@caldrix-ai/caldrix 0.0.0-stage → 0.2.0

package/dist/chunk-PZGTLL5H.js

@@ -0,0 +1 @@
+var o=["probe","fit","architecture","fetch","digest","validate","lowering","backend","request"];function d(n){return typeof n=="string"&&o.includes(n)}var e=class extends Error{kind;missing;constructor(i,t,r){super(t),this.name="CaldrixError",this.kind=i,r!==void 0&&(this.missing=r)}};function s(n){if(n instanceof e)return n;if(typeof n=="object"&&n!==null){let{kind:i,message:t}=n;if(typeof i=="string"&&o.includes(i)&&typeof t=="string")return new e(i,t)}return new e("backend",String(n))}function a(n){return new e("backend",`the WebGPU device failed (${n}); dispose this engine and load the model again`)}export{d as a,e as b,s as c,a as d};

package/dist/index.d.ts

@@ -0,0 +1,252 @@
+/**
+ * The generation options, passed to the host as its camelCase JSON. Any other key reaches the host
+ * and is refused there by name; `signal` is the page's and never leaves the main thread.
+ */
+interface GenerationOptions {
+    /**
+     * Restricts the text to a JSON value matching a schema (a documented subset; see the README) or to
+     * exactly one of some strings. A refusal is a `CaldrixError` of kind `request`.
+     */
+    constraint?: {
+        json: Record<string, unknown>;
+    } | {
+        oneOf: string[];
+    };
+    /** The number of top logprobs to return with each token, at most 20. */
+    logprobs?: number;
+    temperature?: number;
+    topK?: number;
+    topP?: number;
+    minP?: number;
+    repetitionPenalty?: number;
+    seed?: number;
+    maxTokens?: number;
+}
+
+/** The nine error kinds of the caldrix.js API (spec §5.4). */
+type Kind = "probe" | "fit" | "architecture" | "fetch" | "digest" | "validate" | "lowering" | "backend" | "request";
+declare class CaldrixError extends Error {
+    readonly kind: Kind;
+    /** Set by the Hub fetcher when the file is simply absent: not on the Hub (404) or not cached while offline. */
+    readonly missing?: "not-found" | "offline-uncached";
+    constructor(kind: Kind, message: string, missing?: "not-found" | "offline-uncached");
+}
+
+type ChatMessage = {
+    role: "system" | "user" | "assistant";
+    content: string;
+};
+/**
+ * Where one model file came from on this load, and whether its sha256 was checked against the Hub's
+ * `X-Linked-Etag` (§5.2): `verified: false` means the Hub sent no such header for it (a cached entry
+ * carries what was recorded when it was written).
+ */
+interface FileSource {
+    source: "cache" | "network";
+    verified: boolean;
+}
+/**
+ * What the host's `doctor()` returns, plus the worker's facts about the tier it opened; `source` is
+ * the page's to fill. `backend` is the tier (`"wasm"` or `"webgpu"`).
+ */
+interface HostDoctor {
+    backend: string;
+    device: string;
+    dtype: string | null;
+    simd: boolean;
+    threads: number | null;
+    model: string | null;
+    revision: string | null;
+    source: null;
+    line: string;
+    /** WebGPU only, the host's: the context the model was loaded at, and the clamp's line when the default was clamped. */
+    maxContext?: number;
+    contextClamp?: string | null;
+    /** The worker's: the weights' storage form on WebGPU (`null` on wasm). */
+    storage: "f16" | "f32" | null;
+    /**
+     * The worker's: the adapter by its own vendor and architecture strings (`null` on wasm). On
+     * WebGPU `device` is the worker's too: `webgpu (<description>)`, with vendor and architecture
+     * standing in for the description when the browser leaves it empty.
+     */
+    adapter: {
+        vendor: string;
+        architecture: string;
+    } | null;
+    /** The worker's, from the host's `listAliases`: every alias this build loads, by name and file format. */
+    aliases: {
+        name: string;
+        format: "safetensors" | "gguf";
+    }[];
+}
+
+interface LoadOptions {
+    backend?: "auto" | "webgpu" | "wasm" | "webgl";
+    offline?: boolean;
+    onProgress?: (p: {
+        file: string;
+        loaded: number;
+        total: number | null;
+    }) => void;
+    hubBaseUrl?: string;
+    dtype?: "f32" | "bf16";
+    maxContext?: number;
+    /**
+     * The WebGPU tier's weight storage: `"auto"` (the default) stores the block GEMM weights f16 when
+     * the adapter has `shader-f16`, `"f32"` keeps them f32 on any adapter. Under `backend: "auto"` a
+     * load the wasm tier serves ignores it; an explicit `backend: "wasm"` with `"f32"` is refused.
+     */
+    storage?: "auto" | "f32";
+    /**
+     * Cancels the load: before the model is loaded, an abort terminates the engine's worker (and every
+     * fetch in it) and the load rejects with `fetch`; an already-aborted signal rejects before the
+     * worker starts. After the load resolves the signal is ignored.
+     */
+    signal?: AbortSignal;
+    /**
+     * Where the package's own files come from. `worker` is the engine's worker script (default:
+     * `./worker.js` next to this module, i.e. `dist/worker.js`); `hostWasm` is the browser host's
+     * `.wasm` (default: `../host/caldrix_web_bg.wasm` next to the worker, i.e. `host/caldrix_web_bg.wasm`).
+     * A string, absolute or relative, is resolved against the page (`location.href`) on the main
+     * thread, not against the worker. A worker on another origin than the page's starts through a
+     * same-origin module shim that imports it, so that origin must serve the worker, its chunks,
+     * `host/` and the `.wasm` with CORS.
+     *
+     * The CDN fallback, on unless `fallback` is `false`: when the bundle's own copy fails, it is fetched
+     * from jsdelivr at this package's installed version (`cdn`, default
+     * `https://cdn.jsdelivr.net/npm/@caldrix-ai/caldrix@<version>/`), independently for each asset. The worker
+     * falls back when it fails before it is ready (its script, a chunk or the host glue
+     * `host/caldrix_web.js` did not load, or the page refused to construct it); a worker from the CDN
+     * then fetches its `.wasm` from the CDN as well. The `.wasm` falls back, inside the worker, when its
+     * fetch fails (a network error or a non-OK status). Every fallback is reported in `doctor().assets`
+     * and once per load on the console; it is never silent. An explicit `worker` or `hostWasm` is used
+     * as given, with no fallback for that asset.
+     */
+    assets?: {
+        worker?: string | URL;
+        hostWasm?: string | URL;
+        fallback?: boolean;
+        cdn?: string | URL;
+    };
+}
+interface Delta {
+    text: string;
+    /**
+     * The token id this delta carries. Absent only on the single `done` delta a generation yields when
+     * its signal was already aborted at the call: nothing reached the worker, so no token exists.
+     */
+    token?: number;
+    done: boolean;
+    finishReason?: "stop" | "length" | "abort";
+    /** Prompt tokens, generated tokens, and the prompt tokens served from the previous generation's context. */
+    usage?: {
+        promptTokens: number;
+        completionTokens: number;
+        cachedTokens: number;
+    };
+    /** The top logprobs for this token; present only when `logprobs` was asked for. */
+    logprobs?: {
+        token: number;
+        logprob: number;
+    }[];
+}
+/**
+ * The host's doctor object, with `source` filled per file: where it came from and whether its digest
+ * was checked. `backend` is the tier that runs; `why` is how it was chosen (the probe's reason under
+ * `"auto"`, so a fallback to wasm says why -- no adapter, a bf16 request, a quantized alias, or
+ * `webgpu refused: …` when the WebGPU precheck or device open refused a load the wasm tier then served).
+ * `aliases` is every alias this build loads, each with its file format.
+ */
+interface DoctorReport extends Omit<HostDoctor, "source"> {
+    source: Record<string, FileSource>;
+    why: string;
+    /** Where the package's own files came from on this load (`LoadOptions.assets`). */
+    assets: {
+        worker: AssetSource;
+        hostWasm: AssetSource;
+    };
+}
+/**
+ * One of the package's own files: `bundle`, the default beside the bundle that imported caldrix.js;
+ * `cdn`, the CDN fallback (`assets.cdn`); `page`, a URL the page gave (`assets.worker` or
+ * `assets.hostWasm`, or the default `.wasm` beside a page-given worker). `url` is the one fetched.
+ */
+interface AssetSource {
+    source: "bundle" | "cdn" | "page";
+    url: string;
+}
+type GenerateOptions = GenerationOptions & {
+    signal?: AbortSignal;
+};
+declare class Caldrix {
+    /** Load a model by alias into its own worker. */
+    static load(alias: string, opts?: LoadOptions): Promise<Engine>;
+}
+declare class Engine {
+    #private;
+    /** Use `Caldrix.load`. */
+    constructor(worker: Worker, report: DoctorReport);
+    chat(messages: ChatMessage[], opts?: GenerateOptions): AsyncIterable<Delta>;
+    complete(input: string | {
+        ids: number[];
+    }, opts?: GenerateOptions): AsyncIterable<Delta>;
+    doctor(): DoctorReport;
+    /** Release the worker and everything in it. Every later call throws `request`. */
+    dispose(): void;
+}
+
+type WasmFeature = "simd" | "relaxedSimd" | "bulkMemory" | "threads" | "memory64" | "tailCall" | "exceptions";
+type WasmFeatures = Record<WasmFeature, boolean>;
+
+interface WebGpuAdapter {
+    vendor: string;
+    architecture: string;
+    device: string;
+    description: string;
+    features: string[];
+    /** Every numeric limit as the browser gave it (one past `Number.MAX_SAFE_INTEGER` among them). */
+    limits: Record<string, number>;
+    /**
+     * The limits the host cannot read, each as `name = value (why)`: not a safe integer, or not a
+     * number. `hostProbeJson` leaves them out, so the host reports them as missing; `why` names them.
