{
"flat_llm": {
"module_name": "flat_llm",
"source_file": "Codes/framework/compression_engine/flat_llm.py",
"source_lines_count": 394,
"architecture": {
"class_name": "FLATLLMCompressor",
"base_class": "BaseCompressor",
"init_args": "self, model: nn.Module, tokenizer: Optional[Any] = None, keep_ratio: float = 0.7, damping: float = 1e-4, max_error: float = 0.35, target_attention_only: bool = False, fallback_to_original: bool = True, calibration_loader: Optional[Any] = None, calibration_texts: Optional[List[str]] = None, device: Optional[Union[str, torch.device]] = None, structured_24: bool = False, target_modules: Optional[List[str]] = None, cfg: Optional[Any] = None, **kwargs: Any, ) -> None: super().__init__(model, name=\"flat_llm\", **kwargs) cfg = cfg or kwargs.get(\"comp_config\", None) self.cfg = cfg self.tokenizer = tokenizer or getattr(model, \"tokenizer\", None) raw_kr = _cfg_get(cfg, \"flat_llm_keep_ratio\", _cfg_get(cfg, \"keep_ratio\", keep_ratio)) self.keep_ratio = max(0.05, min(float(raw_kr), 1.0)) self.damping = float(_cfg_get(cfg, \"flat_llm_damping\", _cfg_get(cfg, \"damping\", damping))) self.structured_24 = bool(_cfg_get(cfg, \"flat_llm_structured_24\", _cfg_get(cfg, \"structured_24\", structured_24))) self.max_error = float(_cfg_get(cfg, \"flat_llm_max_error\", _cfg_get(cfg, \"max_error\", max_error))) self.target_attention_only = bool( _cfg_get(cfg, \"flat_llm_target_attention_only\", _cfg_get(cfg, \"target_attention_only\", target_attention_only)) ) self.fallback_to_original = bool( _cfg_get(cfg, \"flat_llm_fallback_to_original\", _cfg_get(cfg, \"fallback_to_original\", fallback_to_original)) ) self.calibration_loader = calibration_loader or _cfg_get(cfg, \"calibration_loader\", None) self.calibration_texts = calibration_texts self.target_modules = target_modules self.target_device = ( torch.device(device) if device is not None else next(model.parameters(), torch.tensor(0.0)).device ) self.compressed_layers: List[str] = [] self.fallback_layers: List[str] = [] self.skipped_layers: List[str] = [] self.layer_errors: Dict[str, float] = {} self.layer_keep_ratios: Dict[str, float] = {} def _is_target_module(self, name: str, module: nn.Module) -> bool: name_lower = name.lower() if any(part in name.split(\".\") for part in [\"lora_A\", \"lora_B\", \"base_layer\"]",
"registry_names": [
"flat_llm_compressor",
"flat_llm",
"flat_llm_factory"
]
},
"params": [],
"metrics": {
"results/unified_flat_llm/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": 5.87
},
"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": 27816.89
},
"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": "2026-09-15T12:32:41.194584"
},
"results/unified_flat_llm/roberta/roberta-base_flat_llm_full_metrics.json": {
"model": "roberta-base_flat_llm",
"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_NO_COMPRESSION",
"elapsed_s": 9.02
},
"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": 72,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": 0.0,
"weighted_reconstruction_error": 0.5395,
"per_layer_reconstruction_errors": {
"roberta.encoder.layer.0.attention.self.query": 0.48081703763211586,
"roberta.encoder.layer.0.attention.self.key": 0.4678205793882954,
"roberta.encoder.layer.0.attention.self.value": 0.600996174719836,
"roberta.encoder.layer.0.attention.output.dense": 0.5184713960115622,
"roberta.encoder.layer.0.intermediate.dense": 0.5504354504181636,
"roberta.encoder.layer.0.output.dense": 0.5380613213290566,
"roberta.encoder.layer.1.attention.self.query": 0.5832522478629809,
"roberta.encoder.layer.1.attention.self.key": 0.6117967586196459,
"roberta.encoder.layer.1.attention.self.value": 0.4609478712698367,
"roberta.encoder.layer.1.attention.output.dense": 0.5265832299097785,
"roberta.encoder.layer.1.intermediate.dense": 0.5489612424907715,
"roberta.encoder.layer.1.output.dense": 0.5330325015193179,
"roberta.encoder.layer.2.attention.self.query": 0.5479160240952466,
"roberta.encoder.layer.2.attention.self.key": 0.5634047529319931,
"roberta.encoder.layer.2.attention.self.value": 0.5160298514406528,
"roberta.encoder.layer.2.attention.output.dense": 0.5345467451205721,
"roberta.encoder.layer.2.intermediate.dense": 0.5418580537223934,
"roberta.encoder.layer.2.output.dense": 0.5310974552567965,
"roberta.encoder.layer.3.attention.self.query": 0.5550407767232611,
"roberta.encoder.layer.3.attention.self.key": 0.5405874013567532,
"roberta.encoder.layer.3.attention.self.value": 0.5913338725415708,
"roberta.encoder.layer.3.attention.output.dense": 0.5317781994440105,
"roberta.encoder.layer.3.intermediate.dense": 0.5491246375176064,
"roberta.encoder.layer.3.output.dense": 0.5309126455135799,
"roberta.encoder.layer.4.attention.self.query": 0.4890126837884554,
"roberta.encoder.layer.4.attention.self.key": 0.4901286859782906,
"roberta.encoder.layer.4.attention.self.value": 0.5675768100282815,
"roberta.encoder.layer.4.attention.output.dense": 0.5278989343064305,
"roberta.encoder.layer.4.intermediate.dense": 0.5543334679447507,
"roberta.encoder.layer.4.output.dense": 0.529903285188956,
"roberta.encoder.layer.5.attention.self.query": 0.526445339455892,
"roberta.encoder.layer.5.attention.self.key": 0.5387497693454998,
"roberta.encoder.layer.5.attention.self.value": 0.5614703883476401,
"roberta.encoder.layer.5.attention.output.dense": 0.5311561447395433,
"roberta.encoder.layer.5.intermediate.dense": 0.5453480223638519,
"roberta.encoder.layer.5.output.dense": 0.5287185259870026,
"roberta.encoder.layer.6.attention.self.query": 0.5883239257593531,
"roberta.encoder.layer.6.attention.self.key": 0.5916999268770622,
"roberta.encoder.layer.6.attention.self.value": 0.5311849676931587,
"roberta.encoder.layer.6.attention.output.dense": 0.5525668671502056,
"roberta.encoder.layer.6.intermediate.dense": 0.5500582346147186,
"roberta.encoder.layer.6.output.dense": 0.5307234395256171,
"roberta.encoder.layer.7.attention.self.query": 0.518280522688447,
"roberta.encoder.layer.7.attention.self.key": 0.5129379178034653,
"roberta.encoder.layer.7.attention.self.value": 0.5627828134834333,
"roberta.encoder.layer.7.attention.output.dense": 0.5294931080471172,
"roberta.encoder.layer.7.intermediate.dense": 0.5526663855152963,
"roberta.encoder.layer.7.output.dense": 0.524426411055563,
"roberta.encoder.layer.8.attention.self.query": 0.5474589434670835,
"roberta.encoder.layer.8.attention.self.key": 0.5454838735098068,
"roberta.encoder.layer.8.attention.self.value": 0.5475032849230957,
"roberta.encoder.layer.8.attention.output.dense": 0.5300102369615962,
"roberta.encoder.layer.8.intermediate.dense": 0.5551282287536783,
"roberta.encoder.layer.8.output.dense": 0.5278662949241739,
"roberta.encoder.layer.9.attention.self.query": 0.5510294653295433,
"roberta.encoder.layer.9.attention.self.key": 0.5492205623939821,
"roberta.encoder.layer.9.attention.self.value": 0.5789217200578,
"roberta.encoder.layer.9.attention.output.dense": 0.5301037495666887,
"roberta.encoder.layer.9.intermediate.dense": 0.5430975141591224,
"roberta.encoder.layer.9.output.dense": 0.5291312117878981,
"roberta.encoder.layer.10.attention.self.query": 0.5430404535258238,
"roberta.encoder.layer.10.attention.self.key": 0.5439175943042216,
"roberta.encoder.layer.10.attention.self.value": 0.515404219276337,
"roberta.encoder.layer.10.attention.output.dense": 0.5342768136383318,
"roberta.encoder.layer.10.intermediate.dense": 0.5493738597337353,
"roberta.encoder.layer.10.output.dense": 0.5283292770979253,
"roberta.encoder.layer.11.attention.self.query": 0.5479341233065377,
"roberta.encoder.layer.11.attention.self.key": 0.5512582646630475,
"roberta.encoder.layer.11.attention.self.value": 0.5202785233614646,
"roberta.encoder.layer.11.attention.output.dense": 0.5419589116084491,
"roberta.encoder.layer.11.intermediate.dense": 0.5474309325207576,
"roberta.encoder.layer.11.output.dense": 0.5209231229010328
},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 520.18,
"tokens_per_sec": 28454.01
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 504.62,
"L_256_vram_mb": 507.18,
"L_512_vram_mb": 512.31,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 514)"
},
"timestamp": "2026-09-15T12:32:50.397514"
},
"results/unified_flat_llm/distilbert/distilbert_sst2_flat_llm_full_metrics.json": {
"model": "distilbert_sst2_flat_llm",
"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_NO_COMPRESSION",
"elapsed_s": 4.46
},
"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": 36,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": 0.0,
"weighted_reconstruction_error": 0.5512,
"per_layer_reconstruction_errors": {
"distilbert.transformer.layer.0.attention.q_lin": 0.5659718718329096,
"distilbert.transformer.layer.0.attention.k_lin": 0.574049223556988,
"distilbert.transformer.layer.0.attention.v_lin": 0.5414140106595501,
"distilbert.transformer.layer.0.attention.out_lin": 0.5648154095065941,
"distilbert.transformer.layer.0.ffn.lin1": 0.5471973811569976,
"distilbert.transformer.layer.0.ffn.lin2": 0.5574499265009167,
"distilbert.transformer.layer.1.attention.q_lin": 0.5959530756967731,
"distilbert.transformer.layer.1.attention.k_lin": 0.5961059112685638,
"distilbert.transformer.layer.1.attention.v_lin": 0.527276485429654,
"distilbert.transformer.layer.1.attention.out_lin": 0.5561464883864395,
"distilbert.transformer.layer.1.ffn.lin1": 0.5453296668232669,
"distilbert.transformer.layer.1.ffn.lin2": 0.551031024048419,
