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
|
"""Test different strategies to fix the output length issue."""
import sys
import os
import time
import torch
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from data.longlamp import load_longlamp, select_k_profile_items
from data.templates import build_query_prompt
from models.qwen_wrapper import QwenWrapper
from models.cvh import CVHHead, UnconditionalHead
from adapt.cache_hidden import cache_support_hidden_states
from adapt.fit_theta import fit_theta
from eval.metrics import evaluate_all
def run_with_blend(wrapper, examples, support_sets, head_module, d=64,
beta=0.05, steps=30, lr=0.05, blend_gamma=0.5,
min_new_tokens=64, max_new_tokens=512):
"""Run CVH with logit blending: logits = (1-gamma)*base + gamma*cvh"""
device = 'cuda:1'
lm_head_bias = None
if hasattr(wrapper.model.lm_head, 'bias') and wrapper.model.lm_head.bias is not None:
lm_head_bias = wrapper.model.lm_head.bias.data
predictions = []
theta_norms = []
for i, (ex, support) in enumerate(zip(examples, support_sets)):
cached_h = cache_support_hidden_states(wrapper, support, ex['task'])
if not cached_h:
prompt = build_query_prompt(ex['query_input'], ex['task'])
pred = wrapper.generate_base(prompt, max_new_tokens=max_new_tokens)
predictions.append(pred)
continue
theta = fit_theta(
cached_h=cached_h,
lm_head_weight=wrapper.lm_head_weight,
lm_head_bias=lm_head_bias,
head_module=head_module,
d=d, lr=lr, steps=steps, beta=beta, lam=1e-4,
max_grad_norm=5.0, device=device, verbose=False,
)
theta_norms.append(theta.norm().item())
# Generate with blending
prompt = build_query_prompt(ex['query_input'], ex['task'])
pred = generate_with_blend(
wrapper, prompt, theta, head_module,
gamma=blend_gamma, max_new_tokens=max_new_tokens,
min_new_tokens=min_new_tokens,
)
predictions.append(pred)
del cached_h, theta
torch.cuda.empty_cache()
if (i + 1) % 20 == 0:
print(f" {i+1}/{len(examples)}")
avg_norm = sum(theta_norms) / max(len(theta_norms), 1)
avg_len = sum(len(p.split()) for p in predictions) / max(len(predictions), 1)
return predictions, avg_norm, avg_len
def generate_with_blend(wrapper, input_text, theta, head_module,
gamma=0.5, max_new_tokens=512, min_new_tokens=64):
"""Generate with blended base + CVH logits."""
chat_messages = [
{"role": "system", "content": "You are a helpful writing assistant."},
{"role": "user", "content": input_text},
]
prompt_text = wrapper.tokenizer.apply_chat_template(
chat_messages, tokenize=False, add_generation_prompt=True
)
input_ids = wrapper.tokenizer.encode(prompt_text, return_tensors="pt").to(wrapper.device)
generated_ids = []
past_key_values = None
for step in range(max_new_tokens):
if step == 0:
cur_input = input_ids
else:
cur_input = torch.tensor([[generated_ids[-1]]], device=wrapper.device)
with torch.no_grad():
outputs = wrapper.model(
input_ids=cur_input,
past_key_values=past_key_values,
output_hidden_states=True,
use_cache=True,
return_dict=True,
)
past_key_values = outputs.past_key_values
last_hidden = outputs.hidden_states[-1][:, -1, :] # (1, H)
# Base logits
base_logits = torch.nn.functional.linear(
last_hidden.to(wrapper.lm_head_weight.dtype),
wrapper.lm_head_weight,
wrapper.model.lm_head.bias if hasattr(wrapper.model.lm_head, 'bias') and wrapper.model.lm_head.bias is not None else None,
).float()
# CVH logits
h_prime = head_module.forward_fn(last_hidden.float(), theta)
cvh_logits = torch.nn.functional.linear(
h_prime.to(wrapper.lm_head_weight.dtype),
wrapper.lm_head_weight,
wrapper.model.lm_head.bias if hasattr(wrapper.model.lm_head, 'bias') and wrapper.model.lm_head.bias is not None else None,
).float()
# Blend logits
logits = (1 - gamma) * base_logits + gamma * cvh_logits
# Suppress EOS before min_new_tokens
if step < min_new_tokens and wrapper.tokenizer.eos_token_id is not None:
logits[0, wrapper.tokenizer.eos_token_id] = float('-inf')
next_token = logits.argmax(dim=-1).item()
if next_token == wrapper.tokenizer.eos_token_id:
break
generated_ids.append(next_token)
return wrapper.tokenizer.decode(generated_ids, skip_special_tokens=True)
def main():
N = 50
print(f"Loading data ({N} examples)...")