+     */
+    unrepresentable: string[];
+}
+interface Capabilities {
+    wasm: WasmFeatures;
+    threads: {
+        sharedArrayBuffer: boolean;
+        crossOriginIsolated: boolean;
+        hardwareConcurrency: number | null;
+    };
+    webgl: {
+        available: boolean;
+        maxTextureSize: number | null;
+        floatColorBuffer: boolean | null;
+    };
+    webgpu: WebGpuAdapter | null;
+    memory: {
+        deviceMemoryGiB: number | null;
+    };
+    tiers: {
+        webgpu: {
+            available: boolean;
+            built: true;
+            reason?: string;
+        };
+        webgl: {
+            available: boolean;
+            built: false;
+            reason: "not built";
+        };
+        wasm: {
+            available: true;
+            built: true;
+        };
+    };
+    chosen: "wasm" | "webgpu";
+    why: string;
+}
+declare function probe(): Promise<Capabilities>;
+
+export { type AssetSource, Caldrix, CaldrixError, type Capabilities, type ChatMessage, type Delta, type DoctorReport, Engine, type GenerateOptions, type GenerationOptions, type Kind, type LoadOptions, probe };

package/dist/index.js

@@ -0,0 +1 @@
+import{a as J,b as i,d as H}from"./chunk-PZGTLL5H.js";function v(e){return typeof e=="object"&&e!==null}function se(e){return v(e)&&typeof e.token=="number"&&typeof e.text=="string"&&typeof e.done=="boolean"&&(e.finishReason===null||e.finishReason==="stop"||e.finishReason==="length"||e.finishReason==="abort")&&typeof e.promptTokens=="number"&&typeof e.completionTokens=="number"&&typeof e.cachedTokens=="number"&&(e.logprobs===void 0||Array.isArray(e.logprobs)&&e.logprobs.every(t=>v(t)&&typeof t.token=="number"&&typeof t.logprob=="number"))}function ae(e){return!v(e)||typeof e.url!="string"?!1:e.failed===void 0||v(e.failed)&&typeof e.failed.url=="string"&&typeof e.failed.reason=="string"}function L(e){if(!v(e)||typeof e.id!="number")return!1;switch(e.type){case"ready":return e.id===0&&typeof e.url=="string";case"host":return ae(e.hostWasm);case"progress":return typeof e.file=="string"&&typeof e.loaded=="number"&&(e.total===null||typeof e.total=="number");case"loaded":return v(e.doctor)&&v(e.sources)&&(e.fallback===void 0||typeof e.fallback=="string")&&Object.values(e.sources).every(t=>v(t)&&(t.source==="cache"||t.source==="network")&&typeof t.verified=="boolean");case"delta":return se(e.delta);case"result":return"value"in e;case"error":return J(e.kind)&&typeof e.message=="string"&&(e.fatal===void 0||typeof e.fatal=="string");default:return!1}}function m(e){let t;try{t=JSON.stringify(e)??String(e)}catch{t=String(e)}return t.length>200&&(t=`${t.slice(0,200)}...`),new i("backend",`the worker sent a malformed message: ${t}`)}var x=(...e)=>new Uint8Array(e),ie={simd:x(0,97,115,109,1,0,0,0,1,5,1,96,0,1,123,3,2,1,0,10,10,1,8,0,65,0,253,15,253,98,11),relaxedSimd:x(0,97,115,109,1,0,0,0,1,5,1,96,0,1,123,3,2,1,0,10,15,1,13,0,65,1,253,15,65,2,253,15,253,128,2,11),bulkMemory:x(0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,5,3,1,0,1,10,14,1,12,0,65,0,65,0,65,0,252,10,0,0,11),threads:x(0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,5,4,1,3,1,1,10,11,1,9,0,65,0,254,16,2,0,26,11),memory64:x(0,97,115,109,1,0,0,0,5,3,1,4,1),tailCall:x(0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,10,6,1,4,0,18,0,11),exceptions:x(0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,10,16,1,14,0,2,105,31,64,1,3,0,0,11,0,11,26,11)};function X(){let e={};for(let[t,r]of Object.entries(ie))e[t]=WebAssembly.validate(r);return e}var le=1e3,E=Symbol("timed out");async function K(e){let t,r=new Promise(n=>{t=setTimeout(()=>n(E),le)});try{return await Promise.race([e,r])}finally{clearTimeout(t)}}var z=e=>e instanceof Error?e.message:String(e);function V(){let e=globalThis.navigator;return typeof e=="object"&&e!==null?e:{}}function D(e){return typeof e=="number"&&Number.isFinite(e)?e:null}function ue(){let e={available:!1,maxTextureSize:null,floatColorBuffer:null};try{let t=null;if(typeof OffscreenCanvas<"u"?t=new OffscreenCanvas(1,1):typeof document<"u"&&(t=document.createElement("canvas")),t===null)return e;let r=t,n=r.getContext("webgl2"),o=n??r.getContext("webgl");if(!o)return e;let s=D(o.getParameter(o.MAX_TEXTURE_SIZE)),l=o.getExtension(n?"EXT_color_buffer_float":"WEBGL_color_buffer_float")!==null;return o.getExtension("WEBGL_lose_context")?.loseContext(),{available:!0,maxTextureSize:s,floatColorBuffer:l}}catch{return e}}async function de(e){let t=e.info??{};if(!e.info&&typeof e.requestAdapterInfo=="function")try{t=await e.requestAdapterInfo()}catch{t={}}let r={},n=[];if(e.limits)for(let o in e.limits){let s=e.limits[o];typeof s=="number"?(r[o]=s,Number.isSafeInteger(s)||n.push(`${o} = ${s} (not a safe integer)`)):typeof s!="function"&&n.push(`${o} = ${JSON.stringify(s)??String(s)} (not a number)`)}return{vendor:t.vendor??"",architecture:t.architecture??"",device:t.device??"",description:t.description??"",features:e.features?[...e.features]:[],limits:r,unrepresentable:n}}async function ce(){let e=V().gpu;if(!e||typeof e.requestAdapter!="function")return{webgpu:null,available:!1,reason:"no navigator.gpu"};let t;try{let n=await K(e.requestAdapter());if(n===E)return{webgpu:null,available:!1,reason:"requestAdapter timed out"};t=n}catch(n){return{webgpu:null,available:!1,reason:`requestAdapter failed: ${z(n)}`}}if(!t)return{webgpu:null,available:!1,reason:"requestAdapter returned null"};let r=await de(t);try{let n=await K(t.requestDevice());if(n===E)return{webgpu:r,available:!1,reason:"requestDevice timed out"};n.destroy?.()}catch(n){return{webgpu:r,available:!1,reason:`requestDevice refused: ${z(n)}`}}return{webgpu:r,available:!0}}function fe(e){return e.webgpu.available&&e.webgpu.built?"webgpu":"wasm"}function pe(e,t,r){let n=e.webgpu?.unrepresentable??[],o=n.length>0?`; adapter limits ${n.join(", ")}: the host reads them as missing`:"";return be(e,t,r)+o}function be(e,t,r){if(r==="webgpu"){let o=e.webgpu,s=o.features.includes("shader-f16")?" with shader-f16":"";return`WebGPU adapter present (${o.vendor} ${o.architecture})${s}: webgpu`}let n=e.wasm.simd?"with simd128":"without simd128";return e.webgpu===null?`no WebGPU adapter (${t.reason??"unavailable"}): wasm ${n}`:`WebGPU adapter present but unusable (${t.reason}): wasm ${n}`}async function G(){let e=V(),t=await ce(),r={wasm:X(),threads:{sharedArrayBuffer:typeof SharedArrayBuffer<"u",crossOriginIsolated:globalThis.crossOriginIsolated===!0,hardwareConcurrency:D(e.hardwareConcurrency)},webgl:ue(),webgpu:t.webgpu,memory:{deviceMemoryGiB:D(e.deviceMemory)},tiers:{webgpu:t.available?{available:!0,built:!0}:{available:!1,built:!0,reason:t.reason},webgl:{available:!1,built:!1,reason:"not built"},wasm:{available:!0,built:!0}}};r.tiers.webgl.available=r.webgl.available;let n=fe(r.tiers);return{...r,chosen:n,why:pe(r,t,n)}}function Q(e){let t=e.webgpu===null?null:{vendor:e.webgpu.vendor,architecture:e.webgpu.architecture,device:e.webgpu.device,description:e.webgpu.description,features:e.webgpu.features,limits:Object.fromEntries(Object.entries(e.webgpu.limits).filter(([,r])=>Number.isSafeInteger(r)))};return JSON.stringify({wasm:{simd:e.wasm.simd,relaxedSimd:e.wasm.relaxedSimd,bulkMemory:e.wasm.bulkMemory,threads:e.wasm.threads,memory64:e.wasm.memory64,tailCall:e.wasm.tailCall,exceptions:e.wasm.exceptions},threads:{sharedArrayBuffer:e.threads.sharedArrayBuffer,crossOriginIsolated:e.threads.crossOriginIsolated,hardwareConcurrency:e.threads.hardwareConcurrency},webgpu:t,memory:{deviceMemoryGiB:e.memory.deviceMemoryGiB}})}var me="https://huggingface.co";function ge(e){return typeof location<"u"?new URL(String(e),location.href):new URL(String(e))}var he="https://cdn.jsdelivr.net/npm/@caldrix-ai/caldrix@0.2.0/";function Y(e){return e instanceof Error?e.message:String(e)}function P(e,t){if(t!==void 0){if(typeof t!="string"&&!(t instanceof URL))throw new i("request",`assets.${e} is a string or a URL, not ${typeof t}`);try{return typeof location<"u"?new URL(t,location.href):new URL(t)}catch{throw new i("request",`assets.${e} ${JSON.stringify(String(t))} is not a URL`)}}}function we(e,t){let r=I(e,t),n=null,o=()=>{n!==null&&URL.revokeObjectURL(n),n=null};try{return!r&&!t?{worker:new Worker(new URL("./worker.js",import.meta.url),{type:"module"}),revoke:o,literal:!0}:r?(n=URL.createObjectURL(new Blob([`import ${JSON.stringify(e.href)};`],{type:"text/javascript"})),{worker:new Worker(n,{type:"module"}),revoke:o}):{worker:new Worker(e,{type:"module"}),revoke:o}}catch(s){throw o(),s}}function ye(e){return typeof location<"u"&&e.origin!==location.origin}function I(e,t){return ye(e)&&(t||e.protocol==="http:"||e.protocol==="https:")}function q(e){return!e.chosen&&!I(e.url,!1)?`the package's worker (dist/worker.js as the page's bundler emitted it, or ${e.url.href} unbundled)`:`the engine's worker from ${e.url.href}${I(e.url,e.chosen)?" (through a same-origin shim)":""}`}var C=class{constructor(t){this.reason=t}reason};function ke(e,t,r){return new Promise((n,o)=>{let s;try{s=we(e.url,e.chosen)}catch(d){o(new C(`the page could not construct it: ${Y(d)}`));return}let{worker:l,revoke:f}=s,b=()=>u(r()),u=(d,a=e.url)=>{if(f(),l.onmessage=null,l.onerror=null,t?.removeEventListener("abort",b),d===null)return n({worker:l,url:a});l.terminate(),o(d)};t?.addEventListener("abort",b,{once:!0}),l.onerror=d=>u(new C(d.message||"it did not load")),l.onmessage=d=>{let a=d.data;if(!L(a)||a.id!==0)return u(m(a));if(a.type==="ready")return s.literal?URL.canParse(a.url)?u(null,new URL(a.url)):u(m(a)):u(null);if(a.type==="error")return u(new C(a.message));u(m(a))}})}async function ve(e,t,r,n){let o=[];for(let u of e)try{let{worker:d,url:a}=await ke(u,r,n),g=o[0];return{worker:d,used:u,url:a,...g!==void 0?{failed:{what:q(g.t),reason:g.reason}}:{}}}catch(d){if(!(d instanceof C))throw d;o.push({t:u,reason:d.reason})}let[s,l]=o;if(s!==void 0&&l!