Portuguese
mamba
mamba-3
ssm
muon
muonclip
qk-clip
adamw
optimizer
nope
hybrid-architecture
activation-checkpointing
3d-vision-language
depth-perception
point-cloud
diffusion
stable-diffusion
mixture-of-experts
kohonen
som-attention
random-forest
deep-neural-network
knowledge-distillation
bidirectional-distillation
engram-memory
multimodal
multi-token-prediction
dpo
rlhf
hh-rlhf
pcgrad
long-context
lra-qvit
engineer-agent
integral-state
cython
jev-classifiers
gumbel-softmax
quantization
riemannian-optimization
stiefel-manifold
cython-kernels
gradient-accumulation
vector-quantization
grayscale-detection
joint-perception
self-regulation
signal-table
abmo
ciar
growing-som
pt-br
KHTST v17 KHMAMBAJEPA — Problema 1: MoE EMPACOTADA na quantização (bijeção φ + indexador narrow + floresta B blocos + rollback sem copy_; 3.79–3.91×, Δy ≤ cota — Teos 29.1–29.4); Problema 2: TCE curada (L/d reparametrização pura + limiar invariante ‖e‖²≤θ‖alvo‖² + λ_jepa por razão ponderada — Teos 29.5–29.8); jepa_validator.pyx (Token Rounding exato + AuditoriaMemoriaSonic I1–I4 — Teos 29.10–29.12); ReplaySSM (fronteiras bit-exatas, gradiente exato 0.0 — Teos 29.13–29.14); autoajuste MULTI-MÉTODO RM+PID+bisseção p/ TODOS os parâmetros fixos (15 controladores — Teos 30.1–30.15); vocab 16384 mantido; suíte v17 59✓ + regressão v1–v16 + verificar_tudo 49✓ (part 2)
Browse files
telemetria_out/treino_v17_baseline_v16.jsonl
ADDED
|
@@ -0,0 +1,408 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{"t": 0.27, "passo": null, "atributo": "jepa/ativo", "valor": 1.0}
|
| 2 |
+
{"t": 5.33, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.022999}
|
| 3 |
+
{"t": 5.33, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.019829}
|
| 4 |
+
{"t": 5.33, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.026767}
|
| 5 |
+
{"t": 5.33, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.019829}
|
| 6 |
+
{"t": 5.33, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.109001}
|
| 7 |
+
{"t": 5.33, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.027043}
|
| 8 |
+
{"t": 5.33, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.027043}
|
| 9 |
+
{"t": 5.33, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.724489}
|
| 10 |
+
{"t": 5.33, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.022999}
|
| 11 |
+
{"t": 5.33, "passo": null, "atributo": "som_atn/entropia", "valor": 1.095767}
|
| 12 |
+
{"t": 5.35, "passo": 0, "atributo": "jepa/perda", "valor": 34.716282}
|
| 13 |
+
{"t": 5.35, "passo": 0, "atributo": "jepa/var_latente", "valor": 0.20879}
|
| 14 |
+
{"t": 10.45, "passo": 0, "atributo": "autorreg/lr_escala", "valor": 1.0}
|
| 15 |
+
{"t": 10.45, "passo": 0, "atributo": "autorreg/moe_escala", "valor": 1.0}
|
| 16 |
+
{"t": 10.45, "passo": 0, "atributo": "autorreg/lambda_mtp", "valor": 1.0}
|
| 17 |
+
{"t": 10.45, "passo": 0, "atributo": "autorreg/lambda_moe_lb", "valor": 1.0}
|
| 18 |
+
{"t": 10.45, "passo": 0, "atributo": "autorreg/lambda_som", "valor": 1.0}
|
| 19 |
+
{"t": 10.45, "passo": 0, "atributo": "autorreg/lambda_jepa", "valor": 1.0}
|
| 20 |
+
{"t": 10.45, "passo": 0, "atributo": "muonclip/logit_max", "valor": 0.831882}
|
| 21 |
+
{"t": 10.45, "passo": 1, "perda/treino": 9.669491, "perda/ce": 9.669491, "grad_norm": 234.898438, "lr": 0.000375, "abmo/L_hat": 1.0, "pcgrad/dot": 1.588219, "pcgrad/projetado": 0.0, "conf/h_t": 0.564347, "prs/trust": 0.49, "prs/clip": 5.0, "prs/boost": 1.0, "prs/streak": 0.0, "rpp/rho": 1.0, "rpp/streak": 0.0}
|
| 22 |
+
{"t": 10.45, "passo": 1, "atributo": "mtp/entropia", "valor": 0.693147}
|
| 23 |
+
{"t": 10.45, "passo": 1, "atributo": "mtp/termos_ce", "valor": 106.0}
|
| 24 |
+
{"t": 11.33, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.023324}
|
| 25 |
+
{"t": 11.33, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.01867}
|
| 26 |
+
{"t": 11.33, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.023798}
|
| 27 |
+
{"t": 11.33, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.01867}
|
| 28 |
+
{"t": 11.33, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.110835}
|
| 29 |
+
{"t": 11.33, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.023324}
|
| 30 |
+
{"t": 11.33, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.023324}
|
| 31 |
+
{"t": 11.33, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.734728}
|
| 32 |
+
{"t": 11.33, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.023324}
|
| 33 |
+
{"t": 11.33, "passo": null, "atributo": "som_atn/entropia", "valor": 1.057899}
|
| 34 |
+
{"t": 20.42, "passo": 2, "perda/treino": 9.731093, "perda/ce": 9.731093, "grad_norm": 46.702255, "lr": 0.00075, "abmo/L_hat": 1.000637, "pcgrad/dot": 0.354797, "pcgrad/projetado": 0.0, "conf/h_t": 0.472697, "prs/trust": 0.4802, "prs/clip": 5.0, "prs/boost": 1.0, "prs/streak": 0.0, "rpp/rho": 1.0, "rpp/streak": 0.0}
|
| 35 |
+
{"t": 20.42, "passo": 2, "atributo": "mtp/entropia", "valor": 0.693147}
|
| 36 |
+
{"t": 20.42, "passo": 2, "atributo": "mtp/termos_ce", "valor": 1020.0}
|
| 37 |
+
{"t": 21.94, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.023178}
|
| 38 |
+
{"t": 21.94, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.018226}
|
| 39 |
+
{"t": 21.94, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.024603}
|
| 40 |
+
{"t": 21.94, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.018226}
|
| 41 |
+
{"t": 21.94, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.112861}
|
| 42 |
+
{"t": 21.94, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.024807}
|
| 43 |
+
{"t": 21.94, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.024807}
|
| 44 |
+
{"t": 21.94, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.730113}
|
| 45 |
+
{"t": 21.94, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.023178}
|
| 46 |
+
{"t": 21.94, "passo": null, "atributo": "som_atn/entropia", "valor": 1.070351}
|
| 47 |
+
{"t": 64.11, "passo": 3, "perda/treino": 9.742939, "perda/ce": 9.742939, "grad_norm": 4.748356, "lr": 0.001125, "abmo/L_hat": 1.001333, "pcgrad/dot": -0.158332, "pcgrad/projetado": 1.0, "conf/h_t": 0.437847, "prs/trust": 0.470596, "prs/clip": 4.748356, "prs/boost": 1.073541, "prs/streak": 0.0, "rpp/rho": 1.0, "rpp/streak": 0.0}
|
| 48 |
+
{"t": 64.11, "passo": 3, "atributo": "mtp/entropia", "valor": 0.693147}
|
| 49 |
+
{"t": 64.11, "passo": 3, "atributo": "mtp/termos_ce", "valor": 488.0}
|
| 50 |
+
{"t": 64.82, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.023778}
|
| 51 |
+
{"t": 64.82, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.016391}
|
| 52 |
+
{"t": 64.82, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.022126}
|
| 53 |
+
{"t": 64.82, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.016391}
|
| 54 |
+
{"t": 64.83, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.098017}
|
| 55 |
+
{"t": 64.83, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.025247}
|
| 56 |
+
{"t": 64.83, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.025247}
|
| 57 |
+
{"t": 64.83, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.749023}
|
| 58 |
+
{"t": 64.83, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.023778}
|
| 59 |
+
{"t": 64.83, "passo": null, "atributo": "som_atn/entropia", "valor": 1.026148}
|
| 60 |
+
{"t": 73.86, "passo": 4, "perda/treino": 9.738967, "perda/ce": 9.738967, "grad_norm": 114.646484, "lr": 0.0015, "abmo/L_hat": 1.001918, "pcgrad/dot": -1.079053, "pcgrad/projetado": 1.0, "conf/h_t": 0.414474, "prs/trust": 0.461184, "prs/clip": 5.0, "prs/boost": 1.086236, "prs/streak": 0.0, "rpp/rho": 1.0, "rpp/streak": 0.0}
|
| 61 |
+
{"t": 73.86, "passo": 4, "atributo": "mtp/entropia", "valor": 0.693146}
|
| 62 |
+
{"t": 73.86, "passo": 4, "atributo": "mtp/termos_ce", "valor": 428.0}
|
| 63 |
+
{"t": 74.25, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.022947}
|
| 64 |
+
{"t": 74.25, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.019082}
|
| 65 |
+
{"t": 74.25, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.025758}
|
| 66 |
+
{"t": 74.25, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.019082}
|
| 67 |
+
{"t": 74.25, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.11215}
|
| 68 |
+
{"t": 74.25, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.027592}
|
| 69 |
+
{"t": 74.25, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.027592}
|
| 70 |
+
{"t": 74.25, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.722848}
|
| 71 |
+
{"t": 74.25, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.022947}
|
| 72 |
+
{"t": 74.25, "passo": null, "atributo": "som_atn/entropia", "valor": 1.095989}
|
| 73 |
+
{"t": 79.24, "passo": 5, "perda/treino": 9.703294, "perda/ce": 9.703294, "grad_norm": 197.162613, "lr": 0.001875, "abmo/L_hat": 1.002076, "pcgrad/dot": 1.245448, "pcgrad/projetado": 0.0, "conf/h_t": 0.39863, "prs/trust": 0.45196, "prs/clip": 5.0, "prs/boost": 1.095185, "prs/streak": 0.0, "rpp/rho": 1.0, "rpp/streak": 0.0}
|
| 74 |
+
{"t": 79.24, "passo": 5, "atributo": "mtp/entropia", "valor": 0.693144}
|
| 75 |
+
{"t": 79.24, "passo": 5, "atributo": "mtp/termos_ce", "valor": 112.0}
|
| 76 |
+
{"t": 79.93, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.023219}
|
| 77 |
+
{"t": 79.93, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.018107}
|
| 78 |
+
{"t": 79.93, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.024442}
|
| 79 |
+
{"t": 79.93, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.018107}
|
| 80 |
+
{"t": 79.93, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.111092}
|
| 81 |
+
{"t": 79.93, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.025198}
|
| 82 |
+
{"t": 79.93, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.025198}
|
| 83 |
+
{"t": 79.93, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.731417}
|
| 84 |
+
{"t": 79.93, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.023219}
|
| 85 |
+
{"t": 79.93, "passo": null, "atributo": "som_atn/entropia", "valor": 1.068428}
|
| 86 |
+
{"t": 89.51, "passo": 6, "perda/treino": 9.764006, "perda/ce": 9.764006, "grad_norm": 9.547014, "lr": 0.00225, "abmo/L_hat": 1.002846, "pcgrad/dot": -0.153378, "pcgrad/projetado": 1.0, "conf/h_t": 0.377606, "prs/trust": 0.442921, "prs/clip": 5.0, "prs/boost": 1.095699, "prs/streak": 0.0, "rpp/rho": 1.0, "rpp/streak": 0.0}
|
| 87 |
+
{"t": 89.51, "passo": 6, "atributo": "mtp/entropia", "valor": 0.693141}
|
| 88 |
+
{"t": 89.51, "passo": 6, "atributo": "mtp/termos_ce", "valor": 565.0}
|
| 89 |
+
{"t": 89.84, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.022907}
|
| 90 |
+
{"t": 89.84, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.021036}
|
| 91 |
+
{"t": 89.84, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.028395}
|
| 92 |
+
{"t": 89.84, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.021036}
|
| 93 |
+
{"t": 89.84, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.113728}
|
| 94 |
+
{"t": 89.84, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.024214}
|
| 95 |
+
{"t": 89.84, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.024214}
|
| 96 |
+
{"t": 89.84, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.721565}
