@qvac/sdk 0.8.2 → 0.9.0

package/dist/schemas/finetune.js

@@ -0,0 +1,161 @@
+import { z } from "zod";
+const lrSchedulerValues = ["constant", "cosine", "linear"];
+const finetuneStatusValues = [
+    "IDLE",
+    "RUNNING",
+    "PAUSED",
+    "CANCELLED",
+    "COMPLETED",
+];
+export const finetuneValidationSchema = z
+    .discriminatedUnion("type", [
+    z
+        .object({
+        type: z.literal("none"),
+    })
+        .strict(),
+    z
+        .object({
+        type: z.literal("split"),
+        fraction: z.number().min(0).max(1).optional(),
+    })
+        .strict(),
+    z
+        .object({
+        type: z.literal("dataset"),
+        path: z.string().min(1, "Validation dataset path cannot be empty"),
+    })
+        .strict(),
+])
+    .transform((validation) => validation.type === "split"
+    ? {
+        ...validation,
+        fraction: validation.fraction ?? 0.05,
+    }
+    : validation);
+const finetuneOptionsPayloadSchema = z
+    .object({
+    trainDatasetDir: z.string().min(1, "Training dataset path cannot be empty"),
+    validation: finetuneValidationSchema,
+    outputParametersDir: z.string().min(1, "Output parameters path cannot be empty"),
+    numberOfEpochs: z.number().int().positive().optional(),
+    learningRate: z.number().positive().optional(),
+    contextLength: z.number().int().positive().optional(),
+    batchSize: z.number().int().positive().optional(),
+    microBatchSize: z.number().int().positive().optional(),
+    assistantLossOnly: z.boolean().optional(),
+    loraRank: z.number().int().positive().optional(),
+    loraAlpha: z.number().positive().optional(),
+    loraInitStd: z.number().positive().optional(),
+    loraSeed: z.number().int().optional(),
+    loraModules: z.string().min(1).optional(),
+    checkpointSaveDir: z.string().min(1).optional(),
+    checkpointSaveSteps: z.number().int().min(0).optional(),
+    chatTemplatePath: z.string().min(1).optional(),
+    lrScheduler: z.enum(lrSchedulerValues).optional(),
+    lrMin: z.number().min(0).optional(),
+    warmupRatio: z.number().min(0).max(1).optional(),
+    warmupRatioSet: z.boolean().optional(),
+    warmupSteps: z.number().int().min(0).optional(),
+    warmupStepsSet: z.boolean().optional(),
+    weightDecay: z.number().min(0).optional(),
+})
+    .strict();
+const finetuneModelParamsSchema = z
+    .object({
+    modelId: z.string(),
+})
+    .strict();
+const finetuneRunParamsSchemaBase = finetuneModelParamsSchema
+    .extend({
+    options: finetuneOptionsPayloadSchema,
+})
+    .strict();
+export const finetuneRunParamsSchema = finetuneRunParamsSchemaBase
+    .extend({
+    operation: z.enum(["start", "resume"]).optional(),
+})
+    .strict();
+export const finetuneGetStateParamsSchema = finetuneRunParamsSchemaBase
+    .extend({
