tempest-react-sdk 0.68.0 → 0.70.0

package/dist/vision/core/session.cjs

@@ -1,2 +1,2 @@
-const e=require("../../_virtual/_rolldown/runtime.cjs"),t=require("./exceptions.cjs"),n=require("./graph.cjs"),r=require("./metadata.cjs"),i=require("./providers.cjs");let a=require("onnxruntime-web");a=e.__toESM(a,1);async function o(e){try{let t=await fetch(e);return t.ok?new Uint8Array(await t.arrayBuffer()):(s(e,`HTTP ${t.status} ${t.statusText}`),e)}catch(t){return s(e,t.message),e}}function s(e,t){console.warn(`[@ort-vision-sdk/web] Could not fetch ${e} to read its metadata (${t}). Letting ONNX Runtime load it instead: the model will work, but its baked-in class names are unavailable, so labels fall back to class_0, class_1, ... Pass \`labels\` explicitly to name them.`)}function c(e,t){let n=e.filter(e=>!t.includes(e));n.length!==0&&console.warn(`This browser cannot offer the requested execution provider(s) ${JSON.stringify(n)}; the session will run on ${JSON.stringify(t)}. Inference still produces correct results, on the fallback provider.`)}var l=class e{_session;providers;_metadata;requestedProviders;constructor(e,t,n,r){this._session=e,this.providers=t,this._metadata=n,this.requestedProviders=r}static async create(n,s={}){let l=i.resolveProviders(s.providers),u=await i.detectProviders(l),d=u.length>0?u:[i.FALLBACK_PROVIDER];s.providers!==void 0&&s.providers.length>0&&c(l,d);let f={...s.sessionOptions??{},executionProviders:d},p=s.readMetadata!==!1,m=typeof n==`string`&&p?await o(n):n,h=p&&typeof m!=`string`?r.readModelMetadata(m):{},g;try{g=(typeof m==`string`||m instanceof Uint8Array,await a.InferenceSession.create(m,f))}catch(e){throw new t.ModelLoadError(`Failed to load ONNX model: ${e.message}`,{cause:e})}return new e(g,d,h,l)}get inputNames(){return this._session.inputNames}get inputName(){let e=this._session.inputNames[0];if(e===void 0)throw new t.InferenceError(`Model has no inputs.`);return e}get outputNames(){return this._session.outputNames}get inputShapes(){return n.declaredShapesFrom(this._session.inputMetadata)}get inputShape(){return this.inputShapes[0]??[]}get outputShapes(){return n.declaredShapesFrom(this._session.outputMetadata)}get outputShape(){return this.outputShapes[0]??[]}get metadata(){return this._metadata}async release(){await this._session.release().catch(()=>void 0)}get raw(){return this._session}async run(e){try{return await this._session.run(e)}catch(e){throw new t.InferenceError(`Inference failed: ${e.message}`,{cause:e})}}};exports.OrtSession=l;
+const e=require("../../_virtual/_rolldown/runtime.cjs"),t=require("./exceptions.cjs"),n=require("./dtypes.cjs"),r=require("./graph.cjs"),i=require("./metadata.cjs"),a=require("./providers.cjs");let o=require("onnxruntime-web");o=e.__toESM(o,1);var s=`ort_vision_sdk.graph_optimization`,c=`CAcSDm9ydC12aXNpb24tc2RrOjoKEAoBeBIBeSIISWRlbnRpdHkSBHdhcm1aDwoBeBIKCggIARIECgIIAWIPCgF5EgoKCAgBEgQKAggBQgQKABAN`,l=new Map;function u(e){let t=e.map(a.providerName).join(`,`),n=l.get(t);return n===void 0&&(n=(async()=>{try{let t=Uint8Array.from(atob(c),e=>e.charCodeAt(0));await(await o.InferenceSession.create(t,{executionProviders:e})).release()}catch{return}})(),l.set(t,n)),n}function d(){console.warn(`[@ort-vision-sdk/web] This model was pre-optimized by optimize_model, which targets the WASM backend; on WebGPU some of its fused nodes fall back to the CPU. For WebGPU, ship the original export instead.`)}var f=`ort_vision_sdk.quantization`;function p(e,t){if(t[`ort_vision_sdk.quantization`]===void 0)return[...e];let n=e.filter(e=>a.providerName(e)!==`webgpu`);return n.length===e.length?n:(console.warn(`[@ort-vision-sdk/web] This model was quantized by quantize_model (${t[f]}); ONNX Runtime Web's WebGPU backend cannot run its quantized operators, so it runs on WASM.`),n.length>0?n:[a.FALLBACK_PROVIDER])}var m=`ort-vision-sdk-models`;async function h(e,t){let n=t===null?null:await _(t);if(n!==null){let t=await n.match(e).catch(()=>void 0);if(t!==void 0)return new Uint8Array(await t.arrayBuffer())}try{let t=await fetch(e);if(!t.ok)return v(e,`HTTP ${t.status} ${t.statusText}`),e;let r=await t.arrayBuffer();return n!==null&&await g(n,e,r),new Uint8Array(r)}catch(t){return v(e,t.message),e}}async function g(e,t,n){try{await e.put(t,new Response(n))}catch{return}}async function _(e){if(typeof caches>`u`)return null;try{return await caches.open(e)}catch{return null}}function v(e,t){console.warn(`[@ort-vision-sdk/web] Could not fetch ${e} to read its metadata (${t}). Letting ONNX Runtime load it instead: the model will work, but its baked-in class names are unavailable, so labels fall back to class_0, class_1, ... Pass \`labels\` explicitly to name them.`)}function y(e,t){let n=e.filter(e=>!t.includes(e));n.length!==0&&console.warn(`This browser cannot offer the requested execution provider(s) ${JSON.stringify(n)}; the session will run on ${JSON.stringify(t)}. Inference still produces correct results, on the fallback provider.`)}function b(e){let t=i.readModelInputTypes(e),r={};for(let[e,i]of Object.entries(t))r[e]=n.tensorTypeFor(i);return r}function x(){console.warn(`Could not read any input type from this model's bytes; feeds will be built as float32. A half-precision model will fail at inference with 'Unexpected input data type'. Please report this with the model that produced it.`)}function S(e){let r=Object.entries(e).filter(([,e])=>e===`float16`);if(r.length===0||n.hasFloat16Array())return;let i=r.map(([e])=>e).join(`, `);throw new t.ModelLoadError(`This model declares half-precision input(s) [${i}], but this environment has no Float16Array, which ONNX Runtime requires for a float16 tensor. Use a float32 export of the model, or run in a browser that supports Float16Array.`)}var C=class e{_session;providers;_metadata;requestedProviders;_inputTypes;constructor(e,t,n,r,i={}){this._session=e,this.providers=t,this._metadata=n,this.requestedProviders=r,this._inputTypes=i}_runQueue=Promise.resolve();static async create(n,r={}){let s=a.resolveProviders(r.providers),c=r.readMetadata!==!1,l=r.cache===!0?m:r.cache?r.cache:null,f=typeof n==`string`&&(c||l!==null),g=a.detectProviders(s).then(e=>e.length>0?e:[a.FALLBACK_PROVIDER]),[_,v]=await Promise.all([g,f?h(n,l):Promise.resolve(n),f?g.then(u):Promise.resolve()]),C=c&&typeof v!=`string`?i.readModelMetadata(v):{},w=p(_,C),T=s.map(a.providerName),E=w.map(a.providerName);r.providers!==void 0&&r.providers.length>0&&y(T,_.map(a.providerName)),C[`ort_vision_sdk.graph_optimization`]!==void 0&&E.includes(`webgpu`)&&d();let D={...C[`ort_vision_sdk.graph_optimization`]!==void 0&&r.sessionOptions?.graphOptimizationLevel===void 0?{graphOptimizationLevel:`disabled`}:{},...r.sessionOptions??{},executionProviders:w},O=typeof v==`string`?{}:b(v);typeof v!=`string`&&Object.keys(O).length===0&&x(),S(O);let k;try{k=(typeof v==`string`||v instanceof Uint8Array,await o.InferenceSession.create(v,D))}catch(e){throw new t.ModelLoadError(`Failed to load ONNX model: ${e.message}`,{cause:e})}return new e(k,E,C,T,O)}get inputDtypes(){return this.inputNames.map(e=>this._inputTypes[e]??`float32`)}get inputDtype(){return this._inputTypes[this.inputName]??`float32`}get inputNames(){return this._session.inputNames}get inputName(){let e=this._session.inputNames[0];if(e===void 0)throw new t.InferenceError(`Model has no inputs.`);return e}get outputNames(){return this._session.outputNames}get inputShapes(){return r.declaredShapesFrom(this._session.inputMetadata)}get inputShape(){return this.inputShapes[0]??[]}get outputShapes(){return r.declaredShapesFrom(this._session.outputMetadata)}get outputShape(){return this.outputShapes[0]??[]}get metadata(){return this._metadata}async release(){await this._session.release().catch(()=>void 0)}get raw(){return this._session}run(e){let t=this._runQueue.then(()=>this._runNow(e));return this._runQueue=t.catch(()=>void 0),t}async _runNow(e){try{return await this._session.run(this._asDeclaredTypes(e))}catch(e){throw new t.InferenceError(`Inference failed: ${e.message}`,{cause:e})}}_asDeclaredTypes(e){let t=null;for(let[r,i]of Object.entries(e)){let a=this._inputTypes[r];if(a===void 0||a===i.type||!(i.data instanceof Float32Array))continue;let s=n.toFeedData(i.data,a);s!==i.data&&(t??={...e},t[r]=new o.Tensor(a,s,i.dims))}return t??e}};exports.DEFAULT_MODEL_CACHE=m,exports.GRAPH_OPTIMIZATION_KEY=s,exports.OrtSession=C,exports.QUANTIZATION_KEY=f;
 //# sourceMappingURL=session.cjs.map

package/dist/vision/core/session.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"session.cjs","names":[],"sources":["../../../src/vision/core/session.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * Thin wrapper around `onnxruntime-web` `InferenceSession` with typed metadata.\n */\n\nimport type * as ort from \"onnxruntime-web\";\nimport * as ortRuntime from \"onnxruntime-web\";\n\nimport { InferenceError, ModelLoadError } from \"./exceptions\";\nimport { type DeclaredShape, declaredShapesFrom } from \"./graph\";\nimport { readModelMetadata } from \"./metadata\";\nimport { FALLBACK_PROVIDER, detectProviders, resolveProviders } from \"./providers\";\n\n/** Anything `InferenceSession.create` accepts. */\nexport type ModelSource = string | ArrayBufferLike | Uint8Array;\n\n/**\n * Fetch a model URL as bytes so its metadata can be read.\n *\n * Falls back to the URL itself when the fetch fails, letting ORT try its own\n * load path: losing the metadata map is a downgrade, but failing to load a model\n * that ORT could have fetched would be a regression.\n *\n * @param url Where the `.onnx` lives.\n * @returns The model bytes, or the original URL when they could not be fetched.\n */\nasync function fetchModel(url: string): Promise<Uint8Array | string> {\n    try {\n        const response = await fetch(url);\n        if (!response.ok) {\n            warnMetadataUnavailable(url, `HTTP ${response.status} ${response.statusText}`);\n            return url;\n        }\n        return new Uint8Array(await response.arrayBuffer());\n    } catch (err) {\n        warnMetadataUnavailable(url, (err as Error).message);\n        return url;\n    }\n}\n\n/**\n * Warn that a model's metadata could not be read, and say what that costs.\n *\n * The fallback itself is right — losing the metadata beats failing a load that\n * ORT could have completed on its own — but it used to be silent, and the\n * symptom it produces is remote from the cause: class names come back as\n * `class_0`, `class_1`, ... with nothing anywhere explaining why. Whoever hits\n * this needs to be told that passing `labels` is the way out.\n *\n * @param url The model URL that could not be fetched here.\n * @param reason What went wrong, as reported by `fetch`.\n */\nfunction warnMetadataUnavailable(url: string, reason: string): void {\n    console.warn(\n        `[@ort-vision-sdk/web] Could not fetch ${url} to read its metadata (${reason}). ` +\n            \"Letting ONNX Runtime load it instead: the model will work, but its baked-in \" +\n            \"class names are unavailable, so labels fall back to class_0, class_1, ... \" +\n            \"Pass `labels` explicitly to name them.\",\n    );\n}\n\nexport interface OrtSessionOptions {\n    /**\n     * Execution providers in preference order. `undefined` uses {@link DEFAULT_PROVIDERS}.\n     *\n     * Naming one explicitly also opts into a `console.warn` when this browser\n     * cannot offer it, instead of falling back in silence.\n     */\n    readonly providers?: readonly string[];\n    /** Optional ORT session options forwarded to `InferenceSession.create`. */\n    readonly sessionOptions?: ort.InferenceSession.SessionOptions;\n    /**\n     * Whether to read the model's custom metadata map (`names`, `task`, `imgsz`).\n     * Defaults to `true`.\n     *\n     * The runtime does not expose that map, so it is read from the file itself —\n     * which means a URL model is fetched here and handed to ORT as bytes instead\n     * of letting ORT fetch it. That is the same single download either way, and\n     * it is what lets a task resolve its labels off the model. Set to `false` to\n     * keep the URL path untouched and leave {@link OrtSession.metadata} empty.\n     *\n     * `false` is also the escape hatch when a device cannot afford the bytes: the\n     * fetched buffer is dropped before ORT builds the graph (see\n     * {@link OrtSession.create}), but ORT's own load path still keeps the model out\n     * of reach of anything the SDK holds. A session built this way resolves its\n     * input size from the graph as usual — only the class names are lost, so a\n     * caller taking this route has to pass `labels` itself.\n     */\n    readonly readMetadata?: boolean;\n}\n\n/**\n * Warn when a provider named explicitly by the caller is not going to run.\n *\n * Only explicit requests are worth a warning. The default list exists precisely\n * so that falling from `webgpu` to `wasm` is the expected outcome — but a caller\n * who wrote `providers: [\"webgpu\"]` and lands on WASM has a page several times\n * slower than intended and nothing in the console to explain it.\n *\n * @param requested Providers that were asked for.\n * @param effective Providers that survived capability detection.\n */\nfunction warnOnDroppedProviders(requested: readonly string[], effective: readonly string[]): void {\n    const dropped = requested.filter((provider) => !effective.includes(provider));\n    if (dropped.length === 0) {\n        return;\n    }\n    console.warn(\n        `This browser cannot offer the requested execution provider(s) ${JSON.stringify(dropped)}; ` +\n            `the session will run on ${JSON.stringify(effective)}. Inference still produces correct ` +\n            \"results, on the fallback provider.\",\n    );\n}\n\n/**\n * Wrap an ONNX Runtime Web `InferenceSession` with convenient metadata access.\n *\n * The wrapper exposes input/output names and the shapes the graph declares,\n * manages execution-provider selection, provides a typed {@link OrtSession.run}\n * method, and releases the native session through {@link OrtSession.release}.\n */\nexport class OrtSession {\n    private constructor(\n        private readonly _session: ort.InferenceSession,\n        /**\n         * Execution providers this session is expected to run on.\n         *\n         * The requested list narrowed to what this browser can actually offer — a\n         * `webgpu` entry survives only where an adapter exists. Best-effort: ORT-Web\n         * exposes no way to ask which provider a session ended up on, so an entry\n         * here means \"not ruled out\", not \"confirmed\". See\n         * {@link requestedProviders} for what was asked for.\n         *\n         * When nothing survives — a caller asking for `webgpu` alone on a device\n         * without it — this falls back to {@link FALLBACK_PROVIDER}, which ORT-Web\n         * can always run. Handing ORT the unsatisfiable list instead makes\n         * `InferenceSession.create` reject with \"no available backend found\", so the\n         * page gets no inference at all rather than the slow-but-working fallback\n         * the `console.warn` describes. Measured in a real Chromium, where\n         * `navigator.gpu` exists but yields no adapter.\n         */\n        public readonly providers: readonly string[],\n        private readonly _metadata: Readonly<Record<string, string>>,\n        /**\n         * Execution providers that were asked for, after defaults were applied.\n         *\n         * Kept separate because ORT-Web falls back silently: a page that asks for\n         * `webgpu` on a device without it runs on WASM and is told nothing.\n         */\n        public readonly requestedProviders: readonly string[],\n    ) {}\n\n    /**\n     * Load an ONNX model into an ORT inference session.\n     *\n     * The metadata map is read **before** the session is built, and that order is\n     * load-bearing on memory-constrained devices. ORT copies the model into its\n     * WASM heap and then allocates the graph and the weights on top of that copy;\n     * a `readModelMetadata` call placed after `InferenceSession.create` keeps the\n     * JavaScript-side buffer reachable across the whole build, so a 5 MB model\n     * costs 5 MB of JS heap plus 5 MB of WASM heap plus the weights at the same\n     * instant. Reading first makes the buffer collectable as soon as ORT has copied\n     * it — on a phone that was the difference between a session and\n     * `Can't create a session. failed to allocate a buffer of size N`.\n     *\n     * @param model Either a URL string, or a `Uint8Array`/`ArrayBuffer` containing the model bytes.\n     * @param options Provider list, pass-through `SessionOptions`, and whether to\n     *   read the model's metadata map (see {@link OrtSessionOptions.readMetadata}).\n     * @throws {@link ModelLoadError} if the model cannot be loaded.\n     */\n    static async create(model: ModelSource, options: OrtSessionOptions = {}): Promise<OrtSession> {\n        const requested = resolveProviders(options.providers);\n        const detected = await detectProviders(requested);\n        const providers = detected.length > 0 ? detected : [FALLBACK_PROVIDER];\n        if (options.providers !== undefined && options.providers.length > 0) {\n            warnOnDroppedProviders(requested, providers);\n        }\n        const sessionOptions: ort.InferenceSession.SessionOptions = {\n            ...(options.sessionOptions ?? {}),\n            executionProviders:\n                providers as ort.InferenceSession.SessionOptions[\"executionProviders\"],\n        };\n        const wantsMetadata = options.readMetadata !== false;\n        const source = typeof model === \"string\" && wantsMetadata ? await fetchModel(model) : model;\n        const metadata =\n            wantsMetadata && typeof source !== \"string\" ? readModelMetadata(source) : {};\n\n        let session: ort.InferenceSession;\n        try {\n            if (typeof source === \"string\") {\n                session = await ortRuntime.InferenceSession.create(source, sessionOptions);\n            } else if (source instanceof Uint8Array) {\n                session = await ortRuntime.InferenceSession.create(source, sessionOptions);\n            } else {\n                session = await ortRuntime.InferenceSession.create(\n                    source as ArrayBuffer,\n                    sessionOptions,\n                );\n            }\n        } catch (err) {\n            throw new ModelLoadError(`Failed to load ONNX model: ${(err as Error).message}`, {\n                cause: err,\n            });\n        }\n\n        return new OrtSession(session, providers, metadata, requested);\n    }\n\n    /** Names of the model's inputs, in declaration order. */\n    get inputNames(): readonly string[] {\n        return this._session.inputNames;\n    }\n\n    /** Name of the first (and usually only) input. */\n    get inputName(): string {\n        const name = this._session.inputNames[0];\n        if (name === undefined) {\n            throw new InferenceError(\"Model has no inputs.