File size: 8,404 Bytes
d8bc908 | 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 | #!/usr/bin/env python3
"""
Test BigInt-accumulated ScaledOptum on a ~150M param ternary MLP.
Architecture:
Embedding(288, 2048) → [Repeat: Linear(2048→8192) → ReLU → Linear(8192→2048)] × 5
→ RMSNorm(2048) → Linear(2048→288)
All linear weights use TernaryScaleTensor (packed ternary T + S from optimizer).
Training: predict next byte on TinyShakespeare.
Key metrics:
- Loss trend (should decrease if optimizer works)
- Memory usage (model + optimizer state)
- Effective bits-per-weight
"""
import os, sys, math, gc
sys.path.insert(0, os.path.dirname(__file__))
import torch
import torch.nn as nn
import torch.nn.functional as F
from tscale_mini import TernaryScaleTensor, TernaryRMSNorm, _n_groups
from scaled_optum import ScaledOptum
torch.set_float32_matmul_precision('high')
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(f"Device: {device}")
# ─── Config ───
VOCAB = 288
HIDDEN = 2048
FFN_HIDDEN = 8192
N_LAYERS = 2
GROUP_SIZE = 32
THRESHOLD = 0.05
# ─── Model ───
class TernaryMLP(nn.Module):
"""
Pure ternary MLP with packed weights + ALL-INT persistent state.
No float32/16 anywhere in model buffers.
"""
def __init__(self):
super().__init__()
self.embed = TernaryScaleTensor(VOCAB, HIDDEN, threshold=THRESHOLD,
group_size=GROUP_SIZE)
self.layers = nn.ModuleList()
for i in range(N_LAYERS):
layer = nn.ModuleDict({
'w1': TernaryScaleTensor(HIDDEN, FFN_HIDDEN, threshold=THRESHOLD,
group_size=GROUP_SIZE),
'w2': TernaryScaleTensor(FFN_HIDDEN, HIDDEN, threshold=THRESHOLD,
group_size=GROUP_SIZE),
'norm': TernaryRMSNorm(HIDDEN, group_size=GROUP_SIZE),
})
self.layers.append(layer)
self.final_norm = TernaryRMSNorm(HIDDEN, group_size=GROUP_SIZE)
self.head = TernaryScaleTensor(HIDDEN, VOCAB, threshold=THRESHOLD,
group_size=GROUP_SIZE)
def forward(self, x, targets=None):
B, T = x.shape
emb = self.embed(F.one_hot(x, num_classes=VOCAB).float())
h = emb
for layer in self.layers:
h = layer['w1'](h)
h = F.relu(h)
h = layer['w2'](h)
h = self.final_norm(h)
logits = self.head(h)
if targets is not None:
loss = F.cross_entropy(logits.view(-1, VOCAB), targets.view(-1))
return logits, loss
return logits
def param_counts(self):
total_ternary = 0
total_float = 0
for _, mod in self.named_modules():
if isinstance(mod, (TernaryScaleTensor, TernaryRMSNorm)):
total_ternary += mod.total_ternary_params()
else:
for p in mod.parameters(recurse=False):
total_float += p.numel()
return total_ternary, total_float
def persistent_memory_mb(self):
total = 0
for mod in self.modules():
if isinstance(mod, (TernaryScaleTensor, TernaryRMSNorm)):
total += mod.persistent_memory_mb()
return total
# ─── Data (TinyShakespeare) ───
def load_data(path="/tmp/tinyshakespeare.txt"):
if not os.path.exists(path):
import urllib.request
url = "https://raw.githubusercontent.com/karpathy/char-rnn/master/data/tinyshakespeare/input.txt"
urllib.request.urlretrieve(url, path)
with open(path, 'rb') as f:
data = f.read()
return torch.tensor(list(data), dtype=torch.long)
def get_batch(data, bs, ctx, device=device):
ix = torch.randint(0, len(data) - ctx - 1, (bs,), device='cpu')
x = torch.stack([data[i:i + ctx] for i in ix])
y = torch.stack([data[i + 1:i + ctx + 1] for i in ix])
return x.to(device), y.to(device)
# ─── Test ───
@torch.no_grad()
def compute_loss(model, data, bs=4, ctx=256):
model.eval()
x, y = get_batch(data, bs, ctx)
_, loss = model(x, targets=y)
return loss.item()
def train_step(model, opt, data, bs=2, ctx=128):
model.train()
x, y = get_batch(data, bs, ctx)
logits, loss = model(x, targets=y)
loss.backward()
opt.step()
opt.zero_grad(set_to_none=True)
return loss.item()
def main():
print("Building TernaryMLP...")
