{
"direct_svd": {
"module_name": "direct_svd",
"source_file": "Codes/framework/compression_engine/direct_svd.py",
"source_lines_count": 462,
"architecture": {
"class_name": "DirectSVDCompressor",
"base_class": "BaseCompressor",
"init_args": "self, model: nn.Module, rank: Optional[int] = 32, min_rank: int = 16, rank_ratio: Optional[float] = None, max_error: float = 0.35, fallback_to_original: bool = True, protect_critical: bool = True, lora_r: int = 0, lora_alpha: float = 1.0, skip_attention_qkv: bool = True, target_attention_only: bool = False, max_retries: int = 2, target_modules: Optional[List[str]] = None, cfg: Optional[Any] = None, **kwargs: Any, ) -> None: super().__init__(model, name=\"direct_svd\", **kwargs) cfg_obj = cfg if cfg is not None else kwargs.get(\"comp_config\", None) self.cfg = cfg_obj rank = _cfg_get(cfg_obj, \"direct_svd_rank\", _cfg_get(cfg_obj, \"svd_rank\", _cfg_get(cfg_obj, \"rank\", rank))) min_rank = _cfg_get(cfg_obj, \"min_rank\", min_rank) rank_ratio = _cfg_get(cfg_obj, \"direct_svd_rank_ratio\", _cfg_get(cfg_obj, \"rank_ratio\", rank_ratio)) max_error = _cfg_get(cfg_obj, \"direct_svd_max_error\", _cfg_get(cfg_obj, \"max_error\", max_error)) fallback_to_original = _cfg_get( cfg_obj, \"direct_svd_fallback_to_original\", _cfg_get(cfg_obj, \"fallback_to_original\", fallback_to_original) ) protect_critical = _cfg_get(cfg_obj, \"protect_critical\", protect_critical) lora_r = _cfg_get(cfg_obj, \"lora_r\", _cfg_get(cfg_obj, \"lora_rank\", lora_r)) lora_alpha = _cfg_get(cfg_obj, \"lora_alpha\", lora_alpha) skip_attention_qkv = _cfg_get(cfg_obj, \"skip_attention_qkv\", skip_attention_qkv) target_attention_only = _cfg_get( cfg_obj, \"direct_svd_target_attention_only\", _cfg_get(cfg_obj, \"target_attention_only\", target_attention_only) ) if \"rank\" in kwargs and kwargs[\"rank\"] is not None: rank = kwargs[\"rank\"] self.rank = rank self.min_rank = max(1, int(min_rank)) self.rank_ratio = float(rank_ratio) if rank_ratio is not None else None self.max_error = float(max_error) self.fallback_to_original = bool(fallback_to_original) self.protect_critical = bool(protect_critical) self.lora_r = int(lora_r) self.lora_alpha = float(lora_alpha) self.skip_attention_qkv = bool(skip_attention_qkv) self.target_attention_only = bool(target_attention_only) self.max_retries = int(max_retries) self.target_modules = target_modules self.compressed_layers: List[str] = [] self.fallback_layers: List[str] = [] self.skipped_layers: List[str] = [] self.layer_ranks: Dict[str, int] = {} self.layer_errors: Dict[str, float] = {} def _is_gqa_asymmetric(self, name: str, module: nn.Module) -> bool: if not self.skip_attention_qkv: return False cfg = getattr(self.model, \"config\", None) num_heads = getattr(cfg, \"num_attention_heads\", None) num_kv_heads = getattr(cfg, \"num_key_value_heads\", None) if num_heads is not None and num_kv_heads is not None and num_kv_heads < num_heads: name_lower = name.lower() if any(p in name_lower for p in [\"q_proj\", \"k_proj\", \"v_proj\", \"query\", \"key\", \"value\", \"c_attn\"]",
"registry_names": [
"direct_svd_compressor",
"direct_svd",
"direct_svd_compressor_factory",
"direct_svd_factory"
]
},
"params": [],
"metrics": {
"results/unified_direct_svd/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": 6.01
},
"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": 27189.71
},
"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": "2026-09-15T11:57:09.550344"
},
"results/unified_direct_svd/roberta/roberta-base_direct_svd_full_metrics.json": {
"model": "roberta-base_direct_svd",
"performance": {
"accuracy_pct": 71.1,
"f1_score": 0.7544,
"f1_macro": 0.7017,
"precision": 0.7342,
"recall": 0.708,
"val_loss": 0.6031,
"status": "PASS",
"elapsed_s": 18.12
},
"efficiency": {
"total_params_before": 124647170,
"total_params_after": 121255705,
"trainable_recovery_params": 17687,
"compression_rate": 1.028,
"compression_ratio_pct": 2.72,
"flops_reduction_pct": 25.0,
"num_compressed_layers": 23,
"num_fallback_layers": 49,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": 0.333649210018602,
"weighted_reconstruction_error": 0.4827,
"per_layer_reconstruction_errors": {
"roberta.encoder.layer.0.attention.self.query": 0.3703107387562295,
"roberta.encoder.layer.0.attention.self.key": 0.3394817023607664,
"roberta.encoder.layer.0.attention.self.value": 0.3563514659359639,
"roberta.encoder.layer.0.attention.output.dense": 0.3384400364748302,
"roberta.encoder.layer.0.intermediate.dense": 0.5163187700478369,
"roberta.encoder.layer.0.output.dense": 0.5529538450447853,
"roberta.encoder.layer.1.attention.self.query": 0.36171755836811004,
"roberta.encoder.layer.1.attention.self.key": 0.35123381417844457,
"roberta.encoder.layer.1.attention.self.value": 0.33529264068267295,
"roberta.encoder.layer.1.attention.output.dense": 0.32870197078071106,
"roberta.encoder.layer.1.intermediate.dense": 0.5096122136799608,
"roberta.encoder.layer.1.output.dense": 0.5198930556528292,
"roberta.encoder.layer.2.attention.self.query": 0.34546841597049116,
"roberta.encoder.layer.2.attention.self.key": 0.3480268563364002,
"roberta.encoder.layer.2.attention.self.value": 0.33624750177210216,
"roberta.encoder.layer.2.attention.output.dense": 0.3594448418595497,
"roberta.encoder.layer.2.intermediate.dense": 0.5332086957721524,
"roberta.encoder.layer.2.output.dense": 0.5196850774043572,
"roberta.encoder.layer.3.attention.self.query": 0.3590904741997009,
"roberta.encoder.layer.3.attention.self.key": 0.37341207515193203,
"roberta.encoder.layer.3.attention.self.value": 0.33872203420851466,
"roberta.encoder.layer.3.attention.output.dense": 0.35038398713152374,
"roberta.encoder.layer.3.intermediate.dense": 0.5300345952425514,
"roberta.encoder.layer.3.output.dense": 0.5197195152375925,
"roberta.encoder.layer.4.attention.self.query": 0.35064180386646643,
"roberta.encoder.layer.4.attention.self.key": 0.3568748794430075,
"roberta.encoder.layer.4.attention.self.value": 0.3239212238580438,
"roberta.encoder.layer.4.attention.output.dense": 0.34585865940733085,
"roberta.encoder.layer.4.intermediate.dense": 0.5297949731668853,
"roberta.encoder.layer.4.output.dense": 0.5187560892449229,
"roberta.encoder.layer.5.attention.self.query": 0.3570346789786721,
"roberta.encoder.layer.5.attention.self.key": 0.3638017193477438,
"roberta.encoder.layer.5.attention.self.value": 0.3330388000108121,
"roberta.encoder.layer.5.attention.output.dense": 0.3292133646089158,
"roberta.encoder.layer.5.intermediate.dense": 0.5360194653000937,
"roberta.encoder.layer.5.output.dense": 0.5208389235951708,
"roberta.encoder.layer.6.attention.self.query": 0.34146631261792515,
"roberta.encoder.layer.6.attention.self.key": 0.33123951666887913,
"roberta.encoder.layer.6.attention.self.value": 0.35261781029457534,
"roberta.encoder.layer.6.attention.output.dense": 0.38273852813424336,
"roberta.encoder.layer.6.intermediate.dense": 0.536575612143425,
"roberta.encoder.layer.6.output.dense": 0.518389068566142,
"roberta.encoder.layer.7.attention.self.query": 0.3483886957573061,
"roberta.encoder.layer.7.attention.self.key": 0.3382258149456038,
"roberta.encoder.layer.7.attention.self.value": 0.3539141056185834,
"roberta.encoder.layer.7.attention.output.dense": 0.39183288214773787,
"roberta.encoder.layer.7.intermediate.dense": 0.5462057004812805,
"roberta.encoder.layer.7.output.dense": 0.516263986381828,
"roberta.encoder.layer.8.attention.self.query": 0.3324447322322515,
"roberta.encoder.layer.8.attention.self.key": 0.3274980965078941,
"roberta.encoder.layer.8.attention.self.value": 0.43100527994642474,
"roberta.encoder.layer.8.attention.output.dense": 0.4758618542852608,
"roberta.encoder.layer.8.intermediate.dense": 0.5471793000196414,
"roberta.encoder.layer.8.output.dense": 0.521647221186488,
"roberta.encoder.layer.9.attention.self.query": 0.31622239199084934,
"roberta.encoder.layer.9.attention.self.key": 0.3233658133675395,
"roberta.encoder.layer.9.attention.self.value": 0.4491982881624973,
"roberta.encoder.layer.9.attention.output.dense": 0.4844171842677162,
"roberta.encoder.layer.9.intermediate.dense": 0.5540557790423714,
"roberta.encoder.layer.9.output.dense": 0.5578670985539997,
"roberta.encoder.layer.10.attention.self.query": 0.32379090177334036,
"roberta.encoder.layer.10.attention.self.key": 0.3213244137952926,
"roberta.encoder.layer.10.attention.self.value": 0.44072454960870894,
"roberta.encoder.layer.10.attention.output.dense": 0.49380943054786297,
"roberta.encoder.layer.10.intermediate.dense": 0.5456118880981206,
"roberta.encoder.layer.10.output.dense": 0.5676461008903744,
"roberta.encoder.layer.11.attention.self.query": 0.3275519342993735,
"roberta.encoder.layer.11.attention.self.key": 0.35018315895660346,
"roberta.encoder.layer.11.attention.self.value": 0.43122867819710065,
"roberta.encoder.layer.11.attention.output.dense": 0.47087577181920104,
"roberta.encoder.layer.11.intermediate.dense": 0.5332197157648616,
"roberta.encoder.layer.11.output.dense": 0.540284616736812
},
"per_layer_ranks": {
"roberta.encoder.layer.0.attention.self.query": 288,
"roberta.encoder.layer.0.attention.self.key": 288,
"roberta.encoder.layer.0.attention.self.value": 288,
"roberta.encoder.layer.0.attention.output.dense": 288,
"roberta.encoder.layer.0.intermediate.dense": 288,
"roberta.encoder.layer.0.output.dense": 288,
"roberta.encoder.layer.1.attention.self.query": 288,
"roberta.encoder.layer.1.attention.self.key": 288,
"roberta.encoder.layer.1.attention.self.value": 288,
"roberta.encoder.layer.1.attention.output.dense": 288,
"roberta.encoder.layer.1.intermediate.dense": 288,
"roberta.encoder.layer.1.output.dense": 288,
"roberta.encoder.layer.2.attention.self.query": 288,
"roberta.encoder.layer.2.attention.self.key": 288,
"roberta.encoder.layer.2.attention.self.value": 288,
"roberta.encoder.layer.2.attention.output.dense": 288,
"roberta.encoder.layer.2.intermediate.dense": 288,
"roberta.encoder.layer.2.output.dense": 288,
"roberta.encoder.layer.3.attention.self.query": 288,
"roberta.encoder.layer.3.attention.self.key": 288,
"roberta.encoder.layer.3.attention.self.value": 288,
"roberta.encoder.layer.3.attention.output.dense": 288,
"roberta.encoder.layer.3.intermediate.dense": 288,
"roberta.encoder.layer.3.output.dense": 288,
"roberta.encoder.layer.4.attention.self.query": 288,
"roberta.encoder.layer.4.attention.self.key": 288,
"roberta.encoder.layer.4.attention.self.value": 288,
"roberta.encoder.layer.4.attention.output.dense": 288,
"roberta.encoder.layer.4.intermediate.dense": 288,
"roberta.encoder.layer.4.output.dense": 288,
"roberta.encoder.layer.5.attention.self.query": 288,
"roberta.encoder.layer.5.attention.self.key": 288,
"roberta.encoder.layer.5.attention.self.value": 288,
"roberta.encoder.layer.5.attention.output.dense": 288,
"roberta.encoder.layer.5.intermediate.dense": 288,
"roberta.encoder.layer.5.output.dense": 288,
"roberta.encoder.layer.6.attention.self.query": 288,
"roberta.encoder.layer.6.attention.self.key": 288,
"roberta.encoder.layer.6.attention.self.value": 288,
"roberta.encoder.layer.6.attention.output.dense": 288,
"roberta.encoder.layer.6.intermediate.dense": 288,
"roberta.encoder.layer.6.output.dense": 288,
"roberta.encoder.layer.7.attention.self.query": 288,
"roberta.encoder.layer.7.attention.self.key": 288,
"roberta.encoder.layer.7.attention.self.value": 288,
"roberta.encoder.layer.7.attention.output.dense": 288,
"roberta.encoder.layer.7.intermediate.dense": 288,