\");\n        }\n        return name;\n    }\n\n    /** Names of the model's outputs, in declaration order. */\n    get outputNames(): readonly string[] {\n        return this._session.outputNames;\n    }\n\n    /**\n     * Shapes the graph declares for its inputs, in declaration order.\n     *\n     * Dynamic (symbolic) axes appear as `null`. Empty shapes mean the runtime\n     * reported no metadata — either a non-tensor input, or an `onnxruntime-web`\n     * older than 1.21, which predates input metadata.\n     */\n    get inputShapes(): readonly DeclaredShape[] {\n        return declaredShapesFrom(\n            this._session.inputMetadata as\n                readonly ort.InferenceSession.ValueMetadata[] | undefined,\n        );\n    }\n\n    /**\n     * Shape the graph declares for its first input, dynamic axes as `null`.\n     *\n     * Empty when the runtime reports no metadata for it.\n     */\n    get inputShape(): DeclaredShape {\n        return this.inputShapes[0] ?? [];\n    }\n\n    /**\n     * Shapes the graph declares for its outputs, in declaration order.\n     *\n     * Dynamic (symbolic) axes appear as `null`. Reading them is how a task can\n     * tell how many classes a head emits without being told.\n     */\n    get outputShapes(): readonly DeclaredShape[] {\n        return declaredShapesFrom(\n            this._session.outputMetadata as\n                readonly ort.InferenceSession.ValueMetadata[] | undefined,\n        );\n    }\n\n    /**\n     * Shape the graph declares for its first output, dynamic axes as `null`.\n     *\n     * Empty when the runtime reports no metadata for it.\n     */\n    get outputShape(): DeclaredShape {\n        return this.outputShapes[0] ?? [];\n    }\n\n    /**\n     * The model's custom metadata map — `names`, `task`, `imgsz`, ... for an\n     * Ultralytics export.\n     *\n     * Read from the model's bytes at load time, since the runtime does not expose\n     * it. Empty when the session was created with `readMetadata: false`, from a\n     * URL that could not be fetched here, or from a model carrying no metadata.\n     */\n    get metadata(): Readonly<Record<string, string>> {\n        return this._metadata;\n    }\n\n    /**\n     * Release the native session and free its memory.\n     *\n     * Call it when a session is discarded while the page lives on — rebuilding a\n     * task at a different input size, swapping in a newer model. A failure from\n     * the runtime is ignored: a session being torn down has nothing left to fail\n     * at, and the caller is already moving on.\n     */\n    async release(): Promise<void> {\n        await this._session.release().catch(() => undefined);\n    }\n\n    /** The underlying `onnxruntime-web` session, for advanced use cases. */\n    get raw(): ort.InferenceSession {\n        return this._session;\n    }\n\n    /**\n     * Run inference and return all outputs.\n     *\n     * @param feeds Map of input name to `ort.Tensor`. Keys must match {@link inputNames}.\n     * @throws {@link InferenceError} if ORT raises any error during execution.\n     */\n    async run(feeds: Record<string, ort.Tensor>): Promise<Record<string, ort.Tensor>> {\n        try {\n            const result = await this._session.run(feeds);\n            return result as Record<string, ort.Tensor>;\n        } catch (err) {\n            throw new InferenceError(`Inference failed: ${(err as Error).message}`, { cause: err });\n        }\n    }\n}\n"],"mappings":"0NA0BA,eAAe,EAAW,EAA2C,CACjE,GAAI,CACA,IAAM,EAAW,MAAM,MAAM,CAAG,EAKhC,OAJK,EAAS,GAIP,IAAI,WAAW,MAAM,EAAS,YAAY,CAAC,GAH9C,EAAwB,EAAK,QAAQ,EAAS,OAAO,GAAG,EAAS,YAAY,EACtE,EAGf,OAAS,EAAK,CAEV,OADA,EAAwB,EAAM,EAAc,OAAO,EAC5C,CACX,CACJ,CAcA,SAAS,EAAwB,EAAa,EAAsB,CAChE,QAAQ,KACJ,yCAAyC,EAAI,yBAAyB,EAAO,kMAIjF,CACJ,CA2CA,SAAS,EAAuB,EAA8B,EAAoC,CAC9F,IAAM,EAAU,EAAU,OAAQ,GAAa,CAAC,EAAU,SAAS,CAAQ,CAAC,EACxE,EAAQ,SAAW,GAGvB,QAAQ,KACJ,iEAAiE,KAAK,UAAU,CAAO,EAAE,4BAC1D,KAAK,UAAU,CAAS,EAAE,sEAE7D,CACJ,CASA,IAAa,EAAb,MAAa,CAAW,CAEC,SAkBD,UACC,UAOD,mBA3BpB,YACI,EAkBA,EACA,EAOA,EACF,CA3BmB,KAAA,SAAA,EAkBD,KAAA,UAAA,EACC,KAAA,UAAA,EAOD,KAAA,mBAAA,CACjB,CAoBH,aAAa,OAAO,EAAoB,EAA6B,CAAC,EAAwB,CAC1F,IAAM,EAAY,EAAA,iBAAiB,EAAQ,SAAS,EAC9C,EAAW,MAAM,EAAA,gBAAgB,CAAS,EAC1C,EAAY,EAAS,OAAS,EAAI,EAAW,CAAC,EAAA,iBAAiB,EACjE,EAAQ,YAAc,IAAA,IAAa,EAAQ,UAAU,OAAS,GAC9D,EAAuB,EAAW,CAAS,EAE/C,IAAM,EAAsD,CACxD,GAAI,EAAQ,gBAAkB,CAAC,EAC/B,mBACI,CACR,EACM,EAAgB,EAAQ,eAAiB,GACzC,EAAS,OAAO,GAAU,UAAY,EAAgB,MAAM,EAAW,CAAK,EAAI,EAChF,EACF,GAAiB,OAAO,GAAW,SAAW,EAAA,kBAAkB,CAAM,EAAI,CAAC,EAE3E,EACJ,GAAI,CACA,AAKI,GALA,OAAO,GAAW,UAEX,aAAkB,WADf,MAAM,EAAW,iBAAiB,OAAO,EAAQ,CAAc,EASjF,OAAS,EAAK,CACV,MAAM,IAAI,EAAA,eAAe,8BAA+B,EAAc,UAAW,CAC7E,MAAO,CACX,CAAC,CACL,CAEA,OAAO,IAAI,EAAW,EAAS,EAAW,EAAU,CAAS,CACjE,CAGA,IAAI,YAAgC,CAChC,OAAO,KAAK,SAAS,UACzB,CAGA,IAAI,WAAoB,CACpB,IAAM,EAAO,KAAK,SAAS,WAAW,GACtC,GAAI,IAAS,IAAA,GACT,MAAM,IAAI,EAAA,eAAe,sBAAsB,EAEnD,OAAO,CACX,CAGA,IAAI,aAAiC,CACjC,OAAO,KAAK,SAAS,WACzB,CASA,IAAI,aAAwC,CACxC,OAAO,EAAA,mBACH,KAAK,SAAS,aAElB,CACJ,CAOA,IAAI,YAA4B,CAC5B,OAAO,KAAK,YAAY,IAAM,CAAC,CACnC,CAQA,IAAI,cAAyC,CACzC,OAAO,EAAA,mBACH,KAAK,SAAS,cAElB,CACJ,CAOA,IAAI,aAA6B,CAC7B,OAAO,KAAK,aAAa,IAAM,CAAC,CACpC,CAUA,IAAI,UAA6C,CAC7C,OAAO,KAAK,SAChB,CAUA,MAAM,SAAyB,CAC3B,MAAM,KAAK,SAAS,QAAQ,CAAC,CAAC,UAAY,IAAA,EAAS,CACvD,CAGA,IAAI,KAA4B,CAC5B,OAAO,KAAK,QAChB,CAQA,MAAM,IAAI,EAAwE,CAC9E,GAAI,CAEA,OAAO,MADc,KAAK,SAAS,IAAI,CAAK,CAEhD,OAAS,EAAK,CACV,MAAM,IAAI,EAAA,eAAe,qBAAsB,EAAc,UAAW,CAAE,MAAO,CAAI,CAAC,CAC1F,CACJ,CACJ"}
+{"version":3,"file":"session.cjs","names":[],"sources":["../../../src/vision/core/session.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * Thin wrapper around `onnxruntime-web` `InferenceSession` with typed metadata.\n */\n\nimport type * as ort from \"onnxruntime-web\";\nimport * as ortRuntime from \"onnxruntime-web\";\n\nimport { DEFAULT_TENSOR_TYPE, hasFloat16Array, tensorTypeFor, toFeedData } from \"./dtypes\";\nimport { InferenceError, ModelLoadError } from \"./exceptions\";\nimport { type DeclaredShape, declaredShapesFrom } from \"./graph\";\nimport { readModelInputTypes, readModelMetadata } from \"./metadata\";\nimport {\n    FALLBACK_PROVIDER,\n    type ProviderSpec,\n    detectProviders,\n    providerName,\n    resolveProviders,\n} from \"./providers\";\n\n/** Anything `InferenceSession.create` accepts. */\nexport type ModelSource = string | ArrayBufferLike | Uint8Array;\n\n/**\n * Metadata key naming the level a model was graph-optimized at offline.\n *\n * Written by the Python SDK's `optimize_model`; both test suites pin the same\n * string, so a rename on one side fails. See {@link OrtSession.create} for what\n * the session does when it finds it.\n */\nexport const GRAPH_OPTIMIZATION_KEY = \"ort_vision_sdk.graph_optimization\";\n\n/**\n * An 84-byte ONNX model — one `Identity` on a float `[1]` — used to start the runtime.\n *\n * Generated with `onnx.helper` (opset 13, IR 7, producer `ort-vision-sdk`).\n * See {@link warmRuntime} for why it exists.\n */\nconst RUNTIME_PROBE_MODEL =\n    \"CAcSDm9ydC12aXNpb24tc2RrOjoKEAoBeBIBeSIISWRlbnRpdHkSBHdhcm1aDwoBeBIKCggIARIECgIIAWIPCgF5EgoKCAgBEgQKAggBQgQKABAN\";\n\n/** Runtime start-ups already begun on this page, keyed by provider names. */\nconst runtimeWarmups = new Map<string, Promise<void>>();\n\n/**\n * Start ONNX Runtime's backends for a provider list, once per page.\n *\n * ORT downloads and compiles its WebAssembly binary — 12.8 MB for the plain\n * WASM build, 25.9 MB for the one carrying WebGPU — inside the first\n * `InferenceSession.create`. {@link OrtSession.create} only reaches that call\n * after it has downloaded the model, so on a real network the two largest\n * downloads of a page load ran one after the other. Creating and releasing a\n * session on a tiny model while the real one downloads lets them overlap; the\n * real `create` then finds the runtime ready.\n *\n * A failure here is swallowed: the real `create` runs against the same runtime\n * and reports the same problem with the right model in the message.\n *\n * @param specs Providers the real session will use.\n * @returns Settles when the runtime is up, or when starting it failed.\n */\nfunction warmRuntime(specs: readonly ProviderSpec[]): Promise<void> {\n    const key = specs.map(providerName).join(\",\");\n    let pending = runtimeWarmups.get(key);\n    if (pending === undefined) {\n        pending = (async () => {\n            try {\n                const probe = Uint8Array.from(atob(RUNTIME_PROBE_MODEL), (c) => c.charCodeAt(0));\n                const session = await ortRuntime.InferenceSession.create(probe, {\n                    executionProviders:\n                        specs as ort.InferenceSession.SessionOptions[\"executionProviders\"],\n                });\n                await session.release();\n            } catch {\n                return;\n            }\n        })();\n        runtimeWarmups.set(key, pending);\n    }\n    return pending;\n}\n\n/**\n * Warn that a model optimized offline is about to run on WebGPU.\n *\n * `optimize_model` fuses operators for the CPU, which is what the WASM backend\n * runs. On WebGPU some of those fused nodes have no GPU kernel: on a\n * YOLO11n-seg, ONNX Runtime placed 2 nodes on the CPU for the optimized file\n * and none for the original, so every inference pays GPU↔CPU copies. The SDK\n * cannot swap the file, so it says which one to ship.\n */\nfunction warnPreOptimizedOnWebGpu(): void {\n    console.warn(\n        \"[@ort-vision-sdk/web] This model was pre-optimized by optimize_model, which targets the \" +\n            \"WASM backend; on WebGPU some of its fused nodes fall back to the CPU. For WebGPU, ship \" +\n            \"the original export instead.\",\n    );\n}\n\n/**\n * Metadata key the Python SDK's `quantize_model` writes on an INT8 model.\n *\n * Mirrored by `QUANTIZATION_KEY` in the Python SDK; both test suites pin the\n * string. See {@link keepQuantizedOffWebGpu} for what a session does with it.\n */\nexport const QUANTIZATION_KEY = \"ort_vision_sdk.quantization\";\n\n/**\n * Drop `webgpu` from the providers of a model `quantize_model` produced.\n *\n * ONNX Runtime Web's WebGPU `DequantizeLinear` rejects the quantized bias such\n * a model carries — `scale and zero-point inputs must have the same rank` —\n * so the first `predict()` fails; and a variant with the bias left in float\n * ran but returned different detections than WASM for the same file. The\n * model is fine on WASM, which is also where INT8 pays (52.7 ms vs 68.4 ms on\n * a YOLO11n-seg), so the session runs there and says so.\n *\n * @param specs Providers that survived detection.\n * @param metadata The model's metadata map.\n * @returns `specs` without `webgpu` when the model is marked quantized; the\n *   WASM fallback when nothing else is left.\n */\nfunction keepQuantizedOffWebGpu(\n    specs: readonly ProviderSpec[],\n    metadata: Readonly<Record<string, string>>,\n): ProviderSpec[] {\n    if (metadata[QUANTIZATION_KEY] === undefined) return [...specs];\n    const kept = specs.filter((spec) => providerName(spec) !== \"webgpu\");\n    if (kept.length === specs.length) return kept;\n    console.warn(\n        `[@ort-vision-sdk/web] This model was quantized by quantize_model (${metadata[QUANTIZATION_KEY]}); ` +\n            \"ONNX Runtime Web's WebGPU backend cannot run its quantized operators, so it runs on WASM.\",\n    );\n    return kept.length > 0 ? kept : [FALLBACK_PROVIDER];\n}\n\n/** Cache Storage bucket used when {@link OrtSessionOptions.cache} is `true`. */\nexport const DEFAULT_MODEL_CACHE = \"ort-vision-sdk-models\";\n\n/**\n * Fetch a model URL as bytes so its metadata can be read.\n *\n * Falls back to the URL itself when the fetch fails, letting ORT try its own\n * load path: losing the metadata map is a downgrade, but failing to load a model\n * that ORT could have fetched would be a regression.\n *\n * @param url Where the `.onnx` lives.\n * @param cacheName Cache Storage bucket to read from and fill, or `null` to\n *   always go to the network.\n * @returns The model bytes, or the original URL when they could not be fetched.\n */\nasync function fetchModel(url: string, cacheName: string | null): Promise<Uint8Array | string> {\n    const cache = cacheName === null ? null : await openCache(cacheName);\n    if (cache !== null) {\n        const hit = await cache.match(url).catch(() => undefined);\n        if (hit !== undefined) return new Uint8Array(await hit.arrayBuffer());\n    }\n    try {\n        const response = await fetch(url);\n        if (!response.ok) {\n            warnMetadataUnavailable(url, `HTTP ${response.status} ${response.statusText}`);\n            return url;\n        }\n        const buffer = await response.arrayBuffer();\n        if (cache !== null) await storeModel(cache, url, buffer);\n        return new Uint8Array(buffer);\n    } catch (err) {\n        warnMetadataUnavailable(url, (err as Error).message);\n        return url;\n    }\n}\n\n/**\n * Put a model's bytes into the cache, never failing the load over it.\n *\n * Stores a fresh `Response` built from bytes already read rather than a\n * `clone()` of the network response: one read, and nothing that can throw\n * synchronously outside the guard. Awaited on purpose — a `put` left running\n * would keep its copy of the model reachable while ORT builds the session, the\n * same doubled peak {@link OrtSession.create} orders its reads to avoid.\n *\n * @param cache The open bucket.\n * @param url Cache key.\n * @param buffer The model.\n */\nasync function storeModel(cache: Cache, url: string, buffer: ArrayBuffer): Promise<void> {\n    try {\n        await cache.put(url, new Response(buffer));\n    } catch {\n        return;\n    }\n}\n\n/**\n * Open a Cache Storage bucket, or report that there is none to open.\n *\n * Cache Storage only exists in secure contexts (HTTPS, `localhost`), and a\n * browser may refuse it outright — private windows, blocked site data, an\n * exhausted quota. None of that is worth failing a model load over: the model\n * is fetched from the network as if caching had not been asked for. A failed\n * `put` is swallowed for the same reason — see {@link storeModel}.\n *\n * @param name Bucket name.\n * @returns The cache, or `null` when this environment cannot provide one.\n */\nasync function openCache(name: string): Promise<Cache | null> {\n    if (typeof caches === \"undefined\") return null;\n    try {\n        return await caches.open(name);\n    } catch {\n        return null;\n    }\n}\n\n/**\n * Warn that a model's metadata could not be read, and say what that costs.\n *\n * The fallback itself is right — losing the metadata beats failing a load that\n * ORT could have completed on its own — but it used to be silent, and the\n * symptom it produces is remote from the cause: class names come back as\n * `class_0`, `class_1`, ... with nothing anywhere explaining why. Whoever hits\n * this needs to be told that passing `labels` is the way out.\n *\n * @param url The model URL that could not be fetched here.\n * @param reason What went wrong, as reported by `fetch`.\n */\nfunction warnMetadataUnavailable(url: string, reason: string): void {\n    console.warn(\n        `[@ort-vision-sdk/web] Could not fetch ${url} to read its metadata (${reason}). ` +\n            \"Letting ONNX Runtime load it instead: the model will work, but its baked-in \" +\n            \"class names are unavailable, so labels fall back to class_0, class_1, ... \" +\n            \"Pass `labels` explicitly to name them.\",\n    );\n}\n\nexport interface OrtSessionOptions {\n    /**\n     * Execution providers in preference order. `undefined` uses {@link DEFAULT_PROVIDERS}.\n     *\n     * Each entry is a name (`\"webgpu\"`, `\"wasm\"`) or ORT's config object for that\n     * provider, which passes provider options through untouched —\n     * `{ name: \"webgpu\", preferredLayout: \"NHWC\" }`.\n     *\n     * Naming one explicitly also opts into a `console.warn` when this browser\n     * cannot offer it, instead of falling back in silence.\n     */\n    readonly providers?: readonly ProviderSpec[];\n    /**\n     * Keep a URL model in the browser's Cache Storage between page loads.\n     *\n     * `true` uses the {@link DEFAULT_MODEL_CACHE} bucket; a string names the\n     * bucket. A later `create` with the same URL reads the bytes from the cache\n     * instead of the network — for a model of several megabytes, the bulk of\n     * what a returning visitor waits for. Defaults to `false`.\n     *\n     * The URL is the cache key and nothing expires it: publish a changed model\n     * under a new URL (`yolo.v2.onnx`, `?v=2`) or a new bucket name, or delete\n     * the bucket with `caches.delete(name)`. Ignored for models passed as bytes,\n     * and silently skipped where Cache Storage is unavailable (non-secure\n     * contexts, blocked site data).\n     */\n    readonly cache?: boolean | string;\n    /** Optional ORT session options forwarded to `InferenceSession.create`. */\n    readonly sessionOptions?: ort.InferenceSession.SessionOptions;\n    /**\n     * Whether to read the model's custom metadata map (`names`, `task`, `imgsz`).\n     * Defaults to `true`.\n     *\n     * The runtime does not expose that map, so it is read from the file itself —\n     * which means a URL model is fetched here and handed to ORT as bytes instead\n     * of letting ORT fetch it. That is the same single download either way, and\n     * it is what lets a task resolve its labels off the model. Set to `false` to\n     * keep the URL path untouched (unless {@link cache} is set, which needs the\n     * bytes) and leave {@link OrtSession.metadata} empty.\n     *\n     * `false` is also the escape hatch when a device cannot afford the bytes: the\n     * fetched buffer is dropped before ORT builds the graph (see\n     * {@link OrtSession.create}), but ORT's own load path still keeps the model out\n     * of reach of anything the SDK holds. A session built this way resolves its\n     * input size from the graph as usual — only the class names are lost, so a\n     * caller taking this route has to pass `labels` itself.