File size: 8,994 Bytes
05c5c96 | 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 | #!/usr/bin/env python3
"""
Run a distilled student checkpoint for text generation.
"""
import argparse
import logging
from pathlib import Path
import torch
import torch.nn.functional as F
from transformers import AutoTokenizer
from qwen_distill import QwenDistillationConfig, QwenStudentModel
logging.basicConfig(level=logging.INFO, format="%(asctime)s - %(levelname)s - %(message)s")
logger = logging.getLogger(__name__)
class StudentRunner:
"""Load a trained student checkpoint and generate text."""
def __init__(
self,
checkpoint_path: str,
device: str | None = None,
tokenizer_path: str | None = None,
):
self.checkpoint_path = Path(checkpoint_path)
if not self.checkpoint_path.exists():
raise FileNotFoundError(f"Checkpoint not found: {self.checkpoint_path}")
if device is None:
device = "cuda" if torch.cuda.is_available() else "cpu"
self.device = torch.device(device)
checkpoint = torch.load(self.checkpoint_path, map_location="cpu")
config_data = checkpoint["config"]
config = QwenDistillationConfig()
for key, value in config_data.items():
setattr(config, key, value)
self.config = config
self.model = QwenStudentModel(self.config)
self.model.load_state_dict(checkpoint["model_state_dict"])
self.model.to(self.device)
self.model.eval()
tokenizer_source = self._resolve_tokenizer_source(tokenizer_path)
logger.info("Loading tokenizer from %s", tokenizer_source)
self.tokenizer = AutoTokenizer.from_pretrained(
tokenizer_source,
trust_remote_code=True,
local_files_only=Path(tokenizer_source).exists(),
)
if self.tokenizer.pad_token is None:
self.tokenizer.pad_token = self.tokenizer.eos_token
logger.info(
"Loaded student checkpoint from %s on %s",
self.checkpoint_path,
self.device,
)
def _resolve_tokenizer_source(self, tokenizer_path: str | None) -> str:
if tokenizer_path:
return tokenizer_path
local_teacher = Path("models/teacher")
if local_teacher.exists():
return str(local_teacher)
return self.config.teacher_model_name
def generate(
self,
prompt: str,
max_new_tokens: int = 64,
temperature: float = 0.8,
top_p: float = 0.95,
top_k: int = 50,
repetition_penalty: float = 1.0,
) -> str:
if not prompt.strip():
raise ValueError("Prompt must not be empty.")
encoded = self.tokenizer(prompt, return_tensors="pt", add_special_tokens=True)
input_ids = encoded["input_ids"].to(self.device)
with torch.inference_mode():
for _ in range(max_new_tokens):
window = input_ids[:, -self.config.max_seq_length :]
attention_mask = torch.ones_like(window, device=self.device)
outputs = self.model(window, attention_mask=attention_mask)
next_token_logits = outputs["logits"][:, -1, :]
next_token_logits = self._apply_repetition_penalty(
next_token_logits,
input_ids,
repetition_penalty,
)
next_token = self._sample_token(
next_token_logits,
temperature=temperature,
top_p=top_p,
top_k=top_k,
)
input_ids = torch.cat([input_ids, next_token], dim=-1)
if self.tokenizer.eos_token_id is not None and next_token.item() == self.tokenizer.eos_token_id:
break
return self.tokenizer.decode(input_ids[0], skip_special_tokens=True)
@staticmethod
def _apply_repetition_penalty(
logits: torch.Tensor,
input_ids: torch.Tensor,
repetition_penalty: float,
) -> torch.Tensor:
if repetition_penalty <= 1.0:
return logits
adjusted = logits.clone()
for token_id in torch.unique(input_ids):
token_index = token_id.item()
token_score = adjusted[:, token_index]
adjusted[:, token_index] = torch.where(
token_score < 0,
token_score * repetition_penalty,
