@jax-js/jax 0.1.12 → 0.1.14
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +11 -7
- package/dist/{backend-x-6vqzIM.cjs → backend-VlXzdQvR.cjs} +2111 -1557
- package/dist/{backend-DI-V78Rk.js → backend-apsUOPzb.js} +2111 -1557
- package/dist/index.cjs +10 -1
- package/dist/index.js +10 -1
- package/dist/{webgl-CD3WK_Me.cjs → webgl-C6rCbloA.cjs} +1 -1
- package/dist/{webgl-BhsnpeB0.js → webgl-Hh0FX6oV.js} +1 -1
- package/dist/{webgpu-C2kLdkUh.js → webgpu-BRv5r9Sl.js} +84 -31
- package/dist/{webgpu-C4S8Uq9e.cjs → webgpu-pWnE96Xc.cjs} +84 -31
- package/package.json +1 -1
package/README.md
CHANGED
|
@@ -76,6 +76,7 @@ Demos on the jax-js website:
|
|
|
76
76
|
- [Object detection: D-FINE (ONNX)](https://jax-js.com/d-fine)
|
|
77
77
|
- [Object detection: DETR ResNet-50 (ONNX)](https://jax-js.com/detr-resnet-50)
|
|
78
78
|
- [Fluid simulation (Navier-Stokes)](https://jax-js.com/fluid-sim)
|
|
79
|
+
- [Neural cellular automata](https://jax-js.com/nca-growing)
|
|
79
80
|
- [In-browser REPL](https://jax-js.com/repl)
|
|
80
81
|
- [Matmul benchmark](https://jax-js.com/bench/matmul)
|
|
81
82
|
- [Matvec benchmark](https://jax-js.com/bench/matvec)
|
|
@@ -365,11 +366,14 @@ To see per-kernel traces in browser development tools, call `jax.profiler.startT
|
|
|
365
366
|
|
|
366
367
|
The WebGPU runtime includes an ML compiler with tile-aware optimizations, tuned for indiidual
|
|
367
368
|
browsers. Also, this library uniquely has the `jit()` feature that fuses operations together and
|
|
368
|
-
records an execution graph.
|
|
369
|
-
Apple M4 Max chip ([try it](https://jax-js.com/bench/matmul)).
|
|
369
|
+
records an execution graph.
|
|
370
370
|
|
|
371
|
-
|
|
372
|
-
[
|
|
371
|
+
- _On WebGPU:_ jax-js achieves **over 7000 GFLOP/s** for matrix multiplication on an Apple M4 Max
|
|
372
|
+
chip ([try it](https://jax-js.com/bench/matmul)).
|
|
373
|
+
- _On WebAssembly (CPU):_ jax-js is the fastest multithreaded in-browser matrix multiplication, over
|
|
374
|
+
twice as fast as XNNPACK, and matches **OpenBLAS performance** on Apple Silicon.
|
|
375
|
+
|
|
376
|
+
This is significantly faster than both [TensorFlow.js](https://github.com/tensorflow/tfjs) and
|
|
373
377
|
[ONNX Runtime Web](https://www.npmjs.com/package/onnxruntime-web), which both use handwritten
|
|
374
378
|
libraries of custom kernels.
|
|
375
379
|
|
|
@@ -423,8 +427,8 @@ pnpm check # Run TypeScript type checking
|
|
|
423
427
|
Contributions are welcomed! Some fruitful areas to look into:
|
|
424
428
|
|
|
425
429
|
- Adding support for more JAX functions and operations, see [compatibility table](./FEATURES.md).
|
|
426
|
-
- Improving performance of the WebGPU and Wasm runtimes
|
|
427
|
-
|
|
428
|
-
- Making a fast transformer inference engine
|
|
430
|
+
- Improving performance of the WebGPU and Wasm runtimes and generating better kernels, especially
|
|
431
|
+
for convolutions.
|
|
432
|
+
- Making a fast, general transformer inference engine or model library.
|
|
429
433
|
|
|
430
434
|
You may join our [Discord server](https://discord.gg/BW6YsCd4Tf) and chat with the community.
|