{
"aegis_v1": {
"module_name": "aegis_v1",
"source_file": "Codes/framework/attentions/aegis_v1.py",
"source_lines_count": 814,
"architecture": {
"class_name": "FactorisedLinear",
"base_class": "nn.Module",
"init_args": "self, in_features_or_weight: Union[int, torch.Tensor], out_features_or_bias: Optional[Union[int, torch.Tensor]] = None, energy: float = 0.95, min_rank: int = 4, max_rank: int = 384, rank: Optional[int] = None, bias: bool = True, lora_rank: int = 16, lora_alpha: float = 32.0, **kwargs: Any, ) -> None: super().__init__() self.energy = energy self.min_rank = min_rank self.max_rank = max_rank self.lora_rank = lora_rank self.lora_alpha = lora_alpha self.lora_scaling = (lora_alpha / lora_rank) if lora_rank > 0 else 1.0 if isinstance(in_features_or_weight, torch.Tensor",
"registry_names": [
"aegis_v1_inject",
"AegisAttentionV1",
"aegis_v1_module",
"aegis_v1",
"inject_aegis_v1"
]
},
"params": [
{
"name": "down",
"type": "nn.Linear",
"declaration": "nn.Linear(in_dim, self.rank, bias=False, device=device, dtype=dtype)"
},
{
"name": "up",
"type": "nn.Linear",
"declaration": "nn.Linear(self.rank, out_dim, bias=(b is not None or bias)"
},
{
"name": "lora_A",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.randn(lora_rank, in_dim, device=device, dtype=dtype)"
},
{
"name": "lora_B",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.zeros(out_dim, lora_rank, device=device, dtype=dtype)"
},
{
"name": "down",
"type": "nn.Linear",
"declaration": "nn.Linear(in_dim, self.rank, bias=False)"
},
{
"name": "up",
"type": "nn.Linear",
"declaration": "nn.Linear(self.rank, out_dim, bias=bias)"
},
{
"name": "lora_A",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.randn(lora_rank, in_dim)"
},
{
"name": "lora_B",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.zeros(out_dim, lora_rank)"
}
],
"metrics": {
"results/unified_aegis_v1/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": 7.1
},
"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": 23473.61
},
"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": "2026-09-15T10:12:32.912988"
},
"results/unified_aegis_v1/roberta/roberta-base_aegis_spectron_fp8_full_metrics.json": {
"model": "roberta-base_aegis_spectron_fp8",
"performance": {
"accuracy_pct": 51.61,
"f1_score": 0.6744,
"f1_macro": 0.3662,
"precision": 0.5815,
"recall": 0.5073,
"val_loss": 0.7056,
"status": "PASS_LOW",
"elapsed_s": 47.8
},
"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": 570.54,
"tokens_per_sec": 6246.37
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 534.71,
"L_256_vram_mb": 538.79,
"L_512_vram_mb": 567.11,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 514)"
},
"timestamp": "2026-09-14T22:29:10.156343"
},
"results/unified_aegis_v1/roberta/roberta-base_aegis_v1_full_metrics.json": {
"model": "roberta-base_aegis_v1",
"performance": {
"accuracy_pct": 79.13,
"f1_score": 0.8131,
"f1_macro": 0.7884,
"precision": 0.8034,
"recall": 0.7894,
"val_loss": 0.5087,
"status": "PASS_NO_COMPRESSION",
"elapsed_s": 56.85
},
"efficiency": {
"total_params_before": 124647170,
"total_params_after": 133307330,
"trainable_recovery_params": 1573056,
"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": 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": 384,
"roberta.encoder.layer.0.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.0.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.0.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.1.attention.self": 32,
"roberta.encoder.layer.1.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.1.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.1.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.1.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.2.attention.self": 32,
"roberta.encoder.layer.2.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.2.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.2.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.2.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.3.attention.self": 32,
"roberta.encoder.layer.3.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.3.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.3.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.3.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.4.attention.self": 32,
"roberta.encoder.layer.4.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.4.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.4.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.4.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.5.attention.self": 32,
"roberta.encoder.layer.5.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.5.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.5.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.5.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.6.attention.self": 32,
"roberta.encoder.layer.6.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.6.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.6.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.6.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.7.attention.self": 32,
"roberta.encoder.layer.7.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.7.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.7.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.7.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.8.attention.self": 32,
"roberta.encoder.layer.8.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.8.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.8.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.8.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.9.attention.self": 32,
"roberta.encoder.layer.9.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.9.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.9.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.9.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.10.attention.self": 32,
"roberta.encoder.layer.10.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.10.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.10.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.10.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.11.attention.self": 32,
"roberta.encoder.layer.11.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.11.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.11.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.11.attention.self._wrapped_attention.out_proj": 384
}
},
"hardware": {
"peak_vram_mb": 674.79,
"tokens_per_sec": 6386.38
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 653.48,
"L_256_vram_mb": 656.29,
"L_512_vram_mb": 672.11,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 514)"
},
"timestamp": "2026-09-15T10:13:29.853707"
},
"results/unified_aegis_v1/distilbert/distilbert_sst2_aegis_v1_full_metrics.json": {
"model": "distilbert_sst2_aegis_v1",
"performance": {
"accuracy_pct": 61.81,
"f1_score": 0.7209,
"f1_macro": 0.5583,
"precision": 0.7301,
"recall": 0.6116,
"val_loss": 0.6411,
"status": "PASS_NO_COMPRESSION",
"elapsed_s": 26.77
},
"efficiency": {
"total_params_before": 66955010,
"total_params_after": 67741538,
"trainable_recovery_params": 786528,
"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": 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": 384,
"distilbert.transformer.layer.0.attention._wrapped_attention.k_proj": 384,
"distilbert.transformer.layer.0.attention._wrapped_attention.v_proj": 384,
"distilbert.transformer.layer.0.attention._wrapped_attention.out_proj": 384,
"distilbert.transformer.layer.1.attention": 32,
"distilbert.transformer.layer.1.attention._wrapped_attention.q_proj": 384,
"distilbert.transformer.layer.1.attention._wrapped_attention.k_proj": 384,
"distilbert.transformer.layer.1.attention._wrapped_attention.v_proj": 384,
"distilbert.transformer.layer.1.attention._wrapped_attention.out_proj": 384,
"distilbert.transformer.layer.2.attention": 32,
"distilbert.transformer.layer.2.attention._wrapped_attention.q_proj": 384,
"distilbert.transformer.layer.2.attention._wrapped_attention.k_proj": 384,
"distilbert.transformer.layer.2.attention._wrapped_attention.v_proj": 384,
"distilbert.transformer.layer.2.attention._wrapped_attention.out_proj": 384,
"distilbert.transformer.layer.3.attention": 32,
"distilbert.transformer.layer.3.attention._wrapped_attention.q_proj": 384,
"distilbert.transformer.layer.3.attention._wrapped_attention.k_proj": 384,
"distilbert.transformer.layer.3.attention._wrapped_attention.v_proj": 384,
"distilbert.transformer.layer.3.attention._wrapped_attention.out_proj": 384,
"distilbert.transformer.layer.4.attention": 32,
"distilbert.transformer.layer.4.attention._wrapped_attention.q_proj": 384,
"distilbert.transformer.layer.4.attention._wrapped_attention.k_proj": 384,
"distilbert.transformer.layer.4.attention._wrapped_attention.v_proj": 384,
"distilbert.transformer.layer.4.attention._wrapped_attention.out_proj": 384,
"distilbert.transformer.layer.5.attention": 32,
"distilbert.transformer.layer.5.attention._wrapped_attention.q_proj": 384,
"distilbert.transformer.layer.5.attention._wrapped_attention.k_proj": 384,
"distilbert.transformer.layer.5.attention._wrapped_attention.v_proj": 384,
"distilbert.transformer.layer.5.attention._wrapped_attention.out_proj": 384
}
},
"hardware": {
"peak_vram_mb": 329.48,
"tokens_per_sec": 12321.89
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 308.42,
"L_256_vram_mb": 311.24,
"L_512_vram_mb": "unsupported (exceeds model max_position_embeddings 512)",
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 512)"
},
"timestamp": "2026-09-15T10:11:50.349951"
},
"results/unified_aegis_v1/distilbert/distilbert_sst2_aegis_spectron_fp8_full_metrics.json": {
"model": "distilbert_sst2_aegis_spectron_fp8",
"performance": {
"accuracy_pct": 58.14,
"f1_score": 0.7045,
"f1_macro": 0.4937,
"precision": 0.7194,
"recall": 0.574,
"val_loss": 0.679,
"status": "PASS",
"elapsed_s": 25.24
},
"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": 288.99,
"tokens_per_sec": 11214.11
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 250.66,
"L_256_vram_mb": 254.73,
"L_512_vram_mb": "unsupported (exceeds model max_position_embeddings 512)",
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 512)"
},
"timestamp": "2026-09-14T22:26:28.136943"
},
"results/unified_aegis_v1/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": 3.77
},
"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": 39612.94
},
"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": "2026-09-15T10:11:23.414135"
},
"results/unified_aegis_v1/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": 6.89
},
"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": 23751.25
},
"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": "2026-09-15T10:14:15.733668"
},
"results/unified_aegis_v1/gpt2/gpt2_aegis_v1_full_metrics.json": {
"model": "gpt2_aegis_v1",
"performance": {
"accuracy_pct": 70.99,
"f1_score": 0.7668,