token_score / repetition_penalty,
)
return adjusted
@staticmethod
def _sample_token(
logits: torch.Tensor,
temperature: float,
top_p: float,
top_k: int,
) -> torch.Tensor:
if temperature <= 0:
return torch.argmax(logits, dim=-1, keepdim=True)
scaled_logits = logits / temperature
if top_k > 0:
top_k = min(top_k, scaled_logits.shape[-1])
values, _ = torch.topk(scaled_logits, top_k)
cutoff = values[:, -1].unsqueeze(-1)
scaled_logits = torch.where(
scaled_logits < cutoff,
torch.full_like(scaled_logits, float("-inf")),
scaled_logits,
)
if 0 < top_p < 1.0:
sorted_logits, sorted_indices = torch.sort(scaled_logits, descending=True, dim=-1)
sorted_probs = F.softmax(sorted_logits, dim=-1)
cumulative_probs = torch.cumsum(sorted_probs, dim=-1)
sorted_mask = cumulative_probs > top_p
sorted_mask[..., 1:] = sorted_mask[..., :-1].clone()
sorted_mask[..., 0] = False
removal_mask = torch.zeros_like(sorted_mask, dtype=torch.bool)
removal_mask.scatter_(dim=-1, index=sorted_indices, src=sorted_mask)
scaled_logits = scaled_logits.masked_fill(removal_mask, float("-inf"))
probs = F.softmax(scaled_logits, dim=-1)
return torch.multinomial(probs, num_samples=1)
def build_parser() -> argparse.ArgumentParser:
parser = argparse.ArgumentParser(description="Run a trained student checkpoint.")
parser.add_argument(
"--checkpoint",
default="checkpoints/student_final.pt",
help="Path to the student checkpoint.",
)
parser.add_argument(
"--device",
default=None,
help="Device to run on. Defaults to cuda if available, otherwise cpu.",
)
parser.add_argument(
"--tokenizer-path",
default=None,
help="Optional tokenizer path. Defaults to models/teacher if present.",
)
parser.add_argument(
"--prompt",
default=None,
help="Prompt to generate from.",
)
parser.add_argument(
"--max-new-tokens",
type=int,
default=64,
help="Maximum number of tokens to generate.",
)
parser.add_argument(
"--temperature",
type=float,
default=0.8,
help="Sampling temperature. Use 0 for greedy decoding.",
)
parser.add_argument(
"--top-p",
type=float,
default=0.95,
help="Nucleus sampling threshold.",
)
parser.add_argument(
"--top-k",
type=int,
default=50,
help="Top-k sampling cutoff. Use 0 to disable.",
)
parser.add_argument(
"--repetition-penalty",
type=float,
default=1.1,
help="Penalty for already generated tokens. Use 1.0 to disable.",
)
parser.add_argument(
"--interactive",
action="store_true",
help="Start an interactive prompt loop.",
)
return parser
def interactive_loop(runner: StudentRunner, args: argparse.Namespace) -> None:
print("Interactive mode. Type 'exit' or 'quit' to stop.")
while True:
try:
prompt = input("\nPrompt> ").strip()
except EOFError:
print()
break
if prompt.lower() in {"exit", "quit"}:
break
if not prompt:
continue
output = runner.generate(
prompt=prompt,
max_new_tokens=args.max_new_tokens,
temperature=args.temperature,
top_p=args.top_p,
top_k=args.top_k,
repetition_penalty=args.repetition_penalty,
)
print(f"\n{output}")
def main() -> None:
args = build_parser().parse_args()
runner = StudentRunner(
checkpoint_path=args.checkpoint,
device=args.device,
tokenizer_path=args.tokenizer_path,
)
if args.interactive:
interactive_loop(runner, args)
return
if not args.prompt:
raise SystemExit("Provide --prompt for one-shot generation or use --interactive.")
output = runner.generate(
prompt=args.prompt,
max_new_tokens=args.max_new_tokens,
temperature=args.temperature,
top_p=args.top_p,
top_k=args.top_k,
repetition_penalty=args.repetition_penalty,
)
print(output)
if __name__ == "__main__":
main()
|