AMEVA-Forge Documentation

Release 2.0.0 Initializing WebGPU... [ Live WebGPU Studio ] [ GitHub Repository ]

7. INT4 / INT8 Quantization & Dequantization Lab

Analyzes weight compression, cosine similarity, MSE loss error, and live on-the-fly dequantization dispatch performance.

Quant Scheme: INT4 (AWQ)
VRAM Compression: -75.0%
Cosine Sim: 0.9992
MSE Error: 1.4e-4

Quantization Configuration

Compression & Fidelity Analysis

Original FP32 VRAM
224 MB
Quantized VRAM
56 MB
• Cosine Fidelity: 0.9992 (Preserves >99.9% token ranking)
• Mean Squared Error: 1.4e-4
• On-The-Fly WebGPU Dequant Dispatch: PASS (No VRAM Stall)

Technical & Mathematical Deep-Dive