File size: 12,182 Bytes
0badcf2 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 | #!/usr/bin/env python3
"""Algebraic simplex-linear GenPPL eval for endpoint models.
This decoder matches the supervised bridge:
p_t = (1 - t) * p0 + t * x1
Inference keeps the sampled p0 fixed and replaces the unknown x1 with the
model's current endpoint prediction:
p_{t_next} = (1 - t_next) * p0 + t_next * a_theta(p_t, t).
There is no Dirichlet/Gamma resampling in the loop.
"""
from __future__ import annotations
import argparse
import json
import math
import sys
from pathlib import Path
import torch
import torch.nn.functional as F
from transformers import AutoModelForCausalLM, AutoTokenizer
REPO_ROOT = Path(__file__).resolve().parents[1]
if str(REPO_ROOT) not in sys.path:
sys.path.insert(0, str(REPO_ROOT))
from flowtext_lab.decode import sample_noise_simplex, state_for_model
from flowtext_lab.genppl import filter_generated_texts, summarize_token_diversity
from flowtext_lab.tokenization import BpeTextTokenizer
from eval_lm1b_c1024_fullycoupled_sde_genppl import (
build_model,
collect_special_token_ids,
filter_endpoint_probs,
score_with_gpt2,
)
def lerp(a: float, b: float, t: float) -> float:
return float(a) + float(t) * (float(b) - float(a))
def project_endpoint(
logits: torch.Tensor,
*,
temp: float,
projection: str,
top_k: int,
top_p: float,
banned_ids: list[int],
gumbel_tau: float,
gumbel_noise_scale: float,
eps: float,
) -> torch.Tensor:
endpoint = F.softmax(logits / max(float(temp), eps), dim=-1)
endpoint = filter_endpoint_probs(
endpoint,
top_k=top_k,
top_p=top_p,
banned_ids=banned_ids,
eps=eps,
)
if projection == "soft":
return endpoint
if projection == "argmax":
ids = endpoint.argmax(dim=-1)
return torch.zeros_like(endpoint).scatter_(-1, ids.unsqueeze(-1), 1.0)
if projection == "sample":
ids = torch.multinomial(endpoint.reshape(-1, endpoint.size(-1)), 1).view(*endpoint.shape[:-1])
return torch.zeros_like(endpoint).scatter_(-1, ids.unsqueeze(-1), 1.0)
if projection == "gumbel_softmax":
u = torch.rand_like(endpoint).clamp_(min=eps, max=1.0 - eps)
g = -torch.log(-torch.log(u))
z = (endpoint.clamp_min(eps).log() + float(gumbel_noise_scale) * g) / max(float(gumbel_tau), eps)
y = F.softmax(z, dim=-1).clamp_min(eps)
return y / y.sum(dim=-1, keepdim=True).clamp_min(eps)
raise ValueError(f"unknown endpoint_projection: {projection}")
@torch.inference_mode()
def decode_linear_simplex(
model,
tokenizer: BpeTextTokenizer,
*,
n_samples: int,
batch_size: int,
max_len: int,
steps: int,
seed: int,
device: torch.device,
noise_init: str,
noise_sigma: float,
noise_dirichlet_concentration: float,
endpoint_temp_start: float,
endpoint_temp_end: float,
endpoint_projection: str,
endpoint_top_k: int,
endpoint_top_p: float,
ban_special_tokens: bool,
gumbel_tau_start: float,
gumbel_tau_end: float,
gumbel_noise_scale_start: float,
gumbel_noise_scale_end: float,
final_from: str,
) -> tuple[list[list[int]], list[str], dict]:
torch.manual_seed(seed)
eps = 1e-8
all_ids: list[list[int]] = []
all_texts: list[str] = []
remaining = n_samples
banned_endpoint_ids = collect_special_token_ids(tokenizer) if ban_special_tokens else []
while remaining > 0:
bs = min(batch_size, remaining)
p0 = sample_noise_simplex(
(bs, max_len),
tokenizer.vocab_size,
device,
eps,
noise_mode=noise_init,
target_prob=1.0,
noise_sigma=noise_sigma,
dirichlet_concentration=noise_dirichlet_concentration,
)
probs = p0.clone()
attn = torch.ones((bs, max_len), dtype=torch.bool, device=device)
last_endpoint = probs
for step in range(steps):
