train.py

import os
import random
import numpy as np
import torch
from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForCausalLM, EarlyStoppingCallback
from trl import SFTTrainer, SFTConfig
from peft import LoraConfig
from transformers import BitsAndBytesConfig

# Config from env vars
BASE_MODEL = os.environ.get("BASE_MODEL", "DeepSeek-Coder-V2-Lite-Instruct")
OUTPUT_DIR = os.environ.get("OUTPUT_DIR", "outputs/zenith-lora")
DATA_PATH = os.environ.get("DATA_PATH", "data/zenith.jsonl")
VAL_PATH = os.environ.get("VAL_PATH")
MAX_STEPS = int(os.environ.get("STEPS", 200))
USE_4BIT = os.environ.get("USE_4BIT", "1") == "1"
SEED = int(os.environ.get("SEED", 42))

os.makedirs(OUTPUT_DIR, exist_ok=True)

# Set seeds for reproducibility
random.seed(SEED)
np.random.seed(SEED)
torch.manual_seed(SEED)
if torch.cuda.is_available():
    torch.cuda.manual_seed_all(SEED)

print(f"Loading tokenizer and model from: {BASE_MODEL}")
tokenizer = AutoTokenizer.from_pretrained(BASE_MODEL)
if tokenizer.pad_token is None:
    tokenizer.pad_token = tokenizer.eos_token

# Set compute dtype based on GPU capability
compute_dtype = torch.float16
if torch.cuda.is_available():
    device_cap = torch.cuda.get_device_capability(0)
    if device_cap[0] >= 8:  # Ampere or higher
        print("Using bfloat16 for Ampere GPU")
        compute_dtype = torch.bfloat16

# 4-bit quantization config
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=compute_dtype,
    bnb_4bit_use_double_quant=True,
    llm_int8_enable_fp32_cpu_offload=True,
)

print("Loading model with 4-bit quantization...")
model = AutoModelForCausalLM.from_pretrained(
    BASE_MODEL,
    quantization_config=bnb_config,
    device_map="auto",
    trust_remote_code=True,
)

# Memory-saving configurations
model.config.use_cache = False

data_files = [DATA_PATH, "data/training_data_v2.jsonl"]
print(f"Loading datasets: {data_files}")
raw_train = load_dataset("json", data_files=data_files, split="train")

# Optional external validation file
if VAL_PATH:
    print(f"Loading validation dataset: {VAL_PATH}")
    raw_val = load_dataset("json", data_files=VAL_PATH, split="train")
else:
    split = raw_train.train_test_split(test_size=0.05, seed=SEED)
    raw_train, raw_val = split["train"], split["test"]

# Validate and format examples safely
MAX_SEQ_LEN = int(os.environ.get("MAX_SEQ_LEN", 2048))

def _valid(example):
    msgs = example.get("messages")
    if not isinstance(msgs, list) or not msgs:
        return False
    for m in msgs:
        if not isinstance(m, dict) or "role" not in m or "content" not in m:
            return False
    return True

def _to_text(example):
    try:
        text = tokenizer.apply_chat_template(
            example["messages"], tokenize=False, add_generation_prompt=False
        )
        return {"text": text}
    except Exception:
        return {"text": ""}

train_ds = raw_train.filter(_valid)
val_ds = raw_val.filter(_valid)

train_ds = train_ds.map(_to_text, remove_columns=train_ds.column_names)
val_ds = val_ds.map(_to_text, remove_columns=val_ds.column_names)

# Drop empty or pathological items
train_ds = train_ds.filter(lambda x: isinstance(x.get("text"), str) and len(x["text"]) > 0)
val_ds = val_ds.filter(lambda x: isinstance(x.get("text"), str) and len(x["text"]) > 0)

# LoRA config
peft_config = LoraConfig(
    r=int(os.environ.get("LORA_R", 16)),
    lora_alpha=int(os.environ.get("LORA_ALPHA", 32)),
    lora_dropout=float(os.environ.get("LORA_DROPOUT", 0.05)),
    bias="none",
    task_type="CAUSAL_LM",
)

# Training config - step-based for quick runs with stability
training_args = SFTConfig(
    output_dir=OUTPUT_DIR,
    max_steps=MAX_STEPS,  # Use steps instead of epochs for precise timing
    per_device_train_batch_size=int(os.environ.get("BATCH", 2)),
    gradient_accumulation_steps=int(os.environ.get("GRAD_ACC", 2)),
    learning_rate=float(os.environ.get("LR", 1e-4)),
    lr_scheduler_type=os.environ.get("LR_SCHED", "cosine"),
    warmup_ratio=float(os.environ.get("WARMUP_RATIO", 0.05)),
    weight_decay=float(os.environ.get("WEIGHT_DECAY", 0.01)),
    max_grad_norm=float(os.environ.get("MAX_GRAD_NORM", 1.0)),
    logging_steps=int(os.environ.get("LOG_STEPS", 10)),
    save_steps=int(os.environ.get("SAVE_STEPS", 50)),
    save_total_limit=int(os.environ.get("SAVE_LIMIT", 3)),
    evaluation_strategy="steps",
    eval_steps=int(os.environ.get("EVAL_STEPS", 50)),
    load_best_model_at_end=True,
    metric_for_best_model="eval_loss",
    greater_is_better=False,
    fp16=torch.cuda.is_available(),
    bf16=torch.cuda.is_available() and torch.cuda.get_device_capability(0)[0] >= 8,
    packing=False,
    max_seq_length=MAX_SEQ_LEN,
    dataloader_drop_last=True,
    gradient_checkpointing=True,
    gradient_checkpointing_kwargs={"use_reentrant": False},
    report_to=os.environ.get("REPORT_TO", "none"),
    seed=SEED,
)

print(f"Starting SFT training for {MAX_STEPS} steps...")
trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=train_ds,
    eval_dataset=val_ds,
    peft_config=peft_config,
    args=training_args,
    dataset_text_field="text",
    callbacks=[EarlyStoppingCallback(early_stopping_patience=int(os.environ.get("EARLY_STOP_PATIENCE", 3)))]
)

trainer.train()

print("Saving LoRA adapter...")
trainer.model.save_pretrained(OUTPUT_DIR)
tokenizer.save_pretrained(OUTPUT_DIR)

print(f"✅ ZENITH LoRA adapter saved to: {OUTPUT_DIR}")
print("🎯 World's most advanced autonomous AI development partner ready!")
print("🚀 Ready for Aspetos platform integration!")