File size: 14,879 Bytes
a8b3acb
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
#!/usr/bin/env python3
"""
Review Residuals --- training and evaluation for the paper
"An Update-Conditioned Residual Gate Whose Advantage Emerges at Scale" (Kramer, 2026).

Trains, from scratch on TinyStories, three identity-preserving (additive) residual variants:
  - review_neutral : update scaled by a gate conditioned on BOTH state and proposed update  (ours)
  - highway        : update scaled by a gate conditioned on the state only  (param-matched up)
  - standard       : plain residual, update added with coefficient 1        (param-matched up)
across five model sizes (60M-1B). Resumable; writes per-run validation losses to scaling_v8.csv.
Requires a CUDA GPU (80GB for the 590M/1B sizes). See README.md to reproduce.
"""

#!/usr/bin/env python3
# Review Residuals scaling sweep — RESILIENT, RESUMABLE, disconnect-proof.
# Run on the pod:   nohup python run_scaling.py > scaling.log 2>&1 &
# Watch progress:   tail -f scaling.log
# It resumes automatically: any run already in scaling_partial.csv is skipped.
# Each run is launched in its OWN subprocess, so even a hard crash on one model
# only loses that one model — the sweep keeps going.

import os, sys, csv, time, math, subprocess
os.environ["PYTORCH_CUDA_ALLOC_CONF"]="expandable_segments:True"
CSV="scaling_v8.csv"

# ---- the scaling ladder (identical to the notebook) ----
SIZES=[
  dict(name="1B",   d=1536, L=24, h=16, steps=6000, seeds=[0,1,2]),
  dict(name="590M", d=1280, L=20, h=20, steps=7000, seeds=[0,1,2]),
  dict(name="320M", d=1024, L=16, h=16, steps=8000, seeds=[0,1]),
  dict(name="150M", d=768,  L=12, h=12, steps=8000, seeds=[0,1,2]),
  dict(name="60M",  d=512,  L=8,  h=8,  steps=8000, seeds=[0,1,2]),
]
VARIANTS=["review_neutral","highway","standard"]
# AttnRes is ~8x slower than the others; only run it at the small sizes as anchors.
ATTNRES_SIZES=[]   # no attnres in the recipe-fixed sweep
BLOCK=256; BATCH=64; LR=2e-4; WARMUP=500; N_TEXT=400000   # lower LR + warmup (fixes large-scale divergence)
FIELDS=["size","variant","seed","params_M","steps","val_loss","ece","minutes"]

def done_set():
    s=set()
    if os.path.exists(CSV):
        with open(CSV) as f:
            for r in csv.DictReader(f):
                s.add((r["size"],r["variant"],int(r["seed"])))
    return s

def append_row(row):
    new = not os.path.exists(CSV)
    with open(CSV,"a",newline="") as f:
        w=csv.DictWriter(f,fieldnames=FIELDS)
        if new: w.writeheader()
        w.writerow(row)

# ============================================================ ORCHESTRATOR
def orchestrate():
    print("[orch] installing deps...",flush=True)
    subprocess.run([sys.executable,"-m","pip","install","-q","datasets","transformers","accelerate","matplotlib","pandas"])
    done=done_set()
    plan=[]
    for SZ in SIZES:
        for v in VARIANTS:
            if v=="attnres_plus":
                if SZ["name"] not in ATTNRES_SIZES: continue
                seeds = SZ["seeds"] if SZ["name"]=="60M" else [0]
            else:
                seeds = SZ["seeds"]
            for sd in seeds: plan.append((SZ,v,sd))
    todo=[(SZ,v,sd) for (SZ,v,sd) in plan if (SZ["name"],v,sd) not in done]
    print(f"[orch] {len(done)} runs already done, {len(todo)} to go",flush=True)
    t0=time.time()
    for SZ,v,sd in todo:
        tag=f"{SZ['name']}/{v}/seed{sd}"
        print(f"\n[orch] === launching {tag} (elapsed {(time.time()-t0)/3600:.2f}h) ===",flush=True)
        # isolate each run in a fresh process: a hard crash here cannot kill the sweep
        rc=subprocess.call([sys.executable, os.path.abspath(__file__), "--worker", SZ["name"], v, str(sd)])
        if rc!=0:
            print(f"[orch] !! {tag} exited with code {rc} (logged as failure, continuing)",flush=True)
        else:
            print(f"[orch] ok {tag}",flush=True)
    print(f"\n[orch] SWEEP COMPLETE in {(time.time()-t0)/3600:.2f}h. Results in {CSV}.",flush=True)
    try:
        make_plot()
    except Exception as e:
        print("[orch] plot skipped:",e,flush=True)

