dgain几何profile + M1泄漏谱仪(用户设计,07-30)

--dgain_geo r: 每层读出位移增益=r^l(首层恒1,不含总深度L,加深自动延伸); --dgain_geo_cap封顶。
用户提案"首层不变逐层x2"的实装,三臂电池(2^l | 2^l cap128 | 1.414^l)已排队于仪器ckpt。

--probe_dgspec N: 逐块顽固度谱仪。L_l(g)=mean_b[gEP_l(g)-gBP_l]同批配对差分(batch采样噪声
精确对消,残余=一致性偏差+估计器噪声/sqrtN); 主指标=在BP方向的带符号投影(范数在1e7维被噪声
膨胀,冒烟实测g8读700%,投影无此病且恰为致害分量); 附(-beta,g=1)臂分离奇偶(beta奇函数FD偏差
翻号,位移阈值泄漏为|位移|偶函数不翻); 同时记录每块自由相位移比RMS(d_l)/RMS(z_l)=无量纲律
候选横轴。fp32仪器级(bf16地板致盲,b15教训)。GOV['dgprofile']钩子仅探针置位,训练路径不触。

排队: dgspec C768(泄漏态s250000)+C512对照(应读~0) on GPU1, ladder后补扫C320/C192/C128
=5宽度x12块60点做g*_l x 位移比塌缩检验。

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014FAPDWQ49M5Ye3NpTndTpn

1 files changed, 86 insertions, 2 deletions

diff --git a/ep_run/casc_eq_train.py b/ep_run/casc_eq_train.py
index 0e357c8..a9218a0 100644
--- a/ep_run/casc_eq_train.py
+++ b/ep_run/casc_eq_train.py
@@ -49,7 +49,12 @@ ap.add_argument('--bsign_rand', action='store_true')  # random-sign beta per ste
 ap.add_argument('--bf16', action='store_true')           # cast model to bf16 (E-accumulation + tok_sigma stay fp32) — the x0.5 cost lever, GATE before production
 ap.add_argument('--amp', action='store_true')            # PROPER mixed precision: autocast(bf16) matmuls, fp32 params/states/d/E — amp_gate.py PASSED 2026-07-12 (cos 0.9682 vs fp32 0.9687); --bf16 naive-cast stays DEAD (state quantization, RESULT 11)
 ap.add_argument('--dtop_every', type=int, default=1)    # 1 = exact (DEFAULT, BP-parity); 2 = fast mode (~20% cheaper, ~4% CE tax at high lr)
-ap.add_argument('--dgain_rand', type=float, default=0.0)  # >1: per-STEP log-uniform dgain_top in
+ap.add_argument('--dgain_geo', type=float, default=0.0)   # >0: per-layer geometric read-displacement
+                                                          # gain = geo^l (layer 0 = x1), optionally capped
+ap.add_argument('--dgain_geo_cap', type=float, default=0.0) # >0: cap for the geometric profile
+ap.add_argument('--dgain_rand', type=float, default=0.0)
+ap.add_argument('--probe_dgspec', type=int, default=0)      # >0: M1 spectroscopy, value = n batches; exits before training
+ap.add_argument('--probe_gains', default='1,2,4,8,16,32,64,128,256')  # >1: per-STEP log-uniform dgain_top in
                                                           # [1, this] (spread-spectrum probing of the
                                                           # decade-spread threshold distribution)
 ap.add_argument('--dgain_top', type=float, default=1.0)  # amplify d in STATE FORMATION for blocks
@@ -422,7 +427,14 @@ def relax(z0, zs, ins, outs, y, beta, K, x, bmask=None):
                 else:
                     i = prev.detach().requires_grad_(True)
                 o = blocks[l](i, mask)
-                _dg = args.dgain_all * ((GOV.get('dgcur') or args.dgain_top) if l >= args.L // 2 else 1.0)
+                _dgp = GOV.get('dgprofile')   # probe-only per-layer profile; never set in training
+                if _dgp is not None:
+                    _dg = _dgp.get(l, 1.0)
+                elif args.dgain_geo > 0:
+                    _dg = args.dgain_geo ** l
+                    if args.dgain_geo_cap > 0: _dg = min(_dg, args.dgain_geo_cap)
+                else:
+                    _dg = args.dgain_all * ((GOV.get('dgcur') or args.dgain_top) if l >= args.L // 2 else 1.0)
                 znew = o.detach().float() + (_dg * d[l] if _dg != 1.0 else d[l])
                 # damped (under-relaxed) mixing: geta<1 restores contraction on stiff operators
                 # (wall-2 toolkit); fixed point unchanged (z = z + geta*(o+d-z) <=> z = o+d)
@@ -835,6 +847,78 @@ if args.ddp_grad_test:
     import sys
     sys.exit(0)
 