"distilbert.transformer.layer.2.attention.q_lin": 0.5584381601466808,
"distilbert.transformer.layer.2.attention.k_lin": 0.5570362885445677,
"distilbert.transformer.layer.2.attention.v_lin": 0.5470735244290527,
"distilbert.transformer.layer.2.attention.out_lin": 0.5581598794039241,
"distilbert.transformer.layer.2.ffn.lin1": 0.5534930450135527,
"distilbert.transformer.layer.2.ffn.lin2": 0.5490013263543544,
"distilbert.transformer.layer.3.attention.q_lin": 0.5282794514757287,
"distilbert.transformer.layer.3.attention.k_lin": 0.5290897257359016,
"distilbert.transformer.layer.3.attention.v_lin": 0.5741988746526885,
"distilbert.transformer.layer.3.attention.out_lin": 0.5528623525620557,
"distilbert.transformer.layer.3.ffn.lin1": 0.54723429157497,
"distilbert.transformer.layer.3.ffn.lin2": 0.5498233507270264,
"distilbert.transformer.layer.4.attention.q_lin": 0.5468585045416045,
"distilbert.transformer.layer.4.attention.k_lin": 0.5489145869644949,
"distilbert.transformer.layer.4.attention.v_lin": 0.5442924639138077,
"distilbert.transformer.layer.4.attention.out_lin": 0.5469975597730721,
"distilbert.transformer.layer.4.ffn.lin1": 0.5392504284339205,
"distilbert.transformer.layer.4.ffn.lin2": 0.5625310407350022,
"distilbert.transformer.layer.5.attention.q_lin": 0.5435958738857206,
"distilbert.transformer.layer.5.attention.k_lin": 0.5435670207873065,
"distilbert.transformer.layer.5.attention.v_lin": 0.5426887979308647,
"distilbert.transformer.layer.5.attention.out_lin": 0.5478216840365683,
"distilbert.transformer.layer.5.ffn.lin1": 0.5476384170489461,
"distilbert.transformer.layer.5.ffn.lin2": 0.5496197908325922
},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 300.68,
"tokens_per_sec": 53197.92
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 285.62,
"L_256_vram_mb": 288.19,
"L_512_vram_mb": "unsupported (exceeds model max_position_embeddings 512)",
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 512)"
},
"timestamp": "2026-09-15T12:32:12.779345"
},
"results/unified_flat_llm/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": 2.89
},
"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": 53606.92
},
"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": "2026-09-15T12:32:08.112273"
},
"results/unified_flat_llm/gpt2/gpt2_flat_llm_full_metrics.json": {
"model": "gpt2_flat_llm",
"performance": {
"accuracy_pct": 79.36,
"f1_score": 0.7978,
"f1_macro": 0.7935,
"precision": 0.7935,
"recall": 0.7935,
"val_loss": 0.4875,
"status": "PASS_NO_COMPRESSION",
"elapsed_s": 9.16
},
"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": 48,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": 0.0,
"weighted_reconstruction_error": 0.5528,
"per_layer_reconstruction_errors": {
"transformer.h.0.attn.c_attn": 0.6774935166355116,
"transformer.h.0.attn.c_proj": 0.5640019420709369,
"transformer.h.0.mlp.c_fc": 0.5470148191891724,
"transformer.h.0.mlp.c_proj": 0.5488537186328826,
"transformer.h.1.attn.c_attn": 0.5823175073962819,
"transformer.h.1.attn.c_proj": 0.5172952912758293,
"transformer.h.1.mlp.c_fc": 0.5557861891401422,
"transformer.h.1.mlp.c_proj": 0.5549091523096368,
"transformer.h.2.attn.c_attn": 0.7086699615318554,
"transformer.h.2.attn.c_proj": 0.5398420541916394,
"transformer.h.2.mlp.c_fc": 0.5413749422250856,
"transformer.h.2.mlp.c_proj": 0.5331907620008892,
"transformer.h.3.attn.c_attn": 0.6389096755880905,
"transformer.h.3.attn.c_proj": 0.5354109318006217,
"transformer.h.3.mlp.c_fc": 0.5483258166208166,
"transformer.h.3.mlp.c_proj": 0.5258512836749476,
"transformer.h.4.attn.c_attn": 0.5702221924208564,
"transformer.h.4.attn.c_proj": 0.5315352859354037,
"transformer.h.4.mlp.c_fc": 0.5460088084910387,
"transformer.h.4.mlp.c_proj": 0.5322945051187795,
"transformer.h.5.attn.c_attn": 0.6111341004525097,
"transformer.h.5.attn.c_proj": 0.5349023520820994,
"transformer.h.5.mlp.c_fc": 0.5489664764129568,
"transformer.h.5.mlp.c_proj": 0.5349758006815318,
"transformer.h.6.attn.c_attn": 0.5787893611567113,
"transformer.h.6.attn.c_proj": 0.5336609814081009,
"transformer.h.6.mlp.c_fc": 0.5501376484805612,
"transformer.h.6.mlp.c_proj": 0.5335717213099581,
"transformer.h.7.attn.c_attn": 0.5744195171301719,
"transformer.h.7.attn.c_proj": 0.5323453928521908,
"transformer.h.7.mlp.c_fc": 0.545718911719712,
"transformer.h.7.mlp.c_proj": 0.5366213474450418,
"transformer.h.8.attn.c_attn": 0.563429820516885,
"transformer.h.8.attn.c_proj": 0.5289816121219886,
"transformer.h.8.mlp.c_fc": 0.5463407190032545,
"transformer.h.8.mlp.c_proj": 0.5382752074124773,
"transformer.h.9.attn.c_attn": 0.530023847564622,
"transformer.h.9.attn.c_proj": 0.5309222798393954,
"transformer.h.9.mlp.c_fc": 0.5491462110801424,
"transformer.h.9.mlp.c_proj": 0.5397554983428796,
"transformer.h.10.attn.c_attn": 0.5073025746663753,
"transformer.h.10.attn.c_proj": 0.5202981831086483,
"transformer.h.10.mlp.c_fc": 0.5472513166804839,
"transformer.h.10.mlp.c_proj": 0.5389548404567124,
"transformer.h.11.attn.c_attn": 0.4661294006323907,
"transformer.h.11.attn.c_proj": 0.5374830956393384,
"transformer.h.11.mlp.c_fc": 0.5474166830238908,
"transformer.h.11.mlp.c_proj": 0.5515521443232712
},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 630.16,
"tokens_per_sec": 27417.97
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 510.91,
"L_256_vram_mb": 524.41,
"L_512_vram_mb": 552.17,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 1024)"
},
"timestamp": "2026-09-15T12:33:28.652492"
},
"results/unified_flat_llm/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": 5.91
},
"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": 27570.33
},
"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": "2026-09-15T12:33:19.292846"
},
"results/unified_flat_llm/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,78.56,0.8124,0.7811,0.8051,0.7832,0.4381,5.87,712.34,27816.89,124647170,124647170,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_roberta,2026-09-15T12:32:41.883910\nflat_llm,PASS_NO_COMPRESSION,78.56,0.8124,0.7811,0.8051,0.7832,0.4381,9.03,520.18,28454.01,124647170,124647170,0,1.0,0.0,0.0,0,72,0.0,0.5395,0,0,\"note: zero compressed layers, model grew (LoRA only)\",adapted_roberta,2026-09-15T12:32:51.144859\n",
"results/unified_flat_llm/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,90.71,0.9099,0.907,0.9075,0.9068,0.3027,2.9,389.51,53606.92,66955010,66955010,0,1.0,0.0,0.0,0,0,,,0,0,,legacy_bert,2026-09-15T12:32:08.325766\nflat_llm,PASS_NO_COMPRESSION,90.71,0.9099,0.907,0.9075,0.9068,0.3027,4.46,300.68,53197.92,66955010,66955010,0,1.0,0.0,0.0,0,36,0.0,0.5512,0,0,\"note: zero compressed layers, model grew (LoRA only)\",legacy_bert,2026-09-15T12:32:13.009670\n",
"results/unified_flat_llm/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,79.36,0.7978,0.7935,0.7935,0.7935,0.4875,5.91,793.19,27570.33,124441344,124441344,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_gpt,2026-09-15T12:33:19.990120\nflat_llm,PASS_NO_COMPRESSION,79.36,0.7978,0.7935,0.7935,0.7935,0.4875,9.16,630.16,27417.97,124441344,124441344,0,1.0,0.0,0.0,0,48,0.0,0.5528,0,0,\"note: zero compressed layers, model grew (LoRA only)\",adapted_gpt,2026-09-15T12:33:29.387229\n",
"results/benchmark_summary_v2.csv": "track5_pruning,flat_llm,compressor,bert,sst2,PASS,,10.45,15.24,25.69,0.0,1.0,4386178,4386178,1.0,571.8,0.0,0.6925,0.0,0.0,0.0,0.0647,\"{'loss': [0.6925, 0.6925, 0.6925, 0.6925, 0.6925], 'grad_norm': [0.0647, 0.0647, 0.0647, 0.0647, 0.0647], 'fwd_latency_ms': [10.0925, 10.7067, 10.7593, 9.8914, 10.8183], 'bwd_latency_ms': [15.2705, 15.0293, 15.0545, 16.0744, 14.7758]}\",2026-08-26 03:48:11\ntrack5_pruning,flat_llm,compressor,bert,mrpc,PASS,,10.7,14.57,25.27,0.0,1.0,4386178,4386178,1.0,571.8,0.0,0.686,0.0,0.0,0.0,0.8682,\"{'loss': [0.686, 0.686, 0.686, 0.686, 0.686], 'grad_norm': [0.8682, 0.8682, 0.8682, 0.8682, 0.8682], 'fwd_latency_ms': [10.4886, 11.0163, 10.9035, 10.3867, 10.7255], 'bwd_latency_ms': [14.852, 13.8116, 15.3383, 14.1472, 14.6846]}\",2026-08-26 03:48:12\ntrack5_pruning,flat_llm,compressor,roberta,sst2,PASS,,11.53,15.74,27.27,0.0,1.0,6913282,6913282,1.0,581.7,0.0,0.6932,0.0,0.0,0.0,0.1095,\"{'loss': [0.6932, 0.6932, 0.6932, 0.6932, 0.6932], 'grad_norm': [0.1095, 0.1095, 0.1095, 0.1095, 0.1095], 'fwd_latency_ms': [11.3941, 11.8283, 11.0667, 11.7027, 11.6647], 'bwd_latency_ms': [15.6976, 15.5392, 15.9552, 15.7073, 15.8162]}\",2026-08-26 03:48:12\ntrack5_pruning,flat_llm,compressor,roberta,mrpc,PASS,,12.2,16.81,29.01,0.0,1.0,6913282,6913282,1.0,581.7,0.0,0.6936,0.0,0.0,0.0,0.9335,\"{'loss': [0.6936, 0.6936, 0.6936, 0.6936, 0.6936], 'grad_norm': [0.9335, 0.9335, 0.9335, 0.9335, 0.9335], 'fwd_latency_ms': [11.691, 12.0113, 12.1964, 12.4119, 12.7008], 'bwd_latency_ms': [15.4987, 16.713, 18.1492, 17.7676, 15.934]}\",2026-08-26 03:48:13\ntrack5_pruning,flat_llm,compressor,t5,sst2,PASS,,19.09,33.98,53.07,0.0,1.0,4506368,4506368,1.0,1107.2,0.0,0.7225,0.0,0.0,0.0,2.3611,\"{'loss': [0.7225, 0.7225, 0.7225, 0.7225, 0.7225], 'grad_norm': [2.3611, 2.3611, 2.3611, 2.3611, 2.3611], 'fwd_latency_ms': [15.8761, 19.8705, 19.8557, 19.984, 19.8757], 'bwd_latency_ms': [35.3727, 34.7067, 33.3678, 33.1564, 33.2785]}\",2026-08-26 03:48:14\ntrack5_pruning,flat_llm,compressor,t5,mrpc,PASS,,19.0,32.97,51.97,0.0,1.0,4506368,4506368,1.0,1107.2,0.0,0.7743,0.0,0.0,0.0,4.9226,\"{'loss': [0.7743, 0.7743, 0.7743, 0.7743, 0.7743], 'grad_norm': [4.9226, 4.9226, 4.9226, 4.9226, 4.9226], 'fwd_latency_ms': [16.001, 19.7539, 19.7218, 19.6994, 19.8318], 'bwd_latency_ms': [33.0446, 32.9805, 33.0516, 32.9526, 32.7997]}\",2026-08-26 03:48:15",