|
| 97 |
+
{"t": 89.84, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.022907}
|
| 98 |
+
{"t": 89.84, "passo": null, "atributo": "som_atn/entropia", "valor": 1.099078}
|
| 99 |
+
{"t": 94.34, "passo": 7, "perda/treino": 9.792176, "perda/ce": 9.792176, "grad_norm": 64.380783, "lr": 0.002625, "abmo/L_hat": 1.00383, "pcgrad/dot": 2.924468, "pcgrad/projetado": 0.0, "conf/h_t": 0.3732, "prs/trust": 0.434063, "prs/clip": 5.0, "prs/boost": 1.097514, "prs/streak": 0.0, "rpp/rho": 1.0, "rpp/streak": 0.0}
|
| 100 |
+
{"t": 94.34, "passo": 7, "atributo": "mtp/entropia", "valor": 0.693135}
|
| 101 |
+
{"t": 94.34, "passo": 7, "atributo": "mtp/termos_ce", "valor": 107.0}
|
| 102 |
+
{"t": 94.65, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.022946}
|
| 103 |
+
{"t": 94.65, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.018093}
|
| 104 |
+
{"t": 94.65, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.024423}
|
| 105 |
+
{"t": 94.65, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.018093}
|
| 106 |
+
{"t": 94.65, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.118889}
|
| 107 |
+
{"t": 94.65, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.025901}
|
| 108 |
+
{"t": 94.65, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.025901}
|
| 109 |
+
{"t": 94.65, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.722809}
|
| 110 |
+
{"t": 94.65, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.022946}
|
| 111 |
+
{"t": 94.65, "passo": null, "atributo": "som_atn/entropia", "valor": 1.085602}
|
| 112 |
+
{"t": 98.35, "passo": 8, "perda/treino": 9.67628, "perda/ce": 9.67628, "grad_norm": 368.498993, "lr": 0.003, "abmo/L_hat": 1.003517, "pcgrad/dot": 3.205432, "pcgrad/projetado": 0.0, "conf/h_t": 0.372923, "prs/trust": 0.46236, "prs/clip": 5.0, "prs/boost": 1.10247, "prs/streak": 1.0, "rpp/rho": 1.0, "rpp/streak": 0.0}
|
| 113 |
+
{"t": 98.35, "passo": 8, "atributo": "mtp/entropia", "valor": 0.693126}
|
| 114 |
+
{"t": 98.35, "passo": 8, "atributo": "mtp/termos_ce", "valor": 44.0}
|
| 115 |
+
{"t": 98.7, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.022665}
|
| 116 |
+
{"t": 98.7, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.021197}
|
| 117 |
+
{"t": 98.7, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.028612}
|
| 118 |
+
{"t": 98.7, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.021197}
|
| 119 |
+
{"t": 98.7, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.116839}
|
| 120 |
+
{"t": 98.7, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.026436}
|
| 121 |
+
{"t": 98.7, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.026436}
|
| 122 |
+
{"t": 98.7, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.713954}
|
| 123 |
+
{"t": 98.7, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.022665}
|
| 124 |
+
{"t": 98.7, "passo": null, "atributo": "som_atn/entropia", "valor": 1.120218}
|
| 125 |
+
{"t": 106.36, "passo": 9, "perda/treino": 9.363508, "perda/ce": 9.363508, "grad_norm": 1038.523804, "lr": 0.003, "abmo/L_hat": 1.000001, "pcgrad/dot": 181.47251, "pcgrad/projetado": 0.0, "conf/h_t": 0.421931, "prs/trust": 0.489242, "prs/clip": 5.0, "prs/boost": 1.109692, "prs/streak": 2.0, "rpp/rho": 1.0, "rpp/streak": 1.0}
|
| 126 |
+
{"t": 106.36, "passo": 9, "atributo": "mtp/entropia", "valor": 0.693114}
|
| 127 |
+
{"t": 106.36, "passo": 9, "atributo": "mtp/termos_ce", "valor": 4.0}
|
| 128 |
+
{"t": 106.82, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.023184}
|
| 129 |
+
{"t": 106.82, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.015671}
|
| 130 |
+
{"t": 106.82, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.021153}
|
| 131 |
+
{"t": 106.82, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.015671}
|
| 132 |
+
{"t": 106.82, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.120719}
|
| 133 |
+
{"t": 106.82, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.025054}
|
| 134 |
+
{"t": 106.82, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.025054}
|
| 135 |
+
{"t": 106.82, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.730309}
|
| 136 |
+
{"t": 106.82, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.023184}
|
| 137 |
+
{"t": 106.82, "passo": null, "atributo": "som_atn/entropia", "valor": 1.055598}
|
| 138 |
+
{"t": 113.53, "passo": 10, "perda/treino": 9.813563, "perda/ce": 9.813563, "grad_norm": 41.345501, "lr": 0.003, "abmo/L_hat": 1.001491, "pcgrad/dot": 0.396943, "pcgrad/projetado": 0.0, "conf/h_t": 0.363249, "prs/trust": 0.479457, "prs/clip": 5.0, "prs/boost": 1.102374, "prs/streak": 0.0, "rpp/rho": 1.0, "rpp/streak": 2.0}
|
| 139 |
+
{"t": 113.53, "passo": 10, "atributo": "mtp/entropia", "valor": 0.693101}
|
| 140 |
+
{"t": 113.53, "passo": 10, "atributo": "mtp/termos_ce", "valor": 497.0}
|
| 141 |
+
{"t": 114.1, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.023226}
|
| 142 |
+
{"t": 114.11, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.016113}
|
| 143 |
+
{"t": 114.11, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.021751}
|
| 144 |
+
{"t": 114.11, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.016113}
|
| 145 |
+
{"t": 114.11, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.119971}
|
| 146 |
+
{"t": 114.11, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.023991}
|
| 147 |
+
{"t": 114.11, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.023991}
|
| 148 |
+
{"t": 114.11, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.731619}
|
| 149 |
+
{"t": 114.11, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.023226}
|
| 150 |
+
{"t": 114.11, "passo": null, "atributo": "som_atn/entropia", "valor": 1.052867}
|
| 151 |
+
{"t": 123.12, "passo": 11, "perda/treino": 9.820722, "perda/ce": 9.820722, "grad_norm": 22.319494, "lr": 0.003, "abmo/L_hat": 1.002906, "pcgrad/dot": 0.272932, "pcgrad/projetado": 0.0, "conf/h_t": 0.438027, "prs/trust": 0.469868, "prs/clip": 5.0, "prs/boost": 1.103271, "prs/streak": 0.0, "rpp/rho": 1.0, "rpp/streak": 3.0}
|
| 152 |
+
{"t": 123.12, "passo": 11, "atributo": "mtp/entropia", "valor": 0.693086}
|
| 153 |
+
{"t": 123.12, "passo": 11, "atributo": "mtp/termos_ce", "valor": 616.0}
|
| 154 |
+
{"t": 123.43, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.022894}
|
| 155 |
+
{"t": 123.43, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.01777}
|
| 156 |
+
{"t": 123.43, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.023987}
|
| 157 |
+
{"t": 123.43, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.01777}
|
| 158 |
+
{"t": 123.43, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.119207}
|
| 159 |
+
{"t": 123.43, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.027158}
|
| 160 |
+
{"t": 123.43, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.027158}
|
| 161 |
+
{"t": 123.43, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.721162}
|
| 162 |
+
{"t": 123.43, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.022894}
|
| 163 |
+
{"t": 123.43, "passo": null, "atributo": "som_atn/entropia", "valor": 1.09045}
|
| 164 |
+
{"t": 127.92, "passo": 12, "perda/treino": 9.815102, "perda/ce": 9.815102, "grad_norm": 594.890564, "lr": 0.003, "abmo/L_hat": 1.004121, "pcgrad/dot": 27.486783, "pcgrad/projetado": 0.0, "conf/h_t": 0.429795, "prs/trust": 0.46047, "prs/clip": 5.0, "prs/boost": 1.104182, "prs/streak": 0.0, "rpp/rho": 1.0, "rpp/streak": 4.0}
|
| 165 |
+
{"t": 127.92, "passo": 12, "atributo": "mtp/entropia", "valor": 0.693067}
|
| 166 |
+
{"t": 127.92, "passo": 12, "atributo": "mtp/termos_ce", "valor": 53.0}
|
| 167 |
+
{"t": 128.25, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.022894}
|
| 168 |
+
{"t": 128.25, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.019906}
|
| 169 |
+
{"t": 128.25, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.02687}
|
| 170 |
+
{"t": 128.25, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.019906}
|
| 171 |
+
{"t": 128.25, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.117784}
|
| 172 |
+
{"t": 128.25, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.024283}
|
| 173 |
+
{"t": 128.25, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.024283}
|
| 174 |
+
{"t": 128.25, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.721181}
|
| 175 |
+
{"t": 128.25, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.022894}
|
| 176 |
+
{"t": 128.25, "passo": null, "atributo": "som_atn/entropia", "valor": 1.093193}
|
| 177 |
+
{"t": 132.73, "passo": 13, "perda/treino": 9.607736, "perda/ce": 9.607736, "grad_norm": 63.54044, "lr": 0.003, "abmo/L_hat": 1.00307, "pcgrad/dot": 0.363346, "pcgrad/projetado": 0.0, "conf/h_t": 0.426522, "prs/trust": 0.487447, "prs/clip": 5.0, "prs/boost": 1.107512, "prs/streak": 1.0, "rpp/rho": 1.0, "rpp/streak": 5.0}
|
| 178 |
+
{"t": 132.73, "passo": 13, "atributo": "mtp/entropia", "valor": 0.693047}
|
| 179 |
+
{"t": 132.73, "passo": 13, "atributo": "mtp/termos_ce", "valor": 97.0}
|
| 180 |
+
{"t": 133.29, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.023105}
|
| 181 |
+
{"t": 133.29, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.017786}
|
| 182 |
+
{"t": 133.29, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.024008}
|
| 183 |
+
{"t": 133.29, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.017786}
|
| 184 |
+
{"t": 133.29, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.114572}
|
| 185 |
+
{"t": 133.29, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.025907}
|
| 186 |
+
{"t": 133.29, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.025907}
|
| 187 |
+
{"t": 133.29, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.727824}
|
| 188 |
+
{"t": 133.29, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.023105}
|
| 189 |
+
{"t": 133.29, "passo": null, "atributo": "som_atn/entropia", "valor": 1.075144}
|
| 190 |
+
{"t": 142.71, "passo": 14, "perda/treino": 9.755126, "perda/ce": 9.755126, "grad_norm": 194.987823, "lr": 0.002999, "abmo/L_hat": 1.003649, "pcgrad/dot": 10.524685, "pcgrad/projetado": 0.0, "conf/h_t": 0.425206, "prs/trust": 0.477698, "prs/clip": 5.0, "prs/boost": 1.106162, "prs/streak": 0.0, "rpp/rho": 1.0, "rpp/streak": 6.0}
|
| 191 |
+
{"t": 142.71, "passo": 14, "atributo": "mtp/entropia", "valor": 0.693025}
|
| 192 |
+
{"t": 142.71, "passo": 14, "atributo": "mtp/termos_ce", "valor": 232.0}
|
| 193 |
+
{"t": 143.04, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.022792}
|
| 194 |
+
{"t": 143.04, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.019205}
|
| 195 |
+
{"t": 143.04, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.025923}
|
| 196 |
+
{"t": 143.04, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.019205}
|
| 197 |
+
{"t": 143.04, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.11802}
|
| 198 |
+
{"t": 143.04, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.027054}
|
| 199 |
+
{"t": 143.04, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.027054}
|
| 200 |
+
{"t": 143.04, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.717956}
|
| 201 |
+
{"t": 143.04, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.022792}
|
| 202 |
+
