inferis-ml 1.0.0

package/dist/adapters/onnx.d.cts

@@ -0,0 +1,28 @@
+import { h as ModelAdapterFactory } from '../types-Y6Ytjh7U.cjs';
+
+/**
+ * Adapter for onnxruntime-web.
+ *
+ * @remarks
+ * Provides low-level access to ONNX model inference.
+ * Use for custom models not supported by transformers.js or web-llm.
+ * Input must be pre-processed `OrtTensor` instances.
+ *
+ * @example
+ * ```ts
+ * import { createPool } from 'inferis-ml';
+ * import { onnxAdapter } from 'inferis-ml/adapters/onnx';
+ * import * as ort from 'onnxruntime-web';
+ *
+ * const pool = await createPool({ adapter: onnxAdapter() });
+ * const model = await pool.load('custom', {
+ *   model: 'https://example.com/model.onnx',
+ * });
+ *
+ * const input = new ort.Tensor('float32', data, [1, 3, 224, 224]);
+ * const output = await model.run({ feeds: { input } });
+ * ```
+ */
+declare function onnxAdapter(): ModelAdapterFactory;
+
+export { onnxAdapter };

package/dist/adapters/onnx.d.ts

@@ -0,0 +1,28 @@
+import { h as ModelAdapterFactory } from '../types-Y6Ytjh7U.js';
+
+/**
+ * Adapter for onnxruntime-web.
+ *
+ * @remarks
+ * Provides low-level access to ONNX model inference.
+ * Use for custom models not supported by transformers.js or web-llm.
+ * Input must be pre-processed `OrtTensor` instances.
+ *
+ * @example
+ * ```ts
+ * import { createPool } from 'inferis-ml';
+ * import { onnxAdapter } from 'inferis-ml/adapters/onnx';
+ * import * as ort from 'onnxruntime-web';
+ *
+ * const pool = await createPool({ adapter: onnxAdapter() });
+ * const model = await pool.load('custom', {
+ *   model: 'https://example.com/model.onnx',
+ * });
+ *
+ * const input = new ort.Tensor('float32', data, [1, 3, 224, 224]);
+ * const output = await model.run({ feeds: { input } });
+ * ```
+ */
+declare function onnxAdapter(): ModelAdapterFactory;
+
+export { onnxAdapter };

package/dist/adapters/onnx.js

@@ -0,0 +1,2 @@
+function c(){return {name:"onnx",async create(){let d=await import('onnxruntime-web');return {name:"onnx",estimateMemoryMB(n,e){return e.estimatedMemoryMB??50},async load(n,e,a,s){let o=e.model;s({loaded:0,phase:"downloading",total:0});let t=a==="webgpu"?["webgpu","wasm"]:["wasm"],r=await d.InferenceSession.create(o,{executionProviders:t,graphOptimizationLevel:"all"});return s({loaded:1,phase:"done",total:1}),{instance:r,memoryMB:e.estimatedMemoryMB??50}},async run(n,e,a){let s=n.instance,{feeds:o,outputNames:t}=e;return await s.run(o,t)},async stream(n,e,a,s){let o=n.instance,{feeds:t,outputNames:r}=e,i=await o.run(t,r,s);a(i);},async unload(n){await n.instance.release?.();}}}}}export{c as onnxAdapter};//# sourceMappingURL=onnx.js.map
+//# sourceMappingURL=onnx.js.map

package/dist/adapters/onnx.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../src/adapters/onnx.ts"],"names":["onnxAdapter","ort","_task","config","device","onProgress","modelUrl","executionProviders","session","model","input","_options","feeds","outputNames","onChunk","options","result"],"mappings":"AAyCO,SAASA,GAAmC,CACjD,OAAO,CACL,IAAA,CAAM,OAEN,MAAM,MAAA,EAAgC,CAEpC,IAAMC,EAAM,MAAM,OAAO,iBAAiB,CAAA,CAE1C,OAAO,CACL,IAAA,CAAM,MAAA,CAEN,gBAAA,CAAiBC,EAAeC,CAAAA,CAAyC,CACvE,OAAQA,CAAAA,CAAO,mBAA4C,EAC7D,CAAA,CAEA,MAAM,IAAA,CACJD,EACAC,CAAAA,CACAC,CAAAA,CACAC,EACsB,CACtB,IAAMC,EAAWH,CAAAA,CAAO,KAAA,CAExBE,CAAAA,CAAW,CAAE,OAAQ,CAAA,CAAG,KAAA,CAAO,aAAA,CAAe,KAAA,CAAO,CAAE,CAAC,CAAA,CAExD,IAAME,CAAAA,CAAqBH,IAAW,QAAA,CAClC,CAAC,QAAA,CAAU,MAAM,EACjB,CAAC,MAAM,CAAA,CAELI,CAAAA,CAA4B,MAAMP,CAAAA,CAAI,gBAAA,CAAiB,MAAA,CAAOK,CAAAA,CAAU,CAC5E,kBAAA,CAAAC,CAAAA,CACA,sBAAA,CAAwB,KAC1B,CAAC,CAAA,CAED,OAAAF,EAAW,CAAE,MAAA,CAAQ,EAAG,KAAA,CAAO,MAAA,CAAQ,KAAA,CAAO,CAAE,CAAC,CAAA,CAE1C,CACL,QAAA,CAAUG,CAAAA,CACV,SAAWL,CAAAA,CAAO,iBAAA,EAA4C,EAChE,CACF,EAEA,MAAM,GAAA,CAAIM,CAAAA,CAAoBC,CAAAA,CAAgBC,EAAsC,CAClF,IAAMH,CAAAA,CAAUC,CAAAA,CAAM,SAChB,CAAE,KAAA,CAAAG,CAAAA,CAAO,WAAA,CAAAC,CAAY,CAAA,CAAIH,CAAAA,CAG/B,OADgB,MAAMF,EAAQ,GAAA,CAAII,CAAAA,CAAOC,CAAW,CAEtD,CAAA,CAEA,MAAM,MAAA,CACJJ,CAAAA,CACAC,CAAAA,CACAI,CAAAA,CACAC,EACe,CAGf,IAAMP,CAAAA,CAAUC,CAAAA,CAAM,SAChB,CAAE,KAAA,CAAAG,CAAAA,CAAO,WAAA,CAAAC,CAAY,CAAA,CAAIH,CAAAA,CACzBM,CAAAA,CAAS,MAAMR,EAAQ,GAAA,CAAII,CAAAA,CAAOC,CAAAA,CAAaE,CAAO,EAC5DD,CAAAA,CAAQE,CAAM,EAChB,CAAA,CAEA,MAAM,MAAA,CAAOP,CAAAA,CAAmC,CAE9C,MADgBA,EAAM,QAAA,CACR,OAAA,KAChB,CACF,CACF,CACF,CACF","file":"onnx.js","sourcesContent":["import type {\n  Device,\n  LoadedModel,\n  LoadProgressEvent,\n  ModelAdapter,\n  ModelAdapterFactory,\n} from '../core/types.js';\n\n// eslint-disable-next-line ts/no-explicit-any\ntype InferenceSession = any;\n// eslint-disable-next-line ts/no-explicit-any\ntype Tensor = any;\n\ninterface OnnxInput {\n  feeds: Record<string, Tensor>;\n  outputNames?: string[];\n}\n\n/**\n * Adapter for onnxruntime-web.\n *\n * @remarks\n * Provides low-level access to ONNX model inference.\n * Use for custom models not supported by transformers.js or web-llm.\n * Input must be pre-processed `OrtTensor` instances.\n *\n * @example\n * ```ts\n * import { createPool } from 'inferis-ml';\n * import { onnxAdapter } from 'inferis-ml/adapters/onnx';\n * import * as ort from 'onnxruntime-web';\n *\n * const pool = await createPool({ adapter: onnxAdapter() });\n * const model = await pool.load('custom', {\n *   model: 'https://example.com/model.onnx',\n * });\n *\n * const input = new ort.Tensor('float32', data, [1, 3, 224, 224]);\n * const output = await model.run({ feeds: { input } });\n * ```\n */\nexport function onnxAdapter(): ModelAdapterFactory {\n  return {\n    name: 'onnx',\n\n    async create(): Promise<ModelAdapter> {\n      // @ts-expect-error - optional peer dependency, resolved at runtime inside worker\n      const ort = await import('onnxruntime-web');\n\n      return {\n        name: 'onnx',\n\n        estimateMemoryMB(_task: string, config: Record<string, unknown>): number {\n          return (config.estimatedMemoryMB as number | undefined) ?? 50;\n        },\n\n        async load(\n          _task: string,\n          config: Record<string, unknown>,\n          device: Device,\n          onProgress: (event: LoadProgressEvent) => void,\n        ): Promise<LoadedModel> {\n          const modelUrl = config.model as string;\n\n          onProgress({ loaded: 0, phase: 'downloading', total: 0 });\n\n          const executionProviders = device === 'webgpu'\n            ? ['webgpu', 'wasm'] as const\n            : ['wasm'] as const;\n\n          const session: InferenceSession = await ort.InferenceSession.create(modelUrl, {\n            executionProviders,\n            graphOptimizationLevel: 'all',\n          });\n\n          onProgress({ loaded: 1, phase: 'done', total: 1 });\n\n          return {\n            instance: session,\n            memoryMB: (config.estimatedMemoryMB as number | undefined) ?? 50,\n          };\n        },\n\n        async run(model: LoadedModel, input: unknown, _options?: unknown): Promise<unknown> {\n          const session = model.instance as InferenceSession;\n          const { feeds, outputNames } = input as OnnxInput;\n\n          const results = await session.run(feeds, outputNames);\n          return results;\n        },\n\n        async stream(\n          model: LoadedModel,\n          input: unknown,\n          onChunk: (chunk: unknown) => void,\n          options?: unknown,\n        ): Promise<void> {\n          // ONNX Runtime Web doesn't natively support streaming inference.\n          // Run full inference and emit the result as a single chunk.\n          const session = model.instance as InferenceSession;\n          const { feeds, outputNames } = input as OnnxInput;\n          const result = await session.run(feeds, outputNames, options);\n          onChunk(result);\n        },\n\n        async unload(model: LoadedModel): Promise<void> {\n          const session = model.instance as InferenceSession;\n          await session.release?.();\n        },\n      };\n    },\n  };\n}\n"]}

package/dist/adapters/transformers.cjs

@@ -0,0 +1,2 @@
+'use strict';function k(){return {name:"transformers",async create(){let{pipeline:r,TextStreamer:t,env:p}=await import('@huggingface/transformers');return p.backends.onnx.wasm.proxy=false,{name:"transformers",estimateMemoryMB(n,e){return e.estimatedMemoryMB??100},async load(n,e,o,a){let i=e.model,c=e.dtype??"fp32",d=new Map,w=await r(n,i,{device:o==="webgpu"?"webgpu":"wasm",dtype:c,progress_callback:s=>{if(s.status==="progress"){let m=s.file??"unknown",y=s.total??0;d.set(m,y),a({loaded:s.loaded??0,phase:m,total:y});}else s.status==="initiate"?a({loaded:0,phase:"initiate",total:0}):s.status==="done"&&a({loaded:1,phase:"done",total:1});}}),l=0;for(let s of d.values())l+=s;return {instance:w,memoryMB:Math.ceil(l/(1024*1024))||100}},async run(n,e,o){let a=n.instance,i=Array.isArray(e)?await a(...e,o??{}):await a(e,o??{});return u(i)},async stream(n,e,o,a){let i=n.instance,c=new t(i.tokenizer,{callback_function:d=>o(d),skip_prompt:true,skip_special_tokens:true});await i(e,{...a??{},streamer:c});},async unload(n){await n.instance.dispose?.();}}}}}function u(r){if(r===null||typeof r!="object")return r;if(Array.isArray(r))return r.map(u);let t=r;return typeof t.data<"u"&&Array.isArray(t.dims)?{data:Array.from(t.data),dims:t.dims,size:t.size,type:t.type}:Object.fromEntries(Object.entries(t).map(([p,n])=>[p,u(n)]))}exports.serializeOutput=u;exports.transformersAdapter=k;//# sourceMappingURL=transformers.cjs.map
+//# sourceMappingURL=transformers.cjs.map