+    operation: z.literal("getState"),
+})
+    .strict();
+export const finetuneStopParamsSchema = finetuneModelParamsSchema
+    .extend({
+    operation: z.enum(["pause", "cancel"]),
+})
+    .strict();
+export const finetuneParamsSchema = z.union([
+    finetuneRunParamsSchema,
+    finetuneGetStateParamsSchema,
+    finetuneStopParamsSchema,
+]);
+export const finetuneRunRequestSchema = finetuneRunParamsSchemaBase
+    .extend({
+    type: z.literal("finetune"),
+    operation: z.enum(["start", "resume"]).optional(),
+    withProgress: z.boolean().optional(),
+})
+    .strict();
+export const finetuneGetStateRequestSchema = finetuneRunParamsSchemaBase
+    .extend({
+    type: z.literal("finetune"),
+    operation: z.literal("getState"),
+})
+    .strict();
+export const finetuneStopRequestSchema = finetuneModelParamsSchema
+    .extend({
+    type: z.literal("finetune"),
+    operation: z.enum(["pause", "cancel"]),
+})
+    .strict();
+export const finetuneRequestSchema = z.union([
+    finetuneRunRequestSchema,
+    finetuneGetStateRequestSchema,
+    finetuneStopRequestSchema,
+]);
+export const finetuneStatusSchema = z.enum(finetuneStatusValues);
+export const finetuneProgressSchema = z
+    .object({
+    is_train: z.boolean(),
+    loss: z.xor([z.number(), z.nan()]).nullable(),
+    loss_uncertainty: z.xor([z.number(), z.nan()]).nullable(),
+    accuracy: z.xor([z.number(), z.nan()]).nullable(),
+    accuracy_uncertainty: z.xor([z.number(), z.nan()]).nullable(),
+    global_steps: z.number().int().min(0),
+    current_epoch: z.number().int().min(0),
+    current_batch: z.number().int().min(0),
+    total_batches: z.number().int().min(0),
+    elapsed_ms: z.number().min(0),
+    eta_ms: z.number().min(0),
+})
+    .strict();
+export const finetuneStatsSchema = z
+    .object({
+    train_loss: z.number().optional(),
+    train_loss_uncertainty: z.xor([z.number(), z.nan()]).nullable().optional(),
+    val_loss: z.number().optional(),
+    val_loss_uncertainty: z.xor([z.number(), z.nan()]).nullable().optional(),
+    train_accuracy: z.number().optional(),
+    train_accuracy_uncertainty: z.xor([z.number(), z.nan()]).nullable().optional(),
+    val_accuracy: z.number().optional(),
+    val_accuracy_uncertainty: z.xor([z.number(), z.nan()]).nullable().optional(),
+    learning_rate: z.number().optional(),
+    global_steps: z.number().int().min(0),
+    epochs_completed: z.number().int().min(0),
+})
+    .strict();
+export const finetuneResponseSchema = z
+    .object({
+    type: z.literal("finetune"),
+    status: finetuneStatusSchema,
+    stats: finetuneStatsSchema.optional(),
+})
+    .strict();
+export const finetuneProgressResponseSchema = finetuneProgressSchema
+    .extend({
+    type: z.literal("finetune:progress"),
+    modelId: z.string(),
+})
+    .strict();