# ============================================================ WORKER (one run)
def worker(size_name,variant,seed):
    import math, numpy as np, torch, torch.nn as nn, torch.nn.functional as F
    from datasets import load_dataset; from transformers import GPT2TokenizerFast
    torch.set_float32_matmul_precision("high"); torch.backends.cuda.matmul.allow_tf32=True; torch.backends.cudnn.allow_tf32=True
    device="cuda" if torch.cuda.is_available() else "cpu"; assert device=="cuda","need GPU"
    SZ=[s for s in SIZES if s["name"]==size_name][0]
    tok=GPT2TokenizerFast.from_pretrained("gpt2"); VOCAB=tok.vocab_size

    class RMSNorm(nn.Module):
        def __init__(s,d): super().__init__(); s.g=nn.Parameter(torch.ones(d))
        def forward(s,x): return x*torch.rsqrt(x.pow(2).mean(-1,keepdim=True)+1e-5)*s.g
    class Attn(nn.Module):
        def __init__(s,d,h,block): super().__init__(); s.h=h; s.qkv=nn.Linear(d,3*d); s.proj=nn.Linear(d,d)
        def forward(s,x):
            B,T,d=x.shape; q,k,v=s.qkv(x).split(d,2)
            q=q.view(B,T,s.h,d//s.h).transpose(1,2); k=k.view(B,T,s.h,d//s.h).transpose(1,2); v=v.view(B,T,s.h,d//s.h).transpose(1,2)
            return s.proj(F.scaled_dot_product_attention(q,k,v,is_causal=True).transpose(1,2).reshape(B,T,d))
    class MLP(nn.Module):
        def __init__(s,d): super().__init__(); s.f1=nn.Linear(d,4*d); s.f2=nn.Linear(4*d,d)
        def forward(s,x): return s.f2(F.gelu(s.f1(x)))
    def is_attnres(v): return v in ("attnres","attnres_plus")
    class GPT(nn.Module):
        def __init__(s,variant,d,n_layer,n_head,block,vocab):
            super().__init__(); s.variant=variant
            s.tok=nn.Embedding(vocab,d); s.pos=nn.Embedding(block,d); s.norms=nn.ModuleList(); s.subs=nn.ModuleList()
            for i in range(2*n_layer):
                s.norms.append(RMSNorm(d)); s.subs.append(Attn(d,n_head,block) if i%2==0 else MLP(d))
            nS=2*n_layer
            if variant=="highway": s.gate=nn.ModuleList([nn.Linear(d,d) for _ in range(nS)])
            if variant=="review_neutral":
                s.rgate=nn.ModuleList([nn.Linear(2*d,d) for _ in range(nS)])
                for g in s.rgate: nn.init.zeros_(g.weight); nn.init.zeros_(g.bias)
            if variant=="layerscale": s.ls=nn.ParameterList([nn.Parameter(torch.ones(d)*0.1) for _ in range(nS)])
            if variant=="rezero":     s.rez=nn.ParameterList([nn.Parameter(torch.zeros(1)) for _ in range(nS)])
            if is_attnres(variant): s.dq=nn.Parameter(torch.randn(nS+1,d)*0.02); s.dk=nn.Linear(d,d,bias=False)
            s.lnf=RMSNorm(d); s.head=nn.Linear(d,vocab,bias=False); s.head.weight=s.tok.weight
            # --- GPT-2 / nanoGPT initialization (stable deep training) ---
            def _gpt2(mod):
                if isinstance(mod,nn.Linear):
                    nn.init.normal_(mod.weight,mean=0.0,std=0.02)
                    if mod.bias is not None: nn.init.zeros_(mod.bias)
                elif isinstance(mod,nn.Embedding):
                    nn.init.normal_(mod.weight,mean=0.0,std=0.02)
            s.apply(_gpt2)
            # scale residual-projection outputs by 1/sqrt(2*n_layer)  <-- the key deep-stability fix
            for _n,_p in s.named_parameters():
                if _n.endswith("proj.weight") or _n.endswith("f2.weight"):
                    nn.init.normal_(_p,mean=0.0,std=0.02/math.sqrt(2*n_layer))
            # keep the review gate neutral (must stay zero for r=0.5 start)
            if variant=="review_neutral":
                for g in s.rgate: nn.init.zeros_(g.weight); nn.init.zeros_(g.bias)
        def _rms(s,x): return x*torch.rsqrt(x.pow(2).mean(-1,keepdim=True)+1e-5)
        def _depth_attn(s,M,qi):
            K=s._rms(s.dk(M)); a=(K*qi.view(1,1,1,-1)).sum(-1).softmax(-1).unsqueeze(-1); return (a*M).sum(2)
        def forward(s,idx,targets=None):
            B,T=idx.shape; x0=s.tok(idx)+s.pos(torch.arange(T,device=idx.device))[None]
            if is_attnres(s.variant):
                mem=[x0]
                for i,(nrm,sub) in enumerate(zip(s.norms,s.subs)):
                    mem.append(sub(nrm(s._depth_attn(torch.stack(mem,2),s.dq[i]))))
                h=s._depth_attn(torch.stack(mem,2),s.dq[-1])
            else:
                h=x0
                for i,(nrm,sub) in enumerate(zip(s.norms,s.subs)):
                    u=sub(nrm(h))
                    if s.variant=="highway": g=torch.sigmoid(s.gate[i](h)); h=h+g*u            # additive (identity preserved)
                    elif s.variant=="review_neutral": r=torch.sigmoid(s.rgate[i](torch.cat([s._rms(h),s._rms(u)],-1))); h=h+r*u   # additive (identity preserved)
                    elif s.variant=="layerscale": h=h+s.ls[i]*u
                    elif s.variant=="rezero": h=h+s.rez[i]*u
                    else: h=h+u
            logits=s.head(s.lnf(h))
            loss=F.cross_entropy(logits.view(-1,logits.size(-1)),targets.view(-1)) if targets is not None else None
            return logits,loss

