The Pulse
TRACE Reports 5.4× Faster FP4 Rollouts for MoE Training
A new TRACE paper reports up to 5.4× faster FP4 rollouts than BF16 on one long-output test of a Qwen mixture-of-experts model. Its method uses rollout-side rounding results to guide training and aims to keep benchmark performance near the B
AI.info Team · Written with AI ·
On October 6, researchers posted TRACE, a method for running reinforcement-learning rollouts in four-bit floating point while training the model to follow the same rounding decisions. The paper reports up to 5.4× higher decoding throughput than a BF16 rollout at 128K output length. The result comes from tests on Qwen3.5-35B-A3B, not a general guarantee for other models or hardware.
TRACE targets disagreement between rollout and training
Rollouts are the sequences a language model generates during reinforcement-learning training. Those generations supply the experience used to update the policy, but low-precision rounding can make the rollout and training paths calculate different values. In a mixture-of-experts model, numerical differences can also affect which expert handles a token, adding another source of mismatch.
TRACE records quantization outcomes from rollout and uses them to guide training-side FP4 rounding. The authors pair records with training activations by trajectory, token position and quantization site. To limit the storage and transfer burden, their method retains mantissa and scale information from selected deeper layers rather than saving every quantized value.
“Our results demonstrate that TRACE enables joint FP4 weight/activation and FP4 KV-cache rollout with RL performance comparable to BF16 rollout, while achieving up to 5.4× rollout speedup and strong final FP4 performance compared with post-hoc FP4 quantization of BF16-trained policies.”
Xin Wang et al., “TRACE: Rollout-Guided Quantization-Aware Training for FP4 Reinforcement Learning of MoE Language Models”
One bit carries the rounding hint
In the paper’s default configuration for Qwen3.5-35B-A3B, TRACE keeps a one-bit mantissa hint and scale information for the latter 20 of the model’s 40 MoE layers. The authors report that more than 99% of mismatched quantized values differ by one neighboring FP4 codebook entry, which motivates using a small hint instead of transferring full quantization records. They estimate that this cache takes about 7.5 KB per generated token, roughly six times less than caching the full quantized activation.
On four reasoning benchmarks, TRACE records an average score of 75.3, compared with 74.9 for the BF16 rollout reference and 68.8 for QUADS, another FP4 approach. The comparison covers LiveCodeBench, AIME24, AIME25 and HMMT25. Results on three additional MoE models also sit near the reported BF16 scores: TRACE reaches 33.0 versus 33.4 on DeepSWE, 70.6 versus 68.8 on Terminal-Bench, and 90.2 versus 90.3 on GDPval.
The speed figure is a rollout measurement
The 5.4× peak applies to decoding at 128K output length on Qwen3.5-35B-A3B. The authors measured rollout throughput using SGLang on four GB200 GPUs; the result compares generation speed, not the full reinforcement-learning training process. End-to-end, TRACE raises the RL step time from 664 to 713 in its reported test, a 7.4% increase over vanilla joint FP4 rollout on a separate setup using 72 GB200 GPUs.
The method also leaves some mismatch unresolved. The paper says asynchronous policy updates can make training activations differ from the ones recorded during rollout; TRACE reduces extra discrepancy from inconsistent FP4 rounding but does not eliminate that underlying difference. Its evidence is limited to four MoE models, selected tasks and specified GPU setups, so the headline speedup should be read as a result from those conditions.