package/dist/schemas/get-model-info.d.ts

@@ -42,6 +42,7 @@ export declare const modelInfoSchema: z.ZodObject<{
         parakeet: "parakeet";
         tts: "tts";
         ocr: "ocr";
+        diffusion: "diffusion";
         vad: "vad";
         other: "other";
     }>;
@@ -88,6 +89,7 @@ export declare const getModelInfoResponseSchema: z.ZodObject<{
             parakeet: "parakeet";
             tts: "tts";
             ocr: "ocr";
+            diffusion: "diffusion";
             vad: "vad";
             other: "other";
         }>;

package/dist/schemas/get-model-info.js

@@ -43,6 +43,7 @@ export const modelInfoSchema = z.object({
         "vad",
         "tts",
         "ocr",
+        "diffusion",
         "other",
     ]),
     isCached: z.boolean(),

package/dist/schemas/index.d.ts

@@ -6,6 +6,7 @@ export * from "./delegate";
 export * from "./delete-cache";
 export * from "./download-asset";
 export * from "./embed";
+export * from "./finetune";
 export * from "./load-model";
 export * from "./reload-config";
 export * from "./logging-stream";
@@ -23,12 +24,15 @@ export * from "./text-to-speech";
 export * from "./error";
 export * from "./rag";
 export * from "./ocr";
+export * from "./sdcpp-config";
 export * from "./shard";
+export * from "./suspend";
+export * from "./resume";
 export { SDK_CLIENT_ERROR_CODES } from "./sdk-errors-client";
 export { SDK_SERVER_ERROR_CODES } from "./sdk-errors-server";
 export { REGISTRY_ERROR_CODES } from "./sdk-errors-registry";
 export { qvacConfigSchema, deviceMatchSchema, deviceConfigDefaultsSchema, devicePatternSchema, type QvacConfig, type DeviceMatch, type DeviceConfigDefaults, type DevicePattern, } from "./sdk-config";
-export { PROFILING_KEY, PROFILING_TRAILER_KEY, DELEGATION_BREAKDOWN_KEY, OPERATION_EVENT_KEY, profilerModeSchema, serverBreakdownSchema, delegationBreakdownSchema, operationEventSchema, profilingRequestMetaSchema, profilingResponseMetaSchema, perCallProfilingSchema, type ProfilerMode, type ServerBreakdown, type DelegationBreakdown, type OperationEvent, type ProfilingRequestMeta, type ProfilingResponseMeta, type PerCallProfiling, } from "./profiling";
+export { PROFILING_KEY, PROFILING_TRAILER_KEY, DELEGATION_BREAKDOWN_KEY, OPERATION_EVENT_KEY, MODEL_EXECUTION_KEY, profilerModeSchema, serverBreakdownSchema, delegationBreakdownSchema, operationEventSchema, profilingRequestMetaSchema, profilingResponseMetaSchema, perCallProfilingSchema, type ProfilerMode, type ServerBreakdown, type DelegationBreakdown, type OperationEvent, type ProfilingRequestMeta, type ProfilingResponseMeta, type PerCallProfiling, } from "./profiling";
 export { runtimeContextSchema, type RuntimeContext } from "./runtime-context";
 export * from "./get-model-info";
 export * from "./model-src-utils";

package/dist/schemas/index.js

@@ -7,6 +7,7 @@ export * from "./delegate.js";
 export * from "./delete-cache.js";
 export * from "./download-asset.js";
 export * from "./embed.js";
+export * from "./finetune.js";
 export * from "./load-model.js";
 export * from "./reload-config.js";
 export * from "./logging-stream.js";
@@ -24,12 +25,15 @@ export * from "./text-to-speech.js";
 export * from "./error.js";
 export * from "./rag.js";
 export * from "./ocr.js";
+export * from "./sdcpp-config.js";
 export * from "./shard.js";
+export * from "./suspend.js";
+export * from "./resume.js";
 export { SDK_CLIENT_ERROR_CODES } from "./sdk-errors-client.js";
 export { SDK_SERVER_ERROR_CODES } from "./sdk-errors-server.js";
 export { REGISTRY_ERROR_CODES } from "./sdk-errors-registry.js";
 export { qvacConfigSchema, deviceMatchSchema, deviceConfigDefaultsSchema, devicePatternSchema, } from "./sdk-config.js";
-export { PROFILING_KEY, PROFILING_TRAILER_KEY, DELEGATION_BREAKDOWN_KEY, OPERATION_EVENT_KEY, profilerModeSchema, serverBreakdownSchema, delegationBreakdownSchema, operationEventSchema, profilingRequestMetaSchema, profilingResponseMetaSchema, perCallProfilingSchema, } from "./profiling.js";
+export { PROFILING_KEY, PROFILING_TRAILER_KEY, DELEGATION_BREAKDOWN_KEY, OPERATION_EVENT_KEY, MODEL_EXECUTION_KEY, profilerModeSchema, serverBreakdownSchema, delegationBreakdownSchema, operationEventSchema, profilingRequestMetaSchema, profilingResponseMetaSchema, perCallProfilingSchema, } from "./profiling.js";
 export { runtimeContextSchema } from "./runtime-context.js";
 export * from "./get-model-info.js";
 export * from "./model-src-utils.js";

