File size: 9,723 Bytes
0241b9f | 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 | #!/usr/bin/env python3
from __future__ import annotations
import argparse
import json
import sys
from dataclasses import dataclass
from pathlib import Path
import torch
import torch.nn.functional as F
REPO_ROOT = Path(__file__).resolve().parents[1]
if str(REPO_ROOT) not in sys.path:
sys.path.insert(0, str(REPO_ROOT))
from eval import build_model_from_ckpt
from flowtext_lab.bridges import smooth_onehot
from flowtext_lab.decode import model_time_for_step, sample_noise_simplex, state_for_model
from flowtext_lab.tokenization import BpeTextTokenizer
@dataclass
class DecodeConfig:
label: str
steps: int
damping: float = 1.0
max_gamma: float = 1.0
endpoint_temp: float = 1.0
final_from: str = "state"
def focused_configs(steps: int) -> list[DecodeConfig]:
return [
DecodeConfig("flowmap_t1p00_d1p0", steps, endpoint_temp=1.00, damping=1.0),
DecodeConfig("flowmap_t1p10_d1p0", steps, endpoint_temp=1.10, damping=1.0),
DecodeConfig("flowmap_t1p25_d1p0", steps, endpoint_temp=1.25, damping=1.0),
DecodeConfig("flowmap_t1p40_d1p0", steps, endpoint_temp=1.40, damping=1.0),
DecodeConfig("flowmap_t1p60_d1p0", steps, endpoint_temp=1.60, damping=1.0),
DecodeConfig("flowmap_t1p25_d0p7", steps, endpoint_temp=1.25, damping=0.7),
DecodeConfig("flowmap_t1p40_d0p7", steps, endpoint_temp=1.40, damping=0.7),
DecodeConfig("flowmap_t1p60_d0p7", steps, endpoint_temp=1.60, damping=0.7),
DecodeConfig("flowmap_t1p25_g0p5", steps, endpoint_temp=1.25, damping=1.0, max_gamma=0.5),
DecodeConfig("flowmap_t1p40_g0p5", steps, endpoint_temp=1.40, damping=1.0, max_gamma=0.5),
]
def flowmap_gamma(step: int, steps: int, damping: float, max_gamma: float, eps: float) -> float:
s = step / max(steps, 1)
t_next = (step + 1) / max(steps, 1)
base = (t_next - s) / max(1.0 - s, eps)
gamma = float(damping) * base
return min(gamma, float(max_gamma)) if max_gamma > 0 else gamma
def encode_prompt(tokenizer: BpeTextTokenizer, prompt: str, max_len: int) -> list[int]:
core = list(tokenizer.tokenizer.encode(prompt, add_special_tokens=False).ids)
bos = tokenizer.bos_id
ids = ([bos] if bos is not None and bos >= 0 else []) + core
return ids[:max_len]
def decode_text(tokenizer: BpeTextTokenizer, ids: list[int]) -> str:
return tokenizer.decode(ids, stop_at_eos=False, skip_special_tokens=False)
def build_initial_state(
tokenizer: BpeTextTokenizer,
prompts: list[str],
restarts: int,
max_len: int,
target_prob: float,
eps: float,
noise_init: str,
dirichlet_concentration: float,
device: torch.device,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, list[str]]:
expanded: list[str] = []
prompt_ids: list[list[int]] = []
for prompt in prompts:
ids = encode_prompt(tokenizer, prompt, max_len)
for _ in range(restarts):
expanded.append(prompt)
prompt_ids.append(ids)
batch = len(prompt_ids)
probs = sample_noise_simplex(
(batch, max_len),
tokenizer.vocab_size,
device,
eps,
noise_mode=noise_init,
target_prob=target_prob,
noise_sigma=-1.0,
dirichlet_concentration=dirichlet_concentration,
).float()
attn = torch.ones((batch, max_len), dtype=torch.bool, device=device)
lock = torch.zeros((batch, max_len), dtype=torch.bool, device=device)
lock_probs = torch.zeros((batch, max_len, tokenizer.vocab_size), dtype=torch.float32, device=device)
for row, ids in enumerate(prompt_ids):
if not ids:
continue
ids_t = torch.tensor(ids, dtype=torch.long, device=device).unsqueeze(0)
sp = smooth_onehot(ids_t, tokenizer.vocab_size, target_prob, eps)[0]
probs[row, : len(ids)] = sp
lock_probs[row, : len(ids)] = sp
lock[row, : len(ids)] = True
return probs, attn, lock, lock_probs, expanded
def parse_args() -> argparse.Namespace:
p = argparse.ArgumentParser()
p.add_argument("--checkpoint", required=True)
p.add_argument("--tokenizer_path", required=True)
p.add_argument("--output", required=True)
p.add_argument("--prompts", required=True)
p.add_argument("--prompt", required=True)
p.add_argument("--restarts", type=int, default=20)
p.add_argument("--candidate_index", type=int, required=True)
p.add_argument("--steps", type=int, required=True)
p.add_argument("--config_label", required=True)
p.add_argument("--max_len", type=int, default=128)
p.add_argument("--seed", type=int, default=20260502)
p.add_argument("--target_prob", type=float, default=1.0)
p.add_argument("--noise_init", default="dirichlet")
p.add_argument("--dirichlet_init_concentration", type=float, default=1.0)
p.add_argument("--eps", type=float, default=1e-8)
return p.parse_args()
@torch.no_grad()
def main() -> None:
args = parse_args()
torch.manual_seed(args.seed)
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
tokenizer = BpeTextTokenizer.from_file(args.tokenizer_path)
ckpt = torch.load(args.checkpoint, map_location="cpu")
model = build_model_from_ckpt(ckpt, tokenizer.vocab_size, args.max_len, device)
model.eval()
prompts = args.prompts.split("|")
configs = focused_configs(args.steps)
selected_cfg = None
init = attn = lock = lock_probs = None
expanded: list[str] = []
