wassname/evil_MoE
1
0
Fork 0
mirror of https://github.com/wassname/evil_MoE.git synced 2026-07-24 13:00:36 +08:00
Code Issues Packages Projects Releases Wiki Activity

feat: route2 resid column = cos(deployed delta_S.grad after routing, v_grad)

Browse Source 
Disambiguates qE: high qE + resid~0 = hack stripped cleanly into the
deleted quarantine; high qE + resid>0 = false-negative leak into the
deployed knob. hkgap measures cloud separation, qE measures energy,
neither measures hack-ward alignment of the KEPT gradient. (DeepSeek
review #5 starvation-vs-correct-routing diagnostic.)

Co-Authored-By: Claudypoo <288921227+claudypoo@users.noreply.github.com>
This commit is contained in:
wassname
2026-06-01 03:20:48 +00:00
co-authored by Claudypoo
parent 59f8b6efdc
commit dfc6068896
1 changed files with 11 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/projected_grpo/train.py
+11 -1
View File
@@ -688,6 +688,7 @@ class StepLogger:
cols += [
_Col("tau", 6, "tau", "+.2f", "per-step calibrated route threshold (midpoint of hack vs clean cos clouds)"),
_Col("hkgap", 6, "hkgap", "+.2f", "ema_hack_cos - ema_clean_cos; >0 = v_grad still separates hack from clean (else direction dead)"),
_Col("resid", 6, "resid", "+.2f", "cos(deployed delta_S.grad AFTER routing, v_grad); ~0 = hack stripped cleanly, >0 = leak into deployed knob"),
]
if arm in ("routing", "routing2"):
cols += [
@@ -1147,6 +1148,7 @@ def main(cfg: Config) -> int:
step_flagged: list[float] = []
step_tau: list[float] = [] # per-(prompt,module) calibrated route threshold
step_hkgap: list[float] = [] # ema_hack_cos - ema_clean_cos (discrimination gauge)
step_resid: list[float] = [] # cos(delta_S.grad AFTER routing, v_grad): hack-ward leak into deployed knob
def _route2_grad_filter(info, n_rollouts: int,
hack_anchor: torch.Tensor,
@@ -1194,7 +1196,14 @@ def main(cfg: Config) -> int:
# delta_S keeps only the unflagged. Capacity-balanced: both shape [r].
step_grad_hack[name] = (step_grad_hack[name] + sub.detach().clone()
if name in step_grad_hack else sub.detach().clone())
return g - sub
g_keep = g - sub # the deployed knob's gradient
# Residual hack-ward alignment of the KEPT grad. Disambiguates qE:
# qE high + resid~0 = routing stripped the hack cleanly (dominant
# teacher grad correctly quarantined); qE high + resid>0 = false
# negatives leaked hack-ward grad into the deployed knob (the real
# failure). vg is unit, so this is a plain cosine.
step_resid.append((g_keep @ vg / g_keep.norm().clamp_min(1e-12)).item())
return g_keep
# Split backward into student/teacher only every cos_pre_split_every steps.
# On split steps: 2 backwards per prompt, populates step_grad_s/_t.
@@ -1842,6 +1851,7 @@ def main(cfg: Config) -> int:
"q_egy": q_egy,
"tau": (sum(step_tau) / len(step_tau)) if step_tau else float("nan"),
"hkgap": (sum(step_hkgap) / len(step_hkgap)) if step_hkgap else float("nan"),
"resid": (sum(step_resid) / len(step_resid)) if step_resid else float("nan"),
"lr": sched.get_last_lr()[0],
"cos_pre": diag["mean_cos_pre"],
"cos_pre_s": diag["mean_cos_pre_s"],