search

tenstorrent / pytorch2.0_ttnn

⭐️ TTNN Compiler for PyTorch 2.0 ⭐️ It enables running PyTorch2.0 models on Tenstorrent hardware
https://tenstorrent.github.io/tt-metal/latest/ttnn/
25 stars 5 forks source link

Model Autoencoder (conv) #127

Closed ayerofieiev-tt closed 2 months ago

ayerofieiev-tt commented 2 months ago

Ticket

None

Problem description

This was merged and now reverted as it caused issues. Please try to run locally CC @cthsieh

Add model Autoencoder (conv) to trace.

Solved by writing a new test.

What's changed

Add model Autoencoder (conv).

cthsieh commented 2 months ago

We resolved the issue by manually crafting a convolutional autoencoder model instead of using one from a third-party package.

The third-party package, called "diffusers," is popular and widely used on Hugging Face, with hundreds of projects relying on it. However, the package primarily supports model inference through a "pipeline", which is not compatible with the Torch Dynamo compiler flow.

We've kept the original test but marked it as skipped because using the pipeline is still a valid approach, and we might need to support it in the future. Additionally, we created a new test that bypasses the pipeline, ensuring compatibility with our current compiler flow.