ZhiyuLi-goog issues

Repositories
Issues
Comments

Results 3 issues of


                                            ZhiyuLi-goog

fix dtype bug in adam_pax

# [[Bug] adam_pax has reuse donated buffer warning](https://github.com/google/maxtext/issues/490) Reproduced with `weight_dtype=bfloat16` ```shell python3 MaxText/train.py MaxText/configs/base.yml run_name=run steps=10 weight_dtype=bfloat16 opt_type=adam_pax dataset_type=synthetic enable_checkpointing=false ``` ``` /home/lizhiyu/.local/lib/python3.10/site-packages/jax/_src/interpreters/mlir.py:914: UserWarning: Some donated buffers were not...

update extract_anthropic_prompt

# Description A slightly better prompt split using the common prefix instead of relying on the right-most split. In Anthropic's HH multi-round conversations, the majority of differences occur in the...

better core dump for debugging

## Features upload more detailed info to gcs bucket in failures Switch to `gcloud storage cp`, since it around twice as fast as `gsutil -m cp`. ## Testing / Documentation...