"roberta.encoder.layer.7.output.dense": 288,
"roberta.encoder.layer.8.attention.self.query": 288,
"roberta.encoder.layer.8.attention.self.key": 288,
"roberta.encoder.layer.8.attention.self.value": 288,
"roberta.encoder.layer.8.attention.output.dense": 288,
"roberta.encoder.layer.8.intermediate.dense": 288,
"roberta.encoder.layer.8.output.dense": 288,
"roberta.encoder.layer.9.attention.self.query": 288,
"roberta.encoder.layer.9.attention.self.key": 288,
"roberta.encoder.layer.9.attention.self.value": 288,
"roberta.encoder.layer.9.attention.output.dense": 288,
"roberta.encoder.layer.9.intermediate.dense": 288,
"roberta.encoder.layer.9.output.dense": 288,
"roberta.encoder.layer.10.attention.self.query": 288,
"roberta.encoder.layer.10.attention.self.key": 288,
"roberta.encoder.layer.10.attention.self.value": 288,
"roberta.encoder.layer.10.attention.output.dense": 288,
"roberta.encoder.layer.10.intermediate.dense": 288,
"roberta.encoder.layer.10.output.dense": 288,
"roberta.encoder.layer.11.attention.self.query": 288,
"roberta.encoder.layer.11.attention.self.key": 288,
"roberta.encoder.layer.11.attention.self.value": 288,
"roberta.encoder.layer.11.attention.output.dense": 288,
"roberta.encoder.layer.11.intermediate.dense": 288,
"roberta.encoder.layer.11.output.dense": 288
}
},
"hardware": {
"peak_vram_mb": 567.52,
"tokens_per_sec": 22488.27
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 551.71,
"L_256_vram_mb": 554.28,
"L_512_vram_mb": 559.4,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 514)"
},
"timestamp": "2026-09-15T11:57:27.564750"
},
"results/unified_direct_svd/distilbert/distilbert_sst2_direct_svd_full_metrics.json": {
"model": "distilbert_sst2_direct_svd",
"performance": {
"accuracy_pct": 88.99,
"f1_score": 0.8977,
"f1_macro": 0.8893,
"precision": 0.8957,
"recall": 0.8888,
"val_loss": 0.4521,
"status": "PASS",
"elapsed_s": 9.62
},
"efficiency": {
"total_params_before": 66955010,
"total_params_after": 65333005,
"trainable_recovery_params": 4874507,
"compression_rate": 1.0248,
"compression_ratio_pct": 2.42,
"flops_reduction_pct": 25.0,
"num_compressed_layers": 11,
"num_fallback_layers": 25,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": 0.32805258539915744,
"weighted_reconstruction_error": 0.4929,
"per_layer_reconstruction_errors": {
"distilbert.transformer.layer.0.attention.q_lin": 0.33294541661031835,
"distilbert.transformer.layer.0.attention.k_lin": 0.3265202541608498,
"distilbert.transformer.layer.0.attention.v_lin": 0.3559604153550347,
"distilbert.transformer.layer.0.attention.out_lin": 0.3352648988329486,
"distilbert.transformer.layer.0.ffn.lin1": 0.5717170084644833,
"distilbert.transformer.layer.0.ffn.lin2": 0.5505473029793996,
"distilbert.transformer.layer.1.attention.q_lin": 0.31197218404705623,
"distilbert.transformer.layer.1.attention.k_lin": 0.2890370222440952,
"distilbert.transformer.layer.1.attention.v_lin": 0.3489629867258555,
"distilbert.transformer.layer.1.attention.out_lin": 0.34796077734124503,
"distilbert.transformer.layer.1.ffn.lin1": 0.5548181322884734,
"distilbert.transformer.layer.1.ffn.lin2": 0.5314462363554651,
"distilbert.transformer.layer.2.attention.q_lin": 0.3488535159884333,
"distilbert.transformer.layer.2.attention.k_lin": 0.3538382337756789,
"distilbert.transformer.layer.2.attention.v_lin": 0.35176663591496565,
"distilbert.transformer.layer.2.attention.out_lin": 0.36874548794530815,
"distilbert.transformer.layer.2.ffn.lin1": 0.5535604719779885,
"distilbert.transformer.layer.2.ffn.lin2": 0.5115649996108725,
"distilbert.transformer.layer.3.attention.q_lin": 0.3081674504479119,
"distilbert.transformer.layer.3.attention.k_lin": 0.3102934415071935,
"distilbert.transformer.layer.3.attention.v_lin": 0.40359065599188454,
"distilbert.transformer.layer.3.attention.out_lin": 0.4109987666440988,
"distilbert.transformer.layer.3.ffn.lin1": 0.550859812760822,
"distilbert.transformer.layer.3.ffn.lin2": 0.5527078702792994,
"distilbert.transformer.layer.4.attention.q_lin": 0.350365252447347,
"distilbert.transformer.layer.4.attention.k_lin": 0.3486004914848246,
"distilbert.transformer.layer.4.attention.v_lin": 0.41083232183624463,
"distilbert.transformer.layer.4.attention.out_lin": 0.43395432050394933,
"distilbert.transformer.layer.4.ffn.lin1": 0.5301574971423181,
"distilbert.transformer.layer.4.ffn.lin2": 0.5913154852573249,
"distilbert.transformer.layer.5.attention.q_lin": 0.3565237878224119,
"distilbert.transformer.layer.5.attention.k_lin": 0.3672168487396547,
"distilbert.transformer.layer.5.attention.v_lin": 0.3866447990220312,
"distilbert.transformer.layer.5.attention.out_lin": 0.4223256701252535,
"distilbert.transformer.layer.5.ffn.lin1": 0.5482504297865282,
"distilbert.transformer.layer.5.ffn.lin2": 0.5672468217305555
},
"per_layer_ranks": {
"distilbert.transformer.layer.0.attention.q_lin": 288,
"distilbert.transformer.layer.0.attention.k_lin": 288,
"distilbert.transformer.layer.0.attention.v_lin": 288,
"distilbert.transformer.layer.0.attention.out_lin": 288,
"distilbert.transformer.layer.0.ffn.lin1": 288,
"distilbert.transformer.layer.0.ffn.lin2": 288,
"distilbert.transformer.layer.1.attention.q_lin": 288,
"distilbert.transformer.layer.1.attention.k_lin": 288,
"distilbert.transformer.layer.1.attention.v_lin": 288,
"distilbert.transformer.layer.1.attention.out_lin": 288,
"distilbert.transformer.layer.1.ffn.lin1": 288,
"distilbert.transformer.layer.1.ffn.lin2": 288,
"distilbert.transformer.layer.2.attention.q_lin": 288,
"distilbert.transformer.layer.2.attention.k_lin": 288,
"distilbert.transformer.layer.2.attention.v_lin": 288,
"distilbert.transformer.layer.2.attention.out_lin": 288,
"distilbert.transformer.layer.2.ffn.lin1": 288,
"distilbert.transformer.layer.2.ffn.lin2": 288,
"distilbert.transformer.layer.3.attention.q_lin": 288,
"distilbert.transformer.layer.3.attention.k_lin": 288,
"distilbert.transformer.layer.3.attention.v_lin": 288,
"distilbert.transformer.layer.3.attention.out_lin": 288,
"distilbert.transformer.layer.3.ffn.lin1": 288,
"distilbert.transformer.layer.3.ffn.lin2": 288,
"distilbert.transformer.layer.4.attention.q_lin": 288,
"distilbert.transformer.layer.4.attention.k_lin": 288,
"distilbert.transformer.layer.4.attention.v_lin": 288,
"distilbert.transformer.layer.4.attention.out_lin": 288,
"distilbert.transformer.layer.4.ffn.lin1": 288,
"distilbert.transformer.layer.4.ffn.lin2": 288,
"distilbert.transformer.layer.5.attention.q_lin": 288,
"distilbert.transformer.layer.5.attention.k_lin": 288,
"distilbert.transformer.layer.5.attention.v_lin": 288,
"distilbert.transformer.layer.5.attention.out_lin": 288,
"distilbert.transformer.layer.5.ffn.lin1": 288,
"distilbert.transformer.layer.5.ffn.lin2": 288
}
},
"hardware": {
"peak_vram_mb": 355.29,
"tokens_per_sec": 43766.17
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 340.23,
"L_256_vram_mb": 342.79,
"L_512_vram_mb": "unsupported (exceeds model max_position_embeddings 512)",
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 512)"
},
"timestamp": "2026-09-15T11:56:39.235499"
},
"results/unified_direct_svd/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": 2.99
},
"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": 51084.83
},
"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": "2026-09-15T11:56:29.577632"
},
"results/unified_direct_svd/gpt2/gpt2_direct_svd_full_metrics.json": {
"model": "gpt2_direct_svd",
"performance": {
"accuracy_pct": 68.46,
"f1_score": 0.5827,
"f1_macro": 0.6646,
"precision": 0.7547,
"recall": 0.6893,
"val_loss": 0.5809,
"status": "PASS",
"elapsed_s": 15.06
},
"efficiency": {
"total_params_before": 124441344,
"total_params_after": 123409158,
"trainable_recovery_params": 4614,
"compression_rate": 1.0084,
"compression_ratio_pct": 0.83,
"flops_reduction_pct": 29.17,
"num_compressed_layers": 6,
"num_fallback_layers": 42,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": 0.3114239479225169,
"weighted_reconstruction_error": 0.524,
"per_layer_reconstruction_errors": {
"transformer.h.0.attn.c_attn": 0.4359572020367505,
"transformer.h.0.attn.c_proj": 0.24632889107399192,
"transformer.h.0.mlp.c_fc": 0.5544663538146155,
"transformer.h.0.mlp.c_proj": 0.5604420561104402,
"transformer.h.1.attn.c_attn": 0.42981757107437263,
"transformer.h.1.attn.c_proj": 0.26518290770340747,
"transformer.h.1.mlp.c_fc": 0.5585745638699833,
"transformer.h.1.mlp.c_proj": 0.5834010201501821,
"transformer.h.2.attn.c_attn": 0.4713939852405788,
"transformer.h.2.attn.c_proj": 0.337749362027825,
"transformer.h.2.mlp.c_fc": 0.5486908429297742,
"transformer.h.2.mlp.c_proj": 0.5505523496363166,
"transformer.h.3.attn.c_attn": 0.5088508736089505,
"transformer.h.3.attn.c_proj": 0.35170269651428393,
"transformer.h.3.mlp.c_fc": 0.551661220535963,
"transformer.h.3.mlp.c_proj": 0.5310940934569812,
"transformer.h.4.attn.c_attn": 0.4893240697737358,
"transformer.h.4.attn.c_proj": 0.33817276930938883,
"transformer.h.4.mlp.c_fc": 0.5397920971944391,
"transformer.h.4.mlp.c_proj": 0.5395393092947177,
"transformer.h.5.attn.c_attn": 0.4637313021309327,
"transformer.h.5.attn.c_proj": 0.373032599943282,
"transformer.h.5.mlp.c_fc": 0.5400867249121613,
"transformer.h.5.mlp.c_proj": 0.5472925966574819,
"transformer.h.6.attn.c_attn": 0.49409677027041077,
"transformer.h.6.attn.c_proj": 0.33396203276137054,
"transformer.h.6.mlp.c_fc": 0.5420108271681355,
"transformer.h.6.mlp.c_proj": 0.5530053285816784,
"transformer.h.7.attn.c_attn": 0.4988713253041192,
"transformer.h.7.attn.c_proj": 0.35901756212671,
"transformer.h.7.mlp.c_fc": 0.5463364530349697,
"transformer.h.7.mlp.c_proj": 0.5600505762513335,
"transformer.h.8.attn.c_attn": 0.5397303605224265,
"transformer.h.8.attn.c_proj": 0.3471477246591176,
"transformer.h.8.mlp.c_fc": 0.5549730283254868,
"transformer.h.8.mlp.c_proj": 0.5551377539607433,
"transformer.h.9.attn.c_attn": 0.562183706224178,
"transformer.h.9.attn.c_proj": 0.4080779835075269,
"transformer.h.9.mlp.c_fc": 0.5528578597251608,
"transformer.h.9.mlp.c_proj": 0.5596791821537409,
"transformer.h.10.attn.c_attn": 0.5318939388245411,
"transformer.h.10.attn.c_proj": 0.4560525477285183,
"transformer.h.10.mlp.c_fc": 0.5488322046166907,
"transformer.h.10.mlp.c_proj": 0.5708025478388962,
"transformer.h.11.attn.c_attn": 0.5156999913750293,
"transformer.h.11.attn.c_proj": 0.37547696152560417,
"transformer.h.11.mlp.c_fc": 0.5401652307583459,
"transformer.h.11.mlp.c_proj": 0.5733697989137505
},
"per_layer_ranks": {
"transformer.h.0.attn.c_attn": 288,
"transformer.h.0.attn.c_proj": 192,
"transformer.h.0.mlp.c_fc": 288,
"transformer.h.0.mlp.c_proj": 288,
"transformer.h.1.attn.c_attn": 288,
"transformer.h.1.attn.c_proj": 288,
"transformer.h.1.mlp.c_fc": 288,
"transformer.h.1.mlp.c_proj": 288,
"transformer.h.2.attn.c_attn": 288,
"transformer.h.2.attn.c_proj": 288,
"transformer.h.2.mlp.c_fc": 288,
"transformer.h.2.mlp.c_proj": 288,
"transformer.h.3.attn.c_attn": 288,
"transformer.h.3.attn.c_proj": 288,
"transformer.h.3.mlp.c_fc": 288,
"transformer.h.3.mlp.c_proj": 288,
"transformer.h.4.attn.c_attn": 288,
"transformer.h.4.attn.c_proj": 288,
"transformer.h.4.mlp.c_fc": 288,
"transformer.h.4.mlp.c_proj": 288,
"transformer.h.5.attn.c_attn": 288,
"transformer.h.5.attn.c_proj": 288,
"transformer.h.5.mlp.c_fc": 288,
"transformer.h.5.mlp.c_proj": 288,
"transformer.h.6.attn.c_attn": 288,
"transformer.h.6.attn.c_proj": 288,
"transformer.h.6.mlp.c_fc": 288,
"transformer.h.6.mlp.c_proj": 288,
"transformer.h.7.attn.c_attn": 288,