{"t": 143.04, "passo": null, "atributo": "som_atn/entropia", "valor": 1.104014}
|
| 203 |
+
{"t": 147.6, "passo": 15, "perda/treino": 9.596692, "perda/ce": 9.596692, "grad_norm": 13.990447, "lr": 0.002999, "abmo/L_hat": 1.002531, "pcgrad/dot": -0.953619, "pcgrad/projetado": 1.0, "conf/h_t": 0.44325, "prs/trust": 0.503813, "prs/clip": 5.0, "prs/boost": 1.108348, "prs/streak": 1.0, "rpp/rho": 1.0, "rpp/streak": 7.0}
|
| 204 |
+
{"t": 147.6, "passo": 15, "atributo": "mtp/entropia", "valor": 0.693004}
|
| 205 |
+
{"t": 147.6, "passo": 15, "atributo": "mtp/termos_ce", "valor": 96.0}
|
| 206 |
+
{"t": 148.14, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.022912}
|
| 207 |
+
{"t": 148.14, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.018372}
|
| 208 |
+
{"t": 148.14, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.0248}
|
| 209 |
+
{"t": 148.14, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.018372}
|
| 210 |
+
{"t": 148.14, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.118197}
|
| 211 |
+
{"t": 148.14, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.026351}
|
| 212 |
+
{"t": 148.14, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.026351}
|
| 213 |
+
{"t": 148.14, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.721734}
|
| 214 |
+
{"t": 148.14, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.022912}
|
| 215 |
+
{"t": 148.14, "passo": null, "atributo": "som_atn/entropia", "valor": 1.0905}
|
| 216 |
+
{"t": 157.25, "passo": 16, "perda/treino": 9.846667, "perda/ce": 9.846667, "grad_norm": 4.443491, "lr": 0.002999, "abmo/L_hat": 1.00411, "pcgrad/dot": -0.072342, "pcgrad/projetado": 1.0, "conf/h_t": 0.419451, "prs/trust": 0.493737, "prs/clip": 5.0, "prs/boost": 1.106075, "prs/streak": 0.0, "rpp/rho": 1.0, "rpp/streak": 8.0}
|
| 217 |
+
{"t": 157.25, "passo": 16, "atributo": "mtp/entropia", "valor": 0.692975}
|
| 218 |
+
{"t": 157.25, "passo": 16, "atributo": "mtp/termos_ce", "valor": 304.0}
|
| 219 |
+
{"t": 157.87, "passo": null, "atributo": "ensemble/peso_rf", "valor": 0.339049}
|
| 220 |
+
{"t": 157.87, "passo": null, "atributo": "ensemble/peso_dnn", "valor": 0.329044}
|
| 221 |
+
{"t": 157.87, "passo": null, "atributo": "ensemble/peso_engram", "valor": 0.331907}
|
| 222 |
+
{"t": 157.87, "passo": null, "atributo": "ensemble/acc_ema_kd", "valor": 0.054412}
|
| 223 |
+
{"t": 157.87, "passo": null, "atributo": "ensemble/kd/total", "valor": 3.8074}
|
| 224 |
+
{"t": 157.87, "passo": null, "atributo": "ensemble/kd/acc_ema", "valor": 0.0544}
|
| 225 |
+
{"t": 157.87, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.023024}
|
| 226 |
+
{"t": 157.87, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.018137}
|
| 227 |
+
{"t": 157.87, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.024483}
|
| 228 |
+
{"t": 157.87, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.018137}
|
| 229 |
+
{"t": 157.87, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.112579}
|
| 230 |
+
{"t": 157.87, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.027667}
|
| 231 |
+
{"t": 157.87, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.027667}
|
| 232 |
+
{"t": 157.87, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.725281}
|
| 233 |
+
{"t": 157.87, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.023024}
|
| 234 |
+
{"t": 157.87, "passo": null, "atributo": "som_atn/entropia", "valor": 1.087043}
|
| 235 |
+
{"t": 166.91, "passo": 17, "perda/treino": 9.672022, "perda/ce": 9.672022, "grad_norm": 37.314571, "lr": 0.002998, "abmo/L_hat": 1.003726, "pcgrad/dot": -0.599276, "pcgrad/projetado": 1.0, "conf/h_t": 0.457077, "prs/trust": 0.51905, "prs/clip": 5.0, "prs/boost": 1.108148, "prs/streak": 1.0, "rpp/rho": 1.0, "rpp/streak": 9.0}
|
| 236 |
+
{"t": 166.91, "passo": 17, "atributo": "mtp/entropia", "valor": 0.692937}
|
| 237 |
+
{"t": 166.91, "passo": 17, "atributo": "mtp/termos_ce", "valor": 293.0}
|
| 238 |
+
{"t": 167.22, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.022874}
|
| 239 |
+
{"t": 167.22, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.0187}
|
| 240 |
+
{"t": 167.22, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.025242}
|
| 241 |
+
{"t": 167.22, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.0187}
|
| 242 |
+
{"t": 167.22, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.116883}
|
| 243 |
+
{"t": 167.22, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.027089}
|
| 244 |
+
{"t": 167.22, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.027089}
|
| 245 |
+
{"t": 167.22, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.720549}
|
| 246 |
+
{"t": 167.22, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.022874}
|
| 247 |
+
{"t": 167.22, "passo": null, "atributo": "som_atn/entropia", "valor": 1.09687}
|
| 248 |
+
{"t": 170.81, "passo": 18, "perda/treino": 9.584485, "perda/ce": 9.584485, "grad_norm": 71.603111, "lr": 0.002998, "abmo/L_hat": 1.002474, "pcgrad/dot": 1.566364, "pcgrad/projetado": 0.0, "conf/h_t": 0.474384, "prs/trust": 0.543097, "prs/clip": 5.0, "prs/boost": 1.109196, "prs/streak": 2.0, "rpp/rho": 1.0, "rpp/streak": 0.0}
|
| 249 |
+
{"t": 170.81, "passo": 18, "atributo": "mtp/entropia", "valor": 0.692897}
|
| 250 |
+
{"t": 170.81, "passo": 18, "atributo": "mtp/termos_ce", "valor": 53.0}
|
| 251 |
+
{"t": 171.34, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.023326}
|
| 252 |
+
{"t": 171.34, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.0185}
|
| 253 |
+
{"t": 171.34, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.024808}
|
| 254 |
+
{"t": 171.34, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.0185}
|
| 255 |
+
{"t": 171.34, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.106339}
|
| 256 |
+
{"t": 171.34, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.025216}
|
| 257 |
+
{"t": 171.34, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.025216}
|
| 258 |
+
{"t": 171.34, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.734771}
|
| 259 |
+
{"t": 171.34, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.023326}
|
| 260 |
+
{"t": 171.34, "passo": null, "atributo": "som_atn/entropia", "valor": 1.064465}
|
| 261 |
+
{"t": 179.46, "passo": 19, "perda/treino": 9.743029, "perda/ce": 9.743029, "grad_norm": 1.791735, "lr": 0.002997, "abmo/L_hat": 1.002987, "pcgrad/dot": -0.040954, "pcgrad/projetado": 1.0, "conf/h_t": 0.464889, "prs/trust": 0.532235, "prs/clip": 5.0, "prs/boost": 1.108029, "prs/streak": 0.0, "rpp/rho": 1.0, "rpp/streak": 0.0}
|
| 262 |
+
{"t": 179.46, "passo": 19, "atributo": "mtp/entropia", "valor": 0.692856}
|
| 263 |
+
{"t": 179.46, "passo": 19, "atributo": "mtp/termos_ce", "valor": 520.0}
|
| 264 |
+
{"t": 179.83, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.022873}
|
| 265 |
+
{"t": 179.83, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.019097}
|
| 266 |
+
{"t": 179.83, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.025778}
|
| 267 |
+
{"t": 179.83, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.019097}
|
| 268 |
+
{"t": 179.83, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.117428}
|
| 269 |
+
{"t": 179.83, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.026178}
|
| 270 |
+
{"t": 179.83, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.026178}
|
| 271 |
+
{"t": 179.83, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.720499}
|
| 272 |
+
{"t": 179.83, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.022873}
|
| 273 |
+
{"t": 179.83, "passo": null, "atributo": "som_atn/entropia", "valor": 1.096094}
|
| 274 |
+
{"t": 181.51, "passo": 20, "atributo": "roteador/eq", "valor": 11.656016}
|
| 275 |
+
{"t": 181.51, "passo": 20, "perda/treino": 9.713767, "perda/ce": 9.713767, "grad_norm": 74.328568, "lr": 0.002997, "abmo/L_hat": 1.003146, "pcgrad/dot": 1.175512, "pcgrad/projetado": 0.0, "conf/h_t": 0.48124, "prs/trust": 0.521591, "prs/clip": 5.0, "prs/boost": 1.108501, "prs/streak": 0.0, "rpp/rho": 1.0, "rpp/streak": 0.0}
|
| 276 |
+
{"t": 181.51, "passo": 20, "atributo": "mtp/entropia", "valor": 0.692804}
|
| 277 |
+
{"t": 181.51, "passo": 20, "atributo": "mtp/termos_ce", "valor": 140.0}
|
| 278 |
+
{"t": 182.05, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.022906}
|
| 279 |
+
{"t": 182.05, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.019705}
|
| 280 |
+
{"t": 182.05, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.026599}
|
| 281 |
+
{"t": 182.05, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.019705}
|
| 282 |
+
{"t": 182.05, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.117514}
|
| 283 |
+
{"t": 182.05, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.02456}
|
| 284 |
+
{"t": 182.05, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.02456}
|
| 285 |
+
{"t": 182.05, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.721545}
|
| 286 |
+
{"t": 182.05, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.022906}
|
| 287 |
+
{"t": 182.05, "passo": null, "atributo": "som_atn/entropia", "valor": 1.093089}
|
| 288 |
+
{"t": 191.15, "passo": 20, "atributo": "muonclip/logit_max", "valor": 1.634554}
|
| 289 |
+
{"t": 191.15, "passo": 21, "perda/treino": 9.789917, "perda/ce": 9.789917, "grad_norm": 12.544508, "lr": 0.002996, "abmo/L_hat": 1.004077, "pcgrad/dot": 0.043892, "pcgrad/projetado": 0.0, "conf/h_t": 0.462291, "prs/trust": 0.511159, "prs/clip": 5.0, "prs/boost": 1.108118, "prs/streak": 0.0, "rpp/rho": 1.0, "rpp/streak": 0.0}
|
| 290 |
+
{"t": 191.15, "passo": 21, "atributo": "mtp/entropia", "valor": 0.692751}
|
| 291 |
+
{"t": 191.15, "passo": 21, "atributo": "mtp/termos_ce", "valor": 888.0}
|
| 292 |
+
{"t": 191.94, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.023016}
|
| 293 |
+
{"t": 191.94, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.018769}
|
| 294 |
+
{"t": 191.94, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.025335}
|
| 295 |
+
{"t": 191.94, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.018769}
|
| 296 |
+
{"t": 191.94, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.116855}
|
| 297 |
+
{"t": 191.94, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.024614}
|
| 298 |
+
{"t": 191.94, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.024614}
|
| 299 |
+
{"t": 191.94, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.725013}
|
| 300 |
+
{"t": 191.94, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.023016}
|
| 301 |
+
{"t": 191.94, "passo": null, "atributo": "som_atn/entropia", "valor": 1.081535}
|
| 302 |
+
{"t": 200.82, "passo": 22, "perda/treino": 9.694399, "perda/ce": 9.694399, "grad_norm": 15.782683, "lr": 0.002995, "abmo/L_hat": 1.003927, "pcgrad/dot": 0.111748, "pcgrad/projetado": 0.0, "conf/h_t": 0.468512, "prs/trust": 0.500936, "prs/clip": 5.0, "prs/boost": 1.112651, "prs/streak": 0.0, "rpp/rho": 1.0, "rpp/streak": 0.0}
|
| 303 |
+
{"t": 200.82, "passo": 22, "atributo": "mtp/entropia", "valor": 0.692692}
|
| 304 |
+
{"t": 200.82, "passo": 22, "atributo": "mtp/termos_ce", "valor": 518.0}
|
| 305 |
+
{"t": 201.13, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.022847}
|
| 306 |
+
