File size: 10,176 Bytes
ed1b365
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""Codette LoRA Adapter Test Suite

Tests the newton and davinci adapters:
1. Weight inspection (no base model needed)
2. Full inference comparison (loads base model)

Hardware: Intel Arc 140V (8GB XPU) + 16GB RAM
Strategy: CPU float16 inference with LoRA merge
"""

import os, sys, json, time

# SYCL DLLs
os.environ["PATH"] = r"J:\Lib\site-packages\Library\bin" + os.pathsep + os.environ.get("PATH", "")
os.environ["TOKENIZERS_PARALLELISM"] = "false"

import torch
import safetensors.torch as st
from pathlib import Path

ADAPTER_DIR = Path("J:/codette-training-lab/adapters/hf_download")
NEWTON_DIR = ADAPTER_DIR / "newton"
DAVINCI_DIR = ADAPTER_DIR / "davinci"
BASE_MODEL = "meta-llama/Llama-3.1-8B-Instruct"

# ================================================================
# PHASE 1: Quick Adapter Weight Validation (no base model needed)
# ================================================================
def phase1_weight_inspection():
    print("=" * 60)
    print("PHASE 1: Adapter Weight Inspection")
    print("=" * 60)

    for name, adapter_dir in [("newton", NEWTON_DIR), ("davinci", DAVINCI_DIR)]:
        print(f"\n--- {name.upper()} Adapter ---")

        # Load adapter config
        with open(adapter_dir / "adapter_config.json") as f:
            config = json.load(f)
        print(f"  Base model: {config['base_model_name_or_path']}")
        print(f"  LoRA rank: {config['r']}, alpha: {config['lora_alpha']}")
        print(f"  Targets: {config['target_modules']}")
        print(f"  PEFT version: {config['peft_version']}")

        # Load adapter weights
        weights = st.load_file(str(adapter_dir / "adapter_model.safetensors"))
        print(f"  Weight tensors: {len(weights)}")

        total_params = 0
        layer_stats = {}
        for key, tensor in sorted(weights.items()):
            params = tensor.numel()
            total_params += params
            mean = tensor.float().mean().item()
            std = tensor.float().std().item()
            abs_mean = tensor.float().abs().mean().item()
            nonzero = (tensor != 0).float().mean().item() * 100

            # Group by layer type
            if "lora_A" in key:
                ltype = "lora_A"
            elif "lora_B" in key:
                ltype = "lora_B"
            else:
                ltype = "other"

            if ltype not in layer_stats:
                layer_stats[ltype] = {"count": 0, "means": [], "stds": [], "abs_means": []}
            layer_stats[ltype]["count"] += 1
            layer_stats[ltype]["means"].append(mean)
            layer_stats[ltype]["stds"].append(std)
            layer_stats[ltype]["abs_means"].append(abs_mean)

        print(f"  Total LoRA params: {total_params:,}")
        print(f"  File size: {(adapter_dir / 'adapter_model.safetensors').stat().st_size / 1024**2:.1f} MB")

        for ltype, stats in layer_stats.items():
            avg_mean = sum(stats["means"]) / len(stats["means"])
            avg_std = sum(stats["stds"]) / len(stats["stds"])
            avg_abs = sum(stats["abs_means"]) / len(stats["abs_means"])
            print(f"  {ltype} ({stats['count']} tensors):")
            print(f"    avg mean={avg_mean:.6f}, avg std={avg_std:.6f}, avg |w|={avg_abs:.6f}")

    # Compare newton vs davinci
    print(f"\n--- Weight Divergence (newton vs davinci) ---")
    newton_w = st.load_file(str(NEWTON_DIR / "adapter_model.safetensors"))
    davinci_w = st.load_file(str(DAVINCI_DIR / "adapter_model.safetensors"))

    divergences = []
    for key in sorted(newton_w.keys()):
        if key in davinci_w:
            diff = (newton_w[key].float() - davinci_w[key].float()).abs().mean().item()
            divergences.append((key.split(".")[-2] + "." + key.split(".")[-1], diff))

    divergences.sort(key=lambda x: x[1], reverse=True)
    print(f"  Total shared keys: {len(divergences)}")
    print(f"  Top 5 most divergent layers:")
    for name, div in divergences[:5]:
        print(f"    {name}: {div:.6f}")
    avg_div = sum(d for _, d in divergences) / len(divergences)
    print(f"  Average divergence: {avg_div:.6f}")

    if avg_div > 0.001:
        print(f"  PASS: Adapters learned distinct representations (div={avg_div:.6f} >> 0)")
    else:
        print(f"  WARN: Adapters may be too similar (div={avg_div:.6f})")

    return True


# ================================================================
# PHASE 2: Full Inference Test
# ================================================================
def phase2_inference_test():
    print(f"\n{'=' * 60}")
    print("PHASE 2: Full Inference Test")
    print("=" * 60)

    from transformers import AutoModelForCausalLM, AutoTokenizer
    from peft import PeftModel
    import gc

    # Load tokenizer
    print("Loading tokenizer...")
    tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
    if tokenizer.pad_token is None:
        tokenizer.pad_token = tokenizer.eos_token