cur_t = step / max(steps, 1)
next_t = (step + 1) / max(steps, 1)
t = torch.full((bs,), float(cur_t), dtype=torch.float32, device=device)
logits = model(state_for_model(model, probs, eps), t, attn).float()
endpoint = project_endpoint(
logits,
temp=lerp(endpoint_temp_start, endpoint_temp_end, cur_t),
projection=endpoint_projection,
top_k=endpoint_top_k,
top_p=endpoint_top_p,
banned_ids=banned_endpoint_ids,
gumbel_tau=lerp(gumbel_tau_start, gumbel_tau_end, cur_t),
gumbel_noise_scale=lerp(gumbel_noise_scale_start, gumbel_noise_scale_end, cur_t),
eps=eps,
)
last_endpoint = endpoint
probs = (1.0 - next_t) * p0 + next_t * endpoint
probs = probs.clamp_min(eps)
probs = probs / probs.sum(dim=-1, keepdim=True).clamp_min(eps)
if final_from == "blend_0.5":
final_probs = 0.5 * probs + 0.5 * last_endpoint
ids = final_probs.argmax(dim=-1).detach().cpu().tolist()
elif final_from == "model_t1":
t = torch.ones((bs,), dtype=torch.float32, device=device)
final_logits = model(state_for_model(model, probs, eps), t, attn).float()
ids = final_logits.argmax(dim=-1).detach().cpu().tolist()
else:
raise ValueError(f"unknown final_from: {final_from}")
all_ids.extend(ids)
all_texts.extend(tokenizer.decode(row, stop_at_eos=False, skip_special_tokens=False) for row in ids)
remaining -= bs
print(f"[linear] generated {n_samples - remaining}/{n_samples}", flush=True)
cfg = {
"decode_rule": "linear_simplex_algebraic",
"steps": steps,
"noise_init": noise_init,
"noise_sigma": noise_sigma,
"noise_dirichlet_concentration": noise_dirichlet_concentration,
"endpoint_temp_start": endpoint_temp_start,
"endpoint_temp_end": endpoint_temp_end,
"endpoint_projection": endpoint_projection,
"endpoint_top_k": endpoint_top_k,
"endpoint_top_p": endpoint_top_p,
"ban_special_tokens": ban_special_tokens,
"banned_endpoint_ids": banned_endpoint_ids,
"gumbel_tau_start": gumbel_tau_start,
"gumbel_tau_end": gumbel_tau_end,
"gumbel_noise_scale_start": gumbel_noise_scale_start,
"gumbel_noise_scale_end": gumbel_noise_scale_end,
"final_from": final_from,
"n_samples": n_samples,
"seed": seed,
}
return all_ids, all_texts, cfg
def parse_args() -> argparse.Namespace:
p = argparse.ArgumentParser(description="Linear-simplex algebraic GenPPL eval")
p.add_argument("--checkpoint", required=True)
p.add_argument("--tokenizer_path", required=True)
p.add_argument("--scorer", required=True)
p.add_argument("--out_dir", required=True)
p.add_argument("--n_samples", type=int, default=128)
p.add_argument("--max_len", type=int, default=128)
p.add_argument("--steps", type=int, default=128)
p.add_argument("--batch_size", type=int, default=16)
p.add_argument("--score_batch", type=int, default=8)
p.add_argument("--score_max_length", type=int, default=1024)
p.add_argument("--noise_init", choices=["uniform", "logistic_normal", "dirichlet"], default="logistic_normal")
p.add_argument("--noise_sigma", type=float, default=3.0)
p.add_argument("--noise_dirichlet_concentration", type=float, default=1.0)
p.add_argument("--endpoint_temp_start", type=float, default=1.45)
p.add_argument("--endpoint_temp_end", type=float, default=0.8)
p.add_argument("--endpoint_projection", choices=["soft", "sample", "argmax", "gumbel_softmax"], default="soft")
p.add_argument("--endpoint_top_k", type=int, default=0)
p.add_argument("--endpoint_top_p", type=float, default=1.0)
p.add_argument("--ban_special_tokens", action="store_true")
p.add_argument("--gumbel_tau_start", type=float, default=1.0)
p.add_argument("--gumbel_tau_end", type=float, default=0.2)