{"t": 201.13, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.02023}
|
| 307 |
+
{"t": 201.13, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.027308}
|
| 308 |
+
{"t": 201.13, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.02023}
|
| 309 |
+
{"t": 201.13, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.118963}
|
| 310 |
+
{"t": 201.13, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.023944}
|
| 311 |
+
{"t": 201.13, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.023944}
|
| 312 |
+
{"t": 201.13, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.719687}
|
| 313 |
+
{"t": 201.13, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.022847}
|
| 314 |
+
{"t": 201.13, "passo": null, "atributo": "som_atn/entropia", "valor": 1.097535}
|
| 315 |
+
{"t": 208.85, "passo": 23, "perda/treino": 8.577204, "perda/ce": 8.577204, "grad_norm": 400.755951, "lr": 0.002994, "abmo/L_hat": 0.992238, "pcgrad/dot": 2.146239, "pcgrad/projetado": 0.0, "conf/h_t": 0.487627, "prs/trust": 0.525889, "prs/clip": 5.0, "prs/boost": 1.12171, "prs/streak": 1.0, "rpp/rho": 1.0, "rpp/streak": 1.0}
|
| 316 |
+
{"t": 208.85, "passo": 23, "atributo": "mtp/entropia", "valor": 0.692627}
|
| 317 |
+
{"t": 208.85, "passo": 23, "atributo": "mtp/termos_ce", "valor": 4.0}
|
| 318 |
+
{"t": 209.16, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.022882}
|
| 319 |
+
{"t": 209.16, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.01998}
|
| 320 |
+
{"t": 209.16, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.02697}
|
| 321 |
+
{"t": 209.16, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.01998}
|
| 322 |
+
{"t": 209.16, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.119144}
|
| 323 |
+
{"t": 209.16, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.023692}
|
| 324 |
+
{"t": 209.16, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.023692}
|
| 325 |
+
{"t": 209.16, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.720779}
|
| 326 |
+
{"t": 209.16, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.022882}
|
| 327 |
+
{"t": 209.16, "passo": null, "atributo": "som_atn/entropia", "valor": 1.093482}
|
| 328 |
+
{"t": 216.75, "passo": 24, "perda/treino": 8.478787, "perda/ce": 8.478787, "grad_norm": 20.306791, "lr": 0.002993, "abmo/L_hat": 0.9807, "pcgrad/dot": -4.500329, "pcgrad/projetado": 1.0, "conf/h_t": 0.476761, "prs/trust": 0.549595, "prs/clip": 5.0, "prs/boost": 1.122396, "prs/streak": 2.0, "rpp/rho": 1.0, "rpp/streak": 2.0}
|
| 329 |
+
{"t": 216.75, "passo": 24, "atributo": "mtp/entropia", "valor": 0.692562}
|
| 330 |
+
{"t": 216.75, "passo": 24, "atributo": "mtp/termos_ce", "valor": 4.0}
|
| 331 |
+
{"t": 217.14, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.023048}
|
| 332 |
+
{"t": 217.14, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.019364}
|
| 333 |
+
{"t": 217.14, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.025628}
|
| 334 |
+
{"t": 217.14, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.019364}
|
| 335 |
+
{"t": 217.14, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.112299}
|
| 336 |
+
{"t": 217.14, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.025619}
|
| 337 |
+
{"t": 217.14, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.025619}
|
| 338 |
+
{"t": 217.14, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.726012}
|
| 339 |
+
{"t": 217.14, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.023048}
|
| 340 |
+
{"t": 217.14, "passo": null, "atributo": "som_atn/entropia", "valor": 1.085508}
|
| 341 |
+
{"t": 219.02, "passo": 25, "perda/treino": 9.714931, "perda/ce": 9.714931, "grad_norm": 70.911812, "lr": 0.002993, "abmo/L_hat": 0.9831, "pcgrad/dot": 0.480577, "pcgrad/projetado": 0.0, "conf/h_t": 0.173077, "prs/trust": 0.538603, "prs/clip": 5.0, "prs/boost": 1.112225, "prs/streak": 0.0, "rpp/rho": 1.0, "rpp/streak": 3.0}
|
| 342 |
+
{"t": 219.02, "passo": 25, "atributo": "mtp/entropia", "valor": 0.692493}
|
| 343 |
+
{"t": 219.02, "passo": 25, "atributo": "mtp/termos_ce", "valor": 169.0}
|
| 344 |
+
{"t": 219.55, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.02286}
|
| 345 |
+
{"t": 219.55, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.020402}
|
| 346 |
+
{"t": 219.55, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.02754}
|
| 347 |
+
{"t": 219.55, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.020402}
|
| 348 |
+
{"t": 219.55, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.118462}
|
| 349 |
+
{"t": 219.55, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.023689}
|
| 350 |
+
{"t": 219.55, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.023689}
|
| 351 |
+
{"t": 219.55, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.720094}
|
| 352 |
+
{"t": 219.55, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.02286}
|
| 353 |
+
{"t": 219.55, "passo": null, "atributo": "som_atn/entropia", "valor": 1.096746}
|
| 354 |
+
{"t": 229.04, "passo": 26, "perda/treino": 9.799737, "perda/ce": 9.799737, "grad_norm": 3.102093, "lr": 0.002992, "abmo/L_hat": 0.986137, "pcgrad/dot": -0.188172, "pcgrad/projetado": 1.0, "conf/h_t": 0.468464, "prs/trust": 0.527831, "prs/clip": 5.0, "prs/boost": 1.112321, "prs/streak": 0.0, "rpp/rho": 1.0, "rpp/streak": 4.0}
|
| 355 |
+
{"t": 229.04, "passo": 26, "atributo": "mtp/entropia", "valor": 0.692423}
|
| 356 |
+
{"t": 229.04, "passo": 26, "atributo": "mtp/termos_ce", "valor": 883.0}
|
| 357 |
+
{"t": 229.61, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.023356}
|
| 358 |
+
{"t": 229.61, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.018078}
|
| 359 |
+
{"t": 229.61, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.023516}
|
| 360 |
+
{"t": 229.61, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.018078}
|
| 361 |
+
{"t": 229.61, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.109939}
|
| 362 |
+
{"t": 229.61, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.02398}
|
| 363 |
+
{"t": 229.61, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.02398}
|
| 364 |
+
{"t": 229.61, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.735718}
|
| 365 |
+
{"t": 229.61, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.023356}
|
| 366 |
+
{"t": 229.61, "passo": null, "atributo": "som_atn/entropia", "valor": 1.054939}
|
| 367 |
+
{"t": 239.52, "passo": 27, "perda/treino": 9.458589, "perda/ce": 9.458589, "grad_norm": 17.251015, "lr": 0.002991, "abmo/L_hat": 0.985342, "pcgrad/dot": 0.096223, "pcgrad/projetado": 0.0, "conf/h_t": 0.483185, "prs/trust": 0.551439, "prs/clip": 5.0, "prs/boost": 1.115255, "prs/streak": 1.0, "rpp/rho": 1.0, "rpp/streak": 5.0}
|
| 368 |
+
{"t": 239.52, "passo": 27, "atributo": "mtp/entropia", "valor": 0.69234}
|
| 369 |
+
{"t": 239.52, "passo": 27, "atributo": "mtp/termos_ce", "valor": 527.0}
|
| 370 |
+
{"t": 240.93, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.022983}
|
| 371 |
+
{"t": 240.93, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.018958}
|
| 372 |
+
{"t": 240.93, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.02502}
|
| 373 |
+
{"t": 240.93, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.018958}
|
| 374 |
+
{"t": 240.93, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.113355}
|
| 375 |
+
{"t": 240.93, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.026885}
|
| 376 |
+
{"t": 240.93, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.026885}
|
| 377 |
+
{"t": 240.93, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.723974}
|
| 378 |
+
{"t": 240.93, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.022983}
|
| 379 |
+
{"t": 240.93, "passo": null, "atributo": "som_atn/entropia", "valor": 1.090848}
|
| 380 |
+
{"t": 288.0, "passo": 28, "perda/treino": 9.695085, "perda/ce": 9.695085, "grad_norm": 23.84824, "lr": 0.00299, "abmo/L_hat": 0.987073, "pcgrad/dot": -0.167384, "pcgrad/projetado": 1.0, "conf/h_t": 0.448319, "prs/trust": 0.54041, "prs/clip": 5.0, "prs/boost": 1.112396, "prs/streak": 0.0, "rpp/rho": 1.0, "rpp/streak": 6.0}
|
| 381 |
+
{"t": 288.0, "passo": 28, "atributo": "mtp/entropia", "valor": 0.692252}
|
| 382 |
+
{"t": 288.0, "passo": 28, "atributo": "mtp/termos_ce", "valor": 221.0}
|
| 383 |
+
{"t": 288.8, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.023113}
|
| 384 |
+
{"t": 288.8, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.019089}
|
| 385 |
+
{"t": 288.8, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.024719}
|
| 386 |
+
{"t": 288.8, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.019089}
|
| 387 |
+
{"t": 288.8, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.113342}
|
| 388 |
+
{"t": 288.8, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.024738}
|
| 389 |
+
{"t": 288.8, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.024738}
|
| 390 |
+
{"t": 288.8, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.728058}
|
| 391 |
+
{"t": 288.8, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.023113}
|
| 392 |
+
{"t": 288.8, "passo": null, "atributo": "som_atn/entropia", "valor": 1.077035}
|
| 393 |
+
{"t": 298.15, "passo": 29, "perda/treino": 9.503708, "perda/ce": 9.503708, "grad_norm": 8.924388, "lr": 0.002988, "abmo/L_hat": 0.986651, "pcgrad/dot": 0.002497, "pcgrad/projetado": 0.0, "conf/h_t": 0.48544, "prs/trust": 0.56339, "prs/clip": 5.0, "prs/boost": 1.111458, "prs/streak": 1.0, "rpp/rho": 1.0, "rpp/streak": 7.0}
|
| 394 |
+
{"t": 298.15, "passo": 29, "atributo": "mtp/entropia", "valor": 0.69216}
|
| 395 |
+
{"t": 298.15, "passo": 29, "atributo": "mtp/termos_ce", "valor": 849.0}
|
| 396 |
+
{"t": 299.61, "passo": null, "atributo": "som_atn/peso_som", "valor": 0.023161}
|
| 397 |
+
{"t": 299.61, "passo": null, "atributo": "som_atn/peso_gg", "valor": 0.018201}
|
| 398 |
+
{"t": 299.61, "passo": null, "atributo": "som_atn/peso_gcs", "valor": 0.024378}
|
| 399 |
+
{"t": 299.61, "passo": null, "atributo": "som_atn/peso_gsom", "valor": 0.018201}
|
| 400 |
+
{"t": 299.61, "passo": null, "atributo": "som_atn/peso_hsom", "valor": 0.110916}
|
| 401 |
+
{"t": 299.61, "passo": null, "atributo": "som_atn/peso_tkm", "valor": 0.026193}
|
| 402 |
+
{"t": 299.61, "passo": null, "atributo": "som_atn/peso_rsom", "valor": 0.026193}
|
| 403 |
+
{"t": 299.61, "passo": null, "atributo": "som_atn/peso_cpn", "valor": 0.729596}
|
| 404 |
+
{"t": 299.61, "passo": null, "atributo": "som_atn/peso_ssom", "valor": 0.023161}
|
| 405 |
+
{"t": 299.61, "passo": null, "atributo": "som_atn/entropia", "valor": 1.074913}
|
| 406 |
+
{"t": 341.91, "passo": 30, "perda/treino": 9.761606, "perda/ce": 9.761606, "grad_norm": 94.320343, "lr": 0.002987, "abmo/L_hat": 0.988938, "pcgrad/dot": -1.116409, "pcgrad/projetado": 1.0, "conf/h_t": 0.448685, "prs/trust": 0.552122, "prs/clip": 5.0, "prs/boost": 1.112944, "prs/streak": 0.0, "rpp/rho": 1.0, "rpp/streak": 8.0}
|
| 407 |
+
{"t": 341.91, "passo": 30, "atributo": "mtp/entropia", "valor": 0.692057}
|
| 408 |
+
{"t": 341.91, "passo": 30, "atributo": "mtp/termos_ce", "valor": 409.0}
|
telemetria_out/treino_v17_v17_khmambajepa.jsonl
CHANGED
|
The diff for this file is too large to render.