==void 0){let u=l.t.url.origin;throw new i("backend",`${q(s.t)} did not start (${s.reason}), nor from the CDN fallback ${l.t.url.href} (${l.reason}); serve dist/ and host/ beside the bundle, set assets.worker to where the page serves them, or allow ${u} in the page's Content-Security-Policy (worker-src blob: for the same-origin shim that starts it, script-src ${u} for the modules it imports, connect-src ${u} for the .wasm)`)}let f=s,b=t&&f.t.source==="bundle"?"; the CDN fallback is off (assets.fallback is false)":"";throw new i("backend",`${q(f.t)} did not start: ${f.reason}${b}; caldrix.js runs inference in a Web Worker and has no main-thread mode`)}function We(e,t,r){let n=r!==void 0?`, and with it the host .wasm from ${r}`:"";return`${e.what} did not start (${e.reason}); it was loaded from the CDN fallback ${t}${n}`}function xe(e){return`the host .wasm could not be fetched from ${e.failed.url} (${e.failed.reason}); it was fetched from the CDN fallback ${e.url}`}function Z(e){console.warn(`caldrix.js: ${e}. Serve dist/ and host/ beside the bundle or set assets.worker/assets.hostWasm; assets.fallback: false turns the fallback off.`)}function Re(e){let t={text:e.text,token:e.token,done:e.done};return e.logprobs!==void 0&&(t.logprobs=e.logprobs),e.done&&(e.finishReason!==null&&(t.finishReason=e.finishReason),t.usage={promptTokens:e.promptTokens,completionTokens:e.completionTokens,cachedTokens:e.cachedTokens}),t}var M=class{static async load(t,r={}){let n=r.backend??"auto";if(n==="webgl")throw new i("backend","the webgl backend is not built");if(n!=="auto"&&n!=="webgpu"&&n!=="wasm")throw new i("request",`backend ${JSON.stringify(n)} is not auto, webgpu, wasm or webgl`);if(typeof t!="string"||t==="")throw new i("request","load takes a model alias, a non-empty string");let o=P("worker",r.assets?.worker),s=P("hostWasm",r.assets?.hostWasm),l=r.assets?.fallback??!0;if(typeof l!="boolean")throw new i("request",`assets.fallback is true or false, not ${JSON.stringify(l)}`);let f=P("cdn",r.assets?.cdn)??new URL(he);f.pathname.endsWith("/")||(f.pathname+="/");let b=[o!==void 0?{url:o,chosen:!0,source:"page"}:{url:ge(new URL("./worker.js",import.meta.url)),chosen:!1,source:"bundle"}];l&&o===void 0&&b.push({url:new URL("dist/worker.js",f),chosen:!0,source:"cdn"});let u=l&&s===void 0?new URL("host/caldrix_web_bg.wasm",f).href:void 0,d=r.storage??"auto";if(d!=="auto"&&d!=="f32")throw new i("request",`storage ${JSON.stringify(d)} is not auto or f32`);if(n==="wasm"&&d==="f32")throw new i("request",`storage "f32" is the WebGPU tier's weight storage; backend "wasm" takes none`);let a=()=>new i("fetch",`loading ${t} was aborted`),g=r.signal;if(g?.aborted)throw a();let p=await G(),W,T;if(n==="auto"&&r.dtype==="bf16"&&p.chosen==="webgpu")W="wasm",T=`bf16 requested: wasm (the WebGPU tier runs the f32 plan only); ${p.why}`;else if(n==="auto")W=p.chosen,T=p.why;else if(n==="webgpu"){if(!p.tiers.webgpu.available)throw new i("backend",`backend "webgpu" was asked for, but the probe found no usable WebGPU adapter: ${p.tiers.webgpu.reason??p.why}`);W="webgpu",T=`backend "webgpu" asked for; ${p.why}`}else W="wasm",T=`backend "wasm" asked for; ${p.why}`;let B=p.webgpu?.unrepresentable??[];if(g?.aborted)throw a();let $=await ve(b,!l,g,a),{worker:y}=$,O=$.url,R=null;$.failed!==void 0&&(R=We($.failed,O.href,s===void 0?u:void 0),Z(R));let F=U=>R===null?U:new i(U.kind,`${U.message} (after a CDN fallback: ${R})`,U.missing);if(g?.aborted)throw y.terminate(),F(a());let _=1,ee={id:_,type:"load",alias:t,probeJson:Q(p),request:{backend:W,...r.dtype!==void 0?{dtype:r.dtype}:{},...r.maxContext!==void 0?{maxContext:r.maxContext}:{},...W==="webgpu"&&d==="f32"?{f16:!1}:{}},...n==="auto"&&W==="webgpu"?{auto:!0}:{},offline:r.offline??!1,hubBaseUrl:r.hubBaseUrl??me,...s!==void 0?{hostWasm:s.href}:{},...u!==void 0?{cdnWasm:u}:{}};return new Promise((U,te)=>{let j=()=>k(a()),N=()=>{y.onmessage=null,y.onerror=null,g?.removeEventListener("abort",j)},k=A=>{N(),y.terminate(),te(F(A))},h=null;g?.addEventListener("abort",j,{once:!0}),y.onerror=A=>k(new i("backend",`the engine's worker from ${O.href} failed: ${A.message||"it did not start"}`)),y.onmessage=A=>{let c=A.data;if(!L(c)||c.id!==_)return k(m(c));switch(c.type){case"host":if(h!==null)return k(m(c));h=c.hostWasm,h.failed!==void 0&&(R=xe({...h,failed:h.failed}),Z(R));return;case"progress":try{r.onProgress?.({file:c.file,loaded:c.loaded,total:c.total})}catch(w){k(w instanceof i?w:new i("request",`onProgress threw: ${Y(w)}`))}return;case"loaded":{if(h===null)return k(m(c));N();let w=$.used.source,re=s!==void 0||w==="page"&&h.failed===void 0?"page":h.failed!==void 0||w==="cdn"?"cdn":"bundle",ne={worker:{source:w,url:O.href},hostWasm:{source:re,url:h.url}},oe={...c.doctor,source:c.sources,why:c.fallback!==void 0?`${c.fallback}; ${p.why}`:T,assets:ne};U(new S(y,oe));return}case"error":{let w=c.kind==="probe"&&B.length>0?`; the browser gave ${B.join(", ")}, which the host reads as missing`:"";return k(new i(c.kind,c.message+w))}default:return k(m(c))}},y.postMessage(ee)})}},S=class{#r;#n;#l=2;#t=!1;#e=null;#i=!1;#o=null;constructor(t,r){this.#r=t,this.#n=r,t.onmessage=n=>this.#d(n.data),t.onerror=n=>this.#a(new i("backend",`the engine's worker failed: ${n.message||"it stopped"}`))}chat(t,r={}){return this.#s(),this.#u({messages:t},r)}complete(t,r={}){this.#s();let n;if(typeof t=="string")n={text:t};else if(typeof t=="object"&&t!==null){let o=t;if(o.text!==void 0&&o.ids!==void 0)throw new i("request","complete takes exactly one of text and ids, not both");if(!Array.isArray(o.ids))throw new i("request","complete takes exactly one of text and ids");n={ids:o.ids}}else throw new i("request","complete takes exactly one of text and ids");return this.#u(n,r)}doctor(){this.#s();let t=Object.fromEntries(Object.entries(this.#n.source).map(([s,l])=>[s,{...l}])),{worker:r,hostWasm:n}=this.#n.assets,o=this.#n.aliases.map(s=>({...s}));return{...this.#n,source:t,aliases:o,assets:{worker:{...r},hostWasm:{...n}}}}dispose(){if(!this.#t){this.#t=!0;try{this.#r.postMessage({id:this.#l++,type:"dispose"})}catch{}this.#r.terminate(),this.#o?.sink(new i("request","the engine was disposed"))}}#s(){if(this.#t)throw new i("request","the engine was disposed; load a new one");if(this.#e)throw this.#e}#a(t){this.#e||this.#t||(this.#e=t,this.#r.terminate(),this.#o?.sink(t))}#d(t){if(this.#t||this.#e)return;if(!L(t))return this.#a(m(t));let r=this.#o;if(r===null||t.id!==r.id)return this.#a(m(t));switch(t.type){case"delta":return r.sink(Re(t.delta));case"error":return t.fatal!==void 0&&(this.#e=H(t.fatal)),r.sink(new i(t.kind,t.message));default:return this.#a(m(t))}}async*#u(t,r){if(this.#s(),this.#i)throw new i("request","a generation is already running on this engine");let{signal:n,...o}=r;if(n?.aborted){yield{text:"",done:!0,finishReason:"abort",usage:{promptTokens:0,completionTokens:0,cachedTokens:0}};return}this.#i=!0;let s=this.#l++,l=[],f=null;this.#o={id:s,sink:a=>{l.push(a),f?.()}};let b=()=>{!this.#t&&!this.#e&&this.#r.postMessage({id:s,type:"abort"})},u=!1,d=async()=>{for(;l.length===0;)await new Promise(a=>f=a);return f=null,l.shift()};try{for(this.#r.postMessage({id:s,type:"generate",prompt:t,options:o}),n?.addEventListener("abort",b,{once:!0});;){let a=await d();if(a instanceof i)throw u=!0,a;if(a.done&&(u=!0),yield a,a.done)return}}finally{if(n?.removeEventListener("abort",b),!u)try{for(b();!(this.#t||this.#e);){let a=await d();if(a instanceof i||a.done)break}}catch{}this.#o=null,this.#i=!1}}};export{M as Caldrix,i as CaldrixError,S as Engine,G as probe};

package/dist/worker.js

@@ -0,0 +1 @@
+import{b as c,c as E,d as N}from"./chunk-PZGTLL5H.js";var ke="caldrix.js",ee="x-caldrix-verified",J="x-caldrix-absent";function xe(n){return[...new Uint8Array(n)].map(r=>r.toString(16).padStart(2,"0")).join("")}function ve(n){return n.trim().replace(/^W\//,"").replace(/^"|"$/g,"").toLowerCase()}function R(n){return n instanceof Error?n.message:String(n)}function I(n,r){if(r===""||r==="."||r===".."||/[/\\?#]/.test(r)||/%[0-9a-fA-F]{2}/.test(r)||/[\u0000-\u001f\u007f]/.test(r))throw new c("validate",`${n} ${JSON.stringify(r)} is not a single safe path segment`);return encodeURIComponent(r)}function te(n){return new c("backend",`${n} is not available; caldrix.js needs a secure context`)}function re(n,r,s,e){let t=n.split("/").map(u=>I("repo",u)).join("/");return`${e.endsWith("/")?e.slice(0,-1):e}/${t}/resolve/${I("revision",r)}/${I("file",s)}`}async function ne(n){if(typeof caches>"u")throw te("caches");try{return await caches.open(ke)}catch(r){throw new c("backend",`opening the cache for ${n} failed: ${R(r)}`)}}async function z(n){if(typeof crypto>"u"||!crypto.subtle)throw te("crypto.subtle");return xe(await crypto.subtle.digest("SHA-256",n))}function se(n){return new c("fetch",`${n} is not in the cache and offline is set`,"offline-uncached")}async function oe(n,r,s,e){let t=re(n,r,s,e.hubBaseUrl),i=await ne(s),u;try{u=await i.match(t)}catch(w){throw new c("backend",`reading ${s} from the cache failed: ${R(w)}`)}let a=w=>new c("fetch",`fetching ${s} failed: HTTP 404${w}`,"not-found");if(u&&u.headers.get(J)==="1"){if(e.rememberAbsent)throw a(" (remembered from an earlier load)");u=void 0}let l=!1;if(u){let w=new Uint8Array(await u.arrayBuffer());if(e.sha256===void 0)return{bytes:w,source:"cache",verified:u.headers.get(ee)==="1",cached:!0};if(await z(w)===e.sha256)return{bytes:w,source:"cache",verified:!0,cached:!0};l=!0;try{await i.delete(t)}catch{}}if(e.offline&&l)throw new c("fetch",`${s}: the cached copy fails the alias's sha256 pin and offline is set`,"offline-uncached");if(e.offline)throw se(s);let f=(w,S)=>new c("fetch",`fetching ${s} ${w}`,S),y=()=>{if(e.signal?.aborted)throw f("was aborted")},m;try{m=await fetch(t,{signal:e.signal})}catch(w){throw f(e.signal?.aborted?"was aborted":`failed: ${R(w)}`)}if(m.status===404&&e.rememberAbsent){try{await i.put(t,new Response(null,{headers:{[J]:"1"}}))}catch{}throw a("")}if(!m.ok||!m.body)throw f(`failed: HTTP ${m.status}`,m.status===404?"not-found":void 0);let x=m.headers.get("x-linked-etag"),p=m.headers.get("content-length"),o=p!