Hugging Face halves Qwen3 training-step time on H100 with TRL fix

Top News · Reported · runtimewire.com

For standard-attention models using best-fit-decreasing packing, the TRL fix changes the default batch layout. Hugging Face halved a Qwen3 training step on an H100 by changing how packed batches are arranged for standard-attention models.

Here’s what changed:

• Best-fit-decreasing packed examples now stay in separate batch rows, rather than being flattened into one long row per packed batch.
• The new layout produced step timings on AWS Trainium2, where the previous layout crashed.
• Hugging Face merged the TRL change on November 13, 2026.

One catch: The halved step time comes from a reported Qwen3 H100 test, and the layout change applies to standard-attention models.

Why care? The layout change both halved the reported Qwen3 H100 step time and produced timings on Trainium2, where the old layout crashed.

Sources: runtimewire.com