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

@@ -24,6 +24,9 @@ export declare const llmConfigBaseSchema: z.ZodObject<{
     stop_sequences: z.ZodOptional<z.ZodArray<z.ZodString>>;
     n_discarded: z.ZodOptional<z.ZodNumber>;
     tools: z.ZodOptional<z.ZodBoolean>;
+    "cache-type-k": z.ZodOptional<z.ZodString>;
+    "cache-type-v": z.ZodOptional<z.ZodString>;
+    openclCacheDir: z.ZodOptional<z.ZodString>;
     projectionModelSrc: z.ZodOptional<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
         src: z.ZodString;
         name: z.ZodOptional<z.ZodString>;
@@ -43,6 +46,7 @@ export declare const llmConfigBaseSchema: z.ZodObject<{
             "onnx-tts": "onnx-tts";
             "parakeet-transcription": "parakeet-transcription";
             "onnx-ocr": "onnx-ocr";
+            "sdcpp-generation": "sdcpp-generation";
             llm: "llm";
             whisper: "whisper";
             embeddings: "embeddings";
@@ -50,6 +54,7 @@ export declare const llmConfigBaseSchema: z.ZodObject<{
             parakeet: "parakeet";
             tts: "tts";
             ocr: "ocr";
+            diffusion: "diffusion";
         }>, z.ZodLiteral<"vad">]>>;
     }, z.core.$strip>]>>;
 }, z.core.$strip>;
@@ -79,6 +84,9 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
     stop_sequences: z.ZodOptional<z.ZodArray<z.ZodString>>;
     n_discarded: z.ZodOptional<z.ZodNumber>;
     tools: z.ZodOptional<z.ZodBoolean>;
+    "cache-type-k": z.ZodOptional<z.ZodString>;
+    "cache-type-v": z.ZodOptional<z.ZodString>;
+    openclCacheDir: z.ZodOptional<z.ZodString>;
     projectionModelSrc: z.ZodOptional<z.ZodUnion<readonly [z.ZodString, z.ZodObject<{
         src: z.ZodString;
         name: z.ZodOptional<z.ZodString>;
@@ -98,6 +106,7 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
             "onnx-tts": "onnx-tts";
             "parakeet-transcription": "parakeet-transcription";
             "onnx-ocr": "onnx-ocr";
+            "sdcpp-generation": "sdcpp-generation";
             llm: "llm";
             whisper: "whisper";
             embeddings: "embeddings";
@@ -105,6 +114,7 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
             parakeet: "parakeet";
             tts: "tts";
             ocr: "ocr";
+            diffusion: "diffusion";
         }>, z.ZodLiteral<"vad">]>>;
     }, z.core.$strip>]>>;
 }, z.core.$strip>, z.ZodTransform<{
@@ -126,6 +136,9 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
     stop_sequences?: string[] | undefined;
     n_discarded?: number | undefined;
     tools?: boolean | undefined;
+    "cache-type-k"?: string | undefined;
+    "cache-type-v"?: string | undefined;
+    openclCacheDir?: string | undefined;
     projectionModelSrc?: string | {
         src: string;
         name?: string | undefined;
@@ -137,7 +150,7 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
         engine?: string | undefined;
         expectedSize?: number | undefined;
         sha256Checksum?: string | undefined;
-        addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "vad" | undefined;
+        addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "diffusion" | "vad" | undefined;
     } | undefined;
 }, {
     ctx_size?: number | undefined;
@@ -158,6 +171,9 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
     stop_sequences?: string[] | undefined;
     n_discarded?: number | undefined;
     tools?: boolean | undefined;
+    "cache-type-k"?: string | undefined;
+    "cache-type-v"?: string | undefined;
+    openclCacheDir?: string | undefined;
     projectionModelSrc?: string | {
         src: string;
         name?: string | undefined;
@@ -169,15 +185,15 @@ export declare const llmConfigSchema: z.ZodPipe<z.ZodObject<{
         engine?: string | undefined;
         expectedSize?: number | undefined;
         sha256Checksum?: string | undefined;