"results/all_121_compressors_gpu_matrix.csv": "flat_llm,cuda,PASS,True,True,1.0,127.0,0.175,\nflat_llm_compressor,cuda,PASS,True,True,1.0,127.0,0.176,",
"results/run_quality_all_screening_summary.csv": "run_quality_20260827_052444,flat_llm,DISTILBERT,SST2,PASS,80.85,0.0,80.85,80.85,1.0,0.81,0.6411,71.44,189.36,1848.6,62821.6,1450000000.0,2026-08-27T07:28:19.149768"
},
"notes": []
},
"frandi": {
"module_name": "frandi",
"source_file": "Codes/framework/compression_engine/frandi.py",
"source_lines_count": 16,
"architecture": {
"class_name": null,
"base_class": null,
"init_args": null,
"registry_names": []
},
"params": [],
"metrics": {
"results/unified_frandi/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": 5.82
},
"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": 28082.26
},
"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": "2026-09-15T12:37:19.009939"
},
"results/unified_frandi/roberta/roberta-base_frandi_full_metrics.json": {
"model": "roberta-base_frandi",
"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_NO_COMPRESSION",
"elapsed_s": 10.71
},
"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": 520.68,
"tokens_per_sec": 28218.15
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 504.62,
"L_256_vram_mb": 507.18,
"L_512_vram_mb": 512.31,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 514)"
},
"timestamp": "2026-09-15T12:37:29.768226"
},
"results/unified_frandi/distilbert/distilbert_sst2_frandi_full_metrics.json": {
"model": "distilbert_sst2_frandi",
"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_NO_COMPRESSION",
"elapsed_s": 5.48
},
"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": 301.18,
"tokens_per_sec": 54275.05
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 285.62,
"L_256_vram_mb": 288.19,
"L_512_vram_mb": "unsupported (exceeds model max_position_embeddings 512)",
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 512)"
},
"timestamp": "2026-09-15T12:36:49.650113"
},
"results/unified_frandi/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": 2.97
},
"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": 51948.53
},
"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": "2026-09-15T12:36:43.981074"
},
"results/unified_frandi/gpt2/gpt2_frandi_full_metrics.json": {
"model": "gpt2_frandi",
"performance": {
"accuracy_pct": 79.36,
"f1_score": 0.7978,
"f1_macro": 0.7935,
"precision": 0.7935,
"recall": 0.7935,
"val_loss": 0.4875,
"status": "PASS_NO_COMPRESSION",
"elapsed_s": 9.11
},
"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": 627.16,
"tokens_per_sec": 28528.38
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 510.91,
"L_256_vram_mb": 524.41,
"L_512_vram_mb": 552.17,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 1024)"
},
"timestamp": "2026-09-15T12:38:08.731568"
},
"results/unified_frandi/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": 5.87
},
"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": 27635.78
},
"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": "2026-09-15T12:37:59.578564"
},
"results/unified_frandi/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,78.56,0.8124,0.7811,0.8051,0.7832,0.4381,5.82,712.34,28082.26,124647170,124647170,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_roberta,2026-09-15T12:37:19.689443\nfrandi,PASS_NO_COMPRESSION,78.56,0.8124,0.7811,0.8051,0.7832,0.4381,10.71,520.68,28218.15,124647170,124647170,0,1.0,0.0,0.0,0,0,,,0,0,\"note: zero compressed layers, model grew (LoRA only)\",adapted_roberta,2026-09-15T12:37:30.644485\n",
"results/unified_frandi/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,90.71,0.9099,0.907,0.9075,0.9068,0.3027,2.97,389.51,51948.53,66955010,66955010,0,1.0,0.0,0.0,0,0,,,0,0,,legacy_bert,2026-09-15T12:36:44.189307\nfrandi,PASS_NO_COMPRESSION,90.71,0.9099,0.907,0.9075,0.9068,0.3027,5.48,301.18,54275.05,66955010,66955010,0,1.0,0.0,0.0,0,0,,,0,0,\"note: zero compressed layers, model grew (LoRA only)\",legacy_bert,2026-09-15T12:36:49.925307\n",
"results/unified_frandi/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,79.36,0.7978,0.7935,0.7935,0.7935,0.4875,5.87,793.19,27635.78,124441344,124441344,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_gpt,2026-09-15T12:38:00.260066\nfrandi,PASS_NO_COMPRESSION,79.36,0.7978,0.7935,0.7935,0.7935,0.4875,9.11,627.16,28528.38,124441344,124441344,0,1.0,0.0,0.0,0,0,,,0,0,\"note: zero compressed layers, model grew (LoRA only)\",adapted_gpt,2026-09-15T12:38:09.495568\n",
"results/benchmark_summary_v2.csv": "track1_svd,frandi,compressor,bert,sst2,PASS,,14.23,21.13,35.36,0.0,1.0,4054402,4054402,1.082,581.9,0.0,0.6928,0.0,0.0,0.0,0.0701,\"{'loss': [0.6928, 0.6928, 0.6928, 0.6928, 0.6928], 'grad_norm': [0.0701, 0.0701, 0.0701, 0.0701, 0.0701], 'fwd_latency_ms': [14.4581, 14.1817, 13.9358, 13.6173, 14.9604], 'bwd_latency_ms': [20.0883, 20.9675, 20.7416, 22.9303, 20.9417]}\",2026-08-26 02:33:59\ntrack1_svd,frandi,compressor,bert,mrpc,PASS,,14.0,20.99,34.99,0.0,1.0,4054402,4054402,1.082,581.9,0.0,0.69,0.0,0.0,0.0,0.7128,\"{'loss': [0.69, 0.69, 0.69, 0.69, 0.69], 'grad_norm': [0.7128, 0.7128, 0.7128, 0.7128, 0.7128], 'fwd_latency_ms': [14.1148, 14.279, 14.631, 13.4195, 13.5318], 'bwd_latency_ms': [20.452, 20.7031, 21.1853, 20.9854, 21.6173]}\",2026-08-26 02:34:00\ntrack1_svd,frandi,compressor,roberta,sst2,PASS,,14.36,21.14,35.5,0.0,1.0,6593794,6593794,1.048,591.9,0.0,0.694,0.0,0.0,0.0,0.1935,\"{'loss': [0.694, 0.694, 0.694, 0.694, 0.694], 'grad_norm': [0.1935, 0.1935, 0.1935, 0.1935, 0.1935], 'fwd_latency_ms': [14.1046, 14.7071, 14.4485, 14.1485, 14.4038], 'bwd_latency_ms': [21.5788, 20.514, 21.2877, 21.5634, 20.7694]}\",2026-08-26 02:34:00\ntrack1_svd,frandi,compressor,roberta,mrpc,PASS,,13.86,21.35,35.21,0.0,1.0,6593794,6593794,1.048,591.9,0.0,0.6996,0.0,0.0,0.0,1.3032,\"{'loss': [0.6996, 0.6996, 0.6996, 0.6996, 0.6996], 'grad_norm': [1.3032, 1.3032, 1.3032, 1.3032, 1.3032], 'fwd_latency_ms': [13.7811, 14.1651, 13.8001, 13.781, 13.7912], 'bwd_latency_ms': [20.8428, 21.0845, 21.4308, 20.6661, 22.7049]}\",2026-08-26 02:34:01\ntrack1_svd,frandi,compressor,t5,sst2,PASS,,17.91,31.99,49.9,0.0,1.0,4186880,4186880,1.076,1082.2,0.0,0.7776,0.0,0.0,0.0,1.7821,\"{'loss': [0.7776, 0.7776, 0.7776, 0.7776, 0.7776], 'grad_norm': [1.7821, 1.7821, 1.7821, 1.7821, 1.7821], 'fwd_latency_ms': [14.3032, 18.9018, 19.1218, 18.5323, 18.706], 'bwd_latency_ms': [31.9731, 32.0983, 31.2927, 32.6274, 31.9526]}\",2026-08-26 02:34:02\ntrack1_svd,frandi,compressor,t5,mrpc,PASS,,17.76,32.03,49.79,0.0,1.0,4186880,4186880,1.076,1082.2,0.0,0.8314,0.0,0.0,0.0,2.7308,\"{'loss': [0.8314, 0.8314, 0.8314, 0.8314, 0.8314], 'grad_norm': [2.7308, 2.7308, 2.7308, 2.7308, 2.7308], 'fwd_latency_ms': [14.2653, 18.8894, 18.2182, 18.5456, 18.8704], 'bwd_latency_ms': [32.1958, 31.9625, 31.9977, 31.8642, 32.1315]}\",2026-08-26 02:34:03",
"results/all_121_compressors_gpu_matrix.csv": "frandi,cuda,PASS,True,True,30.96,106.1,0.791,\nfrandi_compressor,cuda,PASS,True,True,30.96,106.1,0.78,",
"results/run_quality_all_screening_summary.csv": "run_quality_20260827_052444,frandi,DISTILBERT,SST2,PASS,64.22,0.0,64.22,64.22,1.0,0.6277,0.5049,46.21,132.54,1352.4,91661.3,1450000000.0,2026-08-27T07:29:48.680266\nrun_quality_20260827_052444,frandi_compressor,DISTILBERT,SST2,PASS,63.99,0.0,63.99,63.99,1.0,0.634,0.3982,48.15,132.38,1352.4,90755.9,1450000000.0,2026-08-27T07:31:18.165504"
},
"notes": []
},
"full_attention": {
"module_name": "full_attention",
"source_file": "Codes/framework/attentions/full_attention.py",
"source_lines_count": 511,
"architecture": {
"class_name": "LoRALinear",
"base_class": "nn.Module",
"init_args": "self, in_features: int, out_features: int, bias: bool = True, lora_rank: int = 16, lora_alpha: float = 32.0, freeze_base: bool = False, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None, ) -> None: super().__init__() factory_kwargs = {\"device\": device, \"dtype\": dtype} self.in_features = in_features self.out_features = out_features 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.freeze_base = freeze_base self.weight = nn.Parameter(torch.empty((out_features, in_features), **factory_kwargs)) if bias: self.bias = nn.Parameter(torch.empty(out_features, **factory_kwargs)) else: self.register_parameter(\"bias\", None) if self.lora_rank > 0: self.lora_A = nn.Parameter(torch.empty((lora_rank, in_features), **factory_kwargs)) self.lora_B = nn.Parameter(torch.empty((out_features, lora_rank), **factory_kwargs)) else: self.register_parameter(\"lora_A\", None) self.register_parameter(\"lora_B\", None) self.reset_parameters() if freeze_base: self.freeze_base_weights() def reset_parameters(self) -> None: nn.init.kaiming_uniform_(self.weight, a=math.sqrt(5)) if self.bias is not None: fan_in, _ = nn.init._calculate_fan_in_and_fan_out(self.weight) bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0 nn.init.uniform_(self.bias, -bound, bound) 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 freeze_base_weights(self) -> None: self.weight.requires_grad = False if self.bias is not None: self.bias.requires_grad = False if self.lora_rank > 0: if self.lora_A is not None: self.lora_A.requires_grad = True if self.lora_B is not None: self.lora_B.requires_grad = True def forward(self, x: torch.Tensor) -> torch.Tensor: out = F.linear(x, self.weight, self.bias) if self.lora_rank > 0 and self.lora_A is not None and self.lora_B is not None: out = out + F.linear(F.linear(x, self.lora_A), self.lora_B) * self.lora_scaling return out @ATTENTION_REGISTRY.register(\"full\", supports_causal=True, supports_bidirectional=True, supports_gqa=True) @ATTENTION_REGISTRY.register(\"full_attention\", supports_causal=True, supports_bidirectional=True, supports_gqa=True) class FullAttention(BaseAttention",
