train.py

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

# ====== CONFIG ======
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_combined.jsonl")
VAL_PATH = os.environ.get("VAL_PATH")
MAX_STEPS = int(os.environ.get("STEPS", 300))
SEED = int(os.environ.get("SEED", 42))

os.makedirs(OUTPUT_DIR, exist_ok=True)

# ====== SEED CONTROL ======
random.seed(SEED)
np.random.seed(SEED)
torch.manual_seed(SEED)
if torch.cuda.is_available():
    torch.cuda.manual_seed_all(SEED)
torch.backends.cuda.matmul.allow_tf32 = True
torch.backends.cudnn.allow_tf32 = True

# ====== TOKENIZER & MODEL ======
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

compute_dtype = torch.float16
if torch.cuda.is_available() and torch.cuda.get_device_capability(0)[0] >= 8:
    compute_dtype = torch.bfloat16
    print("✅ Ampere+ GPU detected — will prefer bf16 where supported.")

bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=compute_dtype,
    bnb_4bit_use_double_quant=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,
)
model.config.use_cache = False

# ====== DATASET ======
data_files = [DATA_PATH]
raw_train = load_dataset("json", data_files=data_files, split="train")

if VAL_PATH and os.path.exists(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"]

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).map(_to_text, remove_columns=raw_train.column_names)
val_ds = raw_val.filter(_valid).map(_to_text, remove_columns=raw_val.column_names)

train_ds = train_ds.filter(lambda x: len(x.get("text", "")) > 0)
val_ds = val_ds.filter(lambda x: len(x.get("text", "")) > 0)

print(f"✅ Training samples: {len(train_ds)}, Validation: {len(val_ds)}")

# ====== LORA CONFIG ======
peft_config = LoraConfig(
    r=int(os.environ.get("LORA_R", 8)),
    lora_alpha=int(os.environ.get("LORA_ALPHA", 16)),
    lora_dropout=float(os.environ.get("LORA_DROPOUT", 0.1)),
    bias="none",
    task_type="CAUSAL_LM",
    target_modules=["q_proj", "v_proj"],  # Required for LoRA injection
)

# ====== EVAL CALLBACK ======
class EvalEveryCallback(TrainerCallback):
    def __init__(self, eval_steps=100):
        self.eval_steps = eval_steps
    def on_step_end(self, args, state, control, **kwargs):
        if state.global_step % self.eval_steps == 0 and state.global_step > 0:
            control.should_evaluate = True
        return control

# ====== TRAINING CONFIG ======
training_args = SFTConfig(
    output_dir=OUTPUT_DIR,
    max_steps=MAX_STEPS,
    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", 5e-5)),
    lr_scheduler_type=os.environ.get("LR_SCHED", "cosine"),
    warmup_ratio=float(os.environ.get("WARMUP_RATIO", 0.1)),
    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", 2)),
    fp16=torch.cuda.is_available() and compute_dtype==torch.float16,
    bf16=torch.cuda.is_available() and compute_dtype==torch.bfloat16,
    gradient_checkpointing=True,
    gradient_checkpointing_kwargs={"use_reentrant": False},
    dataloader_drop_last=True,
    report_to="none",
    seed=SEED,
)

# ====== TRAINER ======
print(f"🏁 Starting Zenith fine-tuning for {MAX_STEPS} steps (~2h config)...")
trainer = SFTTrainer(
    model=model,
    train_dataset=train_ds,
    eval_dataset=val_ds,
    peft_config=peft_config,
    args=training_args,
)

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("🎯 Training complete under ~2 hours.")