+if args.probe_dgspec > 0:
+    # M1 DGAIN SPECTROSCOPY: per-block leak vector vs uniform read-displacement gain.
+    # Paired design: L_l(g) = mean_b[gEP_l(g) - gBP_l] on the SAME batch — batch-sampling
+    # noise cancels exactly in the difference; what survives is the consistent bias (= the
+    # leak) plus estimator-internal noise shrinking 1/sqrt(NB). Extra (-beta, g=1) arm
+    # splits odd-in-beta FD bias from even-in-|displacement| leak. Run WITHOUT --amp: the
+    # bf16 floor is width-blind (b15 lesson); fp32 is the instrument-grade path.
+    assert WORLD == 1, 'probe_dgspec is single-GPU'
+    import json, sys
+    gains = [float(t) for t in args.probe_gains.split(',')]
+    cfgs = [(f'g{g:g}', g, +1.0) for g in gains] + [('g1neg', 1.0, -1.0)]
+    NB = args.probe_dgspec
+    bix = []
+    for blk in blocks:
+        ids = {id(q) for q in blk.parameters()}
+        bix.append([i for i, p in enumerate(all_params) if id(p) in ids])
+    accD = {k: [torch.zeros(sum(all_params[i].numel() for i in ix), device=dev) for ix in bix]
+            for k, _, _ in cfgs}
+    accO = {k: [torch.zeros_like(t) for t in accD[k]] for k, _, _ in cfgs}   # odd-batch half (jackknife)
+    accB = [torch.zeros_like(t) for t in accD[cfgs[0][0]]]
+    disp = [0.0] * args.L                                                   # RMS(d_l)/RMS(z_l) at g=1,+beta
+    def _flat(gl, ix):
+        return torch.cat([(gl[i] if gl[i] is not None else torch.zeros_like(all_params[i]))
+                          .reshape(-1).float() for i in ix])
+    for bi_ in range(NB):
+        x, y = get_batch('train')
+        gbp = bp_gate(x, y)
+        bpfl = [_flat(gbp, ix) for ix in bix]
+        for j, t in enumerate(bpfl): accB[j] += t
+        for key, gA, sgn in cfgs:
+            GOV['dgprofile'] = {l: gA for l in range(args.L)}
+            bt = sgn * args.beta
+            z0, zs, ins, outs = free_states_graphed(x)
+            zp, lo = relax(z0, zs, ins, outs, y, bt, args.K, x)
+            if key == 'g1' :
+                with torch.no_grad():
+                    for l in range(args.L):
+                        disp[l] += float(GOV['_last_d'][l].norm() / max(float(zp[l].norm()), 1e-12)) / NB
+            E = 0.0
+            for z, o in zip(zp, lo): E = E + 0.5 * ((z.detach().float() - o.float()) ** 2).sum()
+            obj = E / (NBT * bt) + obj_loss(readout(zp[-1].detach()).reshape(-1, vocab), y.reshape(-1))
+            gs = torch.autograd.grad(obj, all_params, allow_unused=True)
+            for j, ix in enumerate(bix):
+                d_ = _flat(gs, ix) - bpfl[j]
+                accD[key][j] += d_
+                if bi_ % 2 == 1: accO[key][j] += d_
+        GOV['dgprofile'] = None
+        if (bi_ + 1) % 16 == 0: print(f'[dgspec] batch {bi_+1}/{NB}', flush=True)
+    out = {'disp_ratio': disp}
+    for key, gA, sgn in cfgs:
+        rows = []
+        for j in range(args.L):
+            mB, mD = accB[j] / NB, accD[key][j] / NB
+            ne, no_ = NB - NB // 2, NB // 2
+            mDe, mDo = (accD[key][j] - accO[key][j]) / ne, accO[key][j] / max(no_, 1)
+            rows.append({'block': j,
+                         'rel_leak': float(mD.norm() / max(float(mB.norm()), 1e-12)),
+                         'proj_on_bp': float((mD @ mB) / max(float(mB.norm()) ** 2, 1e-24)),
+                         'half_split': float((mDe - mDo).norm() / max(float(mD.norm()), 1e-12))})
+        out[key] = rows
+        # proj_on_bp is PRIMARY: the norm metric carries a positive noise-dimension bias
+        # (‖mean‖ >= ‖bias‖ inflated by residual noise over ~1e7 coords), the signed projection
+        # is unbiased and is exactly the damaging component; expect proj<0 (missing response).
+        print(f'[dgspec] {key:>7}: proj ' + ' '.join(f'{r["proj_on_bp"]:+.3f}' for r in rows)
+              + ' | rel ' + ' '.join(f'{r["rel_leak"]:.3f}' for r in rows), flush=True)
+    print('[dgspec] disp_ratio ' + ' '.join(f'{v:.2e}' for v in disp), flush=True)
+    Path('runs').mkdir(exist_ok=True)
+    with open(f'runs/dgspec_{args.tag}.json', 'w') as f:
+        json.dump({'args': {k: str(v) for k, v in vars(args).items()}, 'NB': NB, 'curves': out}, f)
+    print(f'[dgspec] DONE -> runs/dgspec_{args.tag}.json', flush=True)
+    sys.exit(0)
+
 best, t0 = 1e9, time.time()
 skips = 0
 for _ in range(start_step): sched.step()   # advance LR schedule to the resumed step