File size: 7,128 Bytes
ced35f2 | 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 | """catgirl-1m standalone inference. Needs only: torch, tokenizers, safetensors.
A full fine-tune of Glint-Research/Glint-2 (1,064,737 params) into a
tsundere-catgirl persona.
Usage:
python generate.py "My name is"
python generate.py --max-new-tokens 120 --temperature 0.7 "User: hey there"
The base is one shared transformer block looped 8 times, with a tiny per-loop
LoRA and a loop-embedding that tells it which iteration it is on. The
checkpoint's loop tables have capacity 16; run it at 8 loops (the value it was
trained at). More loops produce gibberish.
"""
import argparse
import json
import math
import os
import torch
from tokenizers import Tokenizer
from torch import nn
from torch.nn import functional
from safetensors.torch import load_file
LOOPS = 8
ATTENTION_WINDOW = 256
class SwiGlu(nn.Module):
def __init__(self, dim, hidden):
super().__init__()
self.gate_up = nn.Linear(dim, 2 * hidden, bias=False)
self.down = nn.Linear(hidden, dim, bias=False)
def forward(self, x):
gate, up = self.gate_up(x).chunk(2, dim=-1)
return self.down(functional.silu(gate) * up)
class Attention(nn.Module):
def __init__(self, dim, n_heads):
super().__init__()
self.n_heads = n_heads
self.head_dim = dim // n_heads
self.qkv = nn.Linear(dim, 3 * dim, bias=False)
self.out = nn.Linear(dim, dim, bias=False)
def forward(self, x, cos, sin, qkv_delta):
batch, seq_len, dim = x.shape
qkv = self.qkv(x) + qkv_delta
q, k, v = qkv.split(dim, dim=-1)
q = q.view(batch, seq_len, self.n_heads, self.head_dim).transpose(1, 2)
k = k.view(batch, seq_len, self.n_heads, self.head_dim).transpose(1, 2)
v = v.view(batch, seq_len, self.n_heads, self.head_dim).transpose(1, 2)
q, k = apply_rope(q, cos, sin), apply_rope(k, cos, sin)
pos = torch.arange(seq_len, device=x.device)
mask = (pos[None, :] <= pos[:, None]) & (pos[:, None] - pos[None, :] < ATTENTION_WINDOW)
attended = functional.scaled_dot_product_attention(
q, k, v, attn_mask=mask, scale=1.0 / math.sqrt(self.head_dim)
)
return self.out(attended.transpose(1, 2).reshape(batch, seq_len, dim))
def apply_rope(x, cos, sin):
x_even, x_odd = x[..., 0::2], x[..., 1::2]
return torch.stack(
(x_even * cos - x_odd * sin, x_even * sin + x_odd * cos), dim=-1
).flatten(-2)
class Block(nn.Module):
def __init__(self, dim, n_heads, ffn_hidden):
super().__init__()
self.attn_norm = nn.RMSNorm(dim)
self.attn = Attention(dim, n_heads)
self.ffn_norm = nn.RMSNorm(dim)
self.ffn = SwiGlu(dim, ffn_hidden)
def forward(self, x, cos, sin, qkv_delta):
x = x + self.attn(self.attn_norm(x), cos, sin, qkv_delta)
return x + self.ffn(self.ffn_norm(x))
class LoopLora(nn.Module):
def __init__(self, dim, rank, max_loops):
super().__init__()
self.down = nn.ModuleList(nn.Linear(dim, rank, bias=False) for _ in range(max_loops))
self.up = nn.ModuleList(nn.Linear(rank, 3 * dim, bias=False) for _ in range(max_loops))
def forward(self, x, loop_index):
i = min(loop_index, len(self.down) - 1)
return self.up[i](self.down[i](x))
class Indexer(nn.Module):
def __init__(self):
super().__init__()
self.gate = nn.Parameter(torch.tensor([0.1]))
class Glint2(nn.Module):
def __init__(self, cfg, max_loops):
super().__init__()
self.cfg = cfg
self.max_loops = max_loops
self.embed = nn.Embedding(cfg["vocab_size"], cfg["dim"])
self.indexer = Indexer()
self.shared = Block(cfg["dim"], cfg["n_heads"], cfg["ffn_hidden"])
self.loop_lora = LoopLora(cfg["dim"], cfg["lora_rank"], max_loops)
self.loop_embed = nn.Embedding(max_loops, cfg["dim"])
self.final_norm = nn.RMSNorm(cfg["dim"])
def rope(self, seq_len):
head_dim = self.cfg["dim"] // self.cfg["n_heads"]
positions = torch.arange(seq_len, dtype=torch.float32)
inv_freq = 1.0 / (
self.cfg["rope_base"]
** (torch.arange(0, head_dim, 2, dtype=torch.float32) / head_dim)
)
angles = torch.outer(positions, inv_freq)
return torch.cos(angles), torch.sin(angles)
def forward(self, tokens, loops=LOOPS, return_hidden=False):
cos, sin = self.rope(tokens.shape[1])
x = self.embed(tokens)
for loop_index in range(loops):
clamped = min(loop_index, self.max_loops - 1)
gated = x + self.loop_embed.weight[clamped]
delta = self.loop_lora(gated, loop_index)
x = self.shared(gated, cos, sin, delta)
x = self.final_norm(x)
if return_hidden:
return x
return functional.linear(x, self.embed.weight)
def load_model(model_dir):
with open(os.path.join(model_dir, "config.json")) as f:
cfg = json.load(f)
state = load_file(os.path.join(model_dir, "catgirl-1m.safetensors"))
model = Glint2(cfg, max_loops=cfg["max_loops"])
model.load_state_dict(state)
model.eval()
return model
def sample(logits, ids, temperature, top_k, repetition_penalty, repetition_window):
if repetition_penalty != 1.0:
for token in set(ids[-repetition_window:]):
if logits[token] > 0:
logits[token] /= repetition_penalty
else:
logits[token] *= repetition_penalty
if temperature <= 0:
return int(logits.argmax())
values, indices = logits.topk(top_k)
probs = functional.softmax(values / temperature, dim=-1)
return int(indices[torch.multinomial(probs, 1)])
def main():
parser = argparse.ArgumentParser(description="catgirl-1m text generation")
parser.add_argument("prompt")
parser.add_argument("--model-dir", default=os.path.dirname(os.path.abspath(__file__)))
parser.add_argument("--max-new-tokens", type=int, default=100)
parser.add_argument("--temperature", type=float, default=0.15, help="0 = greedy")
parser.add_argument("--top-k", type=int, default=5)
parser.add_argument("--repetition-penalty", type=float, default=1.05)
parser.add_argument("--repetition-window", type=int, default=128)
parser.add_argument("--loops", type=int, default=LOOPS)
args = parser.parse_args()
tokenizer = Tokenizer.from_file(os.path.join(args.model_dir, "tokenizer.json"))
model = load_model(args.model_dir)
ids = tokenizer.encode(args.prompt).ids
print(args.prompt, end="", flush=True)
with torch.no_grad():
for _ in range(args.max_new_tokens):
logits = model(torch.tensor([ids]), loops=args.loops)[0, -1]
token = sample(
logits, ids, args.temperature, args.top_k,
args.repetition_penalty, args.repetition_window,
)
new_text = tokenizer.decode(ids + [token])[len(tokenizer.decode(ids)):]
ids.append(token)
print(new_text, end="", flush=True)
print()
if __name__ == "__main__":
main() |