{
"adasvd_compressor": {
"module_name": "adasvd_compressor",
"source_file": "Codes/framework/compression_engine/adasvd_compressor.py",
"source_lines_count": 488,
"architecture": {
"class_name": "AdaSVDCompressor",
"base_class": "BaseCompressor",
"init_args": "self, model: nn.Module, rank: Optional[int] = None, target_param_ratio: float = 0.7, num_iters: int = 5, max_error: float = 0.40, fallback_to_original: bool = True, target_attention_only: bool = True, min_rank: int = 16, target_modules: Optional[List[str]] = None, comp_config: Any = None, **kwargs: Any, ) -> None: super().__init__(model, name=\"adasvd\", **kwargs) cfg = comp_config or kwargs.get(\"comp_config\", kwargs.get(\"cfg\", getattr(model, \"config\", None))) self.rank = _cfg_get(cfg, \"adasvd_rank\", _cfg_get(cfg, \"rank\", rank)) self.target_param_ratio = float( _cfg_get(cfg, \"adasvd_target_ratio\", _cfg_get(cfg, \"target_param_ratio\", target_param_ratio)) ) self.num_iters = int(_cfg_get(cfg, \"adasvd_iters\", _cfg_get(cfg, \"num_iters\", num_iters))) self.max_error = float(_cfg_get(cfg, \"adasvd_max_error\", _cfg_get(cfg, \"max_error\", max_error))) self.fallback_to_original = bool( _cfg_get(cfg, \"adasvd_fallback_to_original\", _cfg_get(cfg, \"fallback_to_original\", fallback_to_original)) ) self.target_attention_only = bool( _cfg_get(cfg, \"adasvd_target_attention_only\", _cfg_get(cfg, \"target_attention_only\", target_attention_only)) ) self.min_rank = int(_cfg_get(cfg, \"adasvd_min_rank\", _cfg_get(cfg, \"min_rank\", min_rank))) self.target_modules = target_modules self.comp_config = cfg self.compressed_layers: List[str] = [] self.fallback_layers: List[str] = [] self.layer_errors: Dict[str, float] = {} self.layer_ranks: Dict[str, int] = {} self.layer_iters: Dict[str, int] = {} def _is_target_module(self, name: str, module: nn.Module) -> bool: if isinstance(module, SVDLinear",
"registry_names": [
"adasvd"
]
},
"params": [],
"metrics": {
"results/unified_adasvd/roberta/roberta-base_full_full_metrics.json": {
"model": "roberta-base_full",
"performance": {
"accuracy_pct": 78.56,
"f1_score": 0.8124,
"f1_macro": 0.7811,
"precision": 0.8051,
"recall": 0.7832,
"val_loss": 0.4381,
"status": "PASS",
"elapsed_s": 6.77
},
"efficiency": {
"total_params_before": 124647170,
"total_params_after": 124647170,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 712.34,
"tokens_per_sec": 24228.27
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 683.27,
"L_256_vram_mb": 685.84,
"L_512_vram_mb": 699.47,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 514)"
},
"timestamp": "2026-09-15T09:56:11.500945"
},
"results/unified_adasvd/roberta/roberta-base_adasvd_full_metrics.json": {
"model": "roberta-base_adasvd",
"performance": {
"accuracy_pct": 73.39,
"f1_score": 0.7778,
"f1_macro": 0.7232,
"precision": 0.7685,
"recall": 0.7306,
"val_loss": 0.5381,
"status": "PASS",
"elapsed_s": 26.53
},
"efficiency": {
"total_params_before": 124647170,
"total_params_after": 109737258,
"trainable_recovery_params": 8713768,
"compression_rate": 1.1359,
"compression_ratio_pct": 11.96,
"flops_reduction_pct": 63.2,
"num_compressed_layers": 40,
"num_fallback_layers": 8,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": 0.5913941559039471,
"weighted_reconstruction_error": 0.6364,
"per_layer_reconstruction_errors": {
"roberta.encoder.layer.0.attention.self.query": 0.455972929347947,
"roberta.encoder.layer.0.attention.self.key": 0.4377989629219472,
"roberta.encoder.layer.0.attention.self.value": 0.6528715799546899,
"roberta.encoder.layer.0.attention.output.dense": 0.6181080091945617,
"roberta.encoder.layer.1.attention.self.query": 0.5507851348101068,
"roberta.encoder.layer.1.attention.self.key": 0.5445418526049741,
"roberta.encoder.layer.1.attention.self.value": 0.6285729613390068,
"roberta.encoder.layer.1.attention.output.dense": 0.6149361344300587,
"roberta.encoder.layer.2.attention.self.query": 0.5717057309786219,
"roberta.encoder.layer.2.attention.self.key": 0.5767926894779156,
"roberta.encoder.layer.2.attention.self.value": 0.6355413155578186,
"roberta.encoder.layer.2.attention.output.dense": 0.652313549394161,
"roberta.encoder.layer.3.attention.self.query": 0.5620059249830314,
"roberta.encoder.layer.3.attention.self.key": 0.5846928691061228,
"roberta.encoder.layer.3.attention.self.value": 0.6435393603344641,
"roberta.encoder.layer.3.attention.output.dense": 0.6529325522535757,
"roberta.encoder.layer.4.attention.self.query": 0.5558525218786198,
"roberta.encoder.layer.4.attention.self.key": 0.5698482460116029,
"roberta.encoder.layer.4.attention.self.value": 0.6283739962133341,
"roberta.encoder.layer.4.attention.output.dense": 0.6543298298531539,
"roberta.encoder.layer.5.attention.self.query": 0.5746074762623272,
"roberta.encoder.layer.5.attention.self.key": 0.5936990828557264,
"roberta.encoder.layer.5.attention.self.value": 0.6406884607810202,
"roberta.encoder.layer.5.attention.output.dense": 0.6339202611297943,
"roberta.encoder.layer.6.attention.self.query": 0.5627067497330576,
"roberta.encoder.layer.6.attention.self.key": 0.5650050523067174,
"roberta.encoder.layer.6.attention.self.value": 0.6680174309524031,
"roberta.encoder.layer.6.attention.output.dense": 0.682322352516152,
"roberta.encoder.layer.7.attention.self.query": 0.5507184879548199,
"roberta.encoder.layer.7.attention.self.key": 0.5526140495227189,
"roberta.encoder.layer.7.attention.self.value": 0.6559854732744622,
"roberta.encoder.layer.7.attention.output.dense": 0.687257113976049,
"roberta.encoder.layer.8.attention.self.query": 0.5652229686657331,
"roberta.encoder.layer.8.attention.self.key": 0.5677945271740619,
"roberta.encoder.layer.8.attention.self.value": 0.7130425300843204,
"roberta.encoder.layer.8.attention.output.dense": 0.7462517876152233,
"roberta.encoder.layer.9.attention.self.query": 0.5487914433004103,
"roberta.encoder.layer.9.attention.self.key": 0.5625210566219578,
"roberta.encoder.layer.9.attention.self.value": 0.7241673515919304,
"roberta.encoder.layer.9.attention.output.dense": 0.7458869958162637,
"roberta.encoder.layer.10.attention.self.query": 0.5597622495914977,
"roberta.encoder.layer.10.attention.self.key": 0.5651472880233807,
"roberta.encoder.layer.10.attention.self.value": 0.7256538913909379,
"roberta.encoder.layer.10.attention.output.dense": 0.7505458282257228,
"roberta.encoder.layer.11.attention.self.query": 0.5504548639227866,
"roberta.encoder.layer.11.attention.self.key": 0.5770136969470895,
"roberta.encoder.layer.11.attention.self.value": 0.7405519954229905,
"roberta.encoder.layer.11.attention.output.dense": 0.7516753365236131
},
"per_layer_ranks": {
"roberta.encoder.layer.0.attention.self.query": 232,
"roberta.encoder.layer.0.attention.self.key": 221,
"roberta.encoder.layer.0.attention.self.value": 115,
"roberta.encoder.layer.0.attention.output.dense": 115,
"roberta.encoder.layer.1.attention.self.query": 167,
"roberta.encoder.layer.1.attention.self.key": 162,
"roberta.encoder.layer.1.attention.self.value": 115,
"roberta.encoder.layer.1.attention.output.dense": 115,
"roberta.encoder.layer.2.attention.self.query": 150,
"roberta.encoder.layer.2.attention.self.key": 148,
"roberta.encoder.layer.2.attention.self.value": 115,
"roberta.encoder.layer.2.attention.output.dense": 115,
"roberta.encoder.layer.3.attention.self.query": 160,
"roberta.encoder.layer.3.attention.self.key": 158,
"roberta.encoder.layer.3.attention.self.value": 115,
"roberta.encoder.layer.3.attention.output.dense": 115,
"roberta.encoder.layer.4.attention.self.query": 162,
"roberta.encoder.layer.4.attention.self.key": 158,
"roberta.encoder.layer.4.attention.self.value": 115,
"roberta.encoder.layer.4.attention.output.dense": 115,
"roberta.encoder.layer.5.attention.self.query": 154,
"roberta.encoder.layer.5.attention.self.key": 151,
"roberta.encoder.layer.5.attention.self.value": 115,
"roberta.encoder.layer.5.attention.output.dense": 115,
"roberta.encoder.layer.6.attention.self.query": 153,
"roberta.encoder.layer.6.attention.self.key": 146,
"roberta.encoder.layer.6.attention.self.value": 115,
"roberta.encoder.layer.6.attention.output.dense": 115,
"roberta.encoder.layer.7.attention.self.query": 166,
"roberta.encoder.layer.7.attention.self.key": 159,
"roberta.encoder.layer.7.attention.self.value": 115,
"roberta.encoder.layer.7.attention.output.dense": 115,
"roberta.encoder.layer.8.attention.self.query": 149,
"roberta.encoder.layer.8.attention.self.key": 145,
"roberta.encoder.layer.8.attention.self.value": 115,
"roberta.encoder.layer.8.attention.output.dense": 115,
"roberta.encoder.layer.9.attention.self.query": 143,
"roberta.encoder.layer.9.attention.self.key": 142,
"roberta.encoder.layer.9.attention.self.value": 115,
"roberta.encoder.layer.9.attention.output.dense": 115,
"roberta.encoder.layer.10.attention.self.query": 146,
"roberta.encoder.layer.10.attention.self.key": 141,
"roberta.encoder.layer.10.attention.self.value": 115,
"roberta.encoder.layer.10.attention.output.dense": 115,
"roberta.encoder.layer.11.attention.self.query": 149,
"roberta.encoder.layer.11.attention.self.key": 151,
"roberta.encoder.layer.11.attention.self.value": 115,
"roberta.encoder.layer.11.attention.output.dense": 115
}
},
"hardware": {
"peak_vram_mb": 610.89,
"tokens_per_sec": 17145.01
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 595.82,
"L_256_vram_mb": 598.39,
"L_512_vram_mb": 603.52,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 514)"
},
"timestamp": "2026-09-15T09:56:37.912126"
},
"results/unified_adasvd/distilbert/distilbert_sst2_adasvd_full_metrics.json": {
"model": "distilbert_sst2_adasvd",
"performance": {
"accuracy_pct": 86.01,
"f1_score": 0.871,
"f1_macro": 0.8591,
"precision": 0.8671,
"recall": 0.8588,
"val_loss": 0.5242,
"status": "PASS",
"elapsed_s": 14.86
},
"efficiency": {
"total_params_before": 66955010,
"total_params_after": 57613082,
"trainable_recovery_params": 4832280,
"compression_rate": 1.1621,
"compression_ratio_pct": 13.95,
"flops_reduction_pct": 65.99,
"num_compressed_layers": 24,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": 0.6146613516060712,
"weighted_reconstruction_error": 0.6134,
"per_layer_reconstruction_errors": {
"distilbert.transformer.layer.0.attention.q_lin": 0.602676297475002,
"distilbert.transformer.layer.0.attention.k_lin": 0.5978222152975311,
"distilbert.transformer.layer.0.attention.v_lin": 0.6459949938421847,
"distilbert.transformer.layer.0.attention.out_lin": 0.6120172829045932,
"distilbert.transformer.layer.1.attention.q_lin": 0.5110580367721084,
"distilbert.transformer.layer.1.attention.k_lin": 0.49670388030115403,
"distilbert.transformer.layer.1.attention.v_lin": 0.6294398938373562,
"distilbert.transformer.layer.1.attention.out_lin": 0.6409862326819177,
"distilbert.transformer.layer.2.attention.q_lin": 0.5981217943755176,
"distilbert.transformer.layer.2.attention.k_lin": 0.6032224922188253,
"distilbert.transformer.layer.2.attention.v_lin": 0.6279107814128413,
"distilbert.transformer.layer.2.attention.out_lin": 0.6706347235906429,
"distilbert.transformer.layer.3.attention.q_lin": 0.5571223711306362,
"distilbert.transformer.layer.3.attention.k_lin": 0.5528934615675862,
"distilbert.transformer.layer.3.attention.v_lin": 0.6630906611978048,
"distilbert.transformer.layer.3.attention.out_lin": 0.6921638735259316,
"distilbert.transformer.layer.4.attention.q_lin": 0.5929572040507274,
"distilbert.transformer.layer.4.attention.k_lin": 0.5893624852902365,
"distilbert.transformer.layer.4.attention.v_lin": 0.6598243196321351,
"distilbert.transformer.layer.4.attention.out_lin": 0.6845841911639682,
"distilbert.transformer.layer.5.attention.q_lin": 0.5914119397675296,
"distilbert.transformer.layer.5.attention.k_lin": 0.6059360235782776,
"distilbert.transformer.layer.5.attention.v_lin": 0.6427352707813347,
"distilbert.transformer.layer.5.attention.out_lin": 0.6832020121498666
},
"per_layer_ranks": {
"distilbert.transformer.layer.0.attention.q_lin": 120,
"distilbert.transformer.layer.0.attention.k_lin": 118,
"distilbert.transformer.layer.0.attention.v_lin": 115,
"distilbert.transformer.layer.0.attention.out_lin": 115,
"distilbert.transformer.layer.1.attention.q_lin": 147,
"distilbert.transformer.layer.1.attention.k_lin": 141,
"distilbert.transformer.layer.1.attention.v_lin": 115,
"distilbert.transformer.layer.1.attention.out_lin": 115,
"distilbert.transformer.layer.2.attention.q_lin": 134,
"distilbert.transformer.layer.2.attention.k_lin": 134,
"distilbert.transformer.layer.2.attention.v_lin": 120,
"distilbert.transformer.layer.2.attention.out_lin": 115,
"distilbert.transformer.layer.3.attention.q_lin": 133,
"distilbert.transformer.layer.3.attention.k_lin": 133,
"distilbert.transformer.layer.3.attention.v_lin": 138,
"distilbert.transformer.layer.3.attention.out_lin": 130,
"distilbert.transformer.layer.4.attention.q_lin": 138,
"distilbert.transformer.layer.4.attention.k_lin": 138,
"distilbert.transformer.layer.4.attention.v_lin": 141,
"distilbert.transformer.layer.4.attention.out_lin": 135,
"distilbert.transformer.layer.5.attention.q_lin": 139,
"distilbert.transformer.layer.5.attention.k_lin": 140,
"distilbert.transformer.layer.5.attention.v_lin": 143,
"distilbert.transformer.layer.5.attention.out_lin": 137
}
},
"hardware": {
"peak_vram_mb": 230.91,
"tokens_per_sec": 30810.24
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 211.08,
"L_256_vram_mb": 213.71,
"L_512_vram_mb": "unsupported (exceeds model max_position_embeddings 512)",
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 512)"
},
"timestamp": "2026-09-15T09:55:28.510529"
},
"results/unified_adasvd/distilbert/distilbert_sst2_full_full_metrics.json": {
"model": "distilbert_sst2_full",
"performance": {
"accuracy_pct": 90.71,
"f1_score": 0.9099,
"f1_macro": 0.907,
"precision": 0.9075,
"recall": 0.9068,
"val_loss": 0.3028,
"status": "PASS",
"elapsed_s": 3.11
},
"efficiency": {
"total_params_before": 66955010,
"total_params_after": 66955010,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 182.94,
"tokens_per_sec": 48356.1
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 164.38,
"L_256_vram_mb": 167.07,
"L_512_vram_mb": "unsupported (exceeds model max_position_embeddings 512)",
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 512)"
},
"timestamp": "2026-09-15T09:55:13.569544"
},
"results/unified_adasvd/qwen3_5/Qwen3.5-0.8B_full_full_metrics.json": {
"model": "Qwen3.5-0.8B_full",
"performance": {
"accuracy_pct": 67.55,
"f1_score": 0.5501,
"f1_macro": 0.6481,
"precision": 0.7703,
"recall": 0.6808,
"val_loss": 0.6364,
"status": "PASS",
"elapsed_s": 109.37
},
"efficiency": {
"total_params_before": 763217728,
"total_params_after": 763217728,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 10822656,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 1810.59,
"tokens_per_sec": 1295.14
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 1545.79,
"L_256_vram_mb": 1568.28,
"L_512_vram_mb": 1604.86,
"L_1024_vram_mb": 1685.72
},
"timestamp": "2026-09-06T03:59:39.178663"
},
"results/unified_adasvd/gpt2/gpt2_adasvd_full_metrics.json": {
"model": "gpt2_adasvd",
"performance": {
"accuracy_pct": 63.76,
"f1_score": 0.7047,
"f1_macro": 0.6179,
"precision": 0.6649,
"recall": 0.6337,
"val_loss": 0.6401,
"status": "PASS",
"elapsed_s": 19.25
},
"efficiency": {
"total_params_before": 124441344,
"total_params_after": 109692694,
"trainable_recovery_params": 35350,
"compression_rate": 1.1345,
"compression_ratio_pct": 11.85,
"flops_reduction_pct": 54.36,
"num_compressed_layers": 22,
"num_fallback_layers": 2,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": 0.5492219996067473,
"weighted_reconstruction_error": 0.5304,
"per_layer_reconstruction_errors": {
"transformer.h.0.attn.c_attn": 0.4359572020367505,
