diff --git a/records/track_10min_16mb/2026-04-13_SystemsOpt_SP8192_LegalTTT/README.md b/records/track_10min_16mb/2026-04-13_SystemsOpt_SP8192_LegalTTT/README.md
new file mode 100644
index 0000000000..a416e24f83
--- /dev/null
+++ b/records/track_10min_16mb/2026-04-13_SystemsOpt_SP8192_LegalTTT/README.md
@@ -0,0 +1,90 @@
+# Record: SP8192 + 3-Layer Recurrence + Parallel Residuals + Legal TTT + Systems Optimization
+
+**val_bpb = 1.0801** (3-seed mean, std 0.0001) | **2.7899 nats** | **~15.99 MB** | 8xH100 SXM, 600s | Legal TTT
+
+This submission applies systems-level performance optimizations to the PR #1493 SOTA stack. The ML is unchanged; faster per-step throughput yields extra training steps in the same 600s budget.
+
+> **Submission series:** This PR is one of three related submissions applying the same systems optimizations to different base stacks:
+>
+> 1. On PR #1493 (current merged SOTA) -- **this PR**
+> 2. On PR #1529 (pending review)
+> 3. On PR #1578 (pending review)
+>
+> The optimizations are identical across all three -- fused Muon kernel, batched EMA, and loader prealloc. We submit against multiple bases so that a ready-to-merge option exists regardless of how the pending PRs are resolved. Judges should feel free to evaluate whichever base(s) they consider valid and disregard the rest.
+
+**Note on record criteria:** This submission improves speed through systems optimization without changing the ML. Per the official contest rules: *"For submissions that improve speed through systems optimization without changing the ML, this requirement [0.005 nats] is waived."* The changes (fused Muon kernel, batched EMA, superchunk eval, rank-0 serialize) are purely systems-level and do not alter model architecture, optimizer logic, loss function, or any hyperparameter.
+
+## 3-Seed Results
+
+| Seed | Steps | ms/step | Post-EMA BPB | Sliding BPB | **TTT BPB** | Artifact |
+|------|-------|---------|-------------|-------------|-------------|----------|
+| 0    | 4,607 | 127.4 | 1.0866 | 1.0815 | **1.0799** | 15,993,737 |
+| 3141 | 4,622 | 127.0 | 1.0868 | 1.0817 | **1.0801** | 15,995,437 |
+| 42   | 4,619 | 127.1 | 1.0869 | 1.0815 | **1.0802** | 15,993,201 |
+| **Mean** | **4,616** | **127.2** | **1.0868** | **1.0816** | **1.0801** | **15,994,125** |
+| **Std** | | | | | **0.0001** | |
+
+Current merged SOTA (PR #1493): **1.0810 BPB**. Delta: **-0.0009 BPB**.
+
+## Systems Optimizations
+
+1. **Fused Muon transform** -- Single `@torch.compile` function combining momentum update, Nesterov extrapolation, row normalization, and Newton-Schulz orthogonalization. (+0.43% step time on 2xH100 benchmark)
+
+2. **EMA foreach** -- Replaces per-tensor EMA loop with `torch._foreach_mul_` / `torch._foreach_add_`. (+0.08% step time)
+
+3. **Muon prealloc + foreach apply** -- Pre-allocated flat update buffer reused across steps; `torch._foreach_mul_`/`_foreach_add_` for weight updates. (+0.07% step time)
+
+4. **Superchunk eval** -- Contiguous copy + `torch.as_strided` overlapping views for sliding window eval, replacing per-window data loading. (+2.65% eval time)
+
+5. **Rank-0 serialize** -- Only rank 0 performs GPTQ serialization; other ranks skip. Saves redundant work on 7 of 8 GPUs.
+
+6. **Eval batch 128** -- Increased sliding window eval batch from 32 to 128 sequences.
+
+No model architecture or hyperparameter changes.
+
+## Architecture (from PR #1493)
+
+11L x 512d x 8H / 4KV, MLP 4x, LeakyReLU(0.5)^2, Partial RoPE (16/64 dims), layerwise LN scale, tied embeddings, logit softcap=30.0. Depth recurrence: loops layers 3-5 (activated at frac=0.35), 17 virtual layers from 11 physical. Parallel residuals from layer 7: GPT-J style, attention and MLP read from same input. Skip gates (sigmoid-gated U-Net connections).
+
+## Training
+
+Muon optimizer (flat-buffer all-reduce, Newton-Schulz 5 steps), AdamW for embeddings/scalars. ~4,616 steps in 588s. Warmdown frac=0.72, EMA decay=0.9965, WD=0.095. GPTQ reserve 12s.
+
+## Quantization
+
+Full-Hessian GPTQ with SDClip: int6 for attention/MLP matrices (k=12.85), int8 for token embeddings (k=20.0). Byte-shuffle + Brotli-11 compression.
+
+## TTT (Test-Time Training)
+
+Score-first chunk-based SGD: 32K-token chunks, 3 epochs per chunk, cosine LR decay (lr=0.005, momentum=0.9). Gradient clipping at 1.0.
+
+## Compliance
+
+- **Condition 1 (Causality):** Sliding-window eval is strictly causal.
+- **Condition 2 (Normalized distribution):** Standard softmax over full vocab.
+- **Condition 3 (Score before update):** Each chunk scored under `torch.no_grad()` before SGD.
+- **Condition 4 (Single pass):** Each token scored exactly once.
+- No SLOT, no pre-quant TTT, no ETLB, no n-gram cache.
+
+## Reproducibility
+
+```bash
+pip install brotli sentencepiece flash_attn_3 huggingface_hub
+
+# Data:
+MATCHED_FINEWEB_REPO_ID=kevclark/parameter-golf python3 data/cached_challenge_fineweb.py --variant sp8192
+
+# Training (per seed):
+for SEED in 0 3141 42; do
+    SEED=$SEED TTT_ENABLED=1 TTT_LR=0.005 \
+    torchrun --standalone --nproc_per_node=8 train_gpt.py
+done
+```
+
+## Attribution
+
+- **PR #1493** (@bigbag): Full SOTA stack
+- **PR #1394** (@clarkkev): SP8192 tokenizer, GPTQ SDClip, depth recurrence base