    # Load base model on CPU with disk offload to avoid OOM
    print("Loading base model (CPU + disk offload, float16)...")
    os.makedirs("J:/tmp/offload", exist_ok=True)
    start = time.time()
    model = AutoModelForCausalLM.from_pretrained(
        BASE_MODEL,
        dtype=torch.float16,
        device_map={
            "": "cpu",
        },
        low_cpu_mem_usage=True,
    )
    print(f"  Base model loaded in {time.time()-start:.0f}s")

    # Test prompt - same question, different perspectives expected
    test_prompt = "Explain why objects fall to the ground."
    messages = [
        {"role": "system", "content": "You are a helpful assistant. Answer concisely in 2-3 sentences."},
        {"role": "user", "content": test_prompt},
    ]
    input_text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
    inputs = tokenizer(input_text, return_tensors="pt")

    gen_kwargs = dict(
        max_new_tokens=150,
        temperature=0.7,
        top_p=0.9,
        do_sample=True,
        pad_token_id=tokenizer.eos_token_id,
    )

    # --- Base model response ---
    print(f"\n--- BASE MODEL (no adapter) ---")
    print(f"Prompt: {test_prompt}")
    start = time.time()
    with torch.no_grad():
        output = model.generate(**inputs, **gen_kwargs)
    base_response = tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
    print(f"Response ({time.time()-start:.1f}s): {base_response}")

    # --- Newton adapter ---
    print(f"\n--- NEWTON ADAPTER ---")
    print("Loading newton adapter...")
    start = time.time()
    newton_model = PeftModel.from_pretrained(model, str(NEWTON_DIR))
    newton_model.eval()
    print(f"  Adapter loaded in {time.time()-start:.1f}s")

    start = time.time()
    with torch.no_grad():
        output = newton_model.generate(**inputs, **gen_kwargs)
    newton_response = tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
    print(f"Response ({time.time()-start:.1f}s): {newton_response}")

    # Unload newton
    del newton_model
    import gc; gc.collect()

    # --- DaVinci adapter ---
    print(f"\n--- DAVINCI ADAPTER ---")
    print("Loading davinci adapter...")
    start = time.time()
    davinci_model = PeftModel.from_pretrained(model, str(DAVINCI_DIR))
    davinci_model.eval()
    print(f"  Adapter loaded in {time.time()-start:.1f}s")

    start = time.time()
    with torch.no_grad():
        output = davinci_model.generate(**inputs, **gen_kwargs)
    davinci_response = tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
    print(f"Response ({time.time()-start:.1f}s): {davinci_response}")

    del davinci_model
    gc.collect()

    # --- Second test: creative/philosophical prompt ---
    test_prompt2 = "What is the relationship between consciousness and the physical world?"
    messages2 = [
        {"role": "system", "content": "You are a helpful assistant. Answer concisely in 2-3 sentences."},
        {"role": "user", "content": test_prompt2},
    ]
    input_text2 = tokenizer.apply_chat_template(messages2, tokenize=False, add_generation_prompt=True)
    inputs2 = tokenizer(input_text2, return_tensors="pt")

    print(f"\n{'=' * 60}")
    print(f"TEST 2: {test_prompt2}")
    print(f"{'=' * 60}")

    # Newton on philosophical question
    print(f"\n--- NEWTON on consciousness ---")
    newton_model = PeftModel.from_pretrained(model, str(NEWTON_DIR))
    newton_model.eval()
    start = time.time()
    with torch.no_grad():
        output = newton_model.generate(**inputs2, **gen_kwargs)
    response = tokenizer.decode(output[0][inputs2["input_ids"].shape[1]:], skip_special_tokens=True)
    print(f"Response ({time.time()-start:.1f}s): {response}")
    del newton_model; gc.collect()

    # DaVinci on philosophical question
    print(f"\n--- DAVINCI on consciousness ---")
    davinci_model = PeftModel.from_pretrained(model, str(DAVINCI_DIR))
    davinci_model.eval()
    start = time.time()
    with torch.no_grad():
        output = davinci_model.generate(**inputs2, **gen_kwargs)
    response = tokenizer.decode(output[0][inputs2["input_ids"].shape[1]:], skip_special_tokens=True)
    print(f"Response ({time.time()-start:.1f}s): {response}")
    del davinci_model; gc.collect()

    # Cleanup
    del model
    gc.collect()

    print(f"\n{'=' * 60}")
    print("INFERENCE TESTS COMPLETE")
    print(f"{'=' * 60}")
    return True


# ================================================================
# MAIN
# ================================================================
if __name__ == "__main__":
    print("Codette LoRA Adapter Test Suite")
    print(f"PyTorch: {torch.__version__}")
    print(f"XPU: {torch.xpu.is_available()}")
    print(f"Adapters: {ADAPTER_DIR}")
    print()

    # Phase 1 is fast - always run
    phase1_weight_inspection()

    # Phase 2 needs base model download (~16GB) and lots of RAM
    print("\n" + "=" * 60)
    if "--inference" in sys.argv or "--full" in sys.argv:
        phase2_inference_test()
    else:
        print("Skipping inference test (run with --inference to enable)")
        print("  Note: Will download ~16GB base model and needs ~16GB RAM")