==null&&/^\d+$/.test(p.trim())?Number(p):null,d=[],b=0,h=m.body.getReader();try{for(;;){let w;try{w=await h.read()}catch(S){throw f(e.signal?.aborted?"was aborted":`failed: ${R(S)}`)}if(w.done)break;if(d.push(w.value),b+=w.value.byteLength,e.bytes!==void 0&&b>e.bytes)throw new c("fetch",`${s}: the body passed the alias's ${e.bytes} bytes`);e.onProgress?.({file:s,loaded:b,total:o})}}finally{h.cancel().catch(()=>{})}if(y(),o!==null&&o!==b&&!m.headers.get("content-encoding"))throw f(`failed: Content-Length is ${o} but ${b} bytes arrived`);let g=new Uint8Array(b),W=0;for(let w of d)g.set(w,W),W+=w.byteLength;e.onProgress?.({file:s,loaded:b,total:o});let A=!1,G=null;if(e.sha256!==void 0){if(G=await z(g),G!==e.sha256)throw new c("digest",`${s}: sha256 is ${G} but the alias pins ${e.sha256}`);A=!0}if(x!==null){let w=ve(x);if(G??=await z(g),w!==G)throw new c("digest",`${s}: sha256 is ${G} but the Hub's X-Linked-Etag is ${w}`);A=!0}y();try{await i.put(t,new Response(g,{headers:{[ee]:A?"1":"0"}}))}catch{return{bytes:g,source:"network",verified:A,cached:!1}}return{bytes:g,source:"network",verified:A,cached:!0}}async function ie(n,r,s,e,t){if(!Number.isSafeInteger(e)||e<=0)throw new c("request",`a range of ${e} bytes of ${s} is not a positive integer`);let i=re(n,r,s,t.hubBaseUrl),u=await ne(s),a;try{a=await u.match(i)}catch(p){throw new c("backend",`reading ${s} from the cache failed: ${R(p)}`)}if(a&&a.headers.get(J)!=="1"){let p=new Uint8Array(await a.arrayBuffer());if(p.byteLength>=e)return p.slice(0,e)}if(t.offline)throw se(s);let l=p=>new c("fetch",`fetching the first ${e} bytes of ${s} ${p}`),f;try{f=await fetch(i,{headers:{Range:`bytes=0-${e-1}`},signal:t.signal})}catch(p){throw l(t.signal?.aborted?"was aborted":`failed: ${R(p)}`)}if(f.status!==206&&f.status!==200||!f.body)throw l(`failed: HTTP ${f.status}`);let y=new Uint8Array(e),m=0,x=f.body.getReader();try{for(;;){let p;try{p=await x.read()}catch(d){throw l(t.signal?.aborted?"was aborted":`failed: ${R(d)}`)}if(p.done)break;let o=Math.min(p.value.byteLength,e-m);if(y.set(p.value.subarray(0,o),m),m+=p.value.byteLength,m>e||f.status===200&&m===e)break}}finally{x.cancel().catch(()=>{})}if(t.signal?.aborted)throw l("was aborted");if(m<e)throw l(`failed: the HTTP ${f.status} answer carried ${m} bytes, not the ${e} asked for`);if(f.status===206&&m>e)throw l(`failed: the HTTP 206 answer carried more than the ${e} bytes asked for`);return y}var j=null,T=null;function Y(n){return n instanceof Error?n.message:String(n)}async function K(){return await import("../host/caldrix_web.js")}function Pe(n){let r=globalThis.location?.href;return(r!==void 0?new URL(String(n),r):new URL(String(n))).href}async function ae(n){let r;try{r=await fetch(n)}catch(s){throw new Error(Y(s))}if(!r.ok)throw new Error(`${r.status} ${r.statusText}`.trim());return r}async function Q(n,r){if(j===null){let s=await K(),e=n??Pe(new URL("../host/caldrix_web_bg.wasm",import.meta.url)),t,i;try{t=await ae(e),i={url:e}}catch(u){let a=Y(u);if(r===void 0||r===e)throw new c("backend",`the browser host's .wasm could not be fetched from ${e} (${a}); serve host/ beside dist/ or set assets.hostWasm to where the page serves it`);try{t=await ae(r)}catch(l){let f=new URL(r).origin;throw new c("backend",`the browser host's .wasm could not be fetched from ${e} (${a}) nor from the CDN fallback ${r} (${Y(l)}); serve host/ beside dist/, set assets.hostWasm to where the page serves it, or allow ${f} in the page's Content-Security-Policy (connect-src ${f}, the .wasm is fetched)`)}i={url:r,failed:{url:e,reason:a}}}await s.default({module_or_path:t}),j=s,T=i}return j}async function ue(n,r,s){let e=await Q(r,s);return new e.CaldrixWeb(n)}function le(){if(T===null)throw new c("backend","the browser host is not open");return{...T,...T.failed!==void 0?{failed:{...T.failed}}:{}}}async function ce(){return(await Q()).CaldrixWeb}async function fe(n,r){let s=await Q(),e;try{let{instance:u}=await WebAssembly.instantiate(r,{env:{memory:s.CaldrixWeb.memory()}});e=u.exports}catch(u){throw new c("backend",`instantiating the emitted module failed: ${u instanceof Error?u.message:String(u)}`)}let{step:t,prefill:i}=e;if(typeof t!="function"||typeof i!="function")throw new c("backend","the emitted module does not export step and prefill");if(n.attach(t,i),n.argsPtr()===0)throw new c("backend","the host reports no model loaded after load")}var B={MAP_READ:1,COPY_SRC:4,COPY_DST:8,UNIFORM:64,STORAGE:128},$e=1,We=["maxStorageBufferBindingSize","maxBufferSize","maxComputeWorkgroupsPerDimension","maxStorageBuffersPerShaderStage"],de=(n,r)=>Number.isSafeInteger(n)&&n>=0&&n<r,pe=n=>Number.isSafeInteger(n)&&n>=0;function Ae(n,r){let s=t=>new c("backend",`the host's WebGPU program is malformed: ${t}`),e;try{e=JSON.parse(n)}catch(t){throw s(t instanceof Error?t.message:String(t))}if(typeof e!="object"||e===null||!Array.isArray(e.kernels)||!Array.isArray(e.buffers))throw s("no kernels or buffers");if(e.kernels.forEach((t,i)=>{if(!(typeof t?.symbol=="string"&&typeof t.entry=="string"&&Array.isArray(t.buffers)&&t.buffers.every(a=>a==="f32"||a==="f16")&&pe(t.scalars)))throw s(`kernel ${i}`)}),e.buffers.forEach((t,i)=>{if(!(typeof t?.word=="string"&&pe(t.bytes)&&(t.storage==="f32"||t.storage==="f16")&&typeof t.placed=="boolean"))throw s(`buffer ${i}`)}),!de(e.act,e.buffers.length))throw s("act is no buffer");if(!de(e.logits,e.buffers.length))throw s("logits is no buffer");if(!Array.isArray(r)||r.length!==e.kernels.length||!r.every(t=>typeof t=="string"))throw s(`${e.kernels.length} kernels but ${Array.isArray(r)?r.length:"no"} WGSL sources`);return e}function Ge(n,r){let s=[],e=u=>new c("backend",`the host's launch words are malformed: ${u}`),t=0,i=(u,a)=>{if(t+u>n.length)throw e(`launch ${s.length} ends inside its ${a}`);let l=n.subarray(t,t+u);return t+=u,l};for(;t<n.length;){let u=i(1,"kernel")[0],a=r.kernels[u];if(a===void 0)throw e(`launch ${s.length} names kernel ${u}; the program has ${r.kernels.length}`);let l=Array.from(i(i(1,"buffer count")[0],"buffers"));if(l.length!==a.buffers.length)throw e(`a launch of ${a.symbol} binds ${l.length} buffers; the kernel declares ${a.buffers.length}`);let f=l.find(p=>p>=r.buffers.length);if(f!==void 0)throw e(`a launch of ${a.symbol} binds buffer ${f}; the program has ${r.buffers.length}`);let[y,m]=i(2,"grid"),x=i(i(1,"uniform count")[0],"uniform").slice();if(x.length===0!=(a.scalars===0))throw e(`a launch of ${a.symbol} carries ${x.length} uniform words; the kernel declares ${a.scalars} scalars`);s.push({kernel:u,buffers:l,x:y,y:m,uniform:x})}return s}var D=n=>n instanceof Error?n.message:String(n),V=class{constructor(r,s,e,t,i,u,a,l){this.adapter=r;this.description=s;this.host=e;this.device=t;this.program=i;this.pipelines=u;this.buffers=a;this.staging=l;t.lost.then(f=>{let y=`the WebGPU device was lost: ${f.message||f.reason||"no message"}`;!this.#e&&this.#t===null&&(this.#t=y),this.#r?.(y)})}adapter;description;host;device;program;pipelines;buffers;staging;#t=null;#e=!1;#r=null;#n=new Set;get failure(){return this.#t}get storage(){return this.program.buffers.some(r=>r.storage==="f16")?"f16":"f32"}async run(r,s){if(this.#e)throw new c("request","the WebGPU runtime was destroyed (the engine was disposed)");if(this.failure!==null)throw this.#o(this.failure),N(this.failure);let e;try{e=Ge(r,this.program);let l=this.program.buffers[this.program.act];if(s.byteLength>l.bytes)throw new c("backend",`actIn is ${s.byteLength} bytes; the program's ${l.word} is ${l.bytes}`)}catch(l){throw this.#s(D(l),!1)}let t=this.device,i=[],u,a=new Promise((l,f)=>{this.#r=y=>f(new Error(y))});a.catch(()=>null);try{t.pushErrorScope("validation"),t.pushErrorScope("out-of-memory"),t.pushErrorScope("internal");let l=3,f=async()=>{let o=t.popErrorScope(),d=t.popErrorScope(),b=t.popErrorScope();l=0;let[h,g,W]=await Promise.all([o,d,b]);return g?`out of memory: ${g.message}`:W?`validation: ${W.message}`:h?`an internal error: ${h.message}`:null},y,m;try{s.byteLength>0&&t.queue.writeBuffer(this.buffers[this.program.act],0,s);let o=e.map(g=>{let W=g.buffers.map((A,G)=>({binding:G,resource:{buffer:this.buffers[A]}}));if(g.uniform.length>0){let A=t.createBuffer({size:g.uniform.byteLength,usage:B.UNIFORM|B.COPY_DST});i.push(A),this.#n.add(A),t.queue.writeBuffer(A,0,g.uniform),W.push({binding:g.buffers.length,resource:{buffer:A}})}return t.createBindGroup({layout:this.pipelines[g.kernel].getBindGroupLayout(0),entries:W})}),d=t.createCommandEncoder(),b=d.beginComputePass();e.forEach((g,W)=>{b.setPipeline(this.pipelines[g.kernel]),b.setBindGroup(0,o[W]),b.dispatchWorkgroups(g.x,g.y,1)}),b.end();let h=this.buffers[this.program.logits];d.copyBufferToBuffer(h,0,this.staging,0,h.size),t.queue.submit([d.finish()]),y=f(),m=this.staging.mapAsync($e),y.catch(()=>null),m.catch(()=>null)}finally{for(;l-- >0;)t.popErrorScope().catch(()=>null)}if(this.failure!==null)throw new Error(this.failure);let x=await Promise.race([y,a]);if(this.#e)throw new c("request","the engine was disposed during a WebGPU call");if(x!==null)throw new Error(`the device reported ${x}`);try{await Promise.race([m,a])}catch(o){throw this.#e?new c("request","the engine was disposed during a WebGPU call"):new Error(`mapping the logits readback failed: ${D(o)}`)}if(this.#e)throw new c("request","the engine was disposed during a WebGPU call");let p=this.program.buffers[this.program.logits].bytes;u=new Uint8Array(this.staging.getMappedRange(0,this.staging.size).slice(0,p)),this.staging.unmap()}catch(l){throw l instanceof c&&l.kind==="request"?l:this.#s(D(l))}finally{this.#r=null;for(let l of i)this.#n.delete(l),this.#e||l.destroy()}this.host.writeLogits(u)}#s(r,s=!0){return this.#e?new c("request","the engine was disposed during a WebGPU call"):(s&&(this.#t??=r),this.#o(r)??new c("backend",r))}#o(r){try{this.host.failLaunch(r)}catch(s){let e=E(s);if(e.kind==="backend")return e}return null}destroy(){if(!this.#e){this.#e=!0,this.#r?.