p.add_argument("--gumbel_noise_scale_start", type=float, default=1.0)
p.add_argument("--gumbel_noise_scale_end", type=float, default=0.0)
p.add_argument("--final_from", choices=["blend_0.5", "model_t1"], default="model_t1")
p.add_argument("--seed", type=int, default=20260524)
return p.parse_args()
@torch.no_grad()
def main() -> None:
args = parse_args()
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"[load] {args.checkpoint}", flush=True)
ckpt = torch.load(args.checkpoint, map_location="cpu", weights_only=False)
step = ckpt.get("step")
print(f"[ckpt] step={step}", flush=True)
tokenizer = BpeTextTokenizer.from_file(args.tokenizer_path)
model = build_model(ckpt, tokenizer, device)
ids, texts, decode_cfg = decode_linear_simplex(
model,
tokenizer,
n_samples=args.n_samples,
batch_size=args.batch_size,
max_len=args.max_len,
steps=args.steps,
seed=args.seed,
device=device,
noise_init=args.noise_init,
noise_sigma=args.noise_sigma,
noise_dirichlet_concentration=args.noise_dirichlet_concentration,
endpoint_temp_start=args.endpoint_temp_start,
endpoint_temp_end=args.endpoint_temp_end,
endpoint_projection=args.endpoint_projection,
endpoint_top_k=args.endpoint_top_k,
endpoint_top_p=args.endpoint_top_p,
ban_special_tokens=args.ban_special_tokens,
gumbel_tau_start=args.gumbel_tau_start,
gumbel_tau_end=args.gumbel_tau_end,
gumbel_noise_scale_start=args.gumbel_noise_scale_start,
gumbel_noise_scale_end=args.gumbel_noise_scale_end,
final_from=args.final_from,
)
del model
if torch.cuda.is_available():
torch.cuda.empty_cache()
def strip_special(t: str) -> str:
import re
t = t.replace("[CLS]", " ").replace("[SEP]", " ").replace("[PAD]", " ")
t = t.replace("<|endoftext|>", " ")
return re.sub(r"\s+", " ", t).strip()
stripped = [strip_special(t) for t in texts]
kept_raw, _ = filter_generated_texts(texts, min_chars=1, normalize_whitespace=False, drop_empty=True)
kept_stripped, _ = filter_generated_texts(stripped, min_chars=1, normalize_whitespace=True, drop_empty=True)
diversity = summarize_token_diversity(ids).__dict__
print(f"[score] loading scorer: {args.scorer}", flush=True)
scorer_tok = AutoTokenizer.from_pretrained(args.scorer)
if scorer_tok.pad_token_id is None:
scorer_tok.pad_token = scorer_tok.eos_token
scorer_tok.pad_token_id = scorer_tok.eos_token_id
scorer = AutoModelForCausalLM.from_pretrained(args.scorer).to(device).eval()
if getattr(scorer.config, "pad_token_id", None) is None:
scorer.config.pad_token_id = scorer_tok.pad_token_id
raw_ppl = score_with_gpt2(
kept_raw, scorer, scorer_tok,
batch_size=args.score_batch, max_length=args.score_max_length, device=device,
)
stripped_ppl = score_with_gpt2(
kept_stripped, scorer, scorer_tok,
batch_size=args.score_batch, max_length=args.score_max_length, device=device,
)
summary = {
"type": "summary",
"checkpoint": args.checkpoint,
"step": step,
"decode": decode_cfg,
"raw_genppl": raw_ppl,
"stripped_genppl": stripped_ppl,
"diversity": diversity,
}
out_dir = Path(args.out_dir)
out_dir.mkdir(parents=True, exist_ok=True)
out_jsonl = out_dir / f"linear_steps{args.steps}_samples{args.n_samples}_scored.jsonl"
with out_jsonl.open("w", encoding="utf-8") as f:
f.write(json.dumps(summary, ensure_ascii=False) + "\n")
for i, (raw, clean) in enumerate(zip(texts, stripped)):
f.write(json.dumps({"type": "sample", "index": i, "raw_text": raw, "stripped_text": clean}, ensure_ascii=False) + "\n")
print("[summary]", json.dumps(summary, ensure_ascii=False, indent=2), flush=True)
print(f"[done] {out_jsonl}", flush=True)
if __name__ == "__main__":
main()
|