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: Implemented from the paper (arXiv:2504.19874), TurboQuant_mse and TurboQuant_prod.
• QTIP: Implemented from the paper (arXiv:2406.11235) and its reference code, bitshift trellis with tail-biting Viterbi and the 1MAD, 3INST and HYB codes.
• IQ_K (Shifted Tables): Ported from ik_llama.cpp's iqk_quantize.cpp.
• Trellis (KT): Ported from ik_llama.cpp's iqk_quantize.cpp.
• IQ-Quant (Grid-Based): Ported from llama.cpp's ggml-quants.c.

Interface

• Data Generation: Generate synthetic weights using Normal, Laplace, Bimodal, or Outlier-heavy distributions.
• Presets: Laid out by bits per weight (rows) and family (columns), type to filter.
• Rotation: Hadamard, Randomized Hadamard or Random Orthogonal rotation in front of any quant, with a toggle to view the rotated values.
• 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.