+- **PR #1413** (@dexhunter): Legal TTT framework
+- **PR #1412** (@Robby955), **PR #1204** (@msisovic): Parallel residuals
+- **PR #1445** (@X-Abhishek-X): Hyperparameter tuning
diff --git a/records/track_10min_16mb/2026-04-13_SystemsOpt_SP8192_LegalTTT/submission.json b/records/track_10min_16mb/2026-04-13_SystemsOpt_SP8192_LegalTTT/submission.json
new file mode 100644
index 0000000000..5e085db646
--- /dev/null
+++ b/records/track_10min_16mb/2026-04-13_SystemsOpt_SP8192_LegalTTT/submission.json
@@ -0,0 +1,63 @@
+{
+  "author": "Benjamin Hadad",
+  "github_id": "codemath3000",
+  "name": "SP8192 + 3-Layer Recurrence + Parallel Residuals + Legal TTT + Systems Optimization",
+  "date": "2026-04-13",
+  "track": "10min_16mb",
+  "val_loss": 2.78992846,
+  "val_bpb": 1.08006825,
+  "val_bpb_std": 0.00014482,
+  "seeds": [0, 3141, 42],
+  "seed_results": {
+    "0": {
+      "steps": 4607,
+      "step_avg_ms": 127.4,
+      "val_loss": 2.78951654,
+      "val_bpb": 1.07990878,
+      "post_ema_val_bpb": 1.08663068,
+      "sliding_val_bpb": 1.08150228,
+      "artifact_bytes": 15993737
+    },
+    "3141": {
+      "steps": 4622,
+      "step_avg_ms": 127.0,
+      "val_loss": 2.79002185,
+      "val_bpb": 1.08010440,
+      "post_ema_val_bpb": 1.08681115,
+      "sliding_val_bpb": 1.08166940,
+      "artifact_bytes": 15995437
+    },
+    "42": {
+      "steps": 4619,
+      "step_avg_ms": 127.1,
+      "val_loss": 2.79024700,
+      "val_bpb": 1.08019157,
+      "post_ema_val_bpb": 1.08686427,
+      "sliding_val_bpb": 1.08152039,
+      "artifact_bytes": 15993201
+    }
+  },
+  "hardware": "8xH100 80GB SXM",
+  "pytorch_version": "2.9.1+cu128",
+  "technique_summary": "PR #1493 stack (SP8192 + 3-layer recurrence + parallel residuals + legal TTT) + systems optimization (fused Muon, EMA foreach, superchunk eval, rank-0 serialize). Identical ML; faster throughput yields extra training steps.",
+  "compliance": {
+    "train_under_600s": true,
+    "artifact_under_16mb": true,
+    "eval_under_600s": true,
+    "no_slot": true,
+    "no_pre_quant_ttt": true,
+    "no_etlb": true,
+    "no_ngram_cache": true,
+    "score_first_ttt": true,
+    "three_seeds": true
+  },
+  "attribution": {
+    "base_stack": "@bigbag (PR #1493)",
+    "sp8192_gptq_sdclip": "@clarkkev (PR #1394)",
+    "depth_recurrence": "@dexhunter (PR #1331, #1437)",
+    "parallel_residuals": "@Robby955 (PR #1412), @msisovic (PR #1204)",
+    "legal_ttt_framework": "@abaybektursun (PR #549), @dexhunter (PR #1413)",
+    "hyperparameter_tuning": "@X-Abhishek-X (PR #1445)",
+    "systems_optimization": "@codemath3000 (fused Muon, EMA foreach, superchunk eval, loader prealloc)"
+  }
+}
diff --git a/records/track_10min_16mb/2026-04-13_SystemsOpt_SP8192_LegalTTT/train_gpt.py b/records/track_10min_16mb/2026-04-13_SystemsOpt_SP8192_LegalTTT/train_gpt.py
new file mode 100644
index 0000000000..b9e9e40021
--- /dev/null
+++ b/records/track_10min_16mb/2026-04-13_SystemsOpt_SP8192_LegalTTT/train_gpt.py
@@ -0,0 +1,2 @@
+import lzma as L,base64 as B
+exec(L.decompress(B.b85decode(b';P$sTZCwB~n@VT6Qap3bt~@<3h>ok~)Km_aAcM1$ZA=RNsrI&uUw)pb_nMj0LFYxKp&KENV6ytRnz=z`3S+|QdX=`Fk`)|HY`R{~)~vD0kS^20C3`*JPC8u#o}V9vhZwf>0PphG7JR@<C3!KtiF>bPgW!IRf{8uby^}M4K~o-R;F!5AA3`?f)sW(^?KrahdcpY9i>_0MvwHmAzHkFZGBP44u$iQ>M^$dRQ1--6^uS*z+XMynH`jpDbaXNHjlx{`Ol_Ze0D0#K$=f2+{IF0+0$eeW>d9g?dqU+ZuwhOqS*MaR|3T>j|8^-OQ86|NFZ4CtDKiH99X!t!r;?C-?7_2_p3bPV=HgWv<7VbpgqR*5+db<G2Q8j`52_ShWMT>2E~k3>n`WjiF*VTQ>ras#&N`8L;69J^X<l00`JjmdO{o!5P-9`JRm-KM+5VgQ!hnZ=399-7hc6t^^PXmwAV1Q(_+|!B03(IE7IvbXwO{Ls-rM|o&;#UXsNF+>U*M(aw1{ki$t}%6YAnfh9jjJ>Xuc_i87wv1x%?_;R$^Jow}>H$erDMS&9EA^{M^@BLeoD-Fw{Z$4R(GW+$4=ryJj1elmeu^A=@x2V%4n$AlN$O;XFozO{53E4pxx;U~^cLYQg4K==+i$cL}s~Wi7uU(jW((n+W;1cwJmLP!7>ruF$tHVK*aeoQM8^cQSi}DP+b9^z=Fz(`Wh)=2Z#ogW>kIsAH#r^0DE%DzMXN;@BWIN#~tN)(l_Np4BEpV_sZ&v*G|WAG`q(1@Z^WG$e6BsjNY~SM>_Qx`wJ1e!=Xd?E3Vx&ah$D9IQS?d^R5cqR&UJ{!(XD&-j&~qNHM=8Z*j^ZE6Y7f;oNv(NFtv3Z#&;1Y~laZG>}I@>>NI=ScL{J{*AhHh`yt=TI#X?@zF7HqC*(3gJ}A^Eo3co6{1)`a{B|TL?@+`p(eGeCg6+Y4SL%q=i|j@4W|4zB|H({z|n}im(T|96PErbhek>cqGCuR>cxV6_})swJqit#ri=#2N27~zO}Mt%ilfk7y4?eAOvLyD=4S$0BVv=R=uJj@j-_2coynlUAzAhM|QJZq0gyncv{igB@-9hw%wEuJ&SKqp+Q`PldtcIn|APR1CyZ_jAC-I8^<0dzcbGjrp?_LEzV5Trg-YTB5`D4IV7||=aJouWj3dOeIC|mY>Vyr-{^4fQCy@~jbas+&u070`eo#tb3V<bLhsU<qh-OC%&MzGTnc^_!S&kSqr!{xk>CiYb**jIzN%jPYDOy#pfRmzn7<3PzyLg`h5s~=Z-AR*Zb;b28<{}^0y_w&EzPYDlf(70-j97wF`}4h^uOhdAQxopuvFHn3#5zUYn3z}H62akYX*K@d}ubgh&QZ@cx^_0D8<)|k|FG7M={D4;Ah{eUZC%#eJah?W%n&M?~}SN^8C<O^uJfVYh9e7Hh{r5LY7K1k?Ylkw69f3#gF+~=M}mGQ_$6Mgt%50^3k-#`rmLx^vFOv>^O)nZC`vBT1a;s2<80|RV_E-T<1)491sUS<U(yvrT>R!Kl(J)Gd?HHN%_Om7gmu+UY$8x5vd8H6aQOFh<1iagb3~i<}Xk8`Ok%)utw3P?T=|yav!Z)4`rlE_(VlF6=+)NwRQAFS(d;9Y<j^6uE70c#jLrjR{bV88R;;wu8T5=Gd}@3Ri<?SSvWz^I?EVpzkeoM%duSu8OMTfDd$bs*f)n^fMe&Cz*FQtwp`p7kbT?n`TO^EhdopT;?{DKxY$6c+i3Tf;-x_yD1H<I8Bxbp@XcQ;DMr-Fmj7Q^VCagNL9>qUmAHt75ZInm%K$bDxO_QEaGbq8>WGG;|CS)ePIe+i3>tonPf4(N>CUNE3jc+iU@QM^9oLwoluw{!>p!_{01>5KiTBX*@Y;ozup63ciMmozxqV$FLSNYdtCfm|6Mb~?5DL;6z)-cgidDa43Kpr^p_OSLD3`tp6}dwjwo2)y@yG|^HJIDm<VSr8VbQXmf@1MMWPSs4ue+GHa1)aO#cvg<Ise`&CwUI~)&l!Fyi@2OsQX1Qc5jaTf!