weight_converter OOMs: full GLM checkpoint kept in CPU RAM while GPU is idle

[xuxianghong12]

Summary

python -m tilert.models.preprocess.weight_converter --model_type glm-5 converts on CPU. GPU memory stays tiny for the whole run (e.g. nvidia-smi shows the python process at ~742 MiB on GPU). Converted tensors are kept in RAM until every layer is done. RSS grows with the full checkpoint (~700 GB+), and the process is OOM-killed.

What we observe

• Command:

python -m tilert.models.preprocess.weight_converter \
  --model_type glm-5 \
  --model_dir /path/to/GLM-5.1-FP8 \
  --save_dir /path/to/GLM-5.1-FP8-TileRT

• GPU: almost unused.
• RAM: RSS grows without bound as layers accumulate. The job dies with an OOM kill when RAM is exhausted.