Apoorv Khandelwal

Hi. I am training an ObjectNav model on an 8 GPU system (with 20 simulators per GPU, as to saturate GPU memory). I am only at 30-50% utilization per CPU...

Running gradslam at higher processing speed

2

## ❓ Questions and Help I am only able to get ~8.5 FPS when running gradslam step-by-step on GPU. I am wondering if it's possible to run gradslam at a...

Multi-node training with Deepspeed hangs when `full_determinism = True`

2

Hey, as I've described below, I think there are problems training Deepspeed in a multi-node setting when `full_determinism = True` in the `TrainingArguments`. I've replicated this on multiple hardware configurations...

Set FSDP `transformer_layer_cls_to_wrap` to `model._no_split_modules` ?

1

### Feature request Currently, when training with FSDP, the Trainer expects to receive an `fsdp_config` argument specifying `fsdp_transformer_layer_cls_to_wrap`. https://github.com/huggingface/transformers/blob/66954ea25e342fd451c26ec1c295da0b8692086b/src/transformers/trainer.py#L1394-L1406 I am wondering if we can set this automatically, when the...

Currently, when a class is not already in Tango's registry, Tango will import files in hopes to locate that class. In particular, Tango will import all files and modules in...

### 🐛 Describe the bug ## Problem I found that `det_hash` can sometimes hash equal objects differently. This was causing me a lot of grief when trying to manually initializing...

[BUG] Cannot `micropip.install("bokeh>=3.4.0")`

8

### Expected behavior I would like to be able to install `bokeh` in WebAssembly environments (e.g. using [marimo](https://marimo.io)). ```python import micropip await micropip.install("bokeh>=3.4.0") ``` Also relevant to: hvplot/holoviews (cc: @philippjfr)...

Hi, I was wondering how "canonical" `ormsgpack` is: i.e. will any two equal objects result in the same serialized bytes? I found this issue (https://github.com/msgpack/msgpack/issues/215) that says the `msgpack` specification...

Allow `Trainer.get_optimizer_cls_and_kwargs` to be overridden

3

# What does this PR do? Currently, `Trainer` builds an optimizer by loading the optimizer class and arguments from `Trainer.get_optimizer_cls_and_kwargs` in `Trainer.create_optimizer`: https://github.com/huggingface/transformers/blob/6c1d0b069de22d7ed8aa83f733c25045eea0585d/src/transformers/trainer.py#L1076 However, this prevents the `get_optimizer_cls_and_kwargs()` function from...