"f1_macro": 0.6915,
"precision": 0.7639,
"recall": 0.7056,
"val_loss": 0.54,
"status": "PASS_NO_COMPRESSION",
"elapsed_s": 38.41
},
"efficiency": {
"total_params_before": 124441344,
"total_params_after": 126014400,
"trainable_recovery_params": 1573056,
"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": 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": 384,
"transformer.h.0.attn._wrapped_attention.k_proj": 384,
"transformer.h.0.attn._wrapped_attention.v_proj": 384,
"transformer.h.0.attn._wrapped_attention.out_proj": 384,
"transformer.h.1.attn": 32,
"transformer.h.1.attn._wrapped_attention.q_proj": 384,
"transformer.h.1.attn._wrapped_attention.k_proj": 384,
"transformer.h.1.attn._wrapped_attention.v_proj": 384,
"transformer.h.1.attn._wrapped_attention.out_proj": 384,
"transformer.h.2.attn": 32,
"transformer.h.2.attn._wrapped_attention.q_proj": 384,
"transformer.h.2.attn._wrapped_attention.k_proj": 384,
"transformer.h.2.attn._wrapped_attention.v_proj": 384,
"transformer.h.2.attn._wrapped_attention.out_proj": 384,
"transformer.h.3.attn": 32,
"transformer.h.3.attn._wrapped_attention.q_proj": 384,
"transformer.h.3.attn._wrapped_attention.k_proj": 384,
"transformer.h.3.attn._wrapped_attention.v_proj": 384,
"transformer.h.3.attn._wrapped_attention.out_proj": 384,
"transformer.h.4.attn": 32,
"transformer.h.4.attn._wrapped_attention.q_proj": 384,
"transformer.h.4.attn._wrapped_attention.k_proj": 384,
"transformer.h.4.attn._wrapped_attention.v_proj": 384,
"transformer.h.4.attn._wrapped_attention.out_proj": 384,
"transformer.h.5.attn": 32,
"transformer.h.5.attn._wrapped_attention.q_proj": 384,
"transformer.h.5.attn._wrapped_attention.k_proj": 384,
"transformer.h.5.attn._wrapped_attention.v_proj": 384,
"transformer.h.5.attn._wrapped_attention.out_proj": 384,
"transformer.h.6.attn": 32,
"transformer.h.6.attn._wrapped_attention.q_proj": 384,
"transformer.h.6.attn._wrapped_attention.k_proj": 384,
"transformer.h.6.attn._wrapped_attention.v_proj": 384,
"transformer.h.6.attn._wrapped_attention.out_proj": 384,
"transformer.h.7.attn": 32,
"transformer.h.7.attn._wrapped_attention.q_proj": 384,
"transformer.h.7.attn._wrapped_attention.k_proj": 384,
"transformer.h.7.attn._wrapped_attention.v_proj": 384,
"transformer.h.7.attn._wrapped_attention.out_proj": 384,
"transformer.h.8.attn": 32,
"transformer.h.8.attn._wrapped_attention.q_proj": 384,
"transformer.h.8.attn._wrapped_attention.k_proj": 384,
"transformer.h.8.attn._wrapped_attention.v_proj": 384,
"transformer.h.8.attn._wrapped_attention.out_proj": 384,
"transformer.h.9.attn": 32,
"transformer.h.9.attn._wrapped_attention.q_proj": 384,
"transformer.h.9.attn._wrapped_attention.k_proj": 384,
"transformer.h.9.attn._wrapped_attention.v_proj": 384,
"transformer.h.9.attn._wrapped_attention.out_proj": 384,
"transformer.h.10.attn": 32,
"transformer.h.10.attn._wrapped_attention.q_proj": 384,
"transformer.h.10.attn._wrapped_attention.k_proj": 384,
"transformer.h.10.attn._wrapped_attention.v_proj": 384,
"transformer.h.10.attn._wrapped_attention.out_proj": 384,
"transformer.h.11.attn": 32,
"transformer.h.11.attn._wrapped_attention.q_proj": 384,
"transformer.h.11.attn._wrapped_attention.k_proj": 384,
"transformer.h.11.attn._wrapped_attention.v_proj": 384,
"transformer.h.11.attn._wrapped_attention.out_proj": 384
}
},
"hardware": {
"peak_vram_mb": 589.37,
"tokens_per_sec": 9887.94
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 547.69,
"L_256_vram_mb": 552.32,
"L_512_vram_mb": 580.03,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 1024)"
},
"timestamp": "2026-09-15T10:14:54.213996"
},
"results/unified_aegis_v1/gpt2/hf_export/metrics.json": {
"model_name": "gpt2",
"module": "aegis_v1",
"status": "PASS",
"accuracy_pct": 80.85,
"f1": 0.0,
"precision": 0.8361,
"recall": 0.8058,
"eval_loss": 0.0,
"params_total": null,
"params_active": null,
"param_reduction_pct": null,
"peak_vram_mb": 344.49,
"tokens_per_sec": 9004.41,
"elapsed_seconds": null,
"full_metrics": {
"module_name": "aegis_v1",
"status": "PASS",
"accuracy_pct": 80.85,
"f1_score": 0.8348,
"f1_macro": 0.8035,
"precision": 0.8361,
"recall": 0.8058,
"val_loss": 0.422,
"elapsed_s": 42.76,
"peak_vram_mb": 344.49,
"tokens_per_sec": 9004.41,
"total_params_before": 126800640,
"total_params_after": 127488960,
"trainable_recovery_params": 3047616,
"compression_rate": 0.9946,
"compression_ratio_pct": -0.54,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 72,
"num_fallback_layers": 0,
"avg_reconstruction_error": NaN,
"weighted_reconstruction_error": NaN,
"num_injected_attention_layers": 12,
"lora_params": 2654208,
"error": NaN,
"path": "adapted_gpt",
"timestamp": "2026-09-14T22:05:42.775719"
}
},
"results/unified_aegis_v1/distilbert/hf_export/metrics.json": {
"model_name": "distilbert",
"module": "aegis_v1",
"status": "PASS",
"accuracy_pct": 58.14,
"f1": 0.0,
"precision": 0.7194,
"recall": 0.574,
"eval_loss": 0.0,
"params_total": null,
"params_active": null,
"param_reduction_pct": null,
"peak_vram_mb": 288.99,
"tokens_per_sec": 11214.11,
"elapsed_seconds": null,
"full_metrics": {
"module_name": "aegis_spectron_fp8",
"status": "PASS",
"accuracy_pct": 58.14,
"f1_score": 0.7045,
"f1_macro": 0.4937,
"precision": 0.7194,
"recall": 0.574,
"val_loss": 0.679,
"elapsed_s": 25.24,
"peak_vram_mb": 288.99,
"tokens_per_sec": 11214.11,
"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": "2026-09-14T22:26:28.516678"
}
},
"results/unified_aegis_v1/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,78.56,0.8124,0.7811,0.8051,0.7832,0.4381,7.1,712.34,23473.61,124647170,124647170,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_roberta,2026-09-15T10:12:33.792283\naegis_v1,PASS_NO_COMPRESSION,79.13,0.8131,0.7884,0.8034,0.7894,0.5087,56.86,674.79,6386.38,124647170,133307330,1573056,1.0,0.0,0.0,0,0,,,12,1179648,\"note: zero compressed layers, model grew (LoRA only)\",adapted_roberta,2026-09-15T10:13:30.928640\naegis_spectron_fp8,PASS_LOW,51.61,0.6744,0.3662,0.5815,0.5073,0.7056,47.8,570.54,6246.37,124647170,107330390,1520688,1.1613,13.89,0.0,0,0,,,12,1179648,note: evaluated accuracy (51.61%) below chance threshold (56.00%),adapted_roberta,2026-09-14T22:29:11.071570\n",
"results/unified_aegis_v1/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,90.71,0.9099,0.907,0.9075,0.9068,0.3027,3.77,389.51,39612.94,66955010,66955010,0,1.0,0.0,0.0,0,0,,,0,0,,legacy_bert,2026-09-15T10:11:23.675362\naegis_v1,PASS_NO_COMPRESSION,61.81,0.7209,0.5583,0.7301,0.6116,0.6411,26.77,329.48,12321.89,66955010,67741538,786528,1.0,0.0,0.0,0,0,,,6,589824,\"note: zero compressed layers, model grew (LoRA only)\",legacy_bert,2026-09-15T10:11:50.689740\naegis_spectron_fp8,PASS,58.14,0.7045,0.4937,0.7194,0.574,0.679,25.24,288.99,11214.11,66955010,54753068,760344,1.2229,18.22,0.0,0,0,,,6,589824,,legacy_bert,2026-09-14T22:26:28.516678\n",
"results/unified_aegis_v1/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,79.36,0.7978,0.7935,0.7935,0.7935,0.4875,6.9,793.19,23751.25,124441344,124441344,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_gpt,2026-09-15T10:14:16.524127\naegis_v1,PASS_NO_COMPRESSION,70.99,0.7668,0.6915,0.7639,0.7056,0.54,38.41,589.37,9887.94,124441344,126014400,1573056,1.0,0.0,0.0,0,0,,,12,1179648,\"note: zero compressed layers, model grew (LoRA only)\",adapted_gpt,2026-09-15T10:14:55.212237\n",
"results/matrix_test_summary.csv": "qwen3_5,qwen3_5,aegis_v1,attention,PASS,True,True,45.451,5072.9,\ngpt2,gpt2,aegis_v1,attention,PASS,True,True,58.335,1440.7,\nroberta,roberta,aegis_v1,attention,PASS,True,True,56.309,1579.2,\ndistilbert,distilbert,aegis_v1,attention,PASS,True,True,43.797,804.7,",
"results/benchmark_summary_v2.csv": "track2_ssm,aegis_v1,attention,bert,sst2,PASS,,25.3,40.06,65.36,0.0,1.0,4550546,4550546,0.964,1099.6,0.0,0.6963,0.0,0.0,0.0,0.404,\"{'loss': [0.6963, 0.6963, 0.6963, 0.6963, 0.6963], 'grad_norm': [0.404, 0.404, 0.404, 0.404, 0.404], 'fwd_latency_ms': [24.2514, 24.5065, 25.5002, 27.437, 24.8278], 'bwd_latency_ms': [39.871, 39.2432, 42.0176, 39.251, 39.8941]}\",2026-08-26 02:34:15\ntrack2_ssm,aegis_v1,attention,bert,mrpc,PASS,,26.07,40.96,67.03,0.0,1.0,4550546,4550546,0.964,1099.6,0.0,0.7024,0.0,0.0,0.0,1.1639,\"{'loss': [0.7024, 0.7024, 0.7024, 0.7024, 0.7024], 'grad_norm': [1.1639, 1.1639, 1.1639, 1.1639, 1.1639], 'fwd_latency_ms': [24.9212, 28.057, 25.3309, 25.7002, 26.348], 'bwd_latency_ms': [41.0387, 44.9256, 39.5217, 39.5207, 39.8151]}\",2026-08-26 02:34:16\ntrack2_ssm,aegis_v1,attention,roberta,sst2,PASS,,26.01,41.61,67.62,0.0,1.0,7077650,7077650,0.977,1109.5,0.0,0.6928,0.0,0.0,0.0,0.1229,\"{'loss': [0.6928, 0.6928, 0.6928, 0.6928, 0.6928], 'grad_norm': [0.1229, 0.1229, 0.1229, 0.1229, 0.1229], 'fwd_latency_ms': [25.044, 25.386, 26.4434, 26.3856, 26.7935], 'bwd_latency_ms': [40.2117, 40.3475, 40.5797, 40.5937, 46.2944]}\",2026-08-26 02:34:17\ntrack2_ssm,aegis_v1,attention,roberta,mrpc,PASS,,26.14,40.07,66.21,0.0,1.0,7077650,7077650,0.977,1109.5,0.0,0.6924,0.0,0.0,0.0,1.1508,\"{'loss': [0.6924, 0.6924, 0.6924, 0.6924, 0.6924], 'grad_norm': [1.1508, 1.1508, 1.1508, 1.1508, 1.1508], 'fwd_latency_ms': [25.3624, 26.3541, 25.9335, 26.3053, 26.7531], 'bwd_latency_ms': [39.9478, 40.2262, 40.8625, 40.024, 39.2661]}\",2026-08-26 02:34:18\ntrack2_ssm,aegis_v1,attention,t5,sst2,PASS,,26.05,41.09,67.14,0.0,1.0,4671888,4671888,0.965,1161.6,0.0,0.8672,0.0,0.0,0.0,10.5924,\"{'loss': [0.8672, 0.8672, 0.8672, 0.8672, 0.8672], 'grad_norm': [10.5924, 10.5924, 10.5924, 10.5924, 10.5924], 'fwd_latency_ms': [23.9104, 26.1734, 27.3528, 26.0382, 26.7666], 'bwd_latency_ms': [41.0262, 41.1973, 41.2153, 41.016, 41.017]}\",2026-08-26 02:34:18\ntrack2_ssm,aegis_v1,attention,t5,mrpc,PASS,,26.75,41.4,68.15,0.0,1.0,4671888,4671888,0.965,1161.6,0.0,0.6922,0.0,0.0,0.0,8.6896,\"{'loss': [0.6922, 0.6922, 0.6922, 0.6922, 0.6922], 'grad_norm': [8.6896, 8.6896, 8.6896, 8.6896, 8.6896], 'fwd_latency_ms': [24.9622, 26.8394, 26.5849, 27.1512, 28.2057], 'bwd_latency_ms': [41.2369, 40.6017, 41.0951, 42.5731, 41.5173]}\",2026-08-26 02:34:19",
"results/master_aggregated_results.csv": "aegis_v1,0,PASS,78.56,50.92,53.56,100.0,2026-09-11T03:51:13.411537\naegis_v1_module,0,PASS,78.56,50.92,53.56,102.8,2026-09-11T03:52:56.223785",
"results/run_quality_all_screening_summary.csv": "run_quality_20260827_052444,aegis_v1,DISTILBERT,SST2,PASS,50.92,0.0,50.92,50.92,1.0,0.6748,0.2107,134.36,369.55,2027.5,32513.8,1450000000.0,2026-08-27T05:55:21.501855\nrun_quality_20260827_052444,aegis_v1_fn,DISTILBERT,SST2,PASS,50.92,0.0,50.92,50.92,1.0,0.6748,0.2107,133.63,370.3,2027.5,32512.6,1450000000.0,2026-08-27T05:56:58.842085"
},
"notes": []
},
"aegis_v2": {
"module_name": "aegis_v2",
"source_file": "Codes/framework/attentions/aegis_v2.py",
"source_lines_count": 772,
"architecture": {
"class_name": "FactorisedLinear",
"base_class": "nn.Module",
