RESULT 47+48: 涨的量=blocks8-11单模态增益(ρ复刻探针,两血统同构型90%质量,plain 205-210k过1,cent全程平)+b8 logit跑飞70→94; BBP审计=fp32模拟器无加性下界(a=0,overlap 1.000@3e-4)→BBP是硬件设计方程; 三探针入库

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

1 files changed, 130 insertions, 0 deletions

diff --git a/ep_run/probe_statej.py b/ep_run/probe_statej.py
new file mode 100644
index 0000000..f25325b
--- /dev/null
+++ b/ep_run/probe_statej.py
@@ -0,0 +1,130 @@
+"""State-side Jacobian audit: per-block sigma(dO_l/dz_in) at the FREE operating state,
+both lineages, matched steps. Weight-space audit (probe_specaudit) showed all weight scales
+FLAT while the beta-normalized loop gain G=0.9/beta_cap grew ~300x -> hypothesis: the growth
+is the CHAIN PRODUCT of per-block state Jacobians (each +tens-of-%, ^12 = hundreds-x).
+Also dumps per-block max|logit| at the state (softcap calibration).
+FD-JVP + autograd-vjp power iteration on J^T J (no forward-mode through SDPA)."""
+import argparse, pickle
+import numpy as np, torch, torch.nn as nn, torch.nn.functional as F
+from pathlib import Path
+
+ap = argparse.ArgumentParser()
+ap.add_argument('--ckpts', default=('plain:150000,plain:185000,plain:195000,plain:200000,'
+                                    'plain:210000,plain:230000,cent:150000,cent:185000,'
+                                    'cent:195000,cent:200000,cent:210000,cent:230000'))
+ap.add_argument('--iters', type=int, default=25)
+a = ap.parse_args()
+dev = 'cuda'
+torch.manual_seed(7)
+B, T = 8, 256
+
+class RMSNorm(nn.Module):
+    def __init__(self, C, eps=1e-6):
+        super().__init__(); self.g = nn.Parameter(torch.ones(C)); self.eps = eps
+    def forward(self, x):
+        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps) * self.g
+
+class SwiGLU(nn.Module):
+    def __init__(self, C):
+        super().__init__()
+        h = ((8 * C // 3) + 63) // 64 * 64
+        self.w1 = nn.Linear(C, h, bias=False); self.w3 = nn.Linear(C, h, bias=False)
+        self.w2 = nn.Linear(h, C, bias=False)
+    def forward(self, x):
+        return self.w2(F.silu(self.w1(x)) * self.w3(x))
+
+class Olmo2Attn(nn.Module):
+    def __init__(self, C, H, T):
+        super().__init__()
+        self.H, self.hd = H, C // H
+        self.qkv = nn.Linear(C, 3 * C, bias=False); self.proj = nn.Linear(C, C, bias=False)
+        self.qn, self.kn = RMSNorm(C), RMSNorm(C)
+        inv = 1.0 / (500000.0 ** (torch.arange(0, self.hd, 2).float() / self.hd))
+        fr = torch.outer(torch.arange(T).float(), inv)
+        self.register_buffer('rc', fr.cos(), persistent=False)
+        self.register_buffer('rs', fr.sin(), persistent=False)
+    def rope(self, x):
+        Tn = x.shape[2]
+        x1, x2 = x[..., ::2], x[..., 1::2]
+        c, s = self.rc[None, None, :Tn].to(x.dtype), self.rs[None, None, :Tn].to(x.dtype)
+        return torch.stack((x1 * c - x2 * s, x1 * s + x2 * c), dim=-1).flatten(-2)
+    def logits_stats(self, x):
+        Bn, Tn, C = x.shape
+        q, k, _ = self.qkv(x).split(C, dim=2)
+        q, k = self.qn(q), self.kn(k)
+        q = self.rope(q.view(Bn, Tn, self.H, self.hd).transpose(1, 2))
+        k = self.rope(k.view(Bn, Tn, self.H, self.hd).transpose(1, 2))
+        lg = (q @ k.transpose(-2, -1)) / (self.hd ** 0.5)
+        mask = torch.ones(Tn, Tn, dtype=torch.bool, device=x.device).tril()