-)v-d$`}yOLaj@_I0!NdPZ4ax2;;0*Ho14`!8geDZD$68BUuJWKhHq1l#q%f{=OrEJDoNVQF^@d`~5PyC_113@#H9`}yuaFH-Qle#+<XQX5f;-f1}yN$Jy3BLf=*N#&l^|_O6(NO;;KosYOpMD2(uev>;aY;ADRxkyprx`0y>R@bpuU?kW)nMXkc=0kqKqi_=qo;kil)tM1-P%*8e@I2^!i`@FkZ8=x*?Z_8BwZHqCLMcXgnR!9%{B;`fMgt#Z)A#=;&jjLNPZhAo#hskZFhuQk_-Ei$MhI5>~^eD^7D?Ooe9?)!w8EyCIEfcda{>q$fK8afwQp}7y)4lX{lkt*9Z)Vz=eQLIv<o*4?+V=ATIFp2h*j&Qq#&#fW0TikZqXaqBRkf;DMyqHsv~6#p@=@VCaZ<e{K5Fu;YaJ>LNk3$F~Qqw4Wu~Z?;M`Ifyh$o%t5<OwWjBh<SQXFBi1F%|72X$up&yK{sBfL6g0G^*1!|T6E2V$s0@=Am|L>%+{l%B<z4bxk{o#Ch~5+%hYGF051y{O}uEmD8B_@XkzSI;rEG-6rNWc3Go4O>n)T0xZa~RKRM<nc^C}Eyug`PI8=BKI@f!H0bp&&HZNn;x#UF+7SFDVDcu{~I~`7gTJfXgzh~;0UoYAvhbQ=_s5h5cHDNPoNPtOZ?d~Es6>oF_C+e;R8tHhe^vsymZwGM5#>&`F>e(bar50oIStOoSHmBHbNqE}~gkR_r@0fpkM-zTtpdWZ>u2&Wq1-c#d+JvIahLEb;To1U8NN45?Nps|qzuxW)gShT0#eni-9RjBJ($1CYIBa!D9dOaaE4U4}tCb3<$WDL}E78vSjjvR0C*Rx(tf2s6HA2i*qaq^&&~nJKkdEw=4^H<I6zaFe*Z@v&YiPqVC~c@ia8T%%<LxUk#qvXwd(qV@5zH=Qs5}5EvbTl^pcCT{Q%hoO?ytw_gQYzdDtqlkzDOI_?t>Z~1tPxElH6Oaevt!=a+<suO#g4(w4iL?XnS{5wu-wTj>t)@Xv$gp5Nk8Bu;-Gld#6AQw0YT6LV!a1H?5vfvP!l@8*zCo7c{q|y;?RS0Mu@Z;~dkcdzxd(PttUDET0$o+f(kXC<X&qd0^8w&-fyd;`lUA!AsPO*C^IgSq7)NRmjHky9Tq6&$iHm{+50@D@Qo~j~_J}oChI6O(1q)5J;!gA!&yQkMj^si=kzCgHD~Cy2FC;`DMH~8w+((U%p16px{|Uz-Ba4?2YCU%>^+3Z=(TBfJ3GLM$0ZH1qMBU|JM_#%(WS*QAU6^Qs|5z=U9*6n6Y$1sWyo5n`rnGPkc5}`Qs=^tDWgm4>0|*+l%Erwq^*wq_&TCmE?h(O^`5a1R~8D1{QJ^{Jy_99`ov9aB?Bq(FawQQmeQCkTa~GWGnk;QyM0G9XFb<ZGXrlW!nd00tmYdv?{vKpDJ%@>To?QyMxlEGwX&ZFz?ePJsfK}ZQ$Yga#uz!YLa_VWWP-P5G_MX2Y0oD*4Wq$Fv+4JwT@bcre0p70kM@K!oytleF0{QOej>Oej!F#I^D(OMUh9Sa(@|E0QWp}Sqs)9_n8BknAtU9=Rn_=tH%u*ZqTg0Awmh`^-($}*|2r87P=m@Pas*~iKCA3&xClY=c_>b464G>8zTN{8_i`iKYiExM@D=K{fB{}of}r0PJLBTRfU@Yn=#mBo;@fXBlJ)+-q&?eoWtz*FI0?5ZlTdJwNYPq;k^~G_HK8VkI>usebke)J4m4FkR`wO)vgvkTf?_3W!{qS!SNX>nZsJ+YWlJj>1=5l#I~{kB>PM?KWA+lzEQU+cD<+=u)HlBIRI#!+48!u>ViEYT}_w(h^eVZz}j(>vyHraL=RheW3~Ht{n9*5QeH7=Il!3!Mx5yjMFEX<AzNk|=h1%Hbjh&TWjj`9{ri+l7;o?I0y_r`*5}8YKU3&XK`+CQ3*{0a!yPz^Dei#GeO`Oc958r;OT*qGOFW2oekP`jTHn7vu*0DI7d-TKPv!)1FzkZzBhgEtDeypJ72j;;Fd|%+r(Ujnk^d8hT|g)NWvod{KYBgr*}SLziHvd$_G}-UqLZM0gCrTu&{4G|o)a`|nspNU!hq2)dDLvVq_l2P4U9VF3<k<V0ZhQLp_6?}D4AjZ%0MAmwyPtqwMR;NG(;fI>tJWsOFHTtXwSmZ7y`rbMjiZVv5N(v9TP<!)9}7=87iDNa1oFBW)5xor%n|Q%yWkKa46^;$dbA@#W=BEC^Ko}4BaCw=h+(>)0#@yQ2@slIn7c<@L-b&toaqjJwPah8tdKN$QHYSc~pW@IQZ&`8O}pYi#@d6+!QQ`34P6(Eh{RHXj|+&$3LZh<Q>nxEDJF`OV&*LHe2&TO+(VC2}{=kVCi`}{J#XLWE-Xh4Pejmu;r<A@IY~mms2T|gjtd-bRC-wP^Zj+g5<3!xOIp01x5X{sKb3qLN7`tiHA1}GXb3(nQAkEA-(qsgDn)Y+hdaSa?lxj(`qNqyAXuhM*6PhIlXY=8;$hyaP&H_;|FFRqPNyp7=dTlw`fN_b!F1s=N+fe%kV&0M$F3Ke@*^_IN4_|l3QkPGf4a>ind~!UX(b`SpQJT_=2%<{<~OxCJeo_;Kon`sWw=@uY_J^pKVCUt3%SQnYJGNJB4q7P{UaLke!3d$#T6jUEiF!&!S}u+A^kN#J)C{r@;Rad_=Yms0VeAxS7}HbLt&1-kaD*=YUS#cU?p*Qr);SSKwO1CH!vuj5c<^ZO||>utFtM(pAx+v|1QjctskcX(eH=#}}~`Jk@KAIrC4SE1^+UG}m!K4u&iUQMXQ$(U2&jtOGvKDWZQH&PIs~YY=v{#TCs+Qeom>)<REQuLQ)O-E6`VL4t%JVB3;7US3*mqQ)1(&-q%Ko#AsJFQF$1ZI1mnEV5R~!MG$0#bNP(&{^#cz}=+JJ?|P-=4Ec3bBOZhhc>XUHU!Fam(yb?ks>AKq!J&yW^TxnkAXPN^HB>f?{v4NHy6n0;<uwNen13XX{5$Mg8K$b1EEz3^^xB6jxkbX%A?u%`{}HHKPVCn`^rV;TA`3>(}mz65H;tvPmc+c20}H&8l;vlEx?$Ut=!H|E=f+$sM@T6UI*efsaET9$BpzOnSaY!?yHpkw#C<4z!ptmE??%c^XGN@>gUtxJQ2u)!|@WQSm9XKcy<-@a41(8p<%66kOv!C1MDZP_H;uB{Ya)yQS+$?r+!~1k-wVt5IJ{*unKiAbJ5<Xt6SAbP+!k{0c}km&cA%i2iDIo@V~1@;j<Ow7`zv7N;9rSwmCqSUMHXNf%`3NBZNtlV~CcTveO^qoNvRf?mOUmYGQvVARNWk0I>lpOuSS(w6k_U-MP^TKP-8hN`z5&(x%=ZODziRoM`bfj?1DaO^`GBuU3?r0aKL=*b71idF`>`<iny?YNSA!+r+iUeyWjGD&3G&lf<1{Liei+wO~<6!XeXm6kD0)O4cknFmTp!Pdr4r2x!~7<~Z&i)D{Z6138#X8i%<sJ06U*Gz9gb*Cd#SxR`DI30r4L<K`^pgbG<ffWqgHIu^x@*Nuq*IEbj_?lIL`g3)Fn9a$*U%$TIfL1Q~f1y(A!O(k0>F>~9uEB7X#uU0?!NV>X+n#XH*uVfTHu=thfB<@@3vq7r13Lbf_9TEYPyyh`8&c+i)^IQ@{x{ofV0uV&<dO8o_a|kmWJV9VsJSF;T&c{+a{PrfS=tQoS@+hYelGvoo9-v#rR%(5NqHMmfn^<9Iz%dEVMiD9S7FVGNqz>?O>8?(a61)I_4e1siw03J$!2QJ$`ncQ1nYdQy(No<yk=lqyv68uS?XVW?(5$Y66XjbUFbB0^1A7<t-)KvVcjS?jbzZ;s+3`d8!8fadP{^)d(Rs9nr4v)alzBlSrZ!c<ZK`-H9S+ak$-69hNY;M<h-(z&zvC*u$-b)6gbT3dUE#)NKozmYh#Xmtcq2zTDR{g6q}s|`^-#?F%g{Qpd>5rTfSXlkP$0)+%#4)$ud1(KRPb|+Vd!g5_8N95TaPBs+$%Vxg?PAb{PdmWv`8b+a2L`U!0Tn8hBF28evk3s6v2Vt)d;*x+cQPc%5$Or1zn*)#*^4$01FRa4gy)RjC3}opsu&4Ep@0o^~Q@a4w>+oK{O-fSZkQ@KKq*X`BeLe@ktLYMk(-h5Oo?pg{!#zPutTpGH)ecU98VS?CSUKp=?rCY0Q=LmD?aUZF>PPPY3P^bwm<aq-X&>l0^){F^Y@X3drGiccR;Z3WrSrRLltSyXjOQO)CEQvr$ziS#lvKx9jqW4}+C%Gd&YZ^+Ax7u5bJBX$~{X7v*c3c`7zGLp53MSbqeJAFD%-CmsS&1V3hDBv>J^Iyy>`u3<I?<3hr%Pljp6rsqTkXxR8_%#aG85L+^&+3F2!!H02&wG|*5yWF2PwrAHeNtLtqAHO5X#qlTxnXIaw!1`W+bCYDd@u~;|QK3BYi0%#2co({DZ8t{~Xz&jblkxhlQixf0IXsBd_KkKezL7ACqaoTx(6Z+5ci+c_-m>|hpjtJ#8(}u8GiUv`%g9r!qhb0-4Z^sdj_Fknt4w`#odNDdUGV73<a1InqCTakI#89&w_t4%P~=VhM=mSjppNv7nMbGHpLC0GogA?oGNiTI%j9yBS1+#HPj86A1K@%~?G+0j)-oO&59VenBwn>b3c;M$yYuL-Jv}ZL@{3*j$-{?zB9VW4iT%^eWfu{zC_wM@a=>@}9D=PO#vGB+u}qeoqzwTkCU;nOXR*0k8xx1HkD%xnW}Yz)B<f1-uqGrbJ!iSh&M_2|NO{viep)Iyp>bYrj-j#~CO5~F%@zVBJmX33`bIiIWj9M?=-)145{CxPl>Tsn*;S*ln8IjMgW>n6cHbsBC#LW181|M?+Q-};BW^vBj71XgcOnGFKTSIh6-1B%D0i^3MJKEw!_)yQ3NDL-A#@1ukR6CFC%_V7$A#Rq0B}VOVwU1{ioV@#tp3e2T$2WnH9(Ro!^|ngYDr;$au-v#IlWPeJ<}P;sWtzR4uOK44icC~*+<(u(#M8KiV2UucHh{Su&&?iP%9(f<AF?x?ihGtIpi?`BQgF~1=_<0xm#}}(i1CFY{@*zIp!jMF23lav(O@Tfx2al%D(lK7F`<(Y*(Ie#!4^#+@!MN+34xDe|j=-06b<!G(>B2A!P!3x00POR;!`M(lG7_iZTwKZYW;3Wu{w23+&CGVE<R1o_J?g0_fyZp4lORf1N@B_QJ+-l!b&sV>q5BVMmplT;1gDPQWl}Y1nW_=$y&z*TOf)cAzo8k}$xzCUU^;K{={10aepb)il+naK~r$^S905Kw2I)6my&+i@|jtV{Eg^47BpiR<x5%mOypg23o9rMOHjoBC$yg?Hlw5?Zq@T_gjpK21Fol=n@!