fix CDN build

.gitignore

@@ -1,6 +1,7 @@
 node_modules
 .DS_Store
 .vscode
+.claude
 /cache
 /build
 /build-compat

.npmignore

@@ -4,6 +4,11 @@
 /llama.cpp
 /models
 /build
+/build-*
 /cache
+/compat
 /.github
-/.vscode
+/.vscode
+/.claude
+a.out.js
+a.out.wasm

compat/package.json

@@ -1,11 +1,12 @@
 {
   "name": "@wllama/wllama-compat",
-  "version": "3.4.0",
+  "version": "3.4.1",
   "description": "Optional package providing compatibility with older browsers for @wllama/wllama",
   "main": "index.js",
   "type": "module",
   "scripts": {
-    "upload": "npm publish --access public"
+    "upload": "npm publish --access public",
+    "prepublishOnly": "node ../scripts/check_package_size.js"
   },
   "repository": {
     "type": "git",

package.json

@@ -1,6 +1,6 @@
 {
   "name": "@wllama/wllama",
-  "version": "3.4.0",
+  "version": "3.4.1",
   "description": "WebAssembly binding for llama.cpp - Enabling on-browser LLM inference",
   "main": "index.js",
   "type": "module",
@@ -25,7 +25,8 @@
     "test": "vitest",
     "test:firefox": "BROWSER=firefox vitest",
     "test:safari": "BROWSER=safari vitest",
-    "test:wgpu": "WEBGPU=1 vitest"
+    "test:wgpu": "WEBGPU=1 vitest",
+    "prepublishOnly": "node scripts/check_package_size.js"
   },
   "repository": {
     "type": "git",

scripts/check_package_size.js

@@ -0,0 +1,22 @@
+#!/usr/bin/env node
+
+import { execSync } from 'child_process';
+
+const MAX_SIZE = 20 * 1024 * 1024; // 20 MB
+const MAX_FILES = 90;
+
+const result = JSON.parse(execSync('npm pack --dry-run --json 2>/dev/null'));
+const { unpackedSize, entryCount } = result[0];
+
+console.log(`Unpacked size: ${(unpackedSize / 1024 / 1024).toFixed(2)} MB`);
+console.log(`Total files: ${entryCount}`);
+
+if (unpackedSize > MAX_SIZE) {
+  console.error(`ERROR: Unpacked size exceeds 20 MB limit`);
+  process.exit(1);
+}
+
+if (entryCount > MAX_FILES) {
+  console.error(`ERROR: Total files (${entryCount}) exceeds limit of ${MAX_FILES}`);
+  process.exit(1);
+}

src/wasm-from-cdn.ts

@@ -2,12 +2,12 @@
 // Do not edit this file directly
 
 const WasmFromCDN = {
-  default: 'https://cdn.jsdelivr.net/npm/@wllama/wllama@3.4.0/src/wasm/wllama.wasm',
+  default: 'https://cdn.jsdelivr.net/npm/@wllama/wllama@3.4.1/src/wasm/wllama.wasm',
 };
 
 export const WasmCompatFromCDN = {
-  worker: 'https://cdn.jsdelivr.net/npm/@wllama/wllama-compat@3.4.0/wasm/wllama.js',
-  wasm: 'https://cdn.jsdelivr.net/npm/@wllama/wllama-compat@3.4.0/wasm/wllama.wasm',
+  worker: 'https://cdn.jsdelivr.net/npm/@wllama/wllama-compat@3.4.1/wasm/wllama.js',
+  wasm: 'https://cdn.jsdelivr.net/npm/@wllama/wllama-compat@3.4.1/wasm/wllama.wasm',
 };
 
 export default WasmFromCDN;

src/workers-code/generated.ts

@@ -1,7 +1,7 @@
 // This file is auto-generated
 // To re-generate it, run: npm run build:worker
 
