Quantization-Visualizer

Overview

A visualizer for floating-point quantization. It shows how different schemes map raw weights to low-bit integers and the resulting reconstruction error.

Supported Quants

• Primitive (FP Types): Standard floating-point formats (FP32, FP16, BF16, FP8) applied element-wise without additional block-level scaling.
• NVFP4 (Blackwell): NVIDIA’s hardware-native 4-bit format using an E2M1 codebook. Employs hierarchical scaling where blocks of 16 share an FP8 scale, and entire tensors share a global FP32 scale.
• MXFP (OCP Microscaling): Open standard for micro-granularity scaling. Blocks of 32 weights share an 8-bit exponent-only (E8M0) scale factor, supporting sub-8-bit element formats (MXFP4, MXFP6, and MXFP8).
• Symmetric (Block Scale): Block-wise quantization where weights share a single scale factor per block.
• Asymmetric (Block Scale + Zero): Block-wise quantization with both a scale and a minimum offset per block.
• K-Quant (Nested Scales): Superblock structure where sub-block scales are themselves quantized by a shared super-scale.
• TurboQuant (Scalar WHT): SRHT rotation normalizes the coefficient distribution, then a precomputed Lloyd-Max codebook quantizes each coordinate independently. Optional QJL adds a 1-bit residual correction to remove inner-product bias.
• Trellis (Viterbi WHT): SRHT rotation followed by a multi-state Viterbi search that minimizes MSE over the full sequence rather than per value independently, using Ungerboeck set partitioning to increase effective codebook resolution.
• IQ-Quant (Grid-Based): Block quantization using learned grids with per-block scales, auxiliary metadata, and sign bits. Variants: IQ1_S, IQ1_M, IQ2_XXS, IQ2_XS, IQ2_S, IQ3_XXS, IQ3_S, IQ4_XS, IQ4_NL.

Interface

• Data Generation: Generate synthetic weights using Normal, Laplace, Bimodal, or Outlier-heavy distributions.
• Inspector: Hover over bars to see the exact reconstruction math (Q * Scale + Offset) and error stats for that specific block.
• Stats: Live calculation of Bits-Per-Weight (BPW), compression ratio, and global Mean Squared Error (MSE).

License

MIT. (c) Alexander Kurtz 2026.