\n     */\n    readonly readMetadata?: boolean;\n}\n\n/**\n * Warn when a provider named explicitly by the caller is not going to run.\n *\n * Only explicit requests are worth a warning. The default list exists precisely\n * so that falling from `webgpu` to `wasm` is the expected outcome — but a caller\n * who wrote `providers: [\"webgpu\"]` and lands on WASM has a page several times\n * slower than intended and nothing in the console to explain it.\n *\n * @param requested Providers that were asked for.\n * @param effective Providers that survived capability detection.\n */\nfunction warnOnDroppedProviders(requested: readonly string[], effective: readonly string[]): void {\n    const dropped = requested.filter((provider) => !effective.includes(provider));\n    if (dropped.length === 0) {\n        return;\n    }\n    console.warn(\n        `This browser cannot offer the requested execution provider(s) ${JSON.stringify(dropped)}; ` +\n            `the session will run on ${JSON.stringify(effective)}. Inference still produces correct ` +\n            \"results, on the fallback provider.\",\n    );\n}\n\n/**\n * Name the tensor type of every graph input, straight from the model bytes.\n *\n * @param model The `.onnx` file contents.\n * @returns Input name → tensor type name. Empty when the file carries no\n *   readable graph, which callers read as \"assume float32\".\n */\nfunction declaredInputTypes(model: Uint8Array | ArrayBufferLike): Readonly<Record<string, string>> {\n    const declared = readModelInputTypes(model);\n    const named: Record<string, string> = {};\n    for (const [name, elemType] of Object.entries(declared)) {\n        named[name] = tensorTypeFor(elemType);\n    }\n    return named;\n}\n\n/**\n * Warn when the model's bytes were in hand but no input type came back.\n *\n * Every ONNX graph declares at least one input, so an empty result from bytes\n * means the reader failed, not that the model said nothing. Without this the\n * failure is invisible: the session falls back to float32, which is right for\n * most models and wrong for exactly the half-precision ones this reader exists\n * to support — so it looks like nothing is broken until a `predict()` throws\n * `Unexpected input data type`. That is how a 1 MB ceiling on the graph\n * descent shipped in 0.9.0 and reached a consumer.\n */\nfunction warnOnUnreadableTypes(): void {\n    console.warn(\n        \"Could not read any input type from this model's bytes; feeds will be built as float32. \" +\n            \"A half-precision model will fail at inference with 'Unexpected input data type'. \" +\n            \"Please report this with the model that produced it.\",\n    );\n}\n\n/**\n * Refuse a half-precision model in an environment that cannot feed one.\n *\n * ORT requires a real `Float16Array` for a `float16` tensor — the same bits in\n * a `Uint16Array` are rejected — and not every browser has one yet. Without\n * this check the session builds happily and the first `predict()` throws from\n * inside the preprocessing, which is both later and further from the cause.\n *\n * @param inputTypes Tensor type per input name.\n * @throws {@link ModelLoadError} when the graph wants half precision and the\n *   runtime has no `Float16Array`.\n */\nfunction assertFeedableTypes(inputTypes: Readonly<Record<string, string>>): void {\n    const half = Object.entries(inputTypes).filter(([, type]) => type === \"float16\");\n    if (half.length === 0 || hasFloat16Array()) return;\n    const names = half.map(([name]) => name).join(\", \");\n    throw new ModelLoadError(\n        `This model declares half-precision input(s) [${names}], but this environment has no ` +\n            \"Float16Array, which ONNX Runtime requires for a float16 tensor. Use a float32 export \" +\n            \"of the model, or run in a browser that supports Float16Array.\",\n    );\n}\n\n/**\n * Wrap an ONNX Runtime Web `InferenceSession` with convenient metadata access.\n *\n * The wrapper exposes input/output names and the shapes the graph declares,\n * manages execution-provider selection, provides a typed {@link OrtSession.run}\n * method, and releases the native session through {@link OrtSession.release}.\n */\nexport class OrtSession {\n    private constructor(\n        private readonly _session: ort.InferenceSession,\n        /**\n         * Execution providers this session is expected to run on.\n         *\n         * The requested list narrowed to what this browser can actually offer — a\n         * `webgpu` entry survives only where an adapter exists. Best-effort: ORT-Web\n         * exposes no way to ask which provider a session ended up on, so an entry\n         * here means \"not ruled out\", not \"confirmed\". See\n         * {@link requestedProviders} for what was asked for.\n         *\n         * When nothing survives — a caller asking for `webgpu` alone on a device\n         * without it — this falls back to {@link FALLBACK_PROVIDER}, which ORT-Web\n         * can always run. Handing ORT the unsatisfiable list instead makes\n         * `InferenceSession.create` reject with \"no available backend found\", so the\n         * page gets no inference at all rather than the slow-but-working fallback\n         * the `console.warn` describes. Measured in a real Chromium, where\n         * `navigator.gpu` exists but yields no adapter.\n         */\n        public readonly providers: readonly string[],\n        private readonly _metadata: Readonly<Record<string, string>>,\n        /**\n         * Execution providers that were asked for, after defaults were applied.\n         *\n         * Kept separate because ORT-Web falls back silently: a page that asks for\n         * `webgpu` on a device without it runs on WASM and is told nothing.\n         */\n        public readonly requestedProviders: readonly string[],\n        /**\n         * Tensor type each input declares, keyed by input name.\n         *\n         * Read from the model file, not from the session: `inputMetadata` is\n         * `undefined` on `onnxruntime-web` 1.20.1. Empty when the bytes were never\n         * in hand — a URL loaded with `readMetadata: false` — in which case every\n         * feed is built as float32, the behaviour before this existed.\n         */\n        private readonly _inputTypes: Readonly<Record<string, string>> = {},\n    ) {}\n\n    /**\n     * Tail of the run queue: settles when the latest queued run has finished.\n     *\n     * See {@link run} for why runs are serialized.\n     */\n    private _runQueue: Promise<unknown> = Promise.resolve();\n\n    /**\n     * Load an ONNX model into an ORT inference session.\n     *\n     * The metadata map is read **before** the session is built, and that order is\n     * load-bearing on memory-constrained devices. ORT copies the model into its\n     * WASM heap and then allocates the graph and the weights on top of that copy;\n     * a `readModelMetadata` call placed after `InferenceSession.create` keeps the\n     * JavaScript-side buffer reachable across the whole build, so a 5 MB model\n     * costs 5 MB of JS heap plus 5 MB of WASM heap plus the weights at the same\n     * instant. Reading first makes the buffer collectable as soon as ORT has copied\n     * it — on a phone that was the difference between a session and\n     * `Can't create a session. failed to allocate a buffer of size N`.\n     *\n     * The provider probe and the model download run concurrently: neither\n     * depends on the other, and on a real GPU `requestAdapter()` is not free.\n     * When the model is downloaded here, ORT's runtime is started alongside the\n     * download too — see {@link warmRuntime}.\n     *\n     * A model whose metadata carries {@link GRAPH_OPTIMIZATION_KEY} was already\n     * optimized offline by the Python SDK's `optimize_model`, so it is loaded\n     * with `graphOptimizationLevel: \"disabled\"` instead of paying the optimizer a\n     * second time — measured on a YOLO11n-seg under WASM, 29 ms of session\n     * creation down to 12 ms with unchanged inference time. An explicit\n     * `graphOptimizationLevel` in `sessionOptions` always wins. The mark is read\n     * with the rest of the metadata, so `readMetadata: false` skips it too.\n     *\n     * @param model Either a URL string, or a `Uint8Array`/`ArrayBuffer` containing the model bytes.\n     * @param options Provider list, pass-through `SessionOptions`, and whether to\n     *   read the model's metadata map (see {@link OrtSessionOptions.readMetadata}).\n     * @throws {@link ModelLoadError} if the model cannot be loaded.\n     */\n    static async create(model: ModelSource, options: OrtSessionOptions = {}): Promise<OrtSession> {\n        const requested = resolveProviders(options.providers);\n        const wantsMetadata = options.readMetadata !== false;\n        const cacheName =\n            options.cache === true ? DEFAULT_MODEL_CACHE : options.cache ? options.cache : null;\n        const shouldFetch = typeof model === \"string\" && (wantsMetadata || cacheName !== null);\n        const detection = detectProviders(requested).then((detected) =>\n            detected.length > 0 ? detected : [FALLBACK_PROVIDER],\n        );\n        const [detected, source] = await Promise.all([\n            detection,\n            shouldFetch ? fetchModel(model, cacheName) : Promise.resolve(model),\n            shouldFetch ? detection.then(warmRuntime) : Promise.resolve(),\n        ]);\n        const metadata =\n            wantsMetadata && typeof source !== \"string\" ? readModelMetadata(source) : {};\n        const specs = keepQuantizedOffWebGpu(detected, metadata);\n        const requestedNames = requested.map(providerName);\n        const providers = specs.map(providerName);\n        if (options.providers !== undefined && options.providers.length > 0) {\n            warnOnDroppedProviders(requestedNames, detected.map(providerName));\n        }\n        if (metadata[GRAPH_OPTIMIZATION_KEY] !== undefined && providers.includes(\"webgpu\")) {\n            warnPreOptimizedOnWebGpu();\n        }\n        const preOptimized =\n            metadata[GRAPH_OPTIMIZATION_KEY] !== undefined &&\n            options.sessionOptions?.graphOptimizationLevel === undefined;\n        const sessionOptions: ort.InferenceSession.SessionOptions = {\n            ...(preOptimized ? { graphOptimizationLevel: \"disabled\" as const } : {}),\n            ...(options.sessionOptions ?? {}),\n            executionProviders: specs as ort.InferenceSession.SessionOptions[\"executionProviders\"],\n        };\n        const inputTypes = typeof source !== \"string\" ? declaredInputTypes(source) : {};\n        if (typeof source !== \"string\" && Object.keys(inputTypes).length === 0) {\n            warnOnUnreadableTypes();\n        }\n        assertFeedableTypes(inputTypes);\n\n        let session: ort.InferenceSession;\n        try {\n            if (typeof source === \"string\") {\n                session = await ortRuntime.InferenceSession.create(source, sessionOptions);\n            } else if (source instanceof Uint8Array) {\n                session = await ortRuntime.InferenceSession.create(source, sessionOptions);\n            } else {\n                session = await ortRuntime.InferenceSession.create(\n                    source as ArrayBuffer,\n                    sessionOptions,\n                );\n            }\n        } catch (err) {\n            throw new ModelLoadError(`Failed to load ONNX model: ${(err as Error).message}`, {\n                cause: err,\n            });\n        }\n\n        return new OrtSession(session, providers, metadata, requestedNames, inputTypes);\n    }\n\n    /**\n     * Tensor type each input declares, in declaration order.\n     *\n     * `\"float32\"` for a normal export, `\"float16\"` for one exported with\n     * `half=True`. ORT matches a feed's type against this exactly, so the tasks\n     * read it and convert at the feed boundary.\n     */\n    get inputDtypes(): readonly string[] {\n        return this.inputNames.map((name) => this._inputTypes[name] ?? DEFAULT_TENSOR_TYPE);\n    }\n\n    /** Tensor type the first input declares. */\n    get inputDtype(): string {\n        return this._inputTypes[this.inputName] ?? DEFAULT_TENSOR_TYPE;\n    }\n\n    /** Names of the model's inputs, in declaration order. */\n    get inputNames(): readonly string[] {\n        return this._session.inputNames;\n    }\n\n    /** Name of the first (and usually only) input. */\n    get inputName(): string {\n        const name = this._session.inputNames[0];\n        if (name === undefined) {\n            throw new InferenceError(\"Model has no inputs.\");\n        }\n        return name;\n    }\n\n    /** Names of the model's outputs, in declaration order. */\n    get outputNames(): readonly string[] {\n        return this._session.outputNames;\n    }\n\n    /**\n     * Shapes the graph declares for its inputs, in declaration order.\n     *\n     * Dynamic (symbolic) axes appear as `null`. Empty shapes mean the runtime\n     * reported no metadata — either a non-tensor input, or an `onnxruntime-web`\n     * older than 1.21, which predates input metadata.\n     */\n    get inputShapes(): readonly DeclaredShape[] {\n        return declaredShapesFrom(\n            this._session.inputMetadata as\n                readonly ort.InferenceSession.ValueMetadata[] | undefined,\n        );\n    }\n\n    /**\n     * Shape the graph declares for its first input, dynamic axes as `null`.\n     *\n     * Empty when the runtime reports no metadata for it.\n     */\n    get inputShape(): DeclaredShape {\n        return this.inputShapes[0] ?? [];\n    }\n\n    /**\n     * Shapes the graph declares for its outputs, in declaration order.\n     *\n     * Dynamic (symbolic) axes appear as `null`. Reading them is how a task can\n     * tell how many classes a head emits without being told.\n     */\n    get outputShapes(): readonly DeclaredShape[] {\n        return declaredShapesFrom(\n            this._session.outputMetadata as\n                readonly ort.InferenceSession.ValueMetadata[] | undefined,\n        );\n    }\n\n    /**\n     * Shape the graph declares for its first output, dynamic axes as `null`.\n     *\n     * Empty when the runtime reports no metadata for it.\n     */\n    get outputShape(): DeclaredShape {\n        return this.outputShapes[0] ?? [];\n    }\n\n    /**\n     * The model's custom metadata map — `names`, `task`, `imgsz`, ... for an\n     * Ultralytics export.\n     *\n     * Read from the model's bytes at load time, since the runtime does not expose\n     * it. Empty when the session was created with `readMetadata: false`, from a\n     * URL that could not be fetched here, or from a model carrying no metadata.\n     */\n    get metadata(): Readonly<Record<string, string>> {\n        return this._metadata;\n    }\n\n    /**\n     * Release the native session and free its memory.\n     *\n     * Call it when a session is discarded while the page lives on — rebuilding a\n     * task at a different input size, swapping in a newer model. A failure from\n     * the runtime is ignored: a session being torn down has nothing left to fail\n     * at, and the caller is already moving on.\n     */\n    async release(): Promise<void> {\n        await this._session.release().catch(() => undefined);\n    }\n\n    /** The underlying `onnxruntime-web` session, for advanced use cases. */\n    get raw(): ort.InferenceSession {\n        return this._session;\n    }\n\n    /**\n     * Run inference and return all outputs.\n     *\n     * Runs on one session are serialized. ONNX Runtime Web refuses a second\n     * `run` while one is in flight — it rejects with `Session already started` —\n     * so two overlapping `predict()` calls on the same task used to fail. Queued\n     * here, they overlap where they can: while one run executes (in a worker,\n     * with `env.wasm.proxy`, or on the GPU), the next `predict()` decodes and\n     * preprocesses its frame and only waits for its turn at the runtime. A run\n     * that fails does not block the ones behind it.\n     *\n     * @param feeds Map of input name to `ort.Tensor`. Keys must match {@link inputNames}.\n     * @throws {@link InferenceError} if ORT raises any error during execution.\n     */\n    run(feeds: Record<string, ort.Tensor>): Promise<Record<string, ort.Tensor>> {\n        const turn = this._runQueue.then(() => this._runNow(feeds));\n        this._runQueue = turn.catch(() => undefined);\n        return turn;\n    }\n\n    /**\n     * Execute one run immediately; only {@link run}'s queue calls this.\n     *\n     * @param feeds Map of input name to `ort.Tensor`.\n     * @throws {@link InferenceError} if ORT raises any error during execution.\n     */\n    private async _runNow(feeds: Record<string, ort.Tensor>): Promise<Record<string, ort.Tensor>> {\n        try {\n            const result = await this._session.run(this._asDeclaredTypes(feeds));\n            return result as Record<string, ort.Tensor>;\n        } catch (err) {\n            throw new InferenceError(`Inference failed: ${(err as Error).message}`, { cause: err });\n        }\n    }\n\n    /**\n     * Rebuild any feed whose type does not match what its input declares.\n     *\n     * ORT matches feed types against the graph exactly: a float32 tensor against\n     * a `half=True` export fails the run with `Unexpected input data type`. Every\n     * preprocessing path in this SDK produces `Float32Array`, on purpose — the\n     * normalization arithmetic belongs in single precision — so the conversion\n     * happens here, once, at the only boundary all of them pass through. A feed\n     * that already carries the declared type is handed on untouched, so a caller\n     * building its own half tensor pays nothing.\n     *\n     * @param feeds The tensors to run with.\n     * @returns The same mapping, with mismatched float32 feeds converted.\n     */\n    private _asDeclaredTypes(feeds: Record<string, ort.Tensor>): Record<string, ort.Tensor> {\n        let converted: Record<string, ort.Tensor> | null = null;\n        for (const [name, tensor] of Object.entries(feeds)) {\n            const declared = this._inputTypes[name];\n            if (declared === undefined || declared === tensor.type) continue;\n            if (!