#INT4量化

Skipping backpropagation means on-device fine-tuning no longer needs training-grade memory, letting a shared 8B model be personalized overnight on consumer hardware like an iPhone 16 Pro. The 89-100% success-rate jump for a phone-use agent suggests practical deployment paths for on-device agents handling private data. Practitioners working on edge AI, quantization or RLVR-style training should examine FuRA-S as an alternative to LoRA-based ES methods.

The paper shows the quality-price of aggressive quantization is now measured, not guessed: INT4 weights plus a Hadamard-rotated INT4 cache cost under one perplexity point while roughly tripling decode speed on H100. Engineers serving Llama-3.1-class models on Hopper GPUs can use its fragment-layout dequantization and KV-cache placement techniques to cut latency without falling back to padded prefill. Compared with PyTorch's INT4 kernel, which loses to BF16 at batch 32, this design stays fast up to batch 32.