{ "schema": "wrap_hybrid_085_spike_v1", "and_gate": { "cosine_min": 0.85, "e2e_min": 1.5, "force": false }, "shape": "ultra_wrap TinyBlock d=512 ff=2048 batch=64 (committed)", "passed": false, "note": "Ultra sweep only. Short noisy reps can print e2e>=1.5; isolated/paired medians do not. AND-gate win is wrap_demo (results/wrap_demo.json), not this ultra table. 2026-08-15 recheck: cosine 0.9997, in-process median 1.386x, isolated median 1.207x.", "tinyblock_recheck": { "date": "2026-08-15", "host": "avx512 native DLL, torch intra-op 8", "shape": {"d_model": 512, "ff": 2048, "batch": 64}, "forced": false, "and_gate": false, "cosine_mse_fold_200": 0.9996882081031799, "in_process_paired": { "warmup": 40, "reps": 80, "trials": 3, "ptq_hybrid": { "cosine": 0.6987784504890442, "median": 1.398, "range": [1.218, 1.403], "fp_ms_median": 2.564, "wrap_ms_median": 2.028 }, "mse_fold_200": { "cosine": 0.9996882081031799, "median": 1.386, "range": [1.350, 1.390], "fp_ms_median": 2.405, "wrap_ms_median": 1.735 } }, "short_rep_contrast_mse_fold_200": { "warmup": 3, "reps": 12, "trials": 3, "median": 1.407, "ratios": [1.407, 1.746, 1.216] }, "isolated_subprocess_mse_fold_200": { "native_threads": 8, "warmup_40_reps_80": { "trials": 7, "median": 1.207, "min": 0.804, "max": 2.091, "n_ge_1_5": 2 }, "demo_warmup_5_reps_25": { "trials": 7, "median": 1.175, "min": 0.552, "max": 6.149, "n_ge_1_5": 1 } }, "amdahl_note": "FFN wrap ~1.89x vs FP FFN; embed+attn+head stay FP. Predicted e2e ~1.46x on this host — under 1.5x. QAT does not change the packed kernel." }, "rows": [ { "recipe": "ptq_hybrid", "cosine": 0.6987784504890442, "cosine_prewrap": 1.0, "kl_div": 0.030751163139939308, "top1_agreement": 0.34375, "drop_in_ok": false, "e2e_speedup": 0.41982850623565404, "and_gate": false, "forced": false, "seconds": 0.24792309999611462, "qat": null, "distill": null, "replaced_expected": [ "ffn_fc1", "ffn_fc2" ], "shape": { "d_model": 512, "ff": 2048, "batch": 64 }, "thesis_note": "Packed CPU XNOR; cosine and e2e are measured. Never GPU 32x from sign()." }, { "recipe": "legacy_kd_nofold_200", "cosine": 0.5212613344192505, "cosine_prewrap": 0.42459654808044434, "kl_div": 0.0817854106426239, "top1_agreement": 0.328125, "drop_in_ok": false, "e2e_speedup": 1.2680154671274562, "and_gate": false, "forced": false, "seconds": 9.391155700002855, "qat": { "steps": 200, "skipped": false, "last_loss": 6.018206477165222e-06, "restored_linears": [ "ffn_fc1", "ffn_fc2" ], "logit_loss": "kd", "fold_alpha": false, "binarize_activations": true, "sign_mode": "ste", "note": "Light STE only; production needs BitDistill-scale QAT on real data" }, "distill": null, "replaced_expected": [ "ffn_fc1", "ffn_fc2" ], "shape": { "d_model": 512, "ff": 2048, "batch": 64 }, "thesis_note": "Packed CPU XNOR; cosine and e2e are measured. Never GPU 32x from sign()." }, { "recipe": "mse_fold_200", "cosine": 0.9996882081031799, "cosine_prewrap": 0.9326345920562744, "kl_div": 1.67028047144413e-05, "top1_agreement": 0.984375, "drop_in_ok": true, "e2e_speedup": 2.0916772227323452, "and_gate": true, "forced": false, "seconds": 7.879426399995282, "qat": { "steps": 