darwinkernelpanic/luau-codellama-7b-reasoning

base_model: codellama/CodeLlama-7b-hf
model_type: LlamaForCausalLM
tokenizer_type: LlamaTokenizer

# Keep full precision weights (fast on Hopper)
load_in_8bit: false
load_in_4bit: false
strict: false

chat_template: llama3

datasets:
  - path: darwinkernelpanic/luau-reasoning-normalized
    type: chat_template
    conversation: llama3
    field_messages: messages
    add_generation_prompt: true

# Preprocessing workers (CPU). Fine as-is.
num_proc: 16

output_dir: ./outputs/luau-codellama-h200-fast

# ===== LoRA =====
adapter: lora
lora_r: 16
lora_alpha: 32
lora_dropout: 0.05
lora_target_modules:
  - q_proj
  - k_proj
  - v_proj
  - o_proj

# ===== Precision =====
bf16: true
fp16: false
tf32: true

# ===== Sequence / batching =====
sequence_len: 4096
# Keep packing for throughput, but enable length grouping to cut padding
sample_packing: true
group_by_length: true

# Lower micro-batch a bit to kill peak VRAM while staying fast
micro_batch_size: 5
gradient_accumulation_steps: 1

# ===== Training =====
num_epochs: 3
optimizer: adamw_torch
learning_rate: 2e-4
lr_scheduler_type: cosine
warmup_steps: 100

train_on_inputs: false

# Turn on checkpointing — tiny speed hit, big memory win
gradient_checkpointing: true
gradient_clipping: 1.0

# ===== Dataloader =====
# Keep pin_memory, but avoid too many loader workers in Accelerate
dataloader_num_workers: 2
dataloader_pin_memory: true
# Optional: avoid insanely large host->device prefetch
# dataloader_prefetch_factor: 2

# ===== Logging / eval =====
logging_steps: 25
val_set_size: 0.05
# Reduce eval/save frequency to avoid spikes
eval_steps: 1000
save_strategy: steps
save_steps: 1000
save_total_limit: 3

seed: 42

# ===== DeepSpeed =====
# Off for single H200 — overhead not worth it for 7B