Dan Dale

CI testing workflow dependency installation issue

1

## 🐛 Bug CI testing currently failing to install ffmpeg with the following error(s) `Failed to fetch http://azure.archive.ubuntu.com/ubuntu/pool/universe/f/ffmpeg/libavutil56_4.2.4-1ubuntu0.1_amd64.deb 404 Not Found [IP: 40.119.46.219 80]` ... PR is forthcoming that will...

Fine-Tuning Scheduler Tutorial Update for Lightning/PyTorch 2.2.0

2

Validated the Fine-Tuning Scheduler (FTS) tutorial for FTS/Lightning/PyTorch `2.2.0` (as of the final PyTorch RC and the [recent `2.2.0` Lightning commit](https://github.com/Lightning-AI/lightning/tree/c989a97aa196431ecb70e1a0f8c839cdeaba9300)). The only minor change in this PR is to...

`git-lfs` binary not resolving in .git post-checkout hook preventing a number of notebook updates from being published

2

## 🐛 Bug The `Lightning-AI.tutorials [publish]` Azure pipeline [has been failing](https://dev.azure.com/Lightning-AI/Tutorials/_build?definitionId=29&_a=summary) since October 12th. The issue was first observed a few months ago but apparently is rearing its ugly head...

Improve TransformerBridge optimizer compatibility via dual PyTorch/TransformerLens parameter access API

2

Resolves #1141 Firstly, thank you so much for building and maintaining TransformerLens - it's a seminal, foundational and invaluable component enabling increasingly vital open-source interpretability research! This PR fixes an...

Firstly, thank you so much for building and maintaining TransformerLens - it's a seminal, foundational and invaluable component enabling increasingly vital open-source interpretability research! This PR enhances the TransformerLens `generate()`...

Resolves #1140 Firstly, thank you so much for building and maintaining TransformerLens - it's a seminal, foundational and invaluable component enabling increasingly vital open-source interpretability research! This PR resolves device/dtype...

`TransformerBridge.parameters()` returns non-leaf tensors created by `einops.rearrange()`, which cannot be optimized by PyTorch optimizers. This breaks a fundamental PyTorch contract and interferes with users fine-tuning TransformerBridge models or using the...

TransformerLens exhibits device/dtype mismatches that cause failures in several testing contexts, including half-precision (float16/bfloat16) inference and multi-GPU scenarios. These issues manifest as type mismatches in attention operations and device synchronization...