See raw diff
|
|
|
tests/test_khtst_v17.py
CHANGED
|
@@ -1,351 +1,394 @@
|
|
| 1 |
# -*- coding: utf-8 -*-
|
| 2 |
-
"""SUÍTE v17
|
| 3 |
-
ReplaySSM
|
|
|
|
|
|
|
| 4 |
|
| 5 |
-
|
| 6 |
-
1) Teo 29.1 — stack fora do contrato de subclass (reproduz o bug v16);
|
| 7 |
-
slice/narrow DENTRO do opset (o caminho do indexador existe)
|
| 8 |
-
2) Teo 29.2 — INDEXADOR ≡ empilhado: independência por linha da GEMM
|
| 9 |
-
(bit-exato na fatia) e erro ≤ cota h/√12 (Teo 28.1) vs FP32
|
| 10 |
-
3) Teo 29.3 — FLORESTA: B blocos ≡ pool único (índice global exato) e
|
| 11 |
-
E[blocos tocados] ≤ B(1−(1−1/B)^K) respeitado
|
| 12 |
-
4) Teo 29.4 — forward MoE quantizado usa {linear, narrow, dequantize}
|
| 13 |
-
apenas; treino FP32 intacto (nascimento neutro por modo)
|
| 14 |
-
5) Teo 29.5/29.6 — NORMALIZAÇÃO DIMENSIONAL: L_norm·d = L_raw exato e
|
| 15 |
-
λ' = λ·d reproduz o aux da v16 (reparametrização pura)
|
| 16 |
-
6) Teo 29.7 — TAXA invariante de escala: ‖e‖²≤θ‖alvo‖² homogênea
|
| 17 |
-
7) Teo 29.8 — RM λ_jepa: sinal (r*−r)/r* puxa a razão para a banda
|
| 18 |
-
8) Teo 29.11 — TOKEN ROUNDING: n̂=τ⌈n/τ⌉, offsets exatos, GEMM estofado
|
| 19 |
-
≡ GEMM sem estofamento nas linhas reais (bit-exato), ρ(τ) monotônica
|
| 20 |
-
9) Teo 29.12 — AUDITORIA: I1–I4 com marcadores de fase reais
|
| 21 |
-
10) Teo 29.13/29.14 — REPLAYSSM: saída bit-exata vs varredura completa,
|
| 22 |
-
gradiente exato (≤1e-6 rel), memória de ativação menor (medida)
|
| 23 |
-
11) vocab 16384 preservado; pools serializáveis no state_dict
|
| 24 |
-
12) gestor de memória + auditoria integrados ao treinador (smoke)
|
| 25 |
-
"""
|
| 26 |
import os
|
| 27 |
import sys
|
|
|
|
| 28 |
|
| 29 |
-
|
| 30 |
-
|
| 31 |
-
import math
|
| 32 |
import torch
|
| 33 |
-
import torch.nn as nn
|
| 34 |
-
import torch.nn.functional as F
|
| 35 |
|
| 36 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
|
| 38 |
-
_verificacoes = []
|
| 39 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 40 |
|
| 41 |
-
def check(nome, cond, extra=""):
|
| 42 |
-
_verificacoes.append((nome, bool(cond), extra))
|
| 43 |
-
print(f" [{'OK' if cond else 'FALHA'}] {nome} {extra}")
|
| 44 |
|
|
|
|
|
|
|
|
|
|
| 45 |
|
| 46 |
-
|
| 47 |
-
|
| 48 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 49 |
|
| 50 |
-
|
| 51 |
-
|
| 52 |
-
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
try:
|
| 56 |
-
torch.stack([lin.weight, lin.weight])
|
| 57 |
-
except NotImplementedError:
|
| 58 |
-
stack_falha = True
|
| 59 |
-
check("torch.stack em Int8Tensor levanta NotImplementedError", stack_falha)
|
| 60 |
-
slice_ok = False
|
| 61 |
-
try:
|
| 62 |
-
w_f = lin.weight.narrow(0, 0, 8)
|
| 63 |
-
_ = F.linear(torch.randn(4, 32), w_f)
|
| 64 |
-
slice_ok = True
|
| 65 |
-
except Exception:
|
| 66 |
-
pass
|
| 67 |
-
check("narrow + F.linear DENTRO do opset (caminho do indexador)", slice_ok)
|
| 68 |
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 79 |
|
| 80 |
-
|
| 81 |
-
|
| 82 |
-
|
| 83 |
-
|
| 84 |
-
|
| 85 |
-
|
| 86 |
-
moe.eval()
|
| 87 |
-
xb = torch.randn(10, 5, d)
|
| 88 |
-
with torch.no_grad():
|
| 89 |
-
y_ref = moe(xb) # FP32
|
| 90 |
-
ev = EmpacotadorMoE(moe, n_blocos=1).quantizar("int8_wo")
|
| 91 |
-
check("empacotamento respeita cota h/√12 (Teo 28.1)",
|
| 92 |
-
all(v["respeita_cota"] for v in ev["matrizes"].values()),
|
| 93 |
-
f"erro médio {ev['matrizes']['gate']['erro_rel_medio']:.4f}")
|
| 94 |
-
with torch.no_grad():
|
| 95 |
-
y_q = moe(xb)
|
| 96 |
-
delta = float((y_q - y_ref).norm() / (y_ref.norm() + 1e-30))
|
| 97 |
-
check("forward INDEXADOR ≈ FP32 (Δ ≤ 2× erro de quantização)",
|
| 98 |
-
delta < 0.05, f"Δ={delta:.4f}")
|
| 99 |
-
check("toques = especialistas ativos (I4 — SonicMoE)",
|
| 100 |
-
getattr(moe, "sonic_toques", 0) <= getattr(moe, "sonic_n_ativos", 99),
|
| 101 |
-
f"toques={moe.sonic_toques}")
|
| 102 |
-
# floresta B=2 ≡ B=1 (índice global)
|
| 103 |
-
desfazer_pools(moe)
|
| 104 |
-
EmpacotadorMoE(moe, n_blocos=2).quantizar("int8_wo")
|
| 105 |
-
with torch.no_grad():
|
| 106 |
-
y_f2 = moe(xb)
|
| 107 |
-
check("Teo 29.3 — floresta B=2 ≡ B=1 (bit-exato)", torch.equal(y_f2, y_q))
|
| 108 |
-
# treino FP32 intacto
|
| 109 |
-
moe.train()
|
| 110 |
-
y_tr = moe(xb)
|
| 111 |
-
check("Teo 29.4 — treino FP32 intacto (nascimento neutro por modo)",
|
| 112 |
-
bool(torch.isfinite(y_tr).all()) and y_tr.shape == y_ref.shape)
|
| 113 |
-
# pools no state_dict (serialização)
|
| 114 |
-
chaves = [k for k in moe.state_dict().keys() if "pool_" in k]
|
| 115 |
-
check("pools entram no state_dict (serializáveis)", len(chaves) >= 3,
|
| 116 |
-
f"{len(chaves)} chaves")
|
| 117 |
|
| 118 |
-
|
| 119 |
-
|
| 120 |
-
|
| 121 |
-
|
| 122 |
-
|
| 123 |
-
|
| 124 |
-
|
| 125 |
-
|
| 126 |
-
|
| 127 |
-
|
| 128 |
-
|
| 129 |
-
|
| 130 |
-
|
| 131 |
-
|
| 132 |
-
|
| 133 |
-
|
| 134 |
-
L_r, tel_r = mj_raw.perda(lat, gerador=torch.Generator().manual_seed(2026))
|
| 135 |
-
razao = float(L_n.detach()) / max(float(L_r.detach()), 1e-12)
|
| 136 |
-
check("L_normalizada = L_soma / d_latente (exato)",
|
| 137 |
-
abs(razao - 1.0 / d_lat) < 1e-6, f"razão={razao:.6f} vs {1/d_lat:.6f}")
|
| 138 |
-
check("perda v16 era ALTA (soma ≈ 25) e a v17 fica O(1)",
|
| 139 |
-
float(L_r.detach()) > 5.0 and float(L_n.detach()) < 1.0,
|
| 140 |
-
f"raw={float(L_r.detach()):.2f} → norm={float(L_n.detach()):.3f}")
|
| 141 |
-
# gradiente escala exatamente 1/d (Teo 29.6 — linearidade)
|
| 142 |
-
lat1 = lat.clone().requires_grad_(True)
|
| 143 |
-
L1, _ = mj_norm.perda(lat1, gerador=torch.Generator().manual_seed(2026))
|
| 144 |
-
L1.backward()
|
| 145 |
-
g_n = lat1.grad.clone()
|
| 146 |
-
lat2 = lat.clone().requires_grad_(True)
|
| 147 |
-
L2, _ = mj_raw.perda(lat2, gerador=torch.Generator().manual_seed(2026))
|
| 148 |
-
L2.backward()
|
| 149 |
-
g_r = lat2.grad.clone()
|
| 150 |
-
check("∇L_norm = ∇L_raw / d (linearidade — gradiente EXATO)",
|
| 151 |
-
float((g_n - g_r / d_lat).norm() / (g_r.norm() + 1e-12)) < 1e-6)
|
| 152 |
|
| 153 |
-
|
| 154 |
-
|
| 155 |
-
|
| 156 |
-
|
| 157 |
-
|
| 158 |
-
|
| 159 |
-
|
| 160 |
-
|
| 161 |
-
|
| 162 |
-
|
| 163 |
-
|
| 164 |
-
|
| 165 |
-
|
| 166 |
-
|
| 167 |
-
|
| 168 |
-
|
| 169 |
-
|
| 170 |
-
|
| 171 |
-
|
| 172 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 173 |
|
| 174 |
-
|
| 175 |
-
|
| 176 |
-
|
| 177 |
-
|
| 178 |
-
|
| 179 |
-
|
| 180 |
-
|
| 181 |
-
|
| 182 |
-
|
| 183 |
-
|
| 184 |
-
|
| 185 |
-
|
| 186 |
-
|
| 187 |
-
|
|
|
|
|
|
|
| 188 |
|
| 189 |
-
|
| 190 |
-
|
| 191 |
-
|
| 192 |
-
|
| 193 |
-
|
| 194 |
-
|
| 195 |
-
|
| 196 |
-
|
| 197 |
-
|
| 198 |
-
|
| 199 |
-
|
| 200 |
-
|
| 201 |
-
list(off_r) == [0, 3, 3, 10] and list(off_p) == [0, 4, 4, 12])
|
| 202 |
-
check("overhead ρ = Σn̂/Σn", abs(rho - 20 / 15) < 1e-12, f"ρ={rho:.4f}")
|
| 203 |
-
# exatidão do GEMM estofado nas linhas reais
|
| 204 |
-
X = torch.randn(15, 32) # Σn = 15
|
| 205 |
-
W = torch.randn(8, 32)
|
| 206 |
-
buf = torch.zeros(int(off_p[-1] + n_hat[-1]), 32)
|
| 207 |
-
for e in range(4):
|
| 208 |
-
buf[int(off_p[e]):int(off_p[e]) + int(n[e])] = X[int(off_r[e]):int(off_r[e]) + int(n[e])]
|
| 209 |
-
y_padded = F.linear(buf, W)
|
| 210 |
-
y_ref = F.linear(X, W)
|
| 211 |
-
exato = True
|
| 212 |
-
for e in range(4):
|
| 213 |
-
if n[e] == 0:
|
| 214 |
-
continue
|
| 215 |
-
yp = y_padded[int(off_p[e]):int(off_p[e]) + int(n[e])]
|
| 216 |
-
yr = y_ref[int(off_r[e]):int(off_r[e]) + int(n[e])]
|
| 217 |
-
if not torch.equal(yp, yr):
|
| 218 |
-
exato = False
|
| 219 |
-
check("GEMM estofado ≡ GEMM real nas linhas vivas (bit-exato — Teo 29.11)",
|
| 220 |
-
exato)
|
| 221 |
-
# linhas de estofamento: [off_p[e]+n[e], off_p[e]+n̂[e]) são ZERO
|
| 222 |
-
zeros_ok = True
|
| 223 |
-
for e in range(4):
|
| 224 |
-
a0 = int(off_p[e]) + int(n[e])
|
| 225 |
-
a1 = int(off_p[e]) + int(n_hat[e])
|
| 226 |
-
if a1 > a0 and float(buf[a0:a1].abs().max()) != 0.0:
|
| 227 |
-
zeros_ok = False
|
| 228 |
-
if a1 > a0 and float(y_padded[a0:a1].abs().max()) != 0.0:
|
| 229 |
-
zeros_ok = False
|
| 230 |
-
check("linhas estofadas produzem ZERO (linear sem viés)", zeros_ok,
|
| 231 |