"init_args": "self, in_features: int, out_features: int, rank: int = 384, bias: bool = True, lora_rank: int = 16, lora_alpha: float = 32.0, energy: float = 0.95, **kwargs: Any, ) -> None: super().__init__() self.in_features = in_features self.out_features = out_features self.rank = min(rank, in_features, out_features) self.energy = energy self.lora_rank = lora_rank self.lora_alpha = lora_alpha self.lora_scaling = (lora_alpha / lora_rank) if lora_rank > 0 else 1.0 self.down = nn.Linear(in_features, self.rank, bias=False) self.up = nn.Linear(self.rank, out_features, bias=bias) if self.lora_rank > 0: self.lora_A = nn.Parameter(torch.randn(lora_rank, in_features) * 0.01) self.lora_B = nn.Parameter(torch.zeros(out_features, lora_rank)) else: self.register_parameter(\"lora_A\", None) self.register_parameter(\"lora_B\", None) self.reset_parameters() def reset_parameters(self) -> None: nn.init.orthogonal_(self.down.weight) nn.init.orthogonal_(self.up.weight) if self.up.bias is not None: nn.init.zeros_(self.up.bias) if self.lora_rank > 0 and self.lora_A is not None: nn.init.kaiming_uniform_(self.lora_A, a=math.sqrt(5)) self.lora_A.requires_grad = True if self.lora_rank > 0 and self.lora_B is not None: nn.init.zeros_(self.lora_B) self.lora_B.requires_grad = True def decompose_from_weight( self, weight: torch.Tensor, bias: Optional[torch.Tensor] = None, energy: Optional[float] = None, ) -> None: with torch.no_grad(",
"registry_names": [
"aegis_v2_inject",
"inject_aegis_v2",
"aegis_v2"
]
},
"params": [
{
"name": "down",
"type": "nn.Linear",
"declaration": "nn.Linear(in_features, self.rank, bias=False)"
},
{
"name": "up",
"type": "nn.Linear",
"declaration": "nn.Linear(self.rank, out_features, bias=bias)"
},
{
"name": "lora_A",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.randn(lora_rank, in_features)"
},
{
"name": "lora_B",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.zeros(out_features, lora_rank)"
},
{
"name": "fusion_gate",
"type": "nn.Linear",
"declaration": "nn.Linear(self.hidden_size, self.num_heads)"
}
],
"metrics": {
"results/unified_aegis_v2/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": 7.16
},
"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": 22993.16
},
"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": "2026-09-15T10:21:18.040662"
},
"results/unified_aegis_v2/roberta/roberta-base_aegis_v2_full_metrics.json": {
"model": "roberta-base_aegis_v2",
"performance": {
"accuracy_pct": 62.39,
"f1_score": 0.7163,
"f1_macro": 0.5792,
"precision": 0.697,
"recall": 0.6181,
"val_loss": 0.6316,
"status": "PASS",
"elapsed_s": 46.28
},
"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": 591.16,
"tokens_per_sec": 7112.56
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 549.74,
"L_256_vram_mb": 556.23,
"L_512_vram_mb": 582.32,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 514)"
},
"timestamp": "2026-09-15T10:22:04.437819"
},
"results/unified_aegis_v2/distilbert/distilbert_sst2_aegis_v2_full_metrics.json": {
"model": "distilbert_sst2_aegis_v2",
"performance": {
"accuracy_pct": 72.36,
"f1_score": 0.7737,
"f1_macro": 0.7094,
"precision": 0.768,
"recall": 0.7198,
"val_loss": 0.6091,
"status": "PASS",
"elapsed_s": 22.73
},
"efficiency": {
"total_params_before": 66955010,
"total_params_after": 55523882,
"trainable_recovery_params": 1545000,
"compression_rate": 1.2059,
"compression_ratio_pct": 17.07,
"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": 195.7,
"tokens_per_sec": 14953.19
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 153.52,
"L_256_vram_mb": 161.77,
"L_512_vram_mb": "unsupported (exceeds model max_position_embeddings 512)",
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 512)"
},
"timestamp": "2026-09-15T10:20:34.952642"
},
"results/unified_aegis_v2/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": 3.02
},
"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": 51105.1
},
"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": "2026-09-15T10:20:12.018768"
},
"results/unified_aegis_v2/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": 6.96
},
"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": 23739.6
},
"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": "2026-09-15T10:22:50.139145"
},
"results/unified_aegis_v2/gpt2/gpt2_aegis_v2_full_metrics.json": {
"model": "gpt2_aegis_v2",
"performance": {
"accuracy_pct": 52.98,
"f1_score": 0.6782,
"f1_macro": 0.4029,
"precision": 0.6174,
"recall": 0.5215,
"val_loss": 0.6799,
"status": "PASS_LOW",
"elapsed_s": 39.97
},
"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": 8738.48
},
"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": 463.65,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 1024)"
},
"timestamp": "2026-09-15T10:23:30.276275"
},
"results/unified_aegis_v2/gpt2/hf_export/metrics.json": {
"model_name": "gpt2",
"module": "aegis_v2",
"status": "PASS",
"accuracy_pct": 74.89,
"f1": 0.0,
"precision": 0.7669,
"recall": 0.7464,
"eval_loss": 0.0,
"params_total": null,
"params_active": null,
"param_reduction_pct": null,
"peak_vram_mb": 300.59,
"tokens_per_sec": 5372.02,
"elapsed_seconds": null,
"full_metrics": {
"module_name": "aegis_v2",
"status": "PASS",
"accuracy_pct": 74.89,
"f1_score": 0.7812,
"f1_macro": 0.7432,
"precision": 0.7669,
"recall": 0.7464,
"val_loss": 0.5276,
"elapsed_s": 95.78,
"peak_vram_mb": 300.59,
"tokens_per_sec": 5372.02,
"total_params_before": 126800640,
"total_params_after": 103053648,
"trainable_recovery_params": 4564560,
"compression_rate": 1.2304,
"compression_ratio_pct": 18.73,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 72,
"num_fallback_layers": 0,
"avg_reconstruction_error": NaN,
"weighted_reconstruction_error": NaN,
"num_injected_attention_layers": 12,
"lora_params": 2654208,
"error": NaN,
"path": "adapted_gpt",
"timestamp": "2026-09-14T18:13:18.296420"
}
},
"results/unified_aegis_v2/distilbert/hf_export/metrics.json": {
"model_name": "distilbert",
"module": "aegis_v2",
"status": "PASS",
"accuracy_pct": 72.36,
"f1": 0.0,
"precision": 0.768,
"recall": 0.7198,
"eval_loss": 0.0,
"params_total": null,
"params_active": null,
"param_reduction_pct": null,
"peak_vram_mb": 195.7,
"tokens_per_sec": 14953.19,
"elapsed_seconds": null,
"full_metrics": {
"module_name": "aegis_v2",
"status": "PASS",
"accuracy_pct": 72.36,
"f1_score": 0.7737,
"f1_macro": 0.7094,
"precision": 0.768,
"recall": 0.7198,
"val_loss": 0.6091,
"elapsed_s": 22.73,
"peak_vram_mb": 195.7,
"tokens_per_sec": 14953.19,
"total_params_before": 66955010,
"total_params_after": 55523882,
"trainable_recovery_params": 1545000,
"compression_rate": 1.2059,
"compression_ratio_pct": 17.07,
"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": "2026-09-15T10:20:35.208762"
}
},
"results/unified_aegis_v2/roberta/hf_export/metrics.json": {
"model_name": "roberta",
"module": "aegis_v2",
"status": "PASS",
"accuracy_pct": 62.39,
"f1": 0.0,
"precision": 0.697,
"recall": 0.6181,
"eval_loss": 0.0,
"params_total": null,
"params_active": null,
"param_reduction_pct": null,
"peak_vram_mb": 591.16,
"tokens_per_sec": 7112.56,
"elapsed_seconds": null,
"full_metrics": {
"module_name": "aegis_v2",
"status": "PASS",
"accuracy_pct": 62.39,
"f1_score": 0.7163,
"f1_macro": 0.5792,
"precision": 0.697,
"recall": 0.6181,
"val_loss": 0.6316,
"elapsed_s": 46.28,
"peak_vram_mb": 591.16,
"tokens_per_sec": 7112.56,
"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,
"avg_reconstruction_error": NaN,
"weighted_reconstruction_error": NaN,
"num_injected_attention_layers": 12,
"lora_params": 1179648,
"error": NaN,
"path": "adapted_roberta",
"timestamp": "2026-09-15T10:22:05.426774"
}
},
"results/unified_aegis_v2/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,78.56,0.8124,0.7811,0.8051,0.7832,0.4381,7.16,712.34,22993.16,124647170,124647170,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_roberta,2026-09-15T10:21:18.880387\naegis_v2,PASS,62.39,0.7163,0.5792,0.697,0.6181,0.6316,46.28,591.16,7112.56,124647170,108872018,3090000,1.1449,12.66,0.0,0,0,,,12,1179648,,adapted_roberta,2026-09-15T10:22:05.426774\n",
"results/unified_aegis_v2/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,90.71,0.9099,0.907,0.9075,0.9068,0.3028,3.02,182.94,51105.1,66955010,66955010,0,1.0,0.0,0.0,0,0,,,0,0,,legacy_bert,2026-09-15T10:20:12.230121\naegis_v2,PASS,72.36,0.7737,0.7094,0.768,0.7198,0.6091,22.73,195.7,14953.19,66955010,55523882,1545000,1.2059,17.07,0.0,0,0,,,6,589824,,legacy_bert,2026-09-15T10:20:35.208762\n",
"results/unified_aegis_v2/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,79.36,0.7978,0.7935,0.7935,0.7935,0.4875,6.96,793.19,23739.6,124441344,124441344,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_gpt,2026-09-15T10:22:50.993179\naegis_v2,PASS_LOW,52.98,0.6782,0.4029,0.6174,0.5215,0.6799,39.98,476.0,8738.48,124441344,101579088,3090000,1.2251,18.37,0.0,0,0,,,12,1179648,note: evaluated accuracy (52.98%) below chance threshold (56.00%),adapted_gpt,2026-09-15T10:23:31.231827\n",
"results/benchmark_summary_v2.csv": "track2_ssm,aegis_v2,attention,bert,sst2,PASS,,31.66,55.93,87.59,0.0,1.0,4536866,4536866,0.967,1842.9,0.0,0.692,0.0,0.0,0.0,0.1078,\"{'loss': [0.6919, 0.6918, 0.6928, 0.6916, 0.692], 'grad_norm': [0.1033, 0.1085, 0.1121, 0.1076, 0.1075], 'fwd_latency_ms': [31.5372, 31.6417, 31.5357, 32.013, 31.5813], 'bwd_latency_ms': [58.6136, 55.45, 54.3287, 55.6125, 55.6628]}\",2026-08-26 02:34:20\ntrack2_ssm,aegis_v2,attention,bert,mrpc,PASS,,32.58,56.91,89.49,0.0,1.0,4536866,4536866,0.967,1842.9,0.0,0.6826,0.0,0.0,0.0,1.0227,\"{'loss': [0.6823, 0.6838, 0.6815, 0.6822, 0.6834], 'grad_norm': [1.0105, 1.0287, 1.0174, 1.0276, 1.0293], 'fwd_latency_ms': [31.4044, 32.0362, 33.0865, 33.4091, 32.9404], 'bwd_latency_ms': [57.8027, 54.904, 55.9912, 58.3431, 57.5308]}\",2026-08-26 02:34:21\ntrack2_ssm,aegis_v2,attention,roberta,sst2,PASS,,31.81,55.93,87.74,0.0,1.0,7063970,7063970,0.979,1852.8,0.0,0.6934,0.0,0.0,0.0,0.1397,\"{'loss': [0.695, 0.6933, 0.6923, 0.6943, 0.6922], 'grad_norm': [0.1411, 0.132, 0.1448, 0.1384, 0.1424], 'fwd_latency_ms': [31.2033, 31.668, 31.925, 31.982, 32.2502], 'bwd_latency_ms': [57.03, 55.2651, 56.2149, 55.5551, 55.5637]}\",2026-08-26 02:34:22\ntrack2_ssm,aegis_v2,attention,roberta,mrpc,PASS,,32.47,56.98,89.45,0.0,1.0,7063970,7063970,0.979,1852.8,0.0,0.7161,0.0,0.0,0.0,1.4052,\"{'loss': [0.7165, 0.7159, 0.7159, 0.7176, 0.7148], 'grad_norm': [1.4071, 1.4148, 1.4056, 1.3978, 1.4005], 'fwd_latency_ms': [31.3939, 32.4033, 32.4945, 32.9121, 33.1613], 'bwd_latency_ms': [57.7051, 56.2636, 56.6519, 57.0505, 57.242]}\",2026-08-26 02:34:23\ntrack2_ssm,aegis_v2,attention,t5,sst2,PASS,,32.55,57.3,89.85,0.0,1.0,4658208,4658208,0.967,1803.5,0.0,0.6951,0.0,0.0,0.0,0.3099,\"{'loss': [0.695, 0.6946, 0.6948, 0.6957, 0.6952], 'grad_norm': [0.3087, 0.3028, 0.3056, 0.3211, 0.3111], 'fwd_latency_ms': [30.08, 33.1307, 33.3359, 33.1036, 33.1181], 'bwd_latency_ms': [57.8667, 56.9766, 57.3725, 56.8349, 57.4614]}\",2026-08-26 02:34:24\ntrack2_ssm,aegis_v2,attention,t5,mrpc,PASS,,32.66,56.95,89.61,0.0,1.0,4658208,4658208,0.967,1803.5,0.0,0.6356,0.0,0.0,0.0,1.5323,\"{'loss': [0.6357, 0.6365, 0.6346, 0.637, 0.6344], 'grad_norm': [1.5218, 1.5453, 1.5279, 1.5378, 1.5288], 'fwd_latency_ms': [29.8906, 34.3345, 32.5585, 33.2979, 33.2324], 'bwd_latency_ms': [57.2124, 56.8658, 56.4599, 57.0038, 57.188]}\",2026-08-26 02:34:25",