"transformer.h.0.attn.c_proj": 0.4591165032375484,
"transformer.h.1.attn.c_attn": 0.42981757107437263,
"transformer.h.1.attn.c_proj": 0.5283082840662332,
"transformer.h.2.attn.c_attn": 0.4713939852405788,
"transformer.h.2.attn.c_proj": 0.6411698049530289,
"transformer.h.3.attn.c_attn": 0.5088508736089505,
"transformer.h.3.attn.c_proj": 0.6493636918754561,
"transformer.h.4.attn.c_attn": 0.4893240697737358,
"transformer.h.4.attn.c_proj": 0.6413458874592544,
"transformer.h.5.attn.c_attn": 0.4637313021309327,
"transformer.h.5.attn.c_proj": 0.6656301554321922,
"transformer.h.6.attn.c_attn": 0.49409677027041077,
"transformer.h.6.attn.c_proj": 0.630057138023834,
"transformer.h.7.attn.c_attn": 0.4988713253041192,
"transformer.h.7.attn.c_proj": 0.6562152435318196,
"transformer.h.8.attn.c_attn": 0.5397303605224265,
"transformer.h.8.attn.c_proj": 0.6507756392785982,
"transformer.h.9.attn.c_attn": 0.562183706224178,
"transformer.h.9.attn.c_proj": 0.712749788403979,
"transformer.h.10.attn.c_attn": 0.5318939388245411,
"transformer.h.10.attn.c_proj": 0.7323871674447497,
"transformer.h.11.attn.c_attn": 0.5156999913750293,
"transformer.h.11.attn.c_proj": 0.6193505471044521
},
"per_layer_ranks": {
"transformer.h.0.attn.c_attn": 288,
"transformer.h.0.attn.c_proj": 115,
"transformer.h.1.attn.c_attn": 288,
"transformer.h.1.attn.c_proj": 115,
"transformer.h.2.attn.c_attn": 288,
"transformer.h.2.attn.c_proj": 115,
"transformer.h.3.attn.c_attn": 288,
"transformer.h.3.attn.c_proj": 115,
"transformer.h.4.attn.c_attn": 288,
"transformer.h.4.attn.c_proj": 115,
"transformer.h.5.attn.c_attn": 288,
"transformer.h.5.attn.c_proj": 115,
"transformer.h.6.attn.c_attn": 288,
"transformer.h.6.attn.c_proj": 115,
"transformer.h.7.attn.c_attn": 288,
"transformer.h.7.attn.c_proj": 115,
"transformer.h.8.attn.c_attn": 288,
"transformer.h.8.attn.c_proj": 115,
"transformer.h.9.attn.c_attn": 288,
"transformer.h.9.attn.c_proj": 115,
"transformer.h.10.attn.c_attn": 288,
"transformer.h.10.attn.c_proj": 115,
"transformer.h.11.attn.c_attn": 288,
"transformer.h.11.attn.c_proj": 115
}
},
"hardware": {
"peak_vram_mb": 676.71,
"tokens_per_sec": 18628.1
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 490.46,
"L_256_vram_mb": 514.02,
"L_512_vram_mb": 563.97,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 1024)"
},
"timestamp": "2026-09-15T09:57:40.951594"
},
"results/unified_adasvd/gpt2/gpt2_full_full_metrics.json": {
"model": "gpt2_full",
"performance": {
"accuracy_pct": 79.36,
"f1_score": 0.7978,
"f1_macro": 0.7935,
"precision": 0.7935,
"recall": 0.7935,
"val_loss": 0.4875,
"status": "PASS",
"elapsed_s": 7.03
},
"efficiency": {
"total_params_before": 124441344,
"total_params_after": 124441344,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 793.19,
"tokens_per_sec": 22955.86
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 681.07,
"L_256_vram_mb": 691.32,
"L_512_vram_mb": 721.95,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 1024)"
},
"timestamp": "2026-09-15T09:57:21.836156"
},
"results/unified_adasvd/roberta/summary.csv": "module_name,status,accuracy_pct,f1_score,f1_macro,precision,recall,val_loss,elapsed_s,peak_vram_mb,tokens_per_sec,total_params_before,total_params_after,trainable_recovery_params,compression_rate,compression_ratio_pct,flops_reduction_pct,num_compressed_layers,num_fallback_layers,avg_reconstruction_error,weighted_reconstruction_error,num_injected_attention_layers,lora_params,error,path,timestamp\nfull,PASS,78.56,0.8124,0.7811,0.8051,0.7832,0.4381,6.77,712.34,24228.27,124647170,124647170,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_roberta,2026-09-15T09:56:12.278981\nadasvd,PASS,73.39,0.7778,0.7232,0.7685,0.7306,0.5381,26.53,610.89,17145.01,124647170,109737258,8713768,1.1359,11.96,63.2,40,8,0.5913941559039471,0.6364,0,0,,adapted_roberta,2026-09-15T09:56:39.097068\n",
"results/unified_adasvd/distilbert/summary.csv": "module_name,status,accuracy_pct,f1_score,f1_macro,precision,recall,val_loss,elapsed_s,peak_vram_mb,tokens_per_sec,total_params_before,total_params_after,trainable_recovery_params,compression_rate,compression_ratio_pct,flops_reduction_pct,num_compressed_layers,num_fallback_layers,avg_reconstruction_error,weighted_reconstruction_error,num_injected_attention_layers,lora_params,error,path,timestamp\nfull,PASS,90.71,0.9099,0.907,0.9075,0.9068,0.3028,3.11,182.94,48356.1,66955010,66955010,0,1.0,0.0,0.0,0,0,,,0,0,,legacy_bert,2026-09-15T09:55:13.772918\nadasvd,PASS,86.01,0.871,0.8591,0.8671,0.8588,0.5242,14.86,230.91,30810.24,66955010,57613082,4832280,1.1621,13.95,65.99,24,0,0.6146613516060712,0.6134,0,0,,legacy_bert,2026-09-15T09:55:28.859565\n",
"results/unified_adasvd/qwen3_5/summary.csv": "module_name,status,accuracy_pct,f1_score,f1_macro,precision,recall,val_loss,elapsed_s,peak_vram_mb,tokens_per_sec,total_params_before,total_params_after,trainable_recovery_params,compression_rate,compression_ratio_pct,flops_reduction_pct,num_compressed_layers,num_fallback_layers,avg_reconstruction_error,weighted_reconstruction_error,num_injected_attention_layers,lora_params,error,path,timestamp\nfull,PASS,67.55,0.5501,0.6481,0.7703,0.6808,0.6364,109.37,1810.59,1295.14,763217728,763217728,0,1.0,0.0,0.0,0,0,,,0,10822656,,adapted_qwen3_5,2026-09-06T03:59:42.279595\n",
"results/unified_adasvd/gpt2/summary.csv": "module_name,status,accuracy_pct,f1_score,f1_macro,precision,recall,val_loss,elapsed_s,peak_vram_mb,tokens_per_sec,total_params_before,total_params_after,trainable_recovery_params,compression_rate,compression_ratio_pct,flops_reduction_pct,num_compressed_layers,num_fallback_layers,avg_reconstruction_error,weighted_reconstruction_error,num_injected_attention_layers,lora_params,error,path,timestamp\nfull,PASS,79.36,0.7978,0.7935,0.7935,0.7935,0.4875,7.03,793.19,22955.86,124441344,124441344,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_gpt,2026-09-15T09:57:22.599674\nadasvd,PASS,63.76,0.7047,0.6179,0.6649,0.6337,0.6401,19.26,676.71,18628.1,124441344,109692694,35350,1.1345,11.85,54.36,22,2,0.5492219996067473,0.5304,0,0,,adapted_gpt,2026-09-15T09:57:42.109615\n"
},
"notes": []
},
"admm_compressor": {
"module_name": "admm_compressor",
"source_file": "Codes/framework/compression_engine/admm_compressor.py",
"source_lines_count": 422,
"architecture": {
"class_name": "ADMMCompressor",
"base_class": "BaseCompressor",
"init_args": "self, model: nn.Module, sparsity: float = 0.5, quant_bits: int = 4, rho1: float = 0.1, rho2: float = 0.1, max_iter: int = 8, structured_24: bool = True, block_size: Tuple[int, int] = (4, 4), max_error: float = 0.45, fallback_to_original: bool = True, target_attention_only: bool = False, target_modules: Optional[List[str]] = None, comp_config: Any = None, **kwargs: Any, ) -> None: super().__init__(model, name=\"admm\", **kwargs) cfg = comp_config or kwargs.get(\"comp_config\", kwargs.get(\"cfg\", getattr(model, \"config\", None))) self.sparsity = max(0.0, min(float(_cfg_get(cfg, \"admm_sparsity\", _cfg_get(cfg, \"sparsity\", sparsity))), 0.95)) self.quant_bits = int(_cfg_get(cfg, \"admm_quant_bits\", _cfg_get(cfg, \"quant_bits\", quant_bits))) self.rho1 = float(_cfg_get(cfg, \"admm_rho1\", _cfg_get(cfg, \"rho1\", rho1))) self.rho2 = float(_cfg_get(cfg, \"admm_rho2\", _cfg_get(cfg, \"rho2\", rho2))) self.max_iter = int(_cfg_get(cfg, \"admm_max_iter\", _cfg_get(cfg, \"max_iter\", max_iter))) self.structured_24 = bool(_cfg_get(cfg, \"admm_structured_24\", _cfg_get(cfg, \"structured_24\", structured_24))) self.block_size = block_size self.max_error = float(_cfg_get(cfg, \"admm_max_error\", _cfg_get(cfg, \"max_error\", max_error))) self.fallback_to_original = bool( _cfg_get(cfg, \"admm_fallback_to_original\", _cfg_get(cfg, \"fallback_to_original\", fallback_to_original)) ) self.target_attention_only = bool( _cfg_get(cfg, \"admm_target_attention_only\", _cfg_get(cfg, \"target_attention_only\", target_attention_only)) ) self.target_modules = target_modules self.comp_config = cfg self.compressed_layers: List[str] = [] self.fallback_layers: List[str] = [] self.layer_errors: Dict[str, float] = {} self.layer_sparsities: Dict[str, float] = {} self.layer_actual_sparsities: Dict[str, float] = {} def _is_target_module(self, name: str, module: nn.Module) -> bool: # Do not target inner submodules of LoRA or wrappers directly if any(name.endswith(sfx) for sfx in [\".base_layer\", \".lora_A\", \".lora_B\", \".lora_dropout\", \".linear\"]",
"registry_names": [
"admm_compressor",
"admm"
]
},
"params": [],
"metrics": {
"results/unified_admm/roberta/roberta-base_full_full_metrics.json": {
"model": "roberta-base_full",
"performance": {
"accuracy_pct": 78.56,
"f1_score": 0.8124,
"f1_macro": 0.7811,
"precision": 0.8051,
"recall": 0.7832,
"val_loss": 0.4381,
"status": "PASS",
"elapsed_s": 6.89
},
"efficiency": {
"total_params_before": 124647170,
"total_params_after": 124647170,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 712.34,
"tokens_per_sec": 23915.43
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 683.27,
"L_256_vram_mb": 685.84,
"L_512_vram_mb": 699.47,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 514)"
},
"timestamp": "2026-09-15T09:59:16.408989"
},
"results/unified_admm/roberta/roberta-base_admm_full_metrics.json": {
"model": "roberta-base_admm",
"performance": {
"accuracy_pct": 62.04,
"f1_score": 0.6505,
"f1_macro": 0.6176,
"precision": 0.6219,
"recall": 0.619,
"val_loss": 0.7277,
"status": "PASS",
"elapsed_s": 10.24
},
"efficiency": {
"total_params_before": 124647170,
"total_params_after": 124647170,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 27,
"num_fallback_layers": 21,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": 0.4280215595989844,
"weighted_reconstruction_error": 0.4834,
"per_layer_reconstruction_errors": {
"roberta.encoder.layer.0.attention.self.query": 0.41136601993681615,
"roberta.encoder.layer.0.attention.self.key": 0.4366631624476461,
"roberta.encoder.layer.0.attention.self.value": 0.426901481358431,
"roberta.encoder.layer.0.attention.output.dense": 0.7556448949014658,
"roberta.encoder.layer.1.attention.self.query": 0.4114673323533447,
"roberta.encoder.layer.1.attention.self.key": 0.4558870149118198,
"roberta.encoder.layer.1.attention.self.value": 0.4414493424639032,
"roberta.encoder.layer.1.attention.output.dense": 0.7373564704111047,
"roberta.encoder.layer.2.attention.self.query": 0.42032348886437926,
"roberta.encoder.layer.2.attention.self.key": 0.4611841363470059,
"roberta.encoder.layer.2.attention.self.value": 0.5331701340852948,
"roberta.encoder.layer.2.attention.output.dense": 0.592421446196986,
"roberta.encoder.layer.3.attention.self.query": 0.43812791751336516,
"roberta.encoder.layer.3.attention.self.key": 0.46069132927417555,
"roberta.encoder.layer.3.attention.self.value": 0.4802459681433181,
"roberta.encoder.layer.3.attention.output.dense": 0.5951454275624185,
"roberta.encoder.layer.4.attention.self.query": 0.43941697950301795,
"roberta.encoder.layer.4.attention.self.key": 0.4617635328813933,
"roberta.encoder.layer.4.attention.self.value": 0.4078757364843845,
"roberta.encoder.layer.4.attention.output.dense": 0.5447857878099831,
"roberta.encoder.layer.5.attention.self.query": 0.42880828067705073,
"roberta.encoder.layer.5.attention.self.key": 0.4592205083905218,
"roberta.encoder.layer.5.attention.self.value": 0.4030522717004738,
"roberta.encoder.layer.5.attention.output.dense": 0.5500600897298221,
"roberta.encoder.layer.6.attention.self.query": 0.4419771622700167,
"roberta.encoder.layer.6.attention.self.key": 0.45552670988142197,
"roberta.encoder.layer.6.attention.self.value": 0.43530264191634305,
"roberta.encoder.layer.6.attention.output.dense": 0.5145825961512289,
"roberta.encoder.layer.7.attention.self.query": 0.43860469837769916,
"roberta.encoder.layer.7.attention.self.key": 0.42977009019823487,
"roberta.encoder.layer.7.attention.self.value": 0.4103143015802769,
"roberta.encoder.layer.7.attention.output.dense": 0.6464470769497084,
"roberta.encoder.layer.8.attention.self.query": 0.4291950287933315,
"roberta.encoder.layer.8.attention.self.key": 0.4666203098948942,
"roberta.encoder.layer.8.attention.self.value": 0.4189658702731616,
"roberta.encoder.layer.8.attention.output.dense": 0.6248452355812558,
"roberta.encoder.layer.9.attention.self.query": 0.4308300077373004,
"roberta.encoder.layer.9.attention.self.key": 0.4487770497192412,
"roberta.encoder.layer.9.attention.self.value": 0.43040441856502315,
"roberta.encoder.layer.9.attention.output.dense": 0.601603656069653,
"roberta.encoder.layer.10.attention.self.query": 0.4273415198478837,
"roberta.encoder.layer.10.attention.self.key": 0.4431352483882602,
"roberta.encoder.layer.10.attention.self.value": 0.4078836259168287,
"roberta.encoder.layer.10.attention.output.dense": 0.6262126590012314,
"roberta.encoder.layer.11.attention.self.query": 0.43487959420201283,
"roberta.encoder.layer.11.attention.self.key": 0.435898283185562,
"roberta.encoder.layer.11.attention.self.value": 0.42785055489859114,
"roberta.encoder.layer.11.attention.output.dense": 0.6250370219130287
},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 520.18,
"tokens_per_sec": 21347.69
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 504.62,
"L_256_vram_mb": 507.18,
"L_512_vram_mb": 512.31,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 514)"
},
"timestamp": "2026-09-15T09:59:26.731580"
},
"results/unified_admm/distilbert/distilbert_sst2_admm_full_metrics.json": {
"model": "distilbert_sst2_admm",
"performance": {
"accuracy_pct": 90.37,
"f1_score": 0.905,
"f1_macro": 0.9037,
"precision": 0.9036,
"recall": 0.9037,
"val_loss": 0.2858,
"status": "PASS",
"elapsed_s": 4.33
},
"efficiency": {
"total_params_before": 66955010,
"total_params_after": 66955010,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 11,
"num_fallback_layers": 13,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": 0.40874982318885544,
"weighted_reconstruction_error": 0.4895,
"per_layer_reconstruction_errors": {
"distilbert.transformer.layer.0.attention.q_lin": 0.6580605389939275,
"distilbert.transformer.layer.0.attention.k_lin": 0.787776985727936,
"distilbert.transformer.layer.0.attention.v_lin": 0.39606018953896366,
"distilbert.transformer.layer.0.attention.out_lin": 0.7875263865968171,
"distilbert.transformer.layer.1.attention.q_lin": 0.4757240897435198,
"distilbert.transformer.layer.1.attention.k_lin": 0.4875111083514526,
"distilbert.transformer.layer.1.attention.v_lin": 0.4204699911022476,
"distilbert.transformer.layer.1.attention.out_lin": 0.5591444119321567,
"distilbert.transformer.layer.2.attention.q_lin": 0.435239293915547,
"distilbert.transformer.layer.2.attention.k_lin": 0.4263437486138671,
"distilbert.transformer.layer.2.attention.v_lin": 0.39225455220413713,
"distilbert.transformer.layer.2.attention.out_lin": 0.4536393497339716,
"distilbert.transformer.layer.3.attention.q_lin": 0.40295069067664546,
"distilbert.transformer.layer.3.attention.k_lin": 0.4674407322856049,
"distilbert.transformer.layer.3.attention.v_lin": 0.4176402927988247,
"distilbert.transformer.layer.3.attention.out_lin": 0.41443324601261783,
"distilbert.transformer.layer.4.attention.q_lin": 0.39257951915333333,
"distilbert.transformer.layer.4.attention.k_lin": 0.5775680493286506,
"distilbert.transformer.layer.4.attention.v_lin": 0.39301647897423725,
"distilbert.transformer.layer.4.attention.out_lin": 0.47342701317728203,
"distilbert.transformer.layer.5.attention.q_lin": 0.560028591426301,
"distilbert.transformer.layer.5.attention.k_lin": 0.4561895790024111,
"distilbert.transformer.layer.5.attention.v_lin": 0.4052600520869886,
"distilbert.transformer.layer.5.attention.out_lin": 0.5073619945572996
