onnxruntime/js/web/test/data/ops
Arthur Islamov fac3e33da5
[js/web] JSEP Attention & MultiHeadAttention (#17742)
### Description
This is a narrow implementation of Attention/MultiHeadAttention as it
does not support:
a. inputs 5-7 for MHA
b. packed QKV/KV
c. past/present
d. attention mask

But it works well for StableDiffusion and can be extended later. It
reduces VRAM usage as it combines many ops into few
I've updated demo here https://islamov.ai/stable-diffusion-webgpu/ it
takes ~13sec for 1 image with 20 steps on RTX3090Ti and about 25s on M1
Pro
VRAM usage is about 8gb if you don't use img2img

Going to focus on SDXL now

---------

Co-authored-by: Guenther Schmuelling <guschmue@microsoft.com>
Co-authored-by: Yulong Wang <7679871+fs-eire@users.noreply.github.com>
2023-11-17 12:23:52 -08:00
..
_example.jsonc [js/web] allow optional input/output in operator test (#17184) 2023-08-16 11:50:11 -07:00
abs-int32.jsonc [js/webgpu] Include Support for neg.int32 (#17374) 2023-09-06 12:00:16 -07:00
abs.jsonc
absr.jsonc
abss.jsonc [js] enable formatter for more file types (#16888) 2023-07-28 15:46:58 -07:00
acos.jsonc [js] enable formatter for more file types (#16888) 2023-07-28 15:46:58 -07:00
add.jsonc [js] enable formatter for more file types (#16888) 2023-07-28 15:46:58 -07:00
add_int32.jsonc
and.jsonc
asin.jsonc [js] enable formatter for more file types (#16888) 2023-07-28 15:46:58 -07:00
attention.jsonc [js/web] JSEP Attention & MultiHeadAttention (#17742) 2023-11-17 12:23:52 -08:00
bias-add.jsonc [js/web] BiasSplitGelu and BiasAdd kernels (#17161) 2023-10-03 12:20:20 -07:00
bias-split-gelu.jsonc [js/web] BiasSplitGelu and BiasAdd kernels (#17161) 2023-10-03 12:20:20 -07:00
cast.jsonc [js/webgpu] support Cast operator (#16489) 2023-08-18 23:51:03 -07:00
ceil.jsonc
concat.jsonc [js] enable formatter for more file types (#16888) 2023-07-28 15:46:58 -07:00
concat_int32.jsonc [JS/WebGPU] support Concat.int32 operator (#17003) 2023-09-13 00:05:00 -07:00
conv-transpose.jsonc [JS/Web] Add ConvTranspose implementation using MatMul (#17573) 2023-09-29 11:00:44 -07:00
conv.jsonc [js/webgpu] Enable the NCHW ConvMatMul path (#17717) 2023-10-05 00:26:01 -07:00
cos.jsonc [js] enable formatter for more file types (#16888) 2023-07-28 15:46:58 -07:00
depth-to-space.jsonc [js] enable formatter for more file types (#16888) 2023-07-28 15:46:58 -07:00
div.jsonc [js] enable formatter for more file types (#16888) 2023-07-28 15:46:58 -07:00
div_int32.jsonc [js/webgpu] Support int32 type for binary (#16901) 2023-08-18 12:19:01 -07:00
einsum.jsonc [JS/Web] Added Einsum operator support. (#17401) 2023-09-11 15:57:15 -07:00
equal.jsonc
exp.jsonc [js] enable formatter for more file types (#16888) 2023-07-28 15:46:58 -07:00
expand.jsonc [JS/WebGPU] Expand operator fixes (#17137) 2023-08-16 11:24:26 -07:00
floor.jsonc
fused-conv.jsonc [JS/Web]Added FusedConv. (#17766) 2023-11-01 15:34:51 -07:00
gather-elements.jsonc [JS/WebGPU] Support GatherElements kernel (#17243) 2023-08-28 09:55:25 -07:00
