Ovis2-8B-LAYER-PPL / pruning_info.json
Azaz666's picture
Layer-pruned Ovis2-8B (PPL-based (Shortened LLaMA), removed 8 layers, 20.9% reduction)
bda6cd6 verified
{
"base_model": "AIDC-AI/Ovis2-8B",
"method": "layer_pruning_ppl",
"method_name": "PPL-based (Shortened LLaMA)",
"layers_removed": [
4,
5,
7,
8,
10,
11,
13,
15
],
"n_layers_original": 28,
"n_layers_remaining": 20,
"params_before_M": 8935.3,
"params_after_M": 7070.8,
"param_reduction_pct": 20.9,
"benchmarks": {
"vqav2": {
"accuracy": 0.0,
"avg_latency_s": 0.5913,
"peak_memory_mb": 18675.1,
"avg_memory_mb": 18638.7,
"throughput_sps": 1.69,
"avg_power_w": 289.8,
"avg_gpu_util_pct": 94.2,
"n_samples": 50,
"n_evaluated": 50,
"n_skipped": 0,
"all_failed": false,
"zero_accuracy_warning": true,
"metrics": {
"exact_match": 0.0,
"contains": 0.26,
"token_f1": 0.0731,
"bleu": 0.0457,
"rouge_l": 0.0731
}
}
}
}