},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 183.07,
"tokens_per_sec": 50559.49
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 165.01,
"L_256_vram_mb": 167.2,
"L_512_vram_mb": "unsupported (exceeds model max_position_embeddings 512)",
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 512)"
},
"timestamp": "2026-09-15T09:58:34.207565"
},
"results/unified_admm/distilbert/distilbert_sst2_full_full_metrics.json": {
"model": "distilbert_sst2_full",
"performance": {
"accuracy_pct": 90.71,
"f1_score": 0.9099,
"f1_macro": 0.907,
"precision": 0.9075,
"recall": 0.9068,
"val_loss": 0.3028,
"status": "PASS",
"elapsed_s": 2.98
},
"efficiency": {
"total_params_before": 66955010,
"total_params_after": 66955010,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 182.94,
"tokens_per_sec": 51199.68
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 164.38,
"L_256_vram_mb": 167.07,
"L_512_vram_mb": "unsupported (exceeds model max_position_embeddings 512)",
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 512)"
},
"timestamp": "2026-09-15T09:58:29.639880"
},
"results/unified_admm/qwen3_5/Qwen3.5-0.8B_admm_full_metrics.json": {
"model": "Qwen3.5-0.8B_admm",
"performance": {
"accuracy_pct": 88.42,
"f1_score": 0.8947,
"f1_macro": 0.883,
"precision": 0.8955,
"recall": 0.8826,
"val_loss": 1.5309,
"status": "PASS",
"elapsed_s": 1355.44
},
"efficiency": {
"total_params_before": 763217728,
"total_params_after": 763217728,
"trainable_recovery_params": 5776384,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 28,
"num_fallback_layers": 86,
"num_injected_attention_layers": 0,
"lora_params": 10822656,
"avg_reconstruction_error": 0.41579800644162873,
"weighted_reconstruction_error": 0.6669,
"per_layer_reconstruction_errors": {
"model.layers.0.linear_attn.out_proj": 0.8048436257528224,
"model.layers.0.linear_attn.in_proj_qkv": 0.7985198426084548,
"model.layers.0.linear_attn.in_proj_z": 0.857345751055695,
"model.layers.0.linear_attn.in_proj_b": 0.4209942278849084,
"model.layers.0.linear_attn.in_proj_a": 0.3871298336128409,
"model.layers.1.linear_attn.out_proj": 0.8610993440091969,
"model.layers.1.linear_attn.in_proj_qkv": 0.7219953403680267,
"model.layers.1.linear_attn.in_proj_z": 0.4681587368205583,
"model.layers.1.linear_attn.in_proj_b": 0.5094557183798485,
"model.layers.1.linear_attn.in_proj_a": 0.4101455025956409,
"model.layers.2.linear_attn.out_proj": 0.8202639210923023,
"model.layers.2.linear_attn.in_proj_qkv": 0.6057882870773925,
"model.layers.2.linear_attn.in_proj_z": 0.47873575024459686,
"model.layers.2.linear_attn.in_proj_b": 0.46435934197716155,
"model.layers.2.linear_attn.in_proj_a": 0.41323145856043153,
"model.layers.3.self_attn.q_proj": 0.5454813034865768,
"model.layers.3.self_attn.k_proj": 0.4533796565291563,
"model.layers.3.self_attn.v_proj": 0.4843685376730368,
"model.layers.3.self_attn.o_proj": 0.8620346709691896,
"model.layers.4.linear_attn.out_proj": 0.7447042703936904,
"model.layers.4.linear_attn.in_proj_qkv": 0.7069293643426828,
"model.layers.4.linear_attn.in_proj_z": 0.4732143473661153,
"model.layers.4.linear_attn.in_proj_b": 0.43811333729211377,
"model.layers.4.linear_attn.in_proj_a": 0.404387015004295,
"model.layers.5.linear_attn.out_proj": 0.7965420485900966,
"model.layers.5.linear_attn.in_proj_qkv": 0.722061862530207,
"model.layers.5.linear_attn.in_proj_z": 0.5520545179894648,
"model.layers.5.linear_attn.in_proj_b": 0.4648876811357769,
"model.layers.5.linear_attn.in_proj_a": 0.40443054011756446,
"model.layers.6.linear_attn.out_proj": 0.6922351105745191,
"model.layers.6.linear_attn.in_proj_qkv": 0.5629947462911002,
"model.layers.6.linear_attn.in_proj_z": 0.5608667168486609,
"model.layers.6.linear_attn.in_proj_b": 0.4323727223067152,
"model.layers.6.linear_attn.in_proj_a": 0.40757499270604675,
"model.layers.7.self_attn.q_proj": 0.6527976987075234,
"model.layers.7.self_attn.k_proj": 0.46953487670645905,
"model.layers.7.self_attn.v_proj": 0.48334732944702663,
"model.layers.7.self_attn.o_proj": 0.5623418666628308,
"model.layers.8.linear_attn.out_proj": 0.723831495770828,
"model.layers.8.linear_attn.in_proj_qkv": 0.5825963895535734,
"model.layers.8.linear_attn.in_proj_z": 0.8054631447065264,
"model.layers.8.linear_attn.in_proj_b": 0.43988914213466096,
"model.layers.8.linear_attn.in_proj_a": 0.40405372116944627,
"model.layers.9.linear_attn.out_proj": 0.7478618773092672,
"model.layers.9.linear_attn.in_proj_qkv": 0.6617897310082955,
"model.layers.9.linear_attn.in_proj_z": 0.8467090644367338,
"model.layers.9.linear_attn.in_proj_b": 0.419009304446191,
"model.layers.9.linear_attn.in_proj_a": 0.45423378228497513,
"model.layers.10.linear_attn.out_proj": 0.7880243886798519,
"model.layers.10.linear_attn.in_proj_qkv": 0.6191468899334026,
"model.layers.10.linear_attn.in_proj_z": 0.5291104798185097,
"model.layers.10.linear_attn.in_proj_b": 0.4513700158362432,
"model.layers.10.linear_attn.in_proj_a": 0.4108915681550579,
"model.layers.11.self_attn.q_proj": 0.5786481965726237,
"model.layers.11.self_attn.k_proj": 0.46326765971375417,
"model.layers.11.self_attn.v_proj": 0.4616343481739018,
"model.layers.11.self_attn.o_proj": 0.4911279536536986,
"model.layers.12.linear_attn.out_proj": 0.6664749415238315,
"model.layers.12.linear_attn.in_proj_qkv": 0.5992181694318529,
"model.layers.12.linear_attn.in_proj_z": 0.5500502684602033,
"model.layers.12.linear_attn.in_proj_b": 0.4180889057808951,
"model.layers.12.linear_attn.in_proj_a": 0.404633626213282,
"model.layers.13.linear_attn.out_proj": 0.8616983199087361,
"model.layers.13.linear_attn.in_proj_qkv": 0.5718014113699876,
"model.layers.13.linear_attn.in_proj_z": 0.5679831177277522,
"model.layers.13.linear_attn.in_proj_b": 0.4195421802789988,
"model.layers.13.linear_attn.in_proj_a": 0.40609663104978433,
"model.layers.14.linear_attn.out_proj": 0.5793293607405131,
"model.layers.14.linear_attn.in_proj_qkv": 0.8010844473930941,
"model.layers.14.linear_attn.in_proj_z": 0.5475735544770444,
"model.layers.14.linear_attn.in_proj_b": 0.46340548435606055,
"model.layers.14.linear_attn.in_proj_a": 0.4518819753564688,
"model.layers.15.self_attn.q_proj": 0.6573951568839139,
"model.layers.15.self_attn.k_proj": 0.48574098481989275,
"model.layers.15.self_attn.v_proj": 0.6656179324984539,
"model.layers.15.self_attn.o_proj": 0.5266665911430181,
"model.layers.16.linear_attn.out_proj": 0.8067410796122303,
"model.layers.16.linear_attn.in_proj_qkv": 0.6314374565024004,
"model.layers.16.linear_attn.in_proj_z": 0.5577511317194872,
"model.layers.16.linear_attn.in_proj_b": 0.44610666493232704,
"model.layers.16.linear_attn.in_proj_a": 0.3887796652290735,
"model.layers.17.linear_attn.out_proj": 0.6246929321299726,
"model.layers.17.linear_attn.in_proj_qkv": 0.6730174690370689,
"model.layers.17.linear_attn.in_proj_z": 0.46930635838150286,
"model.layers.17.linear_attn.in_proj_b": 0.4252897537436498,
"model.layers.17.linear_attn.in_proj_a": 0.4361585435622759,
"model.layers.18.linear_attn.out_proj": 0.8076517877598156,
"model.layers.18.linear_attn.in_proj_qkv": 0.6253226527222886,
"model.layers.18.linear_attn.in_proj_z": 0.5472128782486748,
"model.layers.18.linear_attn.in_proj_b": 0.430635499741553,
"model.layers.18.linear_attn.in_proj_a": 0.4401114704043207,
"model.layers.19.self_attn.q_proj": 0.5604505226550452,
"model.layers.19.self_attn.k_proj": 0.6238351985188378,
"model.layers.19.self_attn.v_proj": 0.6630952077557101,
"model.layers.19.self_attn.o_proj": 0.558049768896921,
"model.layers.20.linear_attn.out_proj": 0.8337841255459909,
"model.layers.20.linear_attn.in_proj_qkv": 0.8109392558532739,
"model.layers.20.linear_attn.in_proj_z": 0.6515479652847436,
"model.layers.20.linear_attn.in_proj_b": 0.41112796234231197,
"model.layers.20.linear_attn.in_proj_a": 0.4049423291258385,
"model.layers.21.linear_attn.out_proj": 0.7833323974573576,
"model.layers.21.linear_attn.in_proj_qkv": 0.7750768173127515,
"model.layers.21.linear_attn.in_proj_z": 0.6352461151988583,
"model.layers.21.linear_attn.in_proj_b": 0.4102781435370985,
"model.layers.21.linear_attn.in_proj_a": 0.40663565155927517,
"model.layers.22.linear_attn.out_proj": 0.8642979399010761,
"model.layers.22.linear_attn.in_proj_qkv": 0.7369514075584257,
"model.layers.22.linear_attn.in_proj_z": 0.7569131184627718,
"model.layers.22.linear_attn.in_proj_b": 0.4857865748230741,
"model.layers.22.linear_attn.in_proj_a": 0.40169378687900803,
"model.layers.23.self_attn.q_proj": 0.7780953359790628,
"model.layers.23.self_attn.k_proj": 0.5817086547733064,
"model.layers.23.self_attn.v_proj": 0.8287568197637294,
"model.layers.23.self_attn.o_proj": 0.6756134693033635
},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 1851.31,
"tokens_per_sec": 1099.69
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 1587.0,
"L_256_vram_mb": 1608.4,
"L_512_vram_mb": 1636.97,
"L_1024_vram_mb": 1726.48
},
"timestamp": "2026-09-06T04:51:55.093003"
},
"results/unified_admm/qwen3_5/Qwen3.5-0.8B_full_full_metrics.json": {
"model": "Qwen3.5-0.8B_full",
"performance": {
"accuracy_pct": 67.55,
"f1_score": 0.5501,
"f1_macro": 0.6481,
"precision": 0.7703,
"recall": 0.6808,
"val_loss": 0.6364,
"status": "PASS",
"elapsed_s": 110.98
},
"efficiency": {
"total_params_before": 763217728,
"total_params_after": 763217728,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 10822656,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 1810.59,
"tokens_per_sec": 1287.4
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 1545.79,
"L_256_vram_mb": 1568.28,
"L_512_vram_mb": 1604.86,
"L_1024_vram_mb": 1685.72
},
"timestamp": "2026-09-06T04:29:19.367674"
},
"results/unified_admm/gpt2/gpt2_full_full_metrics.json": {
"model": "gpt2_full",
"performance": {
"accuracy_pct": 79.36,
"f1_score": 0.7978,
"f1_macro": 0.7935,
"precision": 0.7935,
"recall": 0.7935,
"val_loss": 0.4875,
"status": "PASS",
"elapsed_s": 6.84
},
"efficiency": {
"total_params_before": 124441344,
"total_params_after": 124441344,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 793.19,
"tokens_per_sec": 23706.62
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 681.07,
"L_256_vram_mb": 691.32,
"L_512_vram_mb": 721.95,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 1024)"
},
"timestamp": "2026-09-15T10:00:10.164382"
},
"results/unified_admm/gpt2/gpt2_admm_full_metrics.json": {
"model": "gpt2_admm",
"performance": {
"accuracy_pct": 79.36,
"f1_score": 0.7978,
"f1_macro": 0.7935,
"precision": 0.7935,
"recall": 0.7935,
"val_loss": 0.4875,
"status": "PASS_NO_COMPRESSION",
"elapsed_s": 9.84
},
"efficiency": {
"total_params_before": 124441344,
"total_params_after": 124441344,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 24,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": 0.0,
"weighted_reconstruction_error": 0.6375,
"per_layer_reconstruction_errors": {
"transformer.h.0.attn.c_attn": 0.5331369025693834,
"transformer.h.0.attn.c_proj": 0.5716791786120512,
"transformer.h.1.attn.c_attn": 0.4512060384684465,
"transformer.h.1.attn.c_proj": 0.850115045916021,
"transformer.h.2.attn.c_attn": 0.5008755601578878,
"transformer.h.2.attn.c_proj": 0.8806542588308877,
"transformer.h.3.attn.c_attn": 0.5651615766239706,
"transformer.h.3.attn.c_proj": 0.8444291022652354,
"transformer.h.4.attn.c_attn": 0.7951295185722521,
"transformer.h.4.attn.c_proj": 0.7432609415444862,
"transformer.h.5.attn.c_attn": 0.49991357341792536,
"transformer.h.5.attn.c_proj": 0.7900097556283429,
"transformer.h.6.attn.c_attn": 0.5546549347448773,
"transformer.h.6.attn.c_proj": 0.6616940278090765,
"transformer.h.7.attn.c_attn": 0.5911012517648465,
"transformer.h.7.attn.c_proj": 0.749312348427571,
"transformer.h.8.attn.c_attn": 0.6345449192178677,
"transformer.h.8.attn.c_proj": 0.8473916455727238,
"transformer.h.9.attn.c_attn": 0.6470749017105901,
"transformer.h.9.attn.c_proj": 0.6084331860105642,
"transformer.h.10.attn.c_attn": 0.6219575473129052,
"transformer.h.10.attn.c_proj": 0.894697971021798,
"transformer.h.11.attn.c_attn": 0.7047311357336467,
"transformer.h.11.attn.c_proj": 0.8609352797924921
},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 628.41,
"tokens_per_sec": 19301.53
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 510.91,
"L_256_vram_mb": 524.41,
"L_512_vram_mb": 552.17,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 1024)"
},
"timestamp": "2026-09-15T10:00:20.254568"
},
"results/unified_admm/roberta/summary.csv": "module_name,status,accuracy_pct,f1_score,f1_macro,precision,recall,val_loss,elapsed_s,peak_vram_mb,tokens_per_sec,total_params_before,total_params_after,trainable_recovery_params,compression_rate,compression_ratio_pct,flops_reduction_pct,num_compressed_layers,num_fallback_layers,avg_reconstruction_error,weighted_reconstruction_error,num_injected_attention_layers,lora_params,error,path,timestamp\nfull,PASS,78.56,0.8124,0.7811,0.8051,0.7832,0.4381,6.89,712.34,23915.43,124647170,124647170,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_roberta,2026-09-15T09:59:17.197780\nadmm,PASS,62.04,0.6505,0.6176,0.6219,0.619,0.7277,10.25,520.18,21347.69,124647170,124647170,0,1.0,0.0,0.0,27,21,0.4280215595989844,0.4834,0,0,,adapted_roberta,2026-09-15T09:59:27.723332\n",
"results/unified_admm/distilbert/summary.csv": "module_name,status,accuracy_pct,f1_score,f1_macro,precision,recall,val_loss,elapsed_s,peak_vram_mb,tokens_per_sec,total_params_before,total_params_after,trainable_recovery_params,compression_rate,compression_ratio_pct,flops_reduction_pct,num_compressed_layers,num_fallback_layers,avg_reconstruction_error,weighted_reconstruction_error,num_injected_attention_layers,lora_params,error,path,timestamp\nfull,PASS,90.71,0.9099,0.907,0.9075,0.9068,0.3028,2.98,182.94,51199.68,66955010,66955010,0,1.0,0.0,0.0,0,0,,,0,0,,legacy_bert,2026-09-15T09:58:29.844088\nadmm,PASS,90.37,0.905,0.9037,0.9036,0.9037,0.2858,4.33,183.07,50559.49,66955010,66955010,0,1.0,0.0,0.0,11,13,0.40874982318885544,0.4895,0,0,,legacy_bert,2026-09-15T09:58:34.430652\n",
"results/unified_admm/qwen3_5/summary.csv": "module_name,status,accuracy_pct,f1_score,f1_macro,precision,recall,val_loss,elapsed_s,peak_vram_mb,tokens_per_sec,total_params_before,total_params_after,trainable_recovery_params,compression_rate,compression_ratio_pct,flops_reduction_pct,num_compressed_layers,num_fallback_layers,avg_reconstruction_error,weighted_reconstruction_error,num_injected_attention_layers,lora_params,error,path,timestamp\nfull,PASS,67.55,0.5501,0.6481,0.7703,0.6808,0.6364,110.98,1810.59,1287.4,763217728,763217728,0,1.0,0.0,0.0,0,0,,,0,10822656,,adapted_qwen3_5,2026-09-06T04:29:22.618333\nadmm,PASS,88.42,0.8947,0.883,0.8955,0.8826,1.5309,1355.44,1851.31,1099.69,763217728,763217728,5776384,1.0,0.0,0.0,28,86,0.41579800644162873,0.6669,0,10822656,,adapted_qwen3_5,2026-09-06T04:51:58.367759\n",
"results/unified_admm/gpt2/summary.csv": "module_name,status,accuracy_pct,f1_score,f1_macro,precision,recall,val_loss,elapsed_s,peak_vram_mb,tokens_per_sec,total_params_before,total_params_after,trainable_recovery_params,compression_rate,compression_ratio_pct,flops_reduction_pct,num_compressed_layers,num_fallback_layers,avg_reconstruction_error,weighted_reconstruction_error,num_injected_attention_layers,lora_params,error,path,timestamp\nfull,PASS,79.36,0.7978,0.7935,0.7935,0.7935,0.4875,6.84,793.19,23706.62,124441344,124441344,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_gpt,2026-09-15T10:00:10.940470\nadmm,PASS_NO_COMPRESSION,79.36,0.7978,0.7935,0.7935,0.7935,0.4875,9.84,628.41,19301.53,124441344,124441344,0,1.0,0.0,0.0,0,24,0.0,0.6375,0,0,\"note: zero compressed layers, model grew (LoRA only)\",adapted_gpt,2026-09-15T10:00:21.029783\n",
"results/all_121_compressors_gpu_matrix.csv": "admm_compressor,cuda,PASS,True,True,1.0,194.2,0.469,"