"transformer.h.7.attn.c_proj": 288,
"transformer.h.7.mlp.c_fc": 288,
"transformer.h.7.mlp.c_proj": 288,
"transformer.h.8.attn.c_attn": 288,
"transformer.h.8.attn.c_proj": 288,
"transformer.h.8.mlp.c_fc": 288,
"transformer.h.8.mlp.c_proj": 288,
"transformer.h.9.attn.c_attn": 288,
"transformer.h.9.attn.c_proj": 288,
"transformer.h.9.mlp.c_fc": 288,
"transformer.h.9.mlp.c_proj": 288,
"transformer.h.10.attn.c_attn": 288,
"transformer.h.10.attn.c_proj": 288,
"transformer.h.10.mlp.c_fc": 288,
"transformer.h.10.mlp.c_proj": 288,
"transformer.h.11.attn.c_attn": 288,
"transformer.h.11.attn.c_proj": 288,
"transformer.h.11.mlp.c_fc": 288,
"transformer.h.11.mlp.c_proj": 288
}
},
"hardware": {
"peak_vram_mb": 646.14,
"tokens_per_sec": 25639.93
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 529.51,
"L_256_vram_mb": 543.02,
"L_512_vram_mb": 570.27,
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 1024)"
},
"timestamp": "2026-09-15T11:58:13.214122"
},
"results/unified_direct_svd/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": 6.0
},
"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": 26772.56
},
"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": "2026-09-15T11:57:58.086112"
},
"results/unified_direct_svd/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,78.56,0.8124,0.7811,0.8051,0.7832,0.4381,6.01,712.34,27189.71,124647170,124647170,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_roberta,2026-09-15T11:57:10.261225\ndirect_svd,PASS,71.1,0.7544,0.7017,0.7342,0.708,0.6031,18.12,567.52,22488.27,124647170,121255705,17687,1.028,2.72,25.0,23,49,0.333649210018602,0.4827,0,0,,adapted_roberta,2026-09-15T11:57:28.638964\n",
"results/unified_direct_svd/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,90.71,0.9099,0.907,0.9075,0.9068,0.3027,2.99,389.51,51084.83,66955010,66955010,0,1.0,0.0,0.0,0,0,,,0,0,,legacy_bert,2026-09-15T11:56:29.791172\ndirect_svd,PASS,88.99,0.8977,0.8893,0.8957,0.8888,0.4521,9.62,355.29,43766.17,66955010,65333005,4874507,1.0248,2.42,25.0,11,25,0.32805258539915744,0.4929,0,0,,legacy_bert,2026-09-15T11:56:39.623735\n",
"results/unified_direct_svd/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,79.36,0.7978,0.7935,0.7935,0.7935,0.4875,6.01,793.19,26772.56,124441344,124441344,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_gpt,2026-09-15T11:57:58.771237\ndirect_svd,PASS,68.46,0.5827,0.6646,0.7547,0.6893,0.5809,15.07,646.14,25639.93,124441344,123409158,4614,1.0084,0.83,29.17,6,42,0.3114239479225169,0.524,0,0,,adapted_gpt,2026-09-15T11:58:14.093825\n",
"results/matrix_test_summary.csv": "qwen3_5,qwen3_5,direct_svd,compressor,PASS,True,True,113.92,4777.1,\ngpt2,gpt2,direct_svd,compressor,PASS,True,True,0.215,1634.6,\nroberta,roberta,direct_svd,compressor,PASS,True,True,4.224,1328.2,\ndistilbert,distilbert,direct_svd,compressor,PASS,True,True,2.167,638.7,",
"results/benchmark_summary_v2.csv": "track1_svd,direct_svd,compressor,bert,sst2,PASS,,13.8,21.02,34.82,0.0,1.0,4054402,4054402,1.082,717.9,0.0,0.6928,0.0,0.0,0.0,0.0701,\"{'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': [13.4431, 13.4669, 14.1316, 14.1405, 13.8082], 'bwd_latency_ms': [21.0086, 21.2413, 21.5977, 20.3537, 20.8857]}\",2026-08-26 02:32:20\ntrack1_svd,direct_svd,compressor,bert,mrpc,PASS,,13.75,20.96,34.71,0.0,1.0,4054402,4054402,1.082,717.9,0.0,0.69,0.0,0.0,0.0,0.7129,\"{'loss': [0.69, 0.69, 0.69, 0.69, 0.69], 'grad_norm': [0.7129, 0.7129, 0.7129, 0.7129, 0.7129], 'fwd_latency_ms': [13.8384, 13.5891, 13.2539, 14.3129, 13.7736], 'bwd_latency_ms': [21.1866, 21.2764, 20.3508, 20.7283, 21.2413]}\",2026-08-26 02:32:20\ntrack1_svd,direct_svd,compressor,roberta,sst2,PASS,,14.82,21.42,36.24,0.0,1.0,6593794,6593794,1.048,727.9,0.0,0.694,0.0,0.0,0.0,0.1935,\"{'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': [14.4465, 15.2783, 14.6228, 15.3099, 14.427], 'bwd_latency_ms': [22.3037, 21.1687, 19.5734, 22.9328, 21.1097]}\",2026-08-26 02:32:21\ntrack1_svd,direct_svd,compressor,roberta,mrpc,PASS,,14.48,21.28,35.76,0.0,1.0,6593794,6593794,1.048,727.9,0.0,0.6996,0.0,0.0,0.0,1.3032,\"{'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': [14.4046, 14.9389, 14.3359, 14.5126, 14.1994], 'bwd_latency_ms': [22.2108, 22.1213, 20.9296, 20.0739, 21.0886]}\",2026-08-26 02:32:22\ntrack1_svd,direct_svd,compressor,t5,sst2,PASS,,17.84,32.32,50.16,0.0,1.0,4186880,4186880,1.076,1082.2,0.0,0.7776,0.0,0.0,0.0,1.7821,\"{'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': [14.293, 18.7902, 18.8088, 18.5783, 18.7157], 'bwd_latency_ms': [32.5744, 32.1306, 32.2486, 32.4147, 32.2542]}\",2026-08-26 02:32:23\ntrack1_svd,direct_svd,compressor,t5,mrpc,PASS,,17.75,32.13,49.88,0.0,1.0,4186880,4186880,1.076,1082.2,0.0,0.8314,0.0,0.0,0.0,2.7308,\"{'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': [14.2746, 18.4909, 18.7261, 18.701, 18.5573], 'bwd_latency_ms': [32.398, 31.8665, 32.5026, 31.9252, 31.9612]}\",2026-08-26 02:32:23",
"results/all_121_compressors_gpu_matrix.csv": "direct_svd,cuda,PASS,True,True,30.96,106.1,0.776,\ndirect_svd_compressor,cuda,PASS,True,True,30.96,106.1,0.768,\ndirect_svd_encoder,cuda,PASS,True,True,30.96,117.6,0.776,",
"results/run_quality_all_screening_summary.csv": "run_quality_20260827_052444,direct_svd,DISTILBERT,SST2,PASS,64.91,0.0,64.91,64.91,1.0,0.6483,0.426,86.82,207.27,1591.9,55710.1,1450000000.0,2026-08-27T07:04:00.286022"
},
"notes": []
},
"dobi": {
"module_name": "dobi",
"source_file": "Codes/framework/compression_engine/dobi.py",
"source_lines_count": 24,
"architecture": {
"class_name": null,
"base_class": null,
"init_args": null,
"registry_names": []
},
"params": [],
"metrics": {
"results/unified_dobi/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": 6.29
},
"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": 25886.74
},
"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": "2026-09-16T00:17:22.147639"
},
"results/unified_dobi/roberta/roberta-base_dobi_svd_full_metrics.json": {
"model": "roberta-base_dobi_svd",
"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": 9.77
},
"efficiency": {
"total_params_before": 124647170,
"total_params_after": 124647170,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 520.18,
"tokens_per_sec": 26585.15
},
"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": "2026-09-16T00:17:31.924533"
},
"results/unified_dobi/distilbert/distilbert_sst2_full_full_metrics.json": {
"model": "distilbert_sst2_full",
"performance": {
"accuracy_pct": 90.71,
"f1_score": 0.9099,
"f1_macro": 0.907,
"precision": 0.9075,
"recall": 0.9068,
"val_loss": 0.3027,
"status": "PASS",
"elapsed_s": 3.02
},
"efficiency": {
"total_params_before": 66955010,
"total_params_after": 66955010,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 389.51,
"tokens_per_sec": 50749.29
},
"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": "2026-09-16T00:16:45.338167"
},
"results/unified_dobi/distilbert/distilbert_sst2_dobi_svd_full_metrics.json": {
"model": "distilbert_sst2_dobi_svd",
"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": 4.75
},
"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": 300.68,
"tokens_per_sec": 49740.42
},
"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": "2026-09-16T00:16:50.239106"
},
"results/unified_dobi/qwen3_5/Qwen3.5-0.8B_dobi_svd_full_metrics.json": {
"model": "Qwen3.5-0.8B_dobi_svd",
"performance": {
"accuracy_pct": 89.45,
"f1_score": 0.895,
"f1_macro": 0.8945,
"precision": 0.8946,
"recall": 0.8947,
"val_loss": 0.3863,
"status": "PASS",
"elapsed_s": 214.18
},
"efficiency": {
"total_params_before": 763217728,
"total_params_after": 763217728,
"trainable_recovery_params": 7285760,
"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": 10822656,
"avg_reconstruction_error": 0.0,
"weighted_reconstruction_error": 0.6949,
"per_layer_reconstruction_errors": {
"model.layers.0.mlp.up_proj": 0.7430297337211321,
"model.layers.0.mlp.down_proj": 0.7394550079989763,
"model.layers.1.mlp.up_proj": 0.733678906950749,
"model.layers.1.mlp.down_proj": 0.718051830179827,
"model.layers.2.mlp.up_proj": 0.7425124407483829,
"model.layers.2.mlp.down_proj": 0.7237513781872895,
"model.layers.3.mlp.up_proj": 0.735582751223258,
"model.layers.3.mlp.down_proj": 0.7104599001183056,
"model.layers.4.mlp.up_proj": 0.7343217734797287,
"model.layers.4.mlp.down_proj": 0.7173982222163473,
"model.layers.5.mlp.up_proj": 0.7262027942216849,
"model.layers.5.mlp.down_proj": 0.7016467383889237,
"model.layers.6.mlp.up_proj": 0.7028053038930362,
"model.layers.6.mlp.down_proj": 0.671106751865246,
"model.layers.7.mlp.up_proj": 0.7120696218768726,
"model.layers.7.mlp.down_proj": 0.6819015521034267,
"model.layers.8.mlp.up_proj": 0.7000975100781869,
"model.layers.8.mlp.down_proj": 0.673564832675475,
"model.layers.9.mlp.up_proj": 0.6844613673069624,
"model.layers.9.mlp.down_proj": 0.6525873349808995,
"model.layers.10.mlp.up_proj": 0.687787585376745,
"model.layers.10.mlp.down_proj": 0.6639056068947407,
"model.layers.11.mlp.up_proj": 0.6833026271695272,
"model.layers.11.mlp.down_proj": 0.6524889495501047,
"model.layers.12.mlp.up_proj": 0.6847964648743877,
"model.layers.12.mlp.down_proj": 0.6552156772622032,
"model.layers.13.mlp.up_proj": 0.6837471265911037,
"model.layers.13.mlp.down_proj": 0.6490350213311139,
"model.layers.14.mlp.up_proj": 0.6795140258827872,
"model.layers.14.mlp.down_proj": 0.6789343477603668,
"model.layers.15.mlp.up_proj": 0.6998216545134613,
"model.layers.15.mlp.down_proj": 0.6832670399511876,
"model.layers.16.mlp.up_proj": 0.6961287066497458,
"model.layers.16.mlp.down_proj": 0.6897547749134175,
"model.layers.17.mlp.up_proj": 0.6995477553452417,
"model.layers.17.mlp.down_proj": 0.6831262040253073,
"model.layers.18.mlp.up_proj": 0.6991997923049648,
"model.layers.18.mlp.down_proj": 0.6776158776356218,
"model.layers.19.mlp.up_proj": 0.6901671841658696,
"model.layers.19.mlp.down_proj": 0.6977915941544546,
"model.layers.20.mlp.up_proj": 0.6914188938665812,
"model.layers.20.mlp.down_proj": 0.6991622298062018,
"model.layers.21.mlp.up_proj": 0.697074915127813,
"model.layers.21.mlp.down_proj": 0.6968834353157872,
"model.layers.22.mlp.up_proj": 0.6888831399165025,
"model.layers.22.mlp.down_proj": 0.6940081363850439,
"model.layers.23.mlp.up_proj": 0.6637766319778717,
"model.layers.23.mlp.down_proj": 0.6842083249487748
},
"per_layer_ranks": {
"model.layers.0.mlp.up_proj": 256,
"model.layers.0.mlp.down_proj": 256,
"model.layers.1.mlp.up_proj": 256,
"model.layers.1.mlp.down_proj": 256,
"model.layers.2.mlp.up_proj": 256,
"model.layers.2.mlp.down_proj": 256,
"model.layers.3.mlp.up_proj": 256,
"model.layers.3.mlp.down_proj": 256,
"model.layers.4.mlp.up_proj": 256,
"model.layers.4.mlp.down_proj": 256,
"model.layers.5.mlp.up_proj": 256,
"model.layers.5.mlp.down_proj": 256,
"model.layers.6.mlp.up_proj": 256,
"model.layers.6.mlp.down_proj": 256,
"model.layers.7.mlp.up_proj": 256,
"model.layers.7.mlp.down_proj": 256,
"model.layers.8.mlp.up_proj": 256,
"model.layers.8.mlp.down_proj": 256,
"model.layers.9.mlp.up_proj": 256,
"model.layers.9.mlp.down_proj": 256,
"model.layers.10.mlp.up_proj": 256,
"model.layers.10.mlp.down_proj": 256,
"model.layers.11.mlp.up_proj": 256,
"model.layers.11.mlp.down_proj": 256,
"model.layers.12.mlp.up_proj": 256,
"model.layers.12.mlp.down_proj": 256,
"model.layers.13.mlp.up_proj": 256,
"model.layers.13.mlp.down_proj": 256,
"model.layers.14.mlp.up_proj": 256,
"model.layers.14.mlp.down_proj": 256,
"model.layers.15.mlp.up_proj": 256,