"registry_names": [
"full_attention",
"full",
"full_attention_inject"
]
},
"params": [
{
"name": "weight",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.empty((out_features, in_features)"
},
{
"name": "bias",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.empty(out_features, **factory_kwargs)"
},
{
"name": "lora_A",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.empty((lora_rank, in_features)"
},
{
"name": "lora_B",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.empty((out_features, lora_rank)"
}
],
"metrics": {
"results/unified_full/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": 6.25
},
"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": 26612.78
},
"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": "2026-09-15T12:39:16.274419"
},
"results/unified_full/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": 2.91
},
"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": 388.01,
"tokens_per_sec": 53818.36
},
"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": "2026-09-15T12:38:41.473314"
},
"results/unified_full/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": 6.24
},
"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": 791.56,
"tokens_per_sec": 26323.5
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 681.69,
"L_256_vram_mb": 690.32,
"L_512_vram_mb": 720.95,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 1024)"
},
"timestamp": "2026-09-15T12:39:52.138306"
},
"results/unified_full/gpt2/hf_export/metrics.json": {
"model_name": "gpt2",
"module": "full",
"status": "PASS",
"accuracy_pct": 79.36,
"f1": 0.0,
"precision": 0.7935,
"recall": 0.7935,
"eval_loss": 0.0,
"params_total": null,
"params_active": null,
"param_reduction_pct": null,
"peak_vram_mb": 791.56,
"tokens_per_sec": 26323.5,
"elapsed_seconds": null,
"full_metrics": {
"module_name": "full",
"status": "PASS",
"accuracy_pct": 79.36,
"f1_score": 0.7978,
"f1_macro": 0.7935,
"precision": 0.7935,
"recall": 0.7935,
"val_loss": 0.4875,
"elapsed_s": 6.24,
"peak_vram_mb": 791.56,
"tokens_per_sec": 26323.5,
"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,
"avg_reconstruction_error": NaN,
"weighted_reconstruction_error": NaN,
"num_injected_attention_layers": 0,
"lora_params": 0,
"error": NaN,
"path": "adapted_gpt",
"timestamp": "2026-09-15T12:39:52.918492"
}
},
"results/unified_full/distilbert/hf_export/metrics.json": {
"model_name": "distilbert",
"module": "full",
"status": "PASS",
"accuracy_pct": 90.71,
"f1": 0.0,
"precision": 0.9075,
"recall": 0.9068,
"eval_loss": 0.0,
"params_total": null,
"params_active": null,
"param_reduction_pct": null,
"peak_vram_mb": 388.01,
"tokens_per_sec": 53818.36,
"elapsed_seconds": null,
"full_metrics": {
"module_name": "full",
"status": "PASS",
"accuracy_pct": 90.71,
"f1_score": 0.9099,
"f1_macro": 0.907,
"precision": 0.9075,
"recall": 0.9068,
"val_loss": 0.3027,
"elapsed_s": 2.91,
"peak_vram_mb": 388.01,
"tokens_per_sec": 53818.36,
"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,
"avg_reconstruction_error": NaN,
"weighted_reconstruction_error": NaN,
"num_injected_attention_layers": 0,
"lora_params": 0,
"error": NaN,
"path": "legacy_bert",
"timestamp": "2026-09-15T12:38:41.760113"
}
},
"results/unified_full/roberta/hf_export/metrics.json": {
"model_name": "roberta",
"module": "full",
"status": "PASS",
"accuracy_pct": 78.56,
"f1": 0.0,
"precision": 0.8051,
"recall": 0.7832,
"eval_loss": 0.0,
"params_total": null,
"params_active": null,
"param_reduction_pct": null,
"peak_vram_mb": 712.34,
"tokens_per_sec": 26612.78,
"elapsed_seconds": null,
"full_metrics": {
"module_name": "full",
"status": "PASS",
"accuracy_pct": 78.56,
"f1_score": 0.8124,
"f1_macro": 0.7811,
"precision": 0.8051,
"recall": 0.7832,
"val_loss": 0.4381,
"elapsed_s": 6.25,
"peak_vram_mb": 712.34,
"tokens_per_sec": 26612.78,
"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,
"avg_reconstruction_error": NaN,
"weighted_reconstruction_error": NaN,
"num_injected_attention_layers": 0,
"lora_params": 0,
"error": NaN,
"path": "adapted_roberta",
"timestamp": "2026-09-15T12:39:17.107708"
}
},
"results/unified_full/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,78.56,0.8124,0.7811,0.8051,0.7832,0.4381,6.25,712.34,26612.78,124647170,124647170,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_roberta,2026-09-15T12:39:17.107708\n",
"results/unified_full/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,90.71,0.9099,0.907,0.9075,0.9068,0.3027,2.91,388.01,53818.36,66955010,66955010,0,1.0,0.0,0.0,0,0,,,0,0,,legacy_bert,2026-09-15T12:38:41.760113\n",
"results/unified_full/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,79.36,0.7978,0.7935,0.7935,0.7935,0.4875,6.24,791.56,26323.5,124441344,124441344,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_gpt,2026-09-15T12:39:52.918492\n"
},
"notes": []
},
"fused_svd_kernel": {
"module_name": "fused_svd_kernel",
"source_file": "Codes/framework/attentions/fused_svd_kernel.py",
"source_lines_count": 1040,
"architecture": {
"class_name": "FusedSVDLinear",
"base_class": "nn.Module",
"init_args": "self, in_features: int, out_features: int, rank: int = 32, bias: bool = True, use_rms_norm: bool = False, eps: float = 1e-6, dtype: torch.dtype = torch.float32, device: Optional[torch.device] = None, ) -> None: super().__init__() self.in_features = in_features self.out_features = out_features self.rank = rank self.eps = eps self.use_rms_norm = use_rms_norm # Low-rank factorized parameters self.V = nn.Parameter(torch.empty(rank, in_features, dtype=dtype, device=device)) self.U = nn.Parameter(torch.empty(out_features, rank, dtype=dtype, device=device)) if use_rms_norm: self.norm_weight = nn.Parameter(torch.ones(rank, dtype=dtype, device=device)) else: self.register_parameter(\"norm_weight\", None) if bias: self.bias = nn.Parameter(torch.empty(out_features, dtype=dtype, device=device)) else: self.register_parameter(\"bias\", None) self.reset_parameters() def reset_parameters(self) -> None: # Symmetrical square-root initialization nn.init.orthogonal_(self.U) nn.init.orthogonal_(self.V) if self.bias is not None: nn.init.zeros_(self.bias) if self.norm_weight is not None: nn.init.ones_(self.norm_weight) @property def weight(self) -> torch.Tensor: \"\"\"Reconstructed effective weight matrix for compatibility.\"\"\" return torch.matmul(self.U, self.V) def forward(self, x: torch.Tensor) -> torch.Tensor: # RESEARCH FIX: Dispatches fused kernel executing down-projection, RMSNorm, and up-projection return fused_svd_linear( x=x, V=self.V, U=self.U, norm_weight=self.norm_weight, bias=self.bias, eps=self.eps, use_rms_norm=self.use_rms_norm, ) def set_low_rank_factors(self, U_factor: torch.Tensor, V_factor: torch.Tensor) -> None: \"\"\"Set precomputed U and V low-rank factors.\"\"\" with torch.no_grad(",
"registry_names": [
"fused_svd_kernel_attention",
"inject_fused_svd_kernel",
"fused_svd_kernel_inject",
"fused_svd_kernel"
]
},
"params": [
{
"name": "V",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.empty(rank, in_features, dtype=dtype, device=device)"
},
{
"name": "U",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.empty(out_features, rank, dtype=dtype, device=device)"
},
{
"name": "norm_weight",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.ones(rank, dtype=dtype, device=device)"
},
{
"name": "bias",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.empty(out_features, dtype=dtype, device=device)"
}
],
"metrics": {
"results/unified_fused_svd_kernel/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": 5.89
},
"efficiency": {
"total_params_before": 124647170,
"total_params_after": 124647170,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 712.34,
"tokens_per_sec": 28038.54
},
"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": "2026-09-15T12:41:50.886821"
},
"results/unified_fused_svd_kernel/roberta/roberta-base_fused_svd_kernel_full_metrics.json": {
"model": "roberta-base_fused_svd_kernel",
"performance": {
"accuracy_pct": 50.92,
"f1_score": 0.6748,
"f1_macro": 0.3374,
"precision": 0.2546,
"recall": 0.5,
"val_loss": 0.7847,
"status": "PASS_LOW",
"elapsed_s": 22.41
},
"efficiency": {
"total_params_before": 124647170,
"total_params_after": 105782018,
"trainable_recovery_params": 2368512,
"compression_rate": 1.1783,
"compression_ratio_pct": 15.13,
"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": {
"roberta.encoder.layer.0.attention.self": 32,
"roberta.encoder.layer.0.attention.self.q_proj": 32,
"roberta.encoder.layer.0.attention.self.k_proj": 32,
"roberta.encoder.layer.0.attention.self.v_proj": 32,
"roberta.encoder.layer.0.attention.self.out_proj": 32,
"roberta.encoder.layer.1.attention.self": 32,
"roberta.encoder.layer.1.attention.self.q_proj": 32,
"roberta.encoder.layer.1.attention.self.k_proj": 32,
"roberta.encoder.layer.1.attention.self.v_proj": 32,