},
"notes": []
},
"admm_pruner": {
"module_name": "admm_pruner",
"source_file": "Codes/framework/compression_engine/admm_pruner.py",
"source_lines_count": 767,
"architecture": {
"class_name": "ADMMMaskedLinear",
"base_class": "nn.Module",
"init_args": "self, in_features: int, out_features: int, bias: bool = True, dtype: torch.dtype = torch.float32, device: Optional[torch.device] = None, ) -> None: super().__init__() self.in_features = in_features self.out_features = out_features self.weight = nn.Parameter(torch.empty(out_features, in_features, dtype=dtype, device=device)) self.register_buffer(\"mask\", torch.ones(out_features, in_features, dtype=dtype, device=device)) if bias: self.bias = nn.Parameter(torch.empty(out_features, dtype=dtype, device=device)) else: self.register_parameter(\"bias\", None) @classmethod def from_linear(cls, linear: nn.Linear, mask: torch.Tensor) -> ADMMMaskedLinear: layer = cls( in_features=linear.in_features, out_features=linear.out_features, bias=linear.bias is not None, dtype=linear.weight.dtype, device=linear.weight.device, ) layer.weight.data.copy_(linear.weight.data) layer.mask.copy_(mask.to(dtype=linear.weight.dtype, device=linear.weight.device)) if linear.bias is not None: layer.bias.data.copy_(linear.bias.data) return layer def forward(self, x: torch.Tensor) -> torch.Tensor: w_masked = ADMMSTEPrune.apply(self.weight, self.mask) return F.linear(x, w_masked, self.bias) @COMPRESSOR_REGISTRY.register(\"admm_pruner\") @COMPRESSOR_REGISTRY.register(\"admm_pruner_distill\") class ADMMPruner(BaseCompressor",
"registry_names": [
"admm_pruner",
"admm_pruner_distill",
"admm_pruner_factory"
]
},
"params": [
{
"name": "weight",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.empty(out_features, in_features, dtype=dtype, device=device)"
},
{
"name": "bias",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.empty(out_features, dtype=dtype, device=device)"
}
],
"metrics": {
"results/unified_admm_pruner/roberta/roberta-base_full_full_metrics.json": {
"model": "roberta-base_full",
"performance": {
"accuracy_pct": 88.76,
"f1_score": 0.8942,
"f1_macro": 0.8872,
"precision": 0.891,
"recall": 0.8868,
"val_loss": 0.295,
"status": "PASS",
"elapsed_s": 5.97
},
"efficiency": {
"total_params_before": 127325954,
"total_params_after": 127325954,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 2678784,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 300.73,
"tokens_per_sec": 30825.8
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 279.78,
"L_256_vram_mb": 282.35,
"L_512_vram_mb": 287.98,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 514)"
},
"timestamp": "2026-09-06T05:08:12.899727"
},
"results/unified_admm_pruner/roberta/roberta-base_admm_pruner_full_metrics.json": {
"model": "roberta-base_admm_pruner",
"performance": {
"accuracy_pct": 80.62,
"f1_score": 0.7966,
"f1_macro": 0.8058,
"precision": 0.8112,
"recall": 0.8073,
"val_loss": 0.4539,
"status": "PASS",
"elapsed_s": 100.32
},
"efficiency": {
"total_params_before": 127325954,
"total_params_after": 127325954,
"trainable_recovery_params": 3270914,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 72,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 2678784,
"avg_reconstruction_error": 0.3513150578015672,
"weighted_reconstruction_error": 0.3495,
"per_layer_reconstruction_errors": {
"roberta.encoder.layer.0.attention.self.query": 0.353880583941378,
"roberta.encoder.layer.0.attention.self.key": 0.3504066634678763,
"roberta.encoder.layer.0.attention.self.value": 0.35180211511657894,
"roberta.encoder.layer.0.attention.output.dense": 0.36439331095959726,
"roberta.encoder.layer.0.intermediate.dense": 0.3481582530469862,
"roberta.encoder.layer.0.output.dense": 0.32019252916635715,
"roberta.encoder.layer.1.attention.self.query": 0.35219396957525934,
"roberta.encoder.layer.1.attention.self.key": 0.3431690779189612,
"roberta.encoder.layer.1.attention.self.value": 0.3493105547002784,
"roberta.encoder.layer.1.attention.output.dense": 0.3580085642022411,
"roberta.encoder.layer.1.intermediate.dense": 0.3463648849742073,
"roberta.encoder.layer.1.output.dense": 0.332440017708914,
"roberta.encoder.layer.2.attention.self.query": 0.35211631423891365,
"roberta.encoder.layer.2.attention.self.key": 0.3429978995662033,
"roberta.encoder.layer.2.attention.self.value": 0.34600523693614466,
"roberta.encoder.layer.2.attention.output.dense": 0.3616817699942833,
"roberta.encoder.layer.2.intermediate.dense": 0.3487217884917807,
"roberta.encoder.layer.2.output.dense": 0.33799742490778906,
"roberta.encoder.layer.3.attention.self.query": 0.3541524452514513,
"roberta.encoder.layer.3.attention.self.key": 0.3457047294204267,
"roberta.encoder.layer.3.attention.self.value": 0.3450806794023792,
"roberta.encoder.layer.3.attention.output.dense": 0.36158487120232585,
"roberta.encoder.layer.3.intermediate.dense": 0.3515371024995805,
"roberta.encoder.layer.3.output.dense": 0.3367983984414318,
"roberta.encoder.layer.4.attention.self.query": 0.35672045298671967,
"roberta.encoder.layer.4.attention.self.key": 0.35046855926491893,
"roberta.encoder.layer.4.attention.self.value": 0.33571094854626354,
"roberta.encoder.layer.4.attention.output.dense": 0.3622037273391773,
"roberta.encoder.layer.4.intermediate.dense": 0.3516536214035121,
"roberta.encoder.layer.4.output.dense": 0.33858950795906195,
"roberta.encoder.layer.5.attention.self.query": 0.3555296989218202,
"roberta.encoder.layer.5.attention.self.key": 0.3478359473927771,
"roberta.encoder.layer.5.attention.self.value": 0.33941072580115766,
"roberta.encoder.layer.5.attention.output.dense": 0.363393926979014,
"roberta.encoder.layer.5.intermediate.dense": 0.3512919780210426,
"roberta.encoder.layer.5.output.dense": 0.3380266191917342,
"roberta.encoder.layer.6.attention.self.query": 0.350677323697209,
"roberta.encoder.layer.6.attention.self.key": 0.34898392415283336,
"roberta.encoder.layer.6.attention.self.value": 0.34137674941530743,
"roberta.encoder.layer.6.attention.output.dense": 0.3658203253731538,
"roberta.encoder.layer.6.intermediate.dense": 0.35354329968127046,
"roberta.encoder.layer.6.output.dense": 0.3340334958557971,
"roberta.encoder.layer.7.attention.self.query": 0.35066698048405615,
"roberta.encoder.layer.7.attention.self.key": 0.3456556791109259,
"roberta.encoder.layer.7.attention.self.value": 0.33939504930471687,
"roberta.encoder.layer.7.attention.output.dense": 0.36130873803069263,
"roberta.encoder.layer.7.intermediate.dense": 0.35480666346268575,
"roberta.encoder.layer.7.output.dense": 0.33814103473495066,
"roberta.encoder.layer.8.attention.self.query": 0.3588952146937881,
"roberta.encoder.layer.8.attention.self.key": 0.3533419842624639,
"roberta.encoder.layer.8.attention.self.value": 0.34280410125881094,
"roberta.encoder.layer.8.attention.output.dense": 0.3616720762672566,
"roberta.encoder.layer.8.intermediate.dense": 0.3602367037306547,
"roberta.encoder.layer.8.output.dense": 0.3399217748785091,
"roberta.encoder.layer.9.attention.self.query": 0.3607093136682335,
"roberta.encoder.layer.9.attention.self.key": 0.3548131935515911,
"roberta.encoder.layer.9.attention.self.value": 0.34346614733748465,
"roberta.encoder.layer.9.attention.output.dense": 0.3618949881703467,
"roberta.encoder.layer.9.intermediate.dense": 0.364401504525545,
"roberta.encoder.layer.9.output.dense": 0.344249579857131,
"roberta.encoder.layer.10.attention.self.query": 0.36345673032456205,
"roberta.encoder.layer.10.attention.self.key": 0.3589836182961219,
"roberta.encoder.layer.10.attention.self.value": 0.3483119630024868,
"roberta.encoder.layer.10.attention.output.dense": 0.3621897733024853,
"roberta.encoder.layer.10.intermediate.dense": 0.36223496631668783,
"roberta.encoder.layer.10.output.dense": 0.3508322530727731,
"roberta.encoder.layer.11.attention.self.query": 0.3622988234797672,
"roberta.encoder.layer.11.attention.self.key": 0.3593407002022056,
"roberta.encoder.layer.11.attention.self.value": 0.3484738427737698,
"roberta.encoder.layer.11.attention.output.dense": 0.3620286644161826,
"roberta.encoder.layer.11.intermediate.dense": 0.3621411990785709,
"roberta.encoder.layer.11.output.dense": 0.3780408830032648
},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 336.39,
"tokens_per_sec": 13333.06
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 306.05,
"L_256_vram_mb": 310.37,
"L_512_vram_mb": 318.58,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 514)"
},
"timestamp": "2026-09-06T05:09:53.429189"
},
"results/unified_admm_pruner/distilbert/distilbert_sst2_admm_pruner_full_metrics.json": {
"model": "distilbert_sst2_admm_pruner",
"performance": {
"accuracy_pct": 84.29,
"f1_score": 0.8603,
"f1_macro": 0.8404,
"precision": 0.8601,
"recall": 0.8409,
"val_loss": 0.6574,
"status": "PASS",
"elapsed_s": 25.31
},
"efficiency": {
"total_params_before": 66955010,
"total_params_after": 66955010,
"trainable_recovery_params": 4135682,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 36,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": 0.35710002246145245,
"weighted_reconstruction_error": 0.3552,
"per_layer_reconstruction_errors": {
"distilbert.transformer.layer.0.attention.q_lin": 0.3535878092207136,
"distilbert.transformer.layer.0.attention.k_lin": 0.35138108041341354,
"distilbert.transformer.layer.0.attention.v_lin": 0.35806251333424427,
"distilbert.transformer.layer.0.attention.out_lin": 0.3618706183179003,
"distilbert.transformer.layer.0.ffn.lin1": 0.36082577764047685,
"distilbert.transformer.layer.0.ffn.lin2": 0.34302148661499643,
"distilbert.transformer.layer.1.attention.q_lin": 0.3536908361277555,
"distilbert.transformer.layer.1.attention.k_lin": 0.3512372004463076,
"distilbert.transformer.layer.1.attention.v_lin": 0.3589342882162706,
"distilbert.transformer.layer.1.attention.out_lin": 0.3630321565442857,
"distilbert.transformer.layer.1.ffn.lin1": 0.35971828270683803,
"distilbert.transformer.layer.1.ffn.lin2": 0.34403754104572953,
"distilbert.transformer.layer.2.attention.q_lin": 0.35867243957249945,
"distilbert.transformer.layer.2.attention.k_lin": 0.3584123158914556,
"distilbert.transformer.layer.2.attention.v_lin": 0.3564895148752842,
"distilbert.transformer.layer.2.attention.out_lin": 0.36394906724839116,
"distilbert.transformer.layer.2.ffn.lin1": 0.35948782951363795,
"distilbert.transformer.layer.2.ffn.lin2": 0.3349396509322286,
"distilbert.transformer.layer.3.attention.q_lin": 0.3596522678082891,
"distilbert.transformer.layer.3.attention.k_lin": 0.3600704827598413,
"distilbert.transformer.layer.3.attention.v_lin": 0.3586865867263614,
"distilbert.transformer.layer.3.attention.out_lin": 0.3644508596482697,
"distilbert.transformer.layer.3.ffn.lin1": 0.36074471288660187,
"distilbert.transformer.layer.3.ffn.lin2": 0.34640995794256274,
"distilbert.transformer.layer.4.attention.q_lin": 0.36092885438401073,
"distilbert.transformer.layer.4.attention.k_lin": 0.35947300586341313,
"distilbert.transformer.layer.4.attention.v_lin": 0.3588137675606668,
"distilbert.transformer.layer.4.attention.out_lin": 0.363455277038118,
"distilbert.transformer.layer.4.ffn.lin1": 0.362920711045996,
"distilbert.transformer.layer.4.ffn.lin2": 0.34626701716664005,
"distilbert.transformer.layer.5.attention.q_lin": 0.358407123124324,
"distilbert.transformer.layer.5.attention.k_lin": 0.36063521716626545,
"distilbert.transformer.layer.5.attention.v_lin": 0.35714409014511134,
"distilbert.transformer.layer.5.attention.out_lin": 0.36417907904061186,
"distilbert.transformer.layer.5.ffn.lin1": 0.3631634533003135,
"distilbert.transformer.layer.5.ffn.lin2": 0.3588479363424607
},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 551.24,
"tokens_per_sec": 43663.37
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 536.67,
"L_256_vram_mb": 539.24,
"L_512_vram_mb": "unsupported (exceeds model max_position_embeddings 512)",
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 512)"
},
"timestamp": "2026-09-06T05:07:23.176304"
},
"results/unified_admm_pruner/distilbert/distilbert_sst2_full_full_metrics.json": {
"model": "distilbert_sst2_full",
"performance": {
"accuracy_pct": 90.48,
"f1_score": 0.9077,
"f1_macro": 0.9047,
"precision": 0.9052,
"recall": 0.9046,
"val_loss": 0.3091,
"status": "PASS",
"elapsed_s": 3.13
},
"efficiency": {
"total_params_before": 66955010,
"total_params_after": 66955010,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 390.7,
"tokens_per_sec": 50442.71
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 373.02,
"L_256_vram_mb": 375.08,
"L_512_vram_mb": "unsupported (exceeds model max_position_embeddings 512)",
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 512)"
},
"timestamp": "2026-09-06T05:06:57.843912"
},
"results/unified_admm_pruner/qwen3_5/Qwen3.5-0.8B_admm_pruner_full_metrics.json": {
"model": "Qwen3.5-0.8B_admm_pruner",
"performance": {
"accuracy_pct": 78.21,
"f1_score": 0.8201,
"f1_macro": 0.772,
"precision": 0.8326,
"recall": 0.7785,
"val_loss": 2.3223,
"status": "PASS",
"elapsed_s": 1382.28
},
"efficiency": {
"total_params_before": 763217728,
"total_params_after": 763217728,
"trainable_recovery_params": 10824704,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 48,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 10822656,
"avg_reconstruction_error": 0.3569369490334629,
"weighted_reconstruction_error": 0.3569,
"per_layer_reconstruction_errors": {
"model.layers.0.mlp.up_proj": 0.3603101785075146,
"model.layers.0.mlp.down_proj": 0.3610792003531197,
"model.layers.1.mlp.up_proj": 0.3599623147048513,
"model.layers.1.mlp.down_proj": 0.35981609341447346,
"model.layers.2.mlp.up_proj": 0.3596083327217418,
"model.layers.2.mlp.down_proj": 0.36047559894273723,
"model.layers.3.mlp.up_proj": 0.35911042158354994,
"model.layers.3.mlp.down_proj": 0.36082650450069265,
"model.layers.4.mlp.up_proj": 0.3583004510144316,
"model.layers.4.mlp.down_proj": 0.3578522123402067,
"model.layers.5.mlp.up_proj": 0.3568423580640812,
"model.layers.5.mlp.down_proj": 0.3581261519876344,
"model.layers.6.mlp.up_proj": 0.35431543580443375,
"model.layers.6.mlp.down_proj": 0.35524933251326996,
"model.layers.7.mlp.up_proj": 0.3557519791440242,
"model.layers.7.mlp.down_proj": 0.3558112854510554,
"model.layers.8.mlp.up_proj": 0.35518668752030713,
"model.layers.8.mlp.down_proj": 0.35543583058153005,
"model.layers.9.mlp.up_proj": 0.35481606340131105,
"model.layers.9.mlp.down_proj": 0.35319447142576466,
"model.layers.10.mlp.up_proj": 0.35485219885421493,
"model.layers.10.mlp.down_proj": 0.35339196831188363,
"model.layers.11.mlp.up_proj": 0.35544637808424484,
"model.layers.11.mlp.down_proj": 0.35438898777060784,
"model.layers.12.mlp.up_proj": 0.3558863117709546,
"model.layers.12.mlp.down_proj": 0.3542939972855905,
"model.layers.13.mlp.up_proj": 0.3563933887380592,
"model.layers.13.mlp.down_proj": 0.3539929945033691,
"model.layers.14.mlp.up_proj": 0.35227735043892106,
"model.layers.14.mlp.down_proj": 0.3465542336373255,
"model.layers.15.mlp.up_proj": 0.35383892612368345,
"model.layers.15.mlp.down_proj": 0.3507349691500875,
"model.layers.16.mlp.up_proj": 0.357002784597431,
"model.layers.16.mlp.down_proj": 0.3574295804843401,
"model.layers.17.mlp.up_proj": 0.3583954429669551,
"model.layers.17.mlp.down_proj": 0.35894381221052035,
"model.layers.18.mlp.up_proj": 0.35946074486992163,
"model.layers.18.mlp.down_proj": 0.3603116798277275,
"model.layers.19.mlp.up_proj": 0.3602171339960996,
"model.layers.19.mlp.down_proj": 0.3600186011509987,
"model.layers.20.mlp.up_proj": 0.36089014231911726,
"model.layers.20.mlp.down_proj": 0.36075581484030134,
"model.layers.21.mlp.up_proj": 0.3604565887695243,
"model.layers.21.mlp.down_proj": 0.35971208604665844,
"model.layers.22.mlp.up_proj": 0.3591936281975854,