(tensor.data instanceof Float32Array)) continue;\n            const data = toFeedData(tensor.data, declared);\n            if (data === tensor.data) continue;\n            converted ??= { ...feeds };\n            converted[name] = new ortRuntime.Tensor(\n                declared as \"float16\",\n                data as unknown as Float32Array,\n                tensor.dims as number[],\n            );\n        }\n        return converted ?? feeds;\n    }\n}\n"],"mappings":"oPA8BA,IAAa,EAAyB,oCAQhC,EACF,mHAGE,EAAiB,IAAI,IAmB3B,SAAS,EAAY,EAA+C,CAChE,IAAM,EAAM,EAAM,IAAI,EAAA,YAAY,CAAC,CAAC,KAAK,GAAG,EACxC,EAAU,EAAe,IAAI,CAAG,EAgBpC,OAfI,IAAY,IAAA,KACZ,GAAW,SAAY,CACnB,GAAI,CACA,IAAM,EAAQ,WAAW,KAAK,KAAK,CAAmB,EAAI,GAAM,EAAE,WAAW,CAAC,CAAC,EAK/E,MAAM,MAJgB,EAAW,iBAAiB,OAAO,EAAO,CAC5D,mBACI,CACR,CAAC,EAAA,CACa,QAAQ,CAC1B,MAAQ,CACJ,MACJ,CACJ,EAAA,CAAG,EACH,EAAe,IAAI,EAAK,CAAO,GAE5B,CACX,CAWA,SAAS,GAAiC,CACtC,QAAQ,KACJ,6MAGJ,CACJ,CAQA,IAAa,EAAmB,8BAiBhC,SAAS,EACL,EACA,EACc,CACd,GAAI,EAAA,iCAA+B,IAAA,GAAW,MAAO,CAAC,GAAG,CAAK,EAC9D,IAAM,EAAO,EAAM,OAAQ,GAAS,EAAA,aAAa,CAAI,IAAM,QAAQ,EAMnE,OALI,EAAK,SAAW,EAAM,OAAe,GACzC,QAAQ,KACJ,qEAAqE,EAAS,GAAkB,6FAEpG,EACO,EAAK,OAAS,EAAI,EAAO,CAAC,EAAA,iBAAiB,EACtD,CAGA,IAAa,EAAsB,wBAcnC,eAAe,EAAW,EAAa,EAAwD,CAC3F,IAAM,EAAQ,IAAc,KAAO,KAAO,MAAM,EAAU,CAAS,EACnE,GAAI,IAAU,KAAM,CAChB,IAAM,EAAM,MAAM,EAAM,MAAM,CAAG,CAAC,CAAC,UAAY,IAAA,EAAS,EACxD,GAAI,IAAQ,IAAA,GAAW,OAAO,IAAI,WAAW,MAAM,EAAI,YAAY,CAAC,CACxE,CACA,GAAI,CACA,IAAM,EAAW,MAAM,MAAM,CAAG,EAChC,GAAI,CAAC,EAAS,GAEV,OADA,EAAwB,EAAK,QAAQ,EAAS,OAAO,GAAG,EAAS,YAAY,EACtE,EAEX,IAAM,EAAS,MAAM,EAAS,YAAY,EAE1C,OADI,IAAU,MAAM,MAAM,EAAW,EAAO,EAAK,CAAM,EAChD,IAAI,WAAW,CAAM,CAChC,OAAS,EAAK,CAEV,OADA,EAAwB,EAAM,EAAc,OAAO,EAC5C,CACX,CACJ,CAeA,eAAe,EAAW,EAAc,EAAa,EAAoC,CACrF,GAAI,CACA,MAAM,EAAM,IAAI,EAAK,IAAI,SAAS,CAAM,CAAC,CAC7C,MAAQ,CACJ,MACJ,CACJ,CAcA,eAAe,EAAU,EAAqC,CAC1D,GAAI,OAAO,OAAW,IAAa,OAAO,KAC1C,GAAI,CACA,OAAO,MAAM,OAAO,KAAK,CAAI,CACjC,MAAQ,CACJ,OAAO,IACX,CACJ,CAcA,SAAS,EAAwB,EAAa,EAAsB,CAChE,QAAQ,KACJ,yCAAyC,EAAI,yBAAyB,EAAO,kMAIjF,CACJ,CA+DA,SAAS,EAAuB,EAA8B,EAAoC,CAC9F,IAAM,EAAU,EAAU,OAAQ,GAAa,CAAC,EAAU,SAAS,CAAQ,CAAC,EACxE,EAAQ,SAAW,GAGvB,QAAQ,KACJ,iEAAiE,KAAK,UAAU,CAAO,EAAE,4BAC1D,KAAK,UAAU,CAAS,EAAE,sEAE7D,CACJ,CASA,SAAS,EAAmB,EAAuE,CAC/F,IAAM,EAAW,EAAA,oBAAoB,CAAK,EACpC,EAAgC,CAAC,EACvC,IAAK,GAAM,CAAC,EAAM,KAAa,OAAO,QAAQ,CAAQ,EAClD,EAAM,GAAQ,EAAA,cAAc,CAAQ,EAExC,OAAO,CACX,CAaA,SAAS,GAA8B,CACnC,QAAQ,KACJ,6NAGJ,CACJ,CAcA,SAAS,EAAoB,EAAoD,CAC7E,IAAM,EAAO,OAAO,QAAQ,CAAU,CAAC,CAAC,QAAQ,EAAG,KAAU,IAAS,SAAS,EAC/E,GAAI,EAAK,SAAW,GAAK,EAAA,gBAAgB,EAAG,OAC5C,IAAM,EAAQ,EAAK,KAAK,CAAC,KAAU,CAAI,CAAC,CAAC,KAAK,IAAI,EAClD,MAAM,IAAI,EAAA,eACN,gDAAgD,EAAM,kLAG1D,CACJ,CASA,IAAa,EAAb,MAAa,CAAW,CAEC,SAkBD,UACC,UAOD,mBASC,YApCrB,YACI,EAkBA,EACA,EAOA,EASA,EAAiE,CAAC,EACpE,CApCmB,KAAA,SAAA,EAkBD,KAAA,UAAA,EACC,KAAA,UAAA,EAOD,KAAA,mBAAA,EASC,KAAA,YAAA,CAClB,CAOH,UAAsC,QAAQ,QAAQ,EAiCtD,aAAa,OAAO,EAAoB,EAA6B,CAAC,EAAwB,CAC1F,IAAM,EAAY,EAAA,iBAAiB,EAAQ,SAAS,EAC9C,EAAgB,EAAQ,eAAiB,GACzC,EACF,EAAQ,QAAU,GAAO,EAAsB,EAAQ,MAAQ,EAAQ,MAAQ,KAC7E,EAAc,OAAO,GAAU,WAAa,GAAiB,IAAc,MAC3E,EAAY,EAAA,gBAAgB,CAAS,CAAC,CAAC,KAAM,GAC/C,EAAS,OAAS,EAAI,EAAW,CAAC,EAAA,iBAAiB,CACvD,EACM,CAAC,EAAU,GAAU,MAAM,QAAQ,IAAI,CACzC,EACA,EAAc,EAAW,EAAO,CAAS,EAAI,QAAQ,QAAQ,CAAK,EAClE,EAAc,EAAU,KAAK,CAAW,EAAI,QAAQ,QAAQ,CAChE,CAAC,EACK,EACF,GAAiB,OAAO,GAAW,SAAW,EAAA,kBAAkB,CAAM,EAAI,CAAC,EACzE,EAAQ,EAAuB,EAAU,CAAQ,EACjD,EAAiB,EAAU,IAAI,EAAA,YAAY,EAC3C,EAAY,EAAM,IAAI,EAAA,YAAY,EACpC,EAAQ,YAAc,IAAA,IAAa,EAAQ,UAAU,OAAS,GAC9D,EAAuB,EAAgB,EAAS,IAAI,EAAA,YAAY,CAAC,EAEjE,EAAA,uCAAqC,IAAA,IAAa,EAAU,SAAS,QAAQ,GAC7E,EAAyB,EAK7B,IAAM,EAAsD,CACxD,GAHA,EAAA,uCAAqC,IAAA,IACrC,EAAQ,gBAAgB,yBAA2B,IAAA,GAEhC,CAAE,uBAAwB,UAAoB,EAAI,CAAC,EACtE,GAAI,EAAQ,gBAAkB,CAAC,EAC/B,mBAAoB,CACxB,EACM,EAAa,OAAO,GAAW,SAAwC,CAAC,EAA9B,EAAmB,CAAM,EACrE,OAAO,GAAW,UAAY,OAAO,KAAK,CAAU,CAAC,CAAC,SAAW,GACjE,EAAsB,EAE1B,EAAoB,CAAU,EAE9B,IAAI,EACJ,GAAI,CACA,AAKI,GALA,OAAO,GAAW,UAEX,aAAkB,WADf,MAAM,EAAW,iBAAiB,OAAO,EAAQ,CAAc,EASjF,OAAS,EAAK,CACV,MAAM,IAAI,EAAA,eAAe,8BAA+B,EAAc,UAAW,CAC7E,MAAO,CACX,CAAC,CACL,CAEA,OAAO,IAAI,EAAW,EAAS,EAAW,EAAU,EAAgB,CAAU,CAClF,CASA,IAAI,aAAiC,CACjC,OAAO,KAAK,WAAW,IAAK,GAAS,KAAK,YAAY,IAAA,SAA4B,CACtF,CAGA,IAAI,YAAqB,CACrB,OAAO,KAAK,YAAY,KAAK,YAAA,SACjC,CAGA,IAAI,YAAgC,CAChC,OAAO,KAAK,SAAS,UACzB,CAGA,IAAI,WAAoB,CACpB,IAAM,EAAO,KAAK,SAAS,WAAW,GACtC,GAAI,IAAS,IAAA,GACT,MAAM,IAAI,EAAA,eAAe,sBAAsB,EAEnD,OAAO,CACX,CAGA,IAAI,aAAiC,CACjC,OAAO,KAAK,SAAS,WACzB,CASA,IAAI,aAAwC,CACxC,OAAO,EAAA,mBACH,KAAK,SAAS,aAElB,CACJ,CAOA,IAAI,YAA4B,CAC5B,OAAO,KAAK,YAAY,IAAM,CAAC,CACnC,CAQA,IAAI,cAAyC,CACzC,OAAO,EAAA,mBACH,KAAK,SAAS,cAElB,CACJ,CAOA,IAAI,aAA6B,CAC7B,OAAO,KAAK,aAAa,IAAM,CAAC,CACpC,CAUA,IAAI,UAA6C,CAC7C,OAAO,KAAK,SAChB,CAUA,MAAM,SAAyB,CAC3B,MAAM,KAAK,SAAS,QAAQ,CAAC,CAAC,UAAY,IAAA,EAAS,CACvD,CAGA,IAAI,KAA4B,CAC5B,OAAO,KAAK,QAChB,CAgBA,IAAI,EAAwE,CACxE,IAAM,EAAO,KAAK,UAAU,SAAW,KAAK,QAAQ,CAAK,CAAC,EAE1D,MADA,MAAK,UAAY,EAAK,UAAY,IAAA,EAAS,EACpC,CACX,CAQA,MAAc,QAAQ,EAAwE,CAC1F,GAAI,CAEA,OAAO,MADc,KAAK,SAAS,IAAI,KAAK,iBAAiB,CAAK,CAAC,CAEvE,OAAS,EAAK,CACV,MAAM,IAAI,EAAA,eAAe,qBAAsB,EAAc,UAAW,CAAE,MAAO,CAAI,CAAC,CAC1F,CACJ,CAgBA,iBAAyB,EAA+D,CACpF,IAAI,EAA+C,KACnD,IAAK,GAAM,CAAC,EAAM,KAAW,OAAO,QAAQ,CAAK,EAAG,CAChD,IAAM,EAAW,KAAK,YAAY,GAElC,GADI,IAAa,IAAA,IAAa,IAAa,EAAO,MAC9C,EAAE,EAAO,gBAAgB,cAAe,SAC5C,IAAM,EAAO,EAAA,WAAW,EAAO,KAAM,CAAQ,EACzC,IAAS,EAAO,OACpB,IAAc,CAAE,GAAG,CAAM,EACzB,EAAU,GAAQ,IAAI,EAAW,OAC7B,EACA,EACA,EAAO,IACX,EACJ,CACA,OAAO,GAAa,CACxB,CACJ"}

package/dist/vision/core/session.js

@@ -1,45 +1,119 @@
 import { InferenceError as e, ModelLoadError as t } from "./exceptions.js";
-import { declaredShapesFrom as n } from "./graph.js";
-import { readModelMetadata as r } from "./metadata.js";
-import { FALLBACK_PROVIDER as i, detectProviders as a, resolveProviders as o } from "./providers.js";
-import * as s from "onnxruntime-web";
+import { hasFloat16Array as n, tensorTypeFor as r, toFeedData as i } from "./dtypes.js";
+import { declaredShapesFrom as a } from "./graph.js";
+import { readModelInputTypes as o, readModelMetadata as s } from "./metadata.js";
+import { FALLBACK_PROVIDER as c, detectProviders as l, providerName as u, resolveProviders as d } from "./providers.js";
+import * as f from "onnxruntime-web";
 //#region src/vision/core/session.ts
-async function c(e) {
+var p = "ort_vision_sdk.graph_optimization", m = "CAcSDm9ydC12aXNpb24tc2RrOjoKEAoBeBIBeSIISWRlbnRpdHkSBHdhcm1aDwoBeBIKCggIARIECgIIAWIPCgF5EgoKCAgBEgQKAggBQgQKABAN", h = /* @__PURE__ */ new Map();
+function g(e) {
+	let t = e.map(u).join(","), n = h.get(t);
+	return n === void 0 && (n = (async () => {
+		try {
+			let t = Uint8Array.from(atob(m), (e) => e.charCodeAt(0));
+			await (await f.InferenceSession.create(t, { executionProviders: e })).release();
+		} catch {
+			return;
+		}
+	})(), h.set(t, n)), n;
+}
+function _() {
+	console.warn("[@ort-vision-sdk/web] This model was pre-optimized by optimize_model, which targets the WASM backend; on WebGPU some of its fused nodes fall back to the CPU. For WebGPU, ship the original export instead.");
+}
+var v = "ort_vision_sdk.quantization";
+function y(e, t) {
+	if (t["ort_vision_sdk.quantization"] === void 0) return [...e];
+	let n = e.filter((e) => u(e) !== "webgpu");
+	return n.length === e.length ? n : (console.warn(`[@ort-vision-sdk/web] This model was quantized by quantize_model (${t[v]}); ONNX Runtime Web's WebGPU backend cannot run its quantized operators, so it runs on WASM.`), n.length > 0 ? n : [c]);
+}
+var b = "ort-vision-sdk-models";
+async function x(e, t) {
+	let n = t === null ? null : await C(t);
+	if (n !== null) {
+		let t = await n.match(e).catch(() => void 0);
+		if (t !== void 0) return new Uint8Array(await t.arrayBuffer());
+	}
 	try {
 		let t = await fetch(e);
-		return t.ok ? new Uint8Array(await t.arrayBuffer()) : (l(e, `HTTP ${t.status} ${t.statusText}`), e);
+		if (!t.ok) return w(e, `HTTP ${t.status} ${t.statusText}`), e;
+		let r = await t.arrayBuffer();
+		return n !== null && await S(n, e, r), new Uint8Array(r);
 	} catch (t) {
-		return l(e, t.message), e;
+		return w(e, t.message), e;
 	}
 }
-function l(e, t) {
+async function S(e, t, n) {
+	try {
+		await e.put(t, new Response(n));
+	} catch {
+		return;
+	}
+}
+async function C(e) {
+	if (typeof caches > "u") return null;
+	try {
+		return await caches.open(e);
+	} catch {
+		return null;
+	}
+}
+function w(e, t) {
 	console.warn(`[@ort-vision-sdk/web] Could not fetch ${e} to read its metadata (${t}). Letting ONNX Runtime load it instead: the model will work, but its baked-in class names are unavailable, so labels fall back to class_0, class_1, ... Pass \`labels\` explicitly to name them.`);
 }
-function u(e, t) {
+function T(e, t) {
 	let n = e.filter((e) => !t.includes(e));
 	n.length !== 0 && console.warn(`This browser cannot offer the requested execution provider(s) ${JSON.stringify(n)}; the session will run on ${JSON.stringify(t)}. Inference still produces correct results, on the fallback provider.`);
 }
-var d = class l {
+function E(e) {
+	let t = o(e), n = {};
+	for (let [e, i] of Object.entries(t)) n[e] = r(i);
+	return n;
+}
+function D() {
+	console.warn("Could not read any input type from this model's bytes; feeds will be built as float32. A half-precision model will fail at inference with 'Unexpected input data type'. Please report this with the model that produced it.");
+}
+function O(e) {
+	let r = Object.entries(e).filter(([, e]) => e === "float16");
+	if (r.length === 0 || n()) return;
+	let i = r.map(([e]) => e).join(", ");
+	throw new t(`This model declares half-precision input(s) [${i}], but this environment has no Float16Array, which ONNX Runtime requires for a float16 tensor. Use a float32 export of the model, or run in a browser that supports Float16Array.`);
+}
+var k = class n {
 	_session;
 	providers;
 	_metadata;
 	requestedProviders;
-	constructor(e, t, n, r) {
-		this._session = e, this.providers = t, this._metadata = n, this.requestedProviders = r;
-	}
-	static async create(e, n = {}) {
-		let d = o(n.providers), f = await a(d), p = f.length > 0 ? f : [i];
-		n.providers !== void 0 && n.providers.length > 0 && u(d, p);
-		let m = {
-			...n.sessionOptions ?? {},
-			executionProviders: p
-		}, h = n.readMetadata !== !1, g = typeof e == "string" && h ? await c(e) : e, _ = h && typeof g != "string" ? r(g) : {}, v;
+	_inputTypes;
+	constructor(e, t, n, r, i = {}) {
+		this._session = e, this.providers = t, this._metadata = n, this.requestedProviders = r, this._inputTypes = i;
+	}
+	_runQueue = Promise.resolve();
+	static async create(e, r = {}) {
+		let i = d(r.providers), a = r.readMetadata !== !1, o = r.cache === !0 ? b : r.cache ? r.cache : null, p = typeof e == "string" && (a || o !== null), m = l(i).then((e) => e.length > 0 ? e : [c]), [h, v] = await Promise.all([
+			m,
+			p ? x(e, o) : Promise.resolve(e),
+			p ? m.then(g) : Promise.resolve()
+		]), S = a && typeof v != "string" ? s(v) : {}, C = y(h, S), w = i.map(u), k = C.map(u);
+		r.providers !== void 0 && r.providers.length > 0 && T(w, h.map(u)), S["ort_vision_sdk.graph_optimization"] !== void 0 && k.includes("webgpu") && _();
+		let A = {
+			...S["ort_vision_sdk.graph_optimization"] !== void 0 && r.sessionOptions?.graphOptimizationLevel === void 0 ? { graphOptimizationLevel: "disabled" } : {},
+			...r.sessionOptions ?? {},
+			executionProviders: C
+		}, j = typeof v == "string" ? {} : E(v);
+		typeof v != "string" && Object.keys(j).length === 0 && D(), O(j);
+		let M;
 		try {
-			v = (typeof g == "string" || g instanceof Uint8Array, await s.InferenceSession.create(g, m));
+			M = (typeof v == "string" || v instanceof Uint8Array, await f.InferenceSession.create(v, A));
 		} catch (e) {
 			throw new t(`Failed to load ONNX model: ${e.message}`, { cause: e });
 		}
-		return new l(v, p, _, d);
+		return new n(M, k, S, w, j);
+	}
+	get inputDtypes() {
+		return this.inputNames.map((e) => this._inputTypes[e] ?? "float32");
+	}
+	get inputDtype() {
+		return this._inputTypes[this.inputName] ?? "float32";
 	}
 	get inputNames() {
 		return this._session.inputNames;
@@ -53,13 +127,13 @@ var d = class l {
 		return this._session.outputNames;
 	}
 	get inputShapes() {
-		return n(this._session.inputMetadata);
+		return a(this._session.inputMetadata);
 	}
 	get inputShape() {
 		return this.inputShapes[0] ?? [];
 	}
 	get outputShapes() {
-		return n(this._session.outputMetadata);
+		return a(this._session.outputMetadata);
 	}
 	get outputShape() {
 		return this.outputShapes[0] ?? [];
@@ -73,15 +147,29 @@ var d = class l {
 	get raw() {
 		return this._session;
 	}
-	async run(t) {
+	run(e) {
+		let t = this._runQueue.then(() => this._runNow(e));
+		return this._runQueue = t.catch(() => void 0), t;
+	}
+	async _runNow(t) {
 		try {
-			return await this._session.run(t);
+			return await this._session.run(this._asDeclaredTypes(t));
 		} catch (t) {
 			throw new e(`Inference failed: ${t.message}`, { cause: t });
 		}
 	}
+	_asDeclaredTypes(e) {
+		let t = null;
+		for (let [n, r] of Object.entries(e)) {
+			let a = this._inputTypes[n];
+			if (a === void 0 || a === r.type || !(r.data instanceof Float32Array)) continue;
+			let o = i(r.data, a);
+			o !== r.data && (t ??= { ...e }, t[n] = new f.Tensor(a, o, r.dims));
+		}
+		return t ?? e;
+	}
 };
 //#endregion
-export { d as OrtSession };
+export { b as DEFAULT_MODEL_CACHE, p as GRAPH_OPTIMIZATION_KEY, k as OrtSession, v as QUANTIZATION_KEY };
 
 //# sourceMappingURL=session.js.map

package/dist/vision/core/session.js.map

@@ -1 +1 @@
-{"version":3,"file":"session.js","names":[],"sources":["../../../src/vision/core/session.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * Thin wrapper around `onnxruntime-web` `InferenceSession` with typed metadata.\n */\n\nimport type * as ort from \"onnxruntime-web\";\nimport * as ortRuntime from \"onnxruntime-web\";\n\nimport { InferenceError, ModelLoadError } from \"./exceptions\";\nimport { type DeclaredShape, declaredShapesFrom } from \"./graph\";\nimport { readModelMetadata } from \"./metadata\";\nimport { FALLBACK_PROVIDER, detectProviders, resolveProviders } from \"./providers\";\n\n/** Anything `InferenceSession.create` accepts. */\nexport type ModelSource = string | ArrayBufferLike | Uint8Array;\n\n/**\n * Fetch a model URL as bytes so its metadata can be read.\n *\n * Falls back to the URL itself when the fetch fails, letting ORT try its own\n * load path: losing the metadata map is a downgrade, but failing to load a model\n * that ORT could have fetched would be a regression.\n *\n * @param url Where the `.onnx` lives.\n * @returns The model bytes, or the original URL when they could not be fetched.\n */\nasync function fetchModel(url: string): Promise<Uint8Array | string> {\n    try {\n        const response = await fetch(url);\n        if (!response.ok) {\n            warnMetadataUnavailable(url, `HTTP ${response.status} ${response.statusText}`);\n            return url;\n        }\n        return new Uint8Array(await response.arrayBuffer());\n    } catch (err) {\n        warnMetadataUnavailable(url, (err as Error).message);\n        return url;\n    }\n}\n\n/**\n * Warn that a model's metadata could not be read, and say what that costs.\n *\n * The fallback itself is right — losing the metadata beats failing a load that\n * ORT could have completed on its own — but it used to be silent, and the\n * symptom it produces is remote from the cause: class names come back as\n * `class_0`, `class_1`, ... with nothing anywhere explaining why. Whoever hits\n * this needs to be told that passing `labels` is the way out.\n *\n * @param url The model URL that could not be fetched here.\n * @param reason What went wrong, as reported by `fetch`.\n */\nfunction warnMetadataUnavailable(url: string, reason: string): void {\n    console.warn(\n        `[@ort-vision-sdk/web] Could not fetch ${url} to read its metadata (${reason}). ` +\n            \"Letting ONNX Runtime load it instead: the model will work, but its baked-in \" +\n            \"class names are unavailable, so labels fall back to class_0, class_1, ... \" +\n            \"Pass `labels` explicitly to name them.\",\n    );\n}\n\nexport interface OrtSessionOptions {\n    /**\n     * Execution providers in preference order. `undefined` uses {@link DEFAULT_PROVIDERS}.\n     *\n     * Naming one explicitly also opts into a `console.warn` when this browser\n     * cannot offer it, instead of falling back in silence.\n     */\n    readonly providers?: readonly string[];\n    /** Optional ORT session options forwarded to `InferenceSession.create`. */\n    readonly sessionOptions?: ort.InferenceSession.SessionOptions;\n    /**\n     * Whether to read the model's custom metadata map (`names`, `task`, `imgsz`).\n     * Defaults to `true`.\n     *\n     * The runtime does not expose that map, so it is read from the file itself —\n     * which means a URL model is fetched here and handed to ORT as bytes instead\n     * of letting ORT fetch it. That is the same single download either way, and\n     * it is what lets a task resolve its labels off the model. Set to `false` to\n     * keep the URL path untouched and leave {@link OrtSession.metadata} empty.\n     *\n     * `false` is also the escape hatch when a device cannot afford the bytes: the\n     * fetched buffer is dropped before ORT builds the graph (see\n     * {@link OrtSession.create}), but ORT's own load path still keeps the model out\n     * of reach of anything the SDK holds. A session built this way resolves its\n     * input size from the graph as usual — only the class names are lost, so a\n     * caller taking this route has to pass `labels` itself.\n     */\n    readonly readMetadata?: boolean;\n}\n\n/**\n * Warn when a provider named explicitly by the caller is not going to run.\n *\n * Only explicit requests are worth a warning. The default list exists precisely\n * so that falling from `webgpu` to `wasm` is the expected outcome — but a caller\n * who wrote `providers: [\"webgpu\"]` and lands on WASM has a page several times\n * slower than intended and nothing in the console to explain it.\n *\n * @param requested Providers that were asked for.\n * @param effective Providers that survived capability detection.\n */\nfunction warnOnDroppedProviders(requested: readonly string[], effective: readonly string[]): void {\n    const dropped = requested.filter((provider) => !effective.includes(provider));\n    if (dropped.length === 0) {\n        return;\n    }\n    console.warn(\n        `This browser cannot offer the requested execution provider(s) ${JSON.stringify(dropped)}; ` +\n            `the session will run on ${JSON.stringify(effective)}. Inference still produces correct ` +\n            \"results, on the fallback provider.