"roberta.encoder.layer.1.attention.self.out_proj": 32,
"roberta.encoder.layer.2.attention.self": 32,
"roberta.encoder.layer.2.attention.self.q_proj": 32,
"roberta.encoder.layer.2.attention.self.k_proj": 32,
"roberta.encoder.layer.2.attention.self.v_proj": 32,
"roberta.encoder.layer.2.attention.self.out_proj": 32,
"roberta.encoder.layer.3.attention.self": 32,
"roberta.encoder.layer.3.attention.self.q_proj": 32,
"roberta.encoder.layer.3.attention.self.k_proj": 32,
"roberta.encoder.layer.3.attention.self.v_proj": 32,
"roberta.encoder.layer.3.attention.self.out_proj": 32,
"roberta.encoder.layer.4.attention.self": 32,
"roberta.encoder.layer.4.attention.self.q_proj": 32,
"roberta.encoder.layer.4.attention.self.k_proj": 32,
"roberta.encoder.layer.4.attention.self.v_proj": 32,
"roberta.encoder.layer.4.attention.self.out_proj": 32,
"roberta.encoder.layer.5.attention.self": 32,
"roberta.encoder.layer.5.attention.self.q_proj": 32,
"roberta.encoder.layer.5.attention.self.k_proj": 32,
"roberta.encoder.layer.5.attention.self.v_proj": 32,
"roberta.encoder.layer.5.attention.self.out_proj": 32,
"roberta.encoder.layer.6.attention.self": 32,
"roberta.encoder.layer.6.attention.self.q_proj": 32,
"roberta.encoder.layer.6.attention.self.k_proj": 32,
"roberta.encoder.layer.6.attention.self.v_proj": 32,
"roberta.encoder.layer.6.attention.self.out_proj": 32,
"roberta.encoder.layer.7.attention.self": 32,
"roberta.encoder.layer.7.attention.self.q_proj": 32,
"roberta.encoder.layer.7.attention.self.k_proj": 32,
"roberta.encoder.layer.7.attention.self.v_proj": 32,
"roberta.encoder.layer.7.attention.self.out_proj": 32,
"roberta.encoder.layer.8.attention.self": 32,
"roberta.encoder.layer.8.attention.self.q_proj": 32,
"roberta.encoder.layer.8.attention.self.k_proj": 32,
"roberta.encoder.layer.8.attention.self.v_proj": 32,
"roberta.encoder.layer.8.attention.self.out_proj": 32,
"roberta.encoder.layer.9.attention.self": 32,
"roberta.encoder.layer.9.attention.self.q_proj": 32,
"roberta.encoder.layer.9.attention.self.k_proj": 32,
"roberta.encoder.layer.9.attention.self.v_proj": 32,
"roberta.encoder.layer.9.attention.self.out_proj": 32,
"roberta.encoder.layer.10.attention.self": 32,
"roberta.encoder.layer.10.attention.self.q_proj": 32,
"roberta.encoder.layer.10.attention.self.k_proj": 32,
"roberta.encoder.layer.10.attention.self.v_proj": 32,
"roberta.encoder.layer.10.attention.self.out_proj": 32,
"roberta.encoder.layer.11.attention.self": 32,
"roberta.encoder.layer.11.attention.self.q_proj": 32,
"roberta.encoder.layer.11.attention.self.k_proj": 32,
"roberta.encoder.layer.11.attention.self.v_proj": 32,
"roberta.encoder.layer.11.attention.self.out_proj": 32
}
},
"hardware": {
"peak_vram_mb": 558.41,
"tokens_per_sec": 17791.54
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 530.35,
"L_256_vram_mb": 533.79,
"L_512_vram_mb": 558.29,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 514)"
},
"timestamp": "2026-09-15T12:42:13.338631"
},
"results/unified_fused_svd_kernel/distilbert/distilbert_sst2_fused_svd_kernel_full_metrics.json": {
"model": "distilbert_sst2_fused_svd_kernel",
"performance": {
"accuracy_pct": 79.93,
"f1_score": 0.8201,
"f1_macro": 0.7966,
"precision": 0.8115,
"recall": 0.7975,
"val_loss": 0.4844,
"status": "PASS",
"elapsed_s": 49.95
},
"efficiency": {
"total_params_before": 66955010,
"total_params_after": 53978882,
"trainable_recovery_params": 1184256,
"compression_rate": 1.2404,
"compression_ratio_pct": 19.38,
"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": {
"distilbert.transformer.layer.0.attention": 32,
"distilbert.transformer.layer.0.attention.q_proj": 32,
"distilbert.transformer.layer.0.attention.k_proj": 32,
"distilbert.transformer.layer.0.attention.v_proj": 32,
"distilbert.transformer.layer.0.attention.out_proj": 32,
"distilbert.transformer.layer.1.attention": 32,
"distilbert.transformer.layer.1.attention.q_proj": 32,
"distilbert.transformer.layer.1.attention.k_proj": 32,
"distilbert.transformer.layer.1.attention.v_proj": 32,
"distilbert.transformer.layer.1.attention.out_proj": 32,
"distilbert.transformer.layer.2.attention": 32,
"distilbert.transformer.layer.2.attention.q_proj": 32,
"distilbert.transformer.layer.2.attention.k_proj": 32,
"distilbert.transformer.layer.2.attention.v_proj": 32,
"distilbert.transformer.layer.2.attention.out_proj": 32,
"distilbert.transformer.layer.3.attention": 32,
"distilbert.transformer.layer.3.attention.q_proj": 32,
"distilbert.transformer.layer.3.attention.k_proj": 32,
"distilbert.transformer.layer.3.attention.v_proj": 32,
"distilbert.transformer.layer.3.attention.out_proj": 32,
"distilbert.transformer.layer.4.attention": 32,
"distilbert.transformer.layer.4.attention.q_proj": 32,
"distilbert.transformer.layer.4.attention.k_proj": 32,
"distilbert.transformer.layer.4.attention.v_proj": 32,
"distilbert.transformer.layer.4.attention.out_proj": 32,
"distilbert.transformer.layer.5.attention": 32,
"distilbert.transformer.layer.5.attention.q_proj": 32,
"distilbert.transformer.layer.5.attention.k_proj": 32,
"distilbert.transformer.layer.5.attention.v_proj": 32,
"distilbert.transformer.layer.5.attention.out_proj": 32
}
},
"hardware": {
"peak_vram_mb": 280.27,
"tokens_per_sec": 3888.43
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 248.86,
"L_256_vram_mb": 252.79,
"L_512_vram_mb": "unsupported (exceeds model max_position_embeddings 512)",
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 512)"
},
"timestamp": "2026-09-15T12:41:21.844338"
},
"results/unified_fused_svd_kernel/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": 2.91
},
"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": 53310.16
},
"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": "2026-09-15T12:40:31.858808"
},
"results/unified_fused_svd_kernel/gpt2/gpt2_fused_svd_kernel_full_metrics.json": {
"model": "gpt2_fused_svd_kernel",
"performance": {
"accuracy_pct": 70.07,
"f1_score": 0.754,
"f1_macro": 0.6859,
"precision": 0.7379,
"recall": 0.6969,
"val_loss": 0.5923,
"status": "PASS",
"elapsed_s": 97.56
},
"efficiency": {
"total_params_before": 124441344,
"total_params_after": 98489088,
"trainable_recovery_params": 1778688,
"compression_rate": 1.2635,
"compression_ratio_pct": 20.86,
"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": {
"transformer.h.0.attn": 32,
"transformer.h.0.attn.q_proj": 32,
"transformer.h.0.attn.k_proj": 32,
"transformer.h.0.attn.v_proj": 32,
"transformer.h.0.attn.out_proj": 32,
"transformer.h.1.attn": 32,
"transformer.h.1.attn.q_proj": 32,
"transformer.h.1.attn.k_proj": 32,
"transformer.h.1.attn.v_proj": 32,
"transformer.h.1.attn.out_proj": 32,
"transformer.h.2.attn": 32,
"transformer.h.2.attn.q_proj": 32,
"transformer.h.2.attn.k_proj": 32,
"transformer.h.2.attn.v_proj": 32,
"transformer.h.2.attn.out_proj": 32,
"transformer.h.3.attn": 32,
"transformer.h.3.attn.q_proj": 32,
"transformer.h.3.attn.k_proj": 32,
"transformer.h.3.attn.v_proj": 32,
"transformer.h.3.attn.out_proj": 32,
"transformer.h.4.attn": 32,
"transformer.h.4.attn.q_proj": 32,
"transformer.h.4.attn.k_proj": 32,
"transformer.h.4.attn.v_proj": 32,
"transformer.h.4.attn.out_proj": 32,
"transformer.h.5.attn": 32,
"transformer.h.5.attn.q_proj": 32,
"transformer.h.5.attn.k_proj": 32,
"transformer.h.5.attn.v_proj": 32,
"transformer.h.5.attn.out_proj": 32,
"transformer.h.6.attn": 32,
"transformer.h.6.attn.q_proj": 32,
"transformer.h.6.attn.k_proj": 32,
"transformer.h.6.attn.v_proj": 32,
"transformer.h.6.attn.out_proj": 32,
"transformer.h.7.attn": 32,
"transformer.h.7.attn.q_proj": 32,
"transformer.h.7.attn.k_proj": 32,
"transformer.h.7.attn.v_proj": 32,
"transformer.h.7.attn.out_proj": 32,
"transformer.h.8.attn": 32,
"transformer.h.8.attn.q_proj": 32,
"transformer.h.8.attn.k_proj": 32,
"transformer.h.8.attn.v_proj": 32,
"transformer.h.8.attn.out_proj": 32,
"transformer.h.9.attn": 32,
"transformer.h.9.attn.q_proj": 32,
"transformer.h.9.attn.k_proj": 32,
"transformer.h.9.attn.v_proj": 32,
"transformer.h.9.attn.out_proj": 32,
"transformer.h.10.attn": 32,
"transformer.h.10.attn.q_proj": 32,
"transformer.h.10.attn.k_proj": 32,
"transformer.h.10.attn.v_proj": 32,
"transformer.h.10.attn.out_proj": 32,
"transformer.h.11.attn": 32,
"transformer.h.11.attn.q_proj": 32,
"transformer.h.11.attn.k_proj": 32,
"transformer.h.11.attn.v_proj": 32,
"transformer.h.11.attn.out_proj": 32
}
},
"hardware": {
"peak_vram_mb": 464.46,
"tokens_per_sec": 1955.65
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 423.59,
"L_256_vram_mb": 428.46,
"L_512_vram_mb": 452.84,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 1024)"
},
"timestamp": "2026-09-15T12:44:21.002427"
},
"results/unified_fused_svd_kernel/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": 5.86
},
"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": 27584.57
},
"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": "2026-09-15T12:42:43.424221"
},
"results/unified_fused_svd_kernel/gpt2/hf_export/metrics.json": {
"model_name": "gpt2",
"module": "fused_svd_kernel",
"status": "PASS",
"accuracy_pct": 70.07,
"f1": 0.0,
"precision": 0.7379,
"recall": 0.6969,
"eval_loss": 0.0,
"params_total": null,
"params_active": null,
"param_reduction_pct": null,
"peak_vram_mb": 464.46,
"tokens_per_sec": 1955.65,
"elapsed_seconds": null,
"full_metrics": {
"module_name": "fused_svd_kernel",
"status": "PASS",
"accuracy_pct": 70.07,
"f1_score": 0.754,
"f1_macro": 0.6859,
"precision": 0.7379,
"recall": 0.6969,
