Nemotron-H 8B Model FP4 Pretraining
UE5M3 FP4 Block Scaling for Stable Language Model Pretraining
This method improves efficiency for large language model training with quantization, potentially reducing costs for AI developers in the region.
The researchers developed a new FP4 pretraining method using E5M3 block scales for the Nemotron-H 8B model. They trained the model on nearly 190 billion tokens, achieving lower training and validation loss compared to NVIDIA's Transformer Engine method. Their approach increased token throughput by 21.2% when removing certain computational steps.
UE5M3 FP4 Block Scaling for Stable Language Model Pretraining
Stable 4-bit floating-point (FP4) pretraining is difficult because the E2M1 payload represents only a narrow range of magnitudes. NVIDIA's Transformer Engine \nv{} recipe addresses this with current-tensor scaling, a randomized Hadamard transform (RHT), and bfloat16 (BF16) final layers, adding work outside the FP4 matrix multiplications. We instead pair E2M1 payloads with unsigned E5M3 (\ue{}) block scales. Their wider range permits periodic tensor scaling, while our recipe applies selective stochastic rounding to backward gradients, omits RHT, and uses FP4 in all eligible internal linears. We pretrain a Nemotron-H 8B model for nearly 190 billion tokens. Compared with Transformer Engine \nv{}, the proposed block-16 recipe finishes with lower final-window training loss and, under their respective quantized-inference policies, lower validation loss measured as held-out negative log-likelihood. Its quantized-inference downstream point estimates are also higher on all three reported aggregates. A native \nv{} execution ablation that jointly removes RHT and the BF16 final-block exemption increases measured model-body token throughput by 21.2\%. These results demonstrate end-to-end software-emulated \uefp{} pretraining with a simpler recipe and motivate native support for \ue{} block scaling.
