File size: 8,160 Bytes
74f2af5
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
#!/usr/bin/env python3
"""Codette LoRA Adapter Training - HuggingFace Jobs (A10G GPU)

Trains all 8 LoRA adapters on Llama 3.1 8B Instruct with QLoRA.
Robust error handling: upload failures won't kill the job.
"""

import json, os, gc, time, torch, traceback
from pathlib import Path
from huggingface_hub import hf_hub_download, HfApi
from datasets import Dataset
from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig
from peft import LoraConfig, get_peft_model, TaskType

try:
    from trl import SFTTrainer, SFTConfig
    USE_NEW_TRL = True
except ImportError:
    from trl import SFTTrainer
    from transformers import TrainingArguments
    USE_NEW_TRL = False

MODEL_NAME = "meta-llama/Llama-3.1-8B-Instruct"
DATASET_REPO = "Raiff1982/codette-training-data"
OUTPUT_REPO = "Raiff1982/codette-lora-adapters"
HF_TOKEN = os.environ.get("HF_TOKEN")

ADAPTERS = [
    ("newton", "newton_reasoning.jsonl", 3),
    ("davinci", "davinci_reasoning.jsonl", 3),
    ("empathy", "empathy_reasoning.jsonl", 3),
    ("philosophy", "philosophy_reasoning.jsonl", 3),
    ("quantum", "quantum_reasoning.jsonl", 3),
    ("consciousness", "consciousness_reasoning.jsonl", 3),
    ("multi_perspective", "multi_perspective_reasoning.jsonl", 3),
    ("systems_architecture", "systems_architecture_reasoning.jsonl", 3),
]

print("=" * 60)
print("Codette LoRA Training - HuggingFace Jobs (A10G GPU)")
print("=" * 60)
print(f"CUDA available: {torch.cuda.is_available()}")
if torch.cuda.is_available():
    print(f"GPU: {torch.cuda.get_device_name(0)}")
    print(f"VRAM: {torch.cuda.get_device_properties(0).total_memory/1024**3:.1f} GB")
print(f"HF Token present: {bool(HF_TOKEN)}")
print(f"USE_NEW_TRL: {USE_NEW_TRL}")

# --- Create output repo ---
api = HfApi(token=HF_TOKEN)
try:
    api.create_repo(OUTPUT_REPO, private=True, token=HF_TOKEN)
    print(f"Created output repo: {OUTPUT_REPO}")
except Exception as e:
    print(f"Output repo status: {e}")

# --- Download datasets ---
print("\nDownloading datasets...")
dataset_dir = Path("/tmp/datasets")
dataset_dir.mkdir(exist_ok=True)
for name, filename, _ in ADAPTERS:
    hf_hub_download(DATASET_REPO, filename, repo_type="dataset", local_dir=str(dataset_dir))
    print(f"  done: {name}")

# --- Load tokenizer ---
print("\nLoading tokenizer...")
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME, token=HF_TOKEN)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

# --- Load model ---
print("Loading model with 4-bit QLoRA...")
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
    bnb_4bit_use_double_quant=True,
)

model = AutoModelForCausalLM.from_pretrained(
    MODEL_NAME,
    quantization_config=bnb_config,
    device_map="auto",
    dtype=torch.bfloat16,
    trust_remote_code=True,
    use_cache=False,
    token=HF_TOKEN,
)
model.gradient_checkpointing_enable()
print(f"Model loaded! GPU: {torch.cuda.memory_allocated()/1024**3:.2f} GB")

# --- Training loop ---
results = {}
failed_uploads = []
total_start = time.time()

for adapter_name, dataset_file, epochs in ADAPTERS:
    print(f"\n{'=' * 60}")
    print(f"TRAINING: {adapter_name} ({epochs} epochs)")
    print(f"{'=' * 60}")
    start = time.time()

    # Load dataset
    dataset_path = dataset_dir / dataset_file
    examples = []
    with open(dataset_path) as f:
        for line in f:
            examples.append(json.loads(line))

    def format_example(ex):
        return {"text": tokenizer.apply_chat_template(ex["messages"], tokenize=False)}

    dataset = Dataset.from_list(examples).map(format_example, remove_columns=["messages"])
    print(f"  Dataset: {len(dataset)} examples")