"model.layers.22.mlp.down_proj": 0.3580071202078602,
"model.layers.23.mlp.up_proj": 0.358116862014818,
"model.layers.23.mlp.down_proj": 0.3499389224606871
},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 1879.76,
"tokens_per_sec": 1102.95
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 1615.45,
"L_256_vram_mb": 1636.85,
"L_512_vram_mb": 1665.13,
"L_1024_vram_mb": 1755.27
},
"timestamp": "2026-09-06T05:39:24.064033"
},
"results/unified_admm_pruner/qwen3_5/Qwen3.5-0.8B_full_full_metrics.json": {
"model": "Qwen3.5-0.8B_full",
"performance": {
"accuracy_pct": 67.55,
"f1_score": 0.5501,
"f1_macro": 0.6481,
"precision": 0.7703,
"recall": 0.6808,
"val_loss": 0.6364,
"status": "PASS",
"elapsed_s": 109.89
},
"efficiency": {
"total_params_before": 763217728,
"total_params_after": 763217728,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 10822656,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 1810.59,
"tokens_per_sec": 1287.97
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 1545.79,
"L_256_vram_mb": 1568.28,
"L_512_vram_mb": 1604.86,
"L_1024_vram_mb": 1685.72
},
"timestamp": "2026-09-06T05:16:21.672104"
},
"results/unified_admm_pruner/gpt2/gpt2_full_full_metrics.json": {
"model": "gpt2_full",
"performance": {
"accuracy_pct": 79.36,
"f1_score": 0.7978,
"f1_macro": 0.7935,
"precision": 0.7935,
"recall": 0.7935,
"val_loss": 0.4875,
"status": "PASS",
"elapsed_s": 7.71
},
"efficiency": {
"total_params_before": 124441344,
"total_params_after": 124441344,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 793.19,
"tokens_per_sec": 20559.85
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 681.07,
"L_256_vram_mb": 691.32,
"L_512_vram_mb": 721.95,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 1024)"
},
"timestamp": "2026-09-15T10:01:09.755969"
},
"results/unified_admm_pruner/gpt2/gpt2_admm_pruner_full_metrics.json": {
"model": "gpt2_admm_pruner",
"performance": {
"accuracy_pct": 51.49,
"f1_score": 0.1278,
"f1_macro": 0.3959,
"precision": 0.6296,
"recall": 0.5232,
"val_loss": 0.8947,
"status": "PASS_LOW",
"elapsed_s": 10.13
},
"efficiency": {
"total_params_before": 124441344,
"total_params_after": 124441344,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 48,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": 0.3564965102455689,
"weighted_reconstruction_error": 0.3545,
"per_layer_reconstruction_errors": {
"transformer.h.0.attn.c_attn": 0.34573750030886846,
"transformer.h.0.attn.c_proj": 0.3623429562000621,
"transformer.h.0.mlp.c_fc": 0.3329107428706859,
"transformer.h.0.mlp.c_proj": 0.32500372110609116,
"transformer.h.1.attn.c_attn": 0.3555693677179189,
"transformer.h.1.attn.c_proj": 0.3540393076308105,
"transformer.h.1.mlp.c_fc": 0.3576601271228946,
"transformer.h.1.mlp.c_proj": 0.3330394193057879,
"transformer.h.2.attn.c_attn": 0.3550603439997944,
"transformer.h.2.attn.c_proj": 0.35901168911274045,
"transformer.h.2.mlp.c_fc": 0.3585415095056415,
"transformer.h.2.mlp.c_proj": 0.33727693592902164,
"transformer.h.3.attn.c_attn": 0.3590091244059882,
"transformer.h.3.attn.c_proj": 0.36384751466734244,
"transformer.h.3.mlp.c_fc": 0.36293421311098123,
"transformer.h.3.mlp.c_proj": 0.3408853516858818,
"transformer.h.4.attn.c_attn": 0.35890129936478965,
"transformer.h.4.attn.c_proj": 0.36320466444056937,
"transformer.h.4.mlp.c_fc": 0.3624299543819203,
"transformer.h.4.mlp.c_proj": 0.34751007855756516,
"transformer.h.5.attn.c_attn": 0.3598443550896834,
"transformer.h.5.attn.c_proj": 0.36218999078252556,
"transformer.h.5.mlp.c_fc": 0.3634651521803743,
"transformer.h.5.mlp.c_proj": 0.35099468117020377,
"transformer.h.6.attn.c_attn": 0.36037117227741217,
"transformer.h.6.attn.c_proj": 0.3627174460017233,
"transformer.h.6.mlp.c_fc": 0.36283887955512323,
"transformer.h.6.mlp.c_proj": 0.35138366621080197,
"transformer.h.7.attn.c_attn": 0.3609113980358935,
"transformer.h.7.attn.c_proj": 0.3625578929284949,
"transformer.h.7.mlp.c_fc": 0.3632534404638959,
"transformer.h.7.mlp.c_proj": 0.3534700579388818,
"transformer.h.8.attn.c_attn": 0.36059546920945407,
"transformer.h.8.attn.c_proj": 0.3650881504770807,
"transformer.h.8.mlp.c_fc": 0.3631176593062812,
"transformer.h.8.mlp.c_proj": 0.35253815294221136,
"transformer.h.9.attn.c_attn": 0.36037262274134296,
"transformer.h.9.attn.c_proj": 0.36388518248446,
"transformer.h.9.mlp.c_fc": 0.36298235750897345,
"transformer.h.9.mlp.c_proj": 0.3520916082337052,
"transformer.h.10.attn.c_attn": 0.35997047638133145,
"transformer.h.10.attn.c_proj": 0.36028775356299336,
"transformer.h.10.mlp.c_fc": 0.36239689218593235,
"transformer.h.10.mlp.c_proj": 0.3504399898107858,
"transformer.h.11.attn.c_attn": 0.35878449698117093,
"transformer.h.11.attn.c_proj": 0.3880565470317182,
"transformer.h.11.mlp.c_fc": 0.3609011965847491,
"transformer.h.11.mlp.c_proj": 0.3414099822847483
},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 628.41,
"tokens_per_sec": 22893.93
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 510.91,
"L_256_vram_mb": 524.41,
"L_512_vram_mb": 552.17,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 1024)"
},
"timestamp": "2026-09-15T10:01:20.105125"
},
"results/unified_admm_pruner/roberta/summary.csv": "module_name,status,accuracy_pct,f1_score,f1_macro,precision,recall,val_loss,elapsed_s,peak_vram_mb,tokens_per_sec,total_params_before,total_params_after,trainable_recovery_params,compression_rate,compression_ratio_pct,flops_reduction_pct,num_compressed_layers,num_fallback_layers,avg_reconstruction_error,weighted_reconstruction_error,num_injected_attention_layers,lora_params,error,path,timestamp\nfull,PASS,88.76,0.8942,0.8872,0.891,0.8868,0.295,5.97,300.73,30825.8,127325954,127325954,0,1.0,0.0,0.0,0,0,,,0,2678784,,adapted_roberta,2026-09-06T05:08:13.785782\nadmm_pruner,PASS,80.62,0.7966,0.8058,0.8112,0.8073,0.4539,100.32,336.39,13333.06,127325954,127325954,3270914,1.0,0.0,0.0,72,0,0.3513150578015672,0.3495,0,2678784,,adapted_roberta,2026-09-06T05:09:54.345720\n",
"results/unified_admm_pruner/distilbert/summary.csv": "module_name,status,accuracy_pct,f1_score,f1_macro,precision,recall,val_loss,elapsed_s,peak_vram_mb,tokens_per_sec,total_params_before,total_params_after,trainable_recovery_params,compression_rate,compression_ratio_pct,flops_reduction_pct,num_compressed_layers,num_fallback_layers,avg_reconstruction_error,weighted_reconstruction_error,num_injected_attention_layers,lora_params,error,path,timestamp\nfull,PASS,90.48,0.9077,0.9047,0.9052,0.9046,0.3091,3.13,390.7,50442.71,66955010,66955010,0,1.0,0.0,0.0,0,0,,,0,0,,legacy_bert,2026-09-06T05:06:58.067448\nadmm_pruner,PASS,84.29,0.8603,0.8404,0.8601,0.8409,0.6574,25.31,551.24,43663.37,66955010,66955010,4135682,1.0,0.0,0.0,36,0,0.35710002246145245,0.3552,0,0,,legacy_bert,2026-09-06T05:07:23.627199\n",
"results/unified_admm_pruner/qwen3_5/summary.csv": "module_name,status,accuracy_pct,f1_score,f1_macro,precision,recall,val_loss,elapsed_s,peak_vram_mb,tokens_per_sec,total_params_before,total_params_after,trainable_recovery_params,compression_rate,compression_ratio_pct,flops_reduction_pct,num_compressed_layers,num_fallback_layers,avg_reconstruction_error,weighted_reconstruction_error,num_injected_attention_layers,lora_params,error,path,timestamp\nfull,PASS,67.55,0.5501,0.6481,0.7703,0.6808,0.6364,109.89,1810.59,1287.97,763217728,763217728,0,1.0,0.0,0.0,0,0,,,0,10822656,,adapted_qwen3_5,2026-09-06T05:16:24.683603\nadmm_pruner,PASS,78.21,0.8201,0.772,0.8326,0.7785,2.3223,1382.28,1879.76,1102.95,763217728,763217728,10824704,1.0,0.0,0.0,48,0,0.3569369490334629,0.3569,0,10822656,,adapted_qwen3_5,2026-09-06T05:39:27.392824\n",
"results/unified_admm_pruner/gpt2/summary.csv": "module_name,status,accuracy_pct,f1_score,f1_macro,precision,recall,val_loss,elapsed_s,peak_vram_mb,tokens_per_sec,total_params_before,total_params_after,trainable_recovery_params,compression_rate,compression_ratio_pct,flops_reduction_pct,num_compressed_layers,num_fallback_layers,avg_reconstruction_error,weighted_reconstruction_error,num_injected_attention_layers,lora_params,error,path,timestamp\nfull,PASS,79.36,0.7978,0.7935,0.7935,0.7935,0.4875,7.71,793.19,20559.85,124441344,124441344,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_gpt,2026-09-15T10:01:10.620503\nadmm_pruner,PASS_LOW,51.49,0.1278,0.3959,0.6296,0.5232,0.8947,10.13,628.41,22893.93,124441344,124441344,0,1.0,0.0,0.0,48,0,0.3564965102455689,0.3545,0,0,note: evaluated accuracy (51.49%) below chance threshold (56.00%),adapted_gpt,2026-09-15T10:01:21.016057\n",
"results/benchmark_summary_v2.csv": "track5_pruning,admm_pruner,compressor,bert,sst2,PASS,,12.85,17.15,30.0,0.0,1.0,4386178,4386178,1.0,573.4,0.0,0.6921,0.0,0.0,0.0,0.019,\"{'loss': [0.6921, 0.6921, 0.6921, 0.6921, 0.6921], 'grad_norm': [0.019, 0.019, 0.019, 0.019, 0.019], 'fwd_latency_ms': [12.4815, 12.7439, 12.9441, 13.2684, 12.8115], 'bwd_latency_ms': [16.4996, 18.9523, 16.7381, 17.0654, 16.5036]}\",2026-08-26 03:47:48\ntrack5_pruning,admm_pruner,compressor,bert,mrpc,PASS,,13.12,17.36,30.48,0.0,1.0,4386178,4386178,1.0,573.4,0.0,0.6778,0.0,0.0,0.0,0.9369,\"{'loss': [0.6778, 0.6778, 0.6778, 0.6778, 0.6778], 'grad_norm': [0.9369, 0.9369, 0.9369, 0.9369, 0.9369], 'fwd_latency_ms': [13.1901, 12.7426, 12.7139, 13.9175, 13.0577], 'bwd_latency_ms': [17.48, 17.4957, 17.0706, 17.6187, 17.1587]}\",2026-08-26 03:47:49\ntrack5_pruning,admm_pruner,compressor,roberta,sst2,PASS,,13.33,17.84,31.17,0.0,1.0,6913282,6913282,1.0,583.3,0.0,0.6942,0.0,0.0,0.0,0.1693,\"{'loss': [0.6942, 0.6942, 0.6942, 0.6942, 0.6942], 'grad_norm': [0.1693, 0.1693, 0.1693, 0.1693, 0.1693], 'fwd_latency_ms': [13.2187, 13.4301, 13.2852, 13.471, 13.2248], 'bwd_latency_ms': [17.4982, 17.6168, 18.174, 17.9032, 18.01]}\",2026-08-26 03:47:50\ntrack5_pruning,admm_pruner,compressor,roberta,mrpc,PASS,,13.13,17.54,30.67,0.0,1.0,6913282,6913282,1.0,583.3,0.0,0.7007,0.0,0.0,0.0,1.1035,\"{'loss': [0.7007, 0.7007, 0.7007, 0.7007, 0.7007], 'grad_norm': [1.1035, 1.1035, 1.1035, 1.1035, 1.1035], 'fwd_latency_ms': [12.9892, 13.1911, 12.9921, 13.1611, 13.3364], 'bwd_latency_ms': [17.3085, 17.3953, 17.3968, 17.8011, 17.7951]}\",2026-08-26 03:47:50\ntrack5_pruning,admm_pruner,compressor,t5,sst2,PASS,,19.65,33.58,53.23,0.0,1.0,4506368,4506368,1.0,1108.7,0.0,0.965,0.0,0.0,0.0,7.3095,\"{'loss': [0.965, 0.965, 0.965, 0.965, 0.965], 'grad_norm': [7.3095, 7.3095, 7.3095, 7.3095, 7.3095], 'fwd_latency_ms': [15.9601, 20.1442, 20.0174, 21.3187, 20.8214], 'bwd_latency_ms': [35.7717, 32.8285, 33.0412, 32.9932, 33.2622]}\",2026-08-26 03:47:51\ntrack5_pruning,admm_pruner,compressor,t5,mrpc,PASS,,19.44,33.01,52.45,0.0,1.0,4506368,4506368,1.0,1108.7,0.0,0.6527,0.0,0.0,0.0,2.4943,\"{'loss': [0.6527, 0.6527, 0.6527, 0.6527, 0.6527], 'grad_norm': [2.4943, 2.4943, 2.4943, 2.4943, 2.4943], 'fwd_latency_ms': [15.9242, 20.3174, 20.4764, 20.1818, 20.3229], 'bwd_latency_ms': [33.0888, 33.1634, 33.0502, 32.7688, 32.9549]}\",2026-08-26 03:47:52",
"results/all_121_compressors_gpu_matrix.csv": "admm_pruner,cuda,PASS,True,True,1.0,284.2,0.485,",
"results/run_quality_all_screening_summary.csv": "run_quality_20260827_052444,admm_pruner,DISTILBERT,SST2,PASS,77.64,0.0,77.64,77.64,1.0,0.7821,0.8359,71.94,187.69,2009.4,63106.4,1450000000.0,2026-08-27T05:46:09.841719",
"results/unified_admm_pruner_4models_summary.csv": "distilbert,distilbert,0.9048,0.8429000000000001,-0.061899999999999955,0.8603,PASS,True,36,0,0.5,0.35710002246145245,0.0002415028122615897,0.0,551.24,results\\unified_admm_pruner\\distilbert,https://huggingface.co/Hooshaai/svd-linear-attention-distilbert-admm-distill\nroberta,roberta,0.8876000000000001,0.8062,-0.08140000000000003,0.7966,PASS,True,72,0,0.5,0.3513150578015672,0.0003741950268805441,0.0,336.39,results\\unified_admm_pruner\\roberta,https://huggingface.co/Hooshaai/svd-linear-attention-roberta-admm-distill\ngpt2,gpt2,0.7936,0.5149,-0.27869999999999995,0.1278,PASS_LOW,False,48,0,0.5,0.3564965102455689,0.0020508204700793917,0.0,628.41,results\\unified_admm_pruner\\gpt2,N/A (Failed Gate)"
},
"notes": []
},
"aegis_hybrid": {
"module_name": "aegis_hybrid",
"source_file": "Codes/framework/attentions/aegis_hybrid.py",
"source_lines_count": 1033,
"architecture": {
"class_name": "FactorizedLinear",
"base_class": "nn.Module",
"init_args": "self, in_features_or_lin1: Union[int, nn.Linear], out_features_or_lin2: Optional[Union[int, nn.Linear]] = None, rank: int = 128, lora_rank: int = 32, bias: bool = True, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None, ) -> None: super().__init__() if isinstance(in_features_or_lin1, nn.Linear) and isinstance(out_features_or_lin2, nn.Linear",
"registry_names": [
"aegis_hybrid",
"aegis_hybrid_inject",
"inject_aegis_hybrid",
"aegis_hybrid_attn",
"aegis_hybrid_attention"
]
},
"params": [
{
"name": "lin1",
"type": "nn.Linear",
"declaration": "nn.Linear(in_features, self.rank, bias=False, **factory_kwargs)"
},
{
"name": "lin2",
"type": "nn.Linear",
"declaration": "nn.Linear(self.rank, out_features, bias=bias, **factory_kwargs)"
},
{
"name": "fusion_gate",
"type": "nn.Linear",
"declaration": "nn.Linear(self.hidden_size, self.num_heads)"
}
],
"metrics": {
"results/unified_aegis_hybrid/roberta/roberta-base_full_full_metrics.json": {
"model": "roberta-base_full",
"performance": {
"accuracy_pct": 78.56,
"f1_score": 0.8124,
"f1_macro": 0.7811,
"precision": 0.8051,
"recall": 0.7832,
"val_loss": 0.4381,
"status": "PASS",
"elapsed_s": 6.85
},
"efficiency": {
"total_params_before": 124647170,
"total_params_after": 124647170,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 712.34,
"tokens_per_sec": 23951.82
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 683.27,
"L_256_vram_mb": 685.84,
"L_512_vram_mb": 699.47,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 514)"
},
"timestamp": "2026-09-15T10:03:13.674317"
},
"results/unified_aegis_hybrid/roberta/roberta-base_aegis_hybrid_full_metrics.json": {
"model": "roberta-base_aegis_hybrid",
"performance": {
"accuracy_pct": 51.26,
"f1_score": 0.6698,
"f1_macro": 0.3699,
"precision": 0.5315,
"recall": 0.5041,
"val_loss": 0.806,
"status": "PASS_LOW",
"elapsed_s": 46.95
},
"efficiency": {
"total_params_before": 124647170,
"total_params_after": 108872018,
"trainable_recovery_params": 3090000,
"compression_rate": 1.1449,
"compression_ratio_pct": 12.66,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 12,
"lora_params": 1179648,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {
"roberta.encoder.layer.0.attention.self": 32,
"roberta.encoder.layer.0.attention.self._wrapped_attention.q_proj": 32,
"roberta.encoder.layer.0.attention.self._wrapped_attention.k_proj": 32,
"roberta.encoder.layer.0.attention.self._wrapped_attention.v_proj": 32,
"roberta.encoder.layer.0.attention.self._wrapped_attention.out_proj": 32,
"roberta.encoder.layer.1.attention.self": 32,
"roberta.encoder.layer.1.attention.self._wrapped_attention.q_proj": 32,
"roberta.encoder.layer.1.attention.self._wrapped_attention.k_proj": 32,
"roberta.encoder.layer.1.attention.self._wrapped_attention.v_proj": 32,
"roberta.encoder.layer.1.attention.self._wrapped_attention.out_proj": 32,
"roberta.encoder.layer.2.attention.self": 32,
"roberta.encoder.layer.2.attention.self._wrapped_attention.q_proj": 32,
"roberta.encoder.layer.2.attention.self._wrapped_attention.k_proj": 32,
"roberta.encoder.layer.2.attention.self._wrapped_attention.v_proj": 32,
"roberta.encoder.layer.2.attention.self._wrapped_attention.out_proj": 32,
"roberta.encoder.layer.3.attention.self": 32,
"roberta.encoder.layer.3.attention.self._wrapped_attention.q_proj": 32,
"roberta.encoder.layer.3.attention.self._wrapped_attention.k_proj": 32,
"roberta.encoder.layer.3.attention.self._wrapped_attention.v_proj": 32,