"model.layers.15.mlp.down_proj": 256,
"model.layers.16.mlp.up_proj": 256,
"model.layers.16.mlp.down_proj": 256,
"model.layers.17.mlp.up_proj": 256,
"model.layers.17.mlp.down_proj": 256,
"model.layers.18.mlp.up_proj": 256,
"model.layers.18.mlp.down_proj": 256,
"model.layers.19.mlp.up_proj": 256,
"model.layers.19.mlp.down_proj": 256,
"model.layers.20.mlp.up_proj": 256,
"model.layers.20.mlp.down_proj": 256,
"model.layers.21.mlp.up_proj": 256,
"model.layers.21.mlp.down_proj": 256,
"model.layers.22.mlp.up_proj": 256,
"model.layers.22.mlp.down_proj": 256,
"model.layers.23.mlp.up_proj": 256,
"model.layers.23.mlp.down_proj": 256
}
},
"hardware": {
"peak_vram_mb": 1850.64,
"tokens_per_sec": 4747.76
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 1587.33,
"L_256_vram_mb": 1608.73,
"L_512_vram_mb": 1636.73,
"L_1024_vram_mb": 1726.87
},
"timestamp": "2026-09-08T19:12:00.070969"
},
"results/unified_dobi/qwen3_5/Qwen3.5-0.8B_full_full_metrics.json": {
"model": "Qwen3.5-0.8B_full",
"performance": {
"accuracy_pct": 67.89,
"f1_score": 0.557,
"f1_macro": 0.6526,
"precision": 0.7722,
"recall": 0.6842,
"val_loss": 0.6349,
"status": "PASS",
"elapsed_s": 25.29
},
"efficiency": {
"total_params_before": 763217728,
"total_params_after": 763217728,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 10822656,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 1796.78,
"tokens_per_sec": 5935.34
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 1532.47,
"L_256_vram_mb": 1556.68,
"L_512_vram_mb": 1591.55,
"L_1024_vram_mb": 1672.53
},
"timestamp": "2026-09-08T19:08:25.790152"
},
"results/unified_dobi/gpt2/gpt2_dobi_svd_full_metrics.json": {
"model": "gpt2_dobi_svd",
"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": 8.6
},
"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": 26552.97
},
"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": "2026-09-16T00:18:18.826661"
},
"results/unified_dobi/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": 6.02
},
"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": 26758.23
},
"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": "2026-09-16T00:18:09.964420"
},
"results/unified_dobi/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,78.56,0.8124,0.7811,0.8051,0.7832,0.4381,6.29,712.34,25886.74,124647170,124647170,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_roberta,2026-09-16T00:17:22.901391\ndobi_svd,PASS_NO_COMPRESSION,78.56,0.8124,0.7811,0.8051,0.7832,0.4381,9.77,520.18,26585.15,124647170,124647170,0,1.0,0.0,0.0,0,0,,,0,0,\"note: zero compressed layers, model grew (LoRA only)\",adapted_roberta,2026-09-16T00:17:32.897460\n",
"results/unified_dobi/distilbert/summary.csv": "module_name,status,accuracy_pct,f1_score,f1_macro,precision,recall,val_loss,elapsed_s,peak_vram_mb,tokens_per_sec,total_params_before,total_params_after,trainable_recovery_params,compression_rate,compression_ratio_pct,flops_reduction_pct,num_compressed_layers,num_fallback_layers,avg_reconstruction_error,weighted_reconstruction_error,num_injected_attention_layers,lora_params,error,path,timestamp\nfull,PASS,90.71,0.9099,0.907,0.9075,0.9068,0.3027,3.02,389.51,50749.29,66955010,66955010,0,1.0,0.0,0.0,0,0,,,0,0,,legacy_bert,2026-09-16T00:16:45.557942\ndobi_svd,PASS_NO_COMPRESSION,90.71,0.9099,0.907,0.9075,0.9068,0.3027,4.75,300.68,49740.42,66955010,66955010,0,1.0,0.0,0.0,0,0,,,0,0,\"note: zero compressed layers, model grew (LoRA only)\",legacy_bert,2026-09-16T00:16:50.531949\n",
"results/unified_dobi/qwen3_5/summary.csv": "module_name,status,accuracy_pct,f1_score,f1_macro,precision,recall,val_loss,elapsed_s,peak_vram_mb,tokens_per_sec,total_params_before,total_params_after,trainable_recovery_params,compression_rate,compression_ratio_pct,flops_reduction_pct,num_compressed_layers,num_fallback_layers,avg_reconstruction_error,weighted_reconstruction_error,num_injected_attention_layers,lora_params,error,path,timestamp\nfull,PASS,67.89,0.557,0.6526,0.7722,0.6842,0.6349,25.29,1796.78,5935.34,763217728,763217728,0,1.0,0.0,0.0,0,0,,,0,10822656,,adapted_qwen3_5,2026-09-08T19:08:26.856697\ndobi_svd,PASS,89.45,0.895,0.8945,0.8946,0.8947,0.3863,214.18,1850.64,4747.76,763217728,763217728,7285760,1.0,0.0,0.0,0,48,0.0,0.6949,0,10822656,,adapted_qwen3_5,2026-09-08T19:12:01.195530\n",
"results/unified_dobi/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,79.36,0.7978,0.7935,0.7935,0.7935,0.4875,6.03,793.19,26758.23,124441344,124441344,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_gpt,2026-09-16T00:18:10.658835\ndobi_svd,PASS_NO_COMPRESSION,79.36,0.7978,0.7935,0.7935,0.7935,0.4875,8.6,627.16,26552.97,124441344,124441344,0,1.0,0.0,0.0,0,0,,,0,0,\"note: zero compressed layers, model grew (LoRA only)\",adapted_gpt,2026-09-16T00:18:19.513805\n",
"results/agnews_unified_dobi/distilbert/distilbert_sst2_dobi_full_metrics.json": {
"model": "distilbert_sst2_dobi",
"performance": {
"accuracy_pct": 70.7,
"f1_score": 0.7041,
"f1_macro": 0.7041,
"precision": 0.7102,
"recall": 0.7026,
"val_loss": 0.8413,
"status": "PASS_NO_COMPRESSION",
"elapsed_s": 2.71
},
"efficiency": {
"total_params_before": 66956548,
"total_params_after": 66956548,
"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": 449.77,
"tokens_per_sec": 244854.42
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 288.34,
"L_256_vram_mb": 290.9,
"L_512_vram_mb": "unsupported (exceeds model max_position_embeddings 512)",
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 512)"
},
"timestamp": "2026-09-16T04:24:33.244502"
},
"results/agnews_unified_dobi/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\ndobi,PASS_NO_COMPRESSION,70.7,0.7041,0.7041,0.7102,0.7026,0.8413,2.72,449.77,244854.42,66956548,66956548,0,1.0,0.0,0.0,0,0,,,0,0,\"note: zero compressed layers, model grew (LoRA only)\",legacy_bert,2026-09-16T04:24:33.430864\n",
"results/mnli_unified_dobi/distilbert/distilbert_sst2_dobi_full_metrics.json": {
"model": "distilbert_sst2_dobi",
"performance": {
"accuracy_pct": 41.1,
"f1_score": 0.3087,
"f1_macro": 0.3087,
"precision": 0.2892,
"recall": 0.4026,
"val_loss": 1.0808,
"status": "PASS_NO_COMPRESSION",
"elapsed_s": 3.35
},
"efficiency": {
"total_params_before": 66955779,
"total_params_after": 66955779,
"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": 319.55,
"tokens_per_sec": 112285.64
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 284.87,
"L_256_vram_mb": 287.43,
"L_512_vram_mb": "unsupported (exceeds model max_position_embeddings 512)",
"L_1024_vram_mb": "unsupported (exceeds model max_position_embeddings 512)"
},
"timestamp": "2026-09-16T03:13:12.128679"
},
"results/mnli_unified_dobi/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\ndobi,PASS_NO_COMPRESSION,41.1,0.3087,0.3087,0.2892,0.4026,1.0808,3.35,319.55,112285.64,66955779,66955779,0,1.0,0.0,0.0,0,0,,,0,0,\"note: zero compressed layers, model grew (LoRA only)\",legacy_bert,2026-09-16T03:13:12.313826\n",
"results/benchmark_summary_v2.csv": "track1_svd,dobi_svd,compressor,bert,sst2,PASS,,15.65,25.21,40.86,0.0,1.0,4054610,4054610,1.082,582.0,0.0,0.6928,0.0,0.0,0.0,0.0671,\"{'loss': [0.6928, 0.6928, 0.6928, 0.6928, 0.6928], 'grad_norm': [0.0671, 0.0671, 0.0671, 0.0671, 0.0671], 'fwd_latency_ms': [15.9059, 16.0228, 15.2461, 15.3268, 15.7502], 'bwd_latency_ms': [25.054, 25.17, 25.2269, 25.3459, 25.235]}\",2026-08-26 02:32:24\ntrack1_svd,dobi_svd,compressor,bert,mrpc,PASS,,15.92,25.48,41.4,0.0,1.0,4054610,4054610,1.082,582.0,0.0,0.6902,0.0,0.0,0.0,0.6702,\"{'loss': [0.6902, 0.6902, 0.6902, 0.6902, 0.6902], 'grad_norm': [0.6702, 0.6702, 0.6702, 0.6702, 0.6702], 'fwd_latency_ms': [15.7756, 16.1996, 16.0163, 15.7822, 15.8037], 'bwd_latency_ms': [24.8412, 26.4035, 25.2852, 25.0269, 25.8339]}\",2026-08-26 02:32:25\ntrack1_svd,dobi_svd,compressor,roberta,sst2,PASS,,16.11,25.03,41.14,0.0,1.0,6593986,6593986,1.048,591.9,0.0,0.694,0.0,0.0,0.0,0.1932,\"{'loss': [0.694, 0.694, 0.694, 0.694, 0.694], 'grad_norm': [0.1932, 0.1932, 0.1932, 0.1932, 0.1932], 'fwd_latency_ms': [16.0573, 15.9117, 16.1847, 15.4126, 16.9917], 'bwd_latency_ms': [25.3072, 25.0968, 24.44, 25.4615, 24.8668]}\",2026-08-26 02:32:26\ntrack1_svd,dobi_svd,compressor,roberta,mrpc,PASS,,16.23,25.53,41.76,0.0,1.0,6593986,6593986,1.048,591.9,0.0,0.6997,0.0,0.0,0.0,1.3027,\"{'loss': [0.6997, 0.6997, 0.6997, 0.6997, 0.6997], 'grad_norm': [1.3027, 1.3027, 1.3027, 1.3027, 1.3027], 'fwd_latency_ms': [15.1572, 16.0162, 18.1689, 15.9686, 15.8379], 'bwd_latency_ms': [25.1027, 24.5874, 27.2205, 25.737, 24.991]}\",2026-08-26 02:32:26\ntrack1_svd,dobi_svd,compressor,t5,sst2,PASS,,18.86,34.07,52.93,0.0,1.0,4187072,4187072,1.076,1082.2,0.0,0.7721,0.0,0.0,0.0,1.4914,\"{'loss': [0.7721, 0.7721, 0.7721, 0.7721, 0.7721], 'grad_norm': [1.4914, 1.4914, 1.4914, 1.4914, 1.4914], 'fwd_latency_ms': [15.7092, 19.9225, 19.8869, 19.626, 19.1612], 'bwd_latency_ms': [34.8141, 34.6326, 34.5214, 33.1527, 33.2528]}\",2026-08-26 02:32:27\ntrack1_svd,dobi_svd,compressor,t5,mrpc,PASS,,18.47,33.1,51.57,0.0,1.0,4187072,4187072,1.076,1082.2,0.0,0.8268,0.0,0.0,0.0,2.3147,\"{'loss': [0.8268, 0.8268, 0.8268, 0.8268, 0.8268], 'grad_norm': [2.3147, 2.3147, 2.3147, 2.3147, 2.3147], 'fwd_latency_ms': [14.5377, 20.0858, 18.9979, 19.3912, 19.3124], 'bwd_latency_ms': [33.5959, 33.3719, 32.6615, 33.0961, 32.7661]}\",2026-08-26 02:32:28",
"results/master_aggregated_results.csv": "dobi,0,FAILED_LOW,0,0,0,167.8,2026-09-11T11:14:07.331307",
"results/all_121_compressors_gpu_matrix.csv": "dobi,cuda,PASS,True,True,30.95,106.1,0.794,\ndobi_svd,cuda,PASS,True,True,30.95,106.1,0.778,",
"results/run_quality_all_screening_summary.csv": "run_quality_20260827_052444,dobi,DISTILBERT,SST2,PASS,64.56,0.0,64.56,64.56,1.0,0.6403,0.4583,48.09,123.92,1304.5,95254.9,1450000000.0,2026-08-27T07:05:29.649461"
},
"notes": []
},
"dynamic_entropy": {
"module_name": "dynamic_entropy",
"source_file": "Codes/framework/compression_engine/dynamic_entropy.py",
"source_lines_count": 32,
"architecture": {
"class_name": null,
"base_class": null,
"init_args": null,
"registry_names": []
},
"params": [],
"metrics": {
"results/unified_dynamic_entropy/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": 6.02
},
"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": 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": 27387.12
},
"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": "2026-09-15T12:02:38.069262"
},
"results/unified_dynamic_entropy/roberta/roberta-base_dynamic_entropy_svd_full_metrics.json": {
"model": "roberta-base_dynamic_entropy_svd",
"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": 12.55
},
"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": 520.18,
"tokens_per_sec": 27253.7
},
"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": "2026-09-15T12:02:50.788828"
},
"results/unified_dynamic_entropy/distilbert/distilbert_sst2_dynamic_entropy_svd_full_metrics.json": {
"model": "distilbert_sst2_dynamic_entropy_svd",
"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": 6.13
},
"efficiency": {
"total_params_before": 66955010,
"total_params_after": 66955010,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 300.68,
"tokens_per_sec": 52783.18
},
"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": "2026-09-15T12:02:08.016345"
},
"results/unified_dynamic_entropy/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": 2.94
},
"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": 51240.09
},
"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": "2026-09-15T12:02:01.712280"