# Reproduce the decode-sweep RNG stream: every config samples a fresh initial
# batch. We consume the same initial batches until the requested config.
for cfg in configs:
init, attn, lock, lock_probs, expanded = build_initial_state(
tokenizer,
prompts,
args.restarts,
args.max_len,
args.target_prob,
args.eps,
args.noise_init,
args.dirichlet_init_concentration,
device,
)
if cfg.label == args.config_label:
selected_cfg = cfg
break
del init, attn, lock, lock_probs
if selected_cfg is None or init is None or attn is None or lock is None or lock_probs is None:
raise ValueError(f"unknown config_label {args.config_label}")
if expanded[args.candidate_index] != args.prompt:
raise ValueError(
f"candidate prompt mismatch: candidate={args.candidate_index} has {expanded[args.candidate_index]!r}, expected {args.prompt!r}"
)
sl = slice(args.candidate_index, args.candidate_index + 1)
probs = init[sl].clone()
attn = attn[sl]
lock = lock[sl]
lock_probs = lock_probs[sl]
last_endpoint = probs
records = []
for step in range(selected_cfg.steps):
t = model_time_for_step("flow", step, selected_cfg.steps, 1, device, dtype=torch.float32)
logits = model(state_for_model(model, probs, args.eps), t, attn).float()
logits = logits / float(selected_cfg.endpoint_temp)
endpoint = F.softmax(logits, dim=-1)
last_endpoint = endpoint
gamma = flowmap_gamma(step, selected_cfg.steps, selected_cfg.damping, selected_cfg.max_gamma, args.eps)
new_probs = probs + gamma * (endpoint - probs)
new_probs = new_probs.clamp_min(args.eps)
new_probs = new_probs / new_probs.sum(dim=-1, keepdim=True).clamp_min(args.eps)
probs = torch.where(lock.unsqueeze(-1), lock_probs, new_probs)
state_top_prob, state_ids = probs[0].max(dim=-1)
endpoint_top_prob, endpoint_ids = endpoint[0].max(dim=-1)
records.append(
{
"step": step,
"gamma": gamma,
"model_t": float(t.item()),
"state_text": decode_text(tokenizer, state_ids.detach().cpu().tolist()),
"endpoint_text": decode_text(tokenizer, endpoint_ids.detach().cpu().tolist()),
"positions": [
{
"pos": pos,
"state_token": tokenizer.decode([int(state_ids[pos].item())], stop_at_eos=False, skip_special_tokens=False),
"state_id": int(state_ids[pos].item()),
"state_top_p": float(state_top_prob[pos].item()),
"endpoint_token": tokenizer.decode([int(endpoint_ids[pos].item())], stop_at_eos=False, skip_special_tokens=False),
"endpoint_id": int(endpoint_ids[pos].item()),
"endpoint_top_p": float(endpoint_top_prob[pos].item()),
}
for pos in range(args.max_len)
],
}
)
if selected_cfg.final_from == "endpoint":
final_dist = torch.where(lock.unsqueeze(-1), lock_probs, last_endpoint)
else:
final_dist = probs
final_dist = final_dist / final_dist.sum(dim=-1, keepdim=True).clamp_min(args.eps)
final_ids = final_dist[0].argmax(dim=-1).detach().cpu().tolist()
payload = {
"checkpoint": args.checkpoint,
"seed": args.seed,
"prompts": prompts,
"prompt": args.prompt,
"restarts": args.restarts,
"candidate_index": args.candidate_index,
"steps": args.steps,
"config": selected_cfg.__dict__,
"final_ids": final_ids,
"final_text": decode_text(tokenizer, final_ids),
"records": records,
}
out = Path(args.output)
out.parent.mkdir(parents=True, exist_ok=True)
out.write_text(json.dumps(payload, ensure_ascii=False), encoding="utf-8")
print(json.dumps({"output": str(out), "final": payload["final_text"]}, ensure_ascii=False, indent=2))
if __name__ == "__main__":
main()
|