("the engine was disposed during a WebGPU call");for(let r of[...this.buffers,this.staging,...this.#n])try{r.destroy()}catch{}this.#n.clear();try{this.device.destroy()}catch{}}}};function Se(n,r){let s=n.limits[r];return typeof s=="number"?s:void 0}function Ee(){return globalThis.navigator?.gpu}async function he(n,r,s=Ee()){let e=f=>new c("backend",f);if(!s)throw e("the worker has no navigator.gpu, though the page's probe found a WebGPU adapter");let t;try{t=await s.requestAdapter()}catch(f){throw e(`the worker's requestAdapter failed: ${D(f)}`)}if(!t)throw e("the worker's requestAdapter returned null, though the page's probe found an adapter");if(r&&!t.features.has("shader-f16"))throw e("the host's precheck chose f16 weight storage, but the worker's adapter has no shader-f16");let i={};for(let f of We){let y=Se(t,f),m=n.webgpu?.limits[f];if(m!==void 0&&(y===void 0||y<m))throw e(`the worker's adapter offers ${f} ${y??"none"}; the host prechecked against the probe's ${m}`);y!==void 0&&(i[f]=y)}let u;try{u=await t.requestDevice({requiredFeatures:r?["shader-f16"]:[],requiredLimits:i})}catch(f){throw e(`requestDevice refused${r?" shader-f16 with":""} the program's limits: ${D(f)}`)}let a=t.info??{};if(!t.info&&typeof t.requestAdapterInfo=="function")try{a=await t.requestAdapterInfo()}catch{a={}}let l=!1;return{device:u,f16:r,adapter:{vendor:a.vendor??"",architecture:a.architecture??""},description:a.description??"",destroy:()=>{if(!l){l=!0;try{u.destroy()}catch{}}}}}async function ge(n,r,s){let e=a=>new c("backend",a),t=r.device,i=[],u=()=>{for(let a of i)a.destroy();r.destroy()};try{let a=Ae(s.program,s.kernels),l=a.buffers.some(x=>x.storage==="f16");if(l!==r.f16)throw e(`the program stores weights ${l?"f16":"f32"}, but the device was opened for ${r.f16?"f16":"f32"} storage`);let f=[];for(let[x,p]of a.kernels.entries()){t.pushErrorScope("validation"),t.pushErrorScope("internal");let o=null,d=null;try{let g=t.createShaderModule({code:s.kernels[x],label:p.symbol});o=t.createComputePipeline({label:p.symbol,layout:"auto",compute:{module:g,entryPoint:p.entry}})}catch(g){d=g}let b=await t.popErrorScope(),h=await t.popErrorScope()??b;if(h)throw e(`kernel ${p.symbol}: creating its pipeline: ${h.message}`);if(o===null)throw e(`kernel ${p.symbol}: creating its pipeline: ${D(d)}`);f.push(o)}for(let[x,p]of a.buffers.entries()){t.pushErrorScope("validation"),t.pushErrorScope("out-of-memory");let o=null;try{let h=t.createBuffer({size:Math.max(4,Math.ceil(p.bytes/4)*4),usage:B.STORAGE|B.COPY_SRC|B.COPY_DST,label:p.word});if(i.push(h),p.placed){let g=n.bufferBytes(x);if(g.byteLength!==p.bytes)throw new c("backend",`buffer ${p.word}: bufferBytes gave ${g.byteLength} bytes; the buffer is ${p.bytes} at ${p.storage} storage`);if(g.byteLength%4!==0)throw new c("backend",`buffer ${p.word}: ${g.byteLength} bytes are not whole 4-byte words`);t.queue.writeBuffer(h,0,g)}}catch(h){o=h}let d=await t.popErrorScope(),b=await t.popErrorScope();if(o!==null)throw o instanceof c?o:E(o);if(d)throw e(`buffer ${p.word} (${p.bytes} bytes): out of memory: ${d.message}`);if(b)throw e(`buffer ${p.word} (${p.bytes} bytes): ${b.message}`)}let y=i[a.logits],m=t.createBuffer({size:y.size,usage:B.MAP_READ|B.COPY_DST,label:"logits readback"});return new V({...r.adapter},r.description,n,t,a,f,i,m)}catch(a){throw u(),a}}var Ce=4194304,Ue="a quantized alias runs on the wasm tier: the WebGPU tier has no quantized kernels in this version",Re=["fit","request","probe"];function be(){return new Promise(n=>{let r=new MessageChannel;r.port1.onmessage=()=>{r.port1.close(),n()},r.port2.postMessage(0)})}async function me(n,r,s){let e,t;try{e=s.kind==="prefill"?n.resolve("prefill",s.n):n.resolve("step",s.pos),t=n.actIn(s.kind==="prefill"?s.n:1)}catch(i){try{n.failLaunch(E(i).message)}catch{}throw i}await r.run(e,t)}async function Be(n,r,s,e){if(await me(n,r,s),!!e.launched())for(n.finishStart();;){if(!await e.between())return;let t=n.beginNext();if("delta"in t){e.delta(JSON.parse(t.delta));return}if(await me(n,r,t.launch),!e.launched())return;let i=JSON.parse(n.finishNext());if(e.delta(i),i.done)return}}function ye(n,r){let s=null,e=null,t=!1,i=null,u=!1,a=[],l=o=>{let d={...o.doctor(),aliases:a.map(A=>({...A}))};if(e===null)return{...d,storage:null,adapter:null};let b=e.adapter,h=`webgpu (${e.description!==""?e.description:`${b.vendor} ${b.architecture}`})`,g=`backend webgpu on ${d.device}`,W=d.line.startsWith(g)?`backend webgpu on ${h}${d.line.slice(g.length)}`:d.line;return{...d,device:h,line:W,storage:e.storage,adapter:{...b}}},f=new c("request","the engine was disposed during its load"),y=()=>{if(u)throw f};async function m(o){if(s!==null||t)throw new c("request","this engine already has a model loaded");t=!0;let d=null;try{d=await n.openHost(o.probeJson,o.hostWasm,o.cdnWasm),r({id:o.id,type:"host",hostWasm:n.wasmUsed()}),y();let b=await n.hostClass(),h=b.resolveAlias(o.alias);a=b.listAliases();let{repo:g,revision:W}=h,A={},G=async(k,v,P)=>{try{let $=await n.fetchFile(g,W,k,{hubBaseUrl:o.hubBaseUrl,offline:o.offline,rememberAbsent:v,...P??{},onProgress:U=>r({id:o.id,type:"progress",file:U.file,loaded:U.loaded,total:U.total})});return y(),A[k]={source:$.source,verified:$.verified},$.bytes}catch($){if(v&&$ instanceof c&&$.missing==="not-found")return null;throw $}},w=async(k,v)=>await G(k,!1,v),S=d,L,q;if(h.format==="gguf"){let k=$=>n.fetchRange(g,W,h.file,$,{hubBaseUrl:o.hubBaseUrl,offline:o.offline}),v=await k(Math.min(Ce,h.bytes));y();for(let $=b.ggufHeaderNeed(v,h.bytes);$!==0;$=b.ggufHeaderNeed(v,h.bytes))v=await k($),y();let P=v;L=$=>S.precheckGguf(P,h.bytes,$),q=async($,U)=>{let H=await w(h.file,{sha256:h.sha256,bytes:h.bytes});return U?S.loadWebGpuGguf(H,$):S.loadGguf(H,$)}}else{let k=await w("config.json"),v=await G("model.safetensors.index.json",!0);L=P=>S.precheck(k,v,P),q=async(P,$)=>{let U=await w("tokenizer.json"),H=await G("tokenizer_config.json",!0),M=[];for(let we of b.weightFiles(v))M.push(await w(we));return $?S.loadWebGpu(k,U,H,M,P):S.load(k,U,H,M,P)}}let C={...o.request,repo:g,revision:W},_,F=null,X=()=>{let k={...C,backend:"wasm"};delete k.f16,C=k},Z=k=>{X();try{L(JSON.stringify(C))}catch(v){let P=E(v);throw new c(P.kind,`${P.message} (backend "auto": WebGPU refused first: ${k.message})`)}_=`webgpu refused: ${k.message}; wasm serves this load`};if(o.auto===!0&&h.format==="gguf")X(),L(JSON.stringify(C)),_=Ue;else if(C.backend==="webgpu"){let k=null;try{k=JSON.parse(L(JSON.stringify(C)))}catch(v){let P=E(v);if(!(o.auto===!0&&Re.includes(P.kind)))throw P;Z(P)}if(k!==null)try{F=await n.openDevice(JSON.parse(o.probeJson),k.f16===!0)}catch(v){let P=E(v);if(o.auto!==!0)throw P;Z(P)}}else L(JSON.stringify(C));try{y();let k=JSON.stringify(C);if(F!==null){let v=await q(k,!0),P=F;F=null,e=await n.openRuntime(S,P,v)}else await n.instantiate(S,await q(k,!1))}finally{F?.destroy()}y(),s=d,r({id:o.id,type:"loaded",doctor:l(d),sources:A,..._!==void 0?{fallback:_}:{}})}catch(b){if(e?.destroy(),e=null,d!==null&&s!==d)try{d.dispose()}catch{}throw b}finally{t=!1}}async function x(o,d,b,h){let g;"messages"in o.prompt?g=d.startChatSplit(JSON.stringify(o.prompt.messages),h):"text"in o.prompt?g=d.startCompleteSplit(o.prompt.text,null,h):g=d.startCompleteSplit(null,Uint32Array.from(o.prompt.ids),h),await Be(d,b,g,{launched:()=>!u,between:async()=>(await be(),!u),delta:W=>r({id:o.id,type:"delta",delta:W})})}async function p(o){if(i!==null)throw new c("request","a generation is already running on this engine");if(s===null)throw new c("request","no model is loaded");let d=s,b=JSON.stringify(o.options);if(e!==null){if(e.failure!==null)throw N(e.failure);i=o.id;try{return await x(o,d,e,b)}catch(h){if(u)return;throw h}finally{i=null}}i=o.id;try{for(("messages"in o.prompt)?d.startChat(JSON.stringify(o.prompt.messages),b):("text"in o.prompt)?d.startComplete(o.prompt.text,null,b):d.startComplete(null,Uint32Array.from(o.prompt.ids),b);;){if(await be(),u)return;let h=JSON.parse(d.next());if(r({id:o.id,type:"delta",delta:h}),h.done)return}}finally{i=null}}return async o=>{try{switch(o.type){case"load":return await m(o);case"generate":return await p(o);case"abort":i===o.id&&s?.abort();return;case"doctor":if(s===null)throw new c("request","no model is loaded");r({id:o.id,type:"result",value:l(s)});return;case"dispose":u=!0,e?.destroy(),e=null,s?.dispose(),s=null,r({id:o.id,type:"result",value:null}),n.close();return}}catch(d){if(u)return;let b=E(d),h=e?.failure??null;r({id:o.id,type:"error",kind:b.kind,message:b.message,...h!==null?{fatal:h}:{}})}}}var O=globalThis,Le=ye({openHost:ue,wasmUsed:le,hostClass:ce,instantiate:fe,openDevice:(n,r)=>he(n,r),openRuntime:ge,fetchFile:oe,fetchRange:ie,close:()=>O.close()},n=>O.postMessage(n));O.onmessage=n=>{Le(n.data)};K().then(()=>O.postMessage({id:0,type:"ready",url:O.location.href}),n=>O.postMessage({id:0,type:"error",kind:"backend",message:`the browser host's glue (host/caldrix_web.js, beside dist/worker.js) did not load (${n instanceof Error?n.message:String(n)})`}));

package/host/caldrix_web.d.ts

@@ -0,0 +1,206 @@
+/* tslint:disable */
+/* eslint-disable */
+
+export class CaldrixWeb {
+    free(): void;
+    [Symbol.dispose](): void;
+    abort(): void;
+    /**
+     * A copy of the first `rows` activation rows the session gathered into `ActA` (f32), for the
+     * runtime to write into the program's `act` buffer before a launch.
+     */
+    actIn(rows: number): Uint8Array;
+    argsPtr(): number;
+    /**
+     * The module's entries: `step(args)` and `prefill(args, n)`, each returning the status code.
+     */
+    attach(step: Function, prefill: Function): void;
+    /**
+     * `{ launch }` (the step to run, then `finishNext`) or `{ delta }` (the last delta, the JSON
+     * string `next` returns; nothing to run).
+     */
+    beginNext(): any;
+    /**
+     * Program buffer `index`'s bytes for the runtime to upload once: the heap's, a block GEMM
+     * weight stored f16 narrowed RNE here (a value past f16 range a `validate` error naming the
+     * tensor).