package/dist/adapters/transformers.cjs.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../src/adapters/transformers.ts"],"names":["transformersAdapter","pipeline","TextStreamer","env","_task","config","task","device","onProgress","model","dtype","fileTotals","pipe","info","file","total","totalBytes","size","input","options","result","serializeOutput","onChunk","streamer","text","value","obj","k","v"],"mappings":"aA8BO,SAASA,GAA2C,CACzD,OAAO,CACL,IAAA,CAAM,eAEN,MAAM,MAAA,EAAgC,CAGpC,GAAM,CAAE,QAAA,CAAAC,CAAAA,CAAU,YAAA,CAAAC,CAAAA,CAAc,IAAAC,CAAI,CAAA,CAAI,MAAM,OAAO,2BAA2B,CAAA,CAGhF,OAAAA,CAAAA,CAAI,QAAA,CAAS,KAAK,IAAA,CAAK,KAAA,CAAQ,KAAA,CAExB,CACL,KAAM,cAAA,CAEN,gBAAA,CAAiBC,CAAAA,CAAeC,CAAAA,CAAyC,CACvE,OAAQA,CAAAA,CAAO,iBAAA,EAA4C,GAC7D,EAEA,MAAM,IAAA,CACJC,CAAAA,CACAD,CAAAA,CACAE,EACAC,CAAAA,CACsB,CACtB,IAAMC,CAAAA,CAAQJ,EAAO,KAAA,CACfK,CAAAA,CAASL,CAAAA,CAAO,KAAA,EAAgC,OAEhDM,CAAAA,CAAa,IAAI,GAAA,CAEjBC,CAAAA,CAAO,MAAMX,CAAAA,CAASK,CAAAA,CAAwCG,CAAAA,CAAO,CACzE,OAAQF,CAAAA,GAAW,QAAA,CAAW,QAAA,CAAW,MAAA,CACzC,MAAAG,CAAAA,CACA,iBAAA,CAAoBG,GAAkC,CACpD,GAAIA,EAAK,MAAA,GAAW,UAAA,CAAY,CAC9B,IAAMC,EAAQD,CAAAA,CAAK,IAAA,EAAmB,SAAA,CAChCE,CAAAA,CAASF,EAAK,KAAA,EAAoB,CAAA,CACxCF,CAAAA,CAAW,GAAA,CAAIG,EAAMC,CAAK,CAAA,CAC1BP,CAAAA,CAAW,CACT,OAASK,CAAAA,CAAK,MAAA,EAAqB,CAAA,CACnC,KAAA,CAAOC,EACP,KAAA,CAAAC,CACF,CAAC,EACH,MACSF,CAAAA,CAAK,MAAA,GAAW,UAAA,CACvBL,CAAAA,CAAW,CAAE,MAAA,CAAQ,CAAA,CAAG,KAAA,CAAO,UAAA,CAAY,MAAO,CAAE,CAAC,CAAA,CAE9CK,CAAAA,CAAK,SAAW,MAAA,EACvBL,CAAAA,CAAW,CAAE,MAAA,CAAQ,EAAG,KAAA,CAAO,MAAA,CAAQ,KAAA,CAAO,CAAE,CAAC,EAErD,CACF,CAAC,CAAA,CAEGQ,EAAa,CAAA,CACjB,IAAA,IAAWC,CAAAA,IAAQN,CAAAA,CAAW,QAAO,CAAGK,CAAAA,EAAcC,CAAAA,CAEtD,OAAO,CACL,QAAA,CAAUL,CAAAA,CACV,QAAA,CAAU,IAAA,CAAK,KAAKI,CAAAA,EAAc,IAAA,CAAO,KAAK,CAAA,EAAK,GACrD,CACF,CAAA,CAEA,MAAM,GAAA,CAAIP,CAAAA,CAAoBS,EAAgBC,CAAAA,CAAqC,CACjF,IAAMP,CAAAA,CAAOH,EAAM,QAAA,CACbW,CAAAA,CAAS,KAAA,CAAM,OAAA,CAAQF,CAAK,CAAA,CAC9B,MAAMN,CAAAA,CAAK,GAAIM,EAAmCC,CAAAA,EAAW,EAAE,CAAA,CAC/D,MAAMP,CAAAA,CAAKM,CAAAA,CAAOC,CAAAA,EAAW,EAAE,CAAA,CACnC,OAAOE,CAAAA,CAAgBD,CAAM,CAC/B,CAAA,CAEA,MAAM,MAAA,CACJX,CAAAA,CACAS,EACAI,CAAAA,CACAH,CAAAA,CACe,CACf,IAAMP,EAAOH,CAAAA,CAAM,QAAA,CAEbc,CAAAA,CAAW,IAAIrB,EAAaU,CAAAA,CAAK,SAAA,CAAW,CAChD,iBAAA,CAAoBY,GAAiBF,CAAAA,CAAQE,CAAI,CAAA,CACjD,WAAA,CAAa,KACb,mBAAA,CAAqB,IACvB,CAAC,CAAA,CAED,MAAMZ,CAAAA,CAAKM,CAAAA,CAAO,CAChB,GAAIC,GAAqB,EAAC,CAC1B,QAAA,CAAAI,CACF,CAAC,EACH,CAAA,CAEA,MAAM,MAAA,CAAOd,CAAAA,CAAmC,CAE9C,MADaA,CAAAA,CAAM,QAAA,CACR,OAAA,KACb,CACF,CACF,CACF,CACF,CAMO,SAASY,CAAAA,CAAgBI,CAAAA,CAAyB,CACvD,GAAIA,CAAAA,GAAU,IAAA,EAAQ,OAAOA,CAAAA,EAAU,SACrC,OAAOA,CAAAA,CAET,GAAI,KAAA,CAAM,QAAQA,CAAK,CAAA,CACrB,OAAOA,CAAAA,CAAM,IAAIJ,CAAe,CAAA,CAElC,IAAMK,CAAAA,CAAMD,EAEZ,OAAI,OAAOC,CAAAA,CAAI,IAAA,CAAS,KAAe,KAAA,CAAM,OAAA,CAAQA,CAAAA,CAAI,IAAI,EACpD,CACL,IAAA,CAAM,KAAA,CAAM,IAAA,CAAKA,EAAI,IAAyB,CAAA,CAC9C,IAAA,CAAMA,CAAAA,CAAI,KACV,IAAA,CAAMA,CAAAA,CAAI,IAAA,CACV,IAAA,CAAMA,EAAI,IACZ,CAAA,CAGK,MAAA,CAAO,WAAA,CACZ,OAAO,OAAA,CAAQA,CAAG,CAAA,CAAE,GAAA,CAAI,CAAC,CAACC,CAAAA,CAAGC,CAAC,CAAA,GAAM,CAACD,CAAAA,CAAGN,CAAAA,CAAgBO,CAAC,CAAC,CAAC,CAC7D,CACF","file":"transformers.cjs","sourcesContent":["import type {\n  Device,\n  LoadedModel,\n  LoadProgressEvent,\n  ModelAdapter,\n  ModelAdapterFactory,\n} from '../core/types.js';\n\n// eslint-disable-next-line ts/no-explicit-any\ntype AnyPipeline = any;\n\n/**\n * Adapter for @huggingface/transformers v3+.\n *\n * @remarks\n * The adapter is instantiated INSIDE the worker via `create()`.\n * The heavy `@huggingface/transformers` library is dynamically imported\n * inside the worker, keeping the main thread bundle lightweight.\n *\n * @example\n * ```ts\n * import { createPool } from 'inferis-ml';\n * import { transformersAdapter } from 'inferis-ml/adapters/transformers';\n *\n * const pool = await createPool({ adapter: transformersAdapter() });\n * const model = await pool.load<number[][]>('feature-extraction', {\n *   model: 'mixedbread-ai/mxbai-embed-xsmall-v1',\n * });\n * ```\n */\nexport function transformersAdapter(): ModelAdapterFactory {\n  return {\n    name: 'transformers',\n\n    async create(): Promise<ModelAdapter> {\n      // Dynamic import — runs inside the worker only\n      // @ts-expect-error - optional peer dependency, resolved at runtime inside worker\n      const { pipeline, TextStreamer, env } = await import('@huggingface/transformers');\n\n      // Use ONNX WASM backend with optimizations\n      env.backends.onnx.wasm.proxy = false;\n\n      return {\n        name: 'transformers',\n\n        estimateMemoryMB(_task: string, config: Record<string, unknown>): number {\n          return (config.estimatedMemoryMB as number | undefined) ?? 100;\n        },\n\n        async load(\n          task: string,\n          config: Record<string, unknown>,\n          device: Device,\n          onProgress: (event: LoadProgressEvent) => void,\n        ): Promise<LoadedModel> {\n          const model = config.model as string;\n          const dtype = (config.dtype as string | undefined) ?? 'fp32';\n\n          const fileTotals = new Map<string, number>();\n\n          const pipe = await pipeline(task as Parameters<typeof pipeline>[0], model, {\n            device: device === 'webgpu' ? 'webgpu' : 'wasm',\n            dtype,\n            progress_callback: (info: Record<string, unknown>) => {\n              if (info.status === 'progress') {\n                const file = (info.file as string) ?? 'unknown';\n                const total = (info.total as number) ?? 0;\n                fileTotals.set(file, total);\n                onProgress({\n                  loaded: (info.loaded as number) ?? 0,\n                  phase: file,\n                  total,\n                });\n              }\n              else if (info.status === 'initiate') {\n                onProgress({ loaded: 0, phase: 'initiate', total: 0 });\n              }\n              else if (info.status === 'done') {\n                onProgress({ loaded: 1, phase: 'done', total: 1 });\n              }\n            },\n          });\n\n          let totalBytes = 0;\n          for (const size of fileTotals.values()) totalBytes += size;\n\n          return {\n            instance: pipe,\n            memoryMB: Math.ceil(totalBytes / (1024 * 1024)) || 100,\n          };\n        },\n\n        async run(model: LoadedModel, input: unknown, options?: unknown): Promise<unknown> {\n          const pipe = model.instance as AnyPipeline;\n          const result = Array.isArray(input)\n            ? await pipe(...(input as [unknown, ...unknown[]]), options ?? {})\n            : await pipe(input, options ?? {});\n          return serializeOutput(result);\n        },\n\n        async stream(\n          model: LoadedModel,\n          input: unknown,\n          onChunk: (chunk: unknown) => void,\n          options?: unknown,\n        ): Promise<void> {\n          const pipe = model.instance as AnyPipeline;\n\n          const streamer = new TextStreamer(pipe.tokenizer, {\n            callback_function: (text: string) => onChunk(text),\n            skip_prompt: true,\n            skip_special_tokens: true,\n          });\n\n          await pipe(input, {\n            ...(options as object ?? {}),\n            streamer,\n          });\n        },\n\n        async unload(model: LoadedModel): Promise<void> {\n          const pipe = model.instance as AnyPipeline;\n          await pipe.dispose?.();\n        },\n      };\n    },\n  };\n}\n\n/**\n * Recursively converts Tensor class instances and typed arrays to plain\n * structured-clone-compatible objects so they can cross the worker boundary.\n */\nexport function serializeOutput(value: unknown): unknown {\n  if (value === null || typeof value !== 'object')\n    return value;\n\n  if (Array.isArray(value))\n    return value.map(serializeOutput);\n\n  const obj = value as Record<string, unknown>;\n\n  if (typeof obj.data !== 'undefined' && Array.isArray(obj.dims)) {\n    return {\n      data: Array.from(obj.data as ArrayLike<number>),\n      dims: obj.dims as number[],\n      size: obj.size,\n      type: obj.type,\n    };\n  }\n\n  return Object.fromEntries(\n    Object.entries(obj).map(([k, v]) => [k, serializeOutput(v)]),\n  );\n}\n"]}

