Prime Intellect added the missing backward pass to DeepSeek's fused MoE kernel, making it usable for training and running end-to-end training steps up to 1.34x faster
AspireAI community · Prime Intellect
MoE training on many GPUs spends a surprising share of its time not doing math but waiting for tokens to travel between the machine holding a token and the machine holding the expert that should process it. DeepSeek shipped a fused 'Mega MoE' kernel in its open-source DeepGEMM library that overlaps that communication with expert compute, but it only implemented the forward pass, so it could not be used to train. Prime Intellect says it built the missing overlapping backward kernel on top, which turns the kernel into something a trainer can actually use end to end.
The reported numbers are specific. Up to 3.11x faster MoE kernels against the default PyTorch implementation, and up to 1.34x faster end-to-end training than Prime Intellect's own Prime-RL baseline, described as 34 percent higher throughput. The benchmarks use 8xH200 per node with Qwen3-30B-A3B across 4 nodes and Qwen3-235B-A22B across 16 nodes, sequence length 16384, expert parallelism 8, FP8, full activation checkpointing, and torch.compile(fullgraph=True), averaging steps 6 to 20 after warmup. Supported data types are bfloat16 everywhere, MXFP8 on Blackwell only, and blockwise FP8 on Hopper only.
This is a vendor engineering post with no outside replication, and the 3.11x figure is the best single-layer case rather than the typical one. The end-to-end comparison is against Prime Intellect's own trainer, so the honest read is 'this kernel is now trainable and fast on Hopper and Blackwell', not 'every MoE training run gets 34 percent cheaper'.
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.