-        addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "vad" | undefined;
+        addon?: "llamacpp-completion" | "whispercpp-transcription" | "llamacpp-embedding" | "nmtcpp-translation" | "onnx-tts" | "parakeet-transcription" | "onnx-ocr" | "sdcpp-generation" | "llm" | "whisper" | "embeddings" | "nmt" | "parakeet" | "tts" | "ocr" | "diffusion" | "vad" | undefined;
     } | 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";
+        gpu: "gpu";
     }>>;
     batchSize: z.ZodOptional<z.ZodNumber>;
     pooling: z.ZodOptional<z.ZodEnum<{
@@ -202,6 +218,7 @@ export declare const embedConfigBaseSchema: z.ZodObject<{
         dedicated: "dedicated";
     }>]>>;
     verbosity: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<0>, z.ZodLiteral<1>, z.ZodLiteral<2>, z.ZodLiteral<3>]>>;
+    openclCacheDir: z.ZodOptional<z.ZodString>;
 }, z.core.$strip>;
 export type EmbedConfigInput = z.infer<typeof embedConfigBaseSchema>;
 export declare const EMBED_CONFIG_DEFAULTS: {
@@ -212,8 +229,8 @@ export declare const EMBED_CONFIG_DEFAULTS: {
 export declare const embedConfigSchema: z.ZodPipe<z.ZodObject<{
     gpuLayers: z.ZodOptional<z.ZodNumber>;
     device: z.ZodOptional<z.ZodEnum<{
-        gpu: "gpu";
         cpu: "cpu";
+        gpu: "gpu";
     }>>;
     batchSize: z.ZodOptional<z.ZodNumber>;
     pooling: z.ZodOptional<z.ZodEnum<{
@@ -238,9 +255,10 @@ export declare const embedConfigSchema: z.ZodPipe<z.ZodObject<{
         dedicated: "dedicated";
     }>]>>;
     verbosity: z.ZodOptional<z.ZodUnion<readonly [z.ZodLiteral<0>, z.ZodLiteral<1>, z.ZodLiteral<2>, z.ZodLiteral<3>]>>;
+    openclCacheDir: z.ZodOptional<z.ZodString>;
 }, z.core.$strip>, z.ZodTransform<{
     gpuLayers: number | undefined;
-    device: "gpu" | "cpu" | undefined;
+    device: "cpu" | "gpu" | undefined;
     batchSize: number | undefined;
     pooling?: "none" | "mean" | "cls" | "last" | "rank" | undefined;
     attention?: "causal" | "non-causal" | undefined;
@@ -248,9 +266,10 @@ export declare const embedConfigSchema: z.ZodPipe<z.ZodObject<{
     flashAttention?: "on" | "off" | "auto" | undefined;
     mainGpu?: number | "integrated" | "dedicated" | undefined;
     verbosity?: 0 | 3 | 2 | 1 | undefined;
+    openclCacheDir?: string | undefined;
 }, {
     gpuLayers?: number | undefined;
-    device?: "gpu" | "cpu" | undefined;
+    device?: "cpu" | "gpu" | undefined;
     batchSize?: number | undefined;
     pooling?: "none" | "mean" | "cls" | "last" | "rank" | undefined;
     attention?: "causal" | "non-causal" | undefined;
@@ -258,6 +277,7 @@ export declare const embedConfigSchema: z.ZodPipe<z.ZodObject<{
     flashAttention?: "on" | "off" | "auto" | undefined;
     mainGpu?: number | "integrated" | "dedicated" | 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

@@ -38,6 +38,13 @@ export const llmConfigBaseSchema = z.object({
     stop_sequences: z.array(z.string()).optional(),
     n_discarded: z.number().optional(),
     tools: z.boolean().optional(),
+    "cache-type-k": z.string().optional(),
+    "cache-type-v": z.string().optional(),
+    /**
+     * Writable directory for OpenCL kernel binary cache. Required on Android
+     * for fast GPU startup.
+     */
+    openclCacheDir: z.string().optional(),
     projectionModelSrc: modelSrcInputSchema.optional(),
 });
 // Default values - typed as partial of the config
@@ -65,6 +72,11 @@ export const embedConfigBaseSchema = z.object({
         .union([z.number().int().min(0), z.enum(["integrated", "dedicated"])])
         .optional(),
     verbosity: verbositySchema.optional(),
+    /**
+     * Writable directory for OpenCL kernel binary cache. Required on Android
+     * for fast GPU startup.
+     */
+    openclCacheDir: z.string().optional(),
 });
 // Default values - typed as partial of the config
 export const EMBED_CONFIG_DEFAULTS = {