@jax-js/jax 0.0.2 → 0.0.3
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 +9 -8
- package/dist/{backend-1eVbAoaV.js → backend-BqDtPGaR.js} +1869 -86
- package/dist/{backend-BK21PBVP.cjs → backend-D2C4MJRP.cjs} +1892 -85
- package/dist/index.cjs +737 -118
- package/dist/index.d.cts +247 -44
- package/dist/index.d.ts +247 -44
- package/dist/index.js +726 -114
- package/dist/{webgpu-JVpVad6g.js → webgpu-CNg9JGva.js} +54 -33
- package/dist/{webgpu-c5Fe8nx8.cjs → webgpu-fqhx41TC.cjs} +54 -33
- package/package.json +7 -6
package/README.md
CHANGED
|
@@ -1,6 +1,6 @@
|
|
|
1
1
|
# jax-js: JAX in pure JavaScript
|
|
2
2
|
|
|
3
|
-
[Website](https://www.ekzhang.com/jax-js/)
|
|
3
|
+
[Website](https://www.ekzhang.com/jax-js/) | [API Reference](https://www.ekzhang.com/jax-js/docs/)
|
|
4
4
|
|
|
5
5
|
This is a machine learning framework for the browser. It aims to bring JAX-style, high-performance
|
|
6
6
|
CPU and GPU kernels to JavaScript, so you can run numerical applications on the web.
|
|
@@ -65,15 +65,14 @@ pnpm test
|
|
|
65
65
|
|
|
66
66
|
## Next on Eric's mind
|
|
67
67
|
|
|
68
|
+
- Finish CLIP inference demo and associated features (depthwise convolution, vmap of gather, etc.)
|
|
68
69
|
- Fix jit-of-grad returning very incorrect result
|
|
69
|
-
- Probably add static_argnums to jit() so that clip and some nn functions have jit added
|
|
70
70
|
- Improve perf of MNIST neural network
|
|
71
|
-
-
|
|
71
|
+
- Optimize conv2d further (maybe blocks -> local dims?)
|
|
72
|
+
- Add fused epilogue to JIT
|
|
72
73
|
- Reduce kernel overhead of constants / inline expressions
|
|
73
74
|
- Investigate why jax-js Matmul is 2x slower on Safari TP than unroll kernel
|
|
74
75
|
- How many threads to create per workgroup, depends on hardware
|
|
75
|
-
- Think about two-stage `cumsum()`
|
|
76
|
-
- Frontend transformations need to match backend type for pureArray() and zeros() calls
|
|
77
76
|
|
|
78
77
|
## Milestones
|
|
79
78
|
|
|
@@ -97,11 +96,13 @@ pnpm test
|
|
|
97
96
|
- [ ] Memory scheduling, buffer allocation (can be tricky)
|
|
98
97
|
- [ ] Demos: Navier-Stokes, neural networks, statistics
|
|
99
98
|
- [x] Features for neural networks
|
|
100
|
-
- [
|
|
99
|
+
- [x] Convolution
|
|
101
100
|
- [x] Random and initializers
|
|
102
|
-
- [
|
|
103
|
-
- [
|
|
101
|
+
- [x] Optimizers (optax package?)
|
|
102
|
+
- [x] Wasm backend (needs malloc)
|
|
103
|
+
- [x] Better memory allocation that frees buffers
|
|
104
104
|
- [ ] SIMD support for Wasm backend
|
|
105
|
+
- [ ] Async / multithreading Wasm support
|
|
105
106
|
- [ ] Device switching with `.to()` between webgpu/cpu/wasm
|
|
106
107
|
- [ ] numpy/jax API compatibility table
|
|
107
108
|
- [ ] Import tfjs models
|