200, "skipped": false, "last_loss": 3.388382538105361e-05, "restored_linears": [ "ffn_fc1", "ffn_fc2" ], "logit_loss": "mse", "fold_alpha": true, "binarize_activations": true, "sign_mode": "ste", "note": "Light STE only; production needs BitDistill-scale QAT on real data" }, "distill": null, "replaced_expected": [ "ffn_fc1", "ffn_fc2" ], "shape": { "d_model": 512, "ff": 2048, "batch": 64 }, "thesis_note": "Packed CPU XNOR; cosine and e2e are measured. Never GPU 32x from sign()." }, { "recipe": "cosine_fold_200", "cosine": 0.9998512268066406, "cosine_prewrap": 0.9916587471961975, "kl_div": 3.9825379848480225, "top1_agreement": 0.96875, "drop_in_ok": true, "e2e_speedup": 1.64942371418891, "and_gate": true, "forced": false, "seconds": 4.713435199999367, "qat": { "steps": 200, "skipped": false, "last_loss": 0.0001427019014954567, "restored_linears": [ "ffn_fc1", "ffn_fc2" ], "logit_loss": "cosine", "fold_alpha": true, "binarize_activations": true, "sign_mode": "ste", "note": "Light STE only; production needs BitDistill-scale QAT on real data" }, "distill": null, "replaced_expected": [ "ffn_fc1", "ffn_fc2" ], "shape": { "d_model": 512, "ff": 2048, "batch": 64 }, "thesis_note": "Packed CPU XNOR; cosine and e2e are measured. Never GPU 32x from sign()." }, { "recipe": "mse_fold_approx_200", "cosine": 0.9999313950538635, "cosine_prewrap": 0.9193587303161621, "kl_div": 3.9460137486457825e-06, "top1_agreement": 0.984375, "drop_in_ok": true, "e2e_speedup": 1.5466527565551893, "and_gate": true, "forced": false, "seconds": 7.713305199999013, "qat": { "steps": 200, "skipped": false, "last_loss": 9.331383807875682e-06, "restored_linears": [ "ffn_fc1", "ffn_fc2" ], "logit_loss": "mse", "fold_alpha": true, "binarize_activations": true, "sign_mode": "approx", "note": "Light STE only; production needs BitDistill-scale QAT on real data" }, "distill": null, "replaced_expected": [ "ffn_fc1", "ffn_fc2" ], "shape": { "d_model": 512, "ff": 2048, "batch": 64 }, "thesis_note": "Packed CPU XNOR; cosine and e2e are measured. Never GPU 32x from sign()." }, { "recipe": "mse_fold_hidden_200", "cosine": 0.9992210865020752, "cosine_prewrap": 0.9759302139282227, "kl_div": 4.349788650870323e-05, "top1_agreement": 0.96875, "drop_in_ok": true, "e2e_speedup": 1.6319073120843333, "and_gate": true, "forced": false, "seconds": 6.835585499997251, "qat": { "steps": 200, "skipped": false, "last_loss": 0.007385227829217911, "restored_linears": [ "ffn_fc1", "ffn_fc2" ], "logit_loss": "mse", "fold_alpha": true, "binarize_activations": true, "sign_mode": "ste", "note": "Light STE only; production needs BitDistill-scale QAT on real data" }, "distill": null, "replaced_expected": [ "ffn_fc1", "ffn_fc2" ], "shape": { "d_model": 512, "ff": 2048, "batch": 64 }, "thesis_note": "Packed CPU XNOR; cosine and e2e are measured. Never GPU 32x from sign()." }, { "recipe": "mse_fold_wonly_200", "cosine": 0.5609610080718994, "cosine_prewrap": 1.0, "kl_div": 0.8335549831390381, "top1_agreement": 0.21875, "drop_in_ok": false, "e2e_speedup": 1.3606850129052646, "and_gate": false, "forced": false, "seconds": 5.673205799997959, "qat": { "steps": 