},
"results/unified_dynamic_entropy/qwen3_5/Qwen3.5-0.8B_dynamic_entropy_svd_full_metrics.json": {
"model": "Qwen3.5-0.8B_dynamic_entropy_svd",
"performance": {
"accuracy_pct": 89.22,
"f1_score": 0.8951,
"f1_macro": 0.8921,
"precision": 0.8924,
"recall": 0.892,
"val_loss": 0.3674,
"status": "PASS",
"elapsed_s": 203.11
},
"efficiency": {
"total_params_before": 763217728,
"total_params_after": 763217728,
"trainable_recovery_params": 7285760,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 10822656,
"avg_reconstruction_error": 0.0,
"weighted_reconstruction_error": 0.6949,
"per_layer_reconstruction_errors": {
"model.layers.0.mlp.up_proj": 0.7430295010046372,
"model.layers.0.mlp.down_proj": 0.7394549494112934,
"model.layers.1.mlp.up_proj": 0.7336789651268197,
"model.layers.1.mlp.down_proj": 0.7180517142499085,
"model.layers.2.mlp.up_proj": 0.7425123796909935,
"model.layers.2.mlp.down_proj": 0.7237513177239042,
"model.layers.3.mlp.up_proj": 0.735582751223258,
"model.layers.3.mlp.down_proj": 0.7104597151887342,
"model.layers.4.mlp.up_proj": 0.7343217121864365,
"model.layers.4.mlp.down_proj": 0.7173980386267597,
"model.layers.5.mlp.up_proj": 0.7262027942216849,
"model.layers.5.mlp.down_proj": 0.7016465574366151,
"model.layers.6.mlp.up_proj": 0.7028051848082598,
"model.layers.6.mlp.down_proj": 0.6711066303023385,
"model.layers.7.mlp.up_proj": 0.7120696218768726,
"model.layers.7.mlp.down_proj": 0.6819013102193774,
"model.layers.8.mlp.up_proj": 0.7000973908872833,
"model.layers.8.mlp.down_proj": 0.673564708806392,
"model.layers.9.mlp.up_proj": 0.6844613082380049,
"model.layers.9.mlp.down_proj": 0.6525871509592649,
"model.layers.10.mlp.up_proj": 0.6877874057528272,
"model.layers.10.mlp.down_proj": 0.6639054817560636,
"model.layers.11.mlp.up_proj": 0.6833024499845082,
"model.layers.11.mlp.down_proj": 0.6524888880331192,
"model.layers.12.mlp.up_proj": 0.6847964044043701,
"model.layers.12.mlp.down_proj": 0.6552153639936379,
"model.layers.13.mlp.up_proj": 0.6837469467782816,
"model.layers.13.mlp.down_proj": 0.6490350213311139,
"model.layers.14.mlp.up_proj": 0.679513907650115,
"model.layers.14.mlp.down_proj": 0.678934284778556,
"model.layers.15.mlp.up_proj": 0.6998216545134613,
"model.layers.15.mlp.down_proj": 0.6832669239522687,
"model.layers.16.mlp.up_proj": 0.696128818579862,
"model.layers.16.mlp.down_proj": 0.6897546583981197,
"model.layers.17.mlp.up_proj": 0.6995476995965355,
"model.layers.17.mlp.down_proj": 0.683126087631102,
"model.layers.18.mlp.up_proj": 0.6991996825295762,
"model.layers.18.mlp.down_proj": 0.6776155874503943,
"model.layers.19.mlp.up_proj": 0.6901671841658696,
"model.layers.19.mlp.down_proj": 0.6977914283490915,
"model.layers.20.mlp.up_proj": 0.6914187893346954,
"model.layers.20.mlp.down_proj": 0.6991621208548929,
"model.layers.21.mlp.up_proj": 0.6970748641642397,
"model.layers.21.mlp.down_proj": 0.6968832737067995,
"model.layers.22.mlp.up_proj": 0.6888830404341886,
"model.layers.22.mlp.down_proj": 0.6940082452533943,
"model.layers.23.mlp.up_proj": 0.6637766768634799,
"model.layers.23.mlp.down_proj": 0.6842082087655157
},
"per_layer_ranks": {
"model.layers.0.mlp.up_proj": 256,
"model.layers.0.mlp.down_proj": 256,
"model.layers.1.mlp.up_proj": 256,
"model.layers.1.mlp.down_proj": 256,
"model.layers.2.mlp.up_proj": 256,
"model.layers.2.mlp.down_proj": 256,
"model.layers.3.mlp.up_proj": 256,
"model.layers.3.mlp.down_proj": 256,
"model.layers.4.mlp.up_proj": 256,
"model.layers.4.mlp.down_proj": 256,
"model.layers.5.mlp.up_proj": 256,
"model.layers.5.mlp.down_proj": 256,
"model.layers.6.mlp.up_proj": 256,
"model.layers.6.mlp.down_proj": 256,
"model.layers.7.mlp.up_proj": 256,
"model.layers.7.mlp.down_proj": 256,
"model.layers.8.mlp.up_proj": 256,
"model.layers.8.mlp.down_proj": 256,
"model.layers.9.mlp.up_proj": 256,
"model.layers.9.mlp.down_proj": 256,
"model.layers.10.mlp.up_proj": 256,
"model.layers.10.mlp.down_proj": 256,
"model.layers.11.mlp.up_proj": 256,
"model.layers.11.mlp.down_proj": 256,
"model.layers.12.mlp.up_proj": 256,
"model.layers.12.mlp.down_proj": 256,
"model.layers.13.mlp.up_proj": 256,
"model.layers.13.mlp.down_proj": 256,
"model.layers.14.mlp.up_proj": 256,
"model.layers.14.mlp.down_proj": 256,
"model.layers.15.mlp.up_proj": 256,
"model.layers.15.mlp.down_proj": 256,
"model.layers.16.mlp.up_proj": 256,
"model.layers.16.mlp.down_proj": 256,
"model.layers.17.mlp.up_proj": 256,
"model.layers.17.mlp.down_proj": 256,
"model.layers.18.mlp.up_proj": 256,
"model.layers.18.mlp.down_proj": 256,
"model.layers.19.mlp.up_proj": 256,
"model.layers.19.mlp.down_proj": 256,
"model.layers.20.mlp.up_proj": 256,
"model.layers.20.mlp.down_proj": 256,
"model.layers.21.mlp.up_proj": 256,
"model.layers.21.mlp.down_proj": 256,
"model.layers.22.mlp.up_proj": 256,
"model.layers.22.mlp.down_proj": 256,
"model.layers.23.mlp.up_proj": 256,
"model.layers.23.mlp.down_proj": 256
}
},
"hardware": {
"peak_vram_mb": 1850.64,
"tokens_per_sec": 4908.12
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 1587.33,
"L_256_vram_mb": 1608.73,
"L_512_vram_mb": 1636.73,
"L_1024_vram_mb": 1726.87
},
"timestamp": "2026-09-08T19:30:14.218168"
},
"results/unified_dynamic_entropy/qwen3_5/Qwen3.5-0.8B_full_full_metrics.json": {
"model": "Qwen3.5-0.8B_full",
"performance": {
"accuracy_pct": 67.89,
"f1_score": 0.557,
"f1_macro": 0.6526,
"precision": 0.7722,
"recall": 0.6842,
"val_loss": 0.6349,
"status": "PASS",
"elapsed_s": 25.08
},
"efficiency": {
"total_params_before": 763217728,
"total_params_after": 763217728,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 0,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 10822656,
"avg_reconstruction_error": null,
"weighted_reconstruction_error": null,
"per_layer_reconstruction_errors": {},
"per_layer_ranks": {}
},
"hardware": {
"peak_vram_mb": 1796.78,
"tokens_per_sec": 6042.7
},
"architecture": {
"span_distribution": {},
"mean_span": 128.0,
"sparsity_ratio_percent": 0.0
},
"scalability": {
"L_128_vram_mb": 1532.47,
"L_256_vram_mb": 1556.68,
"L_512_vram_mb": 1591.55,
"L_1024_vram_mb": 1672.53
},
"timestamp": "2026-09-08T19:26:50.899193"
},
"results/unified_dynamic_entropy/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": 6.02
},
"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": 26678.84
},
"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": "2026-09-15T12:03:21.521980"
},
"results/unified_dynamic_entropy/gpt2/gpt2_dynamic_entropy_svd_full_metrics.json": {
"model": "gpt2_dynamic_entropy_svd",
"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": 11.16
},
"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": 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": 26725.83
},
"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": "2026-09-15T12:03:32.930607"
},
"results/unified_dynamic_entropy/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,78.56,0.8124,0.7811,0.8051,0.7832,0.4381,6.02,712.34,27387.12,124647170,124647170,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_roberta,2026-09-15T12:02:38.831334\ndynamic_entropy_svd,PASS_NO_COMPRESSION,78.56,0.8124,0.7811,0.8051,0.7832,0.4381,12.55,520.18,27253.7,124647170,124647170,0,1.0,0.0,0.0,0,0,,,0,0,\"note: zero compressed layers, model grew (LoRA only)\",adapted_roberta,2026-09-15T12:02:51.642009\n",
"results/unified_dynamic_entropy/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,90.71,0.9099,0.907,0.9075,0.9068,0.3027,2.94,389.51,51240.09,66955010,66955010,0,1.0,0.0,0.0,0,0,,,0,0,,legacy_bert,2026-09-15T12:02:01.923866\ndynamic_entropy_svd,PASS_NO_COMPRESSION,90.71,0.9099,0.907,0.9075,0.9068,0.3027,6.13,300.68,52783.18,66955010,66955010,0,1.0,0.0,0.0,0,0,,,0,0,\"note: zero compressed layers, model grew (LoRA only)\",legacy_bert,2026-09-15T12:02:08.316748\n",
"results/unified_dynamic_entropy/qwen3_5/summary.csv": "module_name,status,accuracy_pct,f1_score,f1_macro,precision,recall,val_loss,elapsed_s,peak_vram_mb,tokens_per_sec,total_params_before,total_params_after,trainable_recovery_params,compression_rate,compression_ratio_pct,flops_reduction_pct,num_compressed_layers,num_fallback_layers,avg_reconstruction_error,weighted_reconstruction_error,num_injected_attention_layers,lora_params,error,path,timestamp\nfull,PASS,67.89,0.557,0.6526,0.7722,0.6842,0.6349,25.08,1796.78,6042.7,763217728,763217728,0,1.0,0.0,0.0,0,0,,,0,10822656,,adapted_qwen3_5,2026-09-08T19:26:51.971051\ndynamic_entropy_svd,PASS,89.22,0.8951,0.8921,0.8924,0.892,0.3674,203.11,1850.64,4908.12,763217728,763217728,7285760,1.0,0.0,0.0,0,0,0.0,0.6949,0,10822656,,adapted_qwen3_5,2026-09-08T19:30:15.404774\n",
"results/unified_dynamic_entropy/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,79.36,0.7978,0.7935,0.7935,0.7935,0.4875,6.02,793.19,26678.84,124441344,124441344,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_gpt,2026-09-15T12:03:22.217920\ndynamic_entropy_svd,PASS_NO_COMPRESSION,79.36,0.7978,0.7935,0.7935,0.7935,0.4875,11.16,627.16,26725.83,124441344,124441344,0,1.0,0.0,0.0,0,0,,,0,0,\"note: zero compressed layers, model grew (LoRA only)\",adapted_gpt,2026-09-15T12:03:33.641450\n",
"results/master_aggregated_results.csv": "dynamic_entropy,0,FAILED_LOW,0,0,0,264.4,2026-09-11T11:18:31.702538",
"results/all_121_compressors_gpu_matrix.csv": "dynamic_entropy,cuda,PASS,True,True,30.96,104.9,1.029,\ndynamic_entropy_svd,cuda,PASS,True,True,30.96,104.9,1.007,"
},
"notes": []
},
"dynamic_rank": {
"module_name": "dynamic_rank",
"source_file": "Codes/framework/compression_engine/dynamic_rank.py",
"source_lines_count": 933,
"architecture": {
"class_name": "DynamicSVDLinear",
"base_class": "nn.Module",
"init_args": "self, in_features: int, out_features: int, rank: int, bias: bool = True, dtype: torch.dtype = torch.float32, device: Optional[torch.device] = None, ) -> None: super().__init__() self.in_features = in_features self.out_features = out_features self.rank = max(1, min(rank, in_features, out_features)) # Factorized low-rank parameter tensors self.V = nn.Parameter( torch.empty(self.rank, in_features, dtype=dtype, device=device) ) self.U = nn.Parameter( torch.empty(out_features, self.rank, dtype=dtype, device=device) ) if bias: self.bias = nn.Parameter( torch.empty(out_features, dtype=dtype, device=device) ) else: self.register_parameter(\"bias\", None) self.reset_parameters() @property def weight(self) -> torch.Tensor: \"\"\"Reconstructed effective weight matrix for compatibility.\"\"\" return torch.matmul(self.U, self.V) def reset_parameters(self) -> None: \"\"\"Kaiming uniform parameter initialization.\"\"\" nn.init.kaiming_uniform_(self.V, a=math.sqrt(5)) nn.init.kaiming_uniform_(self.U, a=math.sqrt(5)) if self.bias is not None: fan_in = self.in_features bound = 1 / math.sqrt(fan_in) if fan_in > 0 else 0 nn.init.uniform_(self.bias, -bound, bound) def forward(self, x: torch.Tensor) -> torch.Tensor: \"\"\" Forward pass executing factorized low-rank multiplication: y = (x @ V^T) @ U^T + bias \"\"\" intermediate = F.linear(x, self.V) out = F.linear(intermediate, self.U, self.bias) return out @classmethod def from_linear( cls, linear: nn.Module, rank: int, covariance: Optional[torch.Tensor] = None, ) -> DynamicSVDLinear: \"\"\" Construct factorized DynamicSVDLinear directly from dense nn.Linear or Conv1D via SVD. Args: linear: Source dense linear or Conv1D layer (or LoRALinear wrapper). rank: Target rank for low-rank factorization. covariance: Optional empirical activation covariance C_X for activation-aware SVD. Returns: DynamicSVDLinear: Initialized factorized low-rank layer. \"\"\" target_mod = getattr(linear, \"base_layer\", linear) cls_name = target_mod.