+        lg = lg.masked_fill(~mask, 0.0)
+        return float(lg.abs().max()), float(lg.abs().quantile(0.999))
+    def forward(self, x):
+        Bn, Tn, C = x.shape
+        q, k, v = self.qkv(x).split(C, dim=2)
+        q, k = self.qn(q), self.kn(k)
+        q = self.rope(q.view(Bn, Tn, self.H, self.hd).transpose(1, 2))
+        k = self.rope(k.view(Bn, Tn, self.H, self.hd).transpose(1, 2))
+        v = v.view(Bn, Tn, self.H, self.hd).transpose(1, 2)
+        y = F.scaled_dot_product_attention(q, k, v, is_causal=True)
+        return self.proj(y.transpose(1, 2).contiguous().view(Bn, Tn, C))
+
+class Olmo2Block(nn.Module):
+    def __init__(self, C, H, T):
+        super().__init__()
+        self.attn = Olmo2Attn(C, H, T); self.ff = SwiGLU(C)
+        self.na, self.nf = RMSNorm(C), RMSNorm(C)
+    def forward(self, z):
+        z = z + self.na(self.attn(z))
+        return z + self.nf(self.ff(z))
+
+def sigma_J(block, z0):
+    """largest singular value of dblock(z)/dz at z0: power iteration on J^T J.
+    Jv by central FD (fp32, scaled eps), J^T u by autograd."""
+    z0 = z0.detach()
+    v = torch.randn_like(z0); v /= v.norm()
+    s = None
+    for _ in range(a.iters):
+        eps = 1e-3 * z0.norm() / (v.norm() * (z0.numel() ** 0.5) + 1e-30) * (z0.numel() ** 0.5)
+        with torch.no_grad():
+            jv = (block(z0 + eps * v) - block(z0 - eps * v)) / (2 * eps)
+        zg = z0.clone().requires_grad_(True)
+        o = block(zg)
+        jtu = torch.autograd.grad((o * jv.detach()).sum(), zg)[0]
+        s = float(jtu.norm().sqrt())          # |J^T J v|^{1/2} -> sigma as v converges
+        v = jtu / (jtu.norm() + 1e-30)
+    return s
+
+first = True
+for spec in a.ckpts.split(','):
+    lineage, step = spec.split(':'); step = int(step)
+    p = f'runs/fw72m_{lineage}_s{step}.pt'
+    try:
+        ck = torch.load(p, map_location='cpu', weights_only=False)
+    except FileNotFoundError:
+        print(f'{lineage} s{step}: MISSING', flush=True); continue
+    cfg = ck['config']; C, H, L = cfg['C'], cfg['H'], cfg['L']
+    if first:
+        DD = Path('/home/yurenh2/ept/ep_run/data') / cfg.get('data', 'fineweb_edu')
+        vocab = pickle.load(open(DD / 'meta.pkl', 'rb'))['vocab_size']
+        data = np.memmap(DD / 'val.bin', dtype=np.uint16, mode='r')
+        ix = torch.randint(len(data) - T - 1, (B,))
+        x = torch.stack([torch.from_numpy(data[i:i + T].astype(np.int64)) for i in ix]).to(dev)
+        first = False
+    tok = nn.Embedding(vocab, C).to(dev); tok.load_state_dict(ck['tok'])
+    blocks = nn.ModuleList([Olmo2Block(C, H, T) for _ in range(L)]).to(dev)
+    blocks.load_state_dict(ck['blocks'], strict=False)
+    with torch.no_grad():
+        zs = [tok(x)]
+        for b in blocks: zs.append(b(zs[-1]))
+    sigs, lgs = [], []
+    for l in range(L):
+        sigs.append(sigma_J(blocks[l], zs[l]))
+        lgs.append(blocks[l].attn.logits_stats(zs[l]))
+    prod = float(np.prod(sigs))
+    top = float(np.prod(sigs[6:]))
+    print(f'{lineage} s{step//1000}k | sig_J per blk ' + ' '.join(f'{s:5.2f}' for s in sigs) +
+          f' | PROD {prod:9.1f} top6 {top:7.1f} | max|logit| ' +
+          ' '.join(f'{m:4.1f}' for m, _ in lgs), flush=True)
+    del tok, blocks, zs
+    torch.cuda.empty_cache()
+print('STATEJ_DONE', flush=True)