    def attnres_width(SZ):
        def est(var,d,L,h):
            nS=2*L; p=VOCAB*d+BLOCK*d+nS*d+L*(4*d*d+4*d)+L*(8*d*d+5*d)+d
            if var=="review_neutral": p+=nS*(2*d*d+d)
            elif var=="attnres_plus": p+=(nS+1)*d+d*d
            return p
        base=est("review_neutral",SZ["d"],SZ["L"],SZ["h"]); d=SZ["d"]
        while est("attnres_plus",d,SZ["L"],SZ["h"])<base: d+=SZ["h"]
        return d
    def highway_width(SZ):
        def est(var,d,L,h):
            nS=2*L; p=VOCAB*d+BLOCK*d+nS*d+L*(4*d*d+4*d)+L*(8*d*d+5*d)+d
            if var=="review_neutral": p+=nS*(2*d*d+d)
            elif var=="highway": p+=nS*(d*d+d)
            return p
        base=est("review_neutral",SZ["d"],SZ["L"],SZ["h"]); d=SZ["d"]
        while est("highway",d,SZ["L"],SZ["h"])<base: d+=SZ["h"]
        return d
    def standard_width(SZ):
        def est(var,d,L,h):
            nS=2*L; p=VOCAB*d+BLOCK*d+nS*d+L*(4*d*d+4*d)+L*(8*d*d+5*d)+d
            if var=="review_neutral": p+=nS*(2*d*d+d)
            return p
        base=est("review_neutral",SZ["d"],SZ["L"],SZ["h"]); d=SZ["d"]
        while est("standard",d,SZ["L"],SZ["h"])<base: d+=SZ["h"]
        return d

    def load_data():
        texts=load_dataset("roneneldan/TinyStories", split=f"train[:{N_TEXT}]")["text"]
        ids=[]
        for i in range(0,len(texts),2000):
            for e in tok(texts[i:i+2000])["input_ids"]: ids.extend(e); ids.append(tok.eos_token_id)
        data=np.array(ids,dtype=np.uint16); sp=int(len(data)*0.97)
        return torch.from_numpy(data[:sp].astype(np.int64)), torch.from_numpy(data[sp:].astype(np.int64))
    def get_batch(t,B,T):
        ix=np.random.randint(0,len(t)-T-1,size=B)
        x=torch.stack([t[i:i+T] for i in ix]); y=torch.stack([t[i+1:i+1+T] for i in ix])
        return x.to(device,non_blocking=True), y.to(device,non_blocking=True)
    @torch.no_grad()
    def evaluate(model,val_t,B,n=80):
        model.eval(); L=[]; C=[]; K=[]
        for _ in range(n):
            x,y=get_batch(val_t,B,BLOCK); lo,l=model(x,y); L.append(l.item())
            p=lo.softmax(-1); c,pr=p.max(-1); C.append(c.flatten().cpu().numpy()); K.append((pr==y).flatten().cpu().numpy())
        C=np.concatenate(C); K=np.concatenate(K).astype(float); e=np.linspace(0,1,16); ece=0
        for i in range(15):
            m=(C>e[i])&(C<=e[i+1])
            if m.sum(): ece+=m.sum()/len(C)*abs(K[m].mean()-C[m].mean())
        return float(np.mean(L)),float(ece)