-export const LIBLLAMA_VERSION = 'b9425-0821c5f';
+export const LIBLLAMA_VERSION = 'b9437-aa46bda';
 
 export const LLAMA_CPP_WORKER_CODE = "// Start the main llama.cpp\nlet wllamaMalloc;\nlet wllamaStart;\nlet wllamaAction;\nlet wllamaExit;\nlet wllamaDebug;\n\nlet Module = null;\nlet isCompat = false;\nlet lastStack = '';\nlet isAborted = false;\nlet hasMultithread = false;\n\n//////////////////////////////////////////////////////////////\n// UTILS\n//////////////////////////////////////////////////////////////\n\n// send message back to main thread\nconst msg = (data, transfer) => postMessage(data, transfer);\n\n// Convert CPP log into JS log\nconst cppLogToJSLog = (line) => {\n  const matched = line.match(/@@(DEBUG|INFO|WARN|ERROR)@@(.*)/);\n  return !!matched\n    ? {\n        level: (matched[1] === 'INFO' ? 'debug' : matched[1]).toLowerCase(),\n        text: matched[2],\n      }\n    : { level: 'log', text: line };\n};\n\nconst getHeapU8 = () => {\n  const buffer = Module.wasmMemory.buffer;\n  return new Uint8Array(buffer);\n};\n\nconst toSizeT = (num) => {\n  return isCompat ? Number(num) : BigInt(num);\n};\n\n// Get module config that forwards stdout/err to main thread\nconst getWModuleConfig = (_argMainScriptBlob) => {\n  var pathConfig = RUN_OPTIONS.pathConfig;\n  var pthreadPoolSize = RUN_OPTIONS.nbThread;\n  var argMainScriptBlob = _argMainScriptBlob;\n\n  isCompat = RUN_OPTIONS.compat;\n  hasMultithread = pthreadPoolSize > 1;\n\n  msg({\n    verb: 'console.debug',\n    args: [\n      `Multithread enabled: ${hasMultithread}, pthreadPoolSize: ${pthreadPoolSize}`,\n    ],\n  });\n\n  if (!pathConfig['wllama.wasm']) {\n    throw new Error('\"wllama.wasm\" is missing in pathConfig');\n  }\n  return {\n    noInitialRun: true,\n    print: function (text) {\n      if (arguments.length > 1)\n        text = Array.prototype.slice.call(arguments).join(' ');\n      msg({ verb: 'console.log', args: [text] });\n    },\n    printErr: function (text) {\n      if (arguments.length > 1)\n        text = Array.prototype.slice.call(arguments).join(' ');\n      if (text.startsWith('@@STACK@@')) {\n        lastStack = text.slice('@@STACK@@'.length);\n        return;\n      }\n      const logLine = cppLogToJSLog(text);\n      msg({ verb: 'console.' + logLine.level, args: [logLine.text] });\n    },\n    locateFile: function (filename, basePath) {\n      const p = pathConfig[filename];\n      const truncate = (str) =>\n        str.length > 128 ? `${str.substr(0, 128)}...` : str;\n      if (filename.match(/wllama\\.worker\\.js/)) {\n        msg({\n          verb: 'console.error',\n          args: [\n            '\"wllama.worker.js\" is removed from v2.2.1. Hint: make sure to clear browser\\'s cache.',\n          ],\n        });\n      } else {\n        msg({\n          verb: 'console.debug',\n          args: [`Loading \"${filename}\" from \"${truncate(p)}\"`],\n        });\n        return p;\n      }\n    },\n    mainScriptUrlOrBlob: hasMultithread\n      ? argMainScriptBlob\n      : 'throw new Error(\"Multithreading is not enabled\")',\n    pthreadPoolSize: hasMultithread ? pthreadPoolSize : 0,\n    wasmMemory: hasMultithread ? getWasmMemory() : null,\n    onAbort: function (message) {\n      isAborted = true;\n      msg({ verb: 'signal.abort', args: ['abort', message, lastStack, null] });\n    },\n    onExit: function (code) {\n      isAborted = true;\n      const callstack = new Error().stack.toString();\n      msg({\n        verb: 'signal.abort',\n        args: ['abort', 'exit(' + code + ')', callstack, null],\n      });\n    },\n  };\n};\n\n// Get the memory to be used by wasm. (Only used in multi-thread mode)\n// Because we have a weird OOM issue on iOS, we need to try some values\n// See: https://github.com/emscripten-core/emscripten/issues/19144\n//      