File size: 9,987 Bytes
a1d6a7c
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
"""
End-to-end validation of the complete Fusion-LLM pipeline.

Tests:
1. FusionModel training - loss decreases over 50 steps
2. ThinkingDialModel generate - different depths produce different logits
3. GRPO training pipeline - train_step completes without errors
4. Full generate loop - model produces valid token sequences

Run: python train/e2e_validation.py
"""
import sys
import torch
import torch.nn as nn
from pathlib import Path

sys.path.insert(0, str(Path(__file__).parent.parent))

from models.fusion_model import FusionModel, FusionConfig
from models.thinking_dial import ThinkingDialModel, ThinkingConfig, GRPOTrainer, GRPOConfig
from train.model_utils import create_local_model

DEVICE = "cuda" if torch.cuda.is_available() else "cpu"
print(f"[E2E] Device: {DEVICE}")
print()

PASS_COUNT = 0
FAIL_COUNT = 0

def check(name, condition, detail=""):
    global PASS_COUNT, FAIL_COUNT
    if condition:
        PASS_COUNT += 1
        print(f"  [PASS] {name}")
    else:
        FAIL_COUNT += 1
        print(f"  [FAIL] {name} {detail}")


def test_training_loss_decrease():
    """Test 1: FusionModel training - loss should decrease"""
    print("\n=== Test 1: FusionModel Training Loss ===")
    
    config = FusionConfig(
        vocab_size=500,
        hidden_size=128,
        num_hidden_layers=2,
        num_attention_heads=4,
        num_key_value_heads=2,
        intermediate_size=256,
        block_size=8,
        latent_dim=16,
        window_size=64,
    )
    model = FusionModel(config)
    model.train()
    model.to(DEVICE)
    
    # Count params
    param_count = sum(p.numel() for p in model.parameters())
    print(f"  Parameters: {param_count:,}")
    
    optimizer = torch.optim.AdamW(model.parameters(), lr=1e-3)
    batch_size, seq_len = 4, 32
    
    # Synthetic data
    input_ids = torch.randint(1, config.vocab_size, (batch_size, seq_len), device=DEVICE)
    
    losses = []
    for step in range(50):
        optimizer.zero_grad()
        outputs = model(input_ids=input_ids, labels=input_ids)
        loss = outputs.loss
        loss.backward()
        optimizer.step()
        losses.append(loss.item())
        
        if (step + 1) % 25 == 0:
            print(f"    Step {step+1:3d}: Loss = {loss.item():.4f}")
    
    initial = losses[0]
    final = losses[-1]
    decreased = final < initial
    print(f"    Initial: {initial:.4f}, Final: {final:.4f}, Delta: {final - initial:+.4f}")
    check("Loss decreased over 50 steps", decreased, f"(final {final:.4f} >= initial {initial:.4f})")
    
    del model, optimizer
    torch.cuda.empty_cache() if torch.cuda.is_available() else None
    return decreased


def test_thinking_dial_different_depths():
    """Test 2: Different thinking_depths produce different logits"""
    print("\n=== Test 2: ThinkingDialModel Depth Sensitivity ===")
    
    config = FusionConfig(
        vocab_size=500,
        hidden_size=128,
        num_hidden_layers=2,
        num_attention_heads=4,
        num_key_value_heads=2,
        intermediate_size=256,
        block_size=8,
        latent_dim=16,
        window_size=64,
    )
    base_model = FusionModel(config)
    base_model.eval()
    base_model.to(DEVICE)
    
    thinking_config = ThinkingConfig(num_thinking_depths=4)
    td_model = ThinkingDialModel(base_model, thinking_config)
    
    input_ids = torch.randint(1, config.vocab_size, (1, 8), device=DEVICE)
    
    with torch.no_grad():
        outputs_depth0 = td_model.generate(input_ids, max_new_tokens=4, thinking_depth=0, do_sample=False)
        outputs_depth3 = td_model.generate(input_ids, max_new_tokens=4, thinking_depth=3, do_sample=False)
    
    same = torch.equal(outputs_depth0, outputs_depth3)
    check("Different depths produce different outputs", not same, "(outputs identical)")
    
    print(f"    Depth 0 output: {outputs_depth0[0, 8:].tolist()}")
    print(f"    Depth 3 output: {outputs_depth3[0, 8:].tolist()}")
    
    del td_model, base_model
    torch.cuda.empty_cache() if torch.cuda.is_available() else None
    return not same


def test_grpo_train_step():
    """Test 3: GRPO train_step completes without errors"""
    print("\n=== Test 3: GRPO Train Step ===")
    
    config = FusionConfig(
        vocab_size=500,
        hidden_size=128,
        num_hidden_layers=2,
        num_attention_heads=4,
        num_key_value_heads=2,
        intermediate_size=256,
        block_size=8,
        latent_dim=16,
        window_size=64,
    )
    base_model = FusionModel(config)
    base_model.train()
    base_model.to(DEVICE)
    
    thinking_config = ThinkingConfig(num_thinking_depths=4)
    td_model = ThinkingDialModel(base_model, thinking_config)
    td_model.train()
    
    grpo_config = GRPOConfig(
        grpo_sample_size=2,
        kl_coef=0.1,
    )
    trainer = GRPOTrainer(td_model, grpo_config=grpo_config, thinking_config=thinking_config)
    
