vLLM says its method keeps train-inference KL at exactly 0 on Nemotron-3.5-Lightning-30B-A3B, a Mamba-2 hybrid with NVFP4, in a filtered SWE-bench Pro training subset.
Training reward rises from 0.56 to 0.59, while the baseline run falls from 0.56 to 0.22, comparing the first and last 10 steps.
Rollout throughput reaches 87.5% of baseline over the first 30 steps, and the post notes these are training results, not held-out benchmark scores.
Part of a thread · earlier post
4/ The method transfers to Nemotron-3.5-Lightning-30B-A3B, a Mamba-2 hybrid with NVFP4.
On a filtered SWE-bench Pro training subset, train–inference KL stays exactly 0. Training reward holds at 0.56 → 0.59, while this baseline run falls from 0.56 → 0.22 (first 10 vs. last 10 steps).
Rollout throughput: 87.5% of baseline over the first 30 steps. These are training results, not held-out benchmark scores.
Source: vLLM · x.comPublished · added here