https://github.com/godotengine/godot/issues/70621\nconst getWasmMemory = () => {\n  let minBytes = 128 * 1024 * 1024;\n  let maxBytes = 4096 * 1024 * 1024;\n  let stepBytes = 128 * 1024 * 1024;\n  while (maxBytes > minBytes) {\n    try {\n      const wasmMemory = new WebAssembly.Memory({\n        initial: toSizeT(minBytes / 65536),\n        maximum: toSizeT(maxBytes / 65536),\n        shared: true,\n        address: isCompat ? undefined : 'i64',\n      });\n      return wasmMemory;\n    } catch (e) {\n      maxBytes -= stepBytes;\n      continue; // retry\n    }\n  }\n  throw new Error('Cannot allocate WebAssembly.Memory');\n};\n\n//////////////////////////////////////////////////////////////\n// HEAPFS PATCH\n//////////////////////////////////////////////////////////////\n\n/**\n * By default, emscripten uses memfs. The way it works is by\n * allocating new Uint8Array in javascript heap. This is not good\n * because it requires files to be copied to wasm heap each time\n * a file is read.\n *\n * HeapFS is an alternative, which resolves this problem by\n * allocating space for file directly inside wasm heap. This\n * allows us to mmap without doing any copy.\n *\n * For llama.cpp, this is great because we use MAP_SHARED\n *\n * Ref: https://github.com/ngxson/wllama/pull/39\n * Ref: https://github.com/emscripten-core/emscripten/blob/main/src/library_memfs.js\n *\n * Note 29/05/2024 @ngxson\n * Due to ftell() being limited to MAX_LONG, we cannot load files bigger than 2^31 bytes (or 2GB)\n * Ref: https://github.com/emscripten-core/emscripten/blob/main/system/lib/libc/musl/src/stdio/ftell.c\n */\n\nconst fsNameToFile = {}; // map Name => File\nconst fsIdToFile = {}; // map ID => File\nlet currFileId = 0;\n\n// Patch and redirect memfs calls to wllama\nconst patchHeapFS = () => {\n  const m = Module;\n  // save functions\n  m.MEMFS.stream_ops._read = m.MEMFS.stream_ops.read;\n  m.MEMFS.stream_ops._write = m.MEMFS.stream_ops.write;\n  m.MEMFS.stream_ops._llseek = m.MEMFS.stream_ops.llseek;\n  m.MEMFS.stream_ops._allocate = m.MEMFS.stream_ops.allocate;\n  m.MEMFS.stream_ops._mmap = m.MEMFS.stream_ops.mmap;\n  m.MEMFS.stream_ops._msync = m.MEMFS.stream_ops.msync;\n\n  const patchStream = (stream) => {\n    const name = stream.node.name;\n    if (fsNameToFile[name]) {\n      const f = fsNameToFile[name];\n      const ptr = Number(f.ptr);\n      stream.node.contents = getHeapU8().subarray(ptr, ptr + f.size);\n      stream.node.usedBytes = f.size;\n    }\n  };\n\n  // replace \"read\" functions\n  m.MEMFS.stream_ops.read = function (\n    stream,\n    buffer,\n    offset,\n    length,\n    position\n  ) {\n    patchStream(stream);\n    return m.MEMFS.stream_ops._read(stream, buffer, offset, length, position);\n  };\n  m.MEMFS.ops_table.file.stream.read = m.MEMFS.stream_ops.read;\n\n  // replace \"llseek\" functions\n  m.MEMFS.stream_ops.llseek = function (stream, offset, whence) {\n    patchStream(stream);\n    return m.MEMFS.stream_ops._llseek(stream, offset, whence);\n  };\n  m.MEMFS.ops_table.file.stream.llseek = m.MEMFS.stream_ops.llseek;\n\n  // replace \"mmap\" functions\n  m.MEMFS.stream_ops.mmap = function (stream, length, position, prot, flags) {\n    patchStream(stream);\n    const name = stream.node.name;\n    if (fsNameToFile[name]) {\n      const f = fsNameToFile[name];\n      const mmapPtr = f.ptr + toSizeT(position);\n      return {\n        ptr: mmapPtr,\n        allocated: false,\n      };\n    } else {\n      return m.MEMFS.stream_ops._mmap(stream, length, position, prot, flags);\n    }\n  };\n  m.MEMFS.ops_table.file.stream.mmap = m.MEMFS.stream_ops.mmap;\n\n  // mount FS\n  m.FS.mkdir('/models');\n  m.FS.mount(m.MEMFS, { root: '.' }, '/models');\n};\n\n// Allocate a new file in wllama heapfs, returns file ID\nconst heapfsAlloc = (name, size, allocBuffer) => {\n  