"results/master_aggregated_results.csv": "aegis_v2,0,PASS,79.82,50.92,51.03,95.5,2026-09-11T03:54:31.745295",
"results/run_quality_all_screening_summary.csv": "run_quality_20260827_052444,aegis_v2,DISTILBERT,SST2,PASS,50.92,0.0,50.92,50.92,1.0,0.6748,0.2291,230.05,592.49,4821.8,19918.7,1450000000.0,2026-08-27T05:59:00.135401"
},
"notes": []
},
"aegis_v3": {
"module_name": "aegis_v3",
"source_file": "Codes/framework/attentions/aegis_v3.py",
"source_lines_count": 789,
"architecture": {
"class_name": "FactorisedLinear",
"base_class": "nn.Module",
"init_args": "self, in_features_or_weight: Union[int, torch.Tensor], out_features_or_bias: Optional[Union[int, torch.Tensor]] = None, energy: float = 0.95, min_rank: int = 4, max_rank: int = 384, rank: Optional[int] = None, bias: bool = True, lora_rank: int = 16, lora_alpha: float = 32.0, **kwargs: Any, ) -> None: super().__init__() self.energy = energy self.min_rank = min_rank self.max_rank = max_rank self.lora_rank = lora_rank self.lora_alpha = lora_alpha self.lora_scaling = (lora_alpha / lora_rank) if lora_rank > 0 else 1.0 if isinstance(in_features_or_weight, torch.Tensor",
"registry_names": [
"aegis_v3_inject",
"aegis_v3_fn",
"AegisAttention",
"aegis_v3",
"aegis_v3_module"
]
},
"params": [
{
"name": "down",
"type": "nn.Linear",
"declaration": "nn.Linear(in_dim, self.rank, bias=False, device=device, dtype=dtype)"
},
{
"name": "up",
"type": "nn.Linear",
"declaration": "nn.Linear(self.rank, out_dim, bias=(b is not None or bias)"
},
{
"name": "lora_A",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.randn(lora_rank, in_dim, device=device, dtype=dtype)"
},
{
"name": "lora_B",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.zeros(out_dim, lora_rank, device=device, dtype=dtype)"
},
{
"name": "down",
"type": "nn.Linear",
"declaration": "nn.Linear(in_dim, self.rank, bias=False)"
},
{
"name": "up",
"type": "nn.Linear",
"declaration": "nn.Linear(self.rank, out_dim, bias=bias)"
},
{
"name": "lora_A",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.randn(lora_rank, in_dim)"
},
{
"name": "lora_B",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.zeros(out_dim, lora_rank)"
}
],
"metrics": {
"results/unified_aegis_v3/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": 6.91
},
"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": 24095.68
},
"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": "2026-09-15T10:26:04.967232"
},
"results/unified_aegis_v3/roberta/roberta-base_aegis_v3_full_metrics.json": {
"model": "roberta-base_aegis_v3",
"performance": {
"accuracy_pct": 55.16,
"f1_score": 0.6943,
"f1_macro": 0.4267,
"precision": 0.7659,
"recall": 0.5432,
"val_loss": 0.8824,
"status": "PASS_NO_COMPRESSION",
"elapsed_s": 57.5
},
"efficiency": {
"total_params_before": 124647170,
"total_params_after": 133307330,
"trainable_recovery_params": 1573056,
"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": 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": 384,
"roberta.encoder.layer.0.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.0.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.0.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.1.attention.self": 32,
"roberta.encoder.layer.1.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.1.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.1.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.1.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.2.attention.self": 32,
"roberta.encoder.layer.2.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.2.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.2.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.2.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.3.attention.self": 32,
"roberta.encoder.layer.3.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.3.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.3.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.3.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.4.attention.self": 32,
"roberta.encoder.layer.4.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.4.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.4.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.4.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.5.attention.self": 32,
"roberta.encoder.layer.5.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.5.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.5.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.5.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.6.attention.self": 32,
"roberta.encoder.layer.6.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.6.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.6.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.6.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.7.attention.self": 32,
"roberta.encoder.layer.7.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.7.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.7.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.7.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.8.attention.self": 32,
"roberta.encoder.layer.8.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.8.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.8.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.8.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.9.attention.self": 32,
"roberta.encoder.layer.9.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.9.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.9.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.9.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.10.attention.self": 32,
"roberta.encoder.layer.10.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.10.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.10.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.10.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.11.attention.self": 32,
"roberta.encoder.layer.11.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.11.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.11.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.11.attention.self._wrapped_attention.out_proj": 384
}
},
"hardware": {
"peak_vram_mb": 688.09,
"tokens_per_sec": 5952.88
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 655.19,
"L_256_vram_mb": 665.1,
"L_512_vram_mb": 710.93,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 514)"
},
"timestamp": "2026-09-15T10:27:02.464741"
},
"results/unified_aegis_v3/distilbert/distilbert_sst2_aegis_v3_full_metrics.json": {
"model": "distilbert_sst2_aegis_v3",
"performance": {
"accuracy_pct": 78.44,
"f1_score": 0.8216,
"f1_macro": 0.7746,
"precision": 0.8339,
"recall": 0.7808,
"val_loss": 0.686,
"status": "PASS_NO_COMPRESSION",
"elapsed_s": 28.11
},
"efficiency": {
"total_params_before": 66955010,
"total_params_after": 67741538,
"trainable_recovery_params": 786528,
"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": 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": 384,
"distilbert.transformer.layer.0.attention._wrapped_attention.k_proj": 384,
"distilbert.transformer.layer.0.attention._wrapped_attention.v_proj": 384,
"distilbert.transformer.layer.0.attention._wrapped_attention.out_proj": 384,
"distilbert.transformer.layer.1.attention": 32,
"distilbert.transformer.layer.1.attention._wrapped_attention.q_proj": 384,
"distilbert.transformer.layer.1.attention._wrapped_attention.k_proj": 384,
"distilbert.transformer.layer.1.attention._wrapped_attention.v_proj": 384,
"distilbert.transformer.layer.1.attention._wrapped_attention.out_proj": 384,
"distilbert.transformer.layer.2.attention": 32,
"distilbert.transformer.layer.2.attention._wrapped_attention.q_proj": 384,
"distilbert.transformer.layer.2.attention._wrapped_attention.k_proj": 384,
"distilbert.transformer.layer.2.attention._wrapped_attention.v_proj": 384,
"distilbert.transformer.layer.2.attention._wrapped_attention.out_proj": 384,
"distilbert.transformer.layer.3.attention": 32,
"distilbert.transformer.layer.3.attention._wrapped_attention.q_proj": 384,
"distilbert.transformer.layer.3.attention._wrapped_attention.k_proj": 384,
"distilbert.transformer.layer.3.attention._wrapped_attention.v_proj": 384,
"distilbert.transformer.layer.3.attention._wrapped_attention.out_proj": 384,
"distilbert.transformer.layer.4.attention": 32,
"distilbert.transformer.layer.4.attention._wrapped_attention.q_proj": 384,
"distilbert.transformer.layer.4.attention._wrapped_attention.k_proj": 384,
"distilbert.transformer.layer.4.attention._wrapped_attention.v_proj": 384,
"distilbert.transformer.layer.4.attention._wrapped_attention.out_proj": 384,
"distilbert.transformer.layer.5.attention": 32,
"distilbert.transformer.layer.5.attention._wrapped_attention.q_proj": 384,
"distilbert.transformer.layer.5.attention._wrapped_attention.k_proj": 384,
"distilbert.transformer.layer.5.attention._wrapped_attention.v_proj": 384,
"distilbert.transformer.layer.5.attention._wrapped_attention.out_proj": 384
}
},
"hardware": {
"peak_vram_mb": 208.28,
"tokens_per_sec": 12530.63
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 170.6,
"L_256_vram_mb": 183.03,
"L_512_vram_mb": "unsupported (exceeds model max_position_embeddings 512)",
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 512)"
},
"timestamp": "2026-09-15T10:25:20.779142"
},
"results/unified_aegis_v3/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": 3.02
},
"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": 51110.39
},
"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": "2026-09-15T10:24:52.492596"
},
"results/unified_aegis_v3/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": 7.01
},
"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": 23517.17
},
"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": "2026-09-15T10:27:47.569320"
},
"results/unified_aegis_v3/gpt2/gpt2_aegis_v3_full_metrics.json": {
"model": "gpt2_aegis_v3",
"performance": {
"accuracy_pct": 54.82,
"f1_score": 0.5397,
"f1_macro": 0.548,
"precision": 0.5488,
"recall": 0.5487,
"val_loss": 0.7411,
"status": "PASS_NO_COMPRESSION",
"elapsed_s": 42.4
},
"efficiency": {
"total_params_before": 124441344,
"total_params_after": 126014400,
"trainable_recovery_params": 1573056,
"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": 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": 384,
"transformer.h.0.attn._wrapped_attention.k_proj": 384,
"transformer.h.0.attn._wrapped_attention.v_proj": 384,
"transformer.h.0.attn._wrapped_attention.out_proj": 384,
"transformer.h.1.attn": 32,
"transformer.h.1.attn._wrapped_attention.q_proj": 384,
"transformer.h.1.attn._wrapped_attention.k_proj": 384,
