Garry Ling

right! I encounter this issue when I using u8 ndarray as passed value in JIT which parameter is u32 ndarray, taichi JIT convert the type automaticly. But in AOT mode,...

https://pytorch.org/executorch/stable/build-run-mediatek-backend.html as reference, I got sdk and mtk adapter header from executorch repo, and mtk adapter api is quit similar as nnapi interface. @jxt1234

@wangzhaode @jxt1234 Thus far, we have successfully implemented basic support for all operators in ResNet101 (such as convolution, addition, raster, etc.), demonstrating that this approach is indeed viable. Moreover, on...

> So can we use this PR for MTK APU acceleration in ResNet model for now? @FranzKafkaYu yep，this pull request is already done all dependency for running restnet on mtk...

> > > So can we use this PR for MTK APU acceleration in ResNet model for now? > > > > > > @FranzKafkaYu yep，this pull request is already...

more: convert to mil, show every shape is correct ``` main[CoreML3](%x: (1, 3, 512, 512, fp32)(Tensor), %grid: (1, 256, 256, 2, fp32)(Tensor)) { block0() { %x_to_fp16: (1, 3, 512, 512,...

> Would that be of any help the [LM Studio has implemented MLX](https://github.com/lmstudio-ai/mlx-engine). And here is [Anemll ANE library](https://github.com/Anemll/Anemll) to work with MLX it is MIT Licensed. And there's [FastMLX](https://github.com/arcee-ai/fastmlx)...

> I'm not understanding the issue. Can you give us some toy code that demonstrates the issue? @TobyRoseman ``` import torch import torch.nn as nn class SimpleCNN(nn.Module): def __init__(self): super(SimpleCNN,...

@YifanShenSZ Is there any timeline or official version will coremltools fix this bug? or If I can contribution for is feature?

> Is there an update on this issue? Maybe a workaround while it's being fixed? @aleksey-chugaev no time for this issue currently, maybe can export using torch jit trace to...