package/dist/adapters/transformers.d.cts

@@ -0,0 +1,29 @@
+import { h as ModelAdapterFactory } from '../types-Y6Ytjh7U.cjs';
+
+/**
+ * Adapter for @huggingface/transformers v3+.
+ *
+ * @remarks
+ * The adapter is instantiated INSIDE the worker via `create()`.
+ * The heavy `@huggingface/transformers` library is dynamically imported
+ * inside the worker, keeping the main thread bundle lightweight.
+ *
+ * @example
+ * ```ts
+ * import { createPool } from 'inferis-ml';
+ * import { transformersAdapter } from 'inferis-ml/adapters/transformers';
+ *
+ * const pool = await createPool({ adapter: transformersAdapter() });
+ * const model = await pool.load<number[][]>('feature-extraction', {
+ *   model: 'mixedbread-ai/mxbai-embed-xsmall-v1',
+ * });
+ * ```
+ */
+declare function transformersAdapter(): ModelAdapterFactory;
+/**
+ * Recursively converts Tensor class instances and typed arrays to plain
+ * structured-clone-compatible objects so they can cross the worker boundary.
+ */
+declare function serializeOutput(value: unknown): unknown;
+
+export { serializeOutput, transformersAdapter };

package/dist/adapters/transformers.d.ts

@@ -0,0 +1,29 @@
+import { h as ModelAdapterFactory } from '../types-Y6Ytjh7U.js';
+
+/**
+ * Adapter for @huggingface/transformers v3+.
+ *
+ * @remarks
+ * The adapter is instantiated INSIDE the worker via `create()`.
+ * The heavy `@huggingface/transformers` library is dynamically imported
+ * inside the worker, keeping the main thread bundle lightweight.
+ *
+ * @example
+ * ```ts
+ * import { createPool } from 'inferis-ml';
+ * import { transformersAdapter } from 'inferis-ml/adapters/transformers';
+ *
+ * const pool = await createPool({ adapter: transformersAdapter() });
+ * const model = await pool.load<number[][]>('feature-extraction', {
+ *   model: 'mixedbread-ai/mxbai-embed-xsmall-v1',
+ * });
+ * ```
+ */
+declare function transformersAdapter(): ModelAdapterFactory;
+/**
+ * Recursively converts Tensor class instances and typed arrays to plain
+ * structured-clone-compatible objects so they can cross the worker boundary.
+ */
+declare function serializeOutput(value: unknown): unknown;
+
+export { serializeOutput, transformersAdapter };

package/dist/adapters/transformers.js

@@ -0,0 +1,2 @@
+function k(){return {name:"transformers",async create(){let{pipeline:r,TextStreamer:t,env:p}=await import('@huggingface/transformers');return p.backends.onnx.wasm.proxy=false,{name:"transformers",estimateMemoryMB(n,e){return e.estimatedMemoryMB??100},async load(n,e,o,a){let i=e.model,c=e.dtype??"fp32",d=new Map,w=await r(n,i,{device:o==="webgpu"?"webgpu":"wasm",dtype:c,progress_callback:s=>{if(s.status==="progress"){let m=s.file??"unknown",y=s.total??0;d.set(m,y),a({loaded:s.loaded??0,phase:m,total:y});}else s.status==="initiate"?a({loaded:0,phase:"initiate",total:0}):s.status==="done"&&a({loaded:1,phase:"done",total:1});}}),l=0;for(let s of d.values())l+=s;return {instance:w,memoryMB:Math.ceil(l/(1024*1024))||100}},async run(n,e,o){let a=n.instance,i=Array.isArray(e)?await a(...e,o??{}):await a(e,o??{});return u(i)},async stream(n,e,o,a){let i=n.instance,c=new t(i.tokenizer,{callback_function:d=>o(d),skip_prompt:true,skip_special_tokens:true});await i(e,{...a??{},streamer:c});},async unload(n){await n.instance.dispose?.();}}}}}function u(r){if(r===null||typeof r!="object")return r;if(Array.isArray(r))return r.map(u);let t=r;return typeof t.data<"u"&&Array.isArray(t.dims)?{data:Array.from(t.data),dims:t.dims,size:t.size,type:t.type}:Object.fromEntries(Object.entries(t).map(([p,n])=>[p,u(n)]))}export{u as serializeOutput,k as transformersAdapter};//# sourceMappingURL=transformers.js.map
+//# sourceMappingURL=transformers.js.map