    print(f"[worker] {size_name} {variant} seed{seed} on {torch.cuda.get_device_name(0)}",flush=True)
    train_t,val_t=load_data()
    d = highway_width(SZ) if variant=="highway" else (standard_width(SZ) if variant=="standard" else SZ["d"])  # param-match baselines UP
    batch=BATCH
    for attempt in range(4):
        try:
            torch.manual_seed(seed); np.random.seed(seed)
            m=GPT(variant,d,SZ["L"],SZ["h"],BLOCK,VOCAB).to(device)
            P=sum(p.numel() for p in m.parameters())/1e6
            opt=torch.optim.AdamW(m.parameters(),lr=LR,weight_decay=0.1,betas=(0.9,0.95))
            def _lrlam(step, T=SZ["steps"], W=WARMUP):
                if step < W: return (step+1)/W                       # linear warmup
                prog=(step-W)/max(1,(T-W)); return 0.1+0.9*0.5*(1+math.cos(math.pi*min(1.0,prog)))  # cosine to 10%
            sch=torch.optim.lr_scheduler.LambdaLR(opt,_lrlam); t0=time.time()
            for step in range(SZ["steps"]):
                x,y=get_batch(train_t,batch,BLOCK); _,loss=m(x,y)
                opt.zero_grad(); loss.backward(); torch.nn.utils.clip_grad_norm_(m.parameters(),1.0); opt.step(); sch.step()
                if step%1000==0: print(f"    step {step}/{SZ['steps']} loss {loss.item():.3f}",flush=True)
            vl,ece=evaluate(m,val_t,batch); mins=(time.time()-t0)/60
            append_row(dict(size=size_name,variant=variant,seed=seed,params_M=round(P,2),
                            steps=SZ["steps"],val_loss=round(vl,4),ece=round(ece,4),minutes=round(mins,1)))
            print(f"[worker] DONE {size_name} {variant} seed{seed} {P:.1f}M val {vl:.4f} ece {ece:.4f} {mins:.1f}min",flush=True)
            return 0
        except RuntimeError as ex:
            if "out of memory" in str(ex).lower() and batch>8:
                torch.cuda.empty_cache(); batch//=2
                print(f"[worker] OOM -> retry at batch {batch}",flush=True)
            else:
                raise
    return 1

# ============================================================ PLOT
def make_plot():
    import pandas as pd, matplotlib; matplotlib.use("Agg"); import matplotlib.pyplot as plt
    df=pd.read_csv(CSV)
    df=df[df["val_loss"]<4.5]   # drop diverged runs from the plot
    agg=df.groupby(["size","variant"]).agg(params_M=("params_M","mean"),val_loss=("val_loss","mean")).reset_index()
    order=[s["name"] for s in SIZES]; agg["o"]=agg["size"].map({n:i for i,n in enumerate(order)}); agg=agg.sort_values("o")
    col={'review_neutral':'#27ae60','attnres_plus':'#8e44ad','highway':'#999'}
    mk={'review_neutral':'o','attnres_plus':'s','highway':'^'}
    fig,ax=plt.subplots(figsize=(8,5.4))
    for v in VARIANTS:
        s=agg[agg.variant==v].sort_values("params_M")
        if len(s): ax.plot(s["params_M"],s["val_loss"],marker=mk[v],color=col[v],lw=2,ms=8,label=v)
    ax.set_xscale("log"); ax.set_xlabel("parameters (millions, log scale)")
    ax.set_ylabel("validation loss  (lower = better)")
    ax.set_title("Review Residuals scaling — loss vs parameters (TinyStories)")
    ax.grid(alpha=.3,which="both"); ax.legend()
    plt.tight_layout(); plt.savefig("scaling_result.png",dpi=140)
    print("[plot] wrote scaling_result.png",flush=True)

if __name__=="__main__":
    if len(sys.argv)>1 and sys.argv[1]=="--worker":
        sys.exit(worker(sys.argv[2], sys.argv[3], int(sys.argv[4])))
    else:
        orchestrate()