Hugging Face· PlatformDeveloper tool

Shipping a Trillion Parameters With a Hub Bucket: Delta Weight Sync in TRL

Hugging Face announces a new delta weight sync feature in TRL that encodes only the changed elements of model weights between RL optimizer steps as a sparse safetensors file, uploads it to a Hugging Face Bucket, and tells vLLM to fetch it, reducing per-step payload from 1.2 GB to 20-35 MB for Qwen3-0.6B.

From the source

We landed a TRL PR that encodes just the changed elements as a sparse safetensors file, uploads it to a Hugging Face Bucket, and tells vLLM to fetch it. On Qwen3-0.6B, the per-step payload drops from 1.2 GB to 20 to 35 MB.

huggingface.co

Announced

27 May 2026 00:00 UTC

Detected

03 Aug 2026 21:12 UTC (+68d after)