package/dist/adapters/transformers.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../src/adapters/transformers.ts"],"names":["transformersAdapter","pipeline","TextStreamer","env","_task","config","task","device","onProgress","model","dtype","fileTotals","pipe","info","file","total","totalBytes","size","input","options","result","serializeOutput","onChunk","streamer","text","value","obj","k","v"],"mappings":"AA8BO,SAASA,GAA2C,CACzD,OAAO,CACL,IAAA,CAAM,eAEN,MAAM,MAAA,EAAgC,CAGpC,GAAM,CAAE,QAAA,CAAAC,CAAAA,CAAU,YAAA,CAAAC,CAAAA,CAAc,IAAAC,CAAI,CAAA,CAAI,MAAM,OAAO,2BAA2B,CAAA,CAGhF,OAAAA,CAAAA,CAAI,QAAA,CAAS,KAAK,IAAA,CAAK,KAAA,CAAQ,KAAA,CAExB,CACL,KAAM,cAAA,CAEN,gBAAA,CAAiBC,CAAAA,CAAeC,CAAAA,CAAyC,CACvE,OAAQA,CAAAA,CAAO,iBAAA,EAA4C,GAC7D,EAEA,MAAM,IAAA,CACJC,CAAAA,CACAD,CAAAA,CACAE,EACAC,CAAAA,CACsB,CACtB,IAAMC,CAAAA,CAAQJ,EAAO,KAAA,CACfK,CAAAA,CAASL,CAAAA,CAAO,KAAA,EAAgC,OAEhDM,CAAAA,CAAa,IAAI,GAAA,CAEjBC,CAAAA,CAAO,MAAMX,CAAAA,CAASK,CAAAA,CAAwCG,CAAAA,CAAO,CACzE,OAAQF,CAAAA,GAAW,QAAA,CAAW,QAAA,CAAW,MAAA,CACzC,MAAAG,CAAAA,CACA,iBAAA,CAAoBG,GAAkC,CACpD,GAAIA,EAAK,MAAA,GAAW,UAAA,CAAY,CAC9B,IAAMC,EAAQD,CAAAA,CAAK,IAAA,EAAmB,SAAA,CAChCE,CAAAA,CAASF,EAAK,KAAA,EAAoB,CAAA,CACxCF,CAAAA,CAAW,GAAA,CAAIG,EAAMC,CAAK,CAAA,CAC1BP,CAAAA,CAAW,CACT,OAASK,CAAAA,CAAK,MAAA,EAAqB,CAAA,CACnC,KAAA,CAAOC,EACP,KAAA,CAAAC,CACF,CAAC,EACH,MACSF,CAAAA,CAAK,MAAA,GAAW,UAAA,CACvBL,CAAAA,CAAW,CAAE,MAAA,CAAQ,CAAA,CAAG,KAAA,CAAO,UAAA,CAAY,MAAO,CAAE,CAAC,CAAA,CAE9CK,CAAAA,CAAK,SAAW,MAAA,EACvBL,CAAAA,CAAW,CAAE,MAAA,CAAQ,EAAG,KAAA,CAAO,MAAA,CAAQ,KAAA,CAAO,CAAE,CAAC,EAErD,CACF,CAAC,CAAA,CAEGQ,EAAa,CAAA,CACjB,IAAA,IAAWC,CAAAA,IAAQN,CAAAA,CAAW,QAAO,CAAGK,CAAAA,EAAcC,CAAAA,CAEtD,OAAO,CACL,QAAA,CAAUL,CAAAA,CACV,QAAA,CAAU,IAAA,CAAK,KAAKI,CAAAA,EAAc,IAAA,CAAO,KAAK,CAAA,EAAK,GACrD,CACF,CAAA,CAEA,MAAM,GAAA,CAAIP,CAAAA,CAAoBS,EAAgBC,CAAAA,CAAqC,CACjF,IAAMP,CAAAA,CAAOH,EAAM,QAAA,CACbW,CAAAA,CAAS,KAAA,CAAM,OAAA,CAAQF,CAAK,CAAA,CAC9B,MAAMN,CAAAA,CAAK,GAAIM,EAAmCC,CAAAA,EAAW,EAAE,CAAA,CAC/D,MAAMP,CAAAA,CAAKM,CAAAA,CAAOC,CAAAA,EAAW,EAAE,CAAA,CACnC,OAAOE,CAAAA,CAAgBD,CAAM,CAC/B,CAAA,CAEA,MAAM,MAAA,CACJX,CAAAA,CACAS,EACAI,CAAAA,CACAH,CAAAA,CACe,CACf,IAAMP,EAAOH,CAAAA,CAAM,QAAA,CAEbc,CAAAA,CAAW,IAAIrB,EAAaU,CAAAA,CAAK,SAAA,CAAW,CAChD,iBAAA,CAAoBY,GAAiBF,CAAAA,CAAQE,CAAI,CAAA,CACjD,WAAA,CAAa,KACb,mBAAA,CAAqB,IACvB,CAAC,CAAA,CAED,MAAMZ,CAAAA,CAAKM,CAAAA,CAAO,CAChB,GAAIC,GAAqB,EAAC,CAC1B,QAAA,CAAAI,CACF,CAAC,EACH,CAAA,CAEA,MAAM,MAAA,CAAOd,CAAAA,CAAmC,CAE9C,MADaA,CAAAA,CAAM,QAAA,CACR,OAAA,KACb,CACF,CACF,CACF,CACF,CAMO,SAASY,CAAAA,CAAgBI,CAAAA,CAAyB,CACvD,GAAIA,CAAAA,GAAU,IAAA,EAAQ,OAAOA,CAAAA,EAAU,SACrC,OAAOA,CAAAA,CAET,GAAI,KAAA,CAAM,QAAQA,CAAK,CAAA,CACrB,OAAOA,CAAAA,CAAM,IAAIJ,CAAe,CAAA,CAElC,IAAMK,CAAAA,CAAMD,EAEZ,OAAI,OAAOC,CAAAA,CAAI,IAAA,CAAS,KAAe,KAAA,CAAM,OAAA,CAAQA,CAAAA,CAAI,IAAI,EACpD,CACL,IAAA,CAAM,KAAA,CAAM,IAAA,CAAKA,EAAI,IAAyB,CAAA,CAC9C,IAAA,CAAMA,CAAAA,CAAI,KACV,IAAA,CAAMA,CAAAA,CAAI,IAAA,CACV,IAAA,CAAMA,EAAI,IACZ,CAAA,CAGK,MAAA,CAAO,WAAA,CACZ,OAAO,OAAA,CAAQA,CAAG,CAAA,CAAE,GAAA,CAAI,CAAC,CAACC,CAAAA,CAAGC,CAAC,CAAA,GAAM,CAACD,CAAAA,CAAGN,CAAAA,CAAgBO,CAAC,CAAC,CAAC,CAC7D,CACF","file":"transformers.js","sourcesContent":["import type {\n  Device,\n  LoadedModel,\n  LoadProgressEvent,\n  ModelAdapter,\n  ModelAdapterFactory,\n} from '../core/types.js';\n\n// eslint-disable-next-line ts/no-explicit-any\ntype AnyPipeline = any;\n\n/**\n * Adapter for @huggingface/transformers v3+.\n *\n * @remarks\n * The adapter is instantiated INSIDE the worker via `create()`.\n * The heavy `@huggingface/transformers` library is dynamically imported\n * inside the worker, keeping the main thread bundle lightweight.\n *\n * @example\n * ```ts\n * import { createPool } from 'inferis-ml';\n * import { transformersAdapter } from 'inferis-ml/adapters/transformers';\n *\n * const pool = await createPool({ adapter: transformersAdapter() });\n * const model = await pool.load<number[][]>('feature-extraction', {\n *   model: 'mixedbread-ai/mxbai-embed-xsmall-v1',\n * });\n * ```\n */\nexport function transformersAdapter(): ModelAdapterFactory {\n  return {\n    name: 'transformers',\n\n    async create(): Promise<ModelAdapter> {\n      // Dynamic import — runs inside the worker only\n      // @ts-expect-error - optional peer dependency, resolved at runtime inside worker\n      const { pipeline, TextStreamer, env } = await import('@huggingface/transformers');\n\n      // Use ONNX WASM backend with optimizations\n      env.backends.onnx.wasm.proxy = false;\n\n      return {\n        name: 'transformers',\n\n        estimateMemoryMB(_task: string, config: Record<string, unknown>): number {\n          return (config.estimatedMemoryMB as number | undefined) ?? 100;\n        },\n\n        async load(\n          task: string,\n          config: Record<string, unknown>,\n          device: Device,\n          onProgress: (event: LoadProgressEvent) => void,\n        ): Promise<LoadedModel> {\n          const model = config.model as string;\n          const dtype = (config.dtype as string | undefined) ?? 'fp32';\n\n          const fileTotals = new Map<string, number>();\n\n          const pipe = await pipeline(task as Parameters<typeof pipeline>[0], model, {\n            device: device === 'webgpu' ? 'webgpu' : 'wasm',\n            dtype,\n            progress_callback: (info: Record<string, unknown>) => {\n              if (info.status === 'progress') {\n                const file = (info.file as string) ?? 'unknown';\n                const total = (info.total as number) ?? 0;\n                fileTotals.set(file, total);\n                onProgress({\n                  loaded: (info.loaded as number) ?? 0,\n                  phase: file,\n                  total,\n                });\n              }\n              else if (info.status === 'initiate') {\n                onProgress({ loaded: 0, phase: 'initiate', total: 0 });\n              }\n              else if (info.status === 'done') {\n                onProgress({ loaded: 1, phase: 'done', total: 1 });\n              }\n            },\n          });\n\n          let totalBytes = 0;\n          for (const size of fileTotals.values()) totalBytes += size;\n\n          return {\n            instance: pipe,\n            memoryMB: Math.ceil(totalBytes / (1024 * 1024)) || 100,\n          };\n        },\n\n        async run(model: LoadedModel, input: unknown, options?: unknown): Promise<unknown> {\n          const pipe = model.instance as AnyPipeline;\n          const result = Array.isArray(input)\n            ? await pipe(...(input as [unknown, ...unknown[]]), options ?? {})\n            : await pipe(input, options ?? {});\n          return serializeOutput(result);\n        },\n\n        async stream(\n          model: LoadedModel,\n          input: unknown,\n          onChunk: (chunk: unknown) => void,\n          options?: unknown,\n        ): Promise<void> {\n          const pipe = model.instance as AnyPipeline;\n\n          const streamer = new TextStreamer(pipe.tokenizer, {\n            callback_function: (text: string) => onChunk(text),\n            skip_prompt: true,\n            skip_special_tokens: true,\n          });\n\n          await pipe(input, {\n            ...(options as object ?? {}),\n            streamer,\n          });\n        },\n\n        async unload(model: LoadedModel): Promise<void> {\n          const pipe = model.instance as AnyPipeline;\n          await pipe.dispose?.();\n        },\n      };\n    },\n  };\n}\n\n/**\n * Recursively converts Tensor class instances and typed arrays to plain\n * structured-clone-compatible objects so they can cross the worker boundary.\n */\nexport function serializeOutput(value: unknown): unknown {\n  if (value === null || typeof value !== 'object')\n    return value;\n\n  if (Array.isArray(value))\n    return value.map(serializeOutput);\n\n  const obj = value as Record<string, unknown>;\n\n  if (typeof obj.data !== 'undefined' && Array.isArray(obj.dims)) {\n    return {\n      data: Array.from(obj.data as ArrayLike<number>),\n      dims: obj.dims as number[],\n      size: obj.size,\n      type: obj.type,\n    };\n  }\n\n  return Object.fromEntries(\n    Object.entries(obj).map(([k, v]) => [k, serializeOutput(v)]),\n  );\n}\n"]}