-
f"pad_total={int(pad)} linhas nulas")
|
| 232 |
-
t_star = escolher_tile(np.array([100, 90, 120, 95], dtype=np.int64))
|
| 233 |
-
check("τ* por medição (argmin ρ̂)", int(t_star) >= 1, f"τ*={t_star}")
|
| 234 |
|
| 235 |
-
|
| 236 |
-
|
| 237 |
-
|
| 238 |
-
|
| 239 |
-
|
| 240 |
-
|
| 241 |
-
|
| 242 |
-
|
| 243 |
-
|
| 244 |
-
|
| 245 |
-
|
| 246 |
-
|
| 247 |
-
|
| 248 |
-
|
| 249 |
-
|
| 250 |
-
|
| 251 |
-
|
| 252 |
-
|
| 253 |
-
|
| 254 |
-
|
| 255 |
-
|
| 256 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 257 |
|
| 258 |
-
|
| 259 |
-
|
| 260 |
-
|
| 261 |
-
|
| 262 |
-
|
| 263 |
-
|
| 264 |
-
|
| 265 |
-
|
| 266 |
-
|
| 267 |
-
|
| 268 |
-
|
| 269 |
-
|
| 270 |
-
|
| 271 |
-
|
| 272 |
-
|
| 273 |
-
|
| 274 |
-
|
| 275 |
-
|
| 276 |
-
|
| 277 |
-
|
| 278 |
-
|
| 279 |
-
|
| 280 |
-
|
| 281 |
-
|
| 282 |
-
|
| 283 |
-
|
| 284 |
-
|
| 285 |
-
|
| 286 |
-
|
| 287 |
-
|
| 288 |
-
|
| 289 |
-
|
| 290 |
-
|
| 291 |
-
total = {"b": 0}
|
| 292 |
-
vistos = set()
|
| 293 |
-
def pack(t):
|
| 294 |
-
try:
|
| 295 |
-
key = (t.untyped_storage().data_ptr(), t.untyped_storage().nbytes())
|
| 296 |
-
if key not in vistos:
|
| 297 |
-
vistos.add(key)
|
| 298 |
-
total["b"] += int(t.untyped_storage().nbytes())
|
| 299 |
-
except Exception:
|
| 300 |
-
total["b"] += t.numel() * t.element_size()
|
| 301 |
-
return t
|
| 302 |
-
def unpack(t):
|
| 303 |
-
return t
|
| 304 |
-
with torch.autograd.graph.saved_tensors_hooks(pack, unpack):
|
| 305 |
-
out = fn(*args)
|
| 306 |
-
out.pow(2).sum().backward()
|
| 307 |
-
return total["b"]
|
| 308 |
-
u1 = u.clone().requires_grad_(True)
|
| 309 |
-
b_full = bytes_salvos(lambda uu: varredura_trapezoidal(uu, delta, A, B_m, C_m), u1)
|
| 310 |
-
u1.grad = None
|
| 311 |
-
b_rep = bytes_salvos(lambda uu: varredura_replay(uu, delta, A, B_m, C_m,
|
| 312 |
-
tam_segmento=8), u1)
|
| 313 |
-
check("Teo 29.14 — ativação guardada pelo replay < caminho completo",
|
| 314 |
-
b_rep < b_full, f"replay={b_rep/1e3:.1f}KB < full={b_full/1e3:.1f}KB")
|
| 315 |
|
| 316 |
-
|
| 317 |
-
|
| 318 |
-
|
| 319 |
-
|
| 320 |
-
|
| 321 |
-
|
| 322 |
-
|
| 323 |
-
|
| 324 |
-
|
| 325 |
-
|
| 326 |
-
|
| 327 |
-
|
| 328 |
-
|
| 329 |
-
|
| 330 |
-
|
| 331 |
-
|
| 332 |
-
|
| 333 |
-
|
| 334 |
-
|
| 335 |
-
|
| 336 |
-
|
| 337 |
-
|
| 338 |
-
|
| 339 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 340 |
|
| 341 |
-
|
| 342 |
-
|
| 343 |
-
|
| 344 |
-
|
| 345 |
-
|
| 346 |
-
|
| 347 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 348 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 349 |
|
| 350 |
-
|
| 351 |
-
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
# -*- coding: utf-8 -*-
|
| 2 |
+
"""SUÍTE v17 KHMAMBAJEPA — Problema 1 (MoE/quantização), Problema 2 (TCE),
|
| 3 |
+
SonicMoE + ReplaySSM (auditoria I1–I4, Token Rounding) e autoajuste
|
| 4 |
+
multi-método (doc 29 Teos 29.1–29.14; doc 30 Teos 30.1–30.15)."""
|
| 5 |
+
from __future__ import annotations
|
| 6 |
|
| 7 |
+
import math
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 8 |
import os
|
| 9 |
import sys
|
| 10 |
+
import tempfile
|
| 11 |
|
| 12 |
+
import numpy as np
|
|
|
|
|
|
|
| 13 |
import torch
|
|
|
|
|
|
|
| 14 |
|
| 15 |
+
sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..", "src"))
|
| 16 |
+
sys.path.insert(0, os.path.join(os.path.dirname(__file__), "..",
|
| 17 |
+
"src", "khtst", "acelerado"))
|
| 18 |
+
|
| 19 |
+
VERDE = 0
|
| 20 |
+
FALHOU = False
|
| 21 |
|
|
|
|
| 22 |
|
| 23 |
+
def check(nome: str, cond: bool, detalhe: str = ""):
|
| 24 |
+
global VERDE, FALHOU
|
| 25 |
+
if cond:
|
| 26 |
+
VERDE += 1
|
| 27 |
+
print(f" ✓ {nome}" + (f" — {detalhe}" if detalhe else ""))
|
| 28 |
+
else:
|
| 29 |
+
FALHOU = True
|
| 30 |
+
print(f" ✗ {nome}" + (f" — {detalhe}" if detalhe else ""))
|
| 31 |
|
|
|
|
|
|
|
|
|
|
| 32 |
|
| 33 |
+
print("=" * 72)
|
| 34 |
+
print("SUÍTE v17 — MoE-pools · TCE · ReplaySSM · Token Rounding · MULTI")
|
| 35 |
+
print("=" * 72)
|
| 36 |
|
| 37 |
+
# ==========================================================================
|
| 38 |
+
print("== 1) PROBLEMA 1 — sonda do opset + empacotador MoE (Teos 29.1–29.4)")
|
| 39 |
+
# ==========================================================================
|
| 40 |
+
from khtst.quanta.moe_empacotador import (sonda_opset_subclass, empacotar_moe,
|
| 41 |
+
desfazer_pools, forward_pools_mlp,
|
| 42 |
+
PoolsMoE)
|
| 43 |
+
from khtst.percepcao.moe import MoEAgrupavel, experts_vetoriais
|
| 44 |
|
| 45 |
+
s = sonda_opset_subclass()
|
| 46 |
+
check("Teo 29.1 (evidência): aten.stack FALHA no subclass",
|
| 47 |
+
"FALHA" in s["stack"], f"stack={s['stack']}")
|
| 48 |
+
check("narrow/slice e F.linear implementados (indexador legal)",
|
| 49 |
+
s["narrow"] == "ok" and s["linear"] == "ok")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 50 |
|
| 51 |
+
torch.manual_seed(2026)
|
| 52 |
+
moe = MoEAgrupavel(64, n_grupos=2, experts_por_grupo=3, d_ff_expert=96,
|
| 53 |
+
n_tarefas=5)
|
| 54 |
+
x = torch.randn(3, 6, 64)
|
| 55 |
+
moe.eval()
|
| 56 |
+
with torch.no_grad():
|
| 57 |
+
y0 = moe(x)
|
| 58 |
+
# bijeção φ: linha (e, i) do empilhado ≡ linha e·f + i do pool
|
| 59 |
+
pesos = [e.w_gate.weight for e in moe.especialistas]
|
| 60 |
+
ev = empacotar_moe(moe, modo="int8_wo", n_blocos=2)
|
| 61 |
+
check("empacotamento floresta B=2 ok",
|
| 62 |
+
ev.get("ok") and ev.get("indexador_bit_exato_linha", False),
|
| 63 |
+
f"compressão {ev.get('compressao', 0):.2f}×")
|
| 64 |
+
pool = moe._pools_quant
|
| 65 |
+
fatia_2 = pool.fatia("gate", 2).dequantize()
|
| 66 |
+
# a fatia é a ÚNICA quantização por linha destas linhas: erro ≤ cota
|
| 67 |
+
ref = pesos[2].float()
|
| 68 |
+
h = float(ref.max() - ref.min()) / 255.0
|
| 69 |
+
cota = h / math.sqrt(12.0) * math.sqrt(ref.numel()) * 1.5
|
| 70 |
+
check("Teo 29.2: fatia(e) reconstrói as linhas do expert e (≤ cota h/√12)",
|
| 71 |
+
float((fatia_2 - ref).norm()) <= cota + 1e-8,
|
| 72 |
+
f"erro {float((fatia_2 - ref).norm()):.3e} ≤ {cota:.3e}")
|
| 73 |
+
# índice 2 níveis
|
| 74 |
+
b, j = pool.localizar(4)
|
| 75 |
+
check("Teo 29.3: índice (bloco, slot) consistente",
|
| 76 |
+
b * pool.experts_por_bloco + j == 4 and 0 <= b < pool.n_blocos)
|
| 77 |
+
y1 = moe(x)
|
| 78 |
+
drel = float((y1 - y0).norm() / (y0.norm() + 1e-30))
|
| 79 |
+
check("forward por pools ≈ FP32 (Teo 28.1/28.2)", drel < 0.05,
|
| 80 |
+
f"Δy = {drel:.2e}")
|
| 81 |
+
check("Teo 29.4-b: pools FORA do state_dict (dados derivados)",
|
| 82 |
+
all("pools" not in k for k in dict(moe.named_parameters())))
|
| 83 |
+
ok_rb = desfazer_pools(moe)
|
| 84 |
+
y2 = moe(x)
|
| 85 |
+
check("Teo 29.4: rollback bit-exato", ok_rb and torch.equal(y0, y2))
|
| 86 |
|
| 87 |
+
# ==========================================================================
|
| 88 |
+
print("== 2) torchao INTEGRAL com MoE EMPACOTADA (não mais excluída)")
|
| 89 |
+
# ==========================================================================
|
| 90 |
+
from khtst.quanta.torchao_quant import quantizar_torchao
|
| 91 |
+
from khtst.nucleo.modelo import KHTSTModel
|
| 92 |
+
from khtst.config import Config
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 93 |
|
| 94 |
+
cfg = Config()
|
| 95 |
+
modelo = KHTSTModel(cfg)
|
| 96 |
+
modelo.eval()
|
| 97 |
+
sonda = torch.randint(0, 100, (1, 16))
|
| 98 |
+
with torch.no_grad():
|
| 99 |
+
evq = quantizar_torchao(modelo, cfg={"modo": "auto", "dim_minimo": 64,
|
| 100 |
+
"empacotar_moe": True,
|
| 101 |
+
"n_blocos_floresta": 2},
|
| 102 |
+
entrada_sonda=sonda)
|
| 103 |
+
check("quantização integral ok", bool(evq.get("ok")),
|
| 104 |
+
f"{evq.get('n_camadas_quantizadas', 0)} camadas")
|
| 105 |
+
check("MoE EMPACOTADA (Teo 28.3-iv revogado)",
|
| 106 |
+
evq.get("n_moe_empacotadas", 0) >= 1,
|
| 107 |
+
f"compressão {evq.get('moe_compressao', 0):.2f}×")
|
| 108 |
+
check("Δy da sonda dentro da cota", evq.get("delta_y_relativo", 1.0) < 0.10,
|
| 109 |
+
f"Δy = {evq.get('delta_y_relativo', float('nan')):.2e}")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 110 |
|
| 111 |
+
