Emin Orhan

Please release the evaluation code as well. "Dirty code" is much better than no code. I'm finding it hard to parse the provided eval logs for ViT-L/16. Please just use...

`max_batch_size` argument in `ModelArgs`

2

I'm just curious what the `max_batch_size` argument does in `ModelArgs`: https://github.com/pytorch/torchtitan/blob/d2a4904f58accc683c17c66a360026cb3c8109af/torchtitan/models/llama/model.py#L32 A quick search suggests that it doesn't actually seem to be used anywhere else in the code base, so...

Triton error on AMD GPUs

8

### 🐛 Describe the bug I'm trying to test this library on an HPC cluster with AMD MI250X GPUs, but I'm getting a weird seemingly Triton-related error specifically when I...

dcp.load fails on checkpoints prior to AdamW refactor

1

I recently upgraded to a nightly pytorch build (`2.7.0.dev20250221+rocm6.3`) and noticed that the dcp checkpoints saved with my prior build could no longer be loaded successfully. The issue seems to...