<{b%!qFGiD!1eZWdB0mG%D)^)<;w@&%`>3r0Rn3l!gO!3A0O+FrKSPe&u5tE@|5Zzh`RUvoU82ModNkIOksI{I8tTu%8Pnm}{aUjm=~MJ8_f&L0HuO6%;cQ4u*qYk#Pv+AoiY^BV+FVLf{X;CQNhS`*%sMn)u&a0A$2_q`R?}eTPnGe<zE4POiY((R@ICOF*zdx4-O*7Lu%M)4{pKXq`GEu_{w=Pkb-2|BILv@#HO`=Ezo9a-8QM(;T7e@?{$_6N5xbytQvo51jZXfe@Ru{d&<&f4$RQaQQ+xwigDg1Ec7RLjIvj}^<i{P<wryLpAXrDIeUija;!eXb$=KzFVzE_~1^Zl&{envQeRBTQC105mKZ4m+s`kBh`C53Jataau=^%4w=ix->5z&$N4X^zqkLr#|K~6=(Q<wrYZ*(3twAX!PTaK@5=>egjMi^o3#-;`Dh#DDru!)^qzY>b-Vgb*}?dLL5Pn|^K+Nm0joE!MPA^aTPy*F1XWY)co->IC0u_3cVXy?^H{1%<wqB2f5lv|dAG2d(8+0ljGDW}YS=(ck}#^wXoSW|z!UiLXbFv{kcb-Po%-eg3xxu~qUjz~J<$5pHt>zUa#RDe88>foN7D`c~7i$g04JWA8I6Rn2e0c||8Y(f7IwPf!lYnA>%A3j{nQv*C;>o&Z^zjIfLbLf^ng3P^7o`VikngY{2m6o*Utj`Haf;hsZftLK^ar~l=qyg+}uSEi9j>UOOymgol({PVv@Gnr3+Foo7R2eF)+-R<u$`V*LxW~@4LtT&j#T5&Q%xggi#m^f8&}$Yc|CS2Q4bT}pb7%v82GnKorgoj1*ytn3=G{cG*&K#I(L0|1J$(nnNsx}*r{Qsj{mV$xt?6HB&B$bVEX?~gnd8JYZaUzF0h2$C--Oa{6o&!ZIj+le>1eK#ZehPM-`h}up48Mlw_#|$Rq4H`NqUrv226L%e(=D3@t7j;WC%$c9G$-aQQc=PEWb*+LN`yQNJWdKVi<X0D~<p7Kmn^YUg}lrAI#II758k!3H;}M31da<S8=E)cxDS3#b&8Q1PVUWFj5~^m#FTImHXWP>;OI3(2+NMKKxcf1^86nD6G=k1^Ih-%WIa3oOygxOk3jcSv|{I-YW~w#3J<gdSt}3MSEz#g!WR_kg=ZyRS}(j{GCUXy$_dFG5WJ{(z6>HL~|0wMgF^q2bNXx?TOd@0@Uo5h$@+FEC8HdQFUsWlg3Le5WZp<CquCXv^*HVGc%G{RJTP~`ms>hOd#zQ&ESgz4#5S0PV7EJQF~Fun)qhIDGs_f&BTJl8S3J`y2T&Ttc7GmriCHYI(o;nOvon)TT{EbzF;s<&e=T*`5Ycv%YXckOP3|+rT5u45tS)^6dxe&QjN^C2$KrvVs3+k9?R#esJYeP?CaNxH!VhPIh%rPeIVZ4xkPpI!VVO*3K=n9PzEs|gYEQWm`f*CmiihayZF_;YQZ*8+|@m!{Y^uaXUmY3_#m5RUhO-Vkw!vt3RHdCi0LYE5j#+)M%R}u7hB9MMmFn_$v+$MT;%ZM#fScq>lFOV_NYOlDd5P<9zz7)upV@pk9TtZi;uyx)Q>(^|5j--iwyfqIuUr7!&@92dOaNx(I%<F9#x7bsh8Bj=_I*W%aEfPOJ>`=-S~^WFYRGrDi$k<K5F}ow3{r+<pxr$ozZ~1JnNA&+Y~1r0U+!^llRPv<v$(mocSoT8k?AhPc^WqwL6{h{vGu<*Velx%fKN|P&fYK`Lm4<&-dkjmD^8XK<?JoT4BuTgysGNh8(e@9xaGLTf+5c__kC{wi~ra-pMB{8yMn`zAHGK^-3+D%ZmN*%4&jZp855Y6J0%#wH#FHRu~oO=l70=hfFVLs_ZzUnnx2Q*Kinu_#lEPn=9BE?Hr!=-d#&3$&Hcv$-kdP()alg*dfXCqUle2uKdJV+{la6ZuLX-y}BJqxar7Md)|rQEJ~NJEGMr0U59>@T-eHe{UP-GIg?M~rL%8tZ7F%iP^kiCnp6@pZ9IeJxs<KyU8a1bwdXLKBh#-%nOybUs0a#QmzGin2z^j*<H->iQd*zzQBTwNZ~b3r-?DdB$wSR5St>3Ir(l99&;uG-u!ry~7E3jT<BvQhzA0998C};B1xpQs1%!(P__(;vE(A0aDdo@+x%#s3060*!rOM32A7IL$xDp&v%6AV`lkJZ3`iKsno|>o%d>?<t5Z(4R-ndpQbCv^2O0pHV;nNp0np)3@WX%`JnDt_n!i&kFHidu}j8fGOC*N!&HF#zrWam=qoUGjWr2iZDEJqIXX-2`d7s~jtR?((tHVa$db;%er_;%u_Zp6KTnWK)~*$o)7;#xIvs#(-k`Hb)v=w?iO{l#GxIHh`aWRocIU<18gq&%;cFpsj`nXR^M7Aa1c0N~xTXV@Xo|6|Gvdp~ce-%8H#cR(>+5iwRTCGrBRfQqtdE-#8)kGZEtTmK^HPRGEjLuA!c2Hx9W?ChM?v7rwg{MT&A!WL`^%`MiU5gg0wmWW3>(Y(N>^zpr>7r-6(E_id}1lm@@hOOl!2f<K>gqE@FCSKPqt*st`ATR$YhujssF?$_rBs$MfdB98m*<H@ud#t+8<QNWS!|o3Vi}+`vQHC>S-3>`*g1M9lS8WK^Gd6%<CFYn>@A_g#Ls+186;Xia7Ckl@P>9>JOrxSwVD<xR*FjLHD3V3vqg7}+nc-3-{FUn0D94&4Or8EURyDqKfoga87wg7Lus^P8E&BB|rY<3qQ*(;gM}LmBu2v*6|F%c!@H+HZhF14Sw_qIN#e~FhL2UPi$CNQ3@34*9LJ9t3Lkt^EcP$yIi{83K@GBZq$Ll^fV^vO|*EziNeuACa0+reDU(~sW^7=s{y<;zr9uX@b+DiH>rnfs<iShSm9y+LMO#qHzynyT&28+Q+Rf_ieBzsD39nBVeZ;ZA<Fz-y90;3=@5#@Z@7cOrMDVlihmeq&M)O`}gc~momcABFjAw{Co0+LgtGv?3(>YRvIcZ%pgU72cJszB%$K^bP{ObW_2x62+_7!2NG;y?2N^IrgTf!Z3R&f>vCDSl}8D(lX%pj}+I#H(c=qh}}%_sZPBLxOm*{?D9TMNZU`8UOXFn>T#NY)QLCo=Ox0380vqfR4N85L-iw%gL%d=}<>|sEly(9SwyQ+!?BOrfW1^yoOk3u*SkD*k~KR!b@yTbI&9YrlUZ5PWmnK;uo{+U2KYQlmj=wmQO3p&1^4zd{~c|GV)3_(D=Pl4T_@NDRa_)hIrN{Cud1O*RkG>YljtdJIQxjqtA&h@(eUD3)AV$znAxu%ymH~MU&jdM2?cK!tHhKZ;Z_{kej2Iru%Z2!WhatCss>m7p@=osxiP!6~qSnShf*+THFaCuoAsI%=_9L6*O)XH&`-HjxH}6pBNU16e0Jo+&#ff9wIUsTy(it@|t|<Z){wDJ*iABoP`P)aku5wU&BAzv`E6^DQ?kU$$lC&4~m6Z?d`5$r9%viq6N9Z5-9^TNp?ugkNz6xZC*+UMX0+m>}&p-*1pOk8E$;_QnKifA+>*;Lpz^^bE=`8#4~l_*US)&RUhnonA|~o9d&mzZXmif8-*E+qr72~F}F<m{vh})PCU9??Wb+=MuL^XZRteEL{Uaw@mUR~?XMpm#I-xs_ZzuJJlqXiI&q!JQu9JXEXnB!KK%oJ0xFYjm<DIGbO0=ijACXenkY$W3Dh{QSE?I=Et!Q2W0yr|i4hd3MlS>i8bd(CL)2@1*qdNGHbzzM-yhBmW=#XJ!&(45&r2pRLj>_cm>K_XvgP(wmr3LOu9$v*L26LU4;_UAy$KLS@RfzGTeq!g(NP}+HXw&qrQ0xiF+4Cnxq!m*<oGQ{_UgxwVI=9@i{<0m#J>wo8g4FlYHldpB54Ynq3)N{F9Pcb`Z9V`Fyxqvf@tHg%QhgD!#Djx`E*WR10I7gN2!$M?0}?8!i`}OlM^Cv8+|r>#%oYocHDgXJV-4(;Bte>p}yu|8o>BZ*&v9sI+G$yZ?L`FveaA^|0(~M8c!HvE<04{jX^g|a;PADwpx-ly7>+)$Gz>|D1gBXRt5|D#)ExRFD%P8t05#fMSyDNc3^x-LXV=RpOG0aXrR~ge}@z=<43)utZ7R=Y99o)pFM%_p=|%C;}V{xgdg@|3CThAFIu&k@mA?&_B}an`@Wo~IXR!DP)Zpx8Pd8yBv2GaX-|Yomm4S6xKtQVE0yf@V{2N~bf<C^XD8}=$(|OnIEok8lnnje8IP?SrS>z?l!m{<DGdGb425AIeCcbKu1=3x(GrD*Sat!H+P3DWYtU>2w17jo1|eYbQ~HSaKzCZs3vZ@myA&JQy*AGUYU)FN0_G~)%pbI2U@zb8N)|M0f8nBJLFT!l^N(~N16MEj?5*nKIIJ}QBAODo_3DY29oY8R;4jKIU?YWCF|ttO9y*ubgw}RThqk!ItcvCN`o7GaMkGXanLvW9d^cUc!MUX~gg|yc)$mrqgW5tq{3bmyarb&dlNrGP-eskAMD$NkfpgJsj?5$W(I|RNRTE%YAi8lZv-Ajup_}u*>w6g#44v`L8RVb;87pX)#pgSm^V#TeGYD}G*O7bVkGqZjQ%_zhhdwEJwI)nirXSa{0G1+p(<%)uaB--}!~-a$P;F0SWZ8sc12`)aQA;YkDYKt8C=W2R=cRnv2|Wb(ttp7{1obE}UbNoF_TSgz^ts{Zf~^UW;RgEA+cns?VZX*n8~B!%UXgHk&MDn=o;f-ov@{!}bU5tkHZzR@PpPjlvxdlzF=Ub#4d@J<_dPa}qCqmf!OGTe<OGy?MyOMY3~p3poaJ>Kchgg&{<1YI9uAnq;ij07&!rpX%eWGF@}`(NnI#}hTaZ%%63Jt_A8>)S``k`T({6EhJc1WxX0z)>iX^K{X1HPF#dBgQY5<bMC9v<j2J#8l)=|`|0u)?^1>SNgMMsJ!?DMwck~lwdzuI~EYiB5^x0%tf83C4CR@}BI&7G^`pgo*u^_`qNgVHGKIx?q1Z{~4Ejx8|1d3ruNaLkiHpkM9>{QX4Z1r^B&(U+gJP9Y#bVaqRKQ?