\",\n    );\n}\n\n/**\n * Wrap an ONNX Runtime Web `InferenceSession` with convenient metadata access.\n *\n * The wrapper exposes input/output names and the shapes the graph declares,\n * manages execution-provider selection, provides a typed {@link OrtSession.run}\n * method, and releases the native session through {@link OrtSession.release}.\n */\nexport class OrtSession {\n    private constructor(\n        private readonly _session: ort.InferenceSession,\n        /**\n         * Execution providers this session is expected to run on.\n         *\n         * The requested list narrowed to what this browser can actually offer — a\n         * `webgpu` entry survives only where an adapter exists. Best-effort: ORT-Web\n         * exposes no way to ask which provider a session ended up on, so an entry\n         * here means \"not ruled out\", not \"confirmed\". See\n         * {@link requestedProviders} for what was asked for.\n         *\n         * When nothing survives — a caller asking for `webgpu` alone on a device\n         * without it — this falls back to {@link FALLBACK_PROVIDER}, which ORT-Web\n         * can always run. Handing ORT the unsatisfiable list instead makes\n         * `InferenceSession.create` reject with \"no available backend found\", so the\n         * page gets no inference at all rather than the slow-but-working fallback\n         * the `console.warn` describes. Measured in a real Chromium, where\n         * `navigator.gpu` exists but yields no adapter.\n         */\n        public readonly providers: readonly string[],\n        private readonly _metadata: Readonly<Record<string, string>>,\n        /**\n         * Execution providers that were asked for, after defaults were applied.\n         *\n         * Kept separate because ORT-Web falls back silently: a page that asks for\n         * `webgpu` on a device without it runs on WASM and is told nothing.\n         */\n        public readonly requestedProviders: readonly string[],\n    ) {}\n\n    /**\n     * Load an ONNX model into an ORT inference session.\n     *\n     * The metadata map is read **before** the session is built, and that order is\n     * load-bearing on memory-constrained devices. ORT copies the model into its\n     * WASM heap and then allocates the graph and the weights on top of that copy;\n     * a `readModelMetadata` call placed after `InferenceSession.create` keeps the\n     * JavaScript-side buffer reachable across the whole build, so a 5 MB model\n     * costs 5 MB of JS heap plus 5 MB of WASM heap plus the weights at the same\n     * instant. Reading first makes the buffer collectable as soon as ORT has copied\n     * it — on a phone that was the difference between a session and\n     * `Can't create a session. failed to allocate a buffer of size N`.\n     *\n     * @param model Either a URL string, or a `Uint8Array`/`ArrayBuffer` containing the model bytes.\n     * @param options Provider list, pass-through `SessionOptions`, and whether to\n     *   read the model's metadata map (see {@link OrtSessionOptions.readMetadata}).\n     * @throws {@link ModelLoadError} if the model cannot be loaded.\n     */\n    static async create(model: ModelSource, options: OrtSessionOptions = {}): Promise<OrtSession> {\n        const requested = resolveProviders(options.providers);\n        const detected = await detectProviders(requested);\n        const providers = detected.length > 0 ? detected : [FALLBACK_PROVIDER];\n        if (options.providers !== undefined && options.providers.length > 0) {\n            warnOnDroppedProviders(requested, providers);\n        }\n        const sessionOptions: ort.InferenceSession.SessionOptions = {\n            ...(options.sessionOptions ?? {}),\n            executionProviders:\n                providers as ort.InferenceSession.SessionOptions[\"executionProviders\"],\n        };\n        const wantsMetadata = options.readMetadata !== false;\n        const source = typeof model === \"string\" && wantsMetadata ? await fetchModel(model) : model;\n        const metadata =\n            wantsMetadata && typeof source !== \"string\" ? readModelMetadata(source) : {};\n\n        let session: ort.InferenceSession;\n        try {\n            if (typeof source === \"string\") {\n                session = await ortRuntime.InferenceSession.create(source, sessionOptions);\n            } else if (source instanceof Uint8Array) {\n                session = await ortRuntime.InferenceSession.create(source, sessionOptions);\n            } else {\n                session = await ortRuntime.InferenceSession.create(\n                    source as ArrayBuffer,\n                    sessionOptions,\n                );\n            }\n        } catch (err) {\n            throw new ModelLoadError(`Failed to load ONNX model: ${(err as Error).message}`, {\n                cause: err,\n            });\n        }\n\n        return new OrtSession(session, providers, metadata, requested);\n    }\n\n    /** Names of the model's inputs, in declaration order. */\n    get inputNames(): readonly string[] {\n        return this._session.inputNames;\n    }\n\n    /** Name of the first (and usually only) input. */\n    get inputName(): string {\n        const name = this._session.inputNames[0];\n        if (name === undefined) {\n            throw new InferenceError(\"Model has no inputs.\");\n        }\n        return name;\n    }\n\n    /** Names of the model's outputs, in declaration order. */\n    get outputNames(): readonly string[] {\n        return this._session.outputNames;\n    }\n\n    /**\n     * Shapes the graph declares for its inputs, in declaration order.\n     *\n     * Dynamic (symbolic) axes appear as `null`. Empty shapes mean the runtime\n     * reported no metadata — either a non-tensor input, or an `onnxruntime-web`\n     * older than 1.21, which predates input metadata.\n     */\n    get inputShapes(): readonly DeclaredShape[] {\n        return declaredShapesFrom(\n            this._session.inputMetadata as\n                readonly ort.InferenceSession.ValueMetadata[] | undefined,\n        );\n    }\n\n    /**\n     * Shape the graph declares for its first input, dynamic axes as `null`.\n     *\n     * Empty when the runtime reports no metadata for it.\n     */\n    get inputShape(): DeclaredShape {\n        return this.inputShapes[0] ?? [];\n    }\n\n    /**\n     * Shapes the graph declares for its outputs, in declaration order.\n     *\n     * Dynamic (symbolic) axes appear as `null`. Reading them is how a task can\n     * tell how many classes a head emits without being told.\n     */\n    get outputShapes(): readonly DeclaredShape[] {\n        return declaredShapesFrom(\n            this._session.outputMetadata as\n                readonly ort.InferenceSession.ValueMetadata[] | undefined,\n        );\n    }\n\n    /**\n     * Shape the graph declares for its first output, dynamic axes as `null`.\n     *\n     * Empty when the runtime reports no metadata for it.\n     */\n    get outputShape(): DeclaredShape {\n        return this.outputShapes[0] ?? [];\n    }\n\n    /**\n     * The model's custom metadata map — `names`, `task`, `imgsz`, ... for an\n     * Ultralytics export.\n     *\n     * Read from the model's bytes at load time, since the runtime does not expose\n     * it. Empty when the session was created with `readMetadata: false`, from a\n     * URL that could not be fetched here, or from a model carrying no metadata.\n     */\n    get metadata(): Readonly<Record<string, string>> {\n        return this._metadata;\n    }\n\n    /**\n     * Release the native session and free its memory.\n     *\n     * Call it when a session is discarded while the page lives on — rebuilding a\n     * task at a different input size, swapping in a newer model. A failure from\n     * the runtime is ignored: a session being torn down has nothing left to fail\n     * at, and the caller is already moving on.\n     */\n    async release(): Promise<void> {\n        await this._session.release().catch(() => undefined);\n    }\n\n    /** The underlying `onnxruntime-web` session, for advanced use cases. */\n    get raw(): ort.InferenceSession {\n        return this._session;\n    }\n\n    /**\n     * Run inference and return all outputs.\n     *\n     * @param feeds Map of input name to `ort.Tensor`. Keys must match {@link inputNames}.\n     * @throws {@link InferenceError} if ORT raises any error during execution.\n     */\n    async run(feeds: Record<string, ort.Tensor>): Promise<Record<string, ort.Tensor>> {\n        try {\n            const result = await this._session.run(feeds);\n            return result as Record<string, ort.Tensor>;\n        } catch (err) {\n            throw new InferenceError(`Inference failed: ${(err as Error).message}`, { cause: err });\n        }\n    }\n}\n"],"mappings":";;;;;;AA0BA,eAAe,EAAW,GAA2C;CACjE,IAAI;EACA,IAAM,IAAW,MAAM,MAAM,CAAG;EAKhC,OAJK,EAAS,KAIP,IAAI,WAAW,MAAM,EAAS,YAAY,CAAC,KAH9C,EAAwB,GAAK,QAAQ,EAAS,OAAO,GAAG,EAAS,YAAY,GACtE;CAGf,SAAS,GAAK;EAEV,OADA,EAAwB,GAAM,EAAc,OAAO,GAC5C;CACX;AACJ;AAcA,SAAS,EAAwB,GAAa,GAAsB;CAChE,QAAQ,KACJ,yCAAyC,EAAI,yBAAyB,EAAO,kMAIjF;AACJ;AA2CA,SAAS,EAAuB,GAA8B,GAAoC;CAC9F,IAAM,IAAU,EAAU,QAAQ,MAAa,CAAC,EAAU,SAAS,CAAQ,CAAC;CACxE,EAAQ,WAAW,KAGvB,QAAQ,KACJ,iEAAiE,KAAK,UAAU,CAAO,EAAE,4BAC1D,KAAK,UAAU,CAAS,EAAE,sEAE7D;AACJ;AASA,IAAa,IAAb,MAAa,EAAW;CAEC;CAkBD;CACC;CAOD;CA3BpB,YACI,GAkBA,GACA,GAOA,GACF;EADkB,AA1BC,KAAA,WAAA,GAkBD,KAAA,YAAA,GACC,KAAA,YAAA,GAOD,KAAA,qBAAA;CACjB;CAoBH,aAAa,OAAO,GAAoB,IAA6B,CAAC,GAAwB;EAC1F,IAAM,IAAY,EAAiB,EAAQ,SAAS,GAC9C,IAAW,MAAM,EAAgB,CAAS,GAC1C,IAAY,EAAS,SAAS,IAAI,IAAW,CAAC,CAAiB;EACrE,AAAI,EAAQ,cAAc,KAAA,KAAa,EAAQ,UAAU,SAAS,KAC9D,EAAuB,GAAW,CAAS;EAE/C,IAAM,IAAsD;GACxD,GAAI,EAAQ,kBAAkB,CAAC;GAC/B,oBACI;EACR,GACM,IAAgB,EAAQ,iBAAiB,IACzC,IAAS,OAAO,KAAU,YAAY,IAAgB,MAAM,EAAW,CAAK,IAAI,GAChF,IACF,KAAiB,OAAO,KAAW,WAAW,EAAkB,CAAM,IAAI,CAAC,GAE3E;EACJ,IAAI;GACA,AAKI,KALA,OAAO,KAAW,YAEX,aAAkB,YADf,MAAM,EAAW,iBAAiB,OAAO,GAAQ,CAAc;EASjF,SAAS,GAAK;GACV,MAAM,IAAI,EAAe,8BAA+B,EAAc,WAAW,EAC7E,OAAO,EACX,CAAC;EACL;EAEA,OAAO,IAAI,EAAW,GAAS,GAAW,GAAU,CAAS;CACjE;CAGA,IAAI,aAAgC;EAChC,OAAO,KAAK,SAAS;CACzB;CAGA,IAAI,YAAoB;EACpB,IAAM,IAAO,KAAK,SAAS,WAAW;EACtC,IAAI,MAAS,KAAA,GACT,MAAM,IAAI,EAAe,sBAAsB;EAEnD,OAAO;CACX;CAGA,IAAI,cAAiC;EACjC,OAAO,KAAK,SAAS;CACzB;CASA,IAAI,cAAwC;EACxC,OAAO,EACH,KAAK,SAAS,aAElB;CACJ;CAOA,IAAI,aAA4B;EAC5B,OAAO,KAAK,YAAY,MAAM,CAAC;CACnC;CAQA,IAAI,eAAyC;EACzC,OAAO,EACH,KAAK,SAAS,cAElB;CACJ;CAOA,IAAI,cAA6B;EAC7B,OAAO,KAAK,aAAa,MAAM,CAAC;CACpC;CAUA,IAAI,WAA6C;EAC7C,OAAO,KAAK;CAChB;CAUA,MAAM,UAAyB;EAC3B,MAAM,KAAK,SAAS,QAAQ,CAAC,CAAC,YAAY,KAAA,CAAS;CACvD;CAGA,IAAI,MAA4B;EAC5B,OAAO,KAAK;CAChB;CAQA,MAAM,IAAI,GAAwE;EAC9E,IAAI;GAEA,OAAO,MADc,KAAK,SAAS,IAAI,CAAK;EAEhD,SAAS,GAAK;GACV,MAAM,IAAI,EAAe,qBAAsB,EAAc,WAAW,EAAE,OAAO,EAAI,CAAC;EAC1F;CACJ;AACJ"}
+{"version":3,"file":"session.js","names":[],"sources":["../../../src/vision/core/session.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * Thin wrapper around `onnxruntime-web` `InferenceSession` with typed metadata.\n */\n\nimport type * as ort from \"onnxruntime-web\";\nimport * as ortRuntime from \"onnxruntime-web\";\n\nimport { DEFAULT_TENSOR_TYPE, hasFloat16Array, tensorTypeFor, toFeedData } from \"./dtypes\";\nimport { InferenceError, ModelLoadError } from \"./exceptions\";\nimport { type DeclaredShape, declaredShapesFrom } from \"./graph\";\nimport { readModelInputTypes, readModelMetadata } from \"./metadata\";\nimport {\n    FALLBACK_PROVIDER,\n    type ProviderSpec,\n    detectProviders,\n    providerName,\n    resolveProviders,\n} from \"./providers\";\n\n/** Anything `InferenceSession.create` accepts. */\nexport type ModelSource = string | ArrayBufferLike | Uint8Array;\n\n/**\n * Metadata key naming the level a model was graph-optimized at offline.\n *\n * Written by the Python SDK's `optimize_model`; both test suites pin the same\n * string, so a rename on one side fails. See {@link OrtSession.create} for what\n * the session does when it finds it.\n */\nexport const GRAPH_OPTIMIZATION_KEY = \"ort_vision_sdk.graph_optimization\";\n\n/**\n * An 84-byte ONNX model — one `Identity` on a float `[1]` — used to start the runtime.\n *\n * Generated with `onnx.helper` (opset 13, IR 7, producer `ort-vision-sdk`).\n * See {@link warmRuntime} for why it exists.\n */\nconst RUNTIME_PROBE_MODEL =\n    \"CAcSDm9ydC12aXNpb24tc2RrOjoKEAoBeBIBeSIISWRlbnRpdHkSBHdhcm1aDwoBeBIKCggIARIECgIIAWIPCgF5EgoKCAgBEgQKAggBQgQKABAN\";\n\n/** Runtime start-ups already begun on this page, keyed by provider names. */\nconst runtimeWarmups = new Map<string, Promise<void>>();\n\n/**\n * Start ONNX Runtime's backends for a provider list, once per page.\n *\n * ORT downloads and compiles its WebAssembly binary — 12.8 MB for the plain\n * WASM build, 25.9 MB for the one carrying WebGPU — inside the first\n * `InferenceSession.create`. {@link OrtSession.create} only reaches that call\n * after it has downloaded the model, so on a real network the two largest\n * downloads of a page load ran one after the other. Creating and releasing a\n * session on a tiny model while the real one downloads lets them overlap; the\n * real `create` then finds the runtime ready.\n *\n * A failure here is swallowed: the real `create` runs against the same runtime\n * and reports the same problem with the right model in the message.\n *\n * @param specs Providers the real session will use.\n * @returns Settles when the runtime is up, or when starting it failed.\n */\nfunction warmRuntime(specs: readonly ProviderSpec[]): Promise<void> {\n    const key = specs.map(providerName).join(\",\");\n    let pending = runtimeWarmups.get(key);\n    if (pending === undefined) {\n        pending = (async () => {\n            try {\n                const probe = Uint8Array.from(atob(RUNTIME_PROBE_MODEL), (c) => c.charCodeAt(0));\n                const session = await ortRuntime.InferenceSession.create(probe, {\n                    executionProviders:\n                        specs as ort.InferenceSession.SessionOptions[\"executionProviders\"],\n                });\n                await session.release();\n            } catch {\n                return;\n            }\n        })();\n        runtimeWarmups.set(key, pending);\n    }\n    return pending;\n}\n\n/**\n * Warn that a model optimized offline is about to run on WebGPU.\n *\n * `optimize_model` fuses operators for the CPU, which is what the WASM backend\n * runs. On WebGPU some of those fused nodes have no GPU kernel: on a\n * YOLO11n-seg, ONNX Runtime placed 2 nodes on the CPU for the optimized file\n * and none for the original, so every inference pays GPU↔CPU copies. The SDK\n * cannot swap the file, so it says which one to ship.\n */\nfunction warnPreOptimizedOnWebGpu(): void {\n    console.warn(\n        \"[@ort-vision-sdk/web] This model was pre-optimized by optimize_model, which targets the \" +\n            \"WASM backend; on WebGPU some of its fused nodes fall back to the CPU. For WebGPU, ship \" +\n            \"the original export instead.\",\n    );\n}\n\n/**\n * Metadata key the Python SDK's `quantize_model` writes on an INT8 model.\n *\n * Mirrored by `QUANTIZATION_KEY` in the Python SDK; both test suites pin the\n * string. See {@link keepQuantizedOffWebGpu} for what a session does with it.\n */\nexport const QUANTIZATION_KEY = \"ort_vision_sdk.quantization\";\n\n/**\n * Drop `webgpu` from the providers of a model `quantize_model` produced.\n *\n * ONNX Runtime Web's WebGPU `DequantizeLinear` rejects the quantized bias such\n * a model carries — `scale and zero-point inputs must have the same rank` —\n * so the first `predict()` fails; and a variant with the bias left in float\n * ran but returned different detections than WASM for the same file. The\n * model is fine on WASM, which is also where INT8 pays (52.7 ms vs 68.4 ms on\n * a YOLO11n-seg), so the session runs there and says so.\n *\n * @param specs Providers that survived detection.\n * @param metadata The model's metadata map.\n * @returns `specs` without `webgpu` when the model is marked quantized; the\n *   WASM fallback when nothing else is left.\n */\nfunction keepQuantizedOffWebGpu(\n    specs: readonly ProviderSpec[],\n    metadata: Readonly<Record<string, string>>,\n): ProviderSpec[] {\n    if (metadata[QUANTIZATION_KEY] === undefined) return [...specs];\n    const kept = specs.filter((spec) => providerName(spec) !== \"webgpu\");\n    if (kept.length === specs.length) return kept;\n    console.warn(\n        `[@ort-vision-sdk/web] This model was quantized by quantize_model (${metadata[QUANTIZATION_KEY]}); ` +\n            \"ONNX Runtime Web's WebGPU backend cannot run its quantized operators, so it runs on WASM.\",\n    );\n    return kept.length > 0 ? kept : [FALLBACK_PROVIDER];\n}\n\n/** Cache Storage bucket used when {@link OrtSessionOptions.cache} is `true`. */\nexport const DEFAULT_MODEL_CACHE = \"ort-vision-sdk-models\";\n\n/**\n * Fetch a model URL as bytes so its metadata can be read.\n *\n * Falls back to the URL itself when the fetch fails, letting ORT try its own\n * load path: losing the metadata map is a downgrade, but failing to load a model\n * that ORT could have fetched would be a regression.\n *\n * @param url Where the `.onnx` lives.\n * @param cacheName Cache Storage bucket to read from and fill, or `null` to\n *   always go to the network.\n * @returns The model bytes, or the original URL when they could not be fetched.\n */\nasync function fetchModel(url: string, cacheName: string | null): Promise<Uint8Array | string> {\n    const cache = cacheName === null ? null : await openCache(cacheName);\n    if (cache !== null) {\n        const hit = await cache.match(url).catch(() => undefined);\n        if (hit !== undefined) return new Uint8Array(await hit.arrayBuffer());\n    }\n    try {\n        const response = await fetch(url);\n        if (!response.ok) {\n            warnMetadataUnavailable(url, `HTTP ${response.status} ${response.statusText}`);\n            return url;\n        }\n        const buffer = await response.arrayBuffer();\n        if (cache !== null) await storeModel(cache, url, buffer);\n        return new Uint8Array(buffer);\n    } catch (err) {\n        warnMetadataUnavailable(url, (err as Error).message);\n        return url;\n    }\n}\n\n/**\n * Put a model's bytes into the cache, never failing the load over it.