cleanup dependencies, TODOs

JetStreamDriver.js

@@ -2098,6 +2098,7 @@ let BENCHMARKS = [
         worstCaseCount: 2,
         tags: ["Default", "Wasm"],
     }),
+    // TODO: Change this to AsyncBenchmark, we don't actually need the Emscripten magic.
     new WasmEMCCBenchmark({
         name: "transformersjs-bert-wasm",
         files: [

transformersjs/README.md

@@ -1,3 +1,12 @@
 - Two tasks: one text/NLP, one audio processing/speech-to-text.
-- Everything in `build/` comes from upstream or is generated.
-- Everything in `util/` is tooling for building, preparing the benchmark.
+- Everything in `build/` is generated or an upstream library.
+- Everything in `util/` is tooling for building and preparing the benchmark.
+
+# Licenses
+
+- Transformers.js: Apache 2.0, https://github.com/huggingface/transformers.js/blob/main/LICENSE
+- ONNX runtime: MIT, https://github.com/microsoft/onnxruntime/blob/main/LICENSE
+- `text-encoding` Polyfill: Unlicense OR Apache 2.0, https://github.com/inexorabletash/text-encoding/blob/master/LICENSE.md
+- Model `DistilBERT base uncased finetuned SST-2`: Apache 2.0, https://huggingface.co/distilbert/distilbert-base-uncased-finetuned-sst-2-english
+- Model `openai/whisper-tiny.en`: Apache 2.0, https://huggingface.co/openai/whisper-tiny.en
+- Audio file for speech-to-text task: Public domain, https://www.jfklibrary.org/learn/about-jfk/historic-speeches/inaugural-address

transformersjs/build.log

@@ -1,9 +1,9 @@
-Built on 2025-02-11T14:56:01Z
+Built on 2025-02-11T16:17:21Z
 Installing Node dependencies...
 
-up to date, audited 153 packages in 489ms
+up to date, audited 85 packages in 418ms
 
-27 packages are looking for funding
+20 packages are looking for funding
   run `npm fund` for details
 
 found 0 vulnerabilities

transformersjs/build.sh

@@ -23,8 +23,6 @@ rm jfk.wav
 echo "Download and run model(s)..." | tee -a "$BUILD_LOG"
 # This automatically places the model files in `build/models/`.
 node util/test-models.mjs
-# TODO(dlehmann): Compress models with zopfli, without header.
-# Uncompress/inflate with https://github.com/binji/raw-wasm/blob/main/inflate/index.html / https://github.com/binji/raw-wasm/blob/main/inflate/inflate.js
 
 echo "Copy library files into build/..." | tee -a "$BUILD_LOG"
 # TextEncoder/TextDecoder polyfill with UTF-16 LE support.
@@ -42,6 +40,10 @@ cp util/node_modules/@huggingface/transformers/dist/transformers.js build/
 # TODO(dlehmann): Measure performance difference between the two.
 cp util/node_modules/onnxruntime-web/dist/ort-wasm-simd-threaded.{mjs,wasm} build/lib/onnxruntime-web/
 
+# TODO: Compress model data (and maybe Wasm modules) with zstd.
+# Either decompress with native APIs available in browsers or JS/Wasm polyfill?
+# E.g., https://github.com/101arrowz/fzstd or https://github.com/fabiospampinato/zstandard-wasm or https://github.com/donmccurdy/zstddec-wasm
+
 # Cleanup node packages.
 # rm -rf util/node_modules/
 

transformersjs/task-whisper.js

@@ -23,7 +23,7 @@ globalThis.initPipeline = async function(pipeline) {
   return await pipeline(
     'automatic-speech-recognition',
     'Xenova/whisper-tiny.en',
-    // FIXME: The non-quantized model is faster!?
+    // Use quantized model because of smaller weights.
     { dtype: 'q8' }
   );
 }

transformersjs/util/package.json

@@ -1,11 +1,10 @@
 {
   "name": "transformersjs-wasm",
-  "version": "1.0.0",
+  "version": "0.0.1",
   "author": "Daniel Lehmann <[email protected]>",
   "license": "MIT",
   "dependencies": {
     "@huggingface/transformers": "^3.3.2",
-    "node-zopfli": "^2.1.4",
     "text-encoding": "^0.7.0",
     "wavefile": "^11.0.0"
   }

transformersjs/util/test-models.mjs

@@ -40,7 +40,6 @@ globalThis.print = function(str) { console.log(str); }
   console.log('Automatic speech recognition with Whisper model.');
   await import('../task-whisper.js');
 
-  // TODO: We could also transcribe `ted_60_16k.wav`, but that takes quite long...
   const inputAudioBuffer = readFileSync('build/inputs/jfk.raw').buffer;
 
   let start = performance.now();

wasm-cli.js

@@ -39,6 +39,7 @@ testList = [
   "zlib-wasm",
   "Kotlin-compose-wasm",
   "transformersjs-bert-wasm",
+  "transformersjs-whisper-wasm",
 ];
 
 // Reuse the full CLI runner, just with the subset of Wasm line items above.