package/dist/adapters/web-llm.cjs

@@ -0,0 +1,2 @@
+'use strict';function m(){return {name:"web-llm",async create(){let{CreateMLCEngine:c}=await import('@mlc-ai/web-llm');return {name:"web-llm",estimateMemoryMB(n,e){return e.estimatedMemoryMB??2e3},async load(n,e,a,t){let s=e.model;return {instance:await c(s,{initProgressCallback:o=>{t({loaded:o.progress,phase:o.text,total:1});}}),memoryMB:e.estimatedMemoryMB??2e3}},async run(n,e,a){let t=n.instance,s=e;return (await t.chat.completions.create({...a??{},messages:s.messages,stream:false})).choices[0]?.message?.content??""},async stream(n,e,a,t){let s=n.instance,r=e,o=await s.chat.completions.create({...t??{},messages:r.messages,stream:true});for await(let d of o){let i=d.choices[0]?.delta?.content;i&&a(i);}},async unload(n){await n.instance.unload?.();}}}}}exports.webLlmAdapter=m;//# sourceMappingURL=web-llm.cjs.map
+//# sourceMappingURL=web-llm.cjs.map

package/dist/adapters/web-llm.cjs.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../src/adapters/web-llm.ts"],"names":["webLlmAdapter","CreateMLCEngine","_task","config","_device","onProgress","modelId","info","model","input","options","engine","chatInput","onChunk","asyncChunks","chunk","delta"],"mappings":"aAiDO,SAASA,GAAqC,CACnD,OAAO,CACL,IAAA,CAAM,UAEN,MAAM,MAAA,EAAgC,CAEpC,GAAM,CAAE,eAAA,CAAAC,CAAgB,EAAI,MAAM,OAAO,iBAAiB,CAAA,CAE1D,OAAO,CACL,IAAA,CAAM,UAEN,gBAAA,CAAiBC,CAAAA,CAAeC,EAAyC,CACvE,OAAQA,EAAO,iBAAA,EAA4C,GAC7D,CAAA,CAEA,MAAM,KACJD,CAAAA,CACAC,CAAAA,CACAC,EACAC,CAAAA,CACsB,CACtB,IAAMC,CAAAA,CAAUH,CAAAA,CAAO,KAAA,CAYvB,OAAO,CACL,QAAA,CAXwB,MAAMF,EAAgBK,CAAAA,CAAS,CACvD,qBAAuBC,CAAAA,EAA6C,CAClEF,CAAAA,CAAW,CACT,OAAQE,CAAAA,CAAK,QAAA,CACb,MAAOA,CAAAA,CAAK,IAAA,CACZ,MAAO,CACT,CAAC,EACH,CACF,CAAC,CAAA,CAIC,QAAA,CAAWJ,EAAO,iBAAA,EAA4C,GAChE,CACF,CAAA,CAEA,MAAM,GAAA,CAAIK,CAAAA,CAAoBC,EAAgBC,CAAAA,CAAqC,CACjF,IAAMC,CAAAA,CAASH,CAAAA,CAAM,SACfI,CAAAA,CAAYH,CAAAA,CAMlB,OAAA,CALiB,MAAME,EAAO,IAAA,CAAK,WAAA,CAAY,OAAO,CACpD,GAAID,GAAqB,EAAC,CAC1B,QAAA,CAAUE,CAAAA,CAAU,SACpB,MAAA,CAAQ,KACV,CAAC,CAAA,EACe,OAAA,CAAQ,CAAC,CAAA,EAAG,OAAA,EAAS,OAAA,EAAW,EAClD,EAEA,MAAM,MAAA,CACJJ,EACAC,CAAAA,CACAI,CAAAA,CACAH,EACe,CACf,IAAMC,CAAAA,CAASH,CAAAA,CAAM,SACfI,CAAAA,CAAYH,CAAAA,CAEZK,EAAc,MAAMH,CAAAA,CAAO,KAAK,WAAA,CAAY,MAAA,CAAO,CACvD,GAAID,GAAqB,EAAC,CAC1B,SAAUE,CAAAA,CAAU,QAAA,CACpB,OAAQ,IACV,CAAC,CAAA,CAED,UAAA,IAAiBG,KAASD,CAAAA,CAAa,CACrC,IAAME,CAAAA,CAAQD,CAAAA,CAAM,QAAQ,CAAC,CAAA,EAAG,KAAA,EAAO,OAAA,CACnCC,GACFH,CAAAA,CAAQG,CAAK,EACjB,CACF,CAAA,CAEA,MAAM,MAAA,CAAOR,CAAAA,CAAmC,CAE9C,MADeA,EAAM,QAAA,CACR,MAAA,KACf,CACF,CACF,CACF,CACF","file":"web-llm.cjs","sourcesContent":["import type {\n  Device,\n  LoadedModel,\n  LoadProgressEvent,\n  ModelAdapter,\n  ModelAdapterFactory,\n} from '../core/types.js';\n\n// eslint-disable-next-line ts/no-explicit-any\ntype MLCEngine = any;\n\ninterface ChatMessage {\n  role: 'system' | 'user' | 'assistant';\n  content: string;\n}\n\ninterface ChatInput {\n  messages: ChatMessage[];\n  temperature?: number;\n  top_p?: number;\n  max_tokens?: number;\n}\n\n/**\n * Adapter for @mlc-ai/web-llm.\n *\n * @remarks\n * Optimized for LLM text generation with WebGPU acceleration.\n * Streaming is natively supported via the OpenAI-compatible chat API.\n *\n * @example\n * ```ts\n * import { createPool } from 'inferis-ml';\n * import { webLlmAdapter } from 'inferis-ml/adapters/web-llm';\n *\n * const pool = await createPool({\n *   adapter: webLlmAdapter(),\n *   defaultDevice: 'webgpu',\n *   maxWorkers: 1,\n * });\n *\n * const llm = await pool.load<string>('text-generation', {\n *   model: 'Llama-3.1-8B-Instruct-q4f32_1-MLC',\n * });\n *\n * const stream = llm.stream({ messages: [{ role: 'user', content: 'Hello' }] });\n * for await (const token of stream) { ... }\n * ```\n */\nexport function webLlmAdapter(): ModelAdapterFactory {\n  return {\n    name: 'web-llm',\n\n    async create(): Promise<ModelAdapter> {\n      // @ts-expect-error - optional peer dependency, resolved at runtime inside worker\n      const { CreateMLCEngine } = await import('@mlc-ai/web-llm');\n\n      return {\n        name: 'web-llm',\n\n        estimateMemoryMB(_task: string, config: Record<string, unknown>): number {\n          return (config.estimatedMemoryMB as number | undefined) ?? 2000;\n        },\n\n        async load(\n          _task: string,\n          config: Record<string, unknown>,\n          _device: Device,\n          onProgress: (event: LoadProgressEvent) => void,\n        ): Promise<LoadedModel> {\n          const modelId = config.model as string;\n\n          const engine: MLCEngine = await CreateMLCEngine(modelId, {\n            initProgressCallback: (info: { progress: number; text: string }) => {\n              onProgress({\n                loaded: info.progress,\n                phase: info.text,\n                total: 1,\n              });\n            },\n          });\n\n          return {\n            instance: engine,\n            memoryMB: (config.estimatedMemoryMB as number | undefined) ?? 2000,\n          };\n        },\n\n        async run(model: LoadedModel, input: unknown, options?: unknown): Promise<unknown> {\n          const engine = model.instance as MLCEngine;\n          const chatInput = input as ChatInput;\n          const response = await engine.chat.completions.create({\n            ...(options as object ?? {}),\n            messages: chatInput.messages,\n            stream: false,\n          });\n          return response.choices[0]?.message?.content ?? '';\n        },\n\n        async stream(\n          model: LoadedModel,\n          input: unknown,\n          onChunk: (chunk: unknown) => void,\n          options?: unknown,\n        ): Promise<void> {\n          const engine = model.instance as MLCEngine;\n          const chatInput = input as ChatInput;\n\n          const asyncChunks = await engine.chat.completions.create({\n            ...(options as object ?? {}),\n            messages: chatInput.messages,\n            stream: true,\n          });\n\n          for await (const chunk of asyncChunks) {\n            const delta = chunk.choices[0]?.delta?.content;\n            if (delta)\n              onChunk(delta);\n          }\n        },\n\n        async unload(model: LoadedModel): Promise<void> {\n          const engine = model.instance as MLCEngine;\n          await engine.unload?.();\n        },\n      };\n    },\n  };\n}\n"]}

