# Colab-Optimized Training Configuration for Stack 2.9
# Target: Google Colab free tier (T4 GPU, 15GB VRAM)
# Model: Qwen/Qwen2.5-Coder-7B (4-bit quantized fits in ~4.5GB)
# Expected runtime: 3-5 hours

model:
  name: "Qwen/Qwen2.5-Coder-7B"  # 7B instead of 32B for Colab
  trust_remote_code: true
  use_flash_attention: false  # T4 doesn't support flash attention well

tokenizer:
  model_max_length: 8192  # Reduced from 131072 for memory
  padding_side: "right"
  truncation_side: "right"

peft:
  peft_type: "LORA"
  task_type: "CAUSAL_LM"
  r: 16  # LoRA rank (lower = faster, good enough for 7B)
  lora_alpha: 32
  lora_dropout: 0.05
  target_modules:
    - "q_proj"
    - "k_proj"
    - "v_proj"
    - "o_proj"
    - "gate_proj"
    - "up_proj"
    - "down_proj"
    # Optional: add "embed_tokens", "lm_head" for full coverage (increases memory)

quantization:
  load_in_4bit: true
  bnb_4bit_compute_dtype: "bfloat16"
  bnb_4bit_quant_type: "nf4"
  bnb_4bit_use_double_quant: true

training:
  output_dir: "./adapters_colab"
  num_train_epochs: 2  # Sufficient for 7B with decent dataset
  per_device_train_batch_size: 1  # Tiny batch for 15GB VRAM
  gradient_accumulation_steps: 16  # Effective batch size = 16
  optim: "paged_adamw_8bit"  # 8-bit optimizer for memory
  learning_rate: 1.0e-4
  weight_decay: 0.01
  warmup_steps: 100
  lr_scheduler_type: "cosine"
  save_steps: 500
  save_total_limit: 2
  logging_steps: 10
  report_to: "none"  # Disable wandb for Colab

  # Memory optimizations
  gradient_checkpointing: true
  fp16: false  # Use bf16 instead if available
  bf16: true  # T4 supports bf16
  max_grad_norm: 1.0
  dataloader_num_workers: 2
  remove_unused_columns: false

data:
  train_file: "./training-data/train.jsonl"
  validation_file: "./training-data/eval.jsonl"
  dataset_format: "chat"  # or "prompt_response"
  max_seq_length: 8192  # Critical for T4 memory
  prompt_template: "chatml"  # Qwen's default template

# Hardware
ddp: false  # Single GPU for Colab

# Misc
seed: 42
push_to_hub: false  # Set to true and add HF token to push during training
hub_model_id: null  # "your-org/stack-2.9-7b-lora"

# Notes:
# - 4-bit quantization + batch size 1 + gradient checkpointing = fits in 15GB
# - If OOM: reduce max_seq_length to 4096 or increase gradient_accumulation_steps
# - If training is slow: increase per_device_train_batch_size to 2 (if memory allows)
# - After training, merge adapter with base model using merge_adapter.py