sahil-theloops/Qwen3-30B-A3B-ckpt-45

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "Qwen/Qwen3-30B-A3B",
    max_seq_length = 1024,   # Context length - can be longer, but uses more memory
    load_in_4bit = True,     # 4bit uses much less memory
    load_in_8bit = False,    # A bit more accurate, uses 2x memory
    full_finetuning = False, # We have full finetuning now!
    # token = "hf_...",      # use one if using gated models
)

model = FastLanguageModel.get_peft_model(
    model,
    r = 8,           # Choose any number > 0! Suggested 8, 16, 32, 64, 128
    target_modules = ["q_proj", "k_proj", "v_proj", "o_proj",
                      "gate_proj", "up_proj", "down_proj",],
    lora_alpha = 8,  # Best to choose alpha = rank or rank*2
    lora_dropout = 0, # Supports any, but = 0 is optimized
    bias = "none",    # Supports any, but = "none" is optimized
    # [NEW] "unsloth" uses 30% less VRAM, fits 2x larger batch sizes!
    use_gradient_checkpointing = False #"unsloth", # True or "unsloth" for very long context
    random_state = 3407,
    use_rslora = False,   # We support rank stabilized LoRA
    loftq_config = None,  # And LoftQ
)

from trl import SFTTrainer, SFTConfig
trainer = SFTTrainer(
    model = model,
    tokenizer = tokenizer,
    train_dataset = combined_dataset,
    dataset_num_proc=4,
    eval_dataset = None, # Can set up evaluation!
    args = SFTConfig(
        dataset_text_field = "text",
        per_device_train_batch_size = 32,
        gradient_accumulation_steps = 4, # Use GA to mimic batch size!
        warmup_steps = 5,
        num_train_epochs = 1, # Set this for 1 full training run.
        # max_steps = 400,
        learning_rate = 8e-5, # Reduce to 2e-5 for long training runs
        logging_steps = 1,
        optim = "adamw_8bit",
        weight_decay = 0.01,
        lr_scheduler_type = "linear",
        seed = 3407,
        max_grad_norm=0.5,
        save_steps=15,
        # report_to="tensorboard"
    ),
)