"val_loss": 0.5923,
"elapsed_s": 97.56,
"peak_vram_mb": 464.46,
"tokens_per_sec": 1955.65,
"total_params_before": 124441344,
"total_params_after": 98489088,
"trainable_recovery_params": 1778688,
"compression_rate": 1.2635,
"compression_ratio_pct": 20.86,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"avg_reconstruction_error": NaN,
"weighted_reconstruction_error": NaN,
"num_injected_attention_layers": 0,
"lora_params": 0,
"error": NaN,
"path": "adapted_gpt",
"timestamp": "2026-09-15T12:44:21.852366"
}
},
"results/unified_fused_svd_kernel/distilbert/hf_export/metrics.json": {
"model_name": "distilbert",
"module": "fused_svd_kernel",
"status": "PASS",
"accuracy_pct": 79.93,
"f1": 0.0,
"precision": 0.8115,
"recall": 0.7975,
"eval_loss": 0.0,
"params_total": null,
"params_active": null,
"param_reduction_pct": null,
"peak_vram_mb": 280.27,
"tokens_per_sec": 3888.43,
"elapsed_seconds": null,
"full_metrics": {
"module_name": "fused_svd_kernel",
"status": "PASS",
"accuracy_pct": 79.93,
"f1_score": 0.8201,
"f1_macro": 0.7966,
"precision": 0.8115,
"recall": 0.7975,
"val_loss": 0.4844,
"elapsed_s": 49.95,
"peak_vram_mb": 280.27,
"tokens_per_sec": 3888.43,
"total_params_before": 66955010,
"total_params_after": 53978882,
"trainable_recovery_params": 1184256,
"compression_rate": 1.2404,
"compression_ratio_pct": 19.38,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"avg_reconstruction_error": NaN,
"weighted_reconstruction_error": NaN,
"num_injected_attention_layers": 0,
"lora_params": 0,
"error": NaN,
"path": "legacy_bert",
"timestamp": "2026-09-15T12:41:22.249540"
}
},
"results/unified_fused_svd_kernel/roberta/hf_export/metrics.json": {
"model_name": "roberta",
"module": "fused_svd_kernel",
"status": "PASS",
"accuracy_pct": 70.07,
"f1": 0.0,
"precision": 0.7201,
"recall": 0.6978,
"eval_loss": 0.0,
"params_total": null,
"params_active": null,
"param_reduction_pct": null,
"peak_vram_mb": 560.48,
"tokens_per_sec": 11835.03,
"elapsed_seconds": null,
"full_metrics": {
"module_name": "fused_svd_kernel",
"status": "PASS",
"accuracy_pct": 70.07,
"f1_score": 0.7439,
"f1_macro": 0.6919,
"precision": 0.7201,
"recall": 0.6978,
"val_loss": 0.5915,
"elapsed_s": 53.35,
"peak_vram_mb": 560.48,
"tokens_per_sec": 11835.03,
"total_params_before": 124647170,
"total_params_after": 107072450,
"trainable_recovery_params": 1882562,
"compression_rate": 1.1641,
"compression_ratio_pct": 14.1,
"flops_reduction_pct": 91.67,
"num_compressed_layers": 48,
"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": "2026-09-11T11:23:23.267525"
}
},
"results/unified_fused_svd_kernel/unified_fusion_attn/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": 5.78
},
"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": 28229.17
},
"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": "2026-09-15T12:45:40.106775"
},
"results/unified_fused_svd_kernel/unified_fusion_attn/roberta/roberta-base_fusion_attn_full_metrics.json": {
"model": "roberta-base_fusion_attn",
"performance": {
"accuracy_pct": 62.61,
"f1_score": 0.5821,
"f1_macro": 0.6219,
"precision": 0.6354,
"recall": 0.6283,
"val_loss": 0.7766,
"status": "PASS_NO_COMPRESSION",
"elapsed_s": 23.04
},
"efficiency": {
"total_params_before": 124647170,
"total_params_after": 133116746,
"trainable_recovery_params": 1382472,
"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.1.attention.self": 32,
"roberta.encoder.layer.2.attention.self": 32,
"roberta.encoder.layer.3.attention.self": 32,
"roberta.encoder.layer.4.attention.self": 32,
"roberta.encoder.layer.5.attention.self": 32,
"roberta.encoder.layer.6.attention.self": 32,
"roberta.encoder.layer.7.attention.self": 32,
"roberta.encoder.layer.8.attention.self": 32,
"roberta.encoder.layer.9.attention.self": 32,
"roberta.encoder.layer.10.attention.self": 32,
"roberta.encoder.layer.11.attention.self": 32
}
},
"hardware": {
"peak_vram_mb": 661.02,
"tokens_per_sec": 13338.03
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 632.96,
"L_256_vram_mb": 635.92,
"L_512_vram_mb": 641.05,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 514)"
},
"timestamp": "2026-09-15T12:46:03.167263"
},
"results/unified_fused_svd_kernel/unified_fusion_attn/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": 2.87
},
"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": 53582.06
},
"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": "2026-09-15T12:45:00.408704"
},
"results/unified_fused_svd_kernel/unified_fusion_attn/distilbert/distilbert_sst2_fusion_attn_full_metrics.json": {
"model": "distilbert_sst2_fusion_attn",
"performance": {
"accuracy_pct": 90.6,
"f1_score": 0.9091,
"f1_macro": 0.9059,
"precision": 0.9065,
"recall": 0.9056,
"val_loss": 0.2599,
"status": "PASS_NO_COMPRESSION",
"elapsed_s": 11.0
},
"efficiency": {
"total_params_before": 66955010,
"total_params_after": 67646246,
"trainable_recovery_params": 691236,
"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.1.attention": 32,
"distilbert.transformer.layer.2.attention": 32,
"distilbert.transformer.layer.3.attention": 32,
"distilbert.transformer.layer.4.attention": 32,
"distilbert.transformer.layer.5.attention": 32
}
},
"hardware": {
"peak_vram_mb": 329.66,
"tokens_per_sec": 26943.9
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 299.04,
"L_256_vram_mb": 301.8,
"L_512_vram_mb": "unsupported (exceeds model max_position_embeddings 512)",
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 512)"
},
"timestamp": "2026-09-15T12:45:11.514082"
},
"results/unified_fused_svd_kernel/unified_fusion_attn/gpt2/gpt2_fusion_attn_full_metrics.json": {
"model": "gpt2_fusion_attn",
"performance": {
"accuracy_pct": 79.36,
"f1_score": 0.7978,
"f1_macro": 0.7935,
"precision": 0.7935,
"recall": 0.7935,
"val_loss": 0.4836,
"status": "PASS_NO_COMPRESSION",
"elapsed_s": 23.76
},
"efficiency": {
"total_params_before": 124441344,
"total_params_after": 125823816,
"trainable_recovery_params": 1382472,
"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.1.attn": 32,
"transformer.h.2.attn": 32,
"transformer.h.3.attn": 32,
"transformer.h.4.attn": 32,
"transformer.h.5.attn": 32,
"transformer.h.6.attn": 32,
"transformer.h.7.attn": 32,
"transformer.h.8.attn": 32,
"transformer.h.9.attn": 32,
"transformer.h.10.attn": 32,
"transformer.h.11.attn": 32
}
},
"hardware": {
"peak_vram_mb": 569.74,
"tokens_per_sec": 12651.27
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 526.8,
"L_256_vram_mb": 531.89,
"L_512_vram_mb": 548.64,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 1024)"
},
"timestamp": "2026-09-15T12:46:56.417145"
},
"results/unified_fused_svd_kernel/unified_fusion_attn/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": 5.88
},
"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": 27799.23
},
"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": "2026-09-15T12:46:32.649246"
},
"results/unified_fused_svd_kernel/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,78.56,0.8124,0.7811,0.8051,0.7832,0.4381,5.89,712.34,28038.54,124647170,124647170,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_roberta,2026-09-15T12:41:51.583974\nfused_svd_kernel,PASS_LOW,50.92,0.6748,0.3374,0.2546,0.5,0.7847,22.41,558.41,17791.54,124647170,105782018,2368512,1.1783,15.13,0.0,0,0,,,0,0,note: evaluated accuracy (50.92%) below chance threshold (56.00%),adapted_roberta,2026-09-15T12:42:14.244443\n",
"results/unified_fused_svd_kernel/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,90.71,0.9099,0.907,0.9075,0.9068,0.3027,2.91,389.51,53310.16,66955010,66955010,0,1.0,0.0,0.0,0,0,,,0,0,,legacy_bert,2026-09-15T12:40:32.074483\nfused_svd_kernel,PASS,79.93,0.8201,0.7966,0.8115,0.7975,0.4844,49.95,280.27,3888.43,66955010,53978882,1184256,1.2404,19.38,0.0,0,0,,,0,0,,legacy_bert,2026-09-15T12:41:22.249540\n",
"results/unified_fused_svd_kernel/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,79.36,0.7978,0.7935,0.7935,0.7935,0.4875,5.86,793.19,27584.57,124441344,124441344,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_gpt,2026-09-15T12:42:44.091671\nfused_svd_kernel,PASS,70.07,0.754,0.6859,0.7379,0.6969,0.5923,97.56,464.46,1955.65,124441344,98489088,1778688,1.2635,20.86,0.0,0,0,,,0,0,,adapted_gpt,2026-09-15T12:44:21.852366\n",
"results/unified_fused_svd_kernel/unified_fusion_attn/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,78.56,0.8124,0.7811,0.8051,0.7832,0.4381,5.78,712.34,28229.17,124647170,124647170,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_roberta,2026-09-15T12:45:40.786625\nfusion_attn,PASS_NO_COMPRESSION,62.61,0.5821,0.6219,0.6354,0.6283,0.7766,23.04,661.02,13338.03,124647170,133116746,1382472,1.0,0.0,0.0,0,0,,,12,1179648,\"note: zero compressed layers, model grew (LoRA only)\",adapted_roberta,2026-09-15T12:46:04.052062\n",
"results/unified_fused_svd_kernel/unified_fusion_attn/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,90.71,0.9099,0.907,0.9075,0.9068,0.3027,2.87,389.51,53582.06,66955010,66955010,0,1.0,0.0,0.0,0,0,,,0,0,,legacy_bert,2026-09-15T12:45:00.613239\nfusion_attn,PASS_NO_COMPRESSION,90.6,0.9091,0.9059,0.9065,0.9056,0.2599,11.0,329.66,26943.9,66955010,67646246,691236,1.0,0.0,0.0,0,0,,,6,589824,\"note: zero compressed layers, model grew (LoRA only)\",legacy_bert,2026-09-15T12:45:11.846166\n",