    input_ids = torch.randint(1, config.vocab_size, (2, 8), device=DEVICE)
    
    try:
        result = trainer.train_step(input_ids, thinking_depth=2)
        check("train_step completed", True)
        check("train_step returned loss", "loss" in result, f"loss={result.get('loss')}")
        check("train_step returned rewards", "mean_reward" in result)
        check("step_count incremented", trainer.step_count == 1, f"step_count={trainer.step_count}")
        print(f"    Loss: {result['loss']:.4f}, Mean Reward: {result['mean_reward']:.4f}")
        # Note: loss can be 0 if all rewards are equal (identical dummy text)
    except Exception as e:
        check("train_step completed", False, f"Exception: {e}")
        import traceback
        traceback.print_exc()
    
    del td_model, base_model, trainer
    torch.cuda.empty_cache() if torch.cuda.is_available() else None
    return True


def test_generate_loop():
    """Test 4: Full generate loop produces valid sequences"""
    print("\n=== Test 4: Full Generate Loop ===")
    
    config = FusionConfig(
        vocab_size=500,
        hidden_size=128,
        num_hidden_layers=2,
        num_attention_heads=4,
        num_key_value_heads=2,
        intermediate_size=256,
        block_size=8,
        latent_dim=16,
        window_size=64,
    )
    model = FusionModel(config)
    model.eval()
    model.to(DEVICE)
    
    input_ids = torch.randint(1, config.vocab_size, (2, 8), device=DEVICE)
    
    with torch.no_grad():
        # Greedy
        out_greedy = model.generate(input_ids, max_new_tokens=16, do_sample=False)
        check("Greedy generate shape correct", out_greedy.shape == (2, 24), f"shape={out_greedy.shape}")
        
        # Sampling
        out_sample = model.generate(input_ids, max_new_tokens=16, do_sample=True, temperature=1.0)
        check("Sampled generate shape correct", out_sample.shape == (2, 24), f"shape={out_sample.shape}")
        
        # Greedy == Greedy (deterministic)
        out_greedy2 = model.generate(input_ids, max_new_tokens=16, do_sample=False)
        check("Greedy is deterministic", torch.equal(out_greedy, out_greedy2))
        
        # All tokens valid
        valid = (out_greedy >= 0).all() and (out_greedy < config.vocab_size).all()
        check("All output tokens valid", valid)
        
        # Prefix preserved
        prefix_match = torch.equal(out_greedy[:, :8], input_ids)
        check("Input prefix preserved", prefix_match)
    
    print(f"    Greedy output[0]: {out_greedy[0].tolist()}")
    print(f"    Sampled output[0]: {out_sample[0].tolist()}")
    
    del model
    torch.cuda.empty_cache() if torch.cuda.is_available() else None
    return True


def test_thinking_dial_with_thinking_depth():
    """Test 5: ThinkingDial generate_with_thinking produces coherent results"""
    print("\n=== Test 5: ThinkingDial Thinking Depth Integration ===")
    torch.manual_seed(42)
    
    config = FusionConfig(
        vocab_size=500,
        hidden_size=128,
        num_hidden_layers=2,
        num_attention_heads=4,
        num_key_value_heads=2,
        intermediate_size=256,
        block_size=8,
        latent_dim=16,
        window_size=64,
    )
    base_model = FusionModel(config)
    base_model.eval()
    base_model.to(DEVICE)
    
    thinking_config = ThinkingConfig(num_thinking_depths=4)
    td_model = ThinkingDialModel(base_model, thinking_config)
    
    input_ids = torch.randint(1, config.vocab_size, (1, 8), device=DEVICE)
    
    results = {}
    for depth in range(4):
        with torch.no_grad():
            out = td_model.generate(input_ids, max_new_tokens=8, thinking_depth=depth, do_sample=False)
            results[depth] = out[0, 8:].tolist()
            print(f"    Depth {depth}: {results[depth]}")
    
    # All depths should produce valid sequences
    all_valid = all(
        all(0 <= t < config.vocab_size for t in results[d])
        for d in range(4)
    )
    check("All depths produce valid tokens", all_valid)
    
    # Note: with random weights, depth=0 vs depth=3 may sometimes produce
    # identical outputs. This is not a bug. Test 2 already verified depth
    # sensitivity in isolation. Here we just verify no crashes.
    check("All 4 depths generated without errors", True)
    
    del td_model, base_model
    torch.cuda.empty_cache() if torch.cuda.is_available() else None
    return True


if __name__ == "__main__":
    print("=" * 60)
    print("Fusion-LLM End-to-End Pipeline Validation")
    print("=" * 60)
    
    test_training_loss_decrease()
    test_thinking_dial_different_depths()
    test_grpo_train_step()
    test_generate_loop()
    test_thinking_dial_with_thinking_depth()
    
    print()
    print("=" * 60)
    print(f"Results: {PASS_COUNT} PASSED, {FAIL_COUNT} FAILED out of {PASS_COUNT + FAIL_COUNT}")
    if FAIL_COUNT == 0:
        print("ALL TESTS PASSED - Pipeline is runtime-verified!")
    else:
        print(f"FAILURES DETECTED - {FAIL_COUNT} test(s) need attention")
    print("=" * 60)
    
    sys.exit(1 if FAIL_COUNT > 0 else 0)