\n *\n * Stores a fresh `Response` built from bytes already read rather than a\n * `clone()` of the network response: one read, and nothing that can throw\n * synchronously outside the guard. Awaited on purpose — a `put` left running\n * would keep its copy of the model reachable while ORT builds the session, the\n * same doubled peak {@link OrtSession.create} orders its reads to avoid.\n *\n * @param cache The open bucket.\n * @param url Cache key.\n * @param buffer The model.\n */\nasync function storeModel(cache: Cache, url: string, buffer: ArrayBuffer): Promise<void> {\n    try {\n        await cache.put(url, new Response(buffer));\n    } catch {\n        return;\n    }\n}\n\n/**\n * Open a Cache Storage bucket, or report that there is none to open.\n *\n * Cache Storage only exists in secure contexts (HTTPS, `localhost`), and a\n * browser may refuse it outright — private windows, blocked site data, an\n * exhausted quota. None of that is worth failing a model load over: the model\n * is fetched from the network as if caching had not been asked for. A failed\n * `put` is swallowed for the same reason — see {@link storeModel}.\n *\n * @param name Bucket name.\n * @returns The cache, or `null` when this environment cannot provide one.\n */\nasync function openCache(name: string): Promise<Cache | null> {\n    if (typeof caches === \"undefined\") return null;\n    try {\n        return await caches.open(name);\n    } catch {\n        return null;\n    }\n}\n\n/**\n * Warn that a model's metadata could not be read, and say what that costs.\n *\n * The fallback itself is right — losing the metadata beats failing a load that\n * ORT could have completed on its own — but it used to be silent, and the\n * symptom it produces is remote from the cause: class names come back as\n * `class_0`, `class_1`, ... with nothing anywhere explaining why. Whoever hits\n * this needs to be told that passing `labels` is the way out.\n *\n * @param url The model URL that could not be fetched here.\n * @param reason What went wrong, as reported by `fetch`.\n */\nfunction warnMetadataUnavailable(url: string, reason: string): void {\n    console.warn(\n        `[@ort-vision-sdk/web] Could not fetch ${url} to read its metadata (${reason}). ` +\n            \"Letting ONNX Runtime load it instead: the model will work, but its baked-in \" +\n            \"class names are unavailable, so labels fall back to class_0, class_1, ... \" +\n            \"Pass `labels` explicitly to name them.\",\n    );\n}\n\nexport interface OrtSessionOptions {\n    /**\n     * Execution providers in preference order. `undefined` uses {@link DEFAULT_PROVIDERS}.\n     *\n     * Each entry is a name (`\"webgpu\"`, `\"wasm\"`) or ORT's config object for that\n     * provider, which passes provider options through untouched —\n     * `{ name: \"webgpu\", preferredLayout: \"NHWC\" }`.\n     *\n     * Naming one explicitly also opts into a `console.warn` when this browser\n     * cannot offer it, instead of falling back in silence.\n     */\n    readonly providers?: readonly ProviderSpec[];\n    /**\n     * Keep a URL model in the browser's Cache Storage between page loads.\n     *\n     * `true` uses the {@link DEFAULT_MODEL_CACHE} bucket; a string names the\n     * bucket. A later `create` with the same URL reads the bytes from the cache\n     * instead of the network — for a model of several megabytes, the bulk of\n     * what a returning visitor waits for. Defaults to `false`.\n     *\n     * The URL is the cache key and nothing expires it: publish a changed model\n     * under a new URL (`yolo.v2.onnx`, `?v=2`) or a new bucket name, or delete\n     * the bucket with `caches.delete(name)`. Ignored for models passed as bytes,\n     * and silently skipped where Cache Storage is unavailable (non-secure\n     * contexts, blocked site data).\n     */\n    readonly cache?: boolean | string;\n    /** Optional ORT session options forwarded to `InferenceSession.create`. */\n    readonly sessionOptions?: ort.InferenceSession.SessionOptions;\n    /**\n     * Whether to read the model's custom metadata map (`names`, `task`, `imgsz`).\n     * Defaults to `true`.\n     *\n     * The runtime does not expose that map, so it is read from the file itself —\n     * which means a URL model is fetched here and handed to ORT as bytes instead\n     * of letting ORT fetch it. That is the same single download either way, and\n     * it is what lets a task resolve its labels off the model. Set to `false` to\n     * keep the URL path untouched (unless {@link cache} is set, which needs the\n     * bytes) and leave {@link OrtSession.metadata} empty.\n     *\n     * `false` is also the escape hatch when a device cannot afford the bytes: the\n     * fetched buffer is dropped before ORT builds the graph (see\n     * {@link OrtSession.create}), but ORT's own load path still keeps the model out\n     * of reach of anything the SDK holds. A session built this way resolves its\n     * input size from the graph as usual — only the class names are lost, so a\n     * caller taking this route has to pass `labels` itself.\n     */\n    readonly readMetadata?: boolean;\n}\n\n/**\n * Warn when a provider named explicitly by the caller is not going to run.\n *\n * Only explicit requests are worth a warning. The default list exists precisely\n * so that falling from `webgpu` to `wasm` is the expected outcome — but a caller\n * who wrote `providers: [\"webgpu\"]` and lands on WASM has a page several times\n * slower than intended and nothing in the console to explain it.\n *\n * @param requested Providers that were asked for.\n * @param effective Providers that survived capability detection.\n */\nfunction warnOnDroppedProviders(requested: readonly string[], effective: readonly string[]): void {\n    const dropped = requested.filter((provider) => !effective.includes(provider));\n    if (dropped.length === 0) {\n        return;\n    }\n    console.warn(\n        `This browser cannot offer the requested execution provider(s) ${JSON.stringify(dropped)}; ` +\n            `the session will run on ${JSON.stringify(effective)}. Inference still produces correct ` +\n            \"results, on the fallback provider.\",\n    );\n}\n\n/**\n * Name the tensor type of every graph input, straight from the model bytes.\n *\n * @param model The `.onnx` file contents.\n * @returns Input name → tensor type name. Empty when the file carries no\n *   readable graph, which callers read as \"assume float32\".\n */\nfunction declaredInputTypes(model: Uint8Array | ArrayBufferLike): Readonly<Record<string, string>> {\n    const declared = readModelInputTypes(model);\n    const named: Record<string, string> = {};\n    for (const [name, elemType] of Object.entries(declared)) {\n        named[name] = tensorTypeFor(elemType);\n    }\n    return named;\n}\n\n/**\n * Warn when the model's bytes were in hand but no input type came back.\n *\n * Every ONNX graph declares at least one input, so an empty result from bytes\n * means the reader failed, not that the model said nothing. Without this the\n * failure is invisible: the session falls back to float32, which is right for\n * most models and wrong for exactly the half-precision ones this reader exists\n * to support — so it looks like nothing is broken until a `predict()` throws\n * `Unexpected input data type`. That is how a 1 MB ceiling on the graph\n * descent shipped in 0.9.0 and reached a consumer.\n */\nfunction warnOnUnreadableTypes(): void {\n    console.warn(\n        \"Could not read any input type from this model's bytes; feeds will be built as float32. \" +\n            \"A half-precision model will fail at inference with 'Unexpected input data type'. \" +\n            \"Please report this with the model that produced it.\",\n    );\n}\n\n/**\n * Refuse a half-precision model in an environment that cannot feed one.\n *\n * ORT requires a real `Float16Array` for a `float16` tensor — the same bits in\n * a `Uint16Array` are rejected — and not every browser has one yet. Without\n * this check the session builds happily and the first `predict()` throws from\n * inside the preprocessing, which is both later and further from the cause.\n *\n * @param inputTypes Tensor type per input name.\n * @throws {@link ModelLoadError} when the graph wants half precision and the\n *   runtime has no `Float16Array`.\n */\nfunction assertFeedableTypes(inputTypes: Readonly<Record<string, string>>): void {\n    const half = Object.entries(inputTypes).filter(([, type]) => type === \"float16\");\n    if (half.length === 0 || hasFloat16Array()) return;\n    const names = half.map(([name]) => name).join(\", \");\n    throw new ModelLoadError(\n        `This model declares half-precision input(s) [${names}], but this environment has no ` +\n            \"Float16Array, which ONNX Runtime requires for a float16 tensor. Use a float32 export \" +\n            \"of the model, or run in a browser that supports Float16Array.\",\n    );\n}\n\n/**\n * Wrap an ONNX Runtime Web `InferenceSession` with convenient metadata access.\n *\n * The wrapper exposes input/output names and the shapes the graph declares,\n * manages execution-provider selection, provides a typed {@link OrtSession.run}\n * method, and releases the native session through {@link OrtSession.release}.\n */\nexport class OrtSession {\n    private constructor(\n        private readonly _session: ort.InferenceSession,\n        /**\n         * Execution providers this session is expected to run on.\n         *\n         * The requested list narrowed to what this browser can actually offer — a\n         * `webgpu` entry survives only where an adapter exists. Best-effort: ORT-Web\n         * exposes no way to ask which provider a session ended up on, so an entry\n         * here means \"not ruled out\", not \"confirmed\". See\n         * {@link requestedProviders} for what was asked for.\n         *\n         * When nothing survives — a caller asking for `webgpu` alone on a device\n         * without it — this falls back to {@link FALLBACK_PROVIDER}, which ORT-Web\n         * can always run. Handing ORT the unsatisfiable list instead makes\n         * `InferenceSession.create` reject with \"no available backend found\", so the\n         * page gets no inference at all rather than the slow-but-working fallback\n         * the `console.warn` describes. Measured in a real Chromium, where\n         * `navigator.gpu` exists but yields no adapter.\n         */\n        public readonly providers: readonly string[],\n        private readonly _metadata: Readonly<Record<string, string>>,\n        /**\n         * Execution providers that were asked for, after defaults were applied.\n         *\n         * Kept separate because ORT-Web falls back silently: a page that asks for\n         * `webgpu` on a device without it runs on WASM and is told nothing.\n         */\n        public readonly requestedProviders: readonly string[],\n        /**\n         * Tensor type each input declares, keyed by input name.\n         *\n         * Read from the model file, not from the session: `inputMetadata` is\n         * `undefined` on `onnxruntime-web` 1.20.1. Empty when the bytes were never\n         * in hand — a URL loaded with `readMetadata: false` — in which case every\n         * feed is built as float32, the behaviour before this existed.\n         */\n        private readonly _inputTypes: Readonly<Record<string, string>> = {},\n    ) {}\n\n    /**\n     * Tail of the run queue: settles when the latest queued run has finished.\n     *\n     * See {@link run} for why runs are serialized.\n     */\n    private _runQueue: Promise<unknown> = Promise.resolve();\n\n    /**\n     * Load an ONNX model into an ORT inference session.\n     *\n     * The metadata map is read **before** the session is built, and that order is\n     * load-bearing on memory-constrained devices. ORT copies the model into its\n     * WASM heap and then allocates the graph and the weights on top of that copy;\n     * a `readModelMetadata` call placed after `InferenceSession.create` keeps the\n     * JavaScript-side buffer reachable across the whole build, so a 5 MB model\n     * costs 5 MB of JS heap plus 5 MB of WASM heap plus the weights at the same\n     * instant. Reading first makes the buffer collectable as soon as ORT has copied\n     * it — on a phone that was the difference between a session and\n     * `Can't create a session. failed to allocate a buffer of size N`.\n     *\n     * The provider probe and the model download run concurrently: neither\n     * depends on the other, and on a real GPU `requestAdapter()` is not free.\n     * When the model is downloaded here, ORT's runtime is started alongside the\n     * download too — see {@link warmRuntime}.\n     *\n     * A model whose metadata carries {@link GRAPH_OPTIMIZATION_KEY} was already\n     * optimized offline by the Python SDK's `optimize_model`, so it is loaded\n     * with `graphOptimizationLevel: \"disabled\"` instead of paying the optimizer a\n     * second time — measured on a YOLO11n-seg under WASM, 29 ms of session\n     * creation down to 12 ms with unchanged inference time. An explicit\n     * `graphOptimizationLevel` in `sessionOptions` always wins. The mark is read\n     * with the rest of the metadata, so `readMetadata: false` skips it too.\n     *\n     * @param model Either a URL string, or a `Uint8Array`/`ArrayBuffer` containing the model bytes.\n     * @param options Provider list, pass-through `SessionOptions`, and whether to\n     *   read the model's metadata map (see {@link OrtSessionOptions.readMetadata}).\n     * @throws {@link ModelLoadError} if the model cannot be loaded.\n     */\n    static async create(model: ModelSource, options: OrtSessionOptions = {}): Promise<OrtSession> {\n        const requested = resolveProviders(options.providers);\n        const wantsMetadata = options.readMetadata !== false;\n        const cacheName =\n            options.cache === true ? DEFAULT_MODEL_CACHE : options.cache ? options.cache : null;\n        const shouldFetch = typeof model === \"string\" && (wantsMetadata || cacheName !== null);\n        const detection = detectProviders(requested).then((detected) =>\n            detected.length > 0 ? detected : [FALLBACK_PROVIDER],\n        );\n        const [detected, source] = await Promise.all([\n            detection,\n            shouldFetch ? fetchModel(model, cacheName) : Promise.resolve(model),\n            shouldFetch ? detection.then(warmRuntime) : Promise.resolve(),\n        ]);\n        const metadata =\n            wantsMetadata && typeof source !== \"string\" ? readModelMetadata(source) : {};\n        const specs = keepQuantizedOffWebGpu(detected, metadata);\n        const requestedNames = requested.map(providerName);\n        const providers = specs.map(providerName);\n        if (options.providers !== undefined && options.providers.length > 0) {\n            warnOnDroppedProviders(requestedNames, detected.map(providerName));\n        }\n        if (metadata[GRAPH_OPTIMIZATION_KEY] !== undefined && providers.includes(\"webgpu\")) {\n            warnPreOptimizedOnWebGpu();\n        }\n        const preOptimized =\n            metadata[GRAPH_OPTIMIZATION_KEY] !== undefined &&\n            options.sessionOptions?.graphOptimizationLevel === undefined;\n        const sessionOptions: ort.InferenceSession.SessionOptions = {\n            ...(preOptimized ? { graphOptimizationLevel: \"disabled\" as const } : {}),\n            ...(options.sessionOptions ?? {}),\n            executionProviders: specs as ort.InferenceSession.SessionOptions[\"executionProviders\"],\n        };\n        const inputTypes = typeof source !== \"string\" ? declaredInputTypes(source) : {};\n        if (typeof source !== \"string\" && Object.keys(inputTypes).length === 0) {\n            warnOnUnreadableTypes();\n        }\n        assertFeedableTypes(inputTypes);\n\n        let session: ort.InferenceSession;\n        try {\n            if (typeof source === \"string\") {\n                session = await ortRuntime.InferenceSession.create(source, sessionOptions);\n            } else if (source instanceof Uint8Array) {\n                session = await ortRuntime.InferenceSession.create(source, sessionOptions);\n            } else {\n                session = await ortRuntime.InferenceSession.create(\n                    source as ArrayBuffer,\n                    sessionOptions,\n                );\n            }\n        } catch (err) {\n            throw new ModelLoadError(`Failed to load ONNX model: ${(err as Error).message}`, {\n                cause: err,\n            });\n        }\n\n        return new OrtSession(session, providers, metadata, requestedNames, inputTypes);\n    }\n\n    /**\n     * Tensor type each input declares, in declaration order.\n     *\n     * `\"float32\"` for a normal export, `\"float16\"` for one exported with\n     * `half=True`. ORT matches a feed's type against this exactly, so the tasks\n     * read it and convert at the feed boundary.\n     */\n    get inputDtypes(): readonly string[] {\n        return this.inputNames.map((name) => this._inputTypes[name] ?? DEFAULT_TENSOR_TYPE);\n    }\n\n    /** Tensor type the first input declares. */\n    get inputDtype(): string {\n        return this._inputTypes[this.inputName] ?? DEFAULT_TENSOR_TYPE;\n    }\n\n    /** Names of the model's inputs, in declaration order. */\n    get inputNames(): readonly string[] {\n        return this._session.inputNames;\n    }\n\n    /** Name of the first (and usually only) input. */\n    get inputName(): string {\n        const name = this._session.inputNames[0];\n        if (name === undefined) {\n            throw new InferenceError(\"Model has no inputs.\");\n        }\n        return name;\n    }\n\n    /** Names of the model's outputs, in declaration order. */\n    get outputNames(): readonly string[] {\n        return this._session.outputNames;\n    }\n\n    /**\n     * Shapes the graph declares for its inputs, in declaration order.\n     *\n     * Dynamic (symbolic) axes appear as `null`. Empty shapes mean the runtime\n     * reported no metadata — either a non-tensor input, or an `onnxruntime-web`\n     * older than 1.21, which predates input metadata.\n     */\n    get inputShapes(): readonly DeclaredShape[] {\n        return declaredShapesFrom(\n            this._session.inputMetadata as\n                readonly ort.InferenceSession.ValueMetadata[] | undefined,\n        );\n    }\n\n    /**\n     * Shape the graph declares for its first input, dynamic axes as `null`.