package/dist/adapters/web-llm.d.cts

@@ -0,0 +1,31 @@
+import { h as ModelAdapterFactory } from '../types-Y6Ytjh7U.cjs';
+
+/**
+ * Adapter for @mlc-ai/web-llm.
+ *
+ * @remarks
+ * Optimized for LLM text generation with WebGPU acceleration.
+ * Streaming is natively supported via the OpenAI-compatible chat API.
+ *
+ * @example
+ * ```ts
+ * import { createPool } from 'inferis-ml';
+ * import { webLlmAdapter } from 'inferis-ml/adapters/web-llm';
+ *
+ * const pool = await createPool({
+ *   adapter: webLlmAdapter(),
+ *   defaultDevice: 'webgpu',
+ *   maxWorkers: 1,
+ * });
+ *
+ * const llm = await pool.load<string>('text-generation', {
+ *   model: 'Llama-3.1-8B-Instruct-q4f32_1-MLC',
+ * });
+ *
+ * const stream = llm.stream({ messages: [{ role: 'user', content: 'Hello' }] });
+ * for await (const token of stream) { ... }
+ * ```
+ */
+declare function webLlmAdapter(): ModelAdapterFactory;
+
+export { webLlmAdapter };

package/dist/adapters/web-llm.d.ts

@@ -0,0 +1,31 @@
+import { h as ModelAdapterFactory } from '../types-Y6Ytjh7U.js';
+
+/**
+ * Adapter for @mlc-ai/web-llm.
+ *
+ * @remarks
+ * Optimized for LLM text generation with WebGPU acceleration.
+ * Streaming is natively supported via the OpenAI-compatible chat API.
+ *
+ * @example
+ * ```ts
+ * import { createPool } from 'inferis-ml';
+ * import { webLlmAdapter } from 'inferis-ml/adapters/web-llm';
+ *
+ * const pool = await createPool({
+ *   adapter: webLlmAdapter(),
+ *   defaultDevice: 'webgpu',
+ *   maxWorkers: 1,
+ * });
+ *
+ * const llm = await pool.load<string>('text-generation', {
+ *   model: 'Llama-3.1-8B-Instruct-q4f32_1-MLC',
+ * });
+ *
+ * const stream = llm.stream({ messages: [{ role: 'user', content: 'Hello' }] });
+ * for await (const token of stream) { ... }
+ * ```
+ */
+declare function webLlmAdapter(): ModelAdapterFactory;
+
+export { webLlmAdapter };

package/dist/adapters/web-llm.js

@@ -0,0 +1,2 @@
+function m(){return {name:"web-llm",async create(){let{CreateMLCEngine:c}=await import('@mlc-ai/web-llm');return {name:"web-llm",estimateMemoryMB(n,e){return e.estimatedMemoryMB??2e3},async load(n,e,a,t){let s=e.model;return {instance:await c(s,{initProgressCallback:o=>{t({loaded:o.progress,phase:o.text,total:1});}}),memoryMB:e.estimatedMemoryMB??2e3}},async run(n,e,a){let t=n.instance,s=e;return (await t.chat.completions.create({...a??{},messages:s.messages,stream:false})).choices[0]?.message?.content??""},async stream(n,e,a,t){let s=n.instance,r=e,o=await s.chat.completions.create({...t??{},messages:r.messages,stream:true});for await(let d of o){let i=d.choices[0]?.delta?.content;i&&a(i);}},async unload(n){await n.instance.unload?.();}}}}}export{m as webLlmAdapter};//# sourceMappingURL=web-llm.js.map
+//# sourceMappingURL=web-llm.js.map

package/dist/adapters/web-llm.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../../src/adapters/web-llm.ts"],"names":["webLlmAdapter","CreateMLCEngine","_task","config","_device","onProgress","modelId","info","model","input","options","engine","chatInput","onChunk","asyncChunks","chunk","delta"],"mappings":"AAiDO,SAASA,GAAqC,CACnD,OAAO,CACL,IAAA,CAAM,UAEN,MAAM,MAAA,EAAgC,CAEpC,GAAM,CAAE,eAAA,CAAAC,CAAgB,EAAI,MAAM,OAAO,iBAAiB,CAAA,CAE1D,OAAO,CACL,IAAA,CAAM,UAEN,gBAAA,CAAiBC,CAAAA,CAAeC,EAAyC,CACvE,OAAQA,EAAO,iBAAA,EAA4C,GAC7D,CAAA,CAEA,MAAM,KACJD,CAAAA,CACAC,CAAAA,CACAC,EACAC,CAAAA,CACsB,CACtB,IAAMC,CAAAA,CAAUH,CAAAA,CAAO,KAAA,CAYvB,OAAO,CACL,QAAA,CAXwB,MAAMF,EAAgBK,CAAAA,CAAS,CACvD,qBAAuBC,CAAAA,EAA6C,CAClEF,CAAAA,CAAW,CACT,OAAQE,CAAAA,CAAK,QAAA,CACb,MAAOA,CAAAA,CAAK,IAAA,CACZ,MAAO,CACT,CAAC,EACH,CACF,CAAC,CAAA,CAIC,QAAA,CAAWJ,EAAO,iBAAA,EAA4C,GAChE,CACF,CAAA,CAEA,MAAM,GAAA,CAAIK,CAAAA,CAAoBC,EAAgBC,CAAAA,CAAqC,CACjF,IAAMC,CAAAA,CAASH,CAAAA,CAAM,SACfI,CAAAA,CAAYH,CAAAA,CAMlB,OAAA,CALiB,MAAME,EAAO,IAAA,CAAK,WAAA,CAAY,OAAO,CACpD,GAAID,GAAqB,EAAC,CAC1B,QAAA,CAAUE,CAAAA,CAAU,SACpB,MAAA,CAAQ,KACV,CAAC,CAAA,EACe,OAAA,CAAQ,CAAC,CAAA,EAAG,OAAA,EAAS,OAAA,EAAW,EAClD,EAEA,MAAM,MAAA,CACJJ,EACAC,CAAAA,CACAI,CAAAA,CACAH,EACe,CACf,IAAMC,CAAAA,CAASH,CAAAA,CAAM,SACfI,CAAAA,CAAYH,CAAAA,CAEZK,EAAc,MAAMH,CAAAA,CAAO,KAAK,WAAA,CAAY,MAAA,CAAO,CACvD,GAAID,GAAqB,EAAC,CAC1B,SAAUE,CAAAA,CAAU,QAAA,CACpB,OAAQ,IACV,CAAC,CAAA,CAED,UAAA,IAAiBG,KAASD,CAAAA,CAAa,CACrC,IAAME,CAAAA,CAAQD,CAAAA,CAAM,QAAQ,CAAC,CAAA,EAAG,KAAA,EAAO,OAAA,CACnCC,GACFH,CAAAA,CAAQG,CAAK,EACjB,CACF,CAAA,CAEA,MAAM,MAAA,CAAOR,CAAAA,CAAmC,CAE9C,MADeA,EAAM,QAAA,CACR,MAAA,KACf,CACF,CACF,CACF,CACF","file":"web-llm.js","sourcesContent":["import type {\n  Device,\n  LoadedModel,\n  LoadProgressEvent,\n  ModelAdapter,\n  ModelAdapterFactory,\n} from '../core/types.js';\n\n// eslint-disable-next-line ts/no-explicit-any\ntype MLCEngine = any;\n\ninterface ChatMessage {\n  role: 'system' | 'user' | 'assistant';\n  content: string;\n}\n\ninterface ChatInput {\n  messages: ChatMessage[];\n  temperature?: number;\n  top_p?: number;\n  max_tokens?: number;\n}\n\n/**\n * Adapter for @mlc-ai/web-llm.\n *\n * @remarks\n * Optimized for LLM text generation with WebGPU acceleration.\n * Streaming is natively supported via the OpenAI-compatible chat API.\n *\n * @example\n * ```ts\n * import { createPool } from 'inferis-ml';\n * import { webLlmAdapter } from 'inferis-ml/adapters/web-llm';\n *\n * const pool = await createPool({\n *   adapter: webLlmAdapter(),\n *   defaultDevice: 'webgpu',\n *   maxWorkers: 1,\n * });\n *\n * const llm = await pool.load<string>('text-generation', {\n *   model: 'Llama-3.1-8B-Instruct-q4f32_1-MLC',\n * });\n *\n * const stream = llm.stream({ messages: [{ role: 'user', content: 'Hello' }] });\n * for await (const token of stream) { ... }\n * ```\n */\nexport function webLlmAdapter(): ModelAdapterFactory {\n  return {\n    name: 'web-llm',\n\n    async create(): Promise<ModelAdapter> {\n      // @ts-expect-error - optional peer dependency, resolved at runtime inside worker\n      const { CreateMLCEngine } = await import('@mlc-ai/web-llm');\n\n      return {\n        name: 'web-llm',\n\n        estimateMemoryMB(_task: string, config: Record<string, unknown>): number {\n          return (config.estimatedMemoryMB as number | undefined) ?? 2000;\n        },\n\n        async load(\n          _task: string,\n          config: Record<string, unknown>,\n          _device: Device,\n          onProgress: (event: LoadProgressEvent) => void,\n        ): Promise<LoadedModel> {\n          const modelId = config.model as string;\n\n          const engine: MLCEngine = await CreateMLCEngine(modelId, {\n            initProgressCallback: (info: { progress: number; text: string }) => {\n              onProgress({\n                loaded: info.progress,\n                phase: info.text,\n                total: 1,\n              });\n            },\n          });\n\n          return {\n            instance: engine,\n            memoryMB: (config.estimatedMemoryMB as number | undefined) ?? 2000,\n          };\n        },\n\n        async run(model: LoadedModel, input: unknown, options?: unknown): Promise<unknown> {\n          const engine = model.instance as MLCEngine;\n          const chatInput = input as ChatInput;\n          const response = await engine.chat.completions.create({\n            ...(options as object ?? {}),\n            messages: chatInput.messages,\n            stream: false,\n          });\n          return response.choices[0]?.message?.content ?? '';\n        },\n\n        async stream(\n          model: LoadedModel,\n          input: unknown,\n          onChunk: (chunk: unknown) => void,\n          options?: unknown,\n        ): Promise<void> {\n          const engine = model.instance as MLCEngine;\n          const chatInput = input as ChatInput;\n\n          const asyncChunks = await engine.chat.completions.create({\n            ...(options as object ?? {}),\n            messages: chatInput.messages,\n            stream: true,\n          });\n\n          for await (const chunk of asyncChunks) {\n            const delta = chunk.choices[0]?.delta?.content;\n            if (delta)\n              onChunk(delta);\n          }\n        },\n\n        async unload(model: LoadedModel): Promise<void> {\n          const engine = model.instance as MLCEngine;\n          await engine.unload?.();\n        },\n      };\n    },\n  };\n}\n"]}