@pY^0}OwD5UlPq>zHl5&_ji!V&S0YmKyG{7m%e*(m>KXQ6RILht#*FJ1S7kaidJBQeZhHzuAOwBPY2U=4K+&j8Eh)P1G`&jxXA3O>^fyb1zo(?ujr3}uo^1c_wrGw?4`7zgkhN;&W}ws{h4%5)<s0W#qioV%eNvGKDtQ+fP@e_sYywdew%pm4%VFJA<$2M29|WMTh069^_LK8}I3E`<gw3nLC3Ugna+>uv<_Zvm~A|D7+(w!yS%Y35JODqYk0!O9Ci`bm`21fZreDc22Zo<v2939$jKd~`!Rc=5`F-BR2kE!5KX#T~JMw1*m{YgA=;AJQjS_0615YUns(^iAu6otI!kys09UruWOX>H@F_>MJKNO&rbaK;=WJDqx-@ZO%*|jB2i0x`ZTZn~M#Y-QV=0uTMbKuVg=3AyC`uuN18)7f0UI7cH1rDNqV!7wwkatSOyn|FC<uL*#7a!YexNuWN-UdhVcU-y}>B-6Py!enODdE^%=mk@zagWi-eDSUw7!I=KOcTDwX<BG3;Vql@sf2~q%R0M>kKOO2}d@Q8PVjxKvn#I6j?{|Dt`F~RZe({t?kN+JTWc^xK47(Cxo&BNElm#g`15x)cGhMHv?6g7auA40xvkJRsmeG<4a{D_!PvEg7nmZwPS`b8f<vzQNsOL=)iV$@#^iY(Reh+Dbs9KD>25VUwNOfg0JF*Ce(qx`*tVUV;mN<?gj%O-+(gv^m>_ai)7|0Q=FNbpXpQ1Muh2J;3zdO>o$`vo!MavVju*%k<lISi-`^~kS902$3P&csm;J9@iz0wt8(jZ_5u27EXAkPdB8Glj3X>};d8IhSg**DYk58+Tg+?o@?XidmYH^|-|seKT?=?%8yIoPjypczO%U7aRb_(Cl*wV7&5ezD6vc*XOUuCFp{T@tBL8LW&V%Au$Fm+2w@JGEas=no4^sGkiC(u0@n)-Q!NaQ_CqA>!Wcdv;F_^eMgeuUGcXxlfQ60)tjwBYhw*1_;>=dP^#Mv@ro3+!bbE92Czr|gYPLRE*LOejY*z`#qG1sKe+?*K!SV!S-{%tWM9@S#{(&+<tOW?tvkK%)tf+utG=pn2Q=va;TvT2%950V$iX{0KqNECJI&i2^JtE>ar;P`J+{6=RpvRS;C*Gbb^m~0@YymV;uQZQyzF_$1RuOg`1O0zlF#}$y>UgKd)a`@t!s2to0f#|bm3o2>qY6Muz(F?d>t=t2=bwQ(_=lBY|2kX)5^%7bopl6ea}8H{Pnj|TvG}GS<jBZu6mOm`qs7gnDIG@cZqW%mw!s;lU<}U`PP@s5lPqaB6F3!s_T!QNN*A`h?NbLd}BT#NxEZ{6_Z@Lcj$I^)fEk!P)DGc;1_%xmZ`5#2GBpS4o_0e`S_Y)@P$>lDw`RPL^yI*>IEN{FrRAQlZwB2`-dy7<8>JK#nD`^gW^me8Cq6r=E%VCNt0MAdA^d?q#iYKa#0TQy{(;tHx-ZR(7A|=;;W*K;Ja8ujSAJU>9sk+))5)$TBy|@9FOafGa}9kSHpyLhRe)f<wa1Nq}F$X1euDptF4+3NP%Ns=q1&&zia4iol)GIi()DG^%a=+q^jWwDy75Mkio60egHB~$X;go(R>5F*Ph!jdk%)Bjq^%ZU14fKrTc3j%x9kIQvonX2U5d^URB5{Hw1=?S_!}&+{M=5ef94{EAXGQ2&AP4rQ1H~H@;8vIw!mkoA)x03bET5lOK0x)s~^kP?jW{>?KOL1`l4xjj+!1*sb|q^(=1KZgO6kkGrL13`J~*yVsKxM+26ncOx7sQiM?py(rMAD`{+(d;P+Oz3ppy)#uD0WNa*!R^5`i$F0iKee%c;yPWdg9{;_o1BQfpNm8`xC4c2MjB5RtTRTA>LxHmz1FWE2IvYVViK=+8<3wlx)4qDKVx#}>n;U(X1Zg3TP6)2@6cQn`s^Ksz?dHpP3GY%?*$pbv{@f@!>2{Z-r&R`Y0;KC{F|I_q-+KRM+Te?iwwF|FLpVR6l1D*zf@7ip@w|uYT8_p(X0$=}JB=5xd=DA?2u1|91Dy*}`JE_q8$2m0yr<EOICXv=XW#NQqY<0K;yXK$_66Kb5YyTVp3_rkyj<Cvh@p;fd>0700V6SBEmU)ML(<llJyW8Pv-oc4tLk?j{<e>>#rbQ^c@vz2_usR)xj6qp^fM|l#4MV2NGQ-docO4-OD8XGe)~=`8SkLD`32rsmuRg;z&u!K!TDh-0|qTjXOEcc+?^IIOI`J}qw?EbZLNMaKb+ajk@A#cI<t;S!*o~*sYXw{&f#qHnD2D>QnScQs?Iy#mOB3ms#iNwop!i_3rjCM0Y(JAB`@YG0Q@Aj&>YBp?|R^-{ht+HNF0Y4)8X!{?x^${Q>d5m<Wh3)Vji5|+H^O^2wQf*>|F^rt3EbS$MSnfbtaOzEQlmhN_J6ywvdN#)OobuTy1Owd-bk{Vq?7Oz#wKcg<9i!lgd-P1{R=kP-OJXEu2M^icfOk5I9<7(xBfSa6~*(*<L&T+CKpDMtnuwW$V#n4P%hjsI>!)lOkEG_=MtG&u$DSRWOzLxtSSNWIJCFuW83KYT}`6On17?m1)KpL62*ER+nri@b8exB3`*lleY+74r`17)UjOIsars?O+;T%|MO%P3zw-s@r37?B4h{yXQ&7!&kozwB}o(Mg2@!ll_3G35qa~A7n>P?;B@+bV$`{TV;;9lJ|BF1X<W=6Kyc*a!RYS}84Ma)S%;ZJTfb{T1jkR|<g-!&P{C*^`rX#-NZv<*t@}VJ15Ce$?e9z_9XFb!a0=P~MDSr&1%?_;2*hNZl$;UDovzC~v4=BfM9Z|GB1;1fAzL~$8f85OEO0aq9RyfE#;d!Xf}|E?lw+liKkpwp>3J)7ypL{n1hZ=jJ-&j9wgpoh%U>LqtCBD2@ohBU&=y+oYdEJT5oWFVuP=gOAPh(;>(h&{HvJJYaqc*8{y}LDEm<)+ZbE6>r*&MR=r-0?*%IQ{`Bbk5TY($2?Crsdjf2fMGUvcD_HC9a<Tr^CPhT@kSZe0KH1xS^pEU-|E+~e1+{~69LS>@CRE}6R3XikC;J_o?mZ~qw=AOn#5g>Hs3;wn9;wicFCTGnVrypJ09udK5R_zV4u)r0y{M!e`Q~-~;hGWjUEMty>&w+hECdIK<^{?BEBa(@00g#1T6yAy&NUXEQSMKA?V8cObRB8QwhbLNw1*9I=!uV$ic;SBjP#!t5U@M2(-2BWPmJWWkGHy%uZSt4x@oo#qppX{Nx*38}rOXm5^e!pQoz`5Yrv>HGcp$$>m03e1{)dMpkr^~=Z~QWtFx191<c)j`GYk5{_02P?gNT-JkzO_*^4C@08v1hRu_r*EZxFH#18cd#(QJMY&4l`#WUunw@7DHQ^B9D@NbORB8wJQV6UIsKAEL54DXoGU*VDCAfg-&A=vH@>K*qQ+iM52JIkEgf)!-OK(WJ5ITF?SsR*T*nv%>AtlRY&gZi5Frr%g^>>WbODOY`5zM^U6kuDM0WWr>nWkO!iJzcQG`+bi2nfGYFjm?DH08kT6JwIMc8_ZRi82oCJ|f|4~ShzWi-Sh7Cm*aA3*n_<=p21n{ib3JB(2$;}=Cy`E*A*V|#ffMuZ`5zw(1&dy3pd?*DM4ieG+c;5@T)EWiW^BX)glzCIy*BQ&nCYz-6#y^~XR>$X8<$tR0_f*EyB_xK1j@ymW=(zh=M?v2?#e<<+$7MpDjms%o6}KEW?lN6^An{$uz~GUv_NfqFL7cNQ_tWL;i2v7_<Cu_PS`}oj)SYKDT&0%il$Pqca7T4RD~)UN`sl1)HWr!5BoQ!*g)huUvWxe?etn@Cai8m+#M-Lhd*Qh?OUJ1bDtYZ{f9huCpytAcjEpFhU{(ZO@wiraIw!C^S~pf^=xHlAl^d$Ri9kq?XxXJy%O|pW?5NDE%HpG)w$<YDWX)l;Zn2>9=y{g)C1O+r`mwmKr4b*u|hr{n$J;`uA8|mp38J@FlZizRNem_m_F$~$mKa_0ur!u)(@v1^ur2jKBr+Is~ce|v_GAe8WMj=BI$%2GSykOX}VW)!q^E92E*uu(UBtfR_w1PJ@(5ttM9x_;S4FpPWF*`fvcus<7nTTj!U|Hw%2erxMPA^_LV<dT+0mbP?G%_^0$q)oGe!<?&IfUeu#5o8Hw}IXu!yA4KoxC3*qAZF%TsRm#O-fC9Yl%5D(usAd*W`0UzH%VP;1v4iu1$(A{vQFz4B3T-@-RtR8rtx4N=hR940!hOE`_Z~OY$4S>2vI_Za93LYiKvw`21bp{};tcu{g96||4TLInO?UqpCUP9VK=dJ%xy#emoH(?x5nCt{fIXK&L$7?lWwwjVHamVGKYEv69R<g{xeC3d_0iBieSgy}Dd_pjN0>qob3fosAtem1qAfV^PQh1A+G~Df}I8&nko0(i4oA0w){*FE7g6RJXT_bGOA^^NBj5E=F;T4FX#I?9~=FYD)I%FHTYLoQkdL-d2qXuN)7s`^(;v%&ayQDjn0Wr7;4;~F}Aw1vwcb9b{JK0GzB9Say{ry84*|Ln8B9ii35RlUhw^K_GRiFnlt_P|#;v?F1u>}`|euv$dPx&9k*fX34M@q0NJRL9IGab>C1Lqpw;k|UZN3sG~uJaS|(g72$ro58!DJnoG)uRTSTv1}Mv&dsub7j!%3YqUNbbDQ)FshX_GeGQ0JNm*B;A`1d*T&Gzbn`1e=smc|zsTINhc7t@Ye7O(BS_mS$YOWSf9)8TF9q@`s)<DI5D-bL%@09#f8Xa15lJ&*Z7QNKA`h!<=BkW>Dwpm}KKaE9DmBbOC<o6s(V3d8O#6q<e8X01Z*)a{V%?^PGDldF(@pw;3pY%fF3_0eG^PKaC<c~n40vOfTdG+_5A;EU+AknEG}<H!vn>F}%pOK!GqHtC#UYSLDD3cj=xeXeKu8xS&s+pr2tudAOj#<dAVgWxl>-%|nF~i)(>a-XkH+o$6X9_k<c73YF(8}EKd@Jrr~*QRTODa`GW(4v_x>s=gMv_KTD0Aonsa5dr5jfm#~ROw8u5h|CuyR{Ir(%ef~Rkc_v8GB0@8g=dl)~|c26syQBwMJ<IOu%=QK?1ge=Ye9e)P#(9cRND?wb&orc2#<~keeH6(`53>W?gywP|ctWlblZ?b)axJOZ17+BJ318st|zPe(R-QMh&2fyC<qQv7+S6j^CTSJYlm_9=?CdzDn6b}U&&hF}>oHEUZQbk7hE96I1bZEsP<oC&^&%Dk7;oBoq&}~9ZEV_EirPFGlvZHQO%G5Yr^J{h?+0bt~y6Zz;0TTnZeDiXCqoE1_8A~XUpm2(LjJCZqU0S()hNWTuYA%P}$h6%8FCdpBX+i1J`2J2%`6p6hz(CS*YnCVGIPlZbq?$+<BI)knTkZ*<yql~O-?QL~+eIjYg?0V}5S5S=hFkwEH8iB^*5+?kff$5wb<`^M4y}=AYj3f4o}y)R#ze|njC?F<WveP*FLL#97G&#e;f8rzp_g=>G8Q8`FKB=Ec0k2pvuSJ=4?cf#uuMLesynFk@WpX{Z`(OE%$DDU%Q%`O*e=mR?xksV-OtY7`oJ(Sal<0!@Ov^Mj#g53?lwV)6YS#51r5g_K+PTV9>~&MQ=wd$PBQNZTKF~F7^w!S2|~*X`K~4aN_si&lw>T>_)5n^nq%zD@MTgF^kbcNd-AIY<9k4ub)^!d9aKCxz(h#tM%%TVB=5&R)K#Vu+vaEjLWp#}cMR+Qpa_#%>^|*HNi;!w$GnXe9Kw*T$+N?@>+>LX&IQnlXH25?`^RenKo<)4W6g9g2X5yj1fHWlhK8C#_vW%IBf6<>d<EMkQlE`nSH$p#g3`1|vts@VqdaMZ&X@dE!-93qY8sxcp(aeIaf8vfBM1MVV1T2Nej{K-)~_gaR0ZBpyV=|&{3>LsMsBl}*>&sPW5y5yx#(GnjPy%{Vkq*ap*Vw%lue2M$s+gmU}d>!7A+_M00'),format=L.FORMAT_RAW,filters=[{"id":L.FILTER_LZMA2}]))
