Initial commit: ept — backprop-free equilibrium transformer (EP)

Code (ep_run/), organized docs (docs/{method,campaign,hardware,outreach,paper}),
analysis scripts (scripts/), ONBOARDING.md entry point. Large data/checkpoints
git-ignored (share separately).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn

1 files changed, 25 insertions, 0 deletions

diff --git a/ep_run/auto_probe.py b/ep_run/auto_probe.py
new file mode 100644
index 0000000..90b969e
--- /dev/null
+++ b/ep_run/auto_probe.py
@@ -0,0 +1,25 @@
+"""Wait for the first converged clean-EP ckpt, run the fixed oracle-adjoint probe on it, report g_EP vs g_transpose."""
+import time, os, subprocess, shutil
+WD = "/home/yurenh2/ept/ep_run"; os.chdir(WD)
+CK, FROZEN = "runs/ep_clean.pt", "runs/ep_clean_probe.pt"
+got = False
+for _ in range(45):                       # up to ~67 min
+    time.sleep(90)
+    if os.path.exists(CK) and os.path.getsize(CK) > 1_000_000:
+        got = True; break
+if not got:
+    print("=== AUTO-PROBE: ep_clean.pt never appeared in ~67min ==="); raise SystemExit
+shutil.copy2(CK, FROZEN)                   # freeze (avoid write race)
+env = dict(os.environ, CUDA_VISIBLE_DEVICES="0", PYTORCH_CUDA_ALLOC_CONF="expandable_segments:True")
+try:
+    r = subprocess.run(["python3", "asym_probe.py", "--ckpt", FROZEN, "--B", "8"],
+                       env=env, capture_output=True, text=True, timeout=1200)
+    txt = r.stdout + "\n" + r.stderr
+except Exception as e:
+    txt = f"probe run error: {e}"
+print("=== AUTO ORACLE PROBE on first clean-EP ckpt ===")
+KEEP = ("z* res", "GMRES", "resid", "cos(", "g_transpose", "g_EP", "g_BPTT", "interpret", "->", "exact", "AsymEP", "# ckpt", "step ")
+DROP = ("UserWarning", "cuBLAS", "warnings.warn", "Triggered", "return Variable", "FutureWarning")
+for line in txt.splitlines():
+    if any(k in line for k in KEEP) and not any(b in line for b in DROP):
+        print(line)