Fucheng Warren Zhu

Setting up hface environment in CausalImage_ImageModelBackbones.R

2

Currently hface environment is set up as follows in `CausalImage_ImageModelBackbones.R` ``` conda create -n hface python=3.11 conda activate hface python3 -m pip install --upgrade transformers torch tensorflow tensorflow_datasets pillow pip...

Complications with Building Backend with Apple Silicon Powered Devices

1

I am not entirely certain, but it seems like Macbook doesn't work with the current default install (at least when I do it manually) since it installs x86-64 versions of...

Error when Following the Build Backend Tutorial

7

**Issue** I encountered an issue while trying to create a new conda environment using the `causalimages::BuildBackend` function. The process fails with the following error: ``` Error while loading conda entry...

This may be a feature of the pipeline rather than a bug, but for RandomProjection ViT or CNN (and most likely other models, but I couldn't run it on my...

bmm, topk, cholesky, linalg.norm, max with out variants set causing r…

3

Fixes #135859, following the approach detailed in the [Dynamic Shapes Manual](https://docs.google.com/document/d/1GgvOe7C8_NVOMLOCwDaYV1mXXyHMXY7ExoewHqooxrs/edit?tab=t.0#heading=h.kzbllkiwjdpm). Adds support for torch.bmm, torch.topk, torch.linalg.cholesky, and torch.linalg.norm. For bmm.out and topk.out, added an out_wrapper decorator in torch/_meta_registrations.py. For...

Bash Mode sends the line you type straight to the persistent shell session backing Claude Code, and is great for quick commands exploring the directory. It can be enabled by...

I often need to visit responses to past prompts, or jump to the start of the prompt when I receive the response. Currently with Claude Code I haven't found a...

# Bug Report: `create_sdk_mcp_server()` passes unsupported `version` parameter causing TypeError ## Summary Agent deployment fails with `TypeError: Server.__init__() got an unexpected keyword argument 'version'` when using `create_sdk_mcp_server()`. The SDK passes...

Include Option to Keep Global Batch Size Constant

10

# Maintain Constant Global Batch Size Upon Failure With the current implementation of `DistributedSampler`, the `global_batch_size` is `group_batch_size * num_replica_group`. It may be more preferable if the DistributedSampler is implemented...

Reuse Valid Accumulated Gradients Upon Failure

4

**Description:** The current failure model in TorchFT handles node failures by zeroing out all accumulated gradients and recalculating them in the subsequent forward/backward pass. This approach, while ensuring correctness, leads...