v3 trainer + English docs + production x8 results

docs/HOW_FRACTUS_IS_TRAINED.md

@@ -1,6 +1,6 @@
 # How Fractus Is Trained
 
-**Updated:** 2026-08-24 (adds optimized v2 path + time-to-finish math)  
+**Updated:** 2026-08-26 (production x8 run live — measured rates replace estimates; adds optimized v2/v3 path + time-to-finish math)  
 **Author:** Philippe-Antoine Robert  
 **Model repo:** https://huggingface.co/thefinalboss/fractus-cte  
 **Dataset:** https://huggingface.co/datasets/thefinalboss/fractus-datasets
@@ -164,12 +164,17 @@ Throughput ~900-1100 tok/s/GPU at B=2. One full phase-2 pass ~4-5 days wall-cloc
 | 900–1100 (baseline, measured) | **4.5 – 5.5 days** |
 | ~1300 (post-opt, ×1.3) | ~3.5 days |
 | ~2000 (×2) | ~2.3 days |
-| ~3000 (×3) | ~1.5 days |
-
-Post-optimization rates are NOT yet measured on GPU — the pod bench
-(benchmarks/bench_attention.py + fast4gpu_boost_v2.py stdout) gives the
-real number within minutes of relaunch. "Dataset complet" = one pass;
-multiply linearly for additional epochs.
+| **~1570 (MEASURED production x8, 2026-08-26)** | **~3.1 days** |
+
+**Production status 2026-08-26:** phase-2 resumed on 8×RTX 5090 with the
+optimized stack (v3 lineage: chunked kernel + BLOCK_CKPT + CE_CHUNK=2048,
+B=8). Sustained **1,554–1,604 tok/s/GPU ≈ 12,600 aggregate**, VRAM
+18.9 GB/32 GB, lb = 14.028 stable. Resume honored gpu0–5 at the x6 positions
+(1.25–1.40M), gpu6=655,360 / gpu7=768,000. Hourly safety sync to
+`checkpoints/x8run/` + `checkpoints/X8_MANIFEST.json`. Projected finish:
+**≈3.1 days from launch**. Deployment automated via
+`scripts/pod_deploy_x8.sh`; full measurements in
+`OPTIMIZATION_2026-08-22.md` §4.
 
 ---