__class__.__name__.lower() is_conv1d = \"conv1d\" in cls_name or hasattr(target_mod, \"nf\") raw_w = target_mod.weight.data if is_conv1d: in_f, out_f = raw_w.shape[0], raw_w.shape[1] W = raw_w.t().float() # [out_features, in_features] else: out_f, in_f = raw_w.shape[0], raw_w.shape[1] W = raw_w.float() # [out_features, in_features] target_rank = max(1, min(rank, in_f, out_f)) new_layer = cls( in_features=in_f, out_features=out_f, rank=target_rank, bias=target_mod.bias is not None, dtype=target_mod.weight.dtype, device=target_mod.weight.device, ) if covariance is not None and covariance.numel() > 0: # RESEARCH FIX: Activation-aware whitening SVD with eigenvalue clamping lambda_i >= 1e-4 cov = covariance.to(device=W.device, dtype=torch.float32) try: # Eigen-decomposition with regularized eigenvalue clamping >= 1e-4 eigvals, eigvecs = torch.linalg.eigh(cov + 1e-4 * torch.eye(in_f, device=W.device)) eigvals_clamped = torch.clamp(eigvals, min=1e-4) cov_sqrt = eigvecs @ torch.diag(torch.sqrt(eigvals_clamped)) @ eigvecs.t() cov_inv_sqrt = ( eigvecs @ torch.diag(1.0 / torch.sqrt(eigvals_clamped)) @ eigvecs.t() ) W_scaled = W @ cov_sqrt U_svd, S_svd, Vh_svd = torch.linalg.svd(W_scaled, full_matrices=False) # Truncate and un-whiten right singular factors sqrt_S = torch.sqrt(torch.clamp(S_svd[:target_rank], min=1e-12)) U_r = U_svd[:, :target_rank] * sqrt_S.unsqueeze(0) V_r = ( sqrt_S.unsqueeze(1) * (Vh_svd[:target_rank, :] @ cov_inv_sqrt) ) # Symmetrically normalize factor magnitudes norm_u = torch.norm(U_r, p=\"fro\").clamp_min(1e-8) norm_v = torch.norm(V_r, p=\"fro\").clamp_min(1e-8) scale = torch.sqrt(norm_u / norm_v) U_r = U_r / scale V_r = V_r * scale except Exception as e: logger.warning( f\"Activation-aware SVD failed ({e}), falling back to standard SVD.\" ) U_svd, S_svd, Vh_svd = torch.linalg.svd(W, full_matrices=False) sqrt_S = torch.sqrt(torch.clamp(S_svd[:target_rank], min=1e-12)) U_r = U_svd[:, :target_rank] * sqrt_S.unsqueeze(0) V_r = sqrt_S.unsqueeze(1) * Vh_svd[:target_rank, :] else: # Standard weight SVD U_svd, S_svd, Vh_svd = torch.linalg.svd(W, full_matrices=False) sqrt_S = torch.sqrt(torch.clamp(S_svd[:target_rank], min=1e-12)) U_r = U_svd[:, :target_rank] * sqrt_S.unsqueeze(0) V_r = sqrt_S.unsqueeze(1) * Vh_svd[:target_rank, :] new_layer.U.data.copy_(U_r.to(dtype=target_mod.weight.dtype)) new_layer.V.data.copy_(V_r.to(dtype=target_mod.weight.dtype)) if target_mod.bias is not None and new_layer.bias is not None: new_layer.bias.data.copy_(target_mod.bias.data) return new_layer def extra_repr(self) -> str: return f\"in_features={self.in_features}, out_features={self.out_features}, rank={self.rank}, bias={self.bias is not None}\" class DynamicEntropySVDCompressor(BaseCompressor",
"registry_names": [
"dynamic_rank_compressor",
"entropy_rank_allocator_class",
"dynamic_rank_factory",
"dynamic_entropy_class",
"dynamic_entropy_svd_class",
"dynamic_entropy",
"dynamic_rank",
"dynamic_rank_inject",
"dynamic_entropy_svd",
"dynamic_rank_class",
"entropy_rank_allocator",
"dynamic_entropy_inject",
"inject_dynamic_rank",
"dynamic_entropy_compressor",
"dynamic_entropy_svd_factory",
"inject_dynamic_entropy",
"dynamic_entropy_factory",
"dynamic_entropy_svd_compressor"
]
},
"params": [],
"metrics": {
"results/benchmark_summary_v2.csv": "track1_svd,dynamic_rank,compressor,bert,sst2,PASS,,13.75,20.72,34.47,0.0,1.0,4054402,4054402,1.082,581.9,0.0,0.6928,0.0,0.0,0.0,0.0701,\"{'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': [13.7833, 13.9396, 14.0605, 13.4042, 13.5813], 'bwd_latency_ms': [19.9187, 20.9725, 20.5509, 21.2706, 20.8895]}\",2026-08-26 02:34:03\ntrack1_svd,dynamic_rank,compressor,bert,mrpc,PASS,,13.71,20.99,34.7,0.0,1.0,4054402,4054402,1.082,581.9,0.0,0.69,0.0,0.0,0.0,0.7128,\"{'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': [14.1396, 13.1303, 14.0134, 13.5196, 13.7676], 'bwd_latency_ms': [21.3251, 20.4349, 21.2504, 20.8031, 21.1384]}\",2026-08-26 02:34:04\ntrack1_svd,dynamic_rank,compressor,roberta,sst2,PASS,,13.92,21.0,34.92,0.0,1.0,6593794,6593794,1.048,591.9,0.0,0.694,0.0,0.0,0.0,0.1935,\"{'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': [14.3944, 13.9263, 13.3978, 13.5269, 14.335], 'bwd_latency_ms': [21.6905, 20.937, 20.7188, 21.2063, 20.4257]}\",2026-08-26 02:34:05\ntrack1_svd,dynamic_rank,compressor,roberta,mrpc,PASS,,12.29,18.64,30.93,0.0,1.0,6593794,6593794,1.048,591.9,0.0,0.6996,0.0,0.0,0.0,1.3032,\"{'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': [11.8874, 12.1484, 12.7091, 12.2536, 12.4268], 'bwd_latency_ms': [18.3094, 18.9094, 18.811, 18.7932, 18.3686]}\",2026-08-26 02:34:06\ntrack1_svd,dynamic_rank,compressor,t5,sst2,PASS,,17.61,31.86,49.47,0.0,1.0,4186880,4186880,1.076,1082.2,0.0,0.7776,0.0,0.0,0.0,1.7821,\"{'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': [14.3145, 18.3587, 19.4294, 18.0572, 17.8667], 'bwd_latency_ms': [31.3784, 31.2794, 34.4222, 31.322, 30.8831]}\",2026-08-26 02:34:06\ntrack1_svd,dynamic_rank,compressor,t5,mrpc,PASS,,18.01,32.29,50.3,0.0,1.0,4186880,4186880,1.076,1082.2,0.0,0.8314,0.0,0.0,0.0,2.7308,\"{'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': [14.2572, 18.5935, 18.903, 19.049, 19.2289], 'bwd_latency_ms': [32.6026, 32.3482, 31.9441, 32.3122, 32.2269]}\",2026-08-26 02:34:07",
"results/all_121_compressors_gpu_matrix.csv": "dynamic_rank,cuda,PASS,True,True,30.96,104.9,1.007,",
"results/run_quality_all_screening_summary.csv": "run_quality_20260827_052444,dynamic_rank,DISTILBERT,SST2,PASS,67.09,0.0,67.09,67.09,1.0,0.6496,0.4824,43.72,121.13,1304.4,99387.4,1450000000.0,2026-08-27T07:07:26.506502"
},
"notes": []
},
"enlighten_compressor": {
"module_name": "enlighten_compressor",
"source_file": "Codes/framework/compression_engine/enlighten_compressor.py",
"source_lines_count": 378,
"architecture": {
"class_name": "ENLightenCompressor",
"base_class": "BaseCompressor",
"init_args": "self, model: nn.Module, rank: int = 64, sparsity: float = 0.5, structured_24: bool = True, max_error: float = 0.35, target_attention_only: bool = False, fallback_to_original: bool = True, max_retries: int = 2, target_modules: Optional[List[str]] = None, cfg: Optional[Any] = None, **kwargs: Any, ) -> None: super().__init__(model, name=\"enlighten\", **kwargs) cfg = cfg or kwargs.get(\"comp_config\", None) self.cfg = cfg self.rank = int(_cfg_get(cfg, \"enlighten_rank\", _cfg_get(cfg, \"rank\", rank))) self.sparsity = float(_cfg_get(cfg, \"enlighten_sparsity\", _cfg_get(cfg, \"sparsity\", sparsity))) self.structured_24 = bool(_cfg_get(cfg, \"enlighten_structured_24\", _cfg_get(cfg, \"structured_24\", structured_24))) self.max_error = float(_cfg_get(cfg, \"enlighten_max_error\", _cfg_get(cfg, \"max_error\", max_error))) self.target_attention_only = bool( _cfg_get(cfg, \"enlighten_target_attention_only\", _cfg_get(cfg, \"target_attention_only\", target_attention_only)) ) self.fallback_to_original = bool( _cfg_get(cfg, \"enlighten_fallback_to_original\", _cfg_get(cfg, \"fallback_to_original\", fallback_to_original)) ) self.max_retries = int(max_retries) self.target_modules = target_modules self.compressed_layers: List[str] = [] self.fallback_layers: List[str] = [] self.skipped_layers: List[str] = [] self.layer_ranks: Dict[str, int] = {} self.layer_sparsities: Dict[str, float] = {} self.layer_errors: Dict[str, float] = {} def _is_target_module(self, name: str, module: nn.Module) -> bool: \"\"\"Determine if a module is targeted for ENLighten compression.\"\"\" name_lower = name.lower() if any(part in name.split(\".\") for part in [\"lora_A\", \"lora_B\", \"base_layer\"]",
"registry_names": [
"enlighten"
]
},
"params": [],
"metrics": {
"results/unified_enlighten/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": 5.92
},
"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": 27357.52
},
"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": "2026-09-15T12:06:54.785202"
},
"results/unified_enlighten/roberta/roberta-base_enlighten_full_metrics.json": {
"model": "roberta-base_enlighten",
"performance": {
"accuracy_pct": 51.03,
"f1_score": 0.6748,
"f1_macro": 0.342,
"precision": 0.5882,
"recall": 0.5012,
"val_loss": 1.3388,
"status": "PASS_LOW",
"elapsed_s": 10.78
},
"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": 72,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": 0.23880489796199525,
"weighted_reconstruction_error": 0.2498,
"per_layer_reconstruction_errors": {
"roberta.encoder.layer.0.attention.self.query": 0.23359410877187917,
"roberta.encoder.layer.0.attention.self.key": 0.2199394167495101,
"roberta.encoder.layer.0.attention.self.value": 0.2259357866210929,
"roberta.encoder.layer.0.attention.output.dense": 0.21450913442945838,
"roberta.encoder.layer.0.intermediate.dense": 0.24719659811986736,
"roberta.encoder.layer.0.output.dense": 0.2524459146000799,
"roberta.encoder.layer.1.attention.self.query": 0.224996364333746,
"roberta.encoder.layer.1.attention.self.key": 0.21926901584043035,
"roberta.encoder.layer.1.attention.self.value": 0.21647183676065984,
"roberta.encoder.layer.1.attention.output.dense": 0.21359114573258903,
"roberta.encoder.layer.1.intermediate.dense": 0.25173964067047944,
"roberta.encoder.layer.1.output.dense": 0.2509060712398215,
"roberta.encoder.layer.2.attention.self.query": 0.2226190745755108,
"roberta.encoder.layer.2.attention.self.key": 0.2216084956783261,
"roberta.encoder.layer.2.attention.self.value": 0.21670281410470554,
"roberta.encoder.layer.2.attention.output.dense": 0.22823324221061733,
"roberta.encoder.layer.2.intermediate.dense": 0.2623928967820663,
"roberta.encoder.layer.2.output.dense": 0.25503942030826493,
"roberta.encoder.layer.3.attention.self.query": 0.2233833518099069,
"roberta.encoder.layer.3.attention.self.key": 0.2300322643046937,
"roberta.encoder.layer.3.attention.self.value": 0.21800157250283148,
"roberta.encoder.layer.3.attention.output.dense": 0.2282865743182875,
"roberta.encoder.layer.3.intermediate.dense": 0.2624802489230301,
"roberta.encoder.layer.3.output.dense": 0.2560982591950917,
"roberta.encoder.layer.4.attention.self.query": 0.22393233238833174,
"roberta.encoder.layer.4.attention.self.key": 0.2260537429946565,
"roberta.encoder.layer.4.attention.self.value": 0.2101052961492022,
"roberta.encoder.layer.4.attention.output.dense": 0.2280954819743462,
"roberta.encoder.layer.4.intermediate.dense": 0.26096697574926003,
"roberta.encoder.layer.4.output.dense": 0.25496070392208114,
"roberta.encoder.layer.5.attention.self.query": 0.22237426388960366,
"roberta.encoder.layer.5.attention.self.key": 0.2277525100175555,
"roberta.encoder.layer.5.attention.self.value": 0.21489850136502825,
"roberta.encoder.layer.5.attention.output.dense": 0.22079788963616528,
"roberta.encoder.layer.5.intermediate.dense": 0.26217289054673154,
"roberta.encoder.layer.5.output.dense": 0.255695276126833,
"roberta.encoder.layer.6.attention.self.query": 0.2192173160404123,
"roberta.encoder.layer.6.attention.self.key": 0.2143906449228088,
"roberta.encoder.layer.6.attention.self.value": 0.2263037444991734,
"roberta.encoder.layer.6.attention.output.dense": 0.2395149647311086,
"roberta.encoder.layer.6.intermediate.dense": 0.2628771617431653,
"roberta.encoder.layer.6.output.dense": 0.2557856291273397,
"roberta.encoder.layer.7.attention.self.query": 0.22549785523079602,