"roberta.encoder.layer.3.attention.self._wrapped_attention.out_proj": 32,
"roberta.encoder.layer.4.attention.self": 32,
"roberta.encoder.layer.4.attention.self._wrapped_attention.q_proj": 32,
"roberta.encoder.layer.4.attention.self._wrapped_attention.k_proj": 32,
"roberta.encoder.layer.4.attention.self._wrapped_attention.v_proj": 32,
"roberta.encoder.layer.4.attention.self._wrapped_attention.out_proj": 32,
"roberta.encoder.layer.5.attention.self": 32,
"roberta.encoder.layer.5.attention.self._wrapped_attention.q_proj": 32,
"roberta.encoder.layer.5.attention.self._wrapped_attention.k_proj": 32,
"roberta.encoder.layer.5.attention.self._wrapped_attention.v_proj": 32,
"roberta.encoder.layer.5.attention.self._wrapped_attention.out_proj": 32,
"roberta.encoder.layer.6.attention.self": 32,
"roberta.encoder.layer.6.attention.self._wrapped_attention.q_proj": 32,
"roberta.encoder.layer.6.attention.self._wrapped_attention.k_proj": 32,
"roberta.encoder.layer.6.attention.self._wrapped_attention.v_proj": 32,
"roberta.encoder.layer.6.attention.self._wrapped_attention.out_proj": 32,
"roberta.encoder.layer.7.attention.self": 32,
"roberta.encoder.layer.7.attention.self._wrapped_attention.q_proj": 32,
"roberta.encoder.layer.7.attention.self._wrapped_attention.k_proj": 32,
"roberta.encoder.layer.7.attention.self._wrapped_attention.v_proj": 32,
"roberta.encoder.layer.7.attention.self._wrapped_attention.out_proj": 32,
"roberta.encoder.layer.8.attention.self": 32,
"roberta.encoder.layer.8.attention.self._wrapped_attention.q_proj": 32,
"roberta.encoder.layer.8.attention.self._wrapped_attention.k_proj": 32,
"roberta.encoder.layer.8.attention.self._wrapped_attention.v_proj": 32,
"roberta.encoder.layer.8.attention.self._wrapped_attention.out_proj": 32,
"roberta.encoder.layer.9.attention.self": 32,
"roberta.encoder.layer.9.attention.self._wrapped_attention.q_proj": 32,
"roberta.encoder.layer.9.attention.self._wrapped_attention.k_proj": 32,
"roberta.encoder.layer.9.attention.self._wrapped_attention.v_proj": 32,
"roberta.encoder.layer.9.attention.self._wrapped_attention.out_proj": 32,
"roberta.encoder.layer.10.attention.self": 32,
"roberta.encoder.layer.10.attention.self._wrapped_attention.q_proj": 32,
"roberta.encoder.layer.10.attention.self._wrapped_attention.k_proj": 32,
"roberta.encoder.layer.10.attention.self._wrapped_attention.v_proj": 32,
"roberta.encoder.layer.10.attention.self._wrapped_attention.out_proj": 32,
"roberta.encoder.layer.11.attention.self": 32,
"roberta.encoder.layer.11.attention.self._wrapped_attention.q_proj": 32,
"roberta.encoder.layer.11.attention.self._wrapped_attention.k_proj": 32,
"roberta.encoder.layer.11.attention.self._wrapped_attention.v_proj": 32,
"roberta.encoder.layer.11.attention.self._wrapped_attention.out_proj": 32
}
},
"hardware": {
"peak_vram_mb": 596.29,
"tokens_per_sec": 7060.79
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 549.74,
"L_256_vram_mb": 557.98,
"L_512_vram_mb": 589.32,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 514)"
},
"timestamp": "2026-09-15T10:04:00.692122"
},
"results/unified_aegis_hybrid/distilbert/distilbert_sst2_full_full_metrics.json": {
"model": "distilbert_sst2_full",
"performance": {
"accuracy_pct": 90.71,
"f1_score": 0.9099,
"f1_macro": 0.907,
"precision": 0.9075,
"recall": 0.9068,
"val_loss": 0.3027,
"status": "PASS",
"elapsed_s": 3.43
},
"efficiency": {
"total_params_before": 66955010,
"total_params_after": 66955010,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 389.51,
"tokens_per_sec": 44854.6
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 371.82,
"L_256_vram_mb": 373.89,
"L_512_vram_mb": "unsupported (exceeds model max_position_embeddings 512)",
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 512)"
},
"timestamp": "2026-09-15T10:02:05.439555"
},
"results/unified_aegis_hybrid/distilbert/distilbert_sst2_aegis_hybrid_full_metrics.json": {
"model": "distilbert_sst2_aegis_hybrid",
"performance": {
"accuracy_pct": 84.98,
"f1_score": 0.8634,
"f1_macro": 0.8483,
"precision": 0.8599,
"recall": 0.8482,
"val_loss": 0.3557,
"status": "PASS",
"elapsed_s": 47.68
},
"efficiency": {
"total_params_before": 66955010,
"total_params_after": 59652626,
"trainable_recovery_params": 2134800,
"compression_rate": 1.1224,
"compression_ratio_pct": 10.91,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 1179648,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {
"distilbert.transformer.layer.0.attention": 128,
"distilbert.transformer.layer.0.attention.q_proj": 128,
"distilbert.transformer.layer.0.attention.k_proj": 128,
"distilbert.transformer.layer.0.attention.v_proj": 128,
"distilbert.transformer.layer.0.attention.out_proj": 128,
"distilbert.transformer.layer.1.attention": 128,
"distilbert.transformer.layer.1.attention.q_proj": 128,
"distilbert.transformer.layer.1.attention.k_proj": 128,
"distilbert.transformer.layer.1.attention.v_proj": 128,
"distilbert.transformer.layer.1.attention.out_proj": 128,
"distilbert.transformer.layer.2.attention": 128,
"distilbert.transformer.layer.2.attention.q_proj": 128,
"distilbert.transformer.layer.2.attention.k_proj": 128,
"distilbert.transformer.layer.2.attention.v_proj": 128,
"distilbert.transformer.layer.2.attention.out_proj": 128,
"distilbert.transformer.layer.3.attention": 128,
"distilbert.transformer.layer.3.attention.q_proj": 128,
"distilbert.transformer.layer.3.attention.k_proj": 128,
"distilbert.transformer.layer.3.attention.v_proj": 128,
"distilbert.transformer.layer.3.attention.out_proj": 128,
"distilbert.transformer.layer.4.attention": 128,
"distilbert.transformer.layer.4.attention.q_proj": 128,
"distilbert.transformer.layer.4.attention.k_proj": 128,
"distilbert.transformer.layer.4.attention.v_proj": 128,
"distilbert.transformer.layer.4.attention.out_proj": 128,
"distilbert.transformer.layer.5.attention": 128,
"distilbert.transformer.layer.5.attention.q_proj": 128,
"distilbert.transformer.layer.5.attention.k_proj": 128,
"distilbert.transformer.layer.5.attention.v_proj": 128,
"distilbert.transformer.layer.5.attention.out_proj": 128
}
},
"hardware": {
"peak_vram_mb": 319.22,
"tokens_per_sec": 3802.96
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 276.42,
"L_256_vram_mb": 283.57,
"L_512_vram_mb": "unsupported (exceeds model max_position_embeddings 512)",
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 512)"
},
"timestamp": "2026-09-17T05:12:36.645097"
},
"results/unified_aegis_hybrid/gpt2/gpt2_full_full_metrics.json": {
"model": "gpt2_full",
"performance": {
"accuracy_pct": 79.36,
"f1_score": 0.7978,
"f1_macro": 0.7935,
"precision": 0.7935,
"recall": 0.7935,
"val_loss": 0.4875,
"status": "PASS",
"elapsed_s": 7.52
},
"efficiency": {
"total_params_before": 124441344,
"total_params_after": 124441344,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 793.19,
"tokens_per_sec": 21011.47
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 681.07,
"L_256_vram_mb": 691.32,
"L_512_vram_mb": 721.95,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 1024)"
},
"timestamp": "2026-09-15T10:04:45.155624"
},
"results/unified_aegis_hybrid/gpt2/gpt2_aegis_hybrid_full_metrics.json": {
"model": "gpt2_aegis_hybrid",
"performance": {
"accuracy_pct": 59.52,
"f1_score": 0.6735,
"f1_macro": 0.5705,
"precision": 0.6155,
"recall": 0.591,
"val_loss": 0.6649,
"status": "PASS",
"elapsed_s": 38.78
},
"efficiency": {
"total_params_before": 124441344,
"total_params_after": 101579088,
"trainable_recovery_params": 3090000,
"compression_rate": 1.2251,
"compression_ratio_pct": 18.37,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 12,
"lora_params": 1179648,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {
"transformer.h.0.attn": 32,
"transformer.h.0.attn._wrapped_attention.q_proj": 32,
"transformer.h.0.attn._wrapped_attention.k_proj": 32,
"transformer.h.0.attn._wrapped_attention.v_proj": 32,
"transformer.h.0.attn._wrapped_attention.out_proj": 32,
"transformer.h.1.attn": 32,
"transformer.h.1.attn._wrapped_attention.q_proj": 32,
"transformer.h.1.attn._wrapped_attention.k_proj": 32,
"transformer.h.1.attn._wrapped_attention.v_proj": 32,
"transformer.h.1.attn._wrapped_attention.out_proj": 32,
"transformer.h.2.attn": 32,
"transformer.h.2.attn._wrapped_attention.q_proj": 32,
"transformer.h.2.attn._wrapped_attention.k_proj": 32,
"transformer.h.2.attn._wrapped_attention.v_proj": 32,
"transformer.h.2.attn._wrapped_attention.out_proj": 32,
"transformer.h.3.attn": 32,
"transformer.h.3.attn._wrapped_attention.q_proj": 32,
"transformer.h.3.attn._wrapped_attention.k_proj": 32,
"transformer.h.3.attn._wrapped_attention.v_proj": 32,
"transformer.h.3.attn._wrapped_attention.out_proj": 32,
"transformer.h.4.attn": 32,
"transformer.h.4.attn._wrapped_attention.q_proj": 32,
"transformer.h.4.attn._wrapped_attention.k_proj": 32,
"transformer.h.4.attn._wrapped_attention.v_proj": 32,
"transformer.h.4.attn._wrapped_attention.out_proj": 32,
"transformer.h.5.attn": 32,
"transformer.h.5.attn._wrapped_attention.q_proj": 32,
"transformer.h.5.attn._wrapped_attention.k_proj": 32,
"transformer.h.5.attn._wrapped_attention.v_proj": 32,
"transformer.h.5.attn._wrapped_attention.out_proj": 32,
"transformer.h.6.attn": 32,
"transformer.h.6.attn._wrapped_attention.q_proj": 32,
"transformer.h.6.attn._wrapped_attention.k_proj": 32,
"transformer.h.6.attn._wrapped_attention.v_proj": 32,
"transformer.h.6.attn._wrapped_attention.out_proj": 32,
"transformer.h.7.attn": 32,
"transformer.h.7.attn._wrapped_attention.q_proj": 32,
"transformer.h.7.attn._wrapped_attention.k_proj": 32,
"transformer.h.7.attn._wrapped_attention.v_proj": 32,
"transformer.h.7.attn._wrapped_attention.out_proj": 32,
"transformer.h.8.attn": 32,
"transformer.h.8.attn._wrapped_attention.q_proj": 32,
"transformer.h.8.attn._wrapped_attention.k_proj": 32,
"transformer.h.8.attn._wrapped_attention.v_proj": 32,
"transformer.h.8.attn._wrapped_attention.out_proj": 32,
"transformer.h.9.attn": 32,
"transformer.h.9.attn._wrapped_attention.q_proj": 32,
"transformer.h.9.attn._wrapped_attention.k_proj": 32,
"transformer.h.9.attn._wrapped_attention.v_proj": 32,
"transformer.h.9.attn._wrapped_attention.out_proj": 32,
"transformer.h.10.attn": 32,
"transformer.h.10.attn._wrapped_attention.q_proj": 32,
"transformer.h.10.attn._wrapped_attention.k_proj": 32,
"transformer.h.10.attn._wrapped_attention.v_proj": 32,
"transformer.h.10.attn._wrapped_attention.out_proj": 32,
"transformer.h.11.attn": 32,
"transformer.h.11.attn._wrapped_attention.q_proj": 32,
"transformer.h.11.attn._wrapped_attention.k_proj": 32,
"transformer.h.11.attn._wrapped_attention.v_proj": 32,
"transformer.h.11.attn._wrapped_attention.out_proj": 32
}
},
"hardware": {
"peak_vram_mb": 476.0,
"tokens_per_sec": 8474.53
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 434.32,
"L_256_vram_mb": 439.39,
"L_512_vram_mb": 470.1,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 1024)"
},
"timestamp": "2026-09-15T10:05:24.021358"
},
"results/unified_aegis_hybrid/gpt2/hf_export/metrics.json": {
"model_name": "gpt2",
"module": "aegis_hybrid",
"status": "PASS",
"accuracy_pct": 59.52,
"f1": 0.0,
"precision": 0.6155,
"recall": 0.591,
"eval_loss": 0.0,
"params_total": null,
"params_active": null,
"param_reduction_pct": null,
"peak_vram_mb": 476.0,
"tokens_per_sec": 8474.53,
"elapsed_seconds": null,
"full_metrics": {
"module_name": "aegis_hybrid",
"status": "PASS",
"accuracy_pct": 59.52,
"f1_score": 0.6735,
"f1_macro": 0.5705,
"precision": 0.6155,
"recall": 0.591,
"val_loss": 0.6649,
"elapsed_s": 38.79,
"peak_vram_mb": 476.0,
"tokens_per_sec": 8474.53,
"total_params_before": 124441344,
"total_params_after": 101579088,
"trainable_recovery_params": 3090000,
"compression_rate": 1.2251,
"compression_ratio_pct": 18.37,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"avg_reconstruction_error": NaN,
"weighted_reconstruction_error": NaN,
"num_injected_attention_layers": 12,
"lora_params": 1179648,
"error": NaN,
"path": "adapted_gpt",
"timestamp": "2026-09-15T10:05:25.000075"
}
},
"results/unified_aegis_hybrid/distilbert/hf_export/metrics.json": {
"model_name": "distilbert",
"module": "aegis_hybrid",
"status": "PASS",
"accuracy_pct": 84.98,
"f1": 0.0,
"precision": 0.8599,
"recall": 0.8482,
"eval_loss": 0.0,
"params_total": null,
"params_active": null,
"param_reduction_pct": null,
"peak_vram_mb": 319.22,
"tokens_per_sec": 3802.96,
"elapsed_seconds": null,
"full_metrics": {
"module_name": "aegis_hybrid",
"status": "PASS",
"accuracy_pct": 84.98,
"f1_score": 0.8634,
"f1_macro": 0.8483,
"precision": 0.8599,
"recall": 0.8482,
"val_loss": 0.3557,
"elapsed_s": 47.68,
"peak_vram_mb": 319.22,
"tokens_per_sec": 3802.96,
"total_params_before": 66955010,
"total_params_after": 59652626,
"trainable_recovery_params": 2134800,
"compression_rate": 1.1224,
"compression_ratio_pct": 10.91,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"avg_reconstruction_error": NaN,
"weighted_reconstruction_error": NaN,
"num_injected_attention_layers": 0,
"lora_params": 1179648,
"error": NaN,
"path": "legacy_bert",
"timestamp": "2026-09-17T05:12:38.532828"
}
},
"results/unified_aegis_hybrid/roberta/summary.csv": "module_name,status,accuracy_pct,f1_score,f1_macro,precision,recall,val_loss,elapsed_s,peak_vram_mb,tokens_per_sec,total_params_before,total_params_after,trainable_recovery_params,compression_rate,compression_ratio_pct,flops_reduction_pct,num_compressed_layers,num_fallback_layers,avg_reconstruction_error,weighted_reconstruction_error,num_injected_attention_layers,lora_params,error,path,timestamp\nfull,PASS,78.56,0.8124,0.7811,0.8051,0.7832,0.4381,6.85,712.34,23951.82,124647170,124647170,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_roberta,2026-09-15T10:03:14.463238\naegis_hybrid,PASS_LOW,51.26,0.6698,0.3699,0.5315,0.5041,0.806,46.95,596.29,7060.79,124647170,108872018,3090000,1.1449,12.66,0.0,0,0,,,12,1179648,note: evaluated accuracy (51.26%) below chance threshold (56.00%),adapted_roberta,2026-09-15T10:04:01.688122\n",
"results/unified_aegis_hybrid/distilbert/summary.csv": "module_name,status,accuracy_pct,f1_score,f1_macro,precision,recall,val_loss,elapsed_s,peak_vram_mb,tokens_per_sec,total_params_before,total_params_after,trainable_recovery_params,compression_rate,compression_ratio_pct,flops_reduction_pct,num_compressed_layers,num_fallback_layers,avg_reconstruction_error,weighted_reconstruction_error,num_injected_attention_layers,lora_params,error,path,timestamp\nfull,PASS,90.71,0.9099,0.907,0.9075,0.9068,0.3027,14.51,388.55,10631.82,66955010,66955010,0,1.0,0.0,0.0,0,0,,,0,0,,legacy_bert,2026-09-17T05:11:49.929002\naegis_hybrid,PASS,84.98,0.8634,0.8483,0.8599,0.8482,0.3557,47.68,319.22,3802.96,66955010,59652626,2134800,1.1224,10.91,0.0,0,0,,,0,1179648,,legacy_bert,2026-09-17T05:12:38.532828\n",
"results/unified_aegis_hybrid/gpt2/summary.csv": "module_name,status,accuracy_pct,f1_score,f1_macro,precision,recall,val_loss,elapsed_s,peak_vram_mb,tokens_per_sec,total_params_before,total_params_after,trainable_recovery_params,compression_rate,compression_ratio_pct,flops_reduction_pct,num_compressed_layers,num_fallback_layers,avg_reconstruction_error,weighted_reconstruction_error,num_injected_attention_layers,lora_params,error,path,timestamp\nfull,PASS,79.36,0.7978,0.7935,0.7935,0.7935,0.4875,7.52,793.19,21011.47,124441344,124441344,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_gpt,2026-09-15T10:04:45.941960\naegis_hybrid,PASS,59.52,0.6735,0.5705,0.6155,0.591,0.6649,38.79,476.0,8474.53,124441344,101579088,3090000,1.2251,18.37,0.0,0,0,,,12,1179648,,adapted_gpt,2026-09-15T10:05:25.000075\n",
"results/matrix_test_summary.csv": "qwen3_5,qwen3_5,aegis_hybrid,attention,PASS,True,True,14.619,4255.0,\ngpt2,gpt2,aegis_hybrid,attention,PASS,True,True,5.286,1442.0,\nroberta,roberta,aegis_hybrid,attention,PASS,True,True,4.52,1587.8,\ndistilbert,distilbert,aegis_hybrid,attention,PASS,True,True,3.532,809.0,",
