ALTO trained a 20B MoE in MXFP4 and got within 0.007 validation loss of the BF16 baseline
AspireAI community · arXiv (AMD)
FP4 arithmetic promises large memory-bandwidth and throughput wins for LLM training, but the quantization error is severe enough to destabilize optimization if left alone. ALTO is an open-source training recipe for GPT-OSS-20B, a 20-billion-parameter mixture-of-experts model, on the MLPerf Small MoE Training benchmark using the MXFP4 microscaling format. It combines hybrid 1D/2D block quantization, randomized Hadamard transforms and stochastic rounding with a simplified weight de-oscillation scheme adapted from TetraJet-v2. On C4 validation at global batch size 16, MXFP4 with RHT, stochastic rounding and de-oscillation reached a validation loss of 3.3350 at 16,128 steps, just 0.007 above the BF16 baseline of 3.3283 and roughly half the gap that plain MXFP4 leaves open at 3.3418. Measured in steps to reach the MLPerf quality target of validation loss 3.34, de-oscillation cut the FP4 convergence overhead against BF16 from +20.0 percent to +5.0 percent. The authors also report negative end-to-end results for differential gradient estimation, outlier clipping and macro-block scaling despite operator-level SNR gains. The honest caveat: all experiments ran on simulated MXFP4 kernels on AMD MI300 hardware, and wall-clock speed was not optimized, so this is a numerics result rather than a throughput result.
Read the story and explore more from the AspireAI community.
Two sparse-autoencoder analyses have resisted scaling: matching semantically similar features across layers, and compressing large feature circuits into interpretable supernodes. This paper argues both are the same underlying problem, estimating semantic distances between SAE features that sit on different activation manifolds. Its representation treats each feature not as a single decoder vector but as an activation-weighted distribution over the hidden states expressing it, projects those distributions into a shared reference space, and compares them with Wasserstein distance. The representation is proved invariant to activation rescaling and stable under perturbation, and recovers true matches under finite-sample margin conditions. Empirically it outperformed decoder-vector and LLM-based baselines and captured subtle functional distinctions between related features, and it compressed large feature circuits into interpretable supernodes automatically. The limitation is that this is interpretability tooling: the paper demonstrates better feature matching and circuit compression, not that downstream interpretability claims become more accurate.