Qwen2.5-Coder 7B/14B/32B: Repo-Level Context (16K-128K) + FIM Native FT

# === Qwen2.5-Coder 7B FT (RTX 4090) ===
from unsloth import FastLanguageModel
from trl import SFTTrainer, SFTConfig
from datasets import load_dataset
 
model, tok = FastLanguageModel.from_pretrained(
    "unsloth/Qwen2.5-Coder-7B-Instruct-bnb-4bit",
    max_seq_length=16384,                       # repo-level için uzun context
    dtype="bfloat16", load_in_4bit=True,
)
model = FastLanguageModel.get_peft_model(
    model, r=64, lora_alpha=128, lora_dropout=0.05,
    target_modules=["q_proj","k_proj","v_proj","o_proj",
                    "gate_proj","up_proj","down_proj"],
    use_gradient_checkpointing="unsloth",
)
 
# Custom repo dataset
# Format: full file content concat (multiple files per sample) +/- FIM
dataset = load_dataset("user/my-python-repo-fim", split="train")
 
cfg = SFTConfig(
    output_dir="qwen-coder-7b-myrepo",
    num_train_epochs=2,
    per_device_train_batch_size=1,             # 16K seq → batch=1
    gradient_accumulation_steps=8,
    learning_rate=1e-4,
    bf16=True, optim="paged_adamw_8bit",
    max_seq_length=16384, packing=True,
    dataset_text_field="text",
    logging_steps=5, report_to="wandb",
)
SFTTrainer(model=model, tokenizer=tok, train_dataset=dataset, args=cfg).train()