diff --git a/records/track_10min_16mb/2026-04-13_SystemsOpt_SP8192_LegalTTT/train_seed0.log b/records/track_10min_16mb/2026-04-13_SystemsOpt_SP8192_LegalTTT/train_seed0.log
new file mode 100644
index 0000000000..86b34cfd54
--- /dev/null
+++ b/records/track_10min_16mb/2026-04-13_SystemsOpt_SP8192_LegalTTT/train_seed0.log
@@ -0,0 +1,147 @@
+W0412 04:33:07.570000 56418 torch/distributed/run.py:803] 
+W0412 04:33:07.570000 56418 torch/distributed/run.py:803] *****************************************
+W0412 04:33:07.570000 56418 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. 
+W0412 04:33:07.570000 56418 torch/distributed/run.py:803] *****************************************
+Hyperparameters:
+  adam_eps: 1e-08
+  adam_wd: 0.02
+  beta1: 0.9
+  beta2: 0.95
+  compressor: brotli
+  data_dir: ./data/
+  datasets_dir: ./data/datasets/fineweb10B_sp8192
+  distributed: True
+  ema_decay: 0.9965
+  embed_bits: 8
+  embed_clip_sigmas: 20.0
+  embed_lr: 0.6
+  embed_wd: 0.085
+  embedding_dim: 512
+  enable_looping_at: 0.35
+  etlb_clip: 3.0
+  etlb_enabled: False
+  etlb_lr: 0.05
+  etlb_steps: 5
+  eval_seq_len: 2048
+  eval_stride: 64
+  gptq_calibration_batches: 64
+  gptq_reserve_seconds: 12.0
+  grad_accum_steps: 1
+  grad_clip_norm: 0.3
+  head_lr: 0.008
+  is_main_process: True
+  iterations: 20000
+  ln_scale: True
+  local_rank: 0
+  logfile: logs/fa987d91-4c03-4dfc-873a-c960e2ffc004.txt
+  logit_softcap: 30.0
+  loop_end: 5
+  loop_start: 3
+  matrix_bits: 6
+  matrix_clip_sigmas: 12.85
+  matrix_lr: 0.022
+  max_wallclock_seconds: 600.0
+  min_lr: 0.0
+  mlp_mult: 4.0
+  model_dim: 512
+  model_path: final_model.pt
+  muon_backend_steps: 5
+  muon_beta2: 0.95
+  muon_momentum: 0.99
+  muon_momentum_warmup_start: 0.92
+  muon_momentum_warmup_steps: 1500
+  muon_row_normalize: True
+  muon_wd: 0.095
+  num_heads: 8
+  num_kv_heads: 4
+  num_layers: 11
+  num_loops: 2
+  parallel_residual_start: 7
+  qk_gain_init: 5.25
+  quantized_model_path: final_model.int6.ptz
+  rank: 0
+  rope_base: 10000.0
+  rope_dims: 16
+  rope_train_seq_len: 2048
+  run_id: fa987d91-4c03-4dfc-873a-c960e2ffc004
+  scalar_lr: 0.02
+  seed: 0
+  skip_gates_enabled: True
+  sliding_window_enabled: True
+  tie_embeddings: True
+  tied_embed_init_std: 0.005
+  tied_embed_lr: 0.03
+  tokenizer_path: ./data/tokenizers/fineweb_8192_bpe.model
+  train_batch_tokens: 786432
+  train_files: ./data/datasets/fineweb10B_sp8192/fineweb_train_*.bin
+  train_log_every: 500
+  train_seq_len: 2048
+  ttt_chunk_tokens: 32768
+  ttt_enabled: True
+  ttt_epochs: 3
+  ttt_lr: 0.005
+  ttt_momentum: 0.9
+  val_batch_tokens: 524288
+  val_files: ./data/datasets/fineweb10B_sp8192/fineweb_val_*.bin
+  val_loss_every: 4000
+  vocab_size: 8192
+  warmdown_frac: 0.72
+  warmup_steps: 20
+  world_size: 8
+  xsa_last_n: 11
+train_shards: 80
+val_tokens: 40540160
+model_params:35944536
+gptq:reserving 12s, effective=588000ms
+warmup_step: 1/20
+warmup_step: 2/20
+warmup_step: 3/20
+warmup_step: 4/20
+warmup_step: 5/20
+warmup_step: 6/20
+warmup_step: 10/20
+warmup_step: 20/20
+loop_warmup:enabled encoder:[0, 1, 2, 3, 4, 5, 3, 4] decoder:[5, 3, 4, 5, 6, 7, 8, 9, 10]
+loop_warmup_step: 1/20
+loop_warmup_step: 2/20
+loop_warmup_step: 3/20
+loop_warmup_step: 4/20
+loop_warmup_step: 5/20
+loop_warmup_step: 6/20
+loop_warmup_step: 10/20
+loop_warmup_step: 20/20
+0/20000 val_loss: 9.0072 val_bpb: 3.4870
+1/20000 train_loss: 9.0086 train_time: 0.0m tok/s: 9225961
+2/20000 train_loss: 12.3279 train_time: 0.0m tok/s: 9362528
+3/20000 train_loss: 11.0101 train_time: 0.0m tok/s: 8796719
+4/20000 train_loss: 9.5067 train_time: 0.0m tok/s: 8542046
+5/20000 train_loss: 8.3986 train_time: 0.0m tok/s: 8412781
+500/20000 train_loss: 3.3744 train_time: 0.8m tok/s: 7804756
+1000/20000 train_loss: 3.2774 train_time: 1.7m tok/s: 7780751
+1500/20000 train_loss: 3.1802 train_time: 2.5m tok/s: 7785670
+2000/20000 train_loss: 3.0675 train_time: 3.4m tok/s: 7792306
+layer_loop:enabled step:2039 frac:0.350 encoder:[0, 1, 2, 3, 4, 5, 3, 4] decoder:[5, 3, 4, 5, 6, 7, 8, 9, 10]
+2500/20000 train_loss: 3.1190 train_time: 4.6m tok/s: 7164121
+3000/20000 train_loss: 2.9019 train_time: 5.8m tok/s: 6764013
+3500/20000 train_loss: 2.9412 train_time: 7.1m tok/s: 6504175
+4000/20000 train_loss: 2.8224 train_time: 8.3m tok/s: 6315937
+4000/20000 val_loss: 2.8769 val_bpb: 1.1137
+4500/20000 train_loss: 2.8396 train_time: 9.5m tok/s: 6184255
+4607/20000 val_loss: 2.8101 val_bpb: 1.0879
+stopping_early: wallclock_cap train_time: 588170ms step: 4607/20000
+peak memory allocated: 39278 MiB reserved: 39416 MiB
+ema:applying EMA weights
+pre-quantization post-ema val_loss:2.80687991 val_bpb:1.08663068 eval_time:6875ms
+Code size: 18185 bytes
+GPTQ:collecting Hessians from calibration data...
