File size: 2,357 Bytes
2d86c06 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 | {
"model_name": "SpecKV-MLP16",
"description": "Lightweight acceptance rate predictor for adaptive speculative decoding gamma selection.",
"architecture": {
"type": "MLPRegressor",
"hidden_layers": [16],
"activation": "relu",
"output": "regression (acceptance rate, 0-1)"
},
"input_features": [
{"name": "mean_draft_entropy", "description": "Mean entropy of draft token distributions (bits)", "dtype": "float32"},
{"name": "mean_draft_confidence", "description": "Mean top-1 probability of draft tokens", "dtype": "float32"},
{"name": "max_draft_entropy", "description": "Max entropy across draft tokens in the step", "dtype": "float32"},
{"name": "min_draft_confidence", "description": "Min top-1 probability across draft tokens in the step", "dtype": "float32"},
{"name": "comp_enc", "description": "Compression level encoding: 0=fp16, 1=int8, 2=nf4", "dtype": "int"},
{"name": "gamma", "description": "Candidate speculation length to evaluate", "dtype": "int"}
],
"output": {
"name": "predicted_acceptance_rate",
"range": [0.0, 1.0],
"description": "Predicted fraction of draft tokens the target model will accept"
},
"usage": {
"gamma_selection": "For each candidate gamma in {2, 4, 6, 8}, predict acceptance rate. Select gamma that maximizes: predicted_ar * gamma + 1",
"overhead": "0.34ms per decision (4 forward passes through the MLP)"
},
"training": {
"data": "5112 step-level records from Phase 2 profiling",
"model_pair": "Llama-3.2-1B-Instruct (draft) / Llama-3.2-3B-Instruct (target)",
"compression_levels": ["fp16", "int8", "nf4"],
"gamma_values": [2, 4, 6, 8],
"tasks": ["code", "math", "chat", "summarization"],
"framework": "scikit-learn 1.x",
"random_seed": 42
},
"performance": {
"test_mse": 0.090,
"test_correlation": 0.685,
"decision_overhead_ms": 0.336,
"improvement_over_fixed4": "56.0%",
"statistical_significance": "p < 0.001 (paired bootstrap, 10K resamples)"
},
"files": {
"speckv_mlp16.pkl": "Full sklearn model (pickle format)",
"speckv_mlp16_weights.npz": "Raw numpy weights for framework-agnostic loading"
},
"license": "MIT",
"paper": "SpecKV: Adaptive Speculative Decoding with Compression-Aware Gamma Selection",
"repository": "https://github.com/Amorfati123/SpecKV"
}
|