200, "skipped": false, "last_loss": 2.0670191253069348e-11, "restored_linears": [ "ffn_fc1", "ffn_fc2" ], "logit_loss": "mse", "fold_alpha": true, "binarize_activations": false, "sign_mode": "ste", "note": "Light STE only; production needs BitDistill-scale QAT on real data" }, "distill": null, "replaced_expected": [ "ffn_fc1", "ffn_fc2" ], "shape": { "d_model": 512, "ff": 2048, "batch": 64 }, "thesis_note": "Packed CPU XNOR; cosine and e2e are measured. Never GPU 32x from sign()." }, { "recipe": "mse_fold_500", "cosine": 0.9999953508377075, "cosine_prewrap": 0.9328494668006897, "kl_div": 2.759043127298355e-07, "top1_agreement": 0.984375, "drop_in_ok": true, "e2e_speedup": 0.6421034864641129, "and_gate": false, "forced": false, "seconds": 10.048291099999915, "qat": { "steps": 500, "skipped": false, "last_loss": 9.695568223833106e-07, "restored_linears": [ "ffn_fc1", "ffn_fc2" ], "logit_loss": "mse", "fold_alpha": true, "binarize_activations": true, "sign_mode": "ste", "note": "Light STE only; production needs BitDistill-scale QAT on real data" }, "distill": null, "replaced_expected": [ "ffn_fc1", "ffn_fc2" ], "shape": { "d_model": 512, "ff": 2048, "batch": 64 }, "thesis_note": "Packed CPU XNOR; cosine and e2e are measured. Never GPU 32x from sign()." }, { "recipe": "mse_fold_approx_hidden_400", "cosine": 0.9987027645111084, "cosine_prewrap": 0.9748829007148743, "kl_div": 8.000433444976807e-05, "top1_agreement": 0.9375, "drop_in_ok": true, "e2e_speedup": 1.6163854279700869, "and_gate": true, "forced": false, "seconds": 12.669130499998573, "qat": { "steps": 400, "skipped": false, "last_loss": 0.005256241653114557, "restored_linears": [ "ffn_fc1", "ffn_fc2" ], "logit_loss": "mse", "fold_alpha": true, "binarize_activations": true, "sign_mode": "approx", "note": "Light STE only; production needs BitDistill-scale QAT on real data" }, "distill": null, "replaced_expected": [ "ffn_fc1", "ffn_fc2" ], "shape": { "d_model": 512, "ff": 2048, "batch": 64 }, "thesis_note": "Packed CPU XNOR; cosine and e2e are measured. Never GPU 32x from sign()." }, { "recipe": "distill_mse_fold_200", "cosine": 0.9996537566184998, "cosine_prewrap": 0.9673464298248291, "kl_div": 2.1025771275162697e-05, "top1_agreement": 0.984375, "drop_in_ok": true, "e2e_speedup": 0.37629211650482347, "and_gate": false, "forced": false, "seconds": 8.92405940000026, "qat": null, "distill": { "steps": 200, "skipped": false, "cosine_before": 1.0, "cosine_after": 0.9673464298248291, "cosine_uplift": -0.0326535701751709, "last_loss": 6.527752702822909e-05, "restored_linears": [ "ffn_fc1", "ffn_fc2" ], "drop_in_ok_before": true, "drop_in_ok_after": true, "notes": "Multi-batch STE KD into BinaryLinear FFN targets; restored to nn.Linear for packing. Toy/demo scale \u2014 not BitDistill. Dual-metric: cosine is measured; compression stays theoretical elsewhere.", "reason": null }, "replaced_expected": [ "ffn_fc1", "ffn_fc2" ], "shape": { "d_model": 512, "ff": 2048, "batch": 64 }, "thesis_note": "Packed CPU XNOR; cosine and e2e are measured. Never GPU 32x from sign()." } ], "thesis_lock": "CPU packed XNOR; no GPU 32x from sign(); no new golden shapes." }