"transformer.h.1.attn._wrapped_attention.v_proj": 384,
"transformer.h.1.attn._wrapped_attention.out_proj": 384,
"transformer.h.2.attn": 32,
"transformer.h.2.attn._wrapped_attention.q_proj": 384,
"transformer.h.2.attn._wrapped_attention.k_proj": 384,
"transformer.h.2.attn._wrapped_attention.v_proj": 384,
"transformer.h.2.attn._wrapped_attention.out_proj": 384,
"transformer.h.3.attn": 32,
"transformer.h.3.attn._wrapped_attention.q_proj": 384,
"transformer.h.3.attn._wrapped_attention.k_proj": 384,
"transformer.h.3.attn._wrapped_attention.v_proj": 384,
"transformer.h.3.attn._wrapped_attention.out_proj": 384,
"transformer.h.4.attn": 32,
"transformer.h.4.attn._wrapped_attention.q_proj": 384,
"transformer.h.4.attn._wrapped_attention.k_proj": 384,
"transformer.h.4.attn._wrapped_attention.v_proj": 384,
"transformer.h.4.attn._wrapped_attention.out_proj": 384,
"transformer.h.5.attn": 32,
"transformer.h.5.attn._wrapped_attention.q_proj": 384,
"transformer.h.5.attn._wrapped_attention.k_proj": 384,
"transformer.h.5.attn._wrapped_attention.v_proj": 384,
"transformer.h.5.attn._wrapped_attention.out_proj": 384,
"transformer.h.6.attn": 32,
"transformer.h.6.attn._wrapped_attention.q_proj": 384,
"transformer.h.6.attn._wrapped_attention.k_proj": 384,
"transformer.h.6.attn._wrapped_attention.v_proj": 384,
"transformer.h.6.attn._wrapped_attention.out_proj": 384,
"transformer.h.7.attn": 32,
"transformer.h.7.attn._wrapped_attention.q_proj": 384,
"transformer.h.7.attn._wrapped_attention.k_proj": 384,
"transformer.h.7.attn._wrapped_attention.v_proj": 384,
"transformer.h.7.attn._wrapped_attention.out_proj": 384,
"transformer.h.8.attn": 32,
"transformer.h.8.attn._wrapped_attention.q_proj": 384,
"transformer.h.8.attn._wrapped_attention.k_proj": 384,
"transformer.h.8.attn._wrapped_attention.v_proj": 384,
"transformer.h.8.attn._wrapped_attention.out_proj": 384,
"transformer.h.9.attn": 32,
"transformer.h.9.attn._wrapped_attention.q_proj": 384,
"transformer.h.9.attn._wrapped_attention.k_proj": 384,
"transformer.h.9.attn._wrapped_attention.v_proj": 384,
"transformer.h.9.attn._wrapped_attention.out_proj": 384,
"transformer.h.10.attn": 32,
"transformer.h.10.attn._wrapped_attention.q_proj": 384,
"transformer.h.10.attn._wrapped_attention.k_proj": 384,
"transformer.h.10.attn._wrapped_attention.v_proj": 384,
"transformer.h.10.attn._wrapped_attention.out_proj": 384,
"transformer.h.11.attn": 32,
"transformer.h.11.attn._wrapped_attention.q_proj": 384,
"transformer.h.11.attn._wrapped_attention.k_proj": 384,
"transformer.h.11.attn._wrapped_attention.v_proj": 384,
"transformer.h.11.attn._wrapped_attention.out_proj": 384
}
},
"hardware": {
"peak_vram_mb": 594.88,
"tokens_per_sec": 8863.42
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 547.9,
"L_256_vram_mb": 558.58,
"L_512_vram_mb": 611.74,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 1024)"
},
"timestamp": "2026-09-15T10:28:30.046806"
},
"results/unified_aegis_v3/gpt2/hf_export/metrics.json": {
"model_name": "gpt2",
"module": "aegis_v3",
"status": "PASS",
"accuracy_pct": 56.31,
"f1": 0.0,
"precision": 0.5758,
"recall": 0.5587,
"eval_loss": 0.0,
"params_total": null,
"params_active": null,
"param_reduction_pct": null,
"peak_vram_mb": 595.01,
"tokens_per_sec": 45270.42,
"elapsed_seconds": null,
"full_metrics": {
"module_name": "aegis_v3",
"status": "PASS",
"accuracy_pct": 56.31,
"f1_score": 0.6495,
"f1_macro": 0.5348,
"precision": 0.5758,
"recall": 0.5587,
"val_loss": 0.7094,
"elapsed_s": 15.93,
"peak_vram_mb": 595.01,
"tokens_per_sec": 45270.42,
"total_params_before": 124441344,
"total_params_after": 126014400,
"trainable_recovery_params": 1574592,
"compression_rate": 0.9875,
"compression_ratio_pct": -1.26,
"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": "2026-09-11T03:56:09.839059"
}
},
"results/unified_aegis_v3/distilbert/hf_export/metrics.json": {
"model_name": "distilbert",
"module": "aegis_v3",
"status": "PASS",
"accuracy_pct": 79.7,
"f1": 0.0,
"precision": 0.8363,
"recall": 0.7939,
"eval_loss": 0.0,
"params_total": null,
"params_active": null,
"param_reduction_pct": null,
"peak_vram_mb": 344.68,
"tokens_per_sec": 12279.99,
"elapsed_seconds": null,
"full_metrics": {
"module_name": "aegis_v3",
"status": "PASS",
"accuracy_pct": 79.7,
"f1_score": 0.829,
"f1_macro": 0.7897,
"precision": 0.8363,
"recall": 0.7939,
"val_loss": 0.692,
"elapsed_s": 27.06,
"peak_vram_mb": 344.68,
"tokens_per_sec": 12279.99,
"total_params_before": 66955010,
"total_params_after": 67741538,
"trainable_recovery_params": 786528,
"compression_rate": 0.9884,
"compression_ratio_pct": -1.17,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 24,
"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": "2026-09-14T22:23:52.352377"
}
},
"results/unified_aegis_v3/roberta/hf_export/metrics.json": {
"model_name": "roberta",
"module": "aegis_v3",
"status": "PASS",
"accuracy_pct": 86.47,
"f1": 0.0,
"precision": 0.866,
"recall": 0.8641,
"eval_loss": 0.0,
"params_total": null,
"params_active": null,
"param_reduction_pct": null,
"peak_vram_mb": 404.47,
"tokens_per_sec": 5921.14,
"elapsed_seconds": null,
"full_metrics": {
"module_name": "aegis_v3",
"status": "PASS",
"accuracy_pct": 86.47,
"f1_score": 0.8706,
"f1_macro": 0.8644,
"precision": 0.866,
"recall": 0.8641,
"val_loss": 0.3717,
"elapsed_s": 82.53,
"peak_vram_mb": 404.47,
"tokens_per_sec": 5921.14,
"total_params_before": 127325954,
"total_params_after": 135101378,
"trainable_recovery_params": 3367104,
"compression_rate": 0.9424,
"compression_ratio_pct": -6.11,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 85,
"num_fallback_layers": 0,
"avg_reconstruction_error": NaN,
"weighted_reconstruction_error": NaN,
"num_injected_attention_layers": 12,
"lora_params": 2973696,
"error": NaN,
"path": "adapted_roberta",
"timestamp": "2026-09-14T11:02:54.072940"
}
},
"results/unified_aegis_v3/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,78.56,0.8124,0.7811,0.8051,0.7832,0.4381,6.91,712.34,24095.68,124647170,124647170,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_roberta,2026-09-15T10:26:05.763299\naegis_v3,PASS_NO_COMPRESSION,55.16,0.6943,0.4267,0.7659,0.5432,0.8824,57.5,688.09,5952.88,124647170,133307330,1573056,1.0,0.0,0.0,0,0,,,12,1179648,\"note: zero compressed layers, model grew (LoRA only)\",adapted_roberta,2026-09-15T10:27:03.548424\n",
"results/unified_aegis_v3/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,90.71,0.9099,0.907,0.9075,0.9068,0.3028,3.02,182.94,51110.39,66955010,66955010,0,1.0,0.0,0.0,0,0,,,0,0,,legacy_bert,2026-09-15T10:24:52.698989\naegis_v3,PASS_NO_COMPRESSION,78.44,0.8216,0.7746,0.8339,0.7808,0.686,28.11,208.28,12530.63,66955010,67741538,786528,1.0,0.0,0.0,0,0,,,6,589824,\"note: zero compressed layers, model grew (LoRA only)\",legacy_bert,2026-09-15T10:25:21.060636\n",
"results/unified_aegis_v3/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,79.36,0.7978,0.7935,0.7935,0.7935,0.4875,7.01,793.19,23517.17,124441344,124441344,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_gpt,2026-09-15T10:27:48.358787\naegis_v3,PASS_NO_COMPRESSION,54.82,0.5397,0.548,0.5488,0.5487,0.7411,42.41,594.88,8863.42,124441344,126014400,1573056,1.0,0.0,0.0,0,0,,,12,1179648,\"note: zero compressed layers, model grew (LoRA only)\",adapted_gpt,2026-09-15T10:28:31.044337\n",
"results/benchmark_summary_v2.csv": "track2_ssm,aegis_v3,attention,bert,sst2,PASS,,37.74,60.73,98.47,0.0,1.0,4583314,4583314,0.957,1634.5,0.0,0.6933,0.0,0.0,0.0,0.193,\"{'loss': [0.6933, 0.6933, 0.6933, 0.6933, 0.6933], 'grad_norm': [0.193, 0.193, 0.193, 0.193, 0.193], 'fwd_latency_ms': [37.7539, 38.275, 37.2725, 37.927, 37.4588], 'bwd_latency_ms': [60.5982, 61.3298, 60.8258, 60.9201, 59.9962]}\",2026-08-26 02:34:26\ntrack2_ssm,aegis_v3,attention,bert,mrpc,PASS,,38.16,61.97,100.13,0.0,1.0,4583314,4583314,0.957,1634.5,0.0,0.682,0.0,0.0,0.0,1.0852,\"{'loss': [0.682, 0.682, 0.682, 0.682, 0.682], 'grad_norm': [1.0852, 1.0852, 1.0852, 1.0852, 1.0852], 'fwd_latency_ms': [37.5859, 38.9337, 38.42, 38.7679, 37.1166], 'bwd_latency_ms': [62.1055, 63.0677, 63.2806, 60.8739, 60.5207]}\",2026-08-26 02:34:27\ntrack2_ssm,aegis_v3,attention,roberta,sst2,PASS,,37.71,61.38,99.09,0.0,1.0,7110418,7110418,0.972,1644.4,0.0,0.6918,0.0,0.0,0.0,0.1983,\"{'loss': [0.6918, 0.6918, 0.6918, 0.6918, 0.6918], 'grad_norm': [0.1983, 0.1983, 0.1983, 0.1983, 0.1983], 'fwd_latency_ms': [36.9759, 37.9829, 37.9962, 37.9234, 37.6617], 'bwd_latency_ms': [61.0341, 61.7097, 61.2596, 61.4523, 61.4405]}\",2026-08-26 02:34:29\ntrack2_ssm,aegis_v3,attention,roberta,mrpc,PASS,,37.81,61.52,99.33,0.0,1.0,7110418,7110418,0.972,1644.4,0.0,0.7047,0.0,0.0,0.0,1.1352,\"{'loss': [0.7047, 0.7047, 0.7047, 0.7047, 0.7047], 'grad_norm': [1.1352, 1.1352, 1.1352, 1.1352, 1.1352], 'fwd_latency_ms': [37.0025, 37.7595, 37.7634, 37.9881, 38.5148], 'bwd_latency_ms': [61.3975, 61.5672, 61.1288, 62.0066, 61.4782]}\",2026-08-26 02:34:30\ntrack2_ssm,aegis_v3,attention,t5,sst2,PASS,,38.23,62.67,100.9,0.0,1.0,4704656,4704656,0.958,1610.8,0.0,1.249,0.0,0.0,0.0,17.2984,\"{'loss': [1.249, 1.249, 1.249, 1.249, 1.249], 'grad_norm': [17.2984, 17.2984, 17.2984, 17.2984, 17.2984], 'fwd_latency_ms': [35.159, 39.0033, 39.3809, 38.9671, 38.6314], 'bwd_latency_ms': [62.9316, 62.104, 62.0927, 63.6262, 62.6067]}\",2026-08-26 02:34:31\ntrack2_ssm,aegis_v3,attention,t5,mrpc,PASS,,38.22,62.48,100.7,0.0,1.0,4704656,4704656,0.958,1610.8,0.0,0.7407,0.0,0.0,0.0,8.4971,\"{'loss': [0.7407, 0.7407, 0.7407, 0.7407, 0.7407], 'grad_norm': [8.4971, 8.4971, 8.4971, 8.4971, 8.4971], 'fwd_latency_ms': [35.0321, 39.0813, 38.9985, 38.8383, 39.1449], 'bwd_latency_ms': [62.4095, 62.2867, 63.2798, 62.256, 62.1743]}\",2026-08-26 02:34:32",
"results/master_aggregated_results.csv": "aegis_v3,0,PASS,83.72,50.92,56.31,113.6,2026-09-11T03:56:25.382032\naegis_v3_module,0,PASS,83.72,50.92,56.31,109.9,2026-09-11T03:58:15.315308",
"results/run_quality_all_screening_summary.csv": "run_quality_20260827_052444,aegis_v3,DISTILBERT,SST2,PASS,50.92,0.0,50.92,50.92,1.0,0.6748,0.2008,173.0,473.84,2248.2,25329.6,1450000000.0,2026-08-27T06:00:52.264422"
},
"notes": []
},
"aegis_v4": {
"module_name": "aegis_v4",
"source_file": "Codes/framework/attentions/aegis_v4.py",
"source_lines_count": 794,
"architecture": {
"class_name": "FactorisedLinear",
"base_class": "nn.Module",
"init_args": "self, in_features_or_weight: Union[int, torch.Tensor], out_features_or_bias: Optional[Union[int, torch.Tensor]] = None, energy: float = 0.95, min_rank: int = 4, max_rank: int = 384, rank: Optional[int] = None, bias: bool = True, lora_rank: int = 16, lora_alpha: float = 32.0, **kwargs: Any, ) -> None: super().__init__() pass self.energy = energy self.min_rank = min_rank self.max_rank = max_rank self.lora_rank = lora_rank self.lora_alpha = lora_alpha self.lora_scaling = (lora_alpha / lora_rank) if lora_rank > 0 else 1.0 if isinstance(in_features_or_weight, torch.Tensor",
"registry_names": [
"aegis_v4_inject",
"aegis_v4_fn",
