        .describe("How to split the model across GPUs: `'none'` (default, single GPU), `'layer'` (pipeline parallelism), or `'row'` (tensor parallelism)."),
-    tensorSplit: z
-        .string()
-        .optional()
-        .describe("Proportions for distributing layers/rows across GPUs, e.g. `'1,1'` (equal) or `'3,1'` (75/25)."),
-    verbosity: verbositySchema
-        .optional()
-        .describe('Native log verbosity: `0`=ERROR, `1`=WARN, `2`=INFO, `3`=DEBUG. Default 0.'),
-    openclCacheDir: z
-        .string()
-        .optional()
-        .describe('Writable directory for the OpenCL kernel binary cache; required on Android for fast GPU startup.')
-});
-// Default values - typed as partial of the config
-export const EMBED_CONFIG_DEFAULTS = {
-    gpuLayers: 99,
-    device: 'gpu',
-    batchSize: 1024
-};
-// Full schema - validates then applies defaults via transform
-export const embedConfigSchema = embedConfigBaseSchema.transform((data) => ({
-    ...EMBED_CONFIG_DEFAULTS,
-    ...data
-}));