model = TernaryMLP().to(device)
total_ternary, total_float = model.param_counts()
total_params = total_ternary + total_float
persistent_mb = model.persistent_memory_mb()
# Breakdown of persistent int storage
t_b = sum(m.T_packed.numel() * m.T_packed.element_size()
for m in model.modules() if isinstance(m, (TernaryScaleTensor, TernaryRMSNorm)))
e_b = sum(m.E.numel() * m.E.element_size()
for m in model.modules() if isinstance(m, (TernaryScaleTensor, TernaryRMSNorm)))
a_b = sum(m.corr_accum.numel() * m.corr_accum.element_size()
for m in model.modules() if isinstance(m, (TernaryScaleTensor, TernaryRMSNorm)))
sc_b = sum(getattr(m, 'step_counter', torch.zeros(1)).numel() * 8
for m in model.modules() if isinstance(m, (TernaryScaleTensor, TernaryRMSNorm)))
bpw = (t_b * 8 + e_b * 8 + a_b * 8 + sc_b * 8) / max(1, total_params)
print(f"\n Total params: {total_params:,}")
print(f" Ternary params: {total_ternary:,} ({total_ternary/max(1,total_params)*100:.1f}%)")
print(f" Float params: {total_float:,}")
print(f" Persistent buffers: {persistent_mb:.2f} MB (ALL INTEGER)")
print(f" T_packed: {t_b/1e6:.2f} MB ({t_b*8/total_ternary:.2f} bpw)")
print(f" E (int8): {e_b/1e6:.2f} MB")
print(f" corr_accum (int64):{a_b/1e6:.2f} MB")
print(f" step_counter: {sc_b/1e6:.2f} MB")
print(f" Effective bpw: {bpw:.2f}")
print(f" Float params (bias): {sum(p.numel()*p.element_size() for p in model.parameters())/1e6:.1f} MB")
# Collect all ternary modules
ternary_modules = [mod for mod in model.modules()
if isinstance(mod, (TernaryScaleTensor, TernaryRMSNorm))]
# Optimizer: pure integer, no float state
dummy = nn.Parameter(torch.zeros(1))
opt = ScaledOptum([dummy], lr=0.3, default_group_size=GROUP_SIZE)
opt.add_ternary_modules(ternary_modules)
n_mods = len(opt.param_groups[0].get('ternary_modules', []))
print(f" Optimizer state: 0 bytes (pure integer, stored on modules)")
print(f" Ternary modules: {n_mods}")
# Data
data = load_data()
train_data = data[:int(0.9 * len(data))]
val_data = data[int(0.9 * len(data)):]
print(f" Train data: {len(train_data):,} bytes")
print(f" Val data: {len(val_data):,} bytes")
# Warmup: nn.Module.parameters() won't find TernaryScaleTensor buffers
# (T_packed etc are buffers, not parameters). The optimizer only sees
# the .S_opt and the norms' float params. That's fine — we handle
# ternary params via hooks, not nn.Parameter.
# Training
N_STEPS = 5000
print(f"\nTraining for {N_STEPS} steps...")
print(f"{'step':>6s} {'loss':>8s} {'bpw':>8s} {'acc%':>6s} {'S_range':>10s} {'VRAM':>6s}")
print("-" * 60)
for step in range(N_STEPS):
loss = train_step(model, opt, train_data, bs=2, ctx=128)
if step % 200 == 0 or step == N_STEPS - 1:
# Compute accuracy
model.eval()
x, y = get_batch(val_data, 1, 128)
logits, _ = model(x, targets=y)
acc = (logits.argmax(-1) == y).float().mean().item()
model.train()
# Get E range and sign_bias for first layer
e_vals = model.layers[0]['w1'].E
e_min, e_max = e_vals.min().item(), e_vals.max().item()
bpw_val = loss / math.log(2)
vram = torch.cuda.max_memory_allocated() / 1e6 if torch.cuda.is_available() else 0
torch.cuda.reset_peak_memory_stats() if torch.cuda.is_available() else None
print(f"{step:6d} {loss:8.4f} {bpw_val:8.3f} {acc*100:5.1f}% "
f"2^{e_min:+3d}–2^{e_max:+3d} {vram:5.0f}MB")
print("\nDone.")
print(f"Final loss: {loss:.4f}")
if __name__ == '__main__':
main()
|