"aegis_v4_module",
"aegis_v4",
"AegisAttention"
]
},
"params": [
{
"name": "down",
"type": "nn.Linear",
"declaration": "nn.Linear(in_dim, self.rank, bias=False, device=device, dtype=dtype)"
},
{
"name": "up",
"type": "nn.Linear",
"declaration": "nn.Linear(self.rank, out_dim, bias=(b is not None or bias)"
},
{
"name": "lora_A",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.randn(lora_rank, in_dim, device=device, dtype=dtype)"
},
{
"name": "lora_B",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.zeros(out_dim, lora_rank, device=device, dtype=dtype)"
},
{
"name": "down",
"type": "nn.Linear",
"declaration": "nn.Linear(in_dim, self.rank, bias=False)"
},
{
"name": "up",
"type": "nn.Linear",
"declaration": "nn.Linear(self.rank, out_dim, bias=bias)"
},
{
"name": "lora_A",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.randn(lora_rank, in_dim)"
},
{
"name": "lora_B",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.zeros(out_dim, lora_rank)"
}
],
"metrics": {
"results/unified_aegis_v4/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": 6.76
},
"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": 24513.32
},
"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": "2026-09-15T10:34:55.564791"
},
"results/unified_aegis_v4/roberta/roberta-base_aegis_v4_full_metrics.json": {
"model": "roberta-base_aegis_v4",
"performance": {
"accuracy_pct": 57.8,
"f1_score": 0.707,
"f1_macro": 0.4765,
"precision": 0.7734,
"recall": 0.5701,
"val_loss": 0.822,
"status": "PASS_NO_COMPRESSION",
"elapsed_s": 55.24
},
"efficiency": {
"total_params_before": 124647170,
"total_params_after": 133307282,
"trainable_recovery_params": 1573008,
"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": 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": 384,
"roberta.encoder.layer.0.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.0.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.0.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.1.attention.self": 32,
"roberta.encoder.layer.1.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.1.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.1.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.1.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.2.attention.self": 32,
"roberta.encoder.layer.2.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.2.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.2.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.2.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.3.attention.self": 32,
"roberta.encoder.layer.3.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.3.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.3.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.3.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.4.attention.self": 32,
"roberta.encoder.layer.4.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.4.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.4.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.4.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.5.attention.self": 32,
"roberta.encoder.layer.5.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.5.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.5.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.5.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.6.attention.self": 32,
"roberta.encoder.layer.6.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.6.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.6.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.6.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.7.attention.self": 32,
"roberta.encoder.layer.7.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.7.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.7.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.7.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.8.attention.self": 32,
"roberta.encoder.layer.8.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.8.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.8.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.8.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.9.attention.self": 32,
"roberta.encoder.layer.9.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.9.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.9.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.9.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.10.attention.self": 32,
"roberta.encoder.layer.10.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.10.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.10.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.10.attention.self._wrapped_attention.out_proj": 384,
"roberta.encoder.layer.11.attention.self": 32,
"roberta.encoder.layer.11.attention.self._wrapped_attention.q_proj": 384,
"roberta.encoder.layer.11.attention.self._wrapped_attention.k_proj": 384,
"roberta.encoder.layer.11.attention.self._wrapped_attention.v_proj": 384,
"roberta.encoder.layer.11.attention.self._wrapped_attention.out_proj": 384
}
},
"hardware": {
"peak_vram_mb": 676.53,
"tokens_per_sec": 6149.86
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 655.14,
"L_256_vram_mb": 657.96,
"L_512_vram_mb": 681.73,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 514)"
},
"timestamp": "2026-09-15T10:35:50.885924"
},
"results/unified_aegis_v4/distilbert/distilbert_sst2_aegis_v4_full_metrics.json": {
"model": "distilbert_sst2_aegis_v4",
"performance": {
"accuracy_pct": 85.21,
"f1_score": 0.8652,
"f1_macro": 0.8506,
"precision": 0.8617,
"recall": 0.8506,
"val_loss": 0.3733,
"status": "PASS_NO_COMPRESSION",
"elapsed_s": 28.4
},
"efficiency": {
"total_params_before": 66955010,
"total_params_after": 67741514,
"trainable_recovery_params": 786504,
"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": 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": 384,
"distilbert.transformer.layer.0.attention._wrapped_attention.k_proj": 384,
"distilbert.transformer.layer.0.attention._wrapped_attention.v_proj": 384,
"distilbert.transformer.layer.0.attention._wrapped_attention.out_proj": 384,
"distilbert.transformer.layer.1.attention": 32,
"distilbert.transformer.layer.1.attention._wrapped_attention.q_proj": 384,
"distilbert.transformer.layer.1.attention._wrapped_attention.k_proj": 384,
"distilbert.transformer.layer.1.attention._wrapped_attention.v_proj": 384,
"distilbert.transformer.layer.1.attention._wrapped_attention.out_proj": 384,
"distilbert.transformer.layer.2.attention": 32,
"distilbert.transformer.layer.2.attention._wrapped_attention.q_proj": 384,
"distilbert.transformer.layer.2.attention._wrapped_attention.k_proj": 384,
"distilbert.transformer.layer.2.attention._wrapped_attention.v_proj": 384,
"distilbert.transformer.layer.2.attention._wrapped_attention.out_proj": 384,
"distilbert.transformer.layer.3.attention": 32,
"distilbert.transformer.layer.3.attention._wrapped_attention.q_proj": 384,
"distilbert.transformer.layer.3.attention._wrapped_attention.k_proj": 384,
"distilbert.transformer.layer.3.attention._wrapped_attention.v_proj": 384,
"distilbert.transformer.layer.3.attention._wrapped_attention.out_proj": 384,
"distilbert.transformer.layer.4.attention": 32,
"distilbert.transformer.layer.4.attention._wrapped_attention.q_proj": 384,
"distilbert.transformer.layer.4.attention._wrapped_attention.k_proj": 384,
"distilbert.transformer.layer.4.attention._wrapped_attention.v_proj": 384,
"distilbert.transformer.layer.4.attention._wrapped_attention.out_proj": 384,
"distilbert.transformer.layer.5.attention": 32,
"distilbert.transformer.layer.5.attention._wrapped_attention.q_proj": 384,
"distilbert.transformer.layer.5.attention._wrapped_attention.k_proj": 384,
"distilbert.transformer.layer.5.attention._wrapped_attention.v_proj": 384,
"distilbert.transformer.layer.5.attention._wrapped_attention.out_proj": 384
}
},
"hardware": {
"peak_vram_mb": 193.86,
"tokens_per_sec": 13166.42
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 168.97,
"L_256_vram_mb": 175.87,
"L_512_vram_mb": "unsupported (exceeds model max_position_embeddings 512)",
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 512)"
},
"timestamp": "2026-09-15T10:34:13.677608"
},
"results/unified_aegis_v4/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": 3.18
},
"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": 46184.49
},
"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": "2026-09-15T10:33:45.115072"
},
"results/unified_aegis_v4/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": 6.88
},
"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": 23723.71
},
"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": "2026-09-15T10:36:34.014838"
},
"results/unified_aegis_v4/gpt2/gpt2_aegis_v4_full_metrics.json": {
"model": "gpt2_aegis_v4",
"performance": {
"accuracy_pct": 70.99,
"f1_score": 0.7668,
"f1_macro": 0.6915,
"precision": 0.7639,
"recall": 0.7056,
"val_loss": 0.54,
"status": "PASS_NO_COMPRESSION",
"elapsed_s": 38.96
},
"efficiency": {
"total_params_before": 124441344,
"total_params_after": 126014352,
"trainable_recovery_params": 1573008,
"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": 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": 384,
"transformer.h.0.attn._wrapped_attention.k_proj": 384,
"transformer.h.0.attn._wrapped_attention.v_proj": 384,
"transformer.h.0.attn._wrapped_attention.out_proj": 384,
"transformer.h.1.attn": 32,
"transformer.h.1.attn._wrapped_attention.q_proj": 384,
"transformer.h.1.attn._wrapped_attention.k_proj": 384,
"transformer.h.1.attn._wrapped_attention.v_proj": 384,
"transformer.h.1.attn._wrapped_attention.out_proj": 384,
"transformer.h.2.attn": 32,
"transformer.h.2.attn._wrapped_attention.q_proj": 384,
"transformer.h.2.attn._wrapped_attention.k_proj": 384,
"transformer.h.2.attn._wrapped_attention.v_proj": 384,
"transformer.h.2.attn._wrapped_attention.out_proj": 384,
"transformer.h.3.attn": 32,
"transformer.h.3.attn._wrapped_attention.q_proj": 384,
"transformer.h.3.attn._wrapped_attention.k_proj": 384,
"transformer.h.3.attn._wrapped_attention.v_proj": 384,
"transformer.h.3.attn._wrapped_attention.out_proj": 384,
"transformer.h.4.attn": 32,
"transformer.h.4.attn._wrapped_attention.q_proj": 384,
"transformer.h.4.attn._wrapped_attention.k_proj": 384,
"transformer.h.4.attn._wrapped_attention.v_proj": 384,
"transformer.h.4.attn._wrapped_attention.out_proj": 384,
"transformer.h.5.attn": 32,
"transformer.h.5.attn._wrapped_attention.q_proj": 384,
"transformer.h.5.attn._wrapped_attention.k_proj": 384,
"transformer.h.5.attn._wrapped_attention.v_proj": 384,
"transformer.h.5.attn._wrapped_attention.out_proj": 384,
"transformer.h.6.attn": 32,
"transformer.h.6.attn._wrapped_attention.q_proj": 384,
"transformer.h.6.attn._wrapped_attention.k_proj": 384,
"transformer.h.6.attn._wrapped_attention.v_proj": 384,
"transformer.h.6.attn._wrapped_attention.out_proj": 384,
"transformer.h.7.attn": 32,