\n     *\n     * Empty when the runtime reports no metadata for it.\n     */\n    get inputShape(): DeclaredShape {\n        return this.inputShapes[0] ?? [];\n    }\n\n    /**\n     * Shapes the graph declares for its outputs, in declaration order.\n     *\n     * Dynamic (symbolic) axes appear as `null`. Reading them is how a task can\n     * tell how many classes a head emits without being told.\n     */\n    get outputShapes(): readonly DeclaredShape[] {\n        return declaredShapesFrom(\n            this._session.outputMetadata as\n                readonly ort.InferenceSession.ValueMetadata[] | undefined,\n        );\n    }\n\n    /**\n     * Shape the graph declares for its first output, dynamic axes as `null`.\n     *\n     * Empty when the runtime reports no metadata for it.\n     */\n    get outputShape(): DeclaredShape {\n        return this.outputShapes[0] ?? [];\n    }\n\n    /**\n     * The model's custom metadata map — `names`, `task`, `imgsz`, ... for an\n     * Ultralytics export.\n     *\n     * Read from the model's bytes at load time, since the runtime does not expose\n     * it. Empty when the session was created with `readMetadata: false`, from a\n     * URL that could not be fetched here, or from a model carrying no metadata.\n     */\n    get metadata(): Readonly<Record<string, string>> {\n        return this._metadata;\n    }\n\n    /**\n     * Release the native session and free its memory.\n     *\n     * Call it when a session is discarded while the page lives on — rebuilding a\n     * task at a different input size, swapping in a newer model. A failure from\n     * the runtime is ignored: a session being torn down has nothing left to fail\n     * at, and the caller is already moving on.\n     */\n    async release(): Promise<void> {\n        await this._session.release().catch(() => undefined);\n    }\n\n    /** The underlying `onnxruntime-web` session, for advanced use cases. */\n    get raw(): ort.InferenceSession {\n        return this._session;\n    }\n\n    /**\n     * Run inference and return all outputs.\n     *\n     * Runs on one session are serialized. ONNX Runtime Web refuses a second\n     * `run` while one is in flight — it rejects with `Session already started` —\n     * so two overlapping `predict()` calls on the same task used to fail. Queued\n     * here, they overlap where they can: while one run executes (in a worker,\n     * with `env.wasm.proxy`, or on the GPU), the next `predict()` decodes and\n     * preprocesses its frame and only waits for its turn at the runtime. A run\n     * that fails does not block the ones behind it.\n     *\n     * @param feeds Map of input name to `ort.Tensor`. Keys must match {@link inputNames}.\n     * @throws {@link InferenceError} if ORT raises any error during execution.\n     */\n    run(feeds: Record<string, ort.Tensor>): Promise<Record<string, ort.Tensor>> {\n        const turn = this._runQueue.then(() => this._runNow(feeds));\n        this._runQueue = turn.catch(() => undefined);\n        return turn;\n    }\n\n    /**\n     * Execute one run immediately; only {@link run}'s queue calls this.\n     *\n     * @param feeds Map of input name to `ort.Tensor`.\n     * @throws {@link InferenceError} if ORT raises any error during execution.\n     */\n    private async _runNow(feeds: Record<string, ort.Tensor>): Promise<Record<string, ort.Tensor>> {\n        try {\n            const result = await this._session.run(this._asDeclaredTypes(feeds));\n            return result as Record<string, ort.Tensor>;\n        } catch (err) {\n            throw new InferenceError(`Inference failed: ${(err as Error).message}`, { cause: err });\n        }\n    }\n\n    /**\n     * Rebuild any feed whose type does not match what its input declares.\n     *\n     * ORT matches feed types against the graph exactly: a float32 tensor against\n     * a `half=True` export fails the run with `Unexpected input data type`. Every\n     * preprocessing path in this SDK produces `Float32Array`, on purpose — the\n     * normalization arithmetic belongs in single precision — so the conversion\n     * happens here, once, at the only boundary all of them pass through. A feed\n     * that already carries the declared type is handed on untouched, so a caller\n     * building its own half tensor pays nothing.\n     *\n     * @param feeds The tensors to run with.\n     * @returns The same mapping, with mismatched float32 feeds converted.\n     */\n    private _asDeclaredTypes(feeds: Record<string, ort.Tensor>): Record<string, ort.Tensor> {\n        let converted: Record<string, ort.Tensor> | null = null;\n        for (const [name, tensor] of Object.entries(feeds)) {\n            const declared = this._inputTypes[name];\n            if (declared === undefined || declared === tensor.type) continue;\n            if (!(tensor.data instanceof Float32Array)) continue;\n            const data = toFeedData(tensor.data, declared);\n            if (data === tensor.data) continue;\n            converted ??= { ...feeds };\n            converted[name] = new ortRuntime.Tensor(\n                declared as \"float16\",\n                data as unknown as Float32Array,\n                tensor.dims as number[],\n            );\n        }\n        return converted ?? feeds;\n    }\n}\n"],"mappings":";;;;;;;AA8BA,IAAa,IAAyB,qCAQhC,IACF,oHAGE,oBAAiB,IAAI,IAA2B;AAmBtD,SAAS,EAAY,GAA+C;CAChE,IAAM,IAAM,EAAM,IAAI,CAAY,CAAC,CAAC,KAAK,GAAG,GACxC,IAAU,EAAe,IAAI,CAAG;CAgBpC,OAfI,MAAY,KAAA,MACZ,KAAW,YAAY;EACnB,IAAI;GACA,IAAM,IAAQ,WAAW,KAAK,KAAK,CAAmB,IAAI,MAAM,EAAE,WAAW,CAAC,CAAC;GAK/E,OAAM,MAJgB,EAAW,iBAAiB,OAAO,GAAO,EAC5D,oBACI,EACR,CAAC,EAAA,CACa,QAAQ;EAC1B,QAAQ;GACJ;EACJ;CACJ,EAAA,CAAG,GACH,EAAe,IAAI,GAAK,CAAO,IAE5B;AACX;AAWA,SAAS,IAAiC;CACtC,QAAQ,KACJ,6MAGJ;AACJ;AAQA,IAAa,IAAmB;AAiBhC,SAAS,EACL,GACA,GACc;CACd,IAAI,EAAA,mCAA+B,KAAA,GAAW,OAAO,CAAC,GAAG,CAAK;CAC9D,IAAM,IAAO,EAAM,QAAQ,MAAS,EAAa,CAAI,MAAM,QAAQ;CAMnE,OALI,EAAK,WAAW,EAAM,SAAe,KACzC,QAAQ,KACJ,qEAAqE,EAAS,GAAkB,6FAEpG,GACO,EAAK,SAAS,IAAI,IAAO,CAAC,CAAiB;AACtD;AAGA,IAAa,IAAsB;AAcnC,eAAe,EAAW,GAAa,GAAwD;CAC3F,IAAM,IAAQ,MAAc,OAAO,OAAO,MAAM,EAAU,CAAS;CACnE,IAAI,MAAU,MAAM;EAChB,IAAM,IAAM,MAAM,EAAM,MAAM,CAAG,CAAC,CAAC,YAAY,KAAA,CAAS;EACxD,IAAI,MAAQ,KAAA,GAAW,OAAO,IAAI,WAAW,MAAM,EAAI,YAAY,CAAC;CACxE;CACA,IAAI;EACA,IAAM,IAAW,MAAM,MAAM,CAAG;EAChC,IAAI,CAAC,EAAS,IAEV,OADA,EAAwB,GAAK,QAAQ,EAAS,OAAO,GAAG,EAAS,YAAY,GACtE;EAEX,IAAM,IAAS,MAAM,EAAS,YAAY;EAE1C,OADI,MAAU,QAAM,MAAM,EAAW,GAAO,GAAK,CAAM,GAChD,IAAI,WAAW,CAAM;CAChC,SAAS,GAAK;EAEV,OADA,EAAwB,GAAM,EAAc,OAAO,GAC5C;CACX;AACJ;AAeA,eAAe,EAAW,GAAc,GAAa,GAAoC;CACrF,IAAI;EACA,MAAM,EAAM,IAAI,GAAK,IAAI,SAAS,CAAM,CAAC;CAC7C,QAAQ;EACJ;CACJ;AACJ;AAcA,eAAe,EAAU,GAAqC;CAC1D,IAAI,OAAO,SAAW,KAAa,OAAO;CAC1C,IAAI;EACA,OAAO,MAAM,OAAO,KAAK,CAAI;CACjC,QAAQ;EACJ,OAAO;CACX;AACJ;AAcA,SAAS,EAAwB,GAAa,GAAsB;CAChE,QAAQ,KACJ,yCAAyC,EAAI,yBAAyB,EAAO,kMAIjF;AACJ;AA+DA,SAAS,EAAuB,GAA8B,GAAoC;CAC9F,IAAM,IAAU,EAAU,QAAQ,MAAa,CAAC,EAAU,SAAS,CAAQ,CAAC;CACxE,EAAQ,WAAW,KAGvB,QAAQ,KACJ,iEAAiE,KAAK,UAAU,CAAO,EAAE,4BAC1D,KAAK,UAAU,CAAS,EAAE,sEAE7D;AACJ;AASA,SAAS,EAAmB,GAAuE;CAC/F,IAAM,IAAW,EAAoB,CAAK,GACpC,IAAgC,CAAC;CACvC,KAAK,IAAM,CAAC,GAAM,MAAa,OAAO,QAAQ,CAAQ,GAClD,EAAM,KAAQ,EAAc,CAAQ;CAExC,OAAO;AACX;AAaA,SAAS,IAA8B;CACnC,QAAQ,KACJ,6NAGJ;AACJ;AAcA,SAAS,EAAoB,GAAoD;CAC7E,IAAM,IAAO,OAAO,QAAQ,CAAU,CAAC,CAAC,QAAQ,GAAG,OAAU,MAAS,SAAS;CAC/E,IAAI,EAAK,WAAW,KAAK,EAAgB,GAAG;CAC5C,IAAM,IAAQ,EAAK,KAAK,CAAC,OAAU,CAAI,CAAC,CAAC,KAAK,IAAI;CAClD,MAAM,IAAI,EACN,gDAAgD,EAAM,kLAG1D;AACJ;AASA,IAAa,IAAb,MAAa,EAAW;CAEC;CAkBD;CACC;CAOD;CASC;CApCrB,YACI,GAkBA,GACA,GAOA,GASA,IAAiE,CAAC,GACpE;EADmB,AAnCA,KAAA,WAAA,GAkBD,KAAA,YAAA,GACC,KAAA,YAAA,GAOD,KAAA,qBAAA,GASC,KAAA,cAAA;CAClB;CAOH,YAAsC,QAAQ,QAAQ;CAiCtD,aAAa,OAAO,GAAoB,IAA6B,CAAC,GAAwB;EAC1F,IAAM,IAAY,EAAiB,EAAQ,SAAS,GAC9C,IAAgB,EAAQ,iBAAiB,IACzC,IACF,EAAQ,UAAU,KAAO,IAAsB,EAAQ,QAAQ,EAAQ,QAAQ,MAC7E,IAAc,OAAO,KAAU,aAAa,KAAiB,MAAc,OAC3E,IAAY,EAAgB,CAAS,CAAC,CAAC,MAAM,MAC/C,EAAS,SAAS,IAAI,IAAW,CAAC,CAAiB,CACvD,GACM,CAAC,GAAU,KAAU,MAAM,QAAQ,IAAI;GACzC;GACA,IAAc,EAAW,GAAO,CAAS,IAAI,QAAQ,QAAQ,CAAK;GAClE,IAAc,EAAU,KAAK,CAAW,IAAI,QAAQ,QAAQ;EAChE,CAAC,GACK,IACF,KAAiB,OAAO,KAAW,WAAW,EAAkB,CAAM,IAAI,CAAC,GACzE,IAAQ,EAAuB,GAAU,CAAQ,GACjD,IAAiB,EAAU,IAAI,CAAY,GAC3C,IAAY,EAAM,IAAI,CAAY;EAIxC,AAHI,EAAQ,cAAc,KAAA,KAAa,EAAQ,UAAU,SAAS,KAC9D,EAAuB,GAAgB,EAAS,IAAI,CAAY,CAAC,GAEjE,EAAA,yCAAqC,KAAA,KAAa,EAAU,SAAS,QAAQ,KAC7E,EAAyB;EAK7B,IAAM,IAAsD;GACxD,GAHA,EAAA,yCAAqC,KAAA,KACrC,EAAQ,gBAAgB,2BAA2B,KAAA,IAEhC,EAAE,wBAAwB,WAAoB,IAAI,CAAC;GACtE,GAAI,EAAQ,kBAAkB,CAAC;GAC/B,oBAAoB;EACxB,GACM,IAAa,OAAO,KAAW,WAAwC,CAAC,IAA9B,EAAmB,CAAM;EAIzE,AAHI,OAAO,KAAW,YAAY,OAAO,KAAK,CAAU,CAAC,CAAC,WAAW,KACjE,EAAsB,GAE1B,EAAoB,CAAU;EAE9B,IAAI;EACJ,IAAI;GACA,AAKI,KALA,OAAO,KAAW,YAEX,aAAkB,YADf,MAAM,EAAW,iBAAiB,OAAO,GAAQ,CAAc;EASjF,SAAS,GAAK;GACV,MAAM,IAAI,EAAe,8BAA+B,EAAc,WAAW,EAC7E,OAAO,EACX,CAAC;EACL;EAEA,OAAO,IAAI,EAAW,GAAS,GAAW,GAAU,GAAgB,CAAU;CAClF;CASA,IAAI,cAAiC;EACjC,OAAO,KAAK,WAAW,KAAK,MAAS,KAAK,YAAY,MAAA,SAA4B;CACtF;CAGA,IAAI,aAAqB;EACrB,OAAO,KAAK,YAAY,KAAK,cAAA;CACjC;CAGA,IAAI,aAAgC;EAChC,OAAO,KAAK,SAAS;CACzB;CAGA,IAAI,YAAoB;EACpB,IAAM,IAAO,KAAK,SAAS,WAAW;EACtC,IAAI,MAAS,KAAA,GACT,MAAM,IAAI,EAAe,sBAAsB;EAEnD,OAAO;CACX;CAGA,IAAI,cAAiC;EACjC,OAAO,KAAK,SAAS;CACzB;CASA,IAAI,cAAwC;EACxC,OAAO,EACH,KAAK,SAAS,aAElB;CACJ;CAOA,IAAI,aAA4B;EAC5B,OAAO,KAAK,YAAY,MAAM,CAAC;CACnC;CAQA,IAAI,eAAyC;EACzC,OAAO,EACH,KAAK,SAAS,cAElB;CACJ;CAOA,IAAI,cAA6B;EAC7B,OAAO,KAAK,aAAa,MAAM,CAAC;CACpC;CAUA,IAAI,WAA6C;EAC7C,OAAO,KAAK;CAChB;CAUA,MAAM,UAAyB;EAC3B,MAAM,KAAK,SAAS,QAAQ,CAAC,CAAC,YAAY,KAAA,CAAS;CACvD;CAGA,IAAI,MAA4B;EAC5B,OAAO,KAAK;CAChB;CAgBA,IAAI,GAAwE;EACxE,IAAM,IAAO,KAAK,UAAU,WAAW,KAAK,QAAQ,CAAK,CAAC;EAE1D,OADA,KAAK,YAAY,EAAK,YAAY,KAAA,CAAS,GACpC;CACX;CAQA,MAAc,QAAQ,GAAwE;EAC1F,IAAI;GAEA,OAAO,MADc,KAAK,SAAS,IAAI,KAAK,iBAAiB,CAAK,CAAC;EAEvE,SAAS,GAAK;GACV,MAAM,IAAI,EAAe,qBAAsB,EAAc,WAAW,EAAE,OAAO,EAAI,CAAC;EAC1F;CACJ;CAgBA,iBAAyB,GAA+D;EACpF,IAAI,IAA+C;EACnD,KAAK,IAAM,CAAC,GAAM,MAAW,OAAO,QAAQ,CAAK,GAAG;GAChD,IAAM,IAAW,KAAK,YAAY;GAElC,IADI,MAAa,KAAA,KAAa,MAAa,EAAO,QAC9C,EAAE,EAAO,gBAAgB,eAAe;GAC5C,IAAM,IAAO,EAAW,EAAO,MAAM,CAAQ;GACzC,MAAS,EAAO,SACpB,MAAc,EAAE,GAAG,EAAM,GACzB,EAAU,KAAQ,IAAI,EAAW,OAC7B,GACA,GACA,EAAO,IACX;EACJ;EACA,OAAO,KAAa;CACxB;AACJ"}

package/dist/vision/fusion.cjs

@@ -1,2 +1,2 @@
-const e=require("./core/metadata.cjs");var t=`detect_classify`,n=`ovs.`,r=`images`,i=`source_image`,a=`letterbox_scale`,o=`letterbox_pad`,s=`boxes`,c=`scores`,l=`classes`,u=`num_detections`,d=`probs`,f=`kind`,p=`sdk_version`,m=`input_size`,h=`crop_size`,g=`crop_source`,_=`max_detections`,v=`conf_threshold`,y=`iou_threshold`,b=`apply_softmax`,x=`detector_names`,S=`classifier_names`,C=`dynamic`;function w(e){if(!e)return null;let t=e.split(`,`);if(t.length!==2)return null;let n=Number(t[0]),r=Number(t[1]);return!Number.isInteger(n)||!Number.isInteger(r)||n<1||r<1?null:[n,r]}function T(e,t){if(!e)return t;let n=Number(e);return Number.isFinite(n)?n:t}function E(n){if(!n)return null;let r={};for(let[e,t]of Object.entries(n))e.startsWith(`ovs.`)&&(r[e.slice(4)]=t);if(r[f]!==`detect_classify`)return null;let i=w(r[m]),a=w(r[h]);if(i===null||a===null)return null;let o=r[_]??C,s=Number(o),c=o===C||!Number.isInteger(s)||s<1?null:s,l=r[g]===`original`?`original`:`detector_input`;return{kind:t,inputSize:i,cropSize:a,cropSource:l,maxDetections:c,confThreshold:T(r[v],.25),iouThreshold:T(r[y],.45),applySoftmax:(r[b]??`1`)!==`0`,detectorNames:e.parseNames(r[x]),classifierNames:e.parseNames(r[S]),sdkVersion:r[p]??``,needsSourceImage:l===`original`}}exports.FUSION_KIND_DETECT_CLASSIFY=t,exports.INPUT_IMAGE=r,exports.INPUT_PAD=o,exports.INPUT_SCALE=a,exports.INPUT_SOURCE=i,exports.METADATA_PREFIX=n,exports.OUTPUT_BOXES=s,exports.OUTPUT_CLASSES=l,exports.OUTPUT_NUM_DETECTIONS=u,exports.OUTPUT_PROBS=d,exports.OUTPUT_SCORES=c,exports.readFusionSpec=E;
+const e=require("./core/metadata.cjs");var t=`detect_classify`,n=`detect_segment_classify`,r=`ovs.`,i=`images`,a=`source_image`,o=`letterbox_scale`,s=`letterbox_pad`,c=`boxes`,l=`scores`,u=`classes`,d=`num_detections`,f=`masks`,p=`probs`,m=`kind`,h=`sdk_version`,g=`input_size`,_=`crop_size`,v=`crop_source`,y=`max_detections`,b=`conf_threshold`,x=`iou_threshold`,S=`apply_softmax`,C=`detector_names`,w=`classifier_names`,T=`segmenter_names`,E=`mask_threshold`,D=`mask_applied`,O=`dynamic`;function k(e){if(!e)return null;let t=e.split(`,`);if(t.length!==2)return null;let n=Number(t[0]),r=Number(t[1]);return!Number.isInteger(n)||!Number.isInteger(r)||n<1||r<1?null:[n,r]}function A(e,t){if(!e)return t;let n=Number(e);return Number.isFinite(n)?n:t}function j(t){if(!t)return null;let r={};for(let[e,n]of Object.entries(t))e.startsWith(`ovs.`)&&(r[e.slice(4)]=n);let i=r[m];if(i!==`detect_classify`&&i!==`detect_segment_classify`)return null;let a=k(r[g]),o=k(r[_]);if(a===null||o===null)return null;let s=r[y]??O,c=Number(s),l=s===O||!Number.isInteger(c)||c<1?null:c,u=r[v]===`original`?`original`:`detector_input`;return{kind:i,inputSize:a,cropSize:o,cropSource:u,maxDetections:l,confThreshold:A(r[b],.25),iouThreshold:A(r[x],.45),applySoftmax:(r[S]??`1`)!==`0`,detectorNames:e.parseNames(r[C]),classifierNames:e.parseNames(r[w]),sdkVersion:r[h]??``,needsSourceImage:u===`original`,hasMasks:i===n,segmenterNames:e.parseNames(r[T]),maskThreshold:A(r[E],.5),maskApplied:(r[D]??`1`)!==`0`}}exports.FUSION_KIND_DETECT_CLASSIFY=t,exports.FUSION_KIND_DETECT_SEGMENT_CLASSIFY=n,exports.INPUT_IMAGE=i,exports.INPUT_PAD=s,exports.INPUT_SCALE=o,exports.INPUT_SOURCE=a,exports.METADATA_PREFIX=r,exports.OUTPUT_BOXES=c,exports.OUTPUT_CLASSES=u,exports.OUTPUT_MASKS=f,exports.OUTPUT_NUM_DETECTIONS=d,exports.OUTPUT_PROBS=p,exports.OUTPUT_SCORES=l,exports.readFusionSpec=j;
 //# sourceMappingURL=fusion.cjs.map

package/dist/vision/fusion.cjs.map

@@ -1 +1 @@
-{"version":3,"file":"fusion.cjs","names":[],"sources":["../../src/vision/fusion.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * The contract a fused pipeline carries inside its own `.onnx` file.\n *\n * A pipeline built by the Python SDK's `ort_vision_sdk.compose` is a single\n * graph that already contains the detector, the crop-and-resize bridge and the\n * classifier. Everything the runtime needs to drive it — the letterbox\n * resolution, the crop resolution, where the crops come from, how many\n * detections the graph emits, the class names of both stages — was decided at\n * fusion time and written into the model's metadata.\n *\n * This module reads it back. It is the browser half of a contract whose other\n * half lives in `ort_vision_sdk/fusion.py`: the same keys, the same encodings,\n * the same fallbacks. A pipeline fused once therefore runs identically in\n * Python and in a browser tab, off the same file.\n *\n * Building a pipeline stays a Python-side build step — there is no ONNX\n * protobuf writer here, and there is no reason for one: fusing is something you\n * do once next to your export pipeline, not in a page load.\n */\n\nimport { parseNames } from \"./core/metadata\";\n\n/**\n * Which tensor the bridge crops the detected boxes out of.\n *\n * - `\"detector_input\"`: the letterboxed tensor already fed to the detector.\n *   The fused graph then has a **single** image input, but a small object is\n *   cropped out of its downscaled copy.\n * - `\"original\"`: a second, full-resolution image input. The bridge undoes the\n *   letterbox transform in-graph and crops at native resolution. Two tensors to\n *   feed, still one session and one model load.\n */\nexport type CropSource = \"detector_input\" | \"original\";\n\n/** Value of the `ovs.kind` metadata key for a detector→classifier pipeline. */\nexport const FUSION_KIND_DETECT_CLASSIFY = \"detect_classify\";\n\n/**\n * Namespace for every metadata key the fusion writes.\n *\n * Namespaced on purpose: the detector's own Ultralytics metadata (`names`,\n * `task`, `imgsz`) is carried over into the fused model, and an un-prefixed key\n * would either collide with it or be mistaken for it.\n */\nexport const METADATA_PREFIX = \"ovs.\";\n\n/** Name of the fused graph's letterboxed detector input, `[1, 3, H, W]` float32 in `[0, 1]`. */\nexport const INPUT_IMAGE = \"images\";\n\n/** Name of the full-resolution input. Present only when `cropSource === \"original\"`. */\nexport const INPUT_SOURCE = \"source_image\";\n\n/** Name of the `[1]` float32 letterbox scale factor. Only with `cropSource === \"original\"`. */\nexport const INPUT_SCALE = \"letterbox_scale\";\n\n/** Name of the `[2]` float32 `[padLeft, padTop]`. Only with `cropSource === \"original\"`. */\nexport const INPUT_PAD = \"letterbox_pad\";\n\n/**\n * Name of the `[K, 4]` float32 xyxy output, in **letterboxed** input pixels.\n *\n * A file fused by `ort-vision-sdk` 0.9.0 or later reports the box that was\n * actually classified: clamped to the image the crop came from, exactly as\n * RoiAlign received it. Older files report the raw box, so one that ran off the\n * frame draws a rectangle wider than the region the classifier saw.\n */\nexport const OUTPUT_BOXES = \"boxes\";\n\n/** Name of the `[K]` float32 detection-confidence output. */\nexport const OUTPUT_SCORES = \"scores\";\n\n/** Name of the `[K]` int64 detector-class output. */\nexport const OUTPUT_CLASSES = \"classes\";\n\n/** Name of the `[1]` int64 output holding how many of the `K` rows are real. */\nexport const OUTPUT_NUM_DETECTIONS = \"num_detections\";\n\n/** Name of the `[K, numClassifierClasses]` float32 classifier output, one row per box. */\nexport const OUTPUT_PROBS = \"probs\";\n\n/** Everything a fused pipeline declares about how it must be driven. */\nexport interface FusionSpec {\n    /** Pipeline family. Only `\"detect_classify\"` exists today. */\n    readonly kind: string;\n    /** `[width, height]` the detector stage expects — the resolution to letterbox to. */\n    readonly inputSize: readonly [number, number];\n    /** `[width, height]` every crop is resampled to inside the graph. */\n    readonly cropSize: readonly [number, number];\n    /** Which tensor the crops are taken from. */\n    readonly cropSource: CropSource;\n    /**\n     * Fixed number of rows `K` every output carries, surplus zero-padded and\n     * counted by {@link OUTPUT_NUM_DETECTIONS}. `null` means the graph emits\n     * exactly as many rows as survived NMS.