| { | |
| "id": "coopq_quantizer", | |
| "name": "CoopQ: Cooperative Game-Theoretic Bit Allocation Quantizer with STE and Per", | |
| "file": "Codes/framework/compression_engine/coopq_quantizer.py", | |
| "type": "Compression Engine", | |
| "category": "Quantization & Mixed Precision", | |
| "paradigm": "quant", | |
| "icon": "\ud83d\udc8e", | |
| "description": "CoopQ: Cooperative Game-Theoretic Bit Allocation Quantizer with STE and Per-Channel Asymmetric Calibration. Registered in COMPRESSOR_REGISTRY as 'coopq'. Mathematical Formulation: 1. Min Bit-Width Floor Enforcement: $$b_l \\ge b_{\\min} = 4, \\quad \\forall l \\in \\{1, \\dots, L\\}$$ 2. Per-Channel Asymmetric Scale & Zero-Point Calibration: $$\\text{scale}_i = \\max\\left( \\frac{\\max_j W_{ij} - \\min_", | |
| "classes": [ | |
| "CoopQSTEFakeQuantize", | |
| "CoopQLinear", | |
| "CoopQQuantizer" | |
| ], | |
| "methods_count": 13, | |
| "math_latex": "b_l \\\\ge b_{\\\\min} = 4, \\\\quad \\\\forall l \\\\in \\\\{1, \\\\dots, L\\\\}", | |
| "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:
- 2.67 kB
- Xet hash:
- 102fec290260b2b3e6a03d020bdf30a5db83d120ed3088183756aad72e28d50f
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.