search

pytorch / executorch

On-device AI across mobile, embedded and edge for PyTorch
https://pytorch.org/executorch/
Other
1.97k stars 325 forks source link

Is it possible to use Qualcomm Backend & ExecuTorch (experimental) training? #5417

Open escorciav opened 1 month ago

escorciav commented 1 month ago

📚 The doc issue

Hi @JacobSzwejbka,

I noticed that you have been participating (or orchestrating) the integration of ExecuTorch training, right? Really cool and quite appealing feat wrt my usual on-device pipeline based on QNN & SNPE where on-device training is not possible (or equally undocumented :thinking: ).

Could you kindly confirm if there is any example using ExecuTorch training with QNN Backends, in particular HTP/NPU?

Thanks in advance :blush:

Kudos to the team & keep shipping :ship: :raised_hands:

Suggest a potential alternative/fix

No response

guangy10 commented 1 month ago

cc: @JacobSzwejbka @cccclai

JacobSzwejbka commented 1 month ago

Currently there is no integration with backends to accelerate ET Training graphs. Its still pretty early days for ET Training, I am writing up a bunch of docs right now actually so you should see things a little more structured under executorch/extension/training over the next week or so. We have a goal to allow backends to integrate with training and generally be composable with the rest of the stack, but no specific plans with any specific backends to announce at this time.

cccclai commented 1 month ago

We're really early stage but we had some early discussion with Qualcomm regarding on device training with QNN. Please stay tuned