\n     */\n    readonly maxDetections: number | null;\n    /** Score threshold baked into the graph's NMS node. */\n    readonly confThreshold: number;\n    /** IoU threshold baked into the graph's NMS node. */\n    readonly iouThreshold: number;\n    /** Whether the classifier stage emits logits that still need a softmax. */\n    readonly applySoftmax: boolean;\n    /** Detector class names in class-id order, or `null` when the fusion recorded none. */\n    readonly detectorNames: readonly string[] | null;\n    /** Classifier class names in class-id order, or `null`. */\n    readonly classifierNames: readonly string[] | null;\n    /** Version of `ort-vision-sdk` that produced the file. */\n    readonly sdkVersion: string;\n    /** Whether driving this pipeline requires feeding the full-resolution input. */\n    readonly needsSourceImage: boolean;\n}\n\nconst KEY_KIND = \"kind\";\nconst KEY_SDK_VERSION = \"sdk_version\";\nconst KEY_INPUT_SIZE = \"input_size\";\nconst KEY_CROP_SIZE = \"crop_size\";\nconst KEY_CROP_SOURCE = \"crop_source\";\nconst KEY_MAX_DETECTIONS = \"max_detections\";\nconst KEY_CONF_THRESHOLD = \"conf_threshold\";\nconst KEY_IOU_THRESHOLD = \"iou_threshold\";\nconst KEY_APPLY_SOFTMAX = \"apply_softmax\";\nconst KEY_DETECTOR_NAMES = \"detector_names\";\nconst KEY_CLASSIFIER_NAMES = \"classifier_names\";\n\nconst DYNAMIC = \"dynamic\";\n\n/**\n * Decode a `\"640,640\"` pair.\n *\n * @param raw The encoded pair.\n * @returns `[width, height]`, or `null` when the value is missing or malformed.\n */\nfunction decodeSize(raw: string | undefined): readonly [number, number] | null {\n    if (!raw) return null;\n    const parts = raw.split(\",\");\n    if (parts.length !== 2) return null;\n    const width = Number(parts[0]);\n    const height = Number(parts[1]);\n    if (!Number.isInteger(width) || !Number.isInteger(height)) return null;\n    if (width < 1 || height < 1) return null;\n    return [width, height];\n}\n\n/**\n * Decode a float, falling back when the value is missing or malformed.\n *\n * @param raw The encoded value.\n * @param fallback Value to use when `raw` cannot be read.\n * @returns The parsed number, or `fallback`.\n */\nfunction decodeFloat(raw: string | undefined, fallback: number): number {\n    if (!raw) return fallback;\n    const value = Number(raw);\n    return Number.isFinite(value) ? value : fallback;\n}\n\n/**\n * Read a pipeline spec out of a model's custom metadata.\n *\n * Individual malformed entries fall back to the value a fusion would have used\n * by default — a single bad float is not a reason to reject an otherwise\n * loadable pipeline. A malformed resolution is fatal, because there is no safe\n * default for one.\n *\n * @param metadata A model's custom metadata map, as read by\n *   {@link readModelMetadata}.\n * @returns The decoded spec, or `null` when the model is not a fused pipeline —\n *   it carries no `ovs.kind` entry, or one naming a pipeline kind this version\n *   does not know how to drive.\n */\nexport function readFusionSpec(\n    metadata: Readonly<Record<string, string>> | undefined,\n): FusionSpec | null {\n    if (!metadata) return null;\n\n    const read: Record<string, string> = {};\n    for (const [key, value] of Object.entries(metadata)) {\n        if (key.startsWith(METADATA_PREFIX)) read[key.slice(METADATA_PREFIX.length)] = value;\n    }\n    if (read[KEY_KIND] !== FUSION_KIND_DETECT_CLASSIFY) return null;\n\n    const inputSize = decodeSize(read[KEY_INPUT_SIZE]);\n    const cropSize = decodeSize(read[KEY_CROP_SIZE]);\n    if (inputSize === null || cropSize === null) return null;\n\n    const rawMax = read[KEY_MAX_DETECTIONS] ?? DYNAMIC;\n    const parsedMax = Number(rawMax);\n    const maxDetections =\n        rawMax === DYNAMIC || !Number.isInteger(parsedMax) || parsedMax < 1 ? null : parsedMax;\n\n    const cropSource: CropSource =\n        read[KEY_CROP_SOURCE] === \"original\" ? \"original\" : \"detector_input\";\n\n    return {\n        kind: FUSION_KIND_DETECT_CLASSIFY,\n        inputSize,\n        cropSize,\n        cropSource,\n        maxDetections,\n        confThreshold: decodeFloat(read[KEY_CONF_THRESHOLD], 0.25),\n        iouThreshold: decodeFloat(read[KEY_IOU_THRESHOLD], 0.45),\n        applySoftmax: (read[KEY_APPLY_SOFTMAX] ?? \"1\") !== \"0\",\n        detectorNames: parseNames(read[KEY_DETECTOR_NAMES]),\n        classifierNames: parseNames(read[KEY_CLASSIFIER_NAMES]),\n        sdkVersion: read[KEY_SDK_VERSION] ?? \"\",\n        needsSourceImage: cropSource === \"original\",\n    };\n}\n"],"mappings":"uCAoCA,IAAa,EAA8B,kBAS9B,EAAkB,OAGlB,EAAc,SAGd,EAAe,eAGf,EAAc,kBAGd,EAAY,gBAUZ,EAAe,QAGf,EAAgB,SAGhB,EAAiB,UAGjB,EAAwB,iBAGxB,EAAe,QAkCtB,EAAW,OACX,EAAkB,cAClB,EAAiB,aACjB,EAAgB,YAChB,EAAkB,cAClB,EAAqB,iBACrB,EAAqB,iBACrB,EAAoB,gBACpB,EAAoB,gBACpB,EAAqB,iBACrB,EAAuB,mBAEvB,EAAU,UAQhB,SAAS,EAAW,EAA2D,CAC3E,GAAI,CAAC,EAAK,OAAO,KACjB,IAAM,EAAQ,EAAI,MAAM,GAAG,EAC3B,GAAI,EAAM,SAAW,EAAG,OAAO,KAC/B,IAAM,EAAQ,OAAO,EAAM,EAAE,EACvB,EAAS,OAAO,EAAM,EAAE,EAG9B,MAFI,CAAC,OAAO,UAAU,CAAK,GAAK,CAAC,OAAO,UAAU,CAAM,GACpD,EAAQ,GAAK,EAAS,EAAU,KAC7B,CAAC,EAAO,CAAM,CACzB,CASA,SAAS,EAAY,EAAyB,EAA0B,CACpE,GAAI,CAAC,EAAK,OAAO,EACjB,IAAM,EAAQ,OAAO,CAAG,EACxB,OAAO,OAAO,SAAS,CAAK,EAAI,EAAQ,CAC5C,CAgBA,SAAgB,EACZ,EACiB,CACjB,GAAI,CAAC,EAAU,OAAO,KAEtB,IAAM,EAA+B,CAAC,EACtC,IAAK,GAAM,CAAC,EAAK,KAAU,OAAO,QAAQ,CAAQ,EAC1C,EAAI,WAAA,MAA0B,IAAG,EAAK,EAAI,MAAM,CAAsB,GAAK,GAEnF,GAAI,EAAK,KAAA,kBAA2C,OAAO,KAE3D,IAAM,EAAY,EAAW,EAAK,EAAe,EAC3C,EAAW,EAAW,EAAK,EAAc,EAC/C,GAAI,IAAc,MAAQ,IAAa,KAAM,OAAO,KAEpD,IAAM,EAAS,EAAK,IAAuB,EACrC,EAAY,OAAO,CAAM,EACzB,EACF,IAAW,GAAW,CAAC,OAAO,UAAU,CAAS,GAAK,EAAY,EAAI,KAAO,EAE3E,EACF,EAAK,KAAqB,WAAa,WAAa,iBAExD,MAAO,CACH,KAAM,EACN,YACA,WACA,aACA,gBACA,cAAe,EAAY,EAAK,GAAqB,GAAI,EACzD,aAAc,EAAY,EAAK,GAAoB,GAAI,EACvD,cAAe,EAAK,IAAsB,OAAS,IACnD,cAAe,EAAA,WAAW,EAAK,EAAmB,EAClD,gBAAiB,EAAA,WAAW,EAAK,EAAqB,EACtD,WAAY,EAAK,IAAoB,GACrC,iBAAkB,IAAe,UACrC,CACJ"}
+{"version":3,"file":"fusion.cjs","names":[],"sources":["../../src/vision/fusion.ts"],"sourcesContent":["/** @generated Vendored from @mauriciobenjamin700/ort-vision-sdk-web. Do not hand-edit — regenerate with `npm run vendor:vision`. */\n/**\n * The contract a fused pipeline carries inside its own `.onnx` file.\n *\n * A pipeline built by the Python SDK's `ort_vision_sdk.compose` is a single\n * graph that already contains the detector, the crop-and-resize bridge and the\n * classifier. Everything the runtime needs to drive it — the letterbox\n * resolution, the crop resolution, where the crops come from, how many\n * detections the graph emits, the class names of both stages — was decided at\n * fusion time and written into the model's metadata.\n *\n * This module reads it back. It is the browser half of a contract whose other\n * half lives in `ort_vision_sdk/fusion.py`: the same keys, the same encodings,\n * the same fallbacks. A pipeline fused once therefore runs identically in\n * Python and in a browser tab, off the same file.\n *\n * Building a pipeline stays a Python-side build step — there is no ONNX\n * protobuf writer here, and there is no reason for one: fusing is something you\n * do once next to your export pipeline, not in a page load.\n */\n\nimport { parseNames } from \"./core/metadata\";\n\n/**\n * Which tensor the bridge crops the detected boxes out of.\n *\n * - `\"detector_input\"`: the letterboxed tensor already fed to the detector.\n *   The fused graph then has a **single** image input, but a small object is\n *   cropped out of its downscaled copy.\n * - `\"original\"`: a second, full-resolution image input. The bridge undoes the\n *   letterbox transform in-graph and crops at native resolution. Two tensors to\n *   feed, still one session and one model load.\n */\nexport type CropSource = \"detector_input\" | \"original\";\n\n/** Value of the `ovs.kind` metadata key for a detector→classifier pipeline. */\nexport const FUSION_KIND_DETECT_CLASSIFY = \"detect_classify\";\n\n/**\n * Value of `ovs.kind` for a detector→segmenter→classifier pipeline.\n *\n * The segmenter runs inside each crop rather than over the whole frame: the\n * detector finds the boxes, the bridge cuts them out, and the segmenter turns\n * each crop into a foreground mask that gates the pixels the classifier sees.\n * Fusing one is a build step on the Python side; the browser only loads the\n * result.\n */\nexport const FUSION_KIND_DETECT_SEGMENT_CLASSIFY = \"detect_segment_classify\";\n\n/**\n * Namespace for every metadata key the fusion writes.\n *\n * Namespaced on purpose: the detector's own Ultralytics metadata (`names`,\n * `task`, `imgsz`) is carried over into the fused model, and an un-prefixed key\n * would either collide with it or be mistaken for it.\n */\nexport const METADATA_PREFIX = \"ovs.\";\n\n/** Name of the fused graph's letterboxed detector input, `[1, 3, H, W]` float32 in `[0, 1]`. */\nexport const INPUT_IMAGE = \"images\";\n\n/** Name of the full-resolution input. Present only when `cropSource === \"original\"`. */\nexport const INPUT_SOURCE = \"source_image\";\n\n/** Name of the `[1]` float32 letterbox scale factor. Only with `cropSource === \"original\"`. */\nexport const INPUT_SCALE = \"letterbox_scale\";\n\n/** Name of the `[2]` float32 `[padLeft, padTop]`. Only with `cropSource === \"original\"`. */\nexport const INPUT_PAD = \"letterbox_pad\";\n\n/**\n * Name of the `[K, 4]` float32 xyxy output, in **letterboxed** input pixels.\n *\n * A file fused by `ort-vision-sdk` 0.9.0 or later reports the box that was\n * actually classified: clamped to the image the crop came from, exactly as\n * RoiAlign received it. Older files report the raw box, so one that ran off the\n * frame draws a rectangle wider than the region the classifier saw.\n */\nexport const OUTPUT_BOXES = \"boxes\";\n\n/** Name of the `[K]` float32 detection-confidence output. */\nexport const OUTPUT_SCORES = \"scores\";\n\n/** Name of the `[K]` int64 detector-class output. */\nexport const OUTPUT_CLASSES = \"classes\";\n\n/** Name of the `[1]` int64 output holding how many of the `K` rows are real. */\nexport const OUTPUT_NUM_DETECTIONS = \"num_detections\";\n\n/** Name of the `[K, numClassifierClasses]` float32 classifier output, one row per box. */\n/**\n * Name of the `(K, 1, cropH, cropW)` float32 mask output. Only on a\n * `detect_segment_classify` pipeline.\n *\n * The mask lives in the **crop's** coordinate space, not the original image's,\n * because that is where the segmenter computed it. Values are 0 or 1 after\n * thresholding; padded rows are all-zero, like every other output.\n */\nexport const OUTPUT_MASKS = \"masks\";\n\nexport const OUTPUT_PROBS = \"probs\";\n\n/** Everything a fused pipeline declares about how it must be driven. */\nexport interface FusionSpec {\n    /** Pipeline family — {@link FUSION_KIND_DETECT_CLASSIFY} or {@link FUSION_KIND_DETECT_SEGMENT_CLASSIFY}. */\n    readonly kind: string;\n    /** `[width, height]` the detector stage expects — the resolution to letterbox to. */\n    readonly inputSize: readonly [number, number];\n    /** `[width, height]` every crop is resampled to inside the graph. */\n    readonly cropSize: readonly [number, number];\n    /** Which tensor the crops are taken from. */\n    readonly cropSource: CropSource;\n    /**\n     * Fixed number of rows `K` every output carries, surplus zero-padded and\n     * counted by {@link OUTPUT_NUM_DETECTIONS}. `null` means the graph emits\n     * exactly as many rows as survived NMS.\n     */\n    readonly maxDetections: number | null;\n    /** Score threshold baked into the graph's NMS node. */\n    readonly confThreshold: number;\n    /** IoU threshold baked into the graph's NMS node. */\n    readonly iouThreshold: number;\n    /** Whether the classifier stage emits logits that still need a softmax. */\n    readonly applySoftmax: boolean;\n    /** Detector class names in class-id order, or `null` when the fusion recorded none. */\n    readonly detectorNames: readonly string[] | null;\n    /** Classifier class names in class-id order, or `null`. */\n    readonly classifierNames: readonly string[] | null;\n    /** Version of `ort-vision-sdk` that produced the file. */\n    readonly sdkVersion: string;\n    /** Whether driving this pipeline requires feeding the full-resolution input. */\n    readonly needsSourceImage: boolean;\n    /** Whether this pipeline emits a {@link OUTPUT_MASKS} tensor. */\n    readonly hasMasks: boolean;\n    /** Class id → name for the segmenter's channels, or `null`. */\n    readonly segmenterNames: readonly string[] | null;\n    /**\n     * Probability above which a segmenter pixel counted as foreground.\n     *\n     * Baked into the graph at fusion time; reported so a caller can say what the\n     * masks mean.\n     */\n    readonly maskThreshold: number;\n    /** Whether the mask gated the crop before the classifier saw it. */\n    readonly maskApplied: boolean;\n}\n\nconst KEY_KIND = \"kind\";\nconst KEY_SDK_VERSION = \"sdk_version\";\nconst KEY_INPUT_SIZE = \"input_size\";\nconst KEY_CROP_SIZE = \"crop_size\";\nconst KEY_CROP_SOURCE = \"crop_source\";\nconst KEY_MAX_DETECTIONS = \"max_detections\";\nconst KEY_CONF_THRESHOLD = \"conf_threshold\";\nconst KEY_IOU_THRESHOLD = \"iou_threshold\";\nconst KEY_APPLY_SOFTMAX = \"apply_softmax\";\nconst KEY_DETECTOR_NAMES = \"detector_names\";\nconst KEY_CLASSIFIER_NAMES = \"classifier_names\";\nconst KEY_SEGMENTER_NAMES = \"segmenter_names\";\nconst KEY_MASK_THRESHOLD = \"mask_threshold\";\nconst KEY_MASK_APPLIED = \"mask_applied\";\n\nconst DYNAMIC = \"dynamic\";\n\n/**\n * Decode a `\"640,640\"` pair.\n *\n * @param raw The encoded pair.\n * @returns `[width, height]`, or `null` when the value is missing or malformed.\n */\nfunction decodeSize(raw: string | undefined): readonly [number, number] | null {\n    if (!raw) return null;\n    const parts = raw.split(\",\");\n    if (parts.length !== 2) return null;\n    const width = Number(parts[0]);\n    const height = Number(parts[1]);\n    if (!Number.isInteger(width) || !Number.isInteger(height)) return null;\n    if (width < 1 || height < 1) return null;\n    return [width, height];\n}\n\n/**\n * Decode a float, falling back when the value is missing or malformed.\n *\n * @param raw The encoded value.\n * @param fallback Value to use when `raw` cannot be read.\n * @returns The parsed number, or `fallback`.\n */\nfunction decodeFloat(raw: string | undefined, fallback: number): number {\n    if (!raw) return fallback;\n    const value = Number(raw);\n    return Number.isFinite(value) ? value : fallback;\n}\n\n/**\n * Read a pipeline spec out of a model's custom metadata.\n *\n * Individual malformed entries fall back to the value a fusion would have used\n * by default — a single bad float is not a reason to reject an otherwise\n * loadable pipeline. A malformed resolution is fatal, because there is no safe\n * default for one.\n *\n * @param metadata A model's custom metadata map, as read by\n *   {@link readModelMetadata}.\n * @returns The decoded spec, or `null` when the model is not a fused pipeline —\n *   it carries no `ovs.kind` entry, or one naming a pipeline kind this version\n *   does not know how to drive.\n */\nexport function readFusionSpec(\n    metadata: Readonly<Record<string, string>> | undefined,\n): FusionSpec | null {\n    if (!metadata) return null;\n\n    const read: Record<string, string> = {};\n    for (const [key, value] of Object.entries(metadata)) {\n        if (key.startsWith(METADATA_PREFIX)) read[key.slice(METADATA_PREFIX.length)] = value;\n    }\n    const kind = read[KEY_KIND];\n    if (kind !== FUSION_KIND_DETECT_CLASSIFY && kind !== FUSION_KIND_DETECT_SEGMENT_CLASSIFY) {\n        return null;\n    }\n\n    const inputSize = decodeSize(read[KEY_INPUT_SIZE]);\n    const cropSize = decodeSize(read[KEY_CROP_SIZE]);\n    if (inputSize === null || cropSize === null) return null;\n\n    const rawMax = read[KEY_MAX_DETECTIONS] ?? DYNAMIC;\n    const parsedMax = Number(rawMax);\n    const maxDetections =\n        rawMax === DYNAMIC || !Number.isInteger(parsedMax) || parsedMax < 1 ? null : parsedMax;\n\n    const cropSource: CropSource =\n        read[KEY_CROP_SOURCE] === \"original\" ? \"original\" : \"detector_input\";\n\n    return {\n        kind,\n        inputSize,\n        cropSize,\n        cropSource,\n        maxDetections,\n        confThreshold: decodeFloat(read[KEY_CONF_THRESHOLD], 0.25),\n        iouThreshold: decodeFloat(read[KEY_IOU_THRESHOLD], 0.45),\n        applySoftmax: (read[KEY_APPLY_SOFTMAX] ?? \"1\") !== \"0\",\n        detectorNames: parseNames(read[KEY_DETECTOR_NAMES]),\n        classifierNames: parseNames(read[KEY_CLASSIFIER_NAMES]),\n        sdkVersion: read[KEY_SDK_VERSION] ?? \"\",\n        needsSourceImage: cropSource === \"original\",\n        hasMasks: kind === FUSION_KIND_DETECT_SEGMENT_CLASSIFY,\n        segmenterNames: parseNames(read[KEY_SEGMENTER_NAMES]),\n        maskThreshold: decodeFloat(read[KEY_MASK_THRESHOLD], 0.5),\n        maskApplied: (read[KEY_MASK_APPLIED] ?? \"1\") !== \"0\",\n    };\n}\n"],"mappings":"uCAoCA,IAAa,EAA8B,kBAW9B,EAAsC,0BAStC,EAAkB,OAGlB,EAAc,SAGd,EAAe,eAGf,EAAc,kBAGd,EAAY,gBAUZ,EAAe,QAGf,EAAgB,SAGhB,EAAiB,UAGjB,EAAwB,iBAWxB,EAAe,QAEf,EAAe,QA+CtB,EAAW,OACX,EAAkB,cAClB,EAAiB,aACjB,EAAgB,YAChB,EAAkB,cAClB,EAAqB,iBACrB,EAAqB,iBACrB,EAAoB,gBACpB,EAAoB,gBACpB,EAAqB,iBACrB,EAAuB,mBACvB,EAAsB,kBACtB,EAAqB,iBACrB,EAAmB,eAEnB,EAAU,UAQhB,SAAS,EAAW,EAA2D,CAC3E,GAAI,CAAC,EAAK,OAAO,KACjB,IAAM,EAAQ,EAAI,MAAM,GAAG,EAC3B,GAAI,EAAM,SAAW,EAAG,OAAO,KAC/B,IAAM,EAAQ,OAAO,EAAM,EAAE,EACvB,EAAS,OAAO,EAAM,EAAE,EAG9B,MAFI,CAAC,OAAO,UAAU,CAAK,GAAK,CAAC,OAAO,UAAU,CAAM,GACpD,EAAQ,GAAK,EAAS,EAAU,KAC7B,CAAC,EAAO,CAAM,CACzB,CASA,SAAS,EAAY,EAAyB,EAA0B,CACpE,GAAI,CAAC,EAAK,OAAO,EACjB,IAAM,EAAQ,OAAO,CAAG,EACxB,OAAO,OAAO,SAAS,CAAK,EAAI,EAAQ,CAC5C,CAgBA,SAAgB,EACZ,EACiB,CACjB,GAAI,CAAC,EAAU,OAAO,KAEtB,IAAM,EAA+B,CAAC,EACtC,IAAK,GAAM,CAAC,EAAK,KAAU,OAAO,QAAQ,CAAQ,EAC1C,EAAI,WAAA,MAA0B,IAAG,EAAK,EAAI,MAAM,CAAsB,GAAK,GAEnF,IAAM,EAAO,EAAK,GAClB,GAAI,IAAA,mBAAwC,IAAA,0BACxC,OAAO,KAGX,IAAM,EAAY,EAAW,EAAK,EAAe,EAC3C,EAAW,EAAW,EAAK,EAAc,EAC/C,GAAI,IAAc,MAAQ,IAAa,KAAM,OAAO,KAEpD,IAAM,EAAS,EAAK,IAAuB,EACrC,EAAY,OAAO,CAAM,EACzB,EACF,IAAW,GAAW,CAAC,OAAO,UAAU,CAAS,GAAK,EAAY,EAAI,KAAO,EAE3E,EACF,EAAK,KAAqB,WAAa,WAAa,iBAExD,MAAO,CACH,OACA,YACA,WACA,aACA,gBACA,cAAe,EAAY,EAAK,GAAqB,GAAI,EACzD,aAAc,EAAY,EAAK,GAAoB,GAAI,EACvD,cAAe,EAAK,IAAsB,OAAS,IACnD,cAAe,EAAA,WAAW,EAAK,EAAmB,EAClD,gBAAiB,EAAA,WAAW,EAAK,EAAqB,EACtD,WAAY,EAAK,IAAoB,GACrC,iBAAkB,IAAe,WACjC,SAAU,IAAS,EACnB,eAAgB,EAAA,WAAW,EAAK,EAAoB,EACpD,cAAe,EAAY,EAAK,GAAqB,EAAG,EACxD,aAAc,EAAK,IAAqB,OAAS,GACrD,CACJ"}