File size: 11,609 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 | #!/usr/bin/env python3
"""
Inference & Evaluation for Qwen-0.8B Student Model
"""
import torch
import torch.nn.functional as F
from transformers import AutoTokenizer
from pathlib import Path
import logging
import time
from typing import Dict, List
logging.basicConfig(level=logging.INFO)
logger = logging.getLogger(__name__)
# ============================================================================
# INFERENCE
# ============================================================================
class StudentInference:
"""Run inference with distilled student model"""
def __init__(self, checkpoint_path: str, device: str = "cuda"):
self.device = torch.device(device)
self.checkpoint_path = checkpoint_path
logger.info(f"Loading checkpoint: {checkpoint_path}")
self.checkpoint = torch.load(checkpoint_path, map_location=device)
self.config = self.checkpoint['config']
# Reconstruct student model
from qwen_distill import QwenDistillationConfig, QwenStudentModel
config_obj = QwenDistillationConfig()
for key, val in self.config.items():
setattr(config_obj, key, val)
self.model = QwenStudentModel(config_obj).to(device)
self.model.load_state_dict(self.checkpoint['model_state_dict'])
self.model.eval()
# Load tokenizer
self.tokenizer = AutoTokenizer.from_pretrained(
config_obj.teacher_model_name,
trust_remote_code=True,
)
self.tokenizer.pad_token = self.tokenizer.eos_token
logger.info(f"✓ Model loaded. Parameters: {sum(p.numel() for p in self.model.parameters())/1e6:.1f}M")
def generate(
self,
prompt: str,
max_length: int = 100,
temperature: float = 0.7,
top_p: float = 0.95,
) -> str:
"""Generate text from prompt"""
input_ids = self.tokenizer.encode(prompt, return_tensors="pt").to(self.device)
with torch.no_grad():
for _ in range(max_length):
outputs = self.model(input_ids)
logits = outputs['logits'][:, -1, :]
# Temperature scaling
logits = logits / temperature
# Top-p sampling
probs = F.softmax(logits, dim=-1)
sorted_probs, sorted_indices = torch.sort(probs, descending=True)
cumsum_probs = torch.cumsum(sorted_probs, dim=-1)
# Remove tokens with cumulative probability > top_p
sorted_indices_to_remove = cumsum_probs > top_p
sorted_indices_to_remove[..., 1:] = sorted_indices_to_remove[..., :-1].clone()
sorted_indices_to_remove[..., 0] = 0
indices_to_remove = sorted_indices[sorted_indices_to_remove]
logits[0, indices_to_remove] = -float('inf')
# Sample
next_token = torch.multinomial(F.softmax(logits, dim=-1), num_samples=1)
input_ids = torch.cat([input_ids, next_token], dim=-1)
if next_token.item() == self.tokenizer.eos_token_id:
break
return self.tokenizer.decode(input_ids[0], skip_special_tokens=True)
def inference_speed_test(self, prompt: str = "The future of AI", num_runs: int = 10):
"""Benchmark inference speed"""
logger.info(f"Running speed test ({num_runs} iterations)...")
input_ids = self.tokenizer.encode(prompt, return_tensors="pt").to(self.device)
# Warmup
with torch.no_grad():
_ = self.model(input_ids)
# Measure
times = []
with torch.no_grad():
for _ in range(num_runs):
torch.cuda.synchronize()
start = time.time()
_ = self.model(input_ids)
torch.cuda.synchronize()
times.append(time.time() - start)
avg_time = sum(times) / len(times) * 1000 # ms
logger.info(f"Average inference time: {avg_time:.1f}ms")
logger.info(f"Throughput: {1000/avg_time:.1f} samples/sec")
return {
'avg_time_ms': avg_time,
'throughput': 1000 / avg_time,
}
# ============================================================================
# EVALUATION
# ============================================================================
class StudentEvaluator:
"""Evaluate student model quality"""
def __init__(self, student_checkpoint: str, teacher_model_name: str, device: str = "cuda"):
self.device = torch.device(device)
self.student_inf = StudentInference(student_checkpoint, device)
# Load teacher
from transformers import AutoModelForCausalLM
logger.info(f"Loading teacher: {teacher_model_name}")
self.teacher = AutoModelForCausalLM.from_pretrained(
teacher_model_name,
torch_dtype=torch.float16,
device_map="auto",
trust_remote_code=True,
)
self.teacher.eval()
self.tokenizer = self.student_inf.tokenizer
def compute_perplexity(self, texts: List[str], max_length: int = 256) -> float:
"""Compute perplexity on text samples"""
total_loss = 0.0