if (size < 1) {\n    throw new Error('File size must be bigger than 0');\n  }\n  const m = Module;\n  const ptr = toSizeT(allocBuffer ? m.mmapAlloc(size) : 0);\n  const file = {\n    ptr: ptr,\n    size: size,\n    id: currFileId++,\n  };\n  fsIdToFile[file.id] = file;\n  fsNameToFile[name] = file;\n  return file.id;\n};\n\n// Add new file to wllama heapfs, return number of written bytes\nconst heapfsWrite = (id, buffer, offset) => {\n  if (fsIdToFile[id]) {\n    const { ptr, size } = fsIdToFile[id];\n    const afterWriteByte = offset + buffer.byteLength;\n    if (afterWriteByte > size) {\n      throw new Error(\n        `File ID ${id} write out of bound, afterWriteByte = ${afterWriteByte} while size = ${size}`\n      );\n    }\n    getHeapU8().set(buffer, Number(ptr) + offset);\n    return buffer.byteLength;\n  } else {\n    throw new Error(`File ID ${id} not found in heapfs`);\n  }\n};\n\n//////////////////////////////////////////////////////////////\n// ASYNC FILE READ\n//////////////////////////////////////////////////////////////\n\nlet isAwaitReading = false;\nlet pendingReadPromise = null;\nlet pendingReadResolve = null;\nlet pendingReadReject = null;\n\nconst _stripModelsPrefix = (path) => path.replace(/^\\/?models\\//, '');\n\n// Called from EM_ASYNC_JS stub in wllama-fs.h (path is already a JS string)\nconst _wllama_js_file_read = async (path, offset, req_size, out_ptr) => {\n  const name = _stripModelsPrefix(path);\n\n  pendingReadPromise = new Promise((res, rej) => {\n    pendingReadResolve = res;\n    pendingReadReject = rej;\n  });\n  isAwaitReading = true;\n\n  postMessage({ verb: 'fs.read_req', args: [name, offset, req_size] });\n\n  let data;\n  try {\n    data = await pendingReadPromise;\n  } finally {\n    isAwaitReading = false;\n    pendingReadResolve = null;\n    pendingReadReject = null;\n  }\n\n  const bytes = new Uint8Array(data);\n  getHeapU8().set(bytes, out_ptr);\n  return toSizeT(bytes.length);\n};\n\n//////////////////////////////////////////////////////////////\n// MAIN CODE\n//////////////////////////////////////////////////////////////\n\nconst callWrapper = (name, ret, args, isAsync) => {\n  const fn = Module.cwrap(\n    name,\n    ret,\n    args,\n    isAsync ? { async: true } : undefined\n  );\n  return async (action, req) => {\n    // console.log(`Calling ${name} with action:`, action, 'and req:', req);\n    let result;\n    try {\n      if (args.length === 2) {\n        result = isAsync ? await fn(action, req) : fn(action, req);\n      } else {\n        result = fn();\n      }\n    } catch (ex) {\n      console.error(ex);\n      throw ex;\n    }\n    return result;\n  };\n};\n\nfunction handleError(err) {\n  // If WASM already aborted, onAbort already sent signal.abort; skip to avoid\n  // re-reporting the resulting WebAssembly.RuntimeError as a JS exception.\n  if (isAborted) return;\n\n  const message = err ? err.message || String(err) : 'Unknown error';\n  const stack = err ? err.stack || String(err) : '';\n  msg({\n    verb: 'signal.abort',\n    args: ['exception', message, stack, err],\n  });\n}\n\nonmessage = async (e) => {\n  if (!e.data) return;\n  const { verb, args, callbackId } = e.data;\n\n  // fs.read_res arrives while wasm is JSPI-suspended; resolve the pending promise.\n  if (verb === 'fs.read_res') {\n    if (pendingReadResolve) {\n      pendingReadResolve(args[0]);\n    }\n    return;\n  }\n\n  // Guard: while awaiting a file read, reject any other incoming task.