File size: 12,311 Bytes
224edbf 6f7a03c 224edbf 6f7a03c 224edbf 6f7a03c 224edbf 6f7a03c 224edbf 6f7a03c 224edbf 6f7a03c 224edbf | 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 | # train_peer_v21.py - pretrain the v21 shared-pool ternary+PEER controller.
# Same data/codon/Muon/WSD machinery as train_peer.py, with the v21 fixes:
# P1 3 shared PEER layers P2 embedding_bag P4 SparseAdam on the pool
# P5 checkpoint ONLY the PEER blocks P6 grad-clip excludes the sparse pool
# Logs a throughput series to <out>/throughput.jsonl for the A/B charts.
import argparse, json, math, os, time, threading, queue
from pathlib import Path
import numpy as np, torch, torch.nn.functional as F
import torch.utils.checkpoint as cp
import sys; sys.path.insert(0, '/root/dna')
from model_dna_peer_v21 import DnaPeerV21, peer_counts, param_groups, compile_
ap = argparse.ArgumentParser()
ap.add_argument('--data-dir', default='/root/dna/data')
ap.add_argument('--out', default='/root/dna/ckpt-peer21'); ap.add_argument('--resume', default='/root/dna/ckpt-peer21/resume.pt')
ap.add_argument('--batch', type=int, default=16); ap.add_argument('--seq', type=int, default=512)
ap.add_argument('--steps', type=int, default=2000000); ap.add_argument('--target-tokens', type=float, default=3e9)
ap.add_argument('--max-hours', type=float, default=0.0, help='stop after N hours (0=off)')
ap.add_argument('--lr', type=float, default=3e-4); ap.add_argument('--muonlr', type=float, default=2e-2)
ap.add_argument('--poollr', type=float, default=0.02) # stateless SGD wants a larger lr
ap.add_argument('--nk', type=int, default=704); ap.add_argument('--topk', type=int, default=8)
ap.add_argument('--pheads', type=int, default=1)
ap.add_argument('--peer-layers', default='4,12') # 2 layers MEASURED 38.2k vs 3 layers 35.7k
ap.add_argument('--rowstep', type=float, default=1.0); ap.add_argument('--warmup', type=int, default=2000)
ap.add_argument('--ckpt-every', type=int, default=2000); ap.add_argument('--hf-every', type=int, default=20000)
ap.add_argument('--hf-repo', default='jaivial/dna-diskchat-2b-peer-v21'); ap.add_argument('--log-every', type=int, default=100)
ap.add_argument('--mtp-frac', type=float, default=0.25)
a = ap.parse_args(); dev = 'cuda'; Path(a.out).mkdir(parents=True, exist_ok=True); torch.manual_seed(1)
torch._dynamo.config.cache_size_limit = 64
torch.backends.cuda.matmul.allow_tf32 = True; torch.backends.cudnn.allow_tf32 = True; torch.backends.cudnn.benchmark = True
m = DnaPeerV21(nk=a.nk, topk=a.topk, pheads=a.pheads,
peer_layers=tuple(int(x) for x in a.peer_layers.split(',')),
chunk=int(os.environ.get('CHUNK_C', '16'))).to(dev)
muon_p, sparse_p, dense_p = param_groups(m)
# P5: compile + checkpoint only where it pays
compile_(m) # compile all but the sparse gathers (inductor cannot lower them)
PEER_SET = set(m.peer_layers)
ctrl, active = peer_counts(m)
CFG = {'ctrl_params': ctrl, 'active_params_per_token': active, 'experts': m.pool.E,
'peer_layers': list(m.peer_layers), 'h_times_k': a.pheads * a.topk, 'nk': a.nk,
'pool_params': sum(p.numel() for p in sparse_p), 'pool_dtype': str(m.pool.up.weight.dtype),
'codon_table_bytes': m.nrows * m.ncod, 'batch': a.batch, 'seq': a.seq, 'variant': 'v21-shared-pool'}
print('PEER21_CONFIG', json.dumps(CFG), flush=True)
def ns5(Gm, steps=int(os.environ.get('NS_STEPS', '3'))):
x = 3.4445; y = -4.7750; z = 2.0315; X = Gm.bfloat16(); tp = Gm.size(0) > Gm.size(1)
if tp: X = X.t()
X = X / (X.norm() + 1e-7)
for _ in range(steps):
A = X @ X.t(); B = y * A + z * (A @ A); X = x * X + B @ X
if tp: X = X.t()
return X.to(Gm.dtype)
class Muon(torch.optim.Optimizer):
def __init__(self, params, lr=2e-2, momentum=0.95): super().__init__(list(params), dict(lr=lr, momentum=momentum))
@torch.no_grad()
def step(self):
for grp in self.param_groups:
for p in grp['params']:
if p.grad is None: continue
st = self.state[p]
if 'm' not in st: st['m'] = torch.zeros_like(p.grad)
buf = st['m']; buf.mul_(grp['momentum']).add_(p.grad); u = p.grad.add(buf, alpha=grp['momentum'])
u = ns5(u); u = u.mul(max(1.0, p.size(0) / p.size(1)) ** 0.5); p.add_(u, alpha=-grp['lr'])
opt = torch.optim.AdamW(dense_p, lr=a.lr, betas=(.9, .95), weight_decay=.1, fused=True)
muon = Muon(muon_p, lr=a.muonlr)