"results/unified_fused_svd_kernel/unified_fusion_attn/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,79.36,0.7978,0.7935,0.7935,0.7935,0.4875,5.88,793.19,27799.23,124441344,124441344,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_gpt,2026-09-15T12:46:33.330820\nfusion_attn,PASS_NO_COMPRESSION,79.36,0.7978,0.7935,0.7935,0.7935,0.4836,23.76,569.74,12651.27,124441344,125823816,1382472,1.0,0.0,0.0,0,0,,,12,1179648,\"note: zero compressed layers, model grew (LoRA only)\",adapted_gpt,2026-09-15T12:46:57.338732\n",
"results/benchmark_summary_v2.csv": "track4_kernel_quant,fused_svd_kernel,attention,bert,sst2,FAIL,\"at 54:17:\ntrack4_kernel_quant,fused_svd_kernel,attention,bert,mrpc,FAIL,\"at 54:17:\ntrack4_kernel_quant,fused_svd_kernel,attention,roberta,sst2,FAIL,\"at 54:17:\ntrack4_kernel_quant,fused_svd_kernel,attention,roberta,mrpc,FAIL,\"at 54:17:\ntrack4_kernel_quant,fused_svd_kernel,attention,t5,sst2,FAIL,\"at 54:17:\ntrack4_kernel_quant,fused_svd_kernel,attention,t5,mrpc,FAIL,\"at 54:17:",
"results/master_aggregated_results.csv": "fused_svd_kernel,0,PASS,0.0,70.07,0.0,214.4,2026-09-11T11:24:56.515321",
"results/run_quality_all_screening_summary.csv": "run_quality_20260827_052444,fused_svd_kernel,DISTILBERT,SST2,PASS,50.92,0.0,50.92,50.92,1.0,0.6748,0.1973,397.98,526.12,1824.6,17729.8,1450000000.0,2026-08-27T07:33:32.086574"
},
"notes": []
},
"fusion_attn": {
"module_name": "fusion_attn",
"source_file": "Codes/framework/attentions/fusion_attn.py",
"source_lines_count": 684,
"architecture": {
"class_name": "LoRALinear",
"base_class": "nn.Module",
"init_args": "self, in_features: int, out_features: int, bias: bool = True, lora_rank: int = 16, lora_alpha: float = 32.0, freeze_base: bool = False, device: Optional[torch.device] = None, dtype: Optional[torch.dtype] = None, ) -> None: super().__init__() factory_kwargs = {\"device\": device, \"dtype\": dtype} self.in_features = in_features self.out_features = out_features 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.freeze_base = freeze_base self.weight = nn.Parameter(torch.empty((out_features, in_features), **factory_kwargs)) if bias: self.bias = nn.Parameter(torch.empty(out_features, **factory_kwargs)) else: self.register_parameter(\"bias\", None) if self.lora_rank > 0: self.lora_A = nn.Parameter(torch.empty((lora_rank, in_features), **factory_kwargs)) self.lora_B = nn.Parameter(torch.empty((out_features, lora_rank), **factory_kwargs)) else: self.register_parameter(\"lora_A\", None) self.register_parameter(\"lora_B\", None) self.reset_parameters() if freeze_base: self.freeze_base_weights() def reset_parameters(self) -> None: nn.init.kaiming_uniform_(self.weight, a=math.sqrt(5)) if self.bias is not None: fan_in, _ = nn.init._calculate_fan_in_and_fan_out(self.weight) bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0 nn.init.uniform_(self.bias, -bound, bound) 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 freeze_base_weights(self) -> None: \"\"\"Freeze base projection parameters so only LoRA parameters adapt.\"\"\" self.weight.requires_grad = False if self.bias is not None: self.bias.requires_grad = False if self.lora_rank > 0 and self.lora_A is not None: self.lora_A.requires_grad = True if self.lora_rank > 0 and self.lora_B is not None: self.lora_B.requires_grad = True def unfreeze_base_weights(self) -> None: \"\"\"Unfreeze base projection parameters.\"\"\" self.weight.requires_grad = True if self.bias is not None: self.bias.requires_grad = True def forward(self, x: torch.Tensor) -> torch.Tensor: out = F.linear(x, self.weight, self.bias) if self.lora_rank > 0 and self.lora_A is not None and self.lora_B is not None: lora_out = F.linear(F.linear(x, self.lora_A), self.lora_B) * self.lora_scaling out = out + lora_out return out @ATTENTION_REGISTRY.register(\"fusion_attn\", supports_causal=True, supports_bidirectional=True, supports_gqa=True) class FUSIONATTN(BaseAttention",
"registry_names": [
"fusion_attn",
"fusion_attn_inject"
]
},
"params": [
{
"name": "weight",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.empty((out_features, in_features)"
},
{
"name": "bias",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.empty(out_features, **factory_kwargs)"
},
{
"name": "lora_A",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.empty((lora_rank, in_features)"
},
{
"name": "lora_B",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.empty((out_features, lora_rank)"
},
{
"name": "gate_temp",
"type": "nn.Parameter",
"declaration": "nn.Parameter(torch.ones(1)"
}
],
"metrics": {
"results/unified_fusion_attn/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": 5.96
},
"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": 27366.6
},
"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": "2026-09-15T15:00:21.659402"
},
"results/unified_fusion_attn/roberta/roberta-base_fusion_attn_full_metrics.json": {
"model": "roberta-base_fusion_attn",
"performance": {
"accuracy_pct": 62.61,
"f1_score": 0.5821,
"f1_macro": 0.6219,
"precision": 0.6354,
"recall": 0.6283,
"val_loss": 0.7766,
"status": "PASS_NO_COMPRESSION",
"elapsed_s": 23.76
},
"efficiency": {
"total_params_before": 124647170,
"total_params_after": 133116746,
"trainable_recovery_params": 1382472,
"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.1.attention.self": 32,
"roberta.encoder.layer.2.attention.self": 32,
"roberta.encoder.layer.3.attention.self": 32,
"roberta.encoder.layer.4.attention.self": 32,
"roberta.encoder.layer.5.attention.self": 32,
"roberta.encoder.layer.6.attention.self": 32,
"roberta.encoder.layer.7.attention.self": 32,
"roberta.encoder.layer.8.attention.self": 32,
"roberta.encoder.layer.9.attention.self": 32,
"roberta.encoder.layer.10.attention.self": 32,
"roberta.encoder.layer.11.attention.self": 32
}
},
"hardware": {
"peak_vram_mb": 661.02,
"tokens_per_sec": 12635.46
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 632.96,
"L_256_vram_mb": 635.92,
"L_512_vram_mb": 641.05,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 514)"
},
"timestamp": "2026-09-15T15:00:45.477667"
},
"results/unified_fusion_attn/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": 2.96
},
"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": 52651.57
},
"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": "2026-09-15T14:59:40.953745"
},
"results/unified_fusion_attn/distilbert/distilbert_sst2_fusion_attn_full_metrics.json": {
"model": "distilbert_sst2_fusion_attn",
"performance": {
"accuracy_pct": 90.6,
"f1_score": 0.9091,
"f1_macro": 0.9059,
"precision": 0.9065,
"recall": 0.9056,
"val_loss": 0.2599,
"status": "PASS_NO_COMPRESSION",
"elapsed_s": 11.4
},
"efficiency": {
"total_params_before": 66955010,
"total_params_after": 67646246,
"trainable_recovery_params": 691236,
"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.1.attention": 32,
"distilbert.transformer.layer.2.attention": 32,
"distilbert.transformer.layer.3.attention": 32,
"distilbert.transformer.layer.4.attention": 32,
"distilbert.transformer.layer.5.attention": 32
}
},
"hardware": {
"peak_vram_mb": 329.66,
"tokens_per_sec": 26220.65
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 299.04,
"L_256_vram_mb": 301.8,
"L_512_vram_mb": "unsupported (exceeds model max_position_embeddings 512)",
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 512)"
},
"timestamp": "2026-09-15T14:59:52.466133"
},
"results/unified_fusion_attn/gpt2/gpt2_fusion_attn_full_metrics.json": {
"model": "gpt2_fusion_attn",
"performance": {
"accuracy_pct": 79.36,
"f1_score": 0.7978,
"f1_macro": 0.7935,
"precision": 0.7935,
"recall": 0.7935,
"val_loss": 0.4836,
"status": "PASS_NO_COMPRESSION",
"elapsed_s": 24.42
},
"efficiency": {
"total_params_before": 124441344,
"total_params_after": 125823816,
"trainable_recovery_params": 1382472,
"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.1.attn": 32,
"transformer.h.2.attn": 32,
"transformer.h.3.attn": 32,
"transformer.h.4.attn": 32,
"transformer.h.5.attn": 32,
"transformer.h.6.attn": 32,
"transformer.h.7.attn": 32,
"transformer.h.8.attn": 32,
"transformer.h.9.attn": 32,
"transformer.h.10.attn": 32,
"transformer.h.11.attn": 32
}
},
"hardware": {
"peak_vram_mb": 569.74,
"tokens_per_sec": 12161.57
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 526.8,
"L_256_vram_mb": 531.89,
"L_512_vram_mb": 548.64,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 1024)"
},
"timestamp": "2026-09-15T15:01:40.372131"
},
"results/unified_fusion_attn/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": 6.08
},
"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": 27066.22
},
"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": "2026-09-15T15:01:15.888702"
},
"results/unified_fusion_attn/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,78.56,0.8124,0.7811,0.8051,0.7832,0.4381,5.96,712.34,27366.6,124647170,124647170,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_roberta,2026-09-15T15:00:22.356732\nfusion_attn,PASS_NO_COMPRESSION,62.61,0.5821,0.6219,0.6354,0.6283,0.7766,23.76,661.02,12635.46,124647170,133116746,1382472,1.0,0.0,0.0,0,0,,,12,1179648,\"note: zero compressed layers, model grew (LoRA only)\",adapted_roberta,2026-09-15T15:00:46.373884\n",
"results/unified_fusion_attn/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,90.71,0.9099,0.907,0.9075,0.9068,0.3027,2.96,389.51,52651.57,66955010,66955010,0,1.0,0.0,0.0,0,0,,,0,0,,legacy_bert,2026-09-15T14:59:41.176791\nfusion_attn,PASS_NO_COMPRESSION,90.6,0.9091,0.9059,0.9065,0.9056,0.2599,11.4,329.66,26220.65,66955010,67646246,691236,1.0,0.0,0.0,0,0,,,6,589824,\"note: zero compressed layers, model grew (LoRA only)\",legacy_bert,2026-09-15T14:59:52.816263\n",