"roberta.encoder.layer.7.attention.self.key": 0.2195705108237813,
"roberta.encoder.layer.7.attention.self.value": 0.2222060091292126,
"roberta.encoder.layer.7.attention.output.dense": 0.24093505658871156,
"roberta.encoder.layer.7.intermediate.dense": 0.2666111313866375,
"roberta.encoder.layer.7.output.dense": 0.2560824856746715,
"roberta.encoder.layer.8.attention.self.query": 0.21982354826665926,
"roberta.encoder.layer.8.attention.self.key": 0.21673932271330462,
"roberta.encoder.layer.8.attention.self.value": 0.24421518241804596,
"roberta.encoder.layer.8.attention.output.dense": 0.2651098438687603,
"roberta.encoder.layer.8.intermediate.dense": 0.2657575809331783,
"roberta.encoder.layer.8.output.dense": 0.25826005759469733,
"roberta.encoder.layer.9.attention.self.query": 0.20971765167296008,
"roberta.encoder.layer.9.attention.self.key": 0.21336315736073083,
"roberta.encoder.layer.9.attention.self.value": 0.24702654568516189,
"roberta.encoder.layer.9.attention.output.dense": 0.26443375941655023,
"roberta.encoder.layer.9.intermediate.dense": 0.26792446942173026,
"roberta.encoder.layer.9.output.dense": 0.2711686298322109,
"roberta.encoder.layer.10.attention.self.query": 0.21760826194370225,
"roberta.encoder.layer.10.attention.self.key": 0.21378155149697245,
"roberta.encoder.layer.10.attention.self.value": 0.2505069606947846,
"roberta.encoder.layer.10.attention.output.dense": 0.2666394445581009,
"roberta.encoder.layer.10.intermediate.dense": 0.26812166245182295,
"roberta.encoder.layer.10.output.dense": 0.2788956762158734,
"roberta.encoder.layer.11.attention.self.query": 0.21611862184670969,
"roberta.encoder.layer.11.attention.self.key": 0.22605492247873765,
"roberta.encoder.layer.11.attention.self.value": 0.2574120648438032,
"roberta.encoder.layer.11.attention.output.dense": 0.26692238429072224,
"roberta.encoder.layer.11.intermediate.dense": 0.26454700814812004,
"roberta.encoder.layer.11.output.dense": 0.2715327218697576
},
"per_layer_ranks": {
"roberta.encoder.layer.0.attention.self.query": 128,
"roberta.encoder.layer.0.attention.self.key": 128,
"roberta.encoder.layer.0.attention.self.value": 128,
"roberta.encoder.layer.0.attention.output.dense": 128,
"roberta.encoder.layer.0.intermediate.dense": 128,
"roberta.encoder.layer.0.output.dense": 128,
"roberta.encoder.layer.1.attention.self.query": 128,
"roberta.encoder.layer.1.attention.self.key": 128,
"roberta.encoder.layer.1.attention.self.value": 128,
"roberta.encoder.layer.1.attention.output.dense": 128,
"roberta.encoder.layer.1.intermediate.dense": 128,
"roberta.encoder.layer.1.output.dense": 128,
"roberta.encoder.layer.2.attention.self.query": 128,
"roberta.encoder.layer.2.attention.self.key": 128,
"roberta.encoder.layer.2.attention.self.value": 128,
"roberta.encoder.layer.2.attention.output.dense": 128,
"roberta.encoder.layer.2.intermediate.dense": 128,
"roberta.encoder.layer.2.output.dense": 128,
"roberta.encoder.layer.3.attention.self.query": 128,
"roberta.encoder.layer.3.attention.self.key": 128,
"roberta.encoder.layer.3.attention.self.value": 128,
"roberta.encoder.layer.3.attention.output.dense": 128,
"roberta.encoder.layer.3.intermediate.dense": 128,
"roberta.encoder.layer.3.output.dense": 128,
"roberta.encoder.layer.4.attention.self.query": 128,
"roberta.encoder.layer.4.attention.self.key": 128,
"roberta.encoder.layer.4.attention.self.value": 128,
"roberta.encoder.layer.4.attention.output.dense": 128,
"roberta.encoder.layer.4.intermediate.dense": 128,
"roberta.encoder.layer.4.output.dense": 128,
"roberta.encoder.layer.5.attention.self.query": 128,
"roberta.encoder.layer.5.attention.self.key": 128,
"roberta.encoder.layer.5.attention.self.value": 128,
"roberta.encoder.layer.5.attention.output.dense": 128,
"roberta.encoder.layer.5.intermediate.dense": 128,
"roberta.encoder.layer.5.output.dense": 128,
"roberta.encoder.layer.6.attention.self.query": 128,
"roberta.encoder.layer.6.attention.self.key": 128,
"roberta.encoder.layer.6.attention.self.value": 128,
"roberta.encoder.layer.6.attention.output.dense": 128,
"roberta.encoder.layer.6.intermediate.dense": 128,
"roberta.encoder.layer.6.output.dense": 128,
"roberta.encoder.layer.7.attention.self.query": 128,
"roberta.encoder.layer.7.attention.self.key": 128,
"roberta.encoder.layer.7.attention.self.value": 128,
"roberta.encoder.layer.7.attention.output.dense": 128,
"roberta.encoder.layer.7.intermediate.dense": 128,
"roberta.encoder.layer.7.output.dense": 128,
"roberta.encoder.layer.8.attention.self.query": 128,
"roberta.encoder.layer.8.attention.self.key": 128,
"roberta.encoder.layer.8.attention.self.value": 128,
"roberta.encoder.layer.8.attention.output.dense": 128,
"roberta.encoder.layer.8.intermediate.dense": 128,
"roberta.encoder.layer.8.output.dense": 128,
"roberta.encoder.layer.9.attention.self.query": 128,
"roberta.encoder.layer.9.attention.self.key": 128,
"roberta.encoder.layer.9.attention.self.value": 128,
"roberta.encoder.layer.9.attention.output.dense": 128,
"roberta.encoder.layer.9.intermediate.dense": 128,
"roberta.encoder.layer.9.output.dense": 128,
"roberta.encoder.layer.10.attention.self.query": 128,
"roberta.encoder.layer.10.attention.self.key": 128,
"roberta.encoder.layer.10.attention.self.value": 128,
"roberta.encoder.layer.10.attention.output.dense": 128,
"roberta.encoder.layer.10.intermediate.dense": 128,
"roberta.encoder.layer.10.output.dense": 128,
"roberta.encoder.layer.11.attention.self.query": 128,
"roberta.encoder.layer.11.attention.self.key": 128,
"roberta.encoder.layer.11.attention.self.value": 128,
"roberta.encoder.layer.11.attention.output.dense": 128,
"roberta.encoder.layer.11.intermediate.dense": 128,
"roberta.encoder.layer.11.output.dense": 128
}
},
"hardware": {
"peak_vram_mb": 520.68,
"tokens_per_sec": 26019.95
},
"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": "2026-09-15T12:07:05.618759"
},
"results/unified_enlighten/distilbert/distilbert_sst2_enlighten_full_metrics.json": {
"model": "distilbert_sst2_enlighten",
"performance": {
"accuracy_pct": 89.22,
"f1_score": 0.8937,
"f1_macro": 0.8922,
"precision": 0.8922,
"recall": 0.8922,
"val_loss": 0.3187,
"status": "PASS",
"elapsed_s": 5.23
},
"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": 36,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": 0.24077659094076578,
"weighted_reconstruction_error": 0.2556,
"per_layer_reconstruction_errors": {
"distilbert.transformer.layer.0.attention.q_lin": 0.21429585820954436,
"distilbert.transformer.layer.0.attention.k_lin": 0.21105491977042795,
"distilbert.transformer.layer.0.attention.v_lin": 0.2248662071805046,
"distilbert.transformer.layer.0.attention.out_lin": 0.21405253730798549,
"distilbert.transformer.layer.0.ffn.lin1": 0.27885617654408384,
"distilbert.transformer.layer.0.ffn.lin2": 0.26476875468436245,
"distilbert.transformer.layer.1.attention.q_lin": 0.19853925567588368,
"distilbert.transformer.layer.1.attention.k_lin": 0.18983492188400491,
"distilbert.transformer.layer.1.attention.v_lin": 0.21998215663663873,
"distilbert.transformer.layer.1.attention.out_lin": 0.22422303634860394,
"distilbert.transformer.layer.1.ffn.lin1": 0.2736337350001992,
"distilbert.transformer.layer.1.ffn.lin2": 0.2575585055295508,
"distilbert.transformer.layer.2.attention.q_lin": 0.2230075591904642,
"distilbert.transformer.layer.2.attention.k_lin": 0.22470490273687524,
"distilbert.transformer.layer.2.attention.v_lin": 0.22196063578770103,
"distilbert.transformer.layer.2.attention.out_lin": 0.23555328913153128,
"distilbert.transformer.layer.2.ffn.lin1": 0.2732771818649986,
"distilbert.transformer.layer.2.ffn.lin2": 0.25027349623841655,
"distilbert.transformer.layer.3.attention.q_lin": 0.20662356452422995,
"distilbert.transformer.layer.3.attention.k_lin": 0.20453210591883017,
"distilbert.transformer.layer.3.attention.v_lin": 0.24929359009301683,
"distilbert.transformer.layer.3.attention.out_lin": 0.2551565237567515,
"distilbert.transformer.layer.3.ffn.lin1": 0.27298085834317165,
"distilbert.transformer.layer.3.ffn.lin2": 0.26513205331484435,
"distilbert.transformer.layer.4.attention.q_lin": 0.22372670887673482,
"distilbert.transformer.layer.4.attention.k_lin": 0.22256568649646052,
"distilbert.transformer.layer.4.attention.v_lin": 0.248877928549884,
"distilbert.transformer.layer.4.attention.out_lin": 0.2557848909747283,
"distilbert.transformer.layer.4.ffn.lin1": 0.2658159810074064,
"distilbert.transformer.layer.4.ffn.lin2": 0.28122206776340686,
"distilbert.transformer.layer.5.attention.q_lin": 0.22386260657032322,
"distilbert.transformer.layer.5.attention.k_lin": 0.22960896179191675,
"distilbert.transformer.layer.5.attention.v_lin": 0.24447707741335825,
"distilbert.transformer.layer.5.attention.out_lin": 0.2571760591681689,
"distilbert.transformer.layer.5.ffn.lin1": 0.2746299504419108,
"distilbert.transformer.layer.5.ffn.lin2": 0.28604752914064746
},
"per_layer_ranks": {
"distilbert.transformer.layer.0.attention.q_lin": 128,
"distilbert.transformer.layer.0.attention.k_lin": 128,
"distilbert.transformer.layer.0.attention.v_lin": 128,
"distilbert.transformer.layer.0.attention.out_lin": 128,
"distilbert.transformer.layer.0.ffn.lin1": 128,
"distilbert.transformer.layer.0.ffn.lin2": 128,
"distilbert.transformer.layer.1.attention.q_lin": 128,
"distilbert.transformer.layer.1.attention.k_lin": 128,
"distilbert.transformer.layer.1.attention.v_lin": 128,
"distilbert.transformer.layer.1.attention.out_lin": 128,
"distilbert.transformer.layer.1.ffn.lin1": 128,
"distilbert.transformer.layer.1.ffn.lin2": 128,
"distilbert.transformer.layer.2.attention.q_lin": 128,
"distilbert.transformer.layer.2.attention.k_lin": 128,
"distilbert.transformer.layer.2.attention.v_lin": 128,
"distilbert.transformer.layer.2.attention.out_lin": 128,
"distilbert.transformer.layer.2.ffn.lin1": 128,
"distilbert.transformer.layer.2.ffn.lin2": 128,
"distilbert.transformer.layer.3.attention.q_lin": 128,
"distilbert.transformer.layer.3.attention.k_lin": 128,
"distilbert.transformer.layer.3.attention.v_lin": 128,
"distilbert.transformer.layer.3.attention.out_lin": 128,
"distilbert.transformer.layer.3.ffn.lin1": 128,
"distilbert.transformer.layer.3.ffn.lin2": 128,
"distilbert.transformer.layer.4.attention.q_lin": 128,
"distilbert.transformer.layer.4.attention.k_lin": 128,
"distilbert.transformer.layer.4.attention.v_lin": 128,
"distilbert.transformer.layer.4.attention.out_lin": 128,
"distilbert.transformer.layer.4.ffn.lin1": 128,
"distilbert.transformer.layer.4.ffn.lin2": 128,
"distilbert.transformer.layer.5.attention.q_lin": 128,
"distilbert.transformer.layer.5.attention.k_lin": 128,
"distilbert.transformer.layer.5.attention.v_lin": 128,
"distilbert.transformer.layer.5.attention.out_lin": 128,
"distilbert.transformer.layer.5.ffn.lin1": 128,
"distilbert.transformer.layer.5.ffn.lin2": 128
}
},
"hardware": {
"peak_vram_mb": 301.18,
"tokens_per_sec": 53181.36
},
"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": "2026-09-15T12:06:25.417745"
},
"results/unified_enlighten/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": 52725.06
},
"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": "2026-09-15T12:06:20.063349"
},
"results/unified_enlighten/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": 5.97
},
"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": 26923.15
},
"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": "2026-09-15T12:07:36.229538"
},
"results/unified_enlighten/gpt2/gpt2_enlighten_full_metrics.json": {