package/dist/index.cjs

@@ -0,0 +1,2 @@
+'use strict';var _documentCurrentScript=typeof document!=='undefined'?document.currentScript:null;var E="inferis:leader",S=class o{role="unknown";listeners=new Set;abortController=null;async start(){if(!this.isSupported())return this.setRole("leader"),"leader";this.abortController=new AbortController;let e=false;return new Promise(t=>{let r=setTimeout(()=>{e||(e=true,this.setRole("leader"),t("leader"));},5e3);navigator.locks.request(E,{signal:this.abortController.signal},async()=>{e||(e=true,clearTimeout(r),this.setRole("leader"),t("leader")),await this.holdLock();}).catch(n=>{n.name;}),navigator.locks.query().then(n=>{if(e)return;let s=n.held?.some(i=>i.name===E)??false,d=n.pending?.some(i=>i.name===E)??false;(s||d)&&(e=true,clearTimeout(r),this.setRole("follower"),t("follower"));}).catch(()=>{});})}holdLock(){return new Promise(e=>{this.abortController&&this.abortController.signal.addEventListener("abort",()=>e(),{once:true});})}stop(){this.abortController?.abort(),this.abortController=null;}onRoleChange(e){return this.listeners.add(e),()=>this.listeners.delete(e)}get currentRole(){return this.role}get isLeader(){return this.role==="leader"}setRole(e){this.role=e;for(let t of this.listeners)try{t(e);}catch{}}static isSupported(){return typeof navigator<"u"&&"locks"in navigator}isSupported(){return o.isSupported()}};var q=class{port;listeners=new Set;constructor(e){let t=new SharedWorker(e,{type:"module"});this.port=t.port,t.onerror=r=>{console.error("[SharedWorkerBridge] worker error:",r.message);},this.port.onmessage=r=>{for(let n of this.listeners)try{n(r.data);}catch{}},this.port.onmessageerror=r=>{console.error("[SharedWorkerBridge] port deserialization error:",r);},this.port.start();}postMessage(e,t){t?.length?this.port.postMessage(e,t):this.port.postMessage(e);}on(e){return this.listeners.add(e),()=>this.listeners.delete(e)}disconnect(){this.listeners.clear(),this.port.close();}static isSupported(){return typeof SharedWorker<"u"}};var $="inferis:bus",j=new Set(["leader-elected","leader-gone","request","response","stream-chunk","stream-end","stream-error"]);function H(o){return typeof o=="object"&&o!==null&&"type"in o&&typeof o.type=="string"&&j.has(o.type)}var R=class{channel;listeners=new Set;constructor(){this.channel=new BroadcastChannel($),this.channel.onmessage=e=>{if(H(e.data))for(let t of this.listeners)try{t(e.data);}catch{}};}send(e){this.channel.postMessage(e);}on(e){return this.listeners.add(e),()=>this.listeners.delete(e)}close(){this.listeners.clear(),this.channel.close();}static isSupported(){return typeof BroadcastChannel<"u"}};var m=class{maxMB;usedMB=0;lruOrder=[];modelSizes=new Map;constructor(e){if(e<=0)throw new RangeError("maxMB must be > 0");this.maxMB=e;}get totalMB(){return this.maxMB}get allocatedMB(){return this.usedMB}get availableMB(){return this.maxMB-this.usedMB}planEviction(e){if(e>this.maxMB)return null;if(e<=this.availableMB)return [];let t=[],r=0;for(let n of this.lruOrder){if(r>=e-this.availableMB)break;let s=this.modelSizes.get(n);s!==void 0&&(t.push(n),r+=s);}return r>=e-this.availableMB?t:null}allocate(e,t){this.modelSizes.has(e)&&this.release(e),this.modelSizes.set(e,t),this.usedMB+=t,this.touch(e);}release(e){let t=this.modelSizes.get(e);if(t===void 0)return;this.modelSizes.delete(e),this.usedMB=Math.max(0,this.usedMB-t);let r=this.lruOrder.indexOf(e);r!==-1&&this.lruOrder.splice(r,1);}touch(e){let t=this.lruOrder.indexOf(e);t!==-1&&this.lruOrder.splice(t,1),this.lruOrder.push(e);}fits(e){return e<=this.availableMB}get lruList(){return this.lruOrder}};var K=new Uint8Array([0,97,115,109,1,0,0,0,1,5,1,96,0,1,123,3,2,1,0,10,10,1,8,0,65,0,253,15,0,0,11]),G=new Uint8Array([0,97,115,109,1,0,0,0,1,4,1,96,0,0,3,2,1,0,5,4,1,3,1,1,10,11,1,9,0,254,1,2,0,65,0,11]);function Y(){return typeof WebAssembly>"u"?{simd:false,supported:false,threads:false}:{simd:WebAssembly.validate(K),supported:true,threads:WebAssembly.validate(G)}}async function Z(){if(typeof navigator>"u"||!("gpu"in navigator))return {adapter:null,isFallback:false,limits:null,supported:false};let o=navigator.gpu;if(o==null)return {adapter:null,isFallback:false,limits:null,supported:false};try{let e=await o.requestAdapter();if(!e)return {adapter:null,isFallback:!1,limits:null,supported:!1};let t=await e.requestAdapterInfo?.()??{},r=e.isFallbackAdapter??!1;return {adapter:{architecture:t.architecture??"",description:t.description??"",device:t.device??"",vendor:t.vendor??""},isFallback:r,limits:{maxBufferSize:e.limits?.maxBufferSize??0,maxStorageBufferBindingSize:e.limits?.maxStorageBufferBindingSize??0},supported:!0}}catch{return {adapter:null,isFallback:false,limits:null,supported:false}}}function V(){return typeof SharedWorker<"u"}function X(){return typeof BroadcastChannel<"u"}function J(){return typeof navigator<"u"&&"locks"in navigator}var b=null;async function A(){if(b)return b;let[o,e]=await Promise.all([Z(),Promise.resolve(Y())]);return b=Object.freeze({broadcastChannel:X(),hardwareConcurrency:typeof navigator<"u"?navigator.hardwareConcurrency??1:1,sharedWorker:V(),wasm:e,webgpu:o,webLocks:J()}),b}function Q(){b=null;}var u=class o extends Error{code;constructor(e,t){super(e),this.name="InferisError",this.code=t;}static fromSerialized(e){let t=new o(e.message,e.code??"UNKNOWN");return t.name=e.name,e.stack&&(t.stack=e.stack),t}serialize(){let e={code:this.code,message:this.message,name:this.name};return this.stack!==void 0?{...e,stack:this.stack}:e}},L=class extends u{modelId;constructor(e,t){super(t,"MODEL_LOAD_ERROR"),this.name="ModelLoadError",this.modelId=e;}},f=class extends u{modelId;constructor(e,t){super(`Model "${e}" is not ready (state: ${t})`,"MODEL_NOT_READY"),this.name="ModelNotReadyError",this.modelId=e;}},C=class extends u{modelId;constructor(e){super(`Model "${e}" has been disposed`,"MODEL_DISPOSED"),this.name="ModelDisposedError",this.modelId=e;}},W=class extends u{modelId;constructor(e,t){super(t,"INFERENCE_ERROR"),this.name="InferenceError",this.modelId=e;}},v=class extends u{requestedMB;budgetMB;constructor(e,t){super(`Cannot load model: requested ${e}MB exceeds memory budget ${t}MB`,"BUDGET_EXCEEDED"),this.name="BudgetExceededError",this.budgetMB=t,this.requestedMB=e;}},M=class extends u{reqId;constructor(e,t){super(`Task "${e}" timed out after ${t}ms`,"TASK_TIMEOUT"),this.name="TaskTimeoutError",this.reqId=e;}},P=class extends u{workerId;constructor(e,t){super(t,"WORKER_ERROR"),this.name="WorkerError",this.workerId=e;}},O=class extends u{modelId;reason;constructor(e,t){super(`GPU device lost for model "${e}": ${t}`,"DEVICE_LOST"),this.name="DeviceLostError",this.modelId=e,this.reason=t;}},w=class extends u{constructor(e,t){super(`Invalid state transition: ${e} -> ${t}`,"INVALID_STATE_TRANSITION"),this.name="InvalidStateTransitionError";}};var D={disposed:new Set,error:new Set(["loading","disposed"]),idle:new Set(["loading","disposed"]),inferring:new Set(["inferring","ready","error","unloading"]),loading:new Set(["ready","error","disposed"]),ready:new Set(["inferring","unloading","error"]),unloading:new Set(["disposed"])};function p(o,e){if(!D[o].has(e))throw new w(o,e);return e}function k(o,e){return D[o].has(e)}function I(o){return o==="ready"}function ee(o){return o==="disposed"}var y=class{models=new Map;static makeId(e,t){return `${e}:${t}`}register(e,t,r){if(this.models.has(e))return this.models.get(e);let n={config:r,device:"wasm",id:e,memoryMB:0,state:"idle",stateListeners:new Set,task:t,workerId:null};return this.models.set(e,n),n}get(e){return this.models.get(e)}has(e){return this.models.has(e)}setState(e,t){let r=this.models.get(e);if(r){r.state=t;for(let n of r.stateListeners)try{n(t);}catch{}}}setLoaded(e,t,r,n){let s=this.models.get(e);s&&(s.device=t,s.memoryMB=r,s.workerId=n);}setUnloaded(e){let t=this.models.get(e);t&&(t.workerId=null,t.memoryMB=0);}subscribe(e,t){let r=this.models.get(e);return r?(r.stateListeners.add(t),()=>r.stateListeners.delete(t)):()=>{}}delete(e){let t=this.models.get(e);t&&(t.stateListeners.clear(),this.models.delete(e));}byState(e){return [...this.models.values()].filter(t=>t.state===e)}byWorker(e){return [...this.models.values()].filter(t=>t.workerId===e)}get size(){return this.models.size}get entries(){return this.models}};var te={high:2,low:0,normal:1},T=class{queues=new Map([["high",[]],["low",[]],["normal",[]]]);workerLoad=new Map;workerModels=new Map;addWorker(e){this.workerLoad.set(e,0),this.workerModels.set(e,new Set);}removeWorker(e){this.workerLoad.delete(e),this.workerModels.delete(e);}notifyModelLoaded(e,t){this.workerModels.get(e)?.add(t);}notifyModelUnloaded(e,t){this.workerModels.get(e)?.delete(t);}enqueue(e,t){let r=this.pickWorker(e.modelId);r!==null?this.dispatch(e,r):this.queues.get(e.priority).push(e);}notifyTaskComplete(e,t){let r=this.workerLoad.get(e);r!==void 0&&this.workerLoad.set(e,Math.max(0,r-1)),this.drainNext(e);}drainNext(e){for(let t of ["high","normal","low"]){let r=this.queues.get(t);if(r.length===0)continue;let n=this.findAffinityTask(e,r),s=n!==-1?n:0,[d]=r.splice(s,1);this.dispatch(d,e);return}}pickWorker(e,t=4){let r=null,n=null,s=1/0;for(let[d,i]of this.workerLoad)i>=t||(this.workerModels.get(d)?.has(e)&&(r===null||i<(this.workerLoad.get(r)??1/0))&&(r=d),i<s&&(s=i,n=d));return r??n}dispatch(e,t){let r=this.workerLoad.get(t)??0;this.workerLoad.set(t,r+1),e.execute(t);}findAffinityTask(e,t){let r=this.workerModels.get(e);return r?t.findIndex(n=>r.has(n.modelId)):-1}get queueDepth(){let e=0;for(let t of this.queues.values())e+=t.length;return e}workerLoadFor(e){return this.workerLoad.get(e)??0}get workerIds(){return [...this.workerLoad.keys()]}reset(){this.workerLoad.clear(),this.workerModels.clear();}static priorityWeight(e){return te[e]}};var re=typeof navigator<"u"?Math.max(1,(navigator.hardwareConcurrency??2)-1):1,ne=2048,se=12e4,N={high:0,low:2,normal:1},x=class o{workers=new Map;registry=new y;scheduler=new T;budget;inferenceWaiters=new Map;pending=new Map;config;caps;resolvedDevice;nextWorkerId=0;nextReqId=0;terminated=false;constructor(e,t){this.config=e,this.caps=t,this.budget=new m(e.maxMemoryMB),this.resolvedDevice=this.resolveDevice(e.defaultDevice,t);}static async create(e){let t=await A(),r={adapter:e.adapter,crossTab:e.crossTab??false,defaultDevice:e.defaultDevice??"auto",maxMemoryMB:e.maxMemoryMB??ne,maxWorkers:e.maxWorkers??re,taskTimeout:e.taskTimeout??se,workerUrl:e.workerUrl??new URL("./worker/dedicated.worker.js",(typeof document === 'undefined' ? require('u' + 'rl').pathToFileURL(__filename).href : (_documentCurrentScript && _documentCurrentScript.tagName.toUpperCase() === 'SCRIPT' && _documentCurrentScript.src || new URL('index.cjs', document.baseURI).href)))},n=new o(r,t);return await n.spawnWorkers(),n}async spawnWorkers(){for(let e=0;e<this.config.maxWorkers;e++)await this.spawnWorker();}async spawnWorker(){let e=this.nextWorkerId++,t=new Worker(this.config.workerUrl,{type:"module"});t.onmessage=n=>{this.handleWorkerMessage(e,n.data);},t.onerror=n=>{console.error(`[inferis] Worker ${e} error:`,n.message),this.handleWorkerCrash(e);};let r={id:e,worker:t};return this.workers.set(e,r),this.scheduler.addWorker(e),t.postMessage({device:this.resolvedDevice,type:"__init__"}),r}handleWorkerMessage(e,t){switch(t.type){case "load-progress":{let r=this.reqIdToModelId.get(t.reqId);if(r){let n=this.registry.get(r);n?.config.onProgress&&n.config.onProgress(t.progress);}break}case "load-complete":{let r=this.reqIdToModelId.get(t.reqId);if(r){let s=this.registry.get(r);s&&(this.registry.setLoaded(r,this.resolvedDevice,t.memoryMB,e),this.budget.allocate(r,t.memoryMB),this.scheduler.notifyModelLoaded(e,r),this.registry.setState(r,p(s.state,"ready"))),this.reqIdToModelId.delete(t.reqId);}let n=this.pending.get(t.reqId);n&&(this.clearPendingTimeout(t.reqId),this.pending.delete(t.reqId),n.resolve(void 0)),this.scheduler.notifyTaskComplete(e);break}case "load-error":{if(t.reqId==="__init__"){console.error("[inferis] Worker init failed:",t.error);break}let r=this.reqIdToModelId.get(t.reqId);if(r){let s=this.registry.get(r);s&&this.registry.setState(r,p(s.state,"error")),this.reqIdToModelId.delete(t.reqId);}let n=this.pending.get(t.reqId);n&&(this.clearPendingTimeout(t.reqId),this.pending.delete(t.reqId),n.reject(u.fromSerialized(t.error))),this.scheduler.notifyTaskComplete(e);break}case "unload-complete":{let r=this.reqIdToModelId.get(t.reqId);r&&(this.budget.release(r),this.scheduler.notifyModelUnloaded(e,r),this.registry.setUnloaded(r),this.registry.setState(r,"disposed"),this.registry.delete(r),this.reqIdToModelId.delete(t.reqId));let