Commit History
-
EBVT instrumentation: expert-usage probe (forward hooks, layer-cake self-check) + baseline X8 103M + harness wiring
Philippe-Antoine Robert- scripts/ebvt_probe.py: EBVTProbe on the 16x128 activation caterpillar (usage cut balances b': spoke = expert count, chain edge = min(L, M-L)); V, H, P_t profile, persistence, support, dominance; identity V = sum_t P_t self-checked on every report; native _expert_hits cross-check by diff (that counter is cumulative). - scripts/ebvt_smoke.py: validation on the real x8run ckpt (native match exact, cumulative x2, reset). - scripts/probe_ciel_ouvert.py: EBVT attached (attach, reset/report in gen_surgical + S1, battery means, S6 continuous state). - docs/2026-09-03-EBVT-X8-BASELINE.md: point 0 of the 4B curve (V=1948686, H=848, P1=8054, pers=1.0, supp=34/2048, dom=0.031, 3-run identical) + kill-test engagement.0840a17 -
docs: re-run verification on HF repo (2026-09-03) + fixed harness
Philippe-Antoine Robert- Full operation re-run from a byte-verified snapshot of HF main @ 5326f88 (92/92 files vs git tree sha1s), HF checkpoint re-downloaded, seed 42. - All S0-S7 metrics within +/-1.0 of 2026-09-02; S0 identical (2.14); same conclusions. Non-bit-identical cause: re-provisioned sandbox torch. - scripts/probe_ciel_ouvert.py: fixed two conversion bugs (leftover _kf alias NameError; repo root missing from sys.path). - docs/X8_RERUN_ON_HF_REPO_2026-09-03.json: raw re-run results.0906b03 -
docs: X8 unified checkpoint — sandbox live operation (2026-09-02)
Philippe-Antoine Robert6 files on top of current main (813f471): - docs/X8_CIEL_OUVERT.md: full journal (EN + FR TL;DR) + push manifest with SHA-256 - docs/X8_UNIFIED_PROBE.json + docs/OPERATION_CIEL_OUVERT_RESULTS.json: raw results - scripts/probe_ciel_ouvert.py: reproducible sandbox harness - docs/MASTER_RUN_LOG.md: §11 entry (merged on top of §10 decode-surgery-I) - docs/DISCOVERY_LOG.md: §7 findings + thesis point 5 (merged on top of §6) Runs on the real HF checkpoint (checkpoints/x8run/FRACTUS_1B_X8_MERGED.pt, ts 2026-08-29) with the real HF code. Gate v2 proposal included.5326f88
-
Upload scripts/fast4gpu_boost_v4.py with huggingface_hub
Verified signaturea25e1b3 -
Upload scripts/hf_push_x8.py with huggingface_hub
Verified signaturebac0aaf
-
v3 trainer + English docs + production x8 results
Verified signature7c7e1e2 -
ops: one-shot x8 pod deploy script
Verified signature63d61fb
-
fix: atomic checkpoint save (tmp+os.replace)
Verified signature612044d
-
opt: cumsum/chunked attention kernels, memory-flat CE, block checkpointing, v2 trainer (proven equivalent, 46 tests)
Verified signature938dd12
-
Upload scripts/prep_kuramoto_fix_resume.py with huggingface_hub
Verified signatureda1fbd7
-
Upload scripts/fast4gpu_boost.py with huggingface_hub
Verified signature9f2620b -
Upload scripts/smoke_load_merged.py with huggingface_hub
Verified signaturebf8ce80 -
Upload scripts/fast4gpu_boost.py with huggingface_hub
Verified signaturec20883a
-
Upload scripts/fast4gpu_stage2_ss.py with huggingface_hub
Verified signaturee484d86 -
Upload scripts/fast4gpu_stage2_fine.py with huggingface_hub
Verified signaturefaa037a -
Upload scripts/fast4gpu_surgery.py with huggingface_hub
Verified signature36e21fe
-
Add production multi-GPU training scripts (B=2 + compile + shards)
Verified signature4119c42 -
Add production multi-GPU training scripts (B=2 + compile + shards)
Verified signaturef937a6c -
Add production multi-GPU training scripts (B=2 + compile + shards)
Verified signature356d264
-
Upload scripts/fractus_self_dataset.py with huggingface_hub
Verified signaturec301656 -
Upload folder using huggingface_hub
Verified signature6c223ff
-
Upload scripts/train_salience_head.py with huggingface_hub
Verified signatureb65956a -
Upload scripts/build_quality_corpus.py with huggingface_hub
Verified signature815ba07 -
Upload scripts/train_1b_gpu.py with huggingface_hub
Verified signaturec711fa2 -
Upload scripts/train_progressive.py with huggingface_hub
Verified signaturea4ddff3