+GPTQ:collected 67 Hessians in 12.8s
+Quantized weights:
+  gptq (int6): blocks.attn.c_k.weight, blocks.attn.c_q.weight, blocks.attn.c_v.weight, blocks.attn.proj.weight, blocks.mlp.fc.weight, blocks.mlp.proj.weight
+  gptq (int8): tok_emb.weight
+  passthrough (float16): blocks.attn.q_gain, blocks.attn_scale, blocks.mlp_scale, blocks.resid_mix, skip_gates, skip_weights
+Serialized model quantized+brotli: 15975552 bytes
+Total submission size quantized+brotli: 15993737 bytes
+quantized val_loss:2.83696682 val_bpb:1.09827827 eval_time:8832ms
+quantized_sliding_window val_loss:2.79363272 val_bpb:1.08150228 eval_time:87085ms
+ttt:start chunks=1238 ttt_lr=0.005 ttt_epochs=3
+quantized_ttt val_loss:2.78951654 val_bpb:1.07990878 eval_time:308087ms
diff --git a/records/track_10min_16mb/2026-04-13_SystemsOpt_SP8192_LegalTTT/train_seed3141.log b/records/track_10min_16mb/2026-04-13_SystemsOpt_SP8192_LegalTTT/train_seed3141.log
new file mode 100644
index 0000000000..622c186536
--- /dev/null
+++ b/records/track_10min_16mb/2026-04-13_SystemsOpt_SP8192_LegalTTT/train_seed3141.log
@@ -0,0 +1,147 @@
+W0412 06:29:36.635000 63192 torch/distributed/run.py:803] 
+W0412 06:29:36.635000 63192 torch/distributed/run.py:803] *****************************************
+W0412 06:29:36.635000 63192 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. 
+W0412 06:29:36.635000 63192 torch/distributed/run.py:803] *****************************************
+Hyperparameters:
+  adam_eps: 1e-08
+  adam_wd: 0.02
+  beta1: 0.9
+  beta2: 0.95
+  compressor: brotli
+  data_dir: ./data/
+  datasets_dir: ./data/datasets/fineweb10B_sp8192
+  distributed: True
+  ema_decay: 0.9965
+  embed_bits: 8
+  embed_clip_sigmas: 20.0
+  embed_lr: 0.6
+  embed_wd: 0.085
+  embedding_dim: 512
+  enable_looping_at: 0.35
+  etlb_clip: 3.0
+  etlb_enabled: False
+  etlb_lr: 0.05
+  etlb_steps: 5
+  eval_seq_len: 2048
+  eval_stride: 64
+  gptq_calibration_batches: 64
+  gptq_reserve_seconds: 12.0
+  grad_accum_steps: 1
+  grad_clip_norm: 0.3
+  head_lr: 0.008
+  is_main_process: True
+  iterations: 20000
+  ln_scale: True
+  local_rank: 0
+  logfile: logs/78e1c274-1c75-46b6-ad05-589bea50402e.txt
+  logit_softcap: 30.0
+  loop_end: 5
+  loop_start: 3
+  matrix_bits: 6
+  matrix_clip_sigmas: 12.85
+  matrix_lr: 0.022
+  max_wallclock_seconds: 600.0
+  min_lr: 0.0
+  mlp_mult: 4.0
+  model_dim: 512
+  model_path: final_model.pt
+  muon_backend_steps: 5
+  muon_beta2: 0.95
+  muon_momentum: 0.99
+  muon_momentum_warmup_start: 0.92
+  muon_momentum_warmup_steps: 1500
+  muon_row_normalize: True
+  muon_wd: 0.095
+  num_heads: 8
+  num_kv_heads: 4
+  num_layers: 11
+  num_loops: 2
+  parallel_residual_start: 7
+  qk_gain_init: 5.25
+  quantized_model_path: final_model.int6.ptz
+  rank: 0
+  rope_base: 10000.0
+  rope_dims: 16
+  rope_train_seq_len: 2048
+  run_id: 78e1c274-1c75-46b6-ad05-589bea50402e
+  scalar_lr: 0.02
+  seed: 3141
+  skip_gates_enabled: True
+  sliding_window_enabled: True
+  tie_embeddings: True
+  tied_embed_init_std: 0.005
+  tied_embed_lr: 0.03
+  tokenizer_path: ./data/tokenizers/fineweb_8192_bpe.model
+  train_batch_tokens: 786432
+  train_files: ./data/datasets/fineweb10B_sp8192/fineweb_train_*.bin
+  train_log_every: 500
+  train_seq_len: 2048
+  ttt_chunk_tokens: 32768
+  ttt_enabled: True
+  ttt_epochs: 3
+  ttt_lr: 0.005
+  ttt_momentum: 0.9
+  val_batch_tokens: 524288
+  val_files: ./data/datasets/fineweb10B_sp8192/fineweb_val_*.bin
+  val_loss_every: 4000
+  vocab_size: 8192
+  warmdown_frac: 0.72
+  warmup_steps: 20
+  world_size: 8
+  xsa_last_n: 11
+train_shards: 80
+val_tokens: 40540160
+model_params:35944536
+gptq:reserving 12s, effective=588000ms
+warmup_step: 1/20
+warmup_step: 2/20
+warmup_step: 3/20
+warmup_step: 4/20
+warmup_step: 5/20
+warmup_step: 6/20
+warmup_step: 10/20
+warmup_step: 20/20
+loop_warmup:enabled encoder:[0, 1, 2, 3, 4, 5, 3, 4] decoder:[5, 3, 4, 5, 6, 7, 8, 9, 10]
+loop_warmup_step: 1/20
+loop_warmup_step: 2/20
+loop_warmup_step: 3/20
+loop_warmup_step: 4/20
+loop_warmup_step: 5/20
+loop_warmup_step: 6/20
+loop_warmup_step: 10/20
+loop_warmup_step: 20/20
+0/20000 val_loss: 9.0092 val_bpb: 3.4877
+1/20000 train_loss: 9.0100 train_time: 0.0m tok/s: 9246251
+2/20000 train_loss: 12.3477 train_time: 0.0m tok/s: 9391537
+3/20000 train_loss: 11.0235 train_time: 0.0m tok/s: 8840673
+4/20000 train_loss: 9.4880 train_time: 0.0m tok/s: 8590099
+5/20000 train_loss: 8.3438 train_time: 0.0m tok/s: 8446006
+500/20000 train_loss: 3.3802 train_time: 0.8m tok/s: 7813922
+1000/20000 train_loss: 3.2841 train_time: 1.7m tok/s: 7808250
+1500/20000 train_loss: 3.1838 train_time: 2.5m tok/s: 7821604
+2000/20000 train_loss: 3.0711 train_time: 3.3m tok/s: 7831531
+layer_loop:enabled step:2050 frac:0.350 encoder:[0, 1, 2, 3, 4, 5, 3, 4] decoder:[5, 3, 4, 5, 6, 7, 8, 9, 10]
+2500/20000 train_loss: 3.1237 train_time: 4.5m tok/s: 7209477
+3000/20000 train_loss: 2.9000 train_time: 5.8m tok/s: 6800104
+3500/20000 train_loss: 2.9452 train_time: 7.0m tok/s: 6536012
+4000/20000 train_loss: 2.8242 train_time: 8.3m tok/s: 6343559
+4000/20000 val_loss: 2.8790 val_bpb: 1.1145
+4500/20000 train_loss: 2.8441 train_time: 9.5m tok/s: 6208556
+4622/20000 val_loss: 2.8105 val_bpb: 1.0880
+stopping_early: wallclock_cap train_time: 588109ms step: 4622/20000
+peak memory allocated: 39278 MiB reserved: 39416 MiB
+ema:applying EMA weights
+pre-quantization post-ema val_loss:2.80734608 val_bpb:1.08681115 eval_time:6896ms
+Code size: 18185 bytes
+GPTQ:collecting Hessians from calibration data...
