File size: 6,880 Bytes
44d9388 |
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 |
#!/usr/bin/env python3
"""
WIRE-SPEED TRANSFORMER - Learns directly from network stream
No batching. No epochs. Just continuous absorption.
Receives tokenized data via stdin from Rust feeder.
Updates weights after every micro-batch (configurable, default 32 tokens).
"""
import sys
import math
import time
import torch
import torch.nn as nn
import torch.nn.functional as F
from collections import deque
# ─────────────────── Config ───────────────────
DEVICE = torch.device("cuda" if torch.cuda.is_available() else "cpu")
torch.backends.cuda.matmul.allow_tf32 = True
# Tiny model for wire-speed updates
CONFIG = {
"d": 256, # embedding dim
"layers": 4, # transformer layers
"heads": 8, # attention heads
"rank": 32, # attention rank (from n.py's tuneable attention)
"vocab": 128256, # DeepSeek V3.2 vocab
"ctx": 512, # context window
}
LR = 1e-4
UPDATE_EVERY = 32 # tokens between weight updates (micro-batch)
PRINT_EVERY = 10000 # tokens between stats
# ─────────────────── Model (simplified from n.py) ───────────────────
class TuneableAttention(nn.Module):
def __init__(self, d, h, r):
super().__init__()
self.h, self.dk, self.r = h, d // h, r
self.qkv = nn.Linear(d, 3 * d, bias=False)
self.U = nn.Parameter(torch.randn(self.dk, r) * 0.02)
self.proj = nn.Linear(d, d, bias=False)
def forward(self, x, mask=None):
B, N, D = x.shape
qkv = self.qkv(x).view(B, N, 3, self.h, self.dk)
q, k, v = qkv.unbind(2) # B, N, h, dk
# Project Q and K through U for tuneable rank
q = (q @ self.U) # B, N, h, r
k = (k @ self.U) # B, N, h, r
# Attention
q, k, v = q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2)
att = (q @ k.transpose(-1, -2)) / math.sqrt(self.r)
if mask is not None:
att = att + mask
att = F.softmax(att, dim=-1)
out = (att @ v).transpose(1, 2).reshape(B, N, D)
return self.proj(out)
class Block(nn.Module):
def __init__(self, d, h, r):
super().__init__()
self.ln1 = nn.LayerNorm(d)
self.attn = TuneableAttention(d, h, r)
self.ln2 = nn.LayerNorm(d)
self.ff = nn.Sequential(
nn.Linear(d, 4 * d),
nn.GELU(),
nn.Linear(4 * d, d)
)
def forward(self, x, mask):
x = x + self.attn(self.ln1(x), mask)
x = x + self.ff(self.ln2(x))
return x
class StreamingTransformer(nn.Module):
def __init__(self, cfg):
super().__init__()
d, L, h, r, V = cfg["d"], cfg["layers"], cfg["heads"], cfg["rank"], cfg["vocab"]
self.emb = nn.Embedding(V, d)
self.blocks = nn.ModuleList([Block(d, h, r) for _ in range(L)])
self.ln = nn.LayerNorm(d)
self.head = nn.Linear(d, V, bias=False)
# Weight tying
self.head.weight = self.emb.weight
def forward(self, x):
B, N = x.shape
# Causal mask
mask = torch.triu(torch.ones(N, N, device=x.device), 1) * -1e9
h = self.emb(x)
for block in self.blocks:
h = block(h, mask)
return self.head(self.ln(h))
def count_params(self):
return sum(p.numel() for p in self.parameters())
# ─────────────────── Online Trainer ───────────────────
class WireSpeedTrainer:
def __init__(self, model, lr=LR):
self.model = model.to(DEVICE)
self.opt = torch.optim.AdamW(model.parameters(), lr=lr, betas=(0.9, 0.95))
self.ctx_size = CONFIG["ctx"]
# Rolling buffer for context
self.buffer = deque(maxlen=self.ctx_size + 1)
# Stats
self.tokens_seen = 0
self.total_loss = 0.0
self.updates = 0
self.start_time = time.time()
def ingest_token(self, token_id):
"""Absorb a single token. Update weights when buffer fills."""
self.buffer.append(token_id)
self.tokens_seen += 1
# Update every N tokens when we have enough context
if len(self.buffer) >= UPDATE_EVERY + 1 and self.tokens_seen % UPDATE_EVERY == 0:
self._update()
# Print stats
if self.tokens_seen % PRINT_EVERY == 0:
self._print_stats()
def _update(self):
"""Single gradient step on current buffer."""
# Convert buffer to tensor
tokens = list(self.buffer)
x = torch.tensor(tokens[:-1], device=DEVICE).unsqueeze(0) # input
y = torch.tensor(tokens[1:], device=DEVICE).unsqueeze(0) # target
# Forward
self.model.train()
logits = self.model(x)
# Loss on last UPDATE_EVERY positions only (most recent)
loss = F.cross_entropy(
logits[:, -UPDATE_EVERY:].reshape(-1, CONFIG["vocab"]),
y[:, -UPDATE_EVERY:].reshape(-1)
)
# Backward
self.opt.zero_grad()
loss.backward()
torch.nn.utils.clip_grad_norm_(self.model.parameters(), 1.0)
self.opt.step()
self.total_loss += loss.item()
self.updates += 1
def _print_stats(self):
elapsed = time.time() - self.start_time
tok_per_sec = self.tokens_seen / elapsed if elapsed > 0 else 0
avg_loss = self.total_loss / max(1, self.updates)
print(f"[{elapsed:.0f}s] {self.tokens_seen:,} tok | {tok_per_sec:.0f} tok/s | "
f"loss={avg_loss:.4f} | updates={self.updates}", flush=True)
# ─────────────────── Main ───────────────────
def main():
print(f"Wire-Speed Transformer", flush=True)
print(f"Config: {CONFIG}", flush=True)
print(f"Device: {DEVICE}", flush=True)
model = StreamingTransformer(CONFIG)
params = model.count_params()
print(f"Parameters: {params:,} ({params/1e6:.1f}M)", flush=True)
trainer = WireSpeedTrainer(model)
print(f"Listening for tokens on stdin...", flush=True)
print(f"Update every {UPDATE_EVERY} tokens, print every {PRINT_EVERY}", flush=True)
# Read token IDs from stdin (one per line from Rust feeder)
for line in sys.stdin:
try:
token_id = int(line.strip())
if 0 <= token_id < CONFIG["vocab"]:
trainer.ingest_token(token_id)
except ValueError:
continue # Skip malformed lines
print(f"Stream ended. Total tokens: {trainer.tokens_seen:,}", flush=True)
if __name__ == "__main__":
main()
|