\n  if (isAwaitReading) {\n    if (callbackId) {\n      msg({\n        callbackId,\n        err: 'Worker is suspended waiting for file data (JSPI)',\n      });\n    }\n    return;\n  }\n\n  if (!callbackId) {\n    msg({ verb: 'console.error', args: ['callbackId is required', e.data] });\n    return;\n  }\n\n  if (verb === 'module.init') {\n    const argMainScriptBlob = args[0];\n    const argUseAsyncFile = args[1];\n    try {\n      Module = getWModuleConfig(argMainScriptBlob);\n      Module.preRun = () => {\n        if (argUseAsyncFile) {\n          Module.ENV['USE_ASYNC_FILE'] = '1';\n        }\n      };\n      Module.onRuntimeInitialized = () => {\n        // async call once module is ready\n        // init FS\n        patchHeapFS();\n        // init cwrap\n        const pointer = isCompat ? 'number' : 'bigint';\n        // TODO: note sure why emscripten cannot bind if there is only 1 argument\n        wllamaMalloc = callWrapper('wllama_malloc', pointer, [\n          'number',\n          pointer,\n        ]);\n        wllamaStart = callWrapper('wllama_start', 'string', [], true);\n        wllamaAction = callWrapper(\n          'wllama_action',\n          pointer,\n          ['string', pointer],\n          true\n        );\n        wllamaExit = callWrapper('wllama_exit', 'string', []);\n        wllamaDebug = callWrapper('wllama_debug', 'string', []);\n        msg({ callbackId, result: null });\n      };\n      wModuleInit();\n    } catch (err) {\n      handleError(err);\n    }\n    return;\n  }\n\n  if (verb === 'fs.alloc') {\n    const argFilename = args[0];\n    const argSize = args[1];\n    const argAllocBuffer = args[2];\n    try {\n      // create blank file\n      const emptyBuffer = new ArrayBuffer(0);\n      Module['FS_createDataFile'](\n        '/models',\n        argFilename,\n        emptyBuffer,\n        true,\n        true,\n        true\n      );\n      // alloc data on heap\n      const fileId = heapfsAlloc(argFilename, argSize, argAllocBuffer);\n      msg({ callbackId, result: { fileId } });\n    } catch (err) {\n      handleError(err);\n    }\n    return;\n  }\n\n  if (verb === 'fs.write') {\n    const argFileId = args[0];\n    const argBuffer = args[1];\n    const argOffset = args[2];\n    try {\n      const writtenBytes = heapfsWrite(argFileId, argBuffer, argOffset);\n      msg({ callbackId, result: { writtenBytes } });\n    } catch (err) {\n      handleError(err);\n    }\n    return;\n  }\n\n  if (verb === 'wllama.start') {\n    try {\n      const result = await wllamaStart();\n      msg({ callbackId, result });\n    } catch (err) {\n      handleError(err);\n    }\n    return;\n  }\n\n  if (verb === 'wllama.action') {\n    const argAction = args[0];\n    const argEncodedMsg = args[1];\n    try {\n      const inputPtr = await wllamaMalloc(toSizeT(argEncodedMsg.byteLength), 0);\n      // copy data to wasm heap\n      const inputBuffer = new Uint8Array(\n        getHeapU8().buffer,\n        Number(inputPtr),\n        argEncodedMsg.byteLength\n      );\n      inputBuffer.set(argEncodedMsg, 0);\n      const outputPtr = await wllamaAction(argAction, inputPtr);\n      // length of output buffer is written at the first 4 bytes of input buffer\n      const outputLen = new Uint32Array(\n        getHeapU8().buffer,\n        Number(inputPtr),\n        1\n      )[0];\n      // copy the output buffer to JS heap\n      const outputBuffer = new Uint8Array(outputLen);\n      const outputSrcView = new Uint8Array(\n        getHeapU8().buffer,\n        Number(outputPtr),\n        outputLen\n      );\n      outputBuffer.set(outputSrcView, 0); // copy it\n      msg({ callbackId, result: outputBuffer }, [outputBuffer.buffer]);\n    } catch (err) {\n      handleError(err);\n    }\n    return;\n  }\n\n  if (verb === 'wllama.exit') {\n    try {\n      const result = await wllamaExit();\n      msg({ callbackId, result });\n    } catch (err) {\n      handleError(err);\n    }\n    return;\n  }\n\n  if (verb === 'wllama.debug') {\n    try {\n      const result = await wllamaDebug();\n      msg({ callbackId, result });\n    } catch (err) {\n      handleError(err);\n    }\n    return;\n  }\n};\n";