+     */
+    bufferBytes(index: number): Uint8Array;
+    /**
+     * Releases the host's heap and everything it holds; every later call (but `dispose` and `abort`,
+     * no-ops, and `argsPtr`, then `0`) is a `request` refusal rather than a use of a freed
+     * object. The page's object itself is freed by `free()` or its finalizer.
+     */
+    dispose(): void;
+    /**
+     * What this host is: `{ backend, device, dtype, simd, threads, model, revision, source, line }`;
+     * `source` is the page's to fill.
+     */
+    doctor(): any;
+    /**
+     * The runtime's launch failed (a device loss, a validation or out-of-memory error): the
+     * generation ends as `next` does on a failed step -- every later split call is the
+     * "finished" refusal -- and `message` is thrown back as a `backend` error.
+     */
+    failLaunch(message: string): void;
+    /**
+     * After a `beginNext` launch ran and `writeLogits` wrote its row: the delta, as `next`'s JSON.
+     */
+    finishNext(): any;
+    /**
+     * After the first launch ran and `writeLogits` wrote its row: picks the first token.
+     */
+    finishStart(): void;
+    /**
+     * [`HEADER_FIRST_RANGE`], the page's first range of a GGUF file (the worker's
+     * `GGUF_HEADER_FIRST_RANGE` is pinned against it).
+     */
+    static ggufHeaderFirstRange(): number;
+    /**
+     * How much of a GGUF file's head the page reads next: `0` when `prefix` holds the whole
+     * header, else the next range's length (doubling, capped at `file_bytes`). The page's first
+     * range is `HEADER_FIRST_RANGE` (4 MiB) or the whole file when smaller.
+     */
+    static ggufHeaderNeed(prefix: Uint8Array, file_bytes: number): number;
+    /**
+     * Every alias the browser build knows: `[{ name, format }]`, `format` `"safetensors"` or
+     * `"gguf"` (what `resolveAlias` answers for that name).
+     */
+    static listAliases(): any;
+    /**
+     * [`CaldrixWeb::load`] over a GGUF file the page verified against the alias's sha256: the
+     * precheck again on the whole file, the import (the tokenizer from the header), the lowering
+     * with the file's per-layer kinds (a quantized file lowers to the wasm tier's quantized
+     * kernels), and the placement.
+     */
+    loadGguf(file: Uint8Array, request_json: string): Uint8Array;
+    /**
+     * [`CaldrixWeb::load_webgpu`] over a GGUF file the page verified (see
+     * [`CaldrixWeb::load_gguf`]); returns `{ program, kernels }`.
+     */
+    loadWebGpuGguf(file: Uint8Array, request_json: string): any;
+    /**
+     * The WebGPU tier's `load`: prechecks the request (`backend: "webgpu"`) against the adapter
+     * again, imports, lowers for WebGPU (plan dtype f32, the block GEMM weights stored f16 when the
+     * precheck chose it), places into the host's heap and resolves the program. Returns
+     * `{ program, kernels }`: `program` the `WebGpuProgram` JSON (`to_json`), `kernels` the WGSL
+     * sources in the same index order. Nothing is attached: the split-phase calls drive it.
+     */
+    loadWebGpu(config_json: Uint8Array, tokenizer_json: Uint8Array, tokenizer_config: Uint8Array | null | undefined, files: Array<any>, request_json: string): any;
+    /**
+     * Prechecks again on the real bytes, imports, lowers and places into the host's heap; returns
+     * the module bytes for the page to instantiate over [`memory`].
+     */
+    load(config_json: Uint8Array, tokenizer_json: Uint8Array, tokenizer_config: Uint8Array | null | undefined, files: Array<any>, request_json: string): Uint8Array;
+    /**
+     * Where the module writes each logits row: [`Placed::logits`], `0` when nothing is loaded. For
+     * a test page that reads the row from [`CaldrixWeb::memory`]; the product never reads it.
+     */
+    logitsPtr(): number;
+    /**
+     * The host's linear memory, which the page instantiates the module over. A static method, not
+     * a free function: a free function's wasm export takes its own name, and `memory` is already
+     * the name of the module's memory export, so the module would not compile.
+     */
+    static memory(): any;
+    constructor(probe_json: string);
+    /**
+     * The next token as JSON: `{ token, text, done, finishReason, promptTokens, cachedTokens, completionTokens }`, plus
+     * `logprobs: [{ token, logprob }, ...]` when the options asked for them and the row has a finite entry.
+     */
+    next(): any;
+    /**
+     * [`CaldrixWeb::precheck`] over a GGUF file's head: `header` is enough of it for the header
+     * to parse ([`CaldrixWeb::gguf_header_need`] answers 0), `file_bytes` the alias's pinned size.
+     */
+    precheckGguf(header: Uint8Array, file_bytes: number, request_json: string): any;
+    precheck(config_json: Uint8Array, index_json: Uint8Array | null | undefined, request_json: string): any;
+    static resolveAlias(name: string): any;
+    /**
+     * The launches of `kind` (`"prefill"` with `n` rows from the argument block's `Pos` word,
+     * `"step"` at position `n_or_pos`) as `launches_to_words`' flat list. A call the program
+     * refuses is the wasm entry's status as a `backend` error ("prefill returned status 2", "step
+     * returned status 1").
+     */
+    resolve(kind: string, n_or_pos: number): Uint32Array;
+    startChatSplit(messages_json: string, options_json: string): any;
+    startChat(messages_json: string, options_json: string): number;
+    startCompleteSplit(text: string | null | undefined, ids: Uint32Array | null | undefined, options_json: string): any;
+    startComplete(text: string | null | undefined, ids: Uint32Array | null | undefined, options_json: string): number;
+    /**
+     * The logits row's length in f32 ([`Placed::vocab`]), `0` when nothing is loaded.
+     */
+    vocab(): number;
+    static weightFiles(index_json?: Uint8Array | null): any;
+    /**
+     * The runtime's readback of the `logits` buffer, written into the host's `Logits` row, where
+     * `finishStart`/`finishNext` read it; a length other than `vocab` f32 is refused.
+     */
+    writeLogits(bytes: Uint8Array): void;
+}
+
+export type InitInput = RequestInfo | URL | Response | BufferSource | WebAssembly.Module;
+
+export interface InitOutput {
+    readonly memory: WebAssembly.Memory;
+    readonly __wbg_caldrixweb_free: (a: number, b: number) => void;
+    readonly caldrixweb_abort: (a: number) => void;
+    readonly caldrixweb_actIn: (a: number, b: number, c: number) => void;
+    readonly caldrixweb_argsPtr: (a: number) => number;
+    readonly caldrixweb_attach: (a: number, b: number, c: number, d: number) => void;
+    readonly caldrixweb_beginNext: (a: number, b: number) => void;
+    readonly caldrixweb_bufferBytes: (a: number, b: number, c: number) => void;
+    readonly caldrixweb_dispose: (a: number) => void;
+    readonly caldrixweb_doctor: (a: number, b: number) => void;
+    readonly caldrixweb_failLaunch: (a: number, b: number, c: number, d: number) => void;
+    readonly caldrixweb_finishNext: (a: number, b: number) => void;
+    readonly caldrixweb_finishStart: (a: number, b: number) => void;
+    readonly caldrixweb_ggufHeaderFirstRange: () => number;
+    readonly caldrixweb_ggufHeaderNeed: (a: number, b: number, c: number, d: number) => void;
+    readonly caldrixweb_listAliases: (a: number) => void;
+    readonly caldrixweb_load: (a: number, b: number, c: number, d: number, e: number, f: number, g: number, h: number, i: number, j: number, k: number) => void;
+    readonly caldrixweb_loadGguf: (a: number, b: number, c: number, d: number, e: number, f: number) => void;
+    readonly caldrixweb_loadWebGpu: (a: number, b: number, c: number, d: number, e: number, f: number, g: number, h: number, i: number, j: number, k: number) => void;
+    readonly caldrixweb_loadWebGpuGguf: (a: number, b: number, c: number, d: number, e: number, f: number) => void;
+    readonly caldrixweb_logitsPtr: (a: number, b: number) => void;
+    readonly caldrixweb_memory: () => number;
+    readonly caldrixweb_new: (a: number, b: number, c: number) => void;
+    readonly caldrixweb_next: (a: number, b: number) => void;
+    readonly caldrixweb_precheck: (a: number, b: number, c: number, d: number, e: number, f: number, g: number, h: number) => void;
+    readonly caldrixweb_precheckGguf: (a: number, b: number, c: number, d: number, e: number, f: number, g: number) => void;
+    readonly caldrixweb_resolve: (a: number, b: number, c: number, d: number, e: number) => void;
+    readonly caldrixweb_resolveAlias: (a: number, b: number, c: number) => void;
+    readonly caldrixweb_startChat: (a: number, b: number, c: number, d: number, e: number, f: number) => void;
+    readonly caldrixweb_startChatSplit: (a: number, b: number, c: number, d: number, e: number, f: number) => void;
+    readonly caldrixweb_startComplete: (a: number, b: number, c: number, d: number, e: number, f: number, g: number, h: number) => void;
+    readonly caldrixweb_startCompleteSplit: (a: number, b: number, c: number, d: number, e: number, f: number, g: number, h: number) => void;
+    readonly caldrixweb_vocab: (a: number, b: number) => void;
+    readonly caldrixweb_weightFiles: (a: number, b: number, c: number) => void;
+    readonly caldrixweb_writeLogits: (a: number, b: number, c: number, d: number) => void;
+    readonly __wbindgen_export: (a: number, b: number) => number;
+    readonly __wbindgen_export2: (a: number, b: number, c: number, d: number) => number;
+    readonly __wbindgen_export3: (a: number) => void;
+    readonly __wbindgen_add_to_stack_pointer: (a: number) => number;
+}
+
+export type SyncInitInput = BufferSource | WebAssembly.Module;
+
+/**
+ * Instantiates the given `module`, which can either be bytes or
+ * a precompiled `WebAssembly.Module`.
+ *
+ * @param {{ module: SyncInitInput }} module - Passing `SyncInitInput` directly is deprecated.
+ *
+ * @returns {InitOutput}
+ */
+export function initSync(module: { module: SyncInitInput } | SyncInitInput): InitOutput;
+
+/**
+ * If `module_or_path` is {RequestInfo} or {URL}, makes a request and
+ * for everything else, calls `WebAssembly.instantiate` directly.
+ *
+ * @param {{ module_or_path: InitInput | Promise<InitInput> }} module_or_path - Passing `InitInput` directly is deprecated.