"results/benchmark_summary_v2.csv": "track2_ssm,aegis_hybrid,attention,bert,sst2,PASS,,31.95,54.41,86.36,0.0,1.0,4536866,4536866,0.967,1842.5,0.0,0.6945,0.0,0.0,0.0,0.1971,\"{'loss': [0.6926, 0.6946, 0.6943, 0.6948, 0.6961], 'grad_norm': [0.1937, 0.192, 0.1973, 0.1933, 0.2094], 'fwd_latency_ms': [31.9785, 31.0587, 31.1721, 31.4606, 34.0945], 'bwd_latency_ms': [55.5934, 53.7134, 54.236, 53.6762, 54.8221]}\",2026-08-26 02:34:54\ntrack2_ssm,aegis_hybrid,attention,bert,mrpc,PASS,,31.84,54.7,86.54,0.0,1.0,4536866,4536866,0.967,1842.5,0.0,0.6786,0.0,0.0,0.0,1.0306,\"{'loss': [0.6782, 0.6787, 0.6773, 0.6795, 0.6792], 'grad_norm': [1.0346, 1.0187, 1.0281, 1.0326, 1.0392], 'fwd_latency_ms': [31.4368, 31.9404, 32.0057, 32.1949, 31.6301], 'bwd_latency_ms': [54.7028, 55.0907, 54.768, 54.7673, 54.1588]}\",2026-08-26 02:34:55\ntrack2_ssm,aegis_hybrid,attention,roberta,sst2,PASS,,32.32,55.99,88.31,0.0,1.0,7063970,7063970,0.979,1852.4,0.0,0.6935,0.0,0.0,0.0,0.1432,\"{'loss': [0.6947, 0.6938, 0.6929, 0.6923, 0.6936], 'grad_norm': [0.1467, 0.1415, 0.1411, 0.1466, 0.14], 'fwd_latency_ms': [31.4852, 32.3999, 32.9334, 32.3546, 32.4164], 'bwd_latency_ms': [55.7036, 55.5543, 57.2099, 55.7347, 55.7549]}\",2026-08-26 02:34:56\ntrack2_ssm,aegis_hybrid,attention,roberta,mrpc,PASS,,32.28,55.65,87.93,0.0,1.0,7063970,7063970,0.979,1852.4,0.0,0.6943,0.0,0.0,0.0,1.1858,\"{'loss': [0.6943, 0.6939, 0.6943, 0.6933, 0.6959], 'grad_norm': [1.1791, 1.1804, 1.1909, 1.187, 1.1917], 'fwd_latency_ms': [31.0692, 32.4472, 32.5595, 32.4316, 32.8852], 'bwd_latency_ms': [55.8939, 55.2337, 55.8669, 55.7981, 55.463]}\",2026-08-26 02:34:57\ntrack2_ssm,aegis_hybrid,attention,t5,sst2,PASS,,32.78,56.9,89.68,0.0,1.0,4658208,4658208,0.967,1803.1,0.0,0.6934,0.0,0.0,0.0,0.0697,\"{'loss': [0.6936, 0.6934, 0.6935, 0.6931, 0.6935], 'grad_norm': [0.069, 0.07, 0.0707, 0.0696, 0.0692], 'fwd_latency_ms': [30.6022, 33.4446, 33.5239, 33.2484, 33.0625], 'bwd_latency_ms': [57.1481, 57.226, 57.0772, 56.0996, 56.9449]}\",2026-08-26 02:34:58\ntrack2_ssm,aegis_hybrid,attention,t5,mrpc,PASS,,32.5,56.15,88.65,0.0,1.0,4658208,4658208,0.967,1803.1,0.0,0.6729,0.0,0.0,0.0,2.2657,\"{'loss': [0.6731, 0.6723, 0.6728, 0.6733, 0.6729], 'grad_norm': [2.2628, 2.2652, 2.2688, 2.273, 2.2585], 'fwd_latency_ms': [30.1681, 33.0696, 33.0331, 32.9575, 33.2786], 'bwd_latency_ms': [56.3671, 55.8171, 56.1999, 56.5008, 55.8424]}\",2026-08-26 02:34:59",
"results/master_aggregated_results.csv": "aegis_hybrid,0,PASS,78.44,50.92,51.03,85.9,2026-09-11T03:48:38.322061",
"results/master_benchmark_summary.csv": "aegis_hybrid,wikitext2,O(N^2) Softmax,11.25,2.4204,54700,4864.6,9.36,52.351,386.604,0.03,SUCCESS\naegis_hybrid,c4,O(N^2) Softmax,11.25,2.4204,54700,4864.6,9.36,52.351,386.604,0.03,SUCCESS",
"results/run_quality_all_screening_summary.csv": "run_quality_20260827_052444,aegis_hybrid,DISTILBERT,SST2,PASS,50.92,0.0,50.92,50.92,1.0,0.6748,0.2224,226.42,588.85,4822.6,20096.3,1450000000.0,2026-08-27T05:48:10.709368\nrun_quality_20260827_052444,aegis_hybrid_fn,DISTILBERT,SST2,PASS,50.92,0.0,50.92,50.92,1.0,0.6748,0.2224,226.98,589.64,4822.6,20063.2,1450000000.0,2026-08-27T05:50:11.438281"
},
"notes": []
},
"aegis_spectron_fp8": {
"module_name": "aegis_spectron_fp8",
"source_file": "Codes/framework/attentions/aegis_spectron_fp8.py",
"source_lines_count": 1022,
"architecture": {
"class_name": "FactorizedLinear",
"base_class": "nn.Module",
"init_args": "self, **kwargs",
"registry_names": [
"inject_aegis_spectron_fp8",
"aegis_spectron_fp8_attn",
"aegis_spectron_fp8_inject",
"aegis_spectron_fp8"
]
},
"params": [
{
"name": "lin1",
"type": "nn.Linear",
"declaration": "nn.Linear(in_features, self.rank, bias=False, **factory_kwargs)"
},
{
"name": "lin2",
"type": "nn.Linear",
"declaration": "nn.Linear(self.rank, out_features, bias=bias, **factory_kwargs)"
},
{
"name": "h_proj",
"type": "nn.Linear",
"declaration": "nn.Linear(config.hidden_size, 3)"
},
{
"name": "linear_A",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.ones(self.all_head_size)"
},
{
"name": "linear_B",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.ones(self.all_head_size)"
},
{
"name": "linear_C",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.ones(self.all_head_size)"
},
{
"name": "decay_param",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.full((self.all_head_size,)"
},
{
"name": "bidir_gate",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.zeros(self.all_head_size)"
},
{
"name": "linear_in",
"type": "nn.Linear",
"declaration": "nn.Linear(self.hidden_size, self.all_head_size, bias=False)"
},
{
"name": "svd_U",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.randn(self.all_head_size, config.svd_rank)"
}
],
"metrics": {
"results/unified_aegis_spectron_fp8/roberta/roberta-base_full_full_metrics.json": {
"model": "roberta-base_full",
"performance": {
"accuracy_pct": 78.56,
"f1_score": 0.8124,
"f1_macro": 0.7811,
"precision": 0.8051,
"recall": 0.7832,
"val_loss": 0.4381,
"status": "PASS",
"elapsed_s": 6.95
},
"efficiency": {
"total_params_before": 124647170,
"total_params_after": 124647170,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 712.34,
"tokens_per_sec": 23592.51
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 683.27,
"L_256_vram_mb": 685.84,
"L_512_vram_mb": 699.47,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 514)"
},
"timestamp": "2026-09-15T10:07:41.486158"
},
"results/unified_aegis_spectron_fp8/roberta/roberta-base_aegis_spectron_fp8_full_metrics.json": {
"model": "roberta-base_aegis_spectron_fp8",
"performance": {
"accuracy_pct": 49.43,
"f1_score": 0.3815,
"f1_macro": 0.4769,
"precision": 0.4974,
"recall": 0.4978,
"val_loss": 0.7801,
"status": "PASS_LOW",
"elapsed_s": 59.05
},
"efficiency": {
"total_params_before": 124647170,
"total_params_after": 107330390,
"trainable_recovery_params": 1520688,
"compression_rate": 1.1613,
"compression_ratio_pct": 13.89,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 12,
"lora_params": 1179648,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {
"roberta.encoder.layer.0.attention.self": 32,
"roberta.encoder.layer.0.attention.self._wrapped_attention.q_proj": 32,
"roberta.encoder.layer.0.attention.self._wrapped_attention.k_proj": 32,
"roberta.encoder.layer.0.attention.self._wrapped_attention.v_proj": 32,
"roberta.encoder.layer.0.attention.self._wrapped_attention.out_proj": 32,
"roberta.encoder.layer.1.attention.self": 32,
"roberta.encoder.layer.1.attention.self._wrapped_attention.q_proj": 32,
"roberta.encoder.layer.1.attention.self._wrapped_attention.k_proj": 32,
"roberta.encoder.layer.1.attention.self._wrapped_attention.v_proj": 32,
"roberta.encoder.layer.1.attention.self._wrapped_attention.out_proj": 32,
"roberta.encoder.layer.2.attention.self": 32,
"roberta.encoder.layer.2.attention.self._wrapped_attention.q_proj": 32,
"roberta.encoder.layer.2.attention.self._wrapped_attention.k_proj": 32,
"roberta.encoder.layer.2.attention.self._wrapped_attention.v_proj": 32,
"roberta.encoder.layer.2.attention.self._wrapped_attention.out_proj": 32,
"roberta.encoder.layer.3.attention.self": 32,
"roberta.encoder.layer.3.attention.self._wrapped_attention.q_proj": 32,
"roberta.encoder.layer.3.attention.self._wrapped_attention.k_proj": 32,
"roberta.encoder.layer.3.attention.self._wrapped_attention.v_proj": 32,
"roberta.encoder.layer.3.attention.self._wrapped_attention.out_proj": 32,
"roberta.encoder.layer.4.attention.self": 32,
"roberta.encoder.layer.4.attention.self._wrapped_attention.q_proj": 32,
"roberta.encoder.layer.4.attention.self._wrapped_attention.k_proj": 32,
"roberta.encoder.layer.4.attention.self._wrapped_attention.v_proj": 32,
"roberta.encoder.layer.4.attention.self._wrapped_attention.out_proj": 32,
"roberta.encoder.layer.5.attention.self": 32,
"roberta.encoder.layer.5.attention.self._wrapped_attention.q_proj": 32,
"roberta.encoder.layer.5.attention.self._wrapped_attention.k_proj": 32,
"roberta.encoder.layer.5.attention.self._wrapped_attention.v_proj": 32,
"roberta.encoder.layer.5.attention.self._wrapped_attention.out_proj": 32,
"roberta.encoder.layer.6.attention.self": 32,
"roberta.encoder.layer.6.attention.self._wrapped_attention.q_proj": 32,
"roberta.encoder.layer.6.attention.self._wrapped_attention.k_proj": 32,
"roberta.encoder.layer.6.attention.self._wrapped_attention.v_proj": 32,
"roberta.encoder.layer.6.attention.self._wrapped_attention.out_proj": 32,
"roberta.encoder.layer.7.attention.self": 32,
"roberta.encoder.layer.7.attention.self._wrapped_attention.q_proj": 32,
"roberta.encoder.layer.7.attention.self._wrapped_attention.k_proj": 32,
"roberta.encoder.layer.7.attention.self._wrapped_attention.v_proj": 32,
"roberta.encoder.layer.7.attention.self._wrapped_attention.out_proj": 32,
"roberta.encoder.layer.8.attention.self": 32,
"roberta.encoder.layer.8.attention.self._wrapped_attention.q_proj": 32,
"roberta.encoder.layer.8.attention.self._wrapped_attention.k_proj": 32,
"roberta.encoder.layer.8.attention.self._wrapped_attention.v_proj": 32,
"roberta.encoder.layer.8.attention.self._wrapped_attention.out_proj": 32,
"roberta.encoder.layer.9.attention.self": 32,
"roberta.encoder.layer.9.attention.self._wrapped_attention.q_proj": 32,
"roberta.encoder.layer.9.attention.self._wrapped_attention.k_proj": 32,
"roberta.encoder.layer.9.attention.self._wrapped_attention.v_proj": 32,
"roberta.encoder.layer.9.attention.self._wrapped_attention.out_proj": 32,
"roberta.encoder.layer.10.attention.self": 32,
"roberta.encoder.layer.10.attention.self._wrapped_attention.q_proj": 32,
"roberta.encoder.layer.10.attention.self._wrapped_attention.k_proj": 32,
"roberta.encoder.layer.10.attention.self._wrapped_attention.v_proj": 32,
"roberta.encoder.layer.10.attention.self._wrapped_attention.out_proj": 32,
"roberta.encoder.layer.11.attention.self": 32,
"roberta.encoder.layer.11.attention.self._wrapped_attention.q_proj": 32,
"roberta.encoder.layer.11.attention.self._wrapped_attention.k_proj": 32,
"roberta.encoder.layer.11.attention.self._wrapped_attention.v_proj": 32,
"roberta.encoder.layer.11.attention.self._wrapped_attention.out_proj": 32
}
},
"hardware": {
"peak_vram_mb": 572.68,
"tokens_per_sec": 4840.56
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 536.4,
"L_256_vram_mb": 540.68,
"L_512_vram_mb": 569.26,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 514)"
},
"timestamp": "2026-09-15T10:08:40.656794"
},
"results/unified_aegis_spectron_fp8/distilbert/distilbert_sst2_aegis_spectron_fp8_full_metrics.json": {
"model": "distilbert_sst2_aegis_spectron_fp8",
"performance": {
"accuracy_pct": 79.59,
"f1_score": 0.8118,
"f1_macro": 0.7944,
"precision": 0.8014,
"recall": 0.7946,
"val_loss": 0.4587,
"status": "PASS",
"elapsed_s": 28.87
},
"efficiency": {
"total_params_before": 66955010,
"total_params_after": 54753068,
"trainable_recovery_params": 760344,
"compression_rate": 1.2229,
"compression_ratio_pct": 18.22,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 6,
"lora_params": 589824,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {
"distilbert.transformer.layer.0.attention": 32,
"distilbert.transformer.layer.0.attention._wrapped_attention.q_proj": 32,
"distilbert.transformer.layer.0.attention._wrapped_attention.k_proj": 32,
"distilbert.transformer.layer.0.attention._wrapped_attention.v_proj": 32,
"distilbert.transformer.layer.0.attention._wrapped_attention.out_proj": 32,
"distilbert.transformer.layer.1.attention": 32,
"distilbert.transformer.layer.1.attention._wrapped_attention.q_proj": 32,
"distilbert.transformer.layer.1.attention._wrapped_attention.k_proj": 32,
"distilbert.transformer.layer.1.attention._wrapped_attention.v_proj": 32,
"distilbert.transformer.layer.1.attention._wrapped_attention.out_proj": 32,
"distilbert.transformer.layer.2.attention": 32,
"distilbert.transformer.layer.2.attention._wrapped_attention.q_proj": 32,
"distilbert.transformer.layer.2.attention._wrapped_attention.k_proj": 32,
"distilbert.transformer.layer.2.attention._wrapped_attention.v_proj": 32,
"distilbert.transformer.layer.2.attention._wrapped_attention.out_proj": 32,
"distilbert.transformer.layer.3.attention": 32,
"distilbert.transformer.layer.3.attention._wrapped_attention.q_proj": 32,
"distilbert.transformer.layer.3.attention._wrapped_attention.k_proj": 32,
"distilbert.transformer.layer.3.attention._wrapped_attention.v_proj": 32,
"distilbert.transformer.layer.3.attention._wrapped_attention.out_proj": 32,
"distilbert.transformer.layer.4.attention": 32,
"distilbert.transformer.layer.4.attention._wrapped_attention.q_proj": 32,
"distilbert.transformer.layer.4.attention._wrapped_attention.k_proj": 32,
"distilbert.transformer.layer.4.attention._wrapped_attention.v_proj": 32,
"distilbert.transformer.layer.4.attention._wrapped_attention.out_proj": 32,
"distilbert.transformer.layer.5.attention": 32,
"distilbert.transformer.layer.5.attention._wrapped_attention.q_proj": 32,
"distilbert.transformer.layer.5.attention._wrapped_attention.k_proj": 32,
"distilbert.transformer.layer.5.attention._wrapped_attention.v_proj": 32,
"distilbert.transformer.layer.5.attention._wrapped_attention.out_proj": 32
}
},
"hardware": {
"peak_vram_mb": 290.29,
"tokens_per_sec": 9858.72
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 250.76,
"L_256_vram_mb": 255.04,
"L_512_vram_mb": "unsupported (exceeds model max_position_embeddings 512)",
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 512)"
},
"timestamp": "2026-09-15T10:06:57.548431"
},
"results/unified_aegis_spectron_fp8/distilbert/distilbert_sst2_full_full_metrics.json": {
"model": "distilbert_sst2_full",
"performance": {
"accuracy_pct": 90.71,
"f1_score": 0.9099,
"f1_macro": 0.907,
"precision": 0.9075,
"recall": 0.9068,
"val_loss": 0.3027,
"status": "PASS",
"elapsed_s": 3.76
},
"efficiency": {
"total_params_before": 66955010,
"total_params_after": 66955010,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 389.51,
"tokens_per_sec": 39628.79
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 371.82,
"L_256_vram_mb": 373.89,
"L_512_vram_mb": "unsupported (exceeds model max_position_embeddings 512)",
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 512)"
},
"timestamp": "2026-09-15T10:06:28.464328"
},
"results/unified_aegis_spectron_fp8/gpt2/gpt2_aegis_spectron_fp8_full_metrics.json": {
"model": "gpt2_aegis_spectron_fp8",
"performance": {
"accuracy_pct": 58.83,
"f1_score": 0.6913,
"f1_macro": 0.5367,
"precision": 0.6423,
"recall": 0.5824,
"val_loss": 0.6673,
"status": "PASS",
"elapsed_s": 55.53
},
"efficiency": {
"total_params_before": 124441344,
"total_params_after": 100037460,
"trainable_recovery_params": 1520688,
"compression_rate": 1.2439,
"compression_ratio_pct": 19.61,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 12,
"lora_params": 1179648,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {
"transformer.h.0.attn": 32,
"transformer.h.0.attn._wrapped_attention.q_proj": 32,
"transformer.h.0.attn._wrapped_attention.k_proj": 32,
"transformer.h.0.attn._wrapped_attention.v_proj": 32,
"transformer.h.0.attn._wrapped_attention.out_proj": 32,
"transformer.h.1.attn": 32,
"transformer.h.1.attn._wrapped_attention.q_proj": 32,
"transformer.h.1.attn._wrapped_attention.k_proj": 32,
"transformer.h.1.attn._wrapped_attention.v_proj": 32,
"transformer.h.1.attn._wrapped_attention.out_proj": 32,
"transformer.h.2.attn": 32,
"transformer.h.2.attn._wrapped_attention.q_proj": 32,
"transformer.h.2.attn._wrapped_attention.k_proj": 32,
"transformer.h.2.attn._wrapped_attention.v_proj": 32,
"transformer.h.2.attn._wrapped_attention.out_proj": 32,
"transformer.h.3.attn": 32,