n=this.pending.get(t.reqId);n&&(this.clearPendingTimeout(t.reqId),this.pending.delete(t.reqId),n.resolve(void 0));break}case "unload-error":{let r=this.reqIdToModelId.get(t.reqId);r&&(this.budget.release(r),this.scheduler.notifyModelUnloaded(e,r),this.registry.setUnloaded(r),this.registry.setState(r,"error"),this.registry.delete(r),this.reqIdToModelId.delete(t.reqId));let n=this.pending.get(t.reqId);n&&(this.clearPendingTimeout(t.reqId),this.pending.delete(t.reqId),n.reject(u.fromSerialized(t.error)));break}case "run-result":{let r=this.pending.get(t.reqId);r&&(this.clearPendingTimeout(t.reqId),this.pending.delete(t.reqId),r.resolve(t.output));let n=this.reqIdToModelId.get(t.reqId);if(n){let s=this.registry.get(n);s&&s.state==="inferring"&&this.registry.setState(n,p(s.state,"ready")),this.reqIdToModelId.delete(t.reqId);}this.scheduler.notifyTaskComplete(e),n&&this.drainInferenceWaiter(n);break}case "run-error":{let r=this.pending.get(t.reqId);r&&(this.clearPendingTimeout(t.reqId),this.pending.delete(t.reqId),r.reject(u.fromSerialized(t.error)));let n=this.reqIdToModelId.get(t.reqId);if(n){let s=this.registry.get(n);s&&s.state==="inferring"&&this.registry.setState(n,p(s.state,"ready")),this.reqIdToModelId.delete(t.reqId);}this.scheduler.notifyTaskComplete(e),n&&this.drainInferenceWaiter(n);break}case "stream-chunk":{this.pending.get(t.reqId)?.streamController?.enqueue(t.chunk);break}case "stream-end":{let r=this.pending.get(t.reqId);r&&(this.clearPendingTimeout(t.reqId),this.pending.delete(t.reqId),r.streamController?.close());let n=this.reqIdToModelId.get(t.reqId);if(n){let s=this.registry.get(n);s&&s.state==="inferring"&&this.registry.setState(n,p(s.state,"ready")),this.reqIdToModelId.delete(t.reqId);}this.scheduler.notifyTaskComplete(e),n&&this.drainInferenceWaiter(n);break}case "stream-error":{let r=this.pending.get(t.reqId);r&&(this.clearPendingTimeout(t.reqId),this.pending.delete(t.reqId),r.streamController?.error(u.fromSerialized(t.error)));let n=this.reqIdToModelId.get(t.reqId);if(n){let s=this.registry.get(n);s&&s.state==="inferring"&&this.registry.setState(n,p(s.state,"ready")),this.reqIdToModelId.delete(t.reqId);}this.scheduler.notifyTaskComplete(e),n&&this.drainInferenceWaiter(n);break}case "device-lost":{let r=this.registry.get(t.modelId);r&&k(r.state,"error")&&this.registry.setState(t.modelId,"error");break}}}reqIdToModelId=new Map;async load(e,t){if(this.terminated)throw new u("Pool has been terminated","POOL_TERMINATED");let r=y.makeId(e,t.model),n=this.registry.get(r);if(n){let a=n.state;if(a==="ready"||a==="loading"||a==="inferring")return this.makeHandle(r)}let s=this.config.adapter,d=t.estimatedMemoryMB??100;if(!t.estimatedMemoryMB){let a=await s.create().catch(()=>null);try{d=a?.estimateMemoryMB(e,t)??100;}finally{await a?.unload({instance:null,memoryMB:0}).catch(()=>{});}}let i=this.budget.planEviction(d);if(i===null)throw new v(d,this.config.maxMemoryMB);for(let a of i)await this.disposeModel(a);return this.registry.register(r,e,t),this.registry.setState(r,p("idle","loading")),await new Promise((a,c)=>{let l=`${this.nextReqId++}`;this.reqIdToModelId.set(l,r),this.pending.set(l,{reject:c,resolve:a}),this.setRequestTimeout(l,this.config.taskTimeout),this.scheduler.enqueue({enqueuedAt:Date.now(),execute:h=>{let g=this.workers.get(h);if(!g){c(new u("Worker not found","WORKER_NOT_FOUND"));return}try{let{estimatedMemoryMB:_,onProgress:ae,...F}=t;g.worker.postMessage({config:F,device:this.resolvedDevice,modelId:r,reqId:l,task:e,type:"load-model"});}catch{this.clearPendingTimeout(l),this.pending.delete(l),this.reqIdToModelId.delete(l),c(new u("Input cannot be serialized for worker transfer","SERIALIZATION_ERROR"));}},modelId:r,priority:"high",reject:c,reqId:l},this.registry.entries);}),this.makeHandle(r)}async disposeModel(e){let t=this.registry.get(e);if(!t||t.state==="disposed"||!k(t.state,"unloading"))return;if(this.registry.setState(e,"unloading"),t.workerId===null){this.budget.release(e),this.registry.delete(e);return}let r=this.workers.get(t.workerId);if(!r){this.budget.release(e),this.registry.delete(e);return}let n=`${this.nextReqId++}`;this.reqIdToModelId.set(n,e),await new Promise((s,d)=>{this.pending.set(n,{reject:d,resolve:s}),this.setRequestTimeout(n,1e4);try{r.worker.postMessage({modelId:e,reqId:n,type:"unload-model"});}catch{this.clearPendingTimeout(n),this.pending.delete(n),this.reqIdToModelId.delete(n),d(new u("Input cannot be serialized for worker transfer","SERIALIZATION_ERROR"));}});}makeHandle(e){let t=this.registry,r=this.disposeModel.bind(this),n=this.runInference.bind(this),s=this.streamInference.bind(this);return {get device(){return t.get(e)?.device??"wasm"},get id(){return e},get memoryMB(){return t.get(e)?.memoryMB??0},get state(){return t.get(e)?.state??"disposed"},dispose:()=>r(e),onStateChange(d){return t.subscribe(e,d)},run(d,i){return n(e,d,i)},stream(d,i){return s(e,d,i)}}}waitForReady(e,t,r){return new Promise((n,s)=>{if(r?.aborted){s(Object.assign(new Error("AbortError"),{name:"AbortError"}));return}if(this.registry.get(e)?.state==="loading"){let l=()=>{h(),s(Object.assign(new Error("AbortError"),{name:"AbortError"}));},h=this.registry.subscribe(e,g=>{g==="ready"?(h(),r?.removeEventListener("abort",l),n()):g!=="inferring"&&g!=="loading"&&(h(),r?.removeEventListener("abort",l),s(new f(e,g)));});r?.addEventListener("abort",l,{once:true});return}let i={enqueuedAt:Date.now(),onAbort:void 0,priority:t,reject:s,resolve:n,signal:r,unsub:void 0},a=this.inferenceWaiters.get(e)??[];a.push(i),a.sort((l,h)=>N[l.priority]-N[h.priority]||l.enqueuedAt-h.enqueuedAt),this.inferenceWaiters.set(e,a);let c=this.registry.subscribe(e,l=>{if(l!=="inferring"&&l!=="loading"&&l!=="ready"){i.unsub=void 0;let h=this.inferenceWaiters.get(e);if(h){let g=h.indexOf(i);g>=0&&h.splice(g,1);}c(),i.signal&&i.onAbort&&i.signal.removeEventListener("abort",i.onAbort),s(new f(e,l));}});if(i.unsub=c,r){let l=()=>{let h=this.inferenceWaiters.get(e);if(h){let g=h.indexOf(i);g>=0&&h.splice(g,1);}i.unsub?.(),i.unsub=void 0,s(Object.assign(new Error("AbortError"),{name:"AbortError"}));};i.onAbort=l,r.addEventListener("abort",l,{once:true});}})}drainInferenceWaiter(e){let t=this.inferenceWaiters.get(e);if(!t||t.length===0)return;let[r]=t.splice(0,1);r.unsub?.(),r.signal&&r.onAbort&&r.signal.removeEventListener("abort",r.onAbort),r.resolve();}async runInference(e,t,r){if(!this.registry.get(e))throw new u(`Model "${e}" not found`,"MODEL_NOT_FOUND");for(;!I(this.registry.get(e).state);){let s=this.registry.get(e).state;if(s==="inferring"||s==="loading")await this.waitForReady(e,r?.priority??"normal",r?.signal);else throw new f(e,s)}if(r?.signal?.aborted)throw Object.assign(new Error("AbortError"),{name:"AbortError"});let n=this.registry.get(e);return this.registry.setState(e,p(n.state,"inferring")),this.budget.touch(e),new Promise((s,d)=>{let i=`${this.nextReqId++}`;this.reqIdToModelId.set(i,e);let a=()=>{(n.workerId!==null?this.workers.get(n.workerId):void 0)?.worker.postMessage({reqId:i,type:"abort"});};r?.signal?.addEventListener("abort",a,{once:true}),this.pending.set(i,{onAbort:a,reject:c=>{r?.signal?.removeEventListener("abort",a),d(c);},resolve:c=>{r?.signal?.removeEventListener("abort",a),s(c);},signal:r?.signal}),this.setRequestTimeout(i,this.config.taskTimeout),this.scheduler.enqueue({enqueuedAt:Date.now(),execute:c=>{let l=this.workers.get(c);if(l)try{let{priority:h,signal:g,..._}=r??{};l.worker.postMessage({input:t,modelId:e,options:_,reqId:i,type:"run"});}catch{this.clearPendingTimeout(i),this.pending.delete(i),this.reqIdToModelId.delete(i),r?.signal?.removeEventListener("abort",a),d(new u("Input cannot be serialized for worker transfer","SERIALIZATION_ERROR"));}},modelId:e,priority:r?.priority??"normal",reject:c=>{r?.signal?.removeEventListener("abort",a),d(c);},reqId:i},this.registry.entries);})}streamInference(e,t,r){if(!this.registry.get(e))return new ReadableStream({start(s){s.error(new u(`Model "${e}" not found`,"MODEL_NOT_FOUND"));}});let n="";return new ReadableStream({start:async s=>{for(;!I(this.registry.get(e).state);){let a=this.registry.get(e).state;if(a==="inferring"||a==="loading")try{await this.waitForReady(e,r?.priority??"normal",r?.signal);}catch(c){s.error(c);return}else {s.error(new f(e,a));return}}let d=this.registry.get(e);this.registry.setState(e,p(d.state,"inferring")),this.budget.touch(e),n=`${this.nextReqId++}`,this.reqIdToModelId.set(n,e);let i=()=>{let a=this.registry.get(e);(a?.workerId!=null?this.workers.get(a.workerId):void 0)?.worker.postMessage({reqId:n,type:"abort"});};r?.signal?.addEventListener("abort",i,{once:true}),this.pending.set(n,{onAbort:i,reject:a=>{r?.signal?.removeEventListener("abort",i),s.error(a);},resolve:()=>{},signal:r?.signal,streamController:s}),this.setRequestTimeout(n,this.config.taskTimeout),this.scheduler.enqueue({enqueuedAt:Date.now(),execute:a=>{let c=this.workers.get(a);if(c)try{let{priority:l,signal:h,...g}=r??{};c.worker.postMessage({input:t,modelId:e,options:g,reqId:n,type:"run-stream"});}catch{this.clearPendingTimeout(n),this.pending.delete(n),this.reqIdToModelId.delete(n),r?.signal?.removeEventListener("abort",i),s.error(new u("Input cannot be serialized for worker transfer","SERIALIZATION_ERROR"));}},modelId:e,priority:r?.priority??"normal",reject:a=>{r?.signal?.removeEventListener("abort",i),s.error(a);},reqId:n},this.registry.entries);},cancel:()=>{if(!n)return;let s=this.registry.get(e);(s?.workerId!=null?this.workers.get(s.workerId):void 0)?.worker.postMessage({reqId:n,type:"abort"});let i=this.pending.get(n);if(i){i.signal&&i.onAbort&&i.signal.removeEventListener("abort",i.onAbort);try{i.streamController?.close();}catch{}this.clearPendingTimeout(n),this.pending.delete(n),this.reqIdToModelId.delete(n);}}})}setRequestTimeout(e,t){let r=setTimeout(()=>{let s=this.pending.get(e);s&&(s.signal&&s.onAbort&&s.signal.removeEventListener("abort",s.onAbort),this.pending.delete(e),this.reqIdToModelId.delete(e),s.reject(new M(e,t)));},t),n=this.pending.get(e);n&&(n.timeoutId=r);}clearPendingTimeout(e){let t=this.pending.get(e);t?.timeoutId!==void 0&&clearTimeout(t.timeoutId);}handleWorkerCrash(e){this.scheduler.removeWorker(e),this.workers.delete(e);let t=new Set;for(let r of this.registry.byWorker(e))k(r.state,"error")&&this.registry.setState(r.id,"error"),t.add(r.id);for(let[r,n]of this.pending){let s=this.reqIdToModelId.get(r);s&&t.has(s)&&(this.clearPendingTimeout(r),n.signal&&n.onAbort&&n.signal.removeEventListener("abort",n.onAbort),this.pending.delete(r),this.reqIdToModelId.delete(r),n.reject(new u("Worker crashed","WORKER_CRASHED")));}}async terminate(){this.terminated=true;for(let[e,t]of this.pending)clearTimeout(t.timeoutId),t.reject(new u("Pool terminated","POOL_TERMINATED")),this.pending.delete(e),this.reqIdToModelId.delete(e);for(let e of this.workers.values())e.worker.terminate();this.workers.clear(),this.scheduler.reset();}capabilities(){return this.caps}resolveDevice(e,t){return e==="auto"?t.webgpu.supported&&!t.webgpu.isFallback?"webgpu":"wasm":e}get _registry(){return this.registry}get _budget(){return this.budget}get _scheduler(){return this.scheduler}};async function ie(o){return x.create(o)}var B=class{listeners=new Set;on(e){return this.listeners.add(e),()=>this.listeners.delete(e)}emit(e){for(let t of this.listeners)try{t(e);}catch{}}clear(){this.listeners.clear();}get listenerCount(){return this.listeners.size}};function z(o){return Symbol.asyncIterator in o?o:{[Symbol.asyncIterator](){let e=o.getReader();return {async next(){let{done:t,value:r}=await e.read();return t?{done:true,value:void 0}:{done:false,value:r}},async return(){return await e.cancel(),{done:true,value:void 0}},async throw(t){throw await e.cancel(t),t}}}}}async function U(o){let e=[];for await(let t of z(o))e.push(t);return e}async function oe(o){return (await U(o)).join("")}exports.BudgetExceededError=v;exports.DeviceLostError=O;exports.InferenceError=W;exports.InferisError=u;exports.InvalidStateTransitionError=w;exports.LeaderElection=S;exports.MemoryBudget=m;exports.ModelDisposedError=C;exports.ModelLoadError=L;exports.ModelNotReadyError=f;exports.ModelRegistry=y;exports.ProgressEmitter=B;exports.Scheduler=T;exports.SharedWorkerBridge=q;exports.TabChannel=R;exports.TaskTimeoutError=M;exports.WorkerError=P;exports.WorkerPool=x;exports.canTransition=k;exports.clearCapabilitiesCache=Q;exports.collectStream=U;exports.collectStreamText=oe;exports.createPool=ie;exports.detectCapabilities=A;exports.isAcceptingInference=I;exports.isTerminal=ee;exports.readableToAsyncIter=z;exports.transition=p;//# sourceMappingURL=index.cjs.map
+//# sourceMappingURL=index.cjs.map