"model": "gpt2_enlighten",
"performance": {
"accuracy_pct": 61.12,
"f1_score": 0.3936,
"f1_macro": 0.5537,
"precision": 0.7577,
"recall": 0.618,
"val_loss": 0.6589,
"status": "PASS",
"elapsed_s": 9.3
},
"efficiency": {
"total_params_before": 124441344,
"total_params_after": 124441344,
"trainable_recovery_params": 0,
"compression_rate": 1.0,
"compression_ratio_pct": 0.0,
"flops_reduction_pct": 0.0,
"num_compressed_layers": 48,
"num_fallback_layers": 0,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": 0.2547934372152875,
"weighted_reconstruction_error": 0.2629,
"per_layer_reconstruction_errors": {
"transformer.h.0.attn.c_attn": 0.2355678266521015,
"transformer.h.0.attn.c_proj": 0.15186002066296647,
"transformer.h.0.mlp.c_fc": 0.2555453915074517,
"transformer.h.0.mlp.c_proj": 0.2651136478878234,
"transformer.h.1.attn.c_attn": 0.23853487237477344,
"transformer.h.1.attn.c_proj": 0.18062418379586606,
"transformer.h.1.mlp.c_fc": 0.26879705933313675,
"transformer.h.1.mlp.c_proj": 0.26917894419503463,
"transformer.h.2.attn.c_attn": 0.25074797917890446,
"transformer.h.2.attn.c_proj": 0.22088759485915146,
"transformer.h.2.mlp.c_fc": 0.26685414225226595,
"transformer.h.2.mlp.c_proj": 0.2624043889161256,
"transformer.h.3.attn.c_attn": 0.2623313436073869,
"transformer.h.3.attn.c_proj": 0.22721794142226606,
"transformer.h.3.mlp.c_fc": 0.27355425991818394,
"transformer.h.3.mlp.c_proj": 0.2582487025394324,
"transformer.h.4.attn.c_attn": 0.25630872676162564,
"transformer.h.4.attn.c_proj": 0.22514832130804174,
"transformer.h.4.mlp.c_fc": 0.2681864998842398,
"transformer.h.4.mlp.c_proj": 0.2621328322719949,
"transformer.h.5.attn.c_attn": 0.25143179589598735,
"transformer.h.5.attn.c_proj": 0.23320374559346774,
"transformer.h.5.mlp.c_fc": 0.26996006654583143,
"transformer.h.5.mlp.c_proj": 0.2669227887051344,
"transformer.h.6.attn.c_attn": 0.26279400069715175,
"transformer.h.6.attn.c_proj": 0.22036374056710564,
"transformer.h.6.mlp.c_fc": 0.2702099213016182,
"transformer.h.6.mlp.c_proj": 0.27032981821439755,
"transformer.h.7.attn.c_attn": 0.26276234360943257,
"transformer.h.7.attn.c_proj": 0.22997262212780875,
"transformer.h.7.mlp.c_fc": 0.27291398753857865,
"transformer.h.7.mlp.c_proj": 0.2743240394068063,
"transformer.h.8.attn.c_attn": 0.27425090656116585,
"transformer.h.8.attn.c_proj": 0.22779414471276963,
"transformer.h.8.mlp.c_fc": 0.27593875661532563,
"transformer.h.8.mlp.c_proj": 0.27309185399752184,
"transformer.h.9.attn.c_attn": 0.2816210618346266,
"transformer.h.9.attn.c_proj": 0.25159612776195706,
"transformer.h.9.mlp.c_fc": 0.2755288980637915,
"transformer.h.9.mlp.c_proj": 0.2743564960120193,
"transformer.h.10.attn.c_attn": 0.2728029498806622,
"transformer.h.10.attn.c_proj": 0.25974431488474486,
"transformer.h.10.mlp.c_fc": 0.27275025720548807,
"transformer.h.10.mlp.c_proj": 0.27808697725628195,
"transformer.h.11.attn.c_attn": 0.2682411657889018,
"transformer.h.11.attn.c_proj": 0.2183378580904109,
"transformer.h.11.mlp.c_fc": 0.2673437704266328,
"transformer.h.11.mlp.c_proj": 0.27416589770940614
},
"per_layer_ranks": {
"transformer.h.0.attn.c_attn": 128,
"transformer.h.0.attn.c_proj": 128,
"transformer.h.0.mlp.c_fc": 128,
"transformer.h.0.mlp.c_proj": 128,
"transformer.h.1.attn.c_attn": 128,
"transformer.h.1.attn.c_proj": 128,
"transformer.h.1.mlp.c_fc": 128,
"transformer.h.1.mlp.c_proj": 128,
"transformer.h.2.attn.c_attn": 128,
"transformer.h.2.attn.c_proj": 128,
"transformer.h.2.mlp.c_fc": 128,
"transformer.h.2.mlp.c_proj": 128,
"transformer.h.3.attn.c_attn": 128,
"transformer.h.3.attn.c_proj": 128,
"transformer.h.3.mlp.c_fc": 128,
"transformer.h.3.mlp.c_proj": 128,
"transformer.h.4.attn.c_attn": 128,
"transformer.h.4.attn.c_proj": 128,
"transformer.h.4.mlp.c_fc": 128,
"transformer.h.4.mlp.c_proj": 128,
"transformer.h.5.attn.c_attn": 128,
"transformer.h.5.attn.c_proj": 128,
"transformer.h.5.mlp.c_fc": 128,
"transformer.h.5.mlp.c_proj": 128,
"transformer.h.6.attn.c_attn": 128,
"transformer.h.6.attn.c_proj": 128,
"transformer.h.6.mlp.c_fc": 128,
"transformer.h.6.mlp.c_proj": 128,
"transformer.h.7.attn.c_attn": 128,
"transformer.h.7.attn.c_proj": 128,
"transformer.h.7.mlp.c_fc": 128,
"transformer.h.7.mlp.c_proj": 128,
"transformer.h.8.attn.c_attn": 128,
"transformer.h.8.attn.c_proj": 128,
"transformer.h.8.mlp.c_fc": 128,
"transformer.h.8.mlp.c_proj": 128,
"transformer.h.9.attn.c_attn": 128,
"transformer.h.9.attn.c_proj": 128,
"transformer.h.9.mlp.c_fc": 128,
"transformer.h.9.mlp.c_proj": 128,
"transformer.h.10.attn.c_attn": 128,
"transformer.h.10.attn.c_proj": 128,
"transformer.h.10.mlp.c_fc": 128,
"transformer.h.10.mlp.c_proj": 128,
"transformer.h.11.attn.c_attn": 128,
"transformer.h.11.attn.c_proj": 128,
"transformer.h.11.mlp.c_fc": 128,
"transformer.h.11.mlp.c_proj": 128
}
},
"hardware": {
"peak_vram_mb": 628.28,
"tokens_per_sec": 25776.19
},
"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": "2026-09-15T12:07:45.828512"
},
"results/unified_enlighten/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,78.56,0.8124,0.7811,0.8051,0.7832,0.4381,5.92,712.34,27357.52,124647170,124647170,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_roberta,2026-09-15T12:06:55.456606\nenlighten,PASS_LOW,51.03,0.6748,0.342,0.5882,0.5012,1.3388,10.78,520.68,26019.95,124647170,124647170,0,1.0,0.0,0.0,72,0,0.23880489796199525,0.2498,0,0,note: evaluated accuracy (51.03%) below chance threshold (56.00%),adapted_roberta,2026-09-15T12:07:06.492378\n",
"results/unified_enlighten/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,90.71,0.9099,0.907,0.9075,0.9068,0.3027,2.87,389.51,52725.06,66955010,66955010,0,1.0,0.0,0.0,0,0,,,0,0,,legacy_bert,2026-09-15T12:06:20.258208\nenlighten,PASS,89.22,0.8937,0.8922,0.8922,0.8922,0.3187,5.23,301.18,53181.36,66955010,66955010,0,1.0,0.0,0.0,36,0,0.24077659094076578,0.2556,0,0,,legacy_bert,2026-09-15T12:06:25.717119\n",
"results/unified_enlighten/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,79.36,0.7978,0.7935,0.7935,0.7935,0.4875,5.97,793.19,26923.15,124441344,124441344,0,1.0,0.0,0.0,0,0,,,0,0,,adapted_gpt,2026-09-15T12:07:36.913605\nenlighten,PASS,61.12,0.3936,0.5537,0.7577,0.618,0.6589,9.3,628.28,25776.19,124441344,124441344,0,1.0,0.0,0.0,48,0,0.2547934372152875,0.2629,0,0,,adapted_gpt,2026-09-15T12:07:46.459856\n"
},
"notes": []
}
}
{
"direct_svd": {
"module_name": "direct_svd",
"source_file": "Codes/framework/compression_engine/direct_svd.py",
"source_lines_count": 462,
"architecture": {
"class_name": "DirectSVDCompressor",
"base_class": "BaseCompressor",
"init_args": "self, model: nn.Module, rank: Optional[int] = 32, min_rank: int = 16, rank_ratio: Optional[float] = None, max_error: float = 0.35, fallback_to_original: bool = True, protect_critical: bool = True, lora_r: int = 0, lora_alpha: float = 1.0, skip_attention_qkv: bool = True, target_attention_only: bool = False, max_retries: int = 2, target_modules: Optional[List[str]] = None, cfg: Optional[Any] = None, **kwargs: Any, ) -> None: super().__init__(model, name=\"direct_svd\", **kwargs) cfg_obj = cfg if cfg is not None else kwargs.get(\"comp_config\", None) self.cfg = cfg_obj rank = _cfg_get(cfg_obj, \"direct_svd_rank\", _cfg_get(cfg_obj, \"svd_rank\", _cfg_get(cfg_obj, \"rank\", rank))) min_rank = _cfg_get(cfg_obj, \"min_rank\", min_rank) rank_ratio = _cfg_get(cfg_obj, \"direct_svd_rank_ratio\", _cfg_get(cfg_obj, \"rank_ratio\", rank_ratio)) max_error = _cfg_get(cfg_obj, \"direct_svd_max_error\", _cfg_get(cfg_obj, \"max_error\", max_error)) fallback_to_original = _cfg_get( cfg_obj, \"direct_svd_fallback_to_original\", _cfg_get(cfg_obj, \"fallback_to_original\", fallback_to_original) ) protect_critical = _cfg_get(cfg_obj, \"protect_critical\", protect_critical) lora_r = _cfg_get(cfg_obj, \"lora_r\", _cfg_get(cfg_obj, \"lora_rank\", lora_r)) lora_alpha = _cfg_get(cfg_obj, \"lora_alpha\", lora_alpha) skip_attention_qkv = _cfg_get(cfg_obj, \"skip_attention_qkv\", skip_attention_qkv) target_attention_only = _cfg_get( cfg_obj, \"direct_svd_target_attention_only\", _cfg_get(cfg_obj, \"target_attention_only\", target_attention_only) ) if \"rank\" in kwargs and kwargs[\"rank\"] is not None: rank = kwargs[\"rank\"] self.rank = rank self.min_rank = max(1, int(min_rank)) self.rank_ratio = float(rank_ratio) if rank_ratio is not None else None self.max_error = float(max_error) self.fallback_to_original = bool(fallback_to_original) self.protect_critical = bool(protect_critical) self.lora_r = int(lora_r) self.lora_alpha = float(lora_alpha) self.skip_attention_qkv = bool(skip_attention_qkv) self.target_attention_only = bool(target_attention_only) self.max_retries = int(max_retries) self.target_modules = target_modules self.compressed_layers: List[str] = [] self.fallback_layers: List[str] = [] self.skipped_layers: List[str] = [] self.layer_ranks: Dict[str, int] = {} self.layer_errors: Dict[str, float] = {} def _is_gqa_asymmetric(self, name: str, module: nn.Module) -> bool: if not self.skip_attention_qkv: return False cfg = getattr(self.model, \"config\", None) num_heads = getattr(cfg, \"num_attention_heads\", None) num_kv_heads = getattr(cfg, \"num_key_value_heads\", None) if num_heads is not None and num_kv_heads is not None and num_kv_heads < num_heads: name_lower = name.lower() if any(p in name_lower for p in [\"q_proj\", \"k_proj\", \"v_proj\", \"query\", \"key\", \"value\", \"c_attn\"]",
"registry_names": [
"direct_svd_compressor",
"direct_svd",
"direct_svd_compressor_factory",
"direct_svd_factory"
]
},
"params": [],
"metrics": {
"results/unified_direct_svd/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": 6.01
},
"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": 27189.71
},
"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": "2026-09-15T11:57:09.550344"
},
"results/unified_direct_svd/roberta/roberta-base_direct_svd_full_metrics.json": {
"model": "roberta-base_direct_svd",
"performance": {
"accuracy_pct": 71.1,
"f1_score": 0.7544,
"f1_macro": 0.7017,
"precision": 0.7342,
"recall": 0.708,
"val_loss": 0.6031,
"status": "PASS",
"elapsed_s": 18.12
},
"efficiency": {
"total_params_before": 124647170,
"total_params_after": 121255705,
"trainable_recovery_params": 17687,
"compression_rate": 1.028,
"compression_ratio_pct": 2.72,
"flops_reduction_pct": 25.0,
"num_compressed_layers": 23,
"num_fallback_layers": 49,
"num_injected_attention_layers": 0,
"lora_params": 0,
"avg_reconstruction_error": 0.333649210018602,
"weighted_reconstruction_error": 0.4827,
"per_layer_reconstruction_errors": {
"roberta.encoder.layer.0.attention.self.query": 0.3703107387562295,
"roberta.encoder.layer.0.attention.self.key": 0.3394817023607664,
"roberta.encoder.layer.0.attention.self.value": 0.3563514659359639,
"roberta.encoder.layer.0.attention.output.dense": 0.3384400364748302,
"roberta.encoder.layer.0.intermediate.dense": 0.5163187700478369,
"roberta.encoder.layer.0.output.dense": 0.5529538450447853,
"roberta.encoder.layer.1.attention.self.query": 0.36171755836811004,
"roberta.encoder.layer.1.attention.self.key": 0.35123381417844457,
"roberta.encoder.layer.1.attention.self.value": 0.33529264068267295,
"roberta.encoder.layer.1.attention.output.dense": 0.32870197078071106,
"roberta.encoder.layer.1.intermediate.dense": 0.5096122136799608,
Xet Storage Details
- Size:
- 133 kB
- Xet hash:
- f030aa2328457517496911e93636e1b01878038d4b1e5b1e28e8c77f91e7d772
·
Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.