executorch
executorch copied to clipboard
pytorch
On-device AI across mobile, embedded and edge for PyTorch
Reviewed By: kimishpatel Differential Revision: D63730600
Summary: A couple tests and helpers that my previous passes didn't catch. After this, the only remaining `//executorch/...` code under the `torch::` namespace is under `backends/...` and `kernels/...` (and in...
Summary: Move the XNNPACK backend out of the `torch::` namespace, and update to avoid using the `torch::` or `exec_aten::` namespaces. Also update utils.h to use a namespace that matches the...
Summary: As title, separate the 16a8w matmul custom annotation from the large chunk of PR. Differential Revision: D63876215
The v0.4.0 release will be cut from the "[release/0.4](https://github.com/pytorch/executorch/tree/release/0.4)" branch for critical fixes to the [v0.4.0](https://github.com/pytorch/executorch/releases/tag/v0.4.0) release. Branch Cut Date: 09/15/2024 Intended Release Date: 10/17/2024 This issue is for tracking...
Fix the `buck2 kill` command. Because of scoping issues, in some cases we only ran "` kill`" because the local value of `$BUCK2` was empty. This should help avoid failures...
Summary: Migrate all `//executorch/...` code to use the new threadpool namespace. Differential Revision: D63783748
For situations where the forward has non-position arguments, such as https://github.com/pytorch/torchtune/blob/3c450ef5f1fbe8237f899e942fd5222491a47ca7/torchtune/modules/transformer.py#L519
Stack from [ghstack](https://github.com/ezyang/ghstack) (oldest at bottom): * __->__ #5861 We should use this option during exporting 1B/3B models as bf16 because KVCache is always fp32. Otherwise, we see regressed performance...
### Summary - Removes redundant steps in the Llama2 export - Factors out checkpointing to be shared with future Llama models (namely 3.2 multimodal) - Comments and orders code more...