| { | |
| "id": "faq_quantizer", | |
| "name": "FAQ (Feature-Activation-Aware Quantization): adjusts quantization scales us", | |
| "file": "Codes/framework/compression_engine/faq_quantizer.py", | |
| "type": "Compression Engine", | |
| "category": "Quantization & Mixed Precision", | |
| "paradigm": "quant", | |
| "icon": "\ud83d\udc8e", | |
| "description": "FAQ (Feature-Activation-Aware Quantization): adjusts quantization scales using activation statistics. Registered as 'faq'.", | |
| "classes": [ | |
| "FAQLinear", | |
| "FAQQuantizer" | |
| ], | |
| "methods_count": 13, | |
| "math_latex": "W = S \\odot (Q_W - Z), \\quad Q_W \\in \\{-8, \\dots, 7\\}, \\quad S = \\text{scale}", | |
| "components": [ | |
| { | |
| "id": "scale_quant", | |
| "name": "Channel / Block Scales (FP8 / INT4)", | |
| "role": "Dynamic Range Quantization", | |
| "shape_transform": "[d_{out} // block_size]", | |
| "color": "#a855f7", | |
| "math_latex": "s = \\frac{\\max(|W|)}{2^{b-1} - 1}, \\quad b=4" | |
| }, | |
| { | |
| "id": "packed_weight", | |
| "name": "Packed Low-Bit Integer Weight Matrix", | |
| "role": "4-bit Memory Footprint", | |
| "shape_transform": "[d_{out} \u00d7 d_{in} // 2]", | |
| "color": "#c084fc", | |
| "math_latex": "q = \\text{clamp}\\left(\\left\\lfloor \\frac{W}{s} \\right\\rceil, -8, 7\\right)" | |
| }, | |
| { | |
| "id": "dequant_kernel", | |
| "name": "Fused GEMM Dequantization Kernel", | |
| "role": "Zero-Overhead Accumulation", | |
| "shape_transform": "TensorCore Fused", | |
| "color": "#e9d5ff", | |
| "math_latex": "\\hat{W} = q \\times s, \\quad Y = X \\hat{W}^\\top" | |
| } | |
| ], | |
| "model_projections": { | |
| "distilbert": { | |
| "hidden_size": 768, | |
| "intermediate_size": 3072, | |
| "baseline_layer_params": 6488064, | |
| "compressed_layer_params": 1622016, | |
| "savings_ratio": 4.0 | |
| }, | |
| "roberta": { | |
| "hidden_size": 768, | |
| "intermediate_size": 3072, | |
| "baseline_layer_params": 6488064, | |
| "compressed_layer_params": 1622016, | |
| "savings_ratio": 4.0 | |
| }, | |
| "gpt2": { | |
| "hidden_size": 768, | |
| "intermediate_size": 3072, | |
| "baseline_layer_params": 6488064, | |
| "compressed_layer_params": 1622016, | |
| "savings_ratio": 4.0 | |
| }, | |
| "qwen": { | |
| "hidden_size": 896, | |
| "intermediate_size": 4864, | |
| "baseline_layer_params": 11124736, | |
| "compressed_layer_params": 2781184, | |
| "savings_ratio": 4.0 | |
| } | |
| } | |
| } |
Xet Storage Details
- Size:
- 2.34 kB
- Xet hash:
- 341dc56ff73dffe54877cea3cb4d9ba9704274656d75b4fbdefef40c0528c0bd
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.