SmolLM3 1.7B: Tiny Tier — Production Model Running on 8GB RAM Devices

# === SmolLM3 1.7B Full FT (RTX 4090) ===
# Full FT mümkün çünkü 1.7B × 2 = 3.4 GB W + grad
from transformers import AutoModelForCausalLM, AutoTokenizer, TrainingArguments
from trl import SFTTrainer, SFTConfig
from datasets import load_dataset
 
model = AutoModelForCausalLM.from_pretrained(
    "HuggingFaceTB/SmolLM3-1.7B-Instruct",
    torch_dtype="bfloat16",
    attn_implementation="flash_attention_2",
    device_map="cuda",
)
tok = AutoTokenizer.from_pretrained("HuggingFaceTB/SmolLM3-1.7B-Instruct")
 
dataset = load_dataset("malhajar/alpaca-gpt4-tr", split="train").map(...)
 
cfg = SFTConfig(
    output_dir="smol-1.7b-tr-fullft",
    num_train_epochs=2,
    per_device_train_batch_size=4,
    gradient_accumulation_steps=4,
    learning_rate=5e-5,                       # Full FT için düşük lr
    warmup_ratio=0.05, lr_scheduler_type="cosine",
    weight_decay=0.01,
    bf16=True, optim="paged_adamw_8bit",
    gradient_checkpointing=True,
    max_seq_length=4096, packing=True,
    dataset_text_field="text",
    logging_steps=5, save_steps=200, report_to="wandb",
)
 
SFTTrainer(model=model, tokenizer=tok, train_dataset=dataset, args=cfg).train()
# 25 dakikada 1 epoch — 4090'ın tadını çıkarıyoruz