+ *
+ * @returns {Promise<InitOutput>}
+ */
+export default function __wbg_init (module_or_path?: { module_or_path: InitInput | Promise<InitInput> } | InitInput | Promise<InitInput>): Promise<InitOutput>;

package/host/caldrix_web.js

@@ -0,0 +1 @@
+class G{__destroy_into_raw(){const e=this.__wbg_ptr;return this.__wbg_ptr=0,N.unregister(this),e}free(){const e=this.__destroy_into_raw();t.__wbg_caldrixweb_free(e,0)}abort(){t.caldrixweb_abort(this.__wbg_ptr)}actIn(e){try{const s=t.__wbindgen_add_to_stack_pointer(-16);t.caldrixweb_actIn(s,this.__wbg_ptr,e);var n=a().getInt32(s+0,!0),_=a().getInt32(s+4,!0),r=a().getInt32(s+8,!0);if(r)throw c(_);return c(n)}finally{t.__wbindgen_add_to_stack_pointer(16)}}argsPtr(){return t.caldrixweb_argsPtr(this.__wbg_ptr)>>>0}attach(e,n){try{const s=t.__wbindgen_add_to_stack_pointer(-16);t.caldrixweb_attach(s,this.__wbg_ptr,y(e),y(n));var _=a().getInt32(s+0,!0),r=a().getInt32(s+4,!0);if(r)throw c(_)}finally{t.__wbindgen_add_to_stack_pointer(16)}}beginNext(){try{const r=t.__wbindgen_add_to_stack_pointer(-16);t.caldrixweb_beginNext(r,this.__wbg_ptr);var e=a().getInt32(r+0,!0),n=a().getInt32(r+4,!0),_=a().getInt32(r+8,!0);if(_)throw c(n);return c(e)}finally{t.__wbindgen_add_to_stack_pointer(16)}}bufferBytes(e){try{const s=t.__wbindgen_add_to_stack_pointer(-16);t.caldrixweb_bufferBytes(s,this.__wbg_ptr,e);var n=a().getInt32(s+0,!0),_=a().getInt32(s+4,!0),r=a().getInt32(s+8,!0);if(r)throw c(_);return c(n)}finally{t.__wbindgen_add_to_stack_pointer(16)}}dispose(){t.caldrixweb_dispose(this.__wbg_ptr)}doctor(){try{const r=t.__wbindgen_add_to_stack_pointer(-16);t.caldrixweb_doctor(r,this.__wbg_ptr);var e=a().getInt32(r+0,!0),n=a().getInt32(r+4,!0),_=a().getInt32(r+8,!0);if(_)throw c(n);return c(e)}finally{t.__wbindgen_add_to_stack_pointer(16)}}failLaunch(e){try{const r=t.__wbindgen_add_to_stack_pointer(-16),s=f(e,t.__wbindgen_export,t.__wbindgen_export2),o=g;t.caldrixweb_failLaunch(r,this.__wbg_ptr,s,o);var n=a().getInt32(r+0,!0),_=a().getInt32(r+4,!0);if(_)throw c(n)}finally{t.__wbindgen_add_to_stack_pointer(16)}}finishNext(){try{const r=t.__wbindgen_add_to_stack_pointer(-16);t.caldrixweb_finishNext(r,this.__wbg_ptr);var e=a().getInt32(r+0,!0),n=a().getInt32(r+4,!0),_=a().getInt32(r+8,!0);if(_)throw c(n);return c(e)}finally{t.__wbindgen_add_to_stack_pointer(16)}}finishStart(){try{const _=t.__wbindgen_add_to_stack_pointer(-16);t.caldrixweb_finishStart(_,this.__wbg_ptr);var e=a().getInt32(_+0,!0),n=a().getInt32(_+4,!0);if(n)throw c(e)}finally{t.__wbindgen_add_to_stack_pointer(16)}}static ggufHeaderFirstRange(){return t.caldrixweb_ggufHeaderFirstRange()}static ggufHeaderNeed(e,n){try{const o=t.__wbindgen_add_to_stack_pointer(-16),d=h(e,t.__wbindgen_export),b=g;t.caldrixweb_ggufHeaderNeed(o,d,b,n);var _=a().getFloat64(o+0,!0),r=a().getInt32(o+8,!0),s=a().getInt32(o+12,!0);if(s)throw c(r);return _}finally{t.__wbindgen_add_to_stack_pointer(16)}}static listAliases(){try{const r=t.__wbindgen_add_to_stack_pointer(-16);t.caldrixweb_listAliases(r);var e=a().getInt32(r+0,!0),n=a().getInt32(r+4,!0),_=a().getInt32(r+8,!0);if(_)throw c(n);return c(e)}finally{t.__wbindgen_add_to_stack_pointer(16)}}loadGguf(e,n){try{const o=t.__wbindgen_add_to_stack_pointer(-16),d=h(e,t.__wbindgen_export),b=g,u=f(n,t.__wbindgen_export,t.__wbindgen_export2),l=g;t.caldrixweb_loadGguf(o,this.__wbg_ptr,d,b,u,l);var _=a().getInt32(o+0,!0),r=a().getInt32(o+4,!0),s=a().getInt32(o+8,!0);if(s)throw c(r);return c(_)}finally{t.__wbindgen_add_to_stack_pointer(16)}}loadWebGpuGguf(e,n){try{const o=t.__wbindgen_add_to_stack_pointer(-16),d=h(e,t.__wbindgen_export),b=g,u=f(n,t.__wbindgen_export,t.__wbindgen_export2),l=g;t.caldrixweb_loadWebGpuGguf(o,this.__wbg_ptr,d,b,u,l);var _=a().getInt32(o+0,!0),r=a().getInt32(o+4,!0),s=a().getInt32(o+8,!0);if(s)throw c(r);return c(_)}finally{t.__wbindgen_add_to_stack_pointer(16)}}loadWebGpu(e,n,_,r,s){try{const w=t.__wbindgen_add_to_stack_pointer(-16),x=h(e,t.__wbindgen_export),I=g,M=h(n,t.__wbindgen_export),T=g;var o=v(_)?0:h(_,t.__wbindgen_export),d=g;const C=f(s,t.__wbindgen_export,t.__wbindgen_export2),j=g;t.caldrixweb_loadWebGpu(w,this.__wbg_ptr,x,I,M,T,o,d,y(r),C,j);var b=a().getInt32(w+0,!0),u=a().getInt32(w+4,!0),l=a().getInt32(w+8,!0);if(l)throw c(u);return c(b)}finally{t.__wbindgen_add_to_stack_pointer(16)}}load(e,n,_,r,s){try{const w=t.__wbindgen_add_to_stack_pointer(-16),x=h(e,t.__wbindgen_export),I=g,M=h(n,t.__wbindgen_export),T=g;var o=v(_)?0:h(_,t.__wbindgen_export),d=g;const C=f(s,t.__wbindgen_export,t.__wbindgen_export2),j=g;t.caldrixweb_load(w,this.__wbg_ptr,x,I,M,T,o,d,y(r),C,j);var b=a().getInt32(w+0,!0),u=a().getInt32(w+4,!0),l=a().getInt32(w+8,!0);if(l)throw c(u);return c(b)}finally{t.__wbindgen_add_to_stack_pointer(16)}}logitsPtr(){try{const r=t.__wbindgen_add_to_stack_pointer(-16);t.caldrixweb_logitsPtr(r,this.__wbg_ptr);var e=a().getInt32(r+0,!0),n=a().getInt32(r+4,!0),_=a().getInt32(r+8,!0);if(_)throw c(n);return e>>>0}finally{t.__wbindgen_add_to_stack_pointer(16)}}static memory(){const e=t.caldrixweb_memory();return c(e)}constructor(e){try{const s=t.__wbindgen_add_to_stack_pointer(-16),o=f(e,t.__wbindgen_export,t.__wbindgen_export2),d=g;t.caldrixweb_new(s,o,d);var n=a().getInt32(s+0,!0),_=a().getInt32(s+4,!0),r=a().getInt32(s+8,!0);if(r)throw c(_);return this.__wbg_ptr=n,N.register(this,this.__wbg_ptr,this),this}finally{t.__wbindgen_add_to_stack_pointer(16)}}next(){try{const r=t.__wbindgen_add_to_stack_pointer(-16);t.caldrixweb_next(r,this.__wbg_ptr);var e=a().getInt32(r+0,!0),n=a().getInt32(r+4,!0),_=a().getInt32(r+8,!0);if(_)throw c(n);return c(e)}finally{t.__wbindgen_add_to_stack_pointer(16)}}precheckGguf(e,n,_){try{const d=t.__wbindgen_add_to_stack_pointer(-16),b=h(e,t.__wbindgen_export),u=g,l=f(_,t.__wbindgen_export,t.__wbindgen_export2),w=g;t.caldrixweb_precheckGguf(d,this.__wbg_ptr,b,u,n,l,w);var r=a().getInt32(d+0,!0),s=a().getInt32(d+4,!0),o=a().getInt32(d+8,!0);if(o)throw c(s);return c(r)}finally{t.__wbindgen_add_to_stack_pointer(16)}}precheck(e,n,_){try{const u=t.__wbindgen_add_to_stack_pointer(-16),l=h(e,t.__wbindgen_export),w=g;var r=v(n)?0:h(n,t.__wbindgen_export),s=g;const x=f(_,t.__wbindgen_export,t.__wbindgen_export2),I=g;t.caldrixweb_precheck(u,this.__wbg_ptr,l,w,r,s,x,I);var o=a().getInt32(u+0,!0),d=a().getInt32(u+4,!0),b=a().getInt32(u+8,!0);if(b)throw c(d);return c(o)}finally{t.__wbindgen_add_to_stack_pointer(16)}}static resolveAlias(e){try{const s=t.__wbindgen_add_to_stack_pointer(-16),o=f(e,t.__wbindgen_export,t.__wbindgen_export2),d=g;t.caldrixweb_resolveAlias(s,o,d);var n=a().getInt32(s+0,!0),_=a().getInt32(s+4,!0),r=a().getInt32(s+8,!0);if(r)throw c(_);return c(n)}finally{t.__wbindgen_add_to_stack_pointer(16)}}resolve(e,n){try{const o=t.__wbindgen_add_to_stack_pointer(-16),d=f(e,t.__wbindgen_export,t.__wbindgen_export2),b=g;t.caldrixweb_resolve(o,this.__wbg_ptr,d,b,n);var _=a().getInt32(o+0,!0),r=a().getInt32(o+4,!0),s=a().getInt32(o+8,!0);if(s)throw c(r);return c(_)}finally{t.__wbindgen_add_to_stack_pointer(16)}}startChatSplit(e,n){try{const o=t.__wbindgen_add_to_stack_pointer(-16),d=f(e,t.__wbindgen_export,t.__wbindgen_export2),b=g,u=f(n,t.__wbindgen_export,t.__wbindgen_export2),l=g;t.caldrixweb_startChatSplit(o,this.__wbg_ptr,d,b,u,l);var _=a().getInt32(o+0,!0),r=a().getInt32(o+4,!0),s=a().getInt32(o+8,!0);if(s)throw c(r);return c(_)}finally{t.__wbindgen_add_to_stack_pointer(16)}}startChat(e,n){try{const o=t.__wbindgen_add_to_stack_pointer(-16),d=f(e,t.__wbindgen_export,t.__wbindgen_export2),b=g,u=f(n,t.__wbindgen_export,t.__wbindgen_export2),l=g;t.caldrixweb_startChat(o,this.__wbg_ptr,d,b,u,l);var _=a().getInt32(o+0,!0),r=a().getInt32(o+4,!0),s=a().getInt32(o+8,!0);if(s)throw c(r);return _>>>0}finally{t.__wbindgen_add_to_stack_pointer(16)}}startCompleteSplit(e,n,_){try{const w=t.__wbindgen_add_to_stack_pointer(-16);var r=v(e)?0:f(e,t.__wbindgen_export,t.__wbindgen_export2),s=g,o=v(n)?0:V(n,t.__wbindgen_export),d=g;const x=f(_,t.__wbindgen_export,t.__wbindgen_export2),I=g;t.caldrixweb_startCompleteSplit(w,this.__wbg_ptr,r,s,o,d,x,I);var b=a().getInt32(w+0,!0),u=a().getInt32(w+4,!0),l=a().getInt32(w+8,!0);if(l)throw c(u);return c(b)}finally{t.__wbindgen_add_to_stack_pointer(16)}}startComplete(e,n,_){try{const w=t.__wbindgen_add_to_stack_pointer(-16);var r=v(e)?0:f(e,t.__wbindgen_export,t.__wbindgen_export2),s=g,o=v(n)?0:V(n,t.__wbindgen_export),d=g;const x=f(_,t.__wbindgen_export,t.__wbindgen_export2),I=g;t.caldrixweb_startComplete(w,this.__wbg_ptr,r,s,o,d,x,I);var b=a().getInt32(w+0,!0),u=a().getInt32(w+4,!0),l=a().getInt32(w+8,!0);if(l)throw c(u);return b>>>0}finally{t.__wbindgen_add_to_stack_pointer(16)}}vocab(){try{const r=t.__wbindgen_add_to_stack_pointer(-16);t.caldrixweb_vocab(r,this.__wbg_ptr);var e=a().getInt32(r+0,!0),n=a().getInt32(r+4,!0),_=a().getInt32(r+8,!0);if(_)throw c(n);return e>>>0}finally{t.__wbindgen_add_to_stack_pointer(16)}}static weightFiles(e){try{const d=t.__wbindgen_add_to_stack_pointer(-16);var n=v(e)?0:h(e,t.__wbindgen_export),_=g;t.caldrixweb_weightFiles(d,n,_);var r=a().getInt32(d+0,!0),s=a().getInt32(d+4,!0),o=a().getInt32(d+8,!0);if(o)throw c(s);return c(r)}finally{t.__wbindgen_add_to_stack_pointer(16)}}writeLogits(e){try{const r=t.__wbindgen_add_to_stack_pointer(-16),s=h(e,t.__wbindgen_export),o=g;t.caldrixweb_writeLogits(r,this.__wbg_ptr,s,o);var n=a().getInt32(r+0,!0),_=a().getInt32(r+4,!0);if(_)throw c(n)}finally{t.