+GPTQ:collected 67 Hessians in 12.8s
+Quantized weights:
+  gptq (int6): blocks.attn.c_k.weight, blocks.attn.c_q.weight, blocks.attn.c_v.weight, blocks.attn.proj.weight, blocks.mlp.fc.weight, blocks.mlp.proj.weight
+  gptq (int8): tok_emb.weight
+  passthrough (float16): blocks.attn.q_gain, blocks.attn_scale, blocks.mlp_scale, blocks.resid_mix, skip_gates, skip_weights
+Serialized model quantized+brotli: 15977252 bytes
+Total submission size quantized+brotli: 15995437 bytes
+quantized val_loss:2.83721177 val_bpb:1.09837309 eval_time:8847ms
+quantized_sliding_window val_loss:2.79406439 val_bpb:1.08166940 eval_time:86618ms
+ttt:start chunks=1238 ttt_lr=0.005 ttt_epochs=3
+quantized_ttt val_loss:2.79002185 val_bpb:1.08010440 eval_time:307147ms
diff --git a/records/track_10min_16mb/2026-04-13_SystemsOpt_SP8192_LegalTTT/train_seed42.log b/records/track_10min_16mb/2026-04-13_SystemsOpt_SP8192_LegalTTT/train_seed42.log
new file mode 100644
index 0000000000..8477221eff
--- /dev/null
+++ b/records/track_10min_16mb/2026-04-13_SystemsOpt_SP8192_LegalTTT/train_seed42.log
@@ -0,0 +1,147 @@
+W0412 03:49:45.337000 53115 torch/distributed/run.py:803] 
+W0412 03:49:45.337000 53115 torch/distributed/run.py:803] *****************************************
+W0412 03:49:45.337000 53115 torch/distributed/run.py:803] Setting OMP_NUM_THREADS environment variable for each process to be 1 in default, to avoid your system being overloaded, please further tune the variable for optimal performance in your application as needed. 
+W0412 03:49:45.337000 53115 torch/distributed/run.py:803] *****************************************
+Hyperparameters:
+  adam_eps: 1e-08
+  adam_wd: 0.02
+  beta1: 0.9
+  beta2: 0.95
+  compressor: brotli
+  data_dir: ./data/
+  datasets_dir: ./data/datasets/fineweb10B_sp8192
+  distributed: True
+  ema_decay: 0.9965
+  embed_bits: 8
+  embed_clip_sigmas: 20.0
+  embed_lr: 0.6
+  embed_wd: 0.085
+  embedding_dim: 512
+  enable_looping_at: 0.35
+  etlb_clip: 3.0
+  etlb_enabled: False
+  etlb_lr: 0.05
+  etlb_steps: 5
+  eval_seq_len: 2048
+  eval_stride: 64
+  gptq_calibration_batches: 64
+  gptq_reserve_seconds: 12.0
+  grad_accum_steps: 1
+  grad_clip_norm: 0.3
+  head_lr: 0.008
+  is_main_process: True
+  iterations: 20000
+  ln_scale: True
+  local_rank: 0
+  logfile: logs/c4ff5b23-f15a-4325-85da-337d08a9590b.txt
+  logit_softcap: 30.0
+  loop_end: 5
+  loop_start: 3
+  matrix_bits: 6
+  matrix_clip_sigmas: 12.85
+  matrix_lr: 0.022
+  max_wallclock_seconds: 600.0
+  min_lr: 0.0
+  mlp_mult: 4.0
+  model_dim: 512
+  model_path: final_model.pt
+  muon_backend_steps: 5
+  muon_beta2: 0.95
+  muon_momentum: 0.99
+  muon_momentum_warmup_start: 0.92
+  muon_momentum_warmup_steps: 1500
+  muon_row_normalize: True
+  muon_wd: 0.095
+  num_heads: 8
+  num_kv_heads: 4
+  num_layers: 11
+  num_loops: 2
+  parallel_residual_start: 7
+  qk_gain_init: 5.25
+  quantized_model_path: final_model.int6.ptz
+  rank: 0
+  rope_base: 10000.0
+  rope_dims: 16
+  rope_train_seq_len: 2048
+  run_id: c4ff5b23-f15a-4325-85da-337d08a9590b
+  scalar_lr: 0.02
+  seed: 42
+  skip_gates_enabled: True
+  sliding_window_enabled: True
+  tie_embeddings: True
+  tied_embed_init_std: 0.005
+  tied_embed_lr: 0.03
+  tokenizer_path: ./data/tokenizers/fineweb_8192_bpe.model
+  train_batch_tokens: 786432
+  train_files: ./data/datasets/fineweb10B_sp8192/fineweb_train_*.bin
+  train_log_every: 500
+  train_seq_len: 2048
+  ttt_chunk_tokens: 32768
+  ttt_enabled: True
+  ttt_epochs: 3
+  ttt_lr: 0.005
+  ttt_momentum: 0.9
+  val_batch_tokens: 524288
+  val_files: ./data/datasets/fineweb10B_sp8192/fineweb_val_*.bin
+  val_loss_every: 4000
+  vocab_size: 8192
+  warmdown_frac: 0.72
+  warmup_steps: 20
+  world_size: 8
+  xsa_last_n: 11
+train_shards: 80
+val_tokens: 40540160
+model_params:35944536
+gptq:reserving 12s, effective=588000ms
+warmup_step: 1/20
+warmup_step: 2/20
+warmup_step: 3/20
+warmup_step: 4/20
+warmup_step: 5/20
+warmup_step: 6/20
+warmup_step: 10/20
+warmup_step: 20/20
+loop_warmup:enabled encoder:[0, 1, 2, 3, 4, 5, 3, 4] decoder:[5, 3, 4, 5, 6, 7, 8, 9, 10]
+loop_warmup_step: 1/20
+loop_warmup_step: 2/20
+loop_warmup_step: 3/20
+loop_warmup_step: 4/20
+loop_warmup_step: 5/20
+loop_warmup_step: 6/20
+loop_warmup_step: 10/20
+loop_warmup_step: 20/20
+0/20000 val_loss: 9.0090 val_bpb: 3.4877
+1/20000 train_loss: 9.0104 train_time: 0.0m tok/s: 9191547
+2/20000 train_loss: 12.3645 train_time: 0.0m tok/s: 9416998
+3/20000 train_loss: 11.0075 train_time: 0.0m tok/s: 8883056
+4/20000 train_loss: 9.4552 train_time: 0.0m tok/s: 8634412
+5/20000 train_loss: 8.3278 train_time: 0.0m tok/s: 8498155
+500/20000 train_loss: 3.3791 train_time: 0.8m tok/s: 7832557
+1000/20000 train_loss: 3.2876 train_time: 1.7m tok/s: 7813307
+1500/20000 train_loss: 3.1843 train_time: 2.5m tok/s: 7816205
+2000/20000 train_loss: 3.0689 train_time: 3.4m tok/s: 7822690
+layer_loop:enabled step:2047 frac:0.350 encoder:[0, 1, 2, 3, 4, 5, 3, 4] decoder:[5, 3, 4, 5, 6, 7, 8, 9, 10]
+2500/20000 train_loss: 3.1243 train_time: 4.6m tok/s: 7199531
+3000/20000 train_loss: 2.9011 train_time: 5.8m tok/s: 6793033
+3500/20000 train_loss: 2.9459 train_time: 7.0m tok/s: 6530046
+4000/20000 train_loss: 2.8242 train_time: 8.3m tok/s: 6338311
+4000/20000 val_loss: 2.8789 val_bpb: 1.1145
+4500/20000 train_loss: 2.8464 train_time: 9.5m tok/s: 6203787
+4619/20000 val_loss: 2.8106 val_bpb: 1.0881
+stopping_early: wallclock_cap train_time: 588109ms step: 4619/20000
+peak memory allocated: 39278 MiB reserved: 39416 MiB
+ema:applying EMA weights
+pre-quantization post-ema val_loss:2.80748328 val_bpb:1.08686427 eval_time:6920ms
+Code size: 18185 bytes
+GPTQ:collecting Hessians from calibration data...
+GPTQ:collected 67 Hessians in 12.8s
+Quantized weights:
+  gptq (int6): blocks.attn.c_k.weight, blocks.attn.c_q.weight, blocks.attn.c_v.weight, blocks.attn.proj.weight, blocks.mlp.fc.weight, blocks.mlp.proj.weight
+  gptq (int8): tok_emb.weight
+  passthrough (float16): blocks.attn.q_gain, blocks.attn_scale, blocks.mlp_scale, blocks.resid_mix, skip_gates, skip_weights
+Serialized model quantized+brotli: 15975016 bytes
+Total submission size quantized+brotli: 15993201 bytes
+quantized val_loss:2.83658359 val_bpb:1.09812991 eval_time:8895ms
+quantized_sliding_window val_loss:2.79367948 val_bpb:1.08152039 eval_time:86709ms
+ttt:start chunks=1238 ttt_lr=0.005 ttt_epochs=3
+quantized_ttt val_loss:2.79024700 val_bpb:1.08019157 eval_time:303873ms