xl weights are F32 not BF16

[DJ05GaZ]

Heads up on the xl weights — they're F32 not BF16

Despite the HuggingFace listing showing "Weights size (bf16)", the downloaded safetensors files are actually F32. You can verify this yourself by reading the header of any shard — every tensor dtype comes back as F32.

If you're self-hosting (RunPod network volume, etc.) this doubles your storage cost unnecessarily, and if your container or pipeline doesn't pre-convert, the model loads in F32 and has to convert to BF16 so your runtime converts on the fly — wasting VRAM and GPU time on every start.

I converted them to BF16 offline using a pure numpy script (no torch required) — 4 shards, ~15 minutes, 50% size reduction, verified dtype after each shard. The converted weights load and generate identically. This saves converting them each time you spin up.