Full-Stack FP4: Stable LLM Pretraining with Quantized Projections, Optimizers, and Attention
TL;DR
Full-Stack FP4 is presented, a modular NVFP4 framework with separate recipes for projections, AdamW states, Root/Muon computation, and attention, and mixed-precision attention that protects a compact projection subspace in BF16 while retaining full-shape NVFP4 computation.
Abstract
Recent NVFP4 pretraining work has primarily optimized Transformer linear projections, leaving persistent optimizer states, optimizer computation, and low-precision attention forward--backward paths less explored. We present \textbf{Full-Stack FP4}, a modular NVFP4 framework with separate recipes for projections, AdamW states, Root/Muon computation, and attention. \textbf{LoRA-SVD} protects a compact projection subspace in BF16 while retaining full-shape NVFP4 computation, reducing the linear-only loss gap from \textbf{1.40\%} to \textbf{0.61\%}. An ordered square-root, tile-mean, and Hadamard pipeline enables stable NVFP4 AdamW momentum storage; shape-dependent coefficients and clipping stabilize direct NVFP4 Root iterations; and mixed-precision attention retains softmax-sensitive operations in BF16. On 3B pretraining with 64B tokens, BF16 and Full-Stack FP4 reach losses of \textbf{2.267} and \textbf{2.286}, a \textbf{0.838\%} gap. Their average zero-shot perplexities are 26.675 and 26.665, respectively, with Full-Stack FP4 averaging 0.10 percentage points lower in accuracy. Native four-block measurements on one RTX 5090 show 2.50--2.83$\times$ Root speedups over optimized BF16 and 37.9--42.5\% lower AdamW peak memory.