num_tokens = 0
self.student_inf.model.eval()
with torch.no_grad():
for text in texts:
enc = self.tokenizer(
text,
max_length=max_length,
truncation=True,
return_tensors="pt",
).to(self.device)
outputs = self.student_inf.model(enc['input_ids'])
logits = outputs['logits']
# Compute cross-entropy loss
loss = F.cross_entropy(
logits[0, :-1, :],
enc['input_ids'][0, 1:],
reduction='mean'
)
total_loss += loss.item()
num_tokens += enc['input_ids'].numel()
perplexity = torch.exp(torch.tensor(total_loss / len(texts))).item()
logger.info(f"Student perplexity: {perplexity:.2f}")
return perplexity
def compute_teacher_perplexity(self, texts: List[str], max_length: int = 256) -> float:
"""Compute perplexity on teacher for comparison"""
total_loss = 0.0
self.teacher.eval()
with torch.no_grad():
for text in texts:
enc = self.tokenizer(
text,
max_length=max_length,
truncation=True,
return_tensors="pt",
).to(self.device)
outputs = self.teacher(enc['input_ids'], output_hidden_states=True)
logits = outputs.logits
loss = F.cross_entropy(
logits[0, :-1, :],
enc['input_ids'][0, 1:],
reduction='mean'
)
total_loss += loss.item()
perplexity = torch.exp(torch.tensor(total_loss / len(texts))).item()
logger.info(f"Teacher perplexity: {perplexity:.2f}")
return perplexity
def top_k_agreement(self, texts: List[str], k: int = 5) -> float:
"""Measure how well student matches teacher top-k predictions"""
match_count = 0
total = 0
self.student_inf.model.eval()
self.teacher.eval()
with torch.no_grad():
for text in texts:
enc = self.tokenizer(
text,
return_tensors="pt",
max_length=256,
truncation=True,
).to(self.device)
student_out = self.student_inf.model(enc['input_ids'])
student_logits = student_out['logits']
teacher_out = self.teacher(enc['input_ids'])
teacher_logits = teacher_out.logits
# Top-k tokens
_, student_topk = torch.topk(student_logits, k, dim=-1)
_, teacher_topk = torch.topk(teacher_logits, k, dim=-1)
# Count matches
matches = (student_topk == teacher_topk).float().sum().item()
match_count += matches
total += student_topk.numel()
agreement = match_count / total if total > 0 else 0.0
logger.info(f"Top-{k} agreement with teacher: {agreement*100:.1f}%")
return agreement
def generate_comparison(self, prompt: str = "The future of AI", max_length: int = 100):
"""Compare student vs teacher generation"""
logger.info(f"\nPrompt: {prompt}\n")
# Student generation
student_text = self.student_inf.generate(prompt, max_length=max_length)
logger.info(f"Student:\n{student_text}\n")
# Teacher generation
input_ids = self.tokenizer.encode(prompt, return_tensors="pt").to(self.device)
with torch.no_grad():
outputs = self.teacher.generate(
input_ids,
max_length=max_length,
temperature=0.7,
top_p=0.95,
)
teacher_text = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
logger.info(f"Teacher:\n{teacher_text}\n")
# ============================================================================
# MAIN
# ============================================================================
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser()
parser.add_argument("--checkpoint", default="checkpoints/student_final.pt", help="Student checkpoint path")
parser.add_argument("--teacher", default="Qwen/Qwen2.5-0.5B", help="Teacher model name")
parser.add_argument("--prompt", default="The future of artificial intelligence", help="Generation prompt")
parser.add_argument("--speed", action="store_true", help="Run speed test")
parser.add_argument("--eval", action="store_true", help="Run evaluation")
args = parser.parse_args()
# Simple generation
logger.info("Loading student model...")
inference = StudentInference(args.checkpoint)
logger.info(f"Generating from prompt: {args.prompt}\n")
text = inference.generate(args.prompt, max_length=100)
print(text)
if args.speed:
logger.info("\nBenchmarking speed...")
inference.inference_speed_test()
if args.eval:
logger.info("\nRunning evaluation...")
evaluator = StudentEvaluator(args.checkpoint, args.teacher)
# Test data
test_texts = [
"Artificial intelligence is transforming industries.",
"Machine learning models require careful tuning.",
"Distillation compresses large models efficiently.",
]
evaluator.compute_perplexity(test_texts)
evaluator.compute_teacher_perplexity(test_texts)
evaluator.top_k_agreement(test_texts, k=5)
evaluator.generate_comparison(args.prompt, max_length=100)
|