# P4: STATELESS sparse SGD on the expert pool (v8 lever T5). SparseAdam was tried
# first and OOMed: it materializes dense exp_avg/exp_avg_sq (4.06 GB) plus a
# sparse_mask temporary, defeating the whole point. This keeps ZERO optimizer state.
@torch.no_grad()
def sparse_sgd(lr):
for p in sparse_p:
g = p.grad
if g is None: continue
if g.is_sparse:
g = g.coalesce()
p.data.index_add_(0, g.indices()[0], g.values(), alpha=-lr)
else:
p.data.add_(g, alpha=-lr)
p.grad = None
print('MUON', sum(p.numel() for p in muon_p), 'ADAM', sum(p.numel() for p in dense_p),
'SPARSE_POOL', sum(p.numel() for p in sparse_p), flush=True)
cur = None; curi = -1
def shard_list(): return sorted([p for p in Path(a.data_dir).glob('shard_*.u16') if not p.with_suffix('.tmp').exists()])
def batch():
global cur, curi
sh = shard_list()
if not sh: raise SystemExit('no shards yet - start stream_data.py first')
i = np.random.randint(0, len(sh))
if i != curi: cur = np.memmap(sh[i], np.uint16, 'r'); curi = i
n = len(cur) - a.seq - 1; ix = np.random.randint(0, n, size=a.batch)
return torch.from_numpy(np.stack([np.asarray(cur[j:j+a.seq+1], np.int64) for j in ix])).to(dev)
def cosf(s):
if s < a.warmup: return s / a.warmup
ds = int(a.steps * 0.8)
if s < ds: return 1.0
p = (s - ds) / max(1, a.steps - ds); return 1.0 - 0.9 * p
def feats(ids):
B, T = ids.shape; x = m.embed(ids)
for i, blk in enumerate(m.blocks):
# P5: recompute only the PEER blocks; the other 13 run without recompute
x = cp.checkpoint(blk, x, m.chunk, use_reentrant=False) if i in PEER_SET else blk(x, m.chunk)
mem, bal, rows = m.route(x.reshape(B * T, m.d)); return m.norm(x + mem.reshape(B, T, m.d)), bal, rows
def save(step):
tmp = a.resume + '.tmp'
torch.save({'model': m.state_dict(), 'opt': opt.state_dict(), 'muon': muon.state_dict(),
'step': step, 'config': m.config()}, tmp)
os.replace(tmp, a.resume)
def hf_push(step):
try:
from huggingface_hub import HfApi
api = HfApi(token=os.environ.get('HUGGING_FACE')); api.create_repo(a.hf_repo, exist_ok=True)
p = f'{a.out}/ckpt_{step}.pt'
torch.save({'model': m.state_dict(), 'config': m.config(), 'step': step}, p)
api.upload_file(path_or_fileobj=p, path_in_repo=f'checkpoints/ckpt_{step}.pt', repo_id=a.hf_repo)
os.remove(p)
tp = Path(a.out) / 'throughput.jsonl'
if tp.exists(): api.upload_file(path_or_fileobj=str(tp), path_in_repo='logs/throughput.jsonl', repo_id=a.hf_repo)
print(f'HF_PUSH step {step}', flush=True)
except Exception as e: print('HF_PUSH_FAIL', str(e)[:140], flush=True)
def chunked_ce(feat, W, tgt, nchunk=8):
"""Tiled cross-entropy (v12 lever A). Never materializes the full [N,V] logit
tensor; mathematically identical to F.cross_entropy(F.linear(feat,W),tgt).
MEASURED: peak VRAM 11.32 -> 10.56 GB at B20, which is what pays for B28."""
