We present \textbf{ITQ3\_S} (Interleaved Ternary Quantization -- Specialized), a novel 3-bit weight quantization format for large language models (LLMs) that integrates \textbf{TurboQuant (TQ)}, a rotation-domain adaptive quantization strategy based on the Fast Walsh-Hadamard Transform (FWHT). Conventional 3-bit quantization methods suffer from catastrophic precision loss caused by heavy-tailed weight distributions and inter-channel outliers. ITQ3\_S addresses this fundamental limitation by pre-rotating the weight space via FWHT prior to quantization, effectively spreading outlier energy across the entire vector and inducing a near-Gaussian distribution amenable to uniform ternary coding. Critically, we derive a mathematically rigorous dequantization procedure that inverts the FWHT exactly using a 256-point Inverse Walsh-Hadamard Transform fused into the CUDA shared-memory loading stage, ensuring zero-error round-trip fidelity between offline quantization and online inference. We prove that for any weight vector $\mathbf{w} \in \mathbb{R}^{256}$ processed by our pipeline, the reconstruction satisfies $\|\hat{\mathbf{w}} - \mathbf{w}\|_2 \leq ε_q$, where $ε_q$ is determined solely by the ternary quantization grid and is strictly smaller than any uniform 3-bit baseline under equal bit-budget constraints. Empirically, on the NVIDIA RTX 5090 (Blackwell architecture), ITQ3\_S achieves perplexity competitive with FP16 baselines while delivering throughput exceeding 1.5$\times$ that of 4-bit alternatives, owing to optimized DP4A and Tensor Core scheduling in the interleaved memory layout. Our results establish ITQ3\_S as a practical, mathematically grounded solution for high-fidelity LLM deployment on consumer-grade hardware.

翻译：暂无翻译