# ==========================================================================
|
| 112 |
+
print("== 3) PROBLEMA 2 — TCE normalizada + limiar invariante (Teos 29.5–29.8)")
|
| 113 |
+
# ==========================================================================
|
| 114 |
+
from khtst.treino.jepa import MotorJepa, ConfigJepa
|
| 115 |
+
torch.manual_seed(3)
|
| 116 |
+
cfg_sum = ConfigJepa(d_modelo=64, d_latente=32, normalizar_dim=False)
|
| 117 |
+
cfg_nd = ConfigJepa(d_modelo=64, d_latente=32, normalizar_dim=True)
|
| 118 |
+
m_sum, m_nd = MotorJepa(None, cfg_sum), MotorJepa(None, cfg_nd)
|
| 119 |
+
for p in m_sum.parameters():
|
| 120 |
+
torch.nn.init.normal_(p, std=0.2)
|
| 121 |
+
m_nd.load_state_dict(m_sum.state_dict())
|
| 122 |
+
m_nd.funil_bar.bar.load_state_dict(m_sum.funil_bar.bar.state_dict())
|
| 123 |
+
h_fixo = torch.randn(2, 8, 64)
|
| 124 |
+
torch.manual_seed(11)
|
| 125 |
+
_, t_sum = m_sum.perda(h_fixo)
|
| 126 |
+
torch.manual_seed(11)
|
| 127 |
+
_, t_nd = m_nd.perda(h_fixo)
|
| 128 |
+
d_lat = 32
|
| 129 |
+
check("Teo 29.5: L_normalizada = L_soma/d EXATO",
|
| 130 |
+
abs(t_nd["jepa/perda"] - t_sum["jepa/perda_soma"] / d_lat) < 1e-6,
|
| 131 |
+
f"{t_nd['jepa/perda']:.4f} vs {t_sum['jepa/perda_soma'] / d_lat:.4f}")
|
| 132 |
+
base = torch.randn(2, 1, 64)
|
| 133 |
+
h_pred = (base + 0.01 * torch.randn(2, 8, 64)).detach()
|
| 134 |
+
ot = torch.optim.Adam(m_nd.parameters(), lr=0.05)
|
| 135 |
+
for _ in range(300):
|
| 136 |
+
L, _ = m_nd.perda(h_pred)
|
| 137 |
+
ot.zero_grad()
|
| 138 |
+
L.backward()
|
| 139 |
+
ot.step()
|
| 140 |
+
m_nd.passo_pos_backward()
|
| 141 |
+
_, t_t = m_nd.perda(h_pred)
|
| 142 |
+
check("Teo 29.6: taxa > 0 com limiar INVARIANTE (v16: ≡ 0)",
|
| 143 |
+
t_t["jepa/taxa_sucesso"] > 0.0,
|
| 144 |
+
f"taxa {t_t['jepa/taxa_sucesso']:.3f}, perda {t_t['jepa/perda']:.3f}")
|
| 145 |
+
_, t_g = m_nd.perda(h_pred * 50.0)
|
| 146 |
+
check("Teo 29.6: homogeneidade — taxa invariante a c = 50",
|
| 147 |
+
abs(t_g["jepa/taxa_sucesso"] - t_t["jepa/taxa_sucesso"]) < 0.25)
|
| 148 |
+
# Teo 29.8: λ_jepa presente no autorregulador (razão ponderada)
|
| 149 |
+
from khtst.nucleo.autorregulador import Autorregulador
|
| 150 |
+
auto = Autorregulador({"ativo": True, "a_cada": 2, "passos_espera": 0})
|
| 151 |
+
auto.registrar_razao_aux({"jepa": 0.9})
|
| 152 |
+
check("Teo 29.8: λ_jepa AUTORREGULÁVEL (razão > ½ ⟹ desce)",
|
| 153 |
+
auto.lambda_escala("jepa") < 1.0,
|
| 154 |
+
f"escala {auto.lambda_escala('jepa'):.4f}")
|
| 155 |
|
| 156 |
+
# ==========================================================================
|
| 157 |
+
print("== 4) jepa_validator — registro C, Token Rounding, auditoria I1–I4")
|
| 158 |
+
# ==========================================================================
|
| 159 |
+
from jepa_validator import (JEPAManager, token_rounding_plano, escolher_tile,
|
| 160 |
+
AuditoriaMemoriaSonic)
|
| 161 |
+
mg = JEPAManager()
|
| 162 |
+
mg.register_variant("Image-JEPA", 10_000, 10_000, 2_000)
|
| 163 |
+
mg.register_variant("V-JEPA-2", 20_000, 20_000, 4_000, moe=5_000, multimodal=1)
|
| 164 |
+
mg.register_variant("Brain-JEPA", 9_000, 9_000, 1_800, moe=3_000)
|
| 165 |
+
v = mg.verify_all()
|
| 166 |
+
check("registro C long long (auditoria)",
|
| 167 |
+
v["total_global"] == 93_800 and v["n_variantes"] == 3,
|
| 168 |
+
f"total {v['total_global']:,}")
|
| 169 |
+
dist = mg.compute_parameter_distribution()
|
| 170 |
+
check("distribuição float64 contínua (Σ=1)",
|
| 171 |
+
dist.dtype == np.float64 and abs(float(dist.sum()) - 1.0) < 1e-12)
|
| 172 |
|
| 173 |
+
n_hat, rho, off = token_rounding_plano(30, 8)
|
| 174 |
+
check("Teo 29.11: n̂ = τ·⌈n/τ⌉ e ρ exato",
|
| 175 |
+
n_hat == 32 and abs(rho - 32 / 30) < 1e-12 and (off == np.arange(30)).all())
|
| 176 |
+
X = torch.randn(30, 64)
|
| 177 |
+
W = torch.randn(32, 64)
|
| 178 |
+
Xp = torch.zeros(32, 64)
|
| 179 |
+
Xp[:30] = X
|
| 180 |
+
check("Teo 29.11: GEMM estofado BIT-EXATO nas linhas vivas",
|
| 181 |
+
torch.equal(torch.nn.functional.linear(X, W),
|
| 182 |
+
torch.nn.functional.linear(Xp, W)[:30]))
|
| 183 |
+
r_tile = escolher_tile(lambda tau: 0.9 + 0.002 * tau, [8, 16, 32], 1000)
|
| 184 |
+
check("tile escolhido por custo efetivo t·ρ", r_tile["tau"] in (8, 16, 32))
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 185 |
|
| 186 |
+
aud = AuditoriaMemoriaSonic()
|
| 187 |
+
from khtst.mamba3.hybrid import Mamba3Camada
|
| 188 |
+
cam = Mamba3Camada(64, d_estado=8, rank_mimo=2)
|
| 189 |
+
cam.train()
|
| 190 |
+
xa = torch.randn(2, 96, 64, requires_grad=True)
|
| 191 |
+
aud.marcar("fwd_ini")
|
| 192 |
+
ya = cam(xa)
|
| 193 |
+
aud.marcar("fwd_fim")
|
| 194 |
+
aud.amostrar()
|
| 195 |
+
aud.registrar_ativos(3)
|
| 196 |
+
aud.marcar("recompute_ini")
|
| 197 |
+
ya.sum().backward()
|
| 198 |
+
aud.marcar("recompute_fim")
|
| 199 |
+
aud.amostrar()
|
| 200 |
+
aud.marcar("pilha_fim")
|
| 201 |
+
for _ in range(6):
|
| 202 |
+
aud.amostrar()
|
| 203 |
+
rel = aud.relatorio()
|
| 204 |
+
check("Teo 29.12 I1: não-vazamento", rel["I1_nao_vazamento"],
|
| 205 |
+
f"drift {rel['crescimento_mb']} MB")
|
| 206 |
+
check("I2: recompute ≤ fwd", rel["I2_recompute_le_fwd"])
|
| 207 |
+
check("I3: sem transiente de empilhamento", rel["I3_sem_transiente_pilha"])
|
| 208 |
+
check("I4: amostras de ativos", aud.n_amostras_ativos() >= 1)
|
| 209 |
+
aud2 = AuditoriaMemoriaSonic()
|
| 210 |
+
lixo = []
|
| 211 |
+
for _ in range(60):
|
| 212 |
+
# 4 MB/amostra por bytearray (malloc/mmap DIRETO) — o pool de cache da
|
| 213 |
+
# CPU do torch e as arenas do glibc absorvem alocações pequenas sem
|
| 214 |
+
# tocar páginas novas; 4 MB × 60 = 240 MB força o crescimento real
|
| 215 |
+
lixo.append(bytearray(4_000_000))
|
| 216 |
+
aud2.amostrar()
|
| 217 |
+
rel2 = aud2.relatorio()
|
| 218 |
+
check("I1 DETETA vazamento sintético (auditoria que reprova)",
|
| 219 |
+
not rel2["I1_nao_vazamento"], f"drift {rel2['crescimento_mb']} MB")
|
| 220 |
+
del lixo
|
| 221 |
|
| 222 |
+
# ==========================================================================
|
| 223 |
+
print("== 5) ReplaySSM — exatidão e integração (Teos 29.13/29.14)")
|
| 224 |
+
# ==========================================================================
|
| 225 |
+
from khtst.mamba3.kernel import varredura_trapezoidal
|
| 226 |
+
from khtst.mamba3.replay import varredura_replay
|
| 227 |
+
torch.manual_seed(7)
|
| 228 |
+
B, L, D, N, R = 2, 128, 16, 4, 2
|
| 229 |
+
u = torch.randn(B, L, D, requires_grad=True)
|
| 230 |
+
delta = torch.sigmoid(torch.randn(B, L, D))
|
| 231 |
+
A = -torch.exp(torch.rand(D, N))
|
| 232 |
+
Bm = torch.randn(B, L, R, N)
|
| 233 |
+
Cm = torch.randn(B, L, R, N)
|
| 234 |
+
of_, hf_ = varredura_trapezoidal(u, delta, A, Bm, Cm, devolve_estado=True)
|
| 235 |
+
or_, hr_ = varredura_replay(u, delta, A, Bm, Cm, 16, devolve_estado=True)
|
| 236 |
+
check("Teo 29.13: saída BIT-EXATA", torch.equal(of_, or_))
|
| 237 |
+
check("Teo 29.13: h_L BIT-EXATO", torch.equal(hf_, hr_))
|
| 238 |
+
u2 = u.detach().clone().requires_grad_(True)
|
| 239 |
+
u3 = u.detach().clone().requires_grad_(True)
|
| 240 |
+
g = torch.randn_like(of_)
|
| 241 |
+
varredura_trapezoidal(u2, delta, A, Bm, Cm).backward(g)
|
| 242 |
+
varredura_replay(u3, delta, A, Bm, Cm, 16).backward(g)
|
| 243 |
+
check("Teo 29.14: gradiente EXATO (erro 0.0)",
|
| 244 |
+
float((u2.grad - u3.grad).abs().max()) == 0.0)
|
| 245 |
+
check("nascimento neutro: S=1 ≡ kernel completo",
|
| 246 |
+
torch.equal(varredura_replay(u.detach(), delta, A, Bm, Cm, 1),
|
| 247 |
+
varredura_trapezoidal(u.detach(), delta, A, Bm, Cm)))
|
| 248 |
+
cfg3 = Config()
|
| 249 |
+
cfg3.modelo.mamba3["replay_segmento"] = 16
|
| 250 |
+
mod_r = KHTSTModel(cfg3)
|
| 251 |
+
segs = [getattr(m, "replay_segmento", 0) for m in mod_r.modules()
|
| 252 |
+
if type(m).__name__ == "Mamba3Camada"]
|
| 253 |
+
check("cfg mamba3.replay_segmento chega às camadas",
|
| 254 |
+
len(segs) > 0 and all(s == 16 for s in segs), f"{len(segs)} camadas")
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 255 |
|
| 256 |
+
# ==========================================================================
|