N = feat.size(0); tot = feat.new_zeros(())
for f, t in zip(feat.chunk(nchunk, 0), tgt.chunk(nchunk, 0)):
tot = tot + F.cross_entropy(F.linear(f, W), t, reduction='sum')
return tot / N
@torch.no_grad()
def valid_loss():
m.eval(); ids = batch()
with torch.autocast('cuda', dtype=torch.bfloat16):
feat, _, _ = feats(ids)
ce = chunked_ce(feat[:, :-1].reshape(-1, m.d), m.embed.weight, ids[:, 1:].reshape(-1))
m.train(); return ce.item()
start = 0
if os.path.exists(a.resume):
ck = torch.load(a.resume, map_location=dev, weights_only=False)
m.load_state_dict(ck['model']); opt.load_state_dict(ck['opt']); muon.load_state_dict(ck['muon'])
start = int(ck['step']); print('RESUME', start, flush=True)
_bq = queue.Queue(maxsize=4)
def _worker():
while True: _bq.put(batch())
for _ in range(2): threading.Thread(target=_worker, daemon=True).start()
def nextbatch(): return _bq.get()
# ---- throughput tracking (the A/B deliverable) ----
TP = Path(a.out) / 'throughput.jsonl'
tp_f = open(TP, 'a')
ts_hist = []
def log_tp(rec):
tp_f.write(json.dumps(rec) + '\n'); tp_f.flush()
m.train(); t0 = time.time(); wall0 = time.time(); seen = 0; step = start
print('LOOP start', start, 'steps', a.steps, flush=True)
for step in range(start + 1, a.steps + 1):
f = cosf(step)
for g in opt.param_groups: g['lr'] = a.lr * f
for g in muon.param_groups: g['lr'] = a.muonlr * f
ids = nextbatch()
opt.zero_grad(set_to_none=True)
for p in sparse_p: p.grad = None
for p in muon_p:
p.grad = None
with torch.autocast('cuda', dtype=torch.bfloat16):
feat, bal, rows = feats(ids)
fl = feat[:, :-1].reshape(-1, m.d); tg = ids[:, 1:].reshape(-1)
l1 = chunked_ce(fl, m.embed.weight, tg)
f2 = feat[:, :-2].reshape(-1, m.d); t2 = ids[:, 2:].reshape(-1)
sel = torch.randint(0, f2.size(0), (int(a.mtp_frac * f2.size(0)),), device=dev)
l2 = chunked_ce(f2[sel], m.head2.weight, t2[sel])
loss = l1 + 0.5 * l2 + 1e-3 * bal
loss.backward()
torch.nn.utils.clip_grad_norm_(muon_p + dense_p, 1.0) # P6: pool excluded
opt.step(); sparse_sgd(a.poollr * f)
if step % int(os.environ.get('MUON_EVERY', '2')) == 0: muon.step()
with torch.no_grad():
leaf = getattr(m, '_leaf', None)
if leaf is not None and leaf.grad is not None:
flat = rows.reshape(-1); g = leaf.grad
gn = g.norm(dim=-1, keepdim=True).clamp_min(1e-8); qstep = m.codebook.detach().std()
rstep = a.rowstep * qstep * f * (m.d ** 0.5); newrows = leaf.detach() - rstep * (g / gn)
uid, inv = torch.unique(flat, return_inverse=True)
acc = torch.zeros(uid.size(0), m.d, device=newrows.device, dtype=newrows.dtype); acc.index_copy_(0, inv, newrows)
if not hasattr(m, '_pu'): m._pu = []; m._pa = []
m._pu.append(uid); m._pa.append(acc)
if step % int(os.environ.get('MUT_EVERY', '16')) == 0:
pu = torch.cat(m._pu); pa = torch.cat(m._pa)
u2, inv2 = torch.unique(pu, return_inverse=True)
a2 = torch.zeros(u2.size(0), m.d, device=pa.device, dtype=pa.dtype); a2.index_copy_(0, inv2, pa)
before = m.codons[u2].clone(); m.mutate_rows(u2, a2)
m._mut = (m.codons[u2] != before).float().mean().item(); m._pu = []; m._pa = []
m._leaf = None
seen += a.batch * (a.seq - 1); toks = step * a.batch * (a.seq - 1)
if step % a.log_every == 0:
torch.cuda.synchronize(); dt = time.time() - t0; ts = seen / dt
ts_hist.append(ts); avg = sum(ts_hist) / len(ts_hist)
print(f'step={step} tokens={toks} ce={l1.item():.4f} ppl={math.exp(min(20, l1.item())):.1f} '
f'bal={bal.item():.3f} mut={getattr(m,"_mut",0):.3f} tok_s={ts:,.0f} avg_tok_s={avg:,.0f}', flush=True)
log_tp({'step': step, 'tokens': toks, 'ce': l1.item(), 'tok_s': ts, 'avg_tok_s': avg,
'elapsed_h': (time.time() - wall0) / 3600, 'variant': 'v21'})
t0 = time.time(); seen = 0
if step % a.ckpt_every == 0:
save(step); vl = valid_loss()
print(f'CKPT step={step} valid_ce={vl:.4f} valid_ppl={math.exp(min(20, vl)):.1f}', flush=True)
log_tp({'step': step, 'tokens': toks, 'valid_ce': vl,
'avg_tok_s': sum(ts_hist)/max(1,len(ts_hist)), 'variant': 'v21'})
if step % a.hf_every == 0: hf_push(step)
if toks >= a.target_tokens:
print(f'TARGET_TOKENS reached {toks}', flush=True); break
if a.max_hours and (time.time() - wall0) / 3600 >= a.max_hours:
print(f'MAX_HOURS reached {(time.time()-wall0)/3600:.2f}', flush=True); break
save(step); torch.save({'model': m.state_dict(), 'config': m.config()}, Path(a.out) / 'base.pt'); hf_push(step)
AVG = sum(ts_hist) / max(1, len(ts_hist))
print('PEER21_BASE_DONE', json.dumps({'valid_ce': valid_loss(), 'steps': step,
'avg_tok_s': AVG, 'tokens': step * a.batch * (a.seq - 1)}), flush=True)
|