__wbindgen_add_to_stack_pointer(16)}}}Symbol.dispose&&(G.prototype[Symbol.dispose]=G.prototype.free);function D(){return{__proto__:null,"./caldrix_web_bg.js":{__proto__:null,__wbg___wbindgen_memory_caa4a6165639c8b5:function(){const e=t.memory;return y(e)},__wbg___wbindgen_number_get_2e0e7dee9f701a71:function(e,n){const _=p(n),r=typeof _=="number"?_:void 0;a().setFloat64(e+8,v(r)?0:r,!0),a().setInt32(e+0,!v(r),!0)},__wbg___wbindgen_string_get_0380ccaa2f57f0d9:function(e,n){const _=p(n),r=typeof _=="string"?_:void 0;var s=v(r)?0:f(r,t.__wbindgen_export,t.__wbindgen_export2),o=g;a().setInt32(e+4,o,!0),a().setInt32(e+0,s,!0)},__wbg___wbindgen_throw_41e9ee4f547fc59a:function(e,n){throw new Error(B(e,n))},__wbg_call_1875a20c43a36133:function(){return U(function(e,n,_,r){const s=p(e).call(p(n),p(_),p(r));return y(s)},arguments)},__wbg_call_187d372bd5fdd4aa:function(){return U(function(e,n,_){const r=p(e).call(p(n),p(_));return y(r)},arguments)},__wbg_getRandomValues_a608c4436c19407a:function(){return U(function(e,n){globalThis.crypto.getRandomValues(E(e,n))},arguments)},__wbg_get_unchecked_288889d017702237:function(e,n){const _=p(e)[n>>>0];return y(_)},__wbg_instanceof_Error_80a725f81f2e102d:function(e){let n;try{n=p(e)instanceof Error}catch{n=!1}return n},__wbg_instanceof_Uint8Array_828cef2aaacafc31:function(e){let n;try{n=p(e)instanceof Uint8Array}catch{n=!1}return n},__wbg_length_7f3c00c40364105e:function(e){return p(e).length},__wbg_length_d4bdea10311bd9cf:function(e){return p(e).length},__wbg_message_5f8387f0c32b90a7:function(e){const n=p(e).message;return y(n)},__wbg_new_617a8cdb8bb1130e:function(){const e=new Object;return y(e)},__wbg_new_ee2291f50781bf1d:function(){const e=new Array;return y(e)},__wbg_new_from_slice_37a3c3a194266cb8:function(e,n){const _=new Uint32Array($(e,n));return y(_)},__wbg_new_from_slice_9a868026ffa4208a:function(e,n){const _=new Uint8Array(E(e,n));return y(_)},__wbg_prototypesetcall_bc27214492979395:function(e,n,_){Uint8Array.prototype.set.call(E(e,n),p(_))},__wbg_push_2baf45db356cf468:function(e,n){return p(e).push(p(n))},__wbg_set_145a351398b48c65:function(){return U(function(e,n,_){return Reflect.set(p(e),p(n),p(_))},arguments)},__wbindgen_generic_0000000000000001:function(e){return y(e)},__wbindgen_generic_0000000000000002:function(e,n){const _=B(e,n);return y(_)},__wbindgen_object_drop_ref:function(e){c(e)}}}}const N=typeof FinalizationRegistry>"u"?{register:()=>{},unregister:()=>{}}:new FinalizationRegistry(i=>t.__wbg_caldrixweb_free(i,1));function y(i){S===k.length&&k.push(k.length+1);const e=S;return S=k[e],k[e]=i,e}function z(i){i<1028||(k[i]=S,S=i)}function $(i,e){return i=i>>>0,P().subarray(i/4,i/4+e)}function E(i,e){return i=i>>>0,A().subarray(i/1,i/1+e)}let m=null;function a(){return(m===null||m.buffer.detached===!0||m.buffer.detached===void 0&&m.buffer!==t.memory.buffer)&&(m=new DataView(t.memory.buffer)),m}function B(i,e){return X(i>>>0,e)}let W=null;function P(){return(W===null||W.byteLength===0)&&(W=new Uint32Array(t.memory.buffer)),W}let R=null;function A(){return(R===null||R.byteLength===0)&&(R=new Uint8Array(t.memory.buffer)),R}function p(i){return k[i]}function U(i,e){try{return i.apply(this,e)}catch(n){t.__wbindgen_export3(y(n))}}let k=new Array(1024).fill(void 0);k.push(void 0,null,!0,!1);let S=k.length;function v(i){return i==null}function V(i,e){const n=e(i.length*4,4)>>>0;return P().set(i,n/4),g=i.length,n}function h(i,e){const n=e(i.length*1,1)>>>0;return A().set(i,n/1),g=i.length,n}function f(i,e,n){if(n===void 0){const d=O.encode(i),b=e(d.length,1)>>>0;return A().subarray(b,b+d.length).set(d),g=d.length,b}let _=i.length,r=e(_,1)>>>0;const s=A();let o=0;for(;o<_;o++){const d=i.charCodeAt(o);if(d>127)break;s[r+o]=d}if(o!==_){o!==0&&(i=i.slice(o)),r=n(r,_,_=o+i.length*3,1)>>>0;const d=A().subarray(r+o,r+_),b=O.encodeInto(i,d);o+=b.written,r=n(r,_,o,1)>>>0}return g=o,r}function c(i){const e=p(i);return z(i),e}let F=new TextDecoder("utf-8",{ignoreBOM:!0,fatal:!0});F.decode();const q=2146435072;let L=0;function X(i,e){return L+=e,L>=q&&(F=new TextDecoder("utf-8",{ignoreBOM:!0,fatal:!0}),F.decode(),L=e),F.decode(A().subarray(i,i+e))}const O=new TextEncoder;"encodeInto"in O||(O.encodeInto=function(i,e){const n=O.encode(i);return e.set(n),{read:i.length,written:n.length}});let g=0,Y,J,t;function H(i,e){return J=i,t=i.exports,Y=e,m=null,W=null,R=null,t}async function K(i,e){if(typeof Response=="function"&&i instanceof Response){if(!i.ok)throw new Error(`failed to fetch Wasm: ${i.status} ${i.statusText} fetching '${i.url}'`);if(typeof WebAssembly.instantiateStreaming=="function")try{return await WebAssembly.instantiateStreaming(i,e)}catch(r){if(n(i.type)&&i.headers.get("Content-Type")!=="application/wasm")console.warn("`WebAssembly.instantiateStreaming` failed because your server does not serve Wasm with `application/wasm` MIME type. Falling back to `WebAssembly.instantiate` which is slower. Original error:\n",r);else throw r}const _=await i.arrayBuffer();return await WebAssembly.instantiate(_,e)}else{const _=await WebAssembly.instantiate(i,e);return _ instanceof WebAssembly.Instance?{instance:_,module:i}:_}function n(_){switch(_){case"basic":case"cors":case"default":return!0}return!1}}function Q(i){if(t!==void 0)return t;i!==void 0&&(Object.getPrototypeOf(i)===Object.prototype?{module:i}=i:console.warn("using deprecated parameters for `initSync()`; pass a single object instead"));const e=D();i instanceof WebAssembly.Module||(i=new WebAssembly.Module(i));const n=new WebAssembly.Instance(i,e);return H(n,i)}async function Z(i){if(t!==void 0)return t;i!==void 0&&(Object.getPrototypeOf(i)===Object.prototype?{module_or_path:i}=i:console.warn("using deprecated parameters for the initialization function; pass a single object instead")),i===void 0&&(i=new URL("caldrix_web_bg.wasm",import.meta.url));const e=D();(typeof i=="string"||typeof Request=="function"&&i instanceof Request||typeof URL=="function"&&i instanceof URL)&&(i=fetch(i));const{instance:n,module:_}=await K(await i,e);return H(n,_)}export{G as CaldrixWeb,Z as default,Q as initSync};

package/host/caldrix_web_bg.wasm.d.ts

@@ -0,0 +1,41 @@
+/* tslint:disable */
+/* eslint-disable */
+export const memory: WebAssembly.Memory;
+export const __wbg_caldrixweb_free: (a: number, b: number) => void;
+export const caldrixweb_abort: (a: number) => void;
+export const caldrixweb_actIn: (a: number, b: number, c: number) => void;
+export const caldrixweb_argsPtr: (a: number) => number;
+export const caldrixweb_attach: (a: number, b: number, c: number, d: number) => void;
+export const caldrixweb_beginNext: (a: number, b: number) => void;
+export const caldrixweb_bufferBytes: (a: number, b: number, c: number) => void;
+export const caldrixweb_dispose: (a: number) => void;
+export const caldrixweb_doctor: (a: number, b: number) => void;
+export const caldrixweb_failLaunch: (a: number, b: number, c: number, d: number) => void;
+export const caldrixweb_finishNext: (a: number, b: number) => void;
+export const caldrixweb_finishStart: (a: number, b: number) => void;
+export const caldrixweb_ggufHeaderFirstRange: () => number;
+export const caldrixweb_ggufHeaderNeed: (a: number, b: number, c: number, d: number) => void;
+export const caldrixweb_listAliases: (a: number) => void;
+export const caldrixweb_load: (a: number, b: number, c: number, d: number, e: number, f: number, g: number, h: number, i: number, j: number, k: number) => void;
+export const caldrixweb_loadGguf: (a: number, b: number, c: number, d: number, e: number, f: number) => void;
+export const caldrixweb_loadWebGpu: (a: number, b: number, c: number, d: number, e: number, f: number, g: number, h: number, i: number, j: number, k: number) => void;
+export const caldrixweb_loadWebGpuGguf: (a: number, b: number, c: number, d: number, e: number, f: number) => void;
+export const caldrixweb_logitsPtr: (a: number, b: number) => void;
+export const caldrixweb_memory: () => number;
+export const caldrixweb_new: (a: number, b: number, c: number) => void;
+export const caldrixweb_next: (a: number, b: number) => void;
+export const caldrixweb_precheck: (a: number, b: number, c: number, d: number, e: number, f: number, g: number, h: number) => void;
+export const caldrixweb_precheckGguf: (a: number, b: number, c: number, d: number, e: number, f: number, g: number) => void;
+export const caldrixweb_resolve: (a: number, b: number, c: number, d: number, e: number) => void;
+export const caldrixweb_resolveAlias: (a: number, b: number, c: number) => void;
+export const caldrixweb_startChat: (a: number, b: number, c: number, d: number, e: number, f: number) => void;
+export const caldrixweb_startChatSplit: (a: number, b: number, c: number, d: number, e: number, f: number) => void;
+export const caldrixweb_startComplete: (a: number, b: number, c: number, d: number, e: number, f: number, g: number, h: number) => void;
+export const caldrixweb_startCompleteSplit: (a: number, b: number, c: number, d: number, e: number, f: number, g: number, h: number) => void;
+export const caldrixweb_vocab: (a: number, b: number) => void;
+export const caldrixweb_weightFiles: (a: number, b: number, c: number) => void;
+export const caldrixweb_writeLogits: (a: number, b: number, c: number, d: number) => void;
+export const __wbindgen_export: (a: number, b: number) => number;
+export const __wbindgen_export2: (a: number, b: number, c: number, d: number) => number;
+export const __wbindgen_export3: (a: number) => void;
+export const __wbindgen_add_to_stack_pointer: (a: number) => number;