| 257 |
+
print("== 6) AUTOAJUSTE MULTI-MÉTODO (Teos 30.1–30.15)")
|
| 258 |
+
# ==========================================================================
|
| 259 |
+
from khtst.nucleo.autorregulador import (AutorregulacaoMulti, ControladorHibrido,
|
| 260 |
+
BissecaoTauQK)
|
| 261 |
+
multi = AutorregulacaoMulti({"ativo": True, "passos_espera": 0})
|
| 262 |
+
check("Teo 30.1: nascimento NEUTRO (todas escalas 1.0)",
|
| 263 |
+
all(abs(v - 1.0) < 1e-12 for v in multi.escalas_adamw().values())
|
| 264 |
+
and abs(multi.escala_moe() - 1.0) < 1e-12
|
| 265 |
+
and abs(multi.escala_som() - 1.0) < 1e-12)
|
| 266 |
+
for _ in range(80):
|
| 267 |
+
multi.atualizar({"cv2": 4.0, "tend_cv2": 1.0, "razao_upd": 3.0,
|
| 268 |
+
"norma_rel": 1.5, "razao_mom": 0.2, "trust": 0.2,
|
| 269 |
+
"velocidade": 0.01, "razao_moe": 0.9, "taxa_som": 0.5,
|
| 270 |
+
"folga_abmo": -0.3, "razao_cresc": 4.0})
|
| 271 |
+
tel_m = multi.telemetria()
|
| 272 |
+
check("Teo 30.3: β₂ desce com CV² alta", tel_m["multi/beta2"] < 1.0)
|
| 273 |
+
check("Teo 30.4: ε sobe com ‖Δp‖/‖g‖ alta", tel_m["multi/eps"] > 1.0)
|
| 274 |
+
check("Teo 30.5: wd sobe com norma relativa > alvo", tel_m["multi/wd"] > 1.0)
|
| 275 |
+
check("Teo 30.6: momentum sobe com razão baixa (loop NEGATIVO)",
|
| 276 |
+
tel_m["multi/momentum"] > 1.0)
|
| 277 |
+
check("Teo 30.8: trust baixo ⟹ η_up ↑ e η_down ↓",
|
| 278 |
+
tel_m["multi/eta_up"] > 1.0 and tel_m["multi/eta_down"] < 1.0)
|
| 279 |
+
check("Teo 30.11: PESO_MOE desce com razão > ½", tel_m["multi/peso_moe"] < 1.0)
|
| 280 |
+
check("Teo 30.12: PESO_SOM sobe com ativação < 90%", tel_m["multi/peso_som"] > 1.0)
|
| 281 |
+
check("Teo 30.13: margem ABMO desce com barreira ativa",
|
| 282 |
+
tel_m["multi/abmo_margem"] < 1.0)
|
| 283 |
+
check("Teo 30.15: crescimento segue razão expansão/poda > 1",
|
| 284 |
+
tel_m["multi/crescimento"] > 1.0)
|
| 285 |
+
tq = BissecaoTauQK((80.0, 120.0), 20.0, 240.0, 100.0, paciencia=1)
|
| 286 |
+
for lm in (200.0, 200.0, 200.0):
|
| 287 |
+
tq.registrar_lmax(lm)
|
| 288 |
+
t_antes = tq.tau
|
| 289 |
+
bracket_antes = tq.tau_max - tq.tau_min
|
| 290 |
+
for lm in (200.0, 200.0):
|
| 291 |
+
tq.registrar_lmax(lm)
|
| 292 |
+
check("Teo 30.7: bisseção contrai o intervalo 1/2 por ajuste",
|
| 293 |
+
tq.tau < t_antes and (tq.tau_max - tq.tau_min) < bracket_antes,
|
| 294 |
+
f"τ {t_antes:.0f} → {tq.tau:.1f}; intervalo {bracket_antes:.0f} → "
|
| 295 |
+
f"{tq.tau_max - tq.tau_min:.0f}")
|
| 296 |
+
ch = ControladorHibrido((0.5, 2.0), b=0.0, c_i=0.05)
|
| 297 |
+
for _ in range(500):
|
| 298 |
+
ch.atualizar(0.0, erro=1.0)
|
| 299 |
+
check("anti-windup: integral satura em 1", abs(ch.I - 1.0) < 1e-9)
|
| 300 |
+
ch2 = ControladorHibrido((0.5, 2.0), b=0.2, c_i=0.1, c_d=0.1)
|
| 301 |
+
thetas = [ch2.atualizar(1.0, erro=1.0) for _ in range(2000)]
|
| 302 |
+
check("Teo 30.1: truncagem ⟹ θ ∈ [θ_min, θ_max] SEMPRE",
|
| 303 |
+
min(thetas) >= 0.5 - 1e-12 and max(thetas) <= 2.0 + 1e-12)
|
| 304 |
+
# escalas chegam ao OTIMIZADOR (hiperparâmetros efetivos)
|
| 305 |
+
from khtst.treino.otimizadores import MuonClipAdamW
|
| 306 |
+
lin = torch.nn.Linear(8, 8)
|
| 307 |
+
opt_t = MuonClipAdamW(lin.parameters(), lr=1e-3)
|
| 308 |
+
opt_t.definir_escalas(beta1=0.9, beta2=1.05, eps=2.0, wd=0.5, momentum=1.04)
|
| 309 |
+
hp = opt_t.hiperparametros_efetivos()
|
| 310 |
+
check("Teo 30.2: hiperparâmetros EFETIVOS = base × escala (com clamps)",
|
| 311 |
+
abs(hp["beta1"] - 0.9 * 0.9) < 1e-9 and abs(hp["eps"] - 2e-8) < 1e-15
|
| 312 |
+
and hp["beta2"] <= 0.9999 and hp["momentum"] <= 0.999)
|
| 313 |
+
g0 = torch.randn(8, 8)
|
| 314 |
+
lin.weight.grad = g0.clone()
|
| 315 |
+
p0 = lin.weight.detach().clone()
|
| 316 |
+
opt_t.step()
|
| 317 |
+
hp2 = opt_t.hiperparametros_efetivos()
|
| 318 |
+
check("step consume os escalados (β/ε/wd/momentum vivos no passo)",
|
| 319 |
+
not torch.equal(p0, lin.weight.detach()))
|
| 320 |
|
| 321 |
+
# ==========================================================================
|
| 322 |
+
print("== 7) INTEGRAÇÃO NO TREINADOR (passos reais: jepa + multi + auditoria)")
|
| 323 |
+
# ==========================================================================
|
| 324 |
+
from khtst.treino.treinador import TreinadorExtensao
|
| 325 |
+
from khtst.dados.tokenizador import TokenizadorKHTST
|
| 326 |
+
from khtst.telemetria.hub import TelemetryHub
|
| 327 |
+
cfg2 = Config()
|
| 328 |
+
cfg2.treino.jepa["lambda_jepa"] = 0.1
|
| 329 |
+
cfg2.modelo.mamba3["replay_segmento"] = 8
|
| 330 |
+
with tempfile.TemporaryDirectory() as td:
|
| 331 |
+
hub = TelemetryHub(os.path.join(td, "tel.jsonl"))
|
| 332 |
+
modelo2 = KHTSTModel(cfg2)
|
| 333 |
+
tk = TokenizadorKHTST()
|
| 334 |
+
regs = [{"tarefa": "lm", "texto": "texto de treino real sintético"},
|
| 335 |
+
{"tarefa": "matematica", "entrada": "quanto é 2+2?",
|
| 336 |
+
"saida": "quatro (4)"},
|
| 337 |
+
{"tarefa": "robo", "entrada": "S:0.1,0.2,0.3",
|
| 338 |
+
"saida": "A:0.4,0.5,0.6"}]
|
| 339 |
+
textos = []
|
| 340 |
+
for r in regs:
|
| 341 |
+
textos += [r.get("texto") or "", r.get("entrada") or "",
|
| 342 |
+
r.get("saida") or ""] * 5
|
| 343 |
+
tk.treinar(iter(textos), vocab=600)
|
| 344 |
+
tr = TreinadorExtensao(cfg2, modelo2, tk, hub, regs)
|
| 345 |
+
check("motor JEPA ativo com λ=0.1", tr.motor_jepa is not None
|
| 346 |
+
and tr.motor_jepa.cfg.normalizar_dim)
|
| 347 |
+
check("auditoria Sonic instanciada (Cython)", tr.auditoria_sonic is not None)
|
| 348 |
+
check("multi ativo e nascido neutro",
|
| 349 |
+
tr.multi.ativo and abs(tr.multi.escala_moe() - 1.0) < 1e-12)
|
| 350 |
+
for _ in range(24):
|
| 351 |
+
tr._passo()
|
| 352 |
+
tel_multi = tr.multi.telemetria()
|
| 353 |
+
check("telemetria multi/* completa (15 controladores)",
|
| 354 |
+
len([k for k in tel_multi if k.startswith("multi/")]) >= 15)
|
| 355 |
+
check("escalas dentro dos limites de projeto",
|
| 356 |
+
all(0.25 <= v <= 4.0 for k, v in tel_multi.items()
|
| 357 |
+
if k not in ("multi/tau_qk", "multi/lmax")),
|
| 358 |
+
f"β₂={tel_multi['multi/beta2']}")
|
| 359 |
+
check("τ_qk da bisseção na faixa [τ_min, τ_max]",
|
| 360 |
+
20.0 <= tel_multi["multi/tau_qk"] <= 240.0)
|
| 361 |
+
tel_j = getattr(tr, "ultima_tel_jepa", {})
|
| 362 |
+
check("TCE normalizada no treino real (perda O(1), não ~25)",
|
| 363 |
+
tel_j.get("jepa/perda", 1e9) < 10.0,
|
| 364 |
+
f"perda {tel_j.get('jepa/perda', float('nan')):.3f}")
|
| 365 |
+
check("taxa de sucesso mensurável (limiar invariante)",
|
| 366 |
+
tel_j.get("jepa/taxa_sucesso", -1.0) >= 0.0)
|
| 367 |
+
rel_t = tr.auditoria_sonic.relatorio()
|
| 368 |
+
check("I1–I4 VERDES em passos reais",
|
| 369 |
+
rel_t["I1_nao_vazamento"] and rel_t["I2_recompute_le_fwd"]
|
| 370 |
+
and rel_t["I3_sem_transiente_pilha"] and rel_t["I4_somente_ativos"])
|
| 371 |
+
raz = tr._razoes_aux
|
| 372 |
+
check("razão PONDERADA do jepa registrada (Teo 29.8)",
|
| 373 |
+
raz.get("jepa") is None or (0 <= raz["jepa"] <= 5.0),
|
| 374 |
+
f"razão {raz.get('jepa')}")
|
| 375 |
|
| 376 |
+
# ==========================================================================
|
| 377 |
+
print("== 8) REGRESSÃO SEMÂNTICA v16 ⟹ v17 (nascimentos neutros)")
|
| 378 |
+
# ==========================================================================
|
| 379 |
+
cfg_n = Config()
|
| 380 |
+
check("replay nasce DESLIGADO (0 = v12 exato)",
|
| 381 |
+
int(cfg_n.modelo.mamba3.get("replay_segmento", 0)) == 0)
|
| 382 |
+
check("multi nasce NEUTRO", all(
|
| 383 |
+
abs(v - 1.0) < 1e-12 for v in AutorregulacaoMulti({}).escalas_adamw().values()))
|
| 384 |
+
check("τ_qk bisseção nasce em τ0 de projeto",
|
| 385 |
+
BissecaoTauQK().tau == 100.0)
|
| 386 |
+
check("empacotador nasce AUSENTE (forward v16 intacto)",
|
| 387 |
+
getattr(MoEAgrupavel(32, n_grupos=1, experts_por_grupo=1, d_ff_expert=32,
|
| 388 |
+
n_tarefas=2), "_pools_quant", None) is None)
|
| 389 |
|
| 390 |
+
print("=" * 72)
|
| 391 |
+
if FALHOU:
|
| 392 |
+
print(f"SUÍTE v17: FALHAS ACIMA ✗ ({VERDE} verdes)")
|
| 393 |
+
sys.exit(1)
|
| 394 |
+
print(f"SUÍTE v17: TODAS VERDES ✓ ({VERDE} verifica��ões)")
|