"transformer.h.3.attn._wrapped_attention.q_proj": 32,
"transformer.h.3.attn._wrapped_attention.k_proj": 32,
"transformer.h.3.attn._wrapped_attention.v_proj": 32,
"transformer.h.3.attn._wrapped_attention.out_proj": 32,
"transformer.h.4.attn": 32,
"transformer.h.4.attn._wrapped_attention.q_proj": 32,
"transformer.h.4.attn._wrapped_attention.k_proj": 32,
"transformer.h.4.attn._wrapped_attention.v_proj": 32,
"transformer.h.4.attn._wrapped_attention.out_proj": 32,
"transformer.h.5.attn": 32,
"transformer.h.5.attn._wrapped_attention.q_proj": 32,
"transformer.h.5.attn._wrapped_attention.k_proj": 32,
"transformer.h.5.attn._wrapped_attention.v_proj": 32,
"transformer.h.5.attn._wrapped_attention.out_proj": 32,
"transformer.h.6.attn": 32,
"transformer.h.6.attn._wrapped_attention.q_proj": 32,
"transformer.h.6.attn._wrapped_attention.k_proj": 32,
"transformer.h.6.attn._wrapped_attention.v_proj": 32,
"transformer.h.6.attn._wrapped_attention.out_proj": 32,
"transformer.h.7.attn": 32,
"transformer.h.7.attn._wrapped_attention.q_proj": 32,
"transformer.h.7.attn._wrapped_attention.k_proj": 32,
"transformer.h.7.attn._wrapped_attention.v_proj": 32,
"transformer.h.7.attn._wrapped_attention.out_proj": 32,
"transformer.h.8.attn": 32,
"transformer.h.8.attn._wrapped_attention.q_proj": 32,
"transformer.h.8.attn._wrapped_attention.k_proj": 32,
"transformer.h.8.attn._wrapped_attention.v_proj": 32,
"transformer.h.8.attn._wrapped_attention.out_proj": 32,
"transformer.h.9.attn": 32,
"transformer.h.9.attn._wrapped_attention.q_proj": 32,
"transformer.h.9.attn._wrapped_attention.k_proj": 32,
"transformer.h.9.attn._wrapped_attention.v_proj": 32,
"transformer.h.9.attn._wrapped_attention.out_proj": 32,
"transformer.h.10.attn": 32,
"transformer.h.10.attn._wrapped_attention.q_proj": 32,
"transformer.h.10.attn._wrapped_attention.k_proj": 32,
"transformer.h.10.attn._wrapped_attention.v_proj": 32,
"transformer.h.10.attn._wrapped_attention.out_proj": 32,
"transformer.h.11.attn": 32,
"transformer.h.11.attn._wrapped_attention.q_proj": 32,
"transformer.h.11.attn._wrapped_attention.k_proj": 32,
"transformer.h.11.attn._wrapped_attention.v_proj": 32,
"transformer.h.11.attn._wrapped_attention.out_proj": 32
}
},
"hardware": {
"peak_vram_mb": 643.49,
"tokens_per_sec": 5217.46
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 447.49,
"L_256_vram_mb": 475.51,
"L_512_vram_mb": 535.55,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 1024)"
},
"timestamp": "2026-09-15T10:10:21.484788"
},
"results/unified_aegis_spectron_fp8/gpt2/gpt2_full_full_metrics.json": {
"model": "gpt2_full",
"performance": {
"accuracy_pct": 79.36,
"f1_score": 0.7978,
"f1_macro": 0.7935,
"precision": 0.7935,
"recall": 0.7935,
"val_loss": 0.4875,
"status": "PASS",
"elapsed_s": 6.98
},
"efficiency": {
"total_params_before": 124441344,
"total_params_after": 124441344,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 793.19,
"tokens_per_sec": 23426.91
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 681.07,
"L_256_vram_mb": 691.32,
"L_512_vram_mb": 721.95,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 1024)"
},
"timestamp": "2026-09-15T10:09:25.767434"
},
"results/unified_aegis_spectron_fp8/gpt2/hf_export/metrics.json": {
"model_name": "gpt2",
"module": "aegis_spectron_fp8",
"status": "PASS",
"accuracy_pct": 58.83,
"f1": 0.0,
"precision": 0.6423,
"recall": 0.5824,
"eval_loss": 0.0,
"params_total": null,
"params_active": null,
"param_reduction_pct": null,
"peak_vram_mb": 643.49,
"tokens_per_sec": 5217.46,
"elapsed_seconds": null,
"full_metrics": {
"module_name": "aegis_spectron_fp8",
"status": "PASS",
"accuracy_pct": 58.83,
"f1_score": 0.6913,
"f1_macro": 0.5367,
"precision": 0.6423,
"recall": 0.5824,
"val_loss": 0.6673,
"elapsed_s": 55.53,
"peak_vram_mb": 643.49,
"tokens_per_sec": 5217.46,
"total_params_before": 124441344,
"total_params_after": 100037460,
"trainable_recovery_params": 1520688,
"compression_rate": 1.2439,
"compression_ratio_pct": 19.61,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"avg_reconstruction_error": NaN,
"weighted_reconstruction_error": NaN,
"num_injected_attention_layers": 12,
"lora_params": 1179648,
"error": NaN,
"path": "adapted_gpt",
"timestamp": "2026-09-15T10:10:22.408494"
}
},
"results/unified_aegis_spectron_fp8/distilbert/hf_export/metrics.json": {
"model_name": "distilbert",
"module": "aegis_spectron_fp8",
"status": "PASS",
"accuracy_pct": 79.59,
"f1": 0.0,
"precision": 0.8014,
"recall": 0.7946,
"eval_loss": 0.0,
"params_total": null,
"params_active": null,
"param_reduction_pct": null,
"peak_vram_mb": 290.29,
"tokens_per_sec": 9858.72,
"elapsed_seconds": null,
"full_metrics": {
"module_name": "aegis_spectron_fp8",
"status": "PASS",
"accuracy_pct": 79.59,
"f1_score": 0.8118,
"f1_macro": 0.7944,
"precision": 0.8014,
"recall": 0.7946,
"val_loss": 0.4587,
"elapsed_s": 28.87,
"peak_vram_mb": 290.29,
"tokens_per_sec": 9858.72,
"total_params_before": 66955010,
"total_params_after": 54753068,
"trainable_recovery_params": 760344,
"compression_rate": 1.2229,
"compression_ratio_pct": 18.22,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"avg_reconstruction_error": NaN,
"weighted_reconstruction_error": NaN,
"num_injected_attention_layers": 6,
"lora_params": 589824,
"error": NaN,
"path": "legacy_bert",
"timestamp": "2026-09-15T10:06:57.863839"
}
},
"results/unified_aegis_spectron_fp8/roberta/summary.csv": "module_name,status,accuracy_pct,f1_score,f1_macro,precision,recall,val_loss,elapsed_s,peak_vram_mb,tokens_per_sec,total_params_before,total_params_after,trainable_recovery_params,compression_rate,compression_ratio_pct,flops_reduction_pct,num_compressed_layers,num_fallback_layers,avg_reconstruction_error,weighted_reconstruction_error,num_injected_attention_layers,lora_params,error,path,timestamp\nfull,PASS,78.56,0.8124,0.7811,0.8051,0.7832,0.4381,6.95,712.34,23592.51,124647170,124647170,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_roberta,2026-09-15T10:07:42.297917\naegis_spectron_fp8,PASS_LOW,49.43,0.3815,0.4769,0.4974,0.4978,0.7801,59.05,572.68,4840.56,124647170,107330390,1520688,1.1613,13.89,0.0,0,0,,,12,1179648,note: evaluated accuracy (49.43%) below chance threshold (56.00%),adapted_roberta,2026-09-15T10:08:41.630327\n",
"results/unified_aegis_spectron_fp8/distilbert/summary.csv": "module_name,status,accuracy_pct,f1_score,f1_macro,precision,recall,val_loss,elapsed_s,peak_vram_mb,tokens_per_sec,total_params_before,total_params_after,trainable_recovery_params,compression_rate,compression_ratio_pct,flops_reduction_pct,num_compressed_layers,num_fallback_layers,avg_reconstruction_error,weighted_reconstruction_error,num_injected_attention_layers,lora_params,error,path,timestamp\nfull,PASS,90.71,0.9099,0.907,0.9075,0.9068,0.3027,3.76,389.51,39628.79,66955010,66955010,0,1.0,0.0,0.0,0,0,,,0,0,,legacy_bert,2026-09-15T10:06:28.747016\naegis_spectron_fp8,PASS,79.59,0.8118,0.7944,0.8014,0.7946,0.4587,28.87,290.29,9858.72,66955010,54753068,760344,1.2229,18.22,0.0,0,0,,,6,589824,,legacy_bert,2026-09-15T10:06:57.863839\n",
"results/unified_aegis_spectron_fp8/gpt2/summary.csv": "module_name,status,accuracy_pct,f1_score,f1_macro,precision,recall,val_loss,elapsed_s,peak_vram_mb,tokens_per_sec,total_params_before,total_params_after,trainable_recovery_params,compression_rate,compression_ratio_pct,flops_reduction_pct,num_compressed_layers,num_fallback_layers,avg_reconstruction_error,weighted_reconstruction_error,num_injected_attention_layers,lora_params,error,path,timestamp\nfull,PASS,79.36,0.7978,0.7935,0.7935,0.7935,0.4875,6.98,793.19,23426.91,124441344,124441344,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_gpt,2026-09-15T10:09:26.610942\naegis_spectron_fp8,PASS,58.83,0.6913,0.5367,0.6423,0.5824,0.6673,55.53,643.49,5217.46,124441344,100037460,1520688,1.2439,19.61,0.0,0,0,,,12,1179648,,adapted_gpt,2026-09-15T10:10:22.408494\n",
"results/benchmark_summary_v2.csv": "track2_ssm,aegis_spectron_fp8,attention,bert,sst2,PASS,,2603.37,1276.67,3880.04,0.0,1.0,4587406,4587406,0.956,1331.6,0.0,0.6927,0.0,0.0,0.0,0.1565,\"{'loss': [0.6927, 0.6928, 0.6927, 0.6927, 0.6927], 'grad_norm': [0.1565, 0.1566, 0.1563, 0.1563, 0.1568], 'fwd_latency_ms': [2423.6453, 2681.7275, 2631.635, 2629.9436, 2649.9097], 'bwd_latency_ms': [1268.8876, 1267.9729, 1271.4896, 1276.6478, 1298.3496]}\",2026-08-26 02:35:00\ntrack2_ssm,aegis_spectron_fp8,attention,bert,mrpc,PASS,,2592.19,1235.97,3828.16,0.0,1.0,4587406,4587406,0.956,1331.6,0.0,0.6996,0.0,0.0,0.0,1.4107,\"{'loss': [0.6996, 0.6996, 0.6996, 0.6995, 0.6996], 'grad_norm': [1.4104, 1.4107, 1.4111, 1.4103, 1.4111], 'fwd_latency_ms': [2620.7395, 2642.0186, 2625.427, 2632.8574, 2439.8857], 'bwd_latency_ms': [1289.7523, 1266.6698, 1276.2734, 1190.345, 1156.8339]}\",2026-08-26 02:35:27\ntrack2_ssm,aegis_spectron_fp8,attention,roberta,sst2,PASS,,2441.38,1224.98,3666.36,0.0,1.0,7114510,7114510,0.972,1341.5,0.0,0.6935,0.0,0.0,0.0,0.1401,\"{'loss': [0.6935, 0.6935, 0.6935, 0.6935, 0.6935], 'grad_norm': [0.1402, 0.1399, 0.1401, 0.14, 0.1402], 'fwd_latency_ms': [2520.8022, 2524.1799, 2329.4751, 2296.9319, 2535.4995], 'bwd_latency_ms': [1261.7517, 1265.3536, 1161.184, 1152.2848, 1284.3127]}\",2026-08-26 02:35:55\ntrack2_ssm,aegis_spectron_fp8,attention,roberta,mrpc,PASS,,2525.18,1227.28,3752.46,0.0,1.0,7114510,7114510,0.972,1341.5,0.0,0.6862,0.0,0.0,0.0,1.2398,\"{'loss': [0.6862, 0.6862, 0.6862, 0.6862, 0.6862], 'grad_norm': [1.2396, 1.2398, 1.2399, 1.24, 1.2395], 'fwd_latency_ms': [2535.4036, 2395.2366, 2489.4756, 2629.2451, 2576.5393], 'bwd_latency_ms': [1172.7189, 1152.8354, 1287.7964, 1268.8739, 1254.1802]}\",2026-08-26 02:36:21\ntrack2_ssm,aegis_spectron_fp8,attention,t5,sst2,PASS,,2737.53,1310.91,4048.44,0.0,1.0,4708748,4708748,0.957,1430.7,0.0,0.692,0.0,0.0,0.0,0.8909,\"{'loss': [0.692, 0.692, 0.692, 0.692, 0.692], 'grad_norm': [0.8908, 0.8907, 0.8914, 0.8905, 0.8909], 'fwd_latency_ms': [2557.5537, 2850.4609, 2785.6423, 2872.4587, 2621.5408], 'bwd_latency_ms': [1255.4827, 1343.2683, 1321.1548, 1406.4403, 1228.1903]}\",2026-08-26 02:36:48\ntrack2_ssm,aegis_spectron_fp8,attention,t5,mrpc,PASS,,3361.68,1210.73,4572.41,0.0,1.0,4708748,4708748,0.957,1430.7,0.0,0.7033,0.0,0.0,0.0,6.9328,\"{'loss': [0.7033, 0.7033, 0.7033, 0.7033, 0.7033], 'grad_norm': [6.9334, 6.9319, 6.9344, 6.9318, 6.9324], 'fwd_latency_ms': [3341.4114, 3357.3223, 3380.0303, 3371.2205, 3358.427], 'bwd_latency_ms': [1207.1083, 1208.8097, 1210.4514, 1209.3563, 1217.9451]}\",2026-08-26 02:37:16",
"results/master_aggregated_results.csv": "aegis_spectron_fp8,0,PASS_LOW,50.92,50.92,50.92,212.0,2026-09-11T10:59:57.016139",
"results/master_benchmark_summary.csv": "aegis_spectron_fp8,wikitext2,O(N^2) Softmax,11.25,2.4204,36511,2290.7,14.023,37.972,180.692,0.07,SUCCESS\naegis_spectron_fp8,c4,O(N^2) Softmax,11.25,2.4204,36511,2290.7,14.023,37.972,180.692,0.07,SUCCESS",
"results/run_quality_all_screening_summary.csv": "run_quality_20260827_052444,aegis_spectron_fp8,DISTILBERT,SST2,PASS,68.81,0.0,68.81,68.81,1.0,0.7318,0.6699,212.24,411.52,3669.2,6566.6,1450000000.0,2026-08-27T05:51:57.777707\nrun_quality_20260827_052444,aegis_spectron_fp8_fn,DISTILBERT,SST2,PASS,68.81,0.0,68.81,68.81,1.0,0.7318,0.6699,211.56,411.2,3669.2,6577.1,1450000000.0,2026-08-27T05:53:43.683998"
},
"notes": []
}
}
{
"adasvd_compressor": {
"module_name": "adasvd_compressor",
"source_file": "Codes/framework/compression_engine/adasvd_compressor.py",
"source_lines_count": 488,
"architecture": {
"class_name": "AdaSVDCompressor",
"base_class": "BaseCompressor",
"init_args": "self, model: nn.Module, rank: Optional[int] = None, target_param_ratio: float = 0.7, num_iters: int = 5, max_error: float = 0.40, fallback_to_original: bool = True, target_attention_only: bool = True, min_rank: int = 16, target_modules: Optional[List[str]] = None, comp_config: Any = None, **kwargs: Any, ) -> None: super().__init__(model, name=\"adasvd\", **kwargs) cfg = comp_config or kwargs.get(\"comp_config\", kwargs.get(\"cfg\", getattr(model, \"config\", None))) self.rank = _cfg_get(cfg, \"adasvd_rank\", _cfg_get(cfg, \"rank\", rank)) self.target_param_ratio = float( _cfg_get(cfg, \"adasvd_target_ratio\", _cfg_get(cfg, \"target_param_ratio\", target_param_ratio)) ) self.num_iters = int(_cfg_get(cfg, \"adasvd_iters\", _cfg_get(cfg, \"num_iters\", num_iters))) self.max_error = float(_cfg_get(cfg, \"adasvd_max_error\", _cfg_get(cfg, \"max_error\", max_error))) self.fallback_to_original = bool( _cfg_get(cfg, \"adasvd_fallback_to_original\", _cfg_get(cfg, \"fallback_to_original\", fallback_to_original)) ) self.target_attention_only = bool( _cfg_get(cfg, \"adasvd_target_attention_only\", _cfg_get(cfg, \"target_attention_only\", target_attention_only)) ) self.min_rank = int(_cfg_get(cfg, \"adasvd_min_rank\", _cfg_get(cfg, \"min_rank\", min_rank))) self.target_modules = target_modules self.comp_config = cfg self.compressed_layers: List[str] = [] self.fallback_layers: List[str] = [] self.layer_errors: Dict[str, float] = {} self.layer_ranks: Dict[str, int] = {} self.layer_iters: Dict[str, int] = {} def _is_target_module(self, name: str, module: nn.Module) -> bool: if isinstance(module, SVDLinear",
"registry_names": [
"adasvd"
]
},
"params": [],
"metrics": {
"results/unified_adasvd/roberta/roberta-base_full_full_metrics.json": {
"model": "roberta-base_full",
"performance": {
"accuracy_pct": 78.56,
"f1_score": 0.8124,
"f1_macro": 0.7811,
"precision": 0.8051,
"recall": 0.7832,
"val_loss": 0.4381,
"status": "PASS",
"elapsed_s": 6.77
},
"efficiency": {
"total_params_before": 124647170,
"total_params_after": 124647170,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 712.34,
"tokens_per_sec": 24228.27
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 683.27,
"L_256_vram_mb": 685.84,
"L_512_vram_mb": 699.47,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 514)"
},
"timestamp": "2026-09-15T09:56:11.500945"
},
"results/unified_adasvd/roberta/roberta-base_adasvd_full_metrics.json": {
"model": "roberta-base_adasvd",
"performance": {
"accuracy_pct": 73.39,
"f1_score": 0.7778,
"f1_macro": 0.7232,
"precision": 0.7685,
"recall": 0.7306,
"val_loss": 0.5381,
"status": "PASS",
"elapsed_s": 26.53
},
"efficiency": {
"total_params_before": 124647170,
"total_params_after": 109737258,
"trainable_recovery_params": 8713768,
"compression_rate": 1.1359,
"compression_ratio_pct": 11.96,
"flops_reduction_pct": 63.2,
"num_compressed_layers": 40,
"num_fallback_layers": 8,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": 0.5913941559039471,
"weighted_reconstruction_error": 0.6364,
"per_layer_reconstruction_errors": {
"roberta.encoder.layer.0.attention.self.query": 0.455972929347947,
"roberta.encoder.layer.0.attention.self.key": 0.4377989629219472,
"roberta.encoder.layer.0.attention.self.value": 0.6528715799546899,
"roberta.encoder.layer.0.attention.output.dense": 0.6181080091945617,
"roberta.encoder.layer.1.attention.self.query": 0.5507851348101068,
"roberta.encoder.layer.1.attention.self.key": 0.5445418526049741,
"roberta.encoder.layer.1.attention.self.value": 0.6285729613390068,
"roberta.encoder.layer.1.attention.output.dense": 0.6149361344300587,
"roberta.encoder.layer.2.attention.self.query": 0.5717057309786219,
"roberta.encoder.layer.2.attention.self.key": 0.5767926894779156,
"roberta.encoder.layer.2.attention.self.value": 0.6355413155578186,
"roberta.encoder.layer.2.attention.output.dense": 0.652313549394161,
"roberta.encoder.layer.3.attention.self.query": 0.5620059249830314,
"roberta.encoder.layer.3.attention.self.key": 0.5846928691061228,
Xet Storage Details
- Size:
- 160 kB
- Xet hash:
- df9c70b346d08442a71dc18cf334da1559a7402e9b01796658fcaf64bd67e81c
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.