The cadence: one anchor every N steps, deltas in between
The trainer emits a full anchor every delta_sync_anchor_interval
steps and a sparse delta on every other step. Watch the bucket fill up. The mocked UI on the right mirrors what you would
see on huggingface.co/buckets/<user>/<run>.
trainer
Llama-3.1-405B · bf16
step
0
idle
aminediroHF/llama405b-async-deltas
0 B · 0 files
⨯ xet
📁anchors/
0 items
📁deltas/
0 items
training steps
anchor (full snapshot, ~810 GB)
delta (sparse patch, ~6 GB)
future step
Xet content-defined chunking deduplicates across anchors automatically. Two consecutive anchors share ~99% of their chunks,
so the actual stored bytes are far smaller than the naive sum of file sizes. The visible total above is the logical sum;
the physical storage cost on the Hub is much lower.