Anton Lozhkov

[WIP] Add Apple M1 tests

1

testing things live, don't merge

[ONNX] PyTorch IO binding for faster GPU inference

2

**Context** Currently [`OnnxStableDiffusionPipeline`](https://github.com/huggingface/diffusers/blob/main/src/diffusers/pipelines/stable_diffusion/pipeline_onnx_stable_diffusion.py#LL34C7-L34C34) perform unnecessary tensor casting between torch and numpy. The downsides of that are: * The pipeline code is harder to maintain: we have to keep track of...

Add from_pretrained telemetry

3

This telemetry will mostly just allow us to gauge the proportion of local vs hub pipeline loading. No potentially sensitive info about the (custom) pipeline class or the model path...

[MPS] Some of the `DPTForDepthEstimation` ops are not supported on MPS

2

To reproduce, remove the mps skips in `StableDiffusiondepth2imgPipelineFastTests`: https://github.com/huggingface/diffusers/blob/cd91fc06fe9513864fca6a57953ca85a7ae7836e/tests/pipelines/stable_diffusion_2/test_stable_diffusion_depth.py#L154 This is probably solvable on the `transformers` side, but opening an issue here too to keep track of the testing :)

### Context The #1526 PR implemented a `PipelineTesterMixin` with common fast tests for pipelines to inherit: saving/loading, casting to fp16, checking tuple outputs, attention slicing, etc. To support `PipelineTesterMixin`, a...

This PR updates the padding mask calculation to be the same as the one in the reference Wav2Vec2 implementation (same commit as listed in SEW's README): https://github.com/pytorch/fairseq/blob/05255f96410e5b1eaf3bf59b767d5b4b7e2c3a35/fairseq/models/wav2vec/wav2vec2.py#L477 For more details...

Config fixes for VLLMModel

4

Just re-adding some tiny but useful features from the base model back to VLLM LMK if you notice anything else!