"transformer.h.7.attn._wrapped_attention.q_proj": 384,
"transformer.h.7.attn._wrapped_attention.k_proj": 384,
"transformer.h.7.attn._wrapped_attention.v_proj": 384,
"transformer.h.7.attn._wrapped_attention.out_proj": 384,
"transformer.h.8.attn": 32,
"transformer.h.8.attn._wrapped_attention.q_proj": 384,
"transformer.h.8.attn._wrapped_attention.k_proj": 384,
"transformer.h.8.attn._wrapped_attention.v_proj": 384,
"transformer.h.8.attn._wrapped_attention.out_proj": 384,
"transformer.h.9.attn": 32,
"transformer.h.9.attn._wrapped_attention.q_proj": 384,
"transformer.h.9.attn._wrapped_attention.k_proj": 384,
"transformer.h.9.attn._wrapped_attention.v_proj": 384,
"transformer.h.9.attn._wrapped_attention.out_proj": 384,
"transformer.h.10.attn": 32,
"transformer.h.10.attn._wrapped_attention.q_proj": 384,
"transformer.h.10.attn._wrapped_attention.k_proj": 384,
"transformer.h.10.attn._wrapped_attention.v_proj": 384,
"transformer.h.10.attn._wrapped_attention.out_proj": 384,
"transformer.h.11.attn": 32,
"transformer.h.11.attn._wrapped_attention.q_proj": 384,
"transformer.h.11.attn._wrapped_attention.k_proj": 384,
"transformer.h.11.attn._wrapped_attention.v_proj": 384,
"transformer.h.11.attn._wrapped_attention.out_proj": 384
}
},
"hardware": {
"peak_vram_mb": 590.29,
"tokens_per_sec": 9810.48
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 547.67,
"L_256_vram_mb": 552.29,
"L_512_vram_mb": 580.75,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 1024)"
},
"timestamp": "2026-09-15T10:37:13.127711"
},
"results/unified_aegis_v4/gpt2/hf_export/metrics.json": {
"model_name": "gpt2",
"module": "aegis_v4",
"status": "PASS",
"accuracy_pct": 69.5,
"f1": 0.0,
"precision": 0.7467,
"recall": 0.6906,
"eval_loss": 0.0,
"params_total": null,
"params_active": null,
"param_reduction_pct": null,
"peak_vram_mb": 348.6,
"tokens_per_sec": 9357.03,
"elapsed_seconds": null,
"full_metrics": {
"module_name": "aegis_v4",
"status": "PASS",
"accuracy_pct": 69.5,
"f1_score": 0.7555,
"f1_macro": 0.675,
"precision": 0.7467,
"recall": 0.6906,
"val_loss": 0.5708,
"elapsed_s": 55.86,
"peak_vram_mb": 348.6,
"tokens_per_sec": 9357.03,
"total_params_before": 126800640,
"total_params_after": 127488960,
"trainable_recovery_params": 3047616,
"compression_rate": 0.9946,
"compression_ratio_pct": -0.54,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 72,
"num_fallback_layers": 0,
"avg_reconstruction_error": NaN,
"weighted_reconstruction_error": NaN,
"num_injected_attention_layers": 12,
"lora_params": 2654208,
"error": NaN,
"path": "adapted_gpt",
"timestamp": "2026-09-14T11:10:29.011778"
}
},
"results/unified_aegis_v4/distilbert/hf_export/metrics.json": {
"model_name": "distilbert",
"module": "aegis_v4",
"status": "PASS",
"accuracy_pct": 83.72,
"f1": 0.0,
"precision": 0.8461,
"recall": 0.8357,
"eval_loss": 0.0,
"params_total": null,
"params_active": null,
"param_reduction_pct": null,
"peak_vram_mb": 212.04,
"tokens_per_sec": 60077.76,
"elapsed_seconds": null,
"full_metrics": {
"module_name": "aegis_v4",
"status": "PASS",
"accuracy_pct": 83.72,
"f1_score": 0.8515,
"f1_macro": 0.8356,
"precision": 0.8461,
"recall": 0.8357,
"val_loss": 0.5274,
"elapsed_s": 9.15,
"peak_vram_mb": 212.04,
"tokens_per_sec": 60077.76,
"total_params_before": 66955010,
"total_params_after": 67741538,
"trainable_recovery_params": 1378658,
"compression_rate": 0.9884,
"compression_ratio_pct": -1.17,
"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": "2026-09-11T03:58:40.204336"
}
},
"results/unified_aegis_v4/roberta/hf_export/metrics.json": {
"model_name": "roberta",
"module": "aegis_v4",
"status": "PASS",
"accuracy_pct": 86.47,
"f1": 0.0,
"precision": 0.866,
"recall": 0.8641,
"eval_loss": 0.0,
"params_total": null,
"params_active": null,
"param_reduction_pct": null,
"peak_vram_mb": 404.47,
"tokens_per_sec": 5891.73,
"elapsed_seconds": null,
"full_metrics": {
"module_name": "aegis_v4",
"status": "PASS",
"accuracy_pct": 86.47,
"f1_score": 0.8706,
"f1_macro": 0.8644,
"precision": 0.866,
"recall": 0.8641,
"val_loss": 0.3717,
"elapsed_s": 82.65,
"peak_vram_mb": 404.47,
"tokens_per_sec": 5891.73,
"total_params_before": 127325954,
"total_params_after": 135101378,
"trainable_recovery_params": 3367104,
"compression_rate": 0.9424,
"compression_ratio_pct": -6.11,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 85,
"num_fallback_layers": 0,
"avg_reconstruction_error": NaN,
"weighted_reconstruction_error": NaN,
"num_injected_attention_layers": 12,
"lora_params": 2973696,
"error": NaN,
"path": "adapted_roberta",
"timestamp": "2026-09-14T11:09:00.441134"
}
},
"results/unified_aegis_v4/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,78.56,0.8124,0.7811,0.8051,0.7832,0.4381,6.76,712.34,24513.32,124647170,124647170,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_roberta,2026-09-15T10:34:56.361686\naegis_v4,PASS_NO_COMPRESSION,57.8,0.707,0.4765,0.7734,0.5701,0.822,55.24,676.53,6149.86,124647170,133307282,1573008,1.0,0.0,0.0,0,0,,,12,1179648,\"note: zero compressed layers, model grew (LoRA only)\",adapted_roberta,2026-09-15T10:35:51.883509\n",
"results/unified_aegis_v4/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,90.71,0.9099,0.907,0.9075,0.9068,0.3028,3.18,182.94,46184.49,66955010,66955010,0,1.0,0.0,0.0,0,0,,,0,0,,legacy_bert,2026-09-15T10:33:45.307345\naegis_v4,PASS_NO_COMPRESSION,85.21,0.8652,0.8506,0.8617,0.8506,0.3733,28.4,193.86,13166.42,66955010,67741514,786504,1.0,0.0,0.0,0,0,,,6,589824,\"note: zero compressed layers, model grew (LoRA only)\",legacy_bert,2026-09-15T10:34:13.938965\n",
"results/unified_aegis_v4/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,79.36,0.7978,0.7935,0.7935,0.7935,0.4875,6.88,793.19,23723.71,124441344,124441344,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_gpt,2026-09-15T10:36:34.838587\naegis_v4,PASS_NO_COMPRESSION,70.99,0.7668,0.6915,0.7639,0.7056,0.54,38.97,590.29,9810.48,124441344,126014352,1573008,1.0,0.0,0.0,0,0,,,12,1179648,\"note: zero compressed layers, model grew (LoRA only)\",adapted_gpt,2026-09-15T10:37:14.085753\n",
"results/benchmark_summary_v2.csv": "track2_ssm,aegis_v4,attention,bert,sst2,PASS,,37.42,61.1,98.52,0.0,1.0,4583314,4583314,0.957,1634.5,0.0,0.6947,0.0,0.0,0.0,0.6923,\"{'loss': [0.6947, 0.6947, 0.6947, 0.6947, 0.6947], 'grad_norm': [0.6923, 0.6923, 0.6923, 0.6923, 0.6923], 'fwd_latency_ms': [37.1455, 37.319, 37.4351, 37.8074, 37.4073], 'bwd_latency_ms': [60.6002, 61.536, 60.8713, 60.9054, 61.5813]}\",2026-08-26 02:34:33\ntrack2_ssm,aegis_v4,attention,bert,mrpc,PASS,,37.64,60.99,98.63,0.0,1.0,4583314,4583314,0.957,1634.5,0.0,0.6777,0.0,0.0,0.0,0.9849,\"{'loss': [0.6777, 0.6777, 0.6777, 0.6777, 0.6777], 'grad_norm': [0.9849, 0.9849, 0.9849, 0.9849, 0.9849], 'fwd_latency_ms': [38.5106, 37.2804, 37.5215, 37.4298, 37.4714], 'bwd_latency_ms': [61.3304, 61.0832, 60.8007, 60.8332, 60.9166]}\",2026-08-26 02:34:34\ntrack2_ssm,aegis_v4,attention,roberta,sst2,PASS,,38.69,62.91,101.6,0.0,1.0,7110418,7110418,0.972,1644.4,0.0,0.6924,0.0,0.0,0.0,11.8915,\"{'loss': [0.6924, 0.6924, 0.6924, 0.6924, 0.6924], 'grad_norm': [11.8915, 11.8915, 11.8915, 11.8915, 11.8915], 'fwd_latency_ms': [37.0545, 39.5123, 39.483, 39.5832, 37.8029], 'bwd_latency_ms': [62.8689, 63.8982, 63.8458, 61.7935, 62.1213]}\",2026-08-26 02:34:35\ntrack2_ssm,aegis_v4,attention,roberta,mrpc,PASS,,37.72,61.24,98.96,0.0,1.0,7110418,7110418,0.972,1644.4,0.0,0.6776,0.0,0.0,0.0,1.0529,\"{'loss': [0.6776, 0.6776, 0.6776, 0.6776, 0.6776], 'grad_norm': [1.0529, 1.0529, 1.0529, 1.0529, 1.0529], 'fwd_latency_ms': [37.2019, 38.1854, 37.4981, 37.9086, 37.7903], 'bwd_latency_ms': [61.0419, 61.3341, 61.1951, 61.2405, 61.3996]}\",2026-08-26 02:34:37\ntrack2_ssm,aegis_v4,attention,t5,sst2,PASS,,37.78,61.68,99.46,0.0,1.0,4704656,4704656,0.958,1610.8,0.0,1.2808,0.0,0.0,0.0,15.8559,\"{'loss': [1.2808, 1.2808, 1.2808, 1.2808, 1.2808], 'grad_norm': [15.8559, 15.8559, 15.8559, 15.8559, 15.8559], 'fwd_latency_ms': [34.9635, 38.1998, 38.1541, 38.0964, 39.488], 'bwd_latency_ms': [61.5048, 61.3926, 61.4616, 62.2499, 61.8051]}\",2026-08-26 02:34:38\ntrack2_ssm,aegis_v4,attention,t5,mrpc,PASS,,38.11,62.52,100.63,0.0,1.0,4704656,4704656,0.958,1610.8,0.0,0.6772,0.0,0.0,0.0,6.6065,\"{'loss': [0.6772, 0.6772, 0.6772, 0.6772, 0.6772], 'grad_norm': [6.6065, 6.6065, 6.6065, 6.6065, 6.6065], 'fwd_latency_ms': [34.9891, 38.8351, 38.7443, 39.3542, 38.6195], 'bwd_latency_ms': [62.7693, 62.396, 62.3828, 63.1169, 61.9526]}\",2026-08-26 02:34:39",
"results/master_aggregated_results.csv": "aegis_v4,0,PASS,83.72,50.92,56.31,107.2,2026-09-11T04:00:02.471388\naegis_v4_module,0,PASS,83.72,50.92,56.31,106.2,2026-09-11T04:01:48.660423",
"results/run_quality_all_screening_summary.csv": "run_quality_20260827_052444,aegis_v4,DISTILBERT,SST2,PASS,50.92,0.0,50.92,50.92,1.0,0.6748,0.201,173.45,473.84,2248.2,25311.5,1450000000.0,2026-08-27T06:02:44.019394\nrun_quality_20260827_052444,aegis_v4_inject,DISTILBERT,SST2,PASS,50.92,0.0,50.92,50.92,1.0,0.6748,0.201,176.17,480.59,2248.2,24946.7,1450000000.0,2026-08-27T06:04:36.675998"
},
"notes": []
},
"aegis_v6": {
"module_name": "aegis_v6",
"source_file": "Codes/framework/attentions/aegis_v6.py",
"source_lines_count": 891,
"architecture": {
"class_name": "FactorisedLowRankProjection",
"base_class": "nn.Module",
"init_args": "self, **kwargs",
"registry_names": [
"aegis_v6",
"inject_aegis_v6",
"aegis_v6_inject",
"aegis_v6_attn"
]
},
"params": [
{
"name": "S",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.ones(self.svd_rank, device=device, dtype=dtype)"
},
{
"name": "A",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.zeros(out_features, lora_rank, device=device, dtype=dtype)"
},
{
"name": "B",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.zeros(lora_rank, in_features, device=device, dtype=dtype)"
},
{
"name": "bias",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.zeros(out_features, device=device, dtype=dtype)"
},
{
"name": "query_gate",
"type": "nn.Linear",
"declaration": "nn.Linear(head_dim, 1, bias=False)"
},
{
"name": "input_gate",
"type": "nn.Linear",
"declaration": "nn.Linear(hidden_size, num_heads, bias=True)"
},
{
"name": "hidden_proj",
"type": "nn.Linear",
"declaration": "nn.Linear(hidden_size, num_heads * head_dim, bias=False)"
},
{
"name": "proj",
"type": "nn.Linear",
"declaration": "nn.Linear(3 * head_dim, 1, bias=True)"
}
],
"metrics": {
"results/unified_aegis_v6/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": 7.06
},
"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": 22893.69
},
"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": "2026-09-15T10:51:39.502149"
},
"results/unified_aegis_v6/roberta/roberta-base_aegis_v6_full_metrics.json": {
"model": "roberta-base_aegis_v6",
"performance": {
"accuracy_pct": 59.98,
"f1_score": 0.6161,
"f1_macro": 0.599,
"precision": 0.5996,
"recall": 0.5992,
"val_loss": 0.6697,
"status": "PASS",
"elapsed_s": 54.56
},
"efficiency": {
"total_params_before": 124647170,
"total_params_after": 104722382,