"results/unified_fusion_attn/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,79.36,0.7978,0.7935,0.7935,0.7935,0.4875,6.08,793.19,27066.22,124441344,124441344,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_gpt,2026-09-15T15:01:16.580745\nfusion_attn,PASS_NO_COMPRESSION,79.36,0.7978,0.7935,0.7935,0.7935,0.4836,24.42,569.74,12161.57,124441344,125823816,1382472,1.0,0.0,0.0,0,0,,,12,1179648,\"note: zero compressed layers, model grew (LoRA only)\",adapted_gpt,2026-09-15T15:01:41.255412\n",
"results/agnews_unified_fusion_attn/distilbert/distilbert_sst2_fusion_attn_full_metrics.json": {
"model": "distilbert_sst2_fusion_attn",
"performance": {
"accuracy_pct": 63.6,
"f1_score": 0.632,
"f1_macro": 0.632,
"precision": 0.678,
"recall": 0.6313,
"val_loss": 0.9736,
"status": "PASS_NO_COMPRESSION",
"elapsed_s": 6.96
},
"efficiency": {
"total_params_before": 66956548,
"total_params_after": 67647784,
"trainable_recovery_params": 691236,
"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": 64,
"distilbert.transformer.layer.1.attention": 64,
"distilbert.transformer.layer.2.attention": 64,
"distilbert.transformer.layer.3.attention": 64,
"distilbert.transformer.layer.4.attention": 64,
"distilbert.transformer.layer.5.attention": 64
}
},
"hardware": {
"peak_vram_mb": 571.06,
"tokens_per_sec": 131906.6
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 301.63,
"L_256_vram_mb": 304.39,
"L_512_vram_mb": "unsupported (exceeds model max_position_embeddings 512)",
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 512)"
},
"timestamp": "2026-09-16T04:03:54.197899"
},
"results/agnews_unified_fusion_attn/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\nfusion_attn,PASS_NO_COMPRESSION,63.6,0.632,0.632,0.678,0.6313,0.9736,6.96,571.06,131906.6,66956548,67647784,691236,1.0,0.0,0.0,0,0,,,6,589824,\"note: zero compressed layers, model grew (LoRA only)\",legacy_bert,2026-09-16T04:03:54.475141\n",
"results/mnli_unified_fusion_attn/roberta/roberta-base_fusion_attn_full_metrics.json": {
"model": "roberta-base_fusion_attn",
"performance": {
"accuracy_pct": 37.3,
"f1_score": 0.2954,
"f1_macro": 0.2954,
"precision": 0.2506,
"recall": 0.3651,
"val_loss": 1.1216,
"status": "PASS_NO_COMPRESSION",
"elapsed_s": 11.75
},
"efficiency": {
"total_params_before": 124647939,
"total_params_after": 133117515,
"trainable_recovery_params": 1382472,
"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": 64,
"roberta.encoder.layer.1.attention.self": 64,
"roberta.encoder.layer.2.attention.self": 64,
"roberta.encoder.layer.3.attention.self": 64,
"roberta.encoder.layer.4.attention.self": 64,
"roberta.encoder.layer.5.attention.self": 64,
"roberta.encoder.layer.6.attention.self": 64,
"roberta.encoder.layer.7.attention.self": 64,
"roberta.encoder.layer.8.attention.self": 64,
"roberta.encoder.layer.9.attention.self": 64,
"roberta.encoder.layer.10.attention.self": 64,
"roberta.encoder.layer.11.attention.self": 64
}
},
"hardware": {
"peak_vram_mb": 693.92,
"tokens_per_sec": 25699.0
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 632.24,
"L_256_vram_mb": 635.2,
"L_512_vram_mb": 640.33,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 514)"
},
"timestamp": "2026-09-16T04:01:43.810206"
},
"results/mnli_unified_fusion_attn/distilbert/distilbert_sst2_fusion_attn_full_metrics.json": {
"model": "distilbert_sst2_fusion_attn",
"performance": {
"accuracy_pct": 34.6,
"f1_score": 0.1888,
"f1_macro": 0.1888,
"precision": 0.2343,
"recall": 0.3392,
"val_loss": 1.1001,
"status": "PASS_NO_COMPRESSION",
"elapsed_s": 5.74
},
"efficiency": {
"total_params_before": 66955779,
"total_params_after": 67647015,
"trainable_recovery_params": 691236,
"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": 64,
"distilbert.transformer.layer.1.attention": 64,
"distilbert.transformer.layer.2.attention": 64,
"distilbert.transformer.layer.3.attention": 64,
"distilbert.transformer.layer.4.attention": 64,
"distilbert.transformer.layer.5.attention": 64
}
},
"hardware": {
"peak_vram_mb": 361.59,
"tokens_per_sec": 53046.99
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 299.28,
"L_256_vram_mb": 302.04,
"L_512_vram_mb": "unsupported (exceeds model max_position_embeddings 512)",
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 512)"
},
"timestamp": "2026-09-16T02:57:55.362922"
},
"results/mnli_unified_fusion_attn/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\nfusion_attn,PASS_NO_COMPRESSION,37.3,0.2954,0.2954,0.2506,0.3651,1.1216,11.75,693.92,25699.0,124647939,133117515,1382472,1.0,0.0,0.0,0,0,,,12,1179648,\"note: zero compressed layers, model grew (LoRA only)\",adapted_roberta,2026-09-16T04:01:44.597169\n",
"results/mnli_unified_fusion_attn/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\nfusion_attn,PASS_NO_COMPRESSION,34.6,0.1888,0.1888,0.2343,0.3392,1.1001,5.74,361.59,53046.99,66955779,67647015,691236,1.0,0.0,0.0,0,0,,,6,589824,\"note: zero compressed layers, model grew (LoRA only)\",legacy_bert,2026-09-16T02:57:55.654471\n",
"results/run_quality_all_screening_summary.csv": "run_quality_20260827_052444,fusion_attn,DISTILBERT,SST2,PASS,77.98,0.0,77.98,77.98,1.0,0.7692,0.7793,98.47,258.77,2008.0,45861.7,1450000000.0,2026-08-27T07:34:44.872225"
},
"notes": []
}
}
{
"flat_llm": {
"module_name": "flat_llm",
"source_file": "Codes/framework/compression_engine/flat_llm.py",
"source_lines_count": 394,
"architecture": {
"class_name": "FLATLLMCompressor",
"base_class": "BaseCompressor",
"init_args": "self, model: nn.Module, tokenizer: Optional[Any] = None, keep_ratio: float = 0.7, damping: float = 1e-4, max_error: float = 0.35, target_attention_only: bool = False, fallback_to_original: bool = True, calibration_loader: Optional[Any] = None, calibration_texts: Optional[List[str]] = None, device: Optional[Union[str, torch.device]] = None, structured_24: bool = False, target_modules: Optional[List[str]] = None, cfg: Optional[Any] = None, **kwargs: Any, ) -> None: super().__init__(model, name=\"flat_llm\", **kwargs) cfg = cfg or kwargs.get(\"comp_config\", None) self.cfg = cfg self.tokenizer = tokenizer or getattr(model, \"tokenizer\", None) raw_kr = _cfg_get(cfg, \"flat_llm_keep_ratio\", _cfg_get(cfg, \"keep_ratio\", keep_ratio)) self.keep_ratio = max(0.05, min(float(raw_kr), 1.0)) self.damping = float(_cfg_get(cfg, \"flat_llm_damping\", _cfg_get(cfg, \"damping\", damping))) self.structured_24 = bool(_cfg_get(cfg, \"flat_llm_structured_24\", _cfg_get(cfg, \"structured_24\", structured_24))) self.max_error = float(_cfg_get(cfg, \"flat_llm_max_error\", _cfg_get(cfg, \"max_error\", max_error))) self.target_attention_only = bool( _cfg_get(cfg, \"flat_llm_target_attention_only\", _cfg_get(cfg, \"target_attention_only\", target_attention_only)) ) self.fallback_to_original = bool( _cfg_get(cfg, \"flat_llm_fallback_to_original\", _cfg_get(cfg, \"fallback_to_original\", fallback_to_original)) ) self.calibration_loader = calibration_loader or _cfg_get(cfg, \"calibration_loader\", None) self.calibration_texts = calibration_texts self.target_modules = target_modules self.target_device = ( torch.device(device) if device is not None else next(model.parameters(), torch.tensor(0.0)).device ) self.compressed_layers: List[str] = [] self.fallback_layers: List[str] = [] self.skipped_layers: List[str] = [] self.layer_errors: Dict[str, float] = {} self.layer_keep_ratios: Dict[str, float] = {} def _is_target_module(self, name: str, module: nn.Module) -> bool: name_lower = name.lower() if any(part in name.split(\".\") for part in [\"lora_A\", \"lora_B\", \"base_layer\"]",
"registry_names": [
"flat_llm_compressor",
"flat_llm",
"flat_llm_factory"
]
},
"params": [],
"metrics": {
"results/unified_flat_llm/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": 5.87
},
"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": 27816.89
},
"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": "2026-09-15T12:32:41.194584"
},
"results/unified_flat_llm/roberta/roberta-base_flat_llm_full_metrics.json": {
"model": "roberta-base_flat_llm",
"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_NO_COMPRESSION",
"elapsed_s": 9.02
},
"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": 72,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": 0.0,
"weighted_reconstruction_error": 0.5395,
"per_layer_reconstruction_errors": {
"roberta.encoder.layer.0.attention.self.query": 0.48081703763211586,
"roberta.encoder.layer.0.attention.self.key": 0.4678205793882954,
"roberta.encoder.layer.0.attention.self.value": 0.600996174719836,
"roberta.encoder.layer.0.attention.output.dense": 0.5184713960115622,
"roberta.encoder.layer.0.intermediate.dense": 0.5504354504181636,
"roberta.encoder.layer.0.output.dense": 0.5380613213290566,
"roberta.encoder.layer.1.attention.self.query": 0.5832522478629809,
"roberta.encoder.layer.1.attention.self.key": 0.6117967586196459,
"roberta.encoder.layer.1.attention.self.value": 0.4609478712698367,
"roberta.encoder.layer.1.attention.output.dense": 0.5265832299097785,
"roberta.encoder.layer.1.intermediate.dense": 0.5489612424907715,
"roberta.encoder.layer.1.output.dense": 0.5330325015193179,
Xet Storage Details
- Size:
- 126 kB
- Xet hash:
- 1ed6a0daaa2f499f7b87af24cd463f67372727009d31ff5edfa0e686bc8f5ccb
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.