examples = load_longlamp('product_review_user', split='val')[:N]
K = 4
support_sets = [select_k_profile_items(ex['profile_items'], K, seed=0) for ex in examples]
references = [ex['target_output'] for ex in examples]
support_texts = [[s['support_output'] for s in ss] for ss in support_sets]
avg_ref_len = sum(len(r.split()) for r in references) / len(references)
print(f"Avg reference length: {avg_ref_len:.0f} words")
print("Loading model...")
wrapper = QwenWrapper('Qwen/Qwen2.5-1.5B-Instruct', device='cuda:1')
H = wrapper.hidden_size
device = 'cuda:1'
# Base with min_new_tokens=200
print("\n=== Base ===")
base_preds = []
for ex in examples:
prompt = build_query_prompt(ex['query_input'], ex['task'])
pred = wrapper.generate_base(prompt, max_new_tokens=512, temperature=0.0)
base_preds.append(pred)
base_r = evaluate_all(base_preds, references, support_texts)
base_len = sum(len(p.split()) for p in base_preds) / len(base_preds)
print(f" R-L: {base_r['rougeL']:.4f}, METEOR: {base_r['meteor']:.4f}, SFD: {base_r['sfd']:.4f}, len: {base_len:.0f}")
results = {'Base': {**base_r, 'avg_len': base_len}}
head = CVHHead(H, d=64, alpha=0.1, basis_seed=42).to(device)
uncond = UnconditionalHead(H, d=64, alpha=0.1, basis_seed=42).to(device)
configs = [
# Standard CVH with higher min_new_tokens
('CVH min=200', head, 1.0, 200),
# Blended CVH with different gammas
('CVH blend=0.3 min=64', head, 0.3, 64),
('CVH blend=0.5 min=64', head, 0.5, 64),
('CVH blend=0.7 min=64', head, 0.7, 64),
('CVH blend=0.5 min=128', head, 0.5, 128),
# Uncond blend
('Uncond blend=0.5 min=64', uncond, 0.5, 64),
]
for name, head_mod, gamma, min_tok in configs:
print(f"\n=== {name} ===")
t0 = time.time()
preds, avg_norm, avg_len = run_with_blend(
wrapper, examples, support_sets, head_mod, d=64,
beta=0.05, steps=30, lr=0.05, blend_gamma=gamma,
min_new_tokens=min_tok, max_new_tokens=512,
)
elapsed = time.time() - t0
r = evaluate_all(preds, references, support_texts)
results[name] = {**r, 'avg_len': avg_len}
print(f" R-L: {r['rougeL']:.4f}, METEOR: {r['meteor']:.4f}, SFD: {r['sfd']:.4f}, "
f"|theta|: {avg_norm:.3f}, len: {avg_len:.0f}, time: {elapsed:.0f}s")
# Summary
print("\n" + "=" * 100)
print(f"{'Config':<30} {'R-1':<8} {'R-L':<8} {'METEOR':<8} {'SFD':<8} {'Len':<6}")
print("-" * 100)
for name, r in results.items():
print(f"{name:<30} {r['rouge1']:<8.4f} {r['rougeL']:<8.4f} "
f"{r['meteor']:<8.4f} {r['sfd']:<8.4f} {r.get('avg_len', 0):<6.0f}")
if __name__ == '__main__':
main()
|