"trainable_recovery_params": 1336524,
"compression_rate": 1.1903,
"compression_ratio_pct": 15.98,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 12,
"lora_params": 0,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 596.26,
"tokens_per_sec": 5763.15
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 559.6,
"L_256_vram_mb": 565.12,
"L_512_vram_mb": 595.83,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 514)"
},
"timestamp": "2026-09-15T10:52:34.036744"
},
"results/unified_aegis_v6/distilbert/distilbert_sst2_aegis_v6_full_metrics.json": {
"model": "distilbert_sst2_aegis_v6",
"performance": {
"accuracy_pct": 87.04,
"f1_score": 0.8791,
"f1_macro": 0.8697,
"precision": 0.8752,
"recall": 0.8694,
"val_loss": 0.336,
"status": "PASS",
"elapsed_s": 27.51
},
"efficiency": {
"total_params_before": 66955010,
"total_params_after": 53449064,
"trainable_recovery_params": 668262,
"compression_rate": 1.2527,
"compression_ratio_pct": 20.17,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 6,
"lora_params": 0,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 301.26,
"tokens_per_sec": 11663.21
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 262.36,
"L_256_vram_mb": 267.87,
"L_512_vram_mb": "unsupported (exceeds model max_position_embeddings 512)",
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 512)"
},
"timestamp": "2026-09-15T10:50:57.084632"
},
"results/unified_aegis_v6/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": 3.41
},
"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": 45229.11
},
"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": "2026-09-15T10:50:29.396557"
},
"results/unified_aegis_v6/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": 6.99
},
"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": 22868.32
},
"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": "2026-09-15T10:53:17.459771"
},
"results/unified_aegis_v6/gpt2/gpt2_aegis_v6_full_metrics.json": {
"model": "gpt2_aegis_v6",
"performance": {
"accuracy_pct": 51.61,
"f1_score": 0.6713,
"f1_macro": 0.377,
"precision": 0.5534,
"recall": 0.5076,
"val_loss": 0.6925,
"status": "PASS_LOW",
"elapsed_s": 49.84
},
"efficiency": {
"total_params_before": 124441344,
"total_params_after": 97429452,
"trainable_recovery_params": 1336524,
"compression_rate": 1.2772,
"compression_ratio_pct": 21.71,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 12,
"lora_params": 0,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 493.7,
"tokens_per_sec": 6170.91
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 453.7,
"L_256_vram_mb": 458.58,
"L_512_vram_mb": 491.99,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 1024)"
},
"timestamp": "2026-09-15T10:54:07.324142"
},
"results/unified_aegis_v6/distilbert/hf_export/metrics.json": {
"model_name": "distilbert",
"module": "aegis_v6",
"status": "PASS",
"accuracy_pct": 87.04,
"f1": 0.0,
"precision": 0.8752,
"recall": 0.8694,
"eval_loss": 0.0,
"params_total": null,
"params_active": null,
"param_reduction_pct": null,
"peak_vram_mb": 301.26,
"tokens_per_sec": 11663.21,
"elapsed_seconds": null,
"full_metrics": {
"module_name": "aegis_v6",
"status": "PASS",
"accuracy_pct": 87.04,
"f1_score": 0.8791,
"f1_macro": 0.8697,
"precision": 0.8752,
"recall": 0.8694,
"val_loss": 0.336,
"elapsed_s": 27.51,
"peak_vram_mb": 301.26,
"tokens_per_sec": 11663.21,
"total_params_before": 66955010,
"total_params_after": 53449064,
"trainable_recovery_params": 668262,
"compression_rate": 1.2527,
"compression_ratio_pct": 20.17,
"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": 0,
"error": NaN,
"path": "legacy_bert",
"timestamp": "2026-09-15T10:50:57.431568"
}
},
"results/unified_aegis_v6/roberta/hf_export/metrics.json": {
"model_name": "roberta",
"module": "aegis_v6",
"status": "PASS",
"accuracy_pct": 59.98,
"f1": 0.0,
"precision": 0.5996,
"recall": 0.5992,
"eval_loss": 0.0,
"params_total": null,
"params_active": null,
"param_reduction_pct": null,
"peak_vram_mb": 596.26,
"tokens_per_sec": 5763.15,
"elapsed_seconds": null,
"full_metrics": {
"module_name": "aegis_v6",
"status": "PASS",
"accuracy_pct": 59.98,
"f1_score": 0.6161,
"f1_macro": 0.599,
"precision": 0.5996,
"recall": 0.5992,
"val_loss": 0.6697,
"elapsed_s": 54.56,
"peak_vram_mb": 596.26,
"tokens_per_sec": 5763.15,
"total_params_before": 124647170,
"total_params_after": 104722382,
"trainable_recovery_params": 1336524,
"compression_rate": 1.1903,
"compression_ratio_pct": 15.98,
"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": 0,
"error": NaN,
"path": "adapted_roberta",
"timestamp": "2026-09-15T10:52:35.104245"
}
},
"results/unified_aegis_v6/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,78.56,0.8124,0.7811,0.8051,0.7832,0.4381,7.06,712.34,22893.69,124647170,124647170,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_roberta,2026-09-15T10:51:40.291119\naegis_v6,PASS,59.98,0.6161,0.599,0.5996,0.5992,0.6697,54.56,596.26,5763.15,124647170,104722382,1336524,1.1903,15.98,0.0,0,0,,,12,0,,adapted_roberta,2026-09-15T10:52:35.104245\n",
"results/unified_aegis_v6/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,90.71,0.9099,0.907,0.9075,0.9068,0.3027,3.41,389.51,45229.11,66955010,66955010,0,1.0,0.0,0.0,0,0,,,0,0,,legacy_bert,2026-09-15T10:50:29.663521\naegis_v6,PASS,87.04,0.8791,0.8697,0.8752,0.8694,0.336,27.51,301.26,11663.21,66955010,53449064,668262,1.2527,20.17,0.0,0,0,,,6,0,,legacy_bert,2026-09-15T10:50:57.431568\n",
"results/unified_aegis_v6/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,79.36,0.7978,0.7935,0.7935,0.7935,0.4875,6.99,793.19,22868.32,124441344,124441344,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_gpt,2026-09-15T10:53:18.234923\naegis_v6,PASS_LOW,51.61,0.6713,0.377,0.5534,0.5076,0.6925,49.84,493.7,6170.91,124441344,97429452,1336524,1.2772,21.71,0.0,0,0,,,12,0,note: evaluated accuracy (51.61%) below chance threshold (56.00%),adapted_gpt,2026-09-15T10:54:08.320001\n",
"results/benchmark_summary_v2.csv": "track2_ssm,aegis_v6,attention,bert,sst2,PASS,,48.06,75.2,123.26,0.0,1.0,4325782,4325782,1.014,1772.3,0.0,0.6929,0.0,0.0,0.0,1.308,\"{'loss': [0.6925, 0.6928, 0.6935, 0.693, 0.6927], 'grad_norm': [1.2524, 1.3228, 1.3753, 1.3249, 1.2645], 'fwd_latency_ms': [46.7272, 49.2839, 48.0622, 48.0325, 48.2013], 'bwd_latency_ms': [75.0353, 77.5766, 74.4084, 74.2935, 74.6888]}\",2026-08-26 02:34:46\ntrack2_ssm,aegis_v6,attention,bert,mrpc,PASS,,48.69,75.33,124.02,0.0,1.0,4325782,4325782,1.014,1772.3,0.0,0.6954,0.0,0.0,0.0,11.4719,\"{'loss': [0.6954, 0.6954, 0.6953, 0.6954, 0.6955], 'grad_norm': [11.4852, 11.4753, 11.4226, 11.4816, 11.4948], 'fwd_latency_ms': [48.6185, 49.5445, 48.6498, 48.6958, 47.9391], 'bwd_latency_ms': [74.4267, 76.5373, 76.6078, 74.2384, 74.8487]}\",2026-08-26 02:34:47\ntrack2_ssm,aegis_v6,attention,roberta,sst2,PASS,,48.45,75.42,123.87,0.0,1.0,6852886,6852886,1.009,1782.2,0.0,0.695,0.0,0.0,0.0,2.0271,\"{'loss': [0.695, 0.6951, 0.6949, 0.6951, 0.6949], 'grad_norm': [1.9807, 1.9701, 2.0978, 2.0295, 2.0572], 'fwd_latency_ms': [47.5907, 48.4665, 48.803, 48.8789, 48.523], 'bwd_latency_ms': [75.411, 75.6097, 75.0158, 75.4397, 75.6094]}\",2026-08-26 02:34:49\ntrack2_ssm,aegis_v6,attention,roberta,mrpc,PASS,,48.87,74.99,123.86,0.0,1.0,6852886,6852886,1.009,1782.2,0.0,0.6774,0.0,0.0,0.0,13.0772,\"{'loss': [0.6774, 0.6774, 0.6775, 0.6775, 0.6774], 'grad_norm': [13.0565, 13.0446, 13.0748, 13.152, 13.0579], 'fwd_latency_ms': [47.531, 48.3032, 49.1152, 50.6939, 48.72], 'bwd_latency_ms': [74.442, 75.1442, 75.1567, 75.0953, 75.1043]}\",2026-08-26 02:34:50\ntrack2_ssm,aegis_v6,attention,t5,sst2,PASS,,48.4,76.83,125.23,0.0,1.0,4447124,4447124,1.013,1820.7,0.0,0.7922,0.0,0.0,0.0,295.8876,\"{'loss': [0.7934, 0.7914, 0.7902, 0.7946, 0.7915], 'grad_norm': [296.1249, 295.3053, 294.8635, 295.9312, 297.2132], 'fwd_latency_ms': [47.999, 48.1539, 48.6287, 49.405, 47.8055], 'bwd_latency_ms': [75.8938, 78.4377, 76.6227, 76.8214, 76.3633]}\",2026-08-26 02:34:51\ntrack2_ssm,aegis_v6,attention,t5,mrpc,PASS,,47.77,75.63,123.4,0.0,1.0,4447124,4447124,1.013,1820.7,0.0,0.6662,0.0,0.0,0.0,189.3947,\"{'loss': [0.6662, 0.667, 0.6661, 0.6651, 0.6666], 'grad_norm': [189.3378, 187.843, 188.3937, 189.892, 191.5068], 'fwd_latency_ms': [47.1378, 47.9928, 48.0537, 48.1307, 47.5584], 'bwd_latency_ms': [75.7538, 75.5712, 75.3993, 75.6782, 75.7252]}\",2026-08-26 02:34:52",
"results/master_aggregated_results.csv": "aegis_v6,0,PASS_LOW,52.87,50.92,50.46,238.3,2026-09-11T11:03:55.367633",
"results/run_quality_all_screening_summary.csv": "run_quality_20260827_052444,aegis_v6,DISTILBERT,SST2,PASS,50.92,0.0,50.92,50.92,1.0,0.6748,0.1982,283.08,775.09,2990.9,15483.4,1450000000.0,2026-08-27T06:08:45.016231"
},
"notes": []
}
}
{
"aegis_v1": {
"module_name": "aegis_v1",
"source_file": "Codes/framework/attentions/aegis_v1.py",
"source_lines_count": 814,
"architecture": {
"class_name": "FactorisedLinear",
"base_class": "nn.Module",
"init_args": "self, in_features_or_weight: Union[int, torch.Tensor], out_features_or_bias: Optional[Union[int, torch.Tensor]] = None, energy: float = 0.95, min_rank: int = 4, max_rank: int = 384, rank: Optional[int] = None, bias: bool = True, lora_rank: int = 16, lora_alpha: float = 32.0, **kwargs: Any, ) -> None: super().__init__() self.energy = energy self.min_rank = min_rank self.max_rank = max_rank self.lora_rank = lora_rank self.lora_alpha = lora_alpha self.lora_scaling = (lora_alpha / lora_rank) if lora_rank > 0 else 1.0 if isinstance(in_features_or_weight, torch.Tensor",
"registry_names": [
"aegis_v1_inject",
"AegisAttentionV1",
"aegis_v1_module",
"aegis_v1",
"inject_aegis_v1"
]
},
"params": [
{
"name": "down",
"type": "nn.Linear",
"declaration": "nn.Linear(in_dim, self.rank, bias=False, device=device, dtype=dtype)"
},
{
"name": "up",
"type": "nn.Linear",
"declaration": "nn.Linear(self.rank, out_dim, bias=(b is not None or bias)"
},
{
"name": "lora_A",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.randn(lora_rank, in_dim, device=device, dtype=dtype)"
},
{
"name": "lora_B",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.zeros(out_dim, lora_rank, device=device, dtype=dtype)"
},
{
"name": "down",
"type": "nn.Linear",
"declaration": "nn.Linear(in_dim, self.rank, bias=False)"
},
{
"name": "up",
"type": "nn.Linear",
"declaration": "nn.Linear(self.rank, out_dim, bias=bias)"
},
{
"name": "lora_A",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.randn(lora_rank, in_dim)"
},
{
"name": "lora_B",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.zeros(out_dim, lora_rank)"
}
],
"metrics": {
"results/unified_aegis_v1/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": 7.1
},
"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": 23473.61
},
"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,
Xet Storage Details
- Size:
- 159 kB
- Xet hash:
- b9191af678c546c250614adf833c5315ee5a0462712a864b7252d786e58d8bc8
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.