    # Configure LoRA
    lora_config = LoraConfig(
        r=16, lora_alpha=32, lora_dropout=0.05,
        target_modules=["q_proj", "k_proj", "v_proj", "o_proj"],
        task_type=TaskType.CAUSAL_LM, bias="none",
    )
    peft_model = get_peft_model(model, lora_config)
    trainable = sum(p.numel() for p in peft_model.parameters() if p.requires_grad)
    total_params = sum(p.numel() for p in peft_model.parameters())
    print(f"  LoRA: {trainable:,}/{total_params:,} trainable")

    output_dir = f"/tmp/adapters/{adapter_name}"

    # Configure trainer
    if USE_NEW_TRL:
        training_args = SFTConfig(
            output_dir=output_dir,
            num_train_epochs=epochs,
            per_device_train_batch_size=2,
            gradient_accumulation_steps=4,
            learning_rate=2e-4,
            warmup_ratio=0.03,
            logging_steps=10,
            save_steps=500,
            bf16=True,
            report_to="none",
            dataset_text_field="text",
            max_length=2048,
        )
        trainer = SFTTrainer(
            model=peft_model,
            args=training_args,
            train_dataset=dataset,
            processing_class=tokenizer,
        )
    else:
        training_args = TrainingArguments(
            output_dir=output_dir,
            num_train_epochs=epochs,
            per_device_train_batch_size=2,
            gradient_accumulation_steps=4,
            learning_rate=2e-4,
            warmup_ratio=0.03,
            logging_steps=10,
            save_steps=500,
            bf16=True,
            report_to="none",
        )
        trainer = SFTTrainer(
            model=peft_model,
            args=training_args,
            train_dataset=dataset,
            tokenizer=tokenizer,
            dataset_text_field="text",
            max_seq_length=2048,
        )

    # Train
    print(f"  Training...")
    result = trainer.train()
    elapsed = time.time() - start
    print(f"  DONE! Loss: {result.training_loss:.4f}, Steps: {result.global_step}, Time: {elapsed:.0f}s")

    # Save locally
    peft_model.save_pretrained(output_dir)
    tokenizer.save_pretrained(output_dir)
    print(f"  Saved locally to {output_dir}")

    # Upload (with error handling - don't crash the job!)
    try:
        api.upload_folder(
            folder_path=output_dir,
            path_in_repo=adapter_name,
            repo_id=OUTPUT_REPO,
            token=HF_TOKEN,
        )
        print(f"  Uploaded to {OUTPUT_REPO}/{adapter_name}")
    except Exception as e:
        print(f"  WARNING: Upload failed for {adapter_name}: {e}")
        failed_uploads.append(adapter_name)

    results[adapter_name] = {
        "loss": result.training_loss,
        "steps": result.global_step,
        "time_seconds": elapsed,
    }

    # Cleanup for next adapter
    try:
        model = peft_model.unload()
    except:
        model = peft_model.base_model.model
    del peft_model, trainer, dataset
    gc.collect()
    torch.cuda.empty_cache()

# --- Summary ---
total_elapsed = time.time() - total_start
print(f"\n{'=' * 60}")
print("ALL 8 ADAPTERS TRAINED!")
print(f"Total time: {total_elapsed/60:.1f} minutes")
print(f"{'=' * 60}")
for name, r in results.items():
    print(f"  {name}: loss={r['loss']:.4f}, steps={r['steps']}, time={r['time_seconds']:.0f}s")

# --- Retry failed uploads ---
if failed_uploads:
    print(f"\nRetrying {len(failed_uploads)} failed uploads...")
    for adapter_name in failed_uploads:
        output_dir = f"/tmp/adapters/{adapter_name}"
        try:
            api.upload_folder(
                folder_path=output_dir,
                path_in_repo=adapter_name,
                repo_id=OUTPUT_REPO,
                token=HF_TOKEN,
            )
            print(f"  Retry SUCCESS: {adapter_name}")
        except Exception as e:
            print(f"  Retry FAILED: {adapter_name}: {e}")

# --- Upload results summary ---
try:
    with open("/tmp/training_results.json", "w") as f:
        json.dump(results, f, indent=2)
    api.upload_file(
        path_or_fileobj="/tmp/training_results.json",
        path_in_repo="training_results.json",
        repo_id=OUTPUT_REPO,
        token=HF_TOKEN,
    )
    print("Results uploaded.")
except Exception as e:
    print(f"Results upload failed: {e}")
    print("Results JSON:")
    print(json.dumps(results, indent=2))

print(f"\nAdapters: https://huggingface.co/{OUTPUT_REPO}")