gather.jsonc
gelu.jsonc [js/web] update op test schema (#16921) 2023-08-03 14:20:20 -07:00
gemm.jsonc
global-average-pool.jsonc [js] enable formatter for more file types (#16888) 2023-07-28 15:46:58 -07:00
greater.jsonc
identity.jsonc
image-scaler.jsonc [js] enable formatter for more file types (#16888) 2023-07-28 15:46:58 -07:00
instance-norm.jsonc [js/webgpu] Optimize InstanceNormalization (#17491) 2023-09-14 17:03:18 -07:00
layer-norm.jsonc [JS/Web] The bias input is optional, not required, for LayerNormalization operator (#17143) 2023-08-16 10:41:20 -07:00
leaky-relu.jsonc
less.jsonc
log.jsonc [js] enable formatter for more file types (#16888) 2023-07-28 15:46:58 -07:00
matmul-broadcast.jsonc [js/webgpu] add matmul broadcast tests (#17335) 2023-09-05 20:41:46 -07:00
matmul.jsonc [js/webgpu] Optimize matmul (#16969) 2023-08-29 12:40:57 -07:00
mul.jsonc [js] enable formatter for more file types (#16888) 2023-07-28 15:46:58 -07:00
mul_int32.jsonc
multi-head-attention.jsonc [js/web] JSEP Attention & MultiHeadAttention (#17742) 2023-11-17 12:23:52 -08:00
neg-int32.jsonc [js/webgpu] Include Support for neg.int32 (#17374) 2023-09-06 12:00:16 -07:00
neg.jsonc
not.jsonc
or.jsonc
pad-big.jsonc [js/web] update op test schema (#16921) 2023-08-03 14:20:20 -07:00
pad.jsonc [js/web] update op test schema (#16921) 2023-08-03 14:20:20 -07:00
pow-big-number.jsonc [js/web] update op test schema (#16921) 2023-08-03 14:20:20 -07:00
pow.jsonc
pow_int32.jsonc [js/webgpu] Support int32 type for binary (#16901) 2023-08-18 12:19:01 -07:00
reduce-min.jsonc
relu.jsonc
reshape-int32.jsonc
reshape-pack.jsonc
reshape.jsonc [js/webgpu] Fix reshape int32 test case (#17113) 2023-08-15 21:18:13 -07:00
resize-pack.jsonc [js/web] update op test schema (#16921) 2023-08-03 14:20:20 -07:00
shape.jsonc
sin.jsonc [js] enable formatter for more file types (#16888) 2023-07-28 15:46:58 -07:00
skip-layer-norm.jsonc [JS/Web] Added SkipLayerNormalization operator. (#17102) 2023-08-18 09:59:03 -07:00
slice.jsonc [JS/Web]Added uniforms support to Slice op. (#18422) 2023-11-16 09:44:13 -08:00
softmax.jsonc [js/webgpu] Support uniform for softmax (#18345) 2023-11-09 11:19:23 -08:00
split.jsonc [js/web] update op test schema (#16921) 2023-08-03 14:20:20 -07:00
sqrt.jsonc
sub.jsonc [js] enable formatter for more file types (#16888) 2023-07-28 15:46:58 -07:00
sub_int32.jsonc
tan.jsonc [js] enable formatter for more file types (#16888) 2023-07-28 15:46:58 -07:00
tile.jsonc [JS/WebGPU] Support Tile operator (#17123) 2023-08-18 10:07:21 -07:00
transpose.jsonc [js/webgpu] Fix the transpose error when dims > 4D (#18027) 2023-10-23 11:02:19 -07:00
transpose_int32_uint32.jsonc [js/WebGPU] Support int32 Transpose in WebGPU (#16952) 2023-08-02 16:27:24 -07:00
upsample.jsonc [js] enable formatter for more file types (#16888) 2023-07-28 15:46:58 -07:00
where.jsonc [js/webgpu] Support where (#17544) 2023-10-03 14:28:21 -07:00
where_broadcast.jsonc [js/webgpu] Support where (#17544) 2023-10-03 14:28:21 -07:00
xor.jsonc