base_model: axolotl-quants/Llama-4-Scout-17B-16E-Linearized-bnb-nf4-bf16
model_type: Llama4ForConditionalGeneration
# Automatically upload checkpoint and final model to HF
# hub_model_id: username/custom_model_name

plugins:
  - axolotl.integrations.liger.LigerPlugin
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin

liger_glu_activation: true
liger_rms_norm: true
liger_layer_norm: true

llama4_linearized_experts: true  # needed with custom linearized experts model
load_in_4bit: true
adapter: qlora
lora_r: 32
lora_alpha: 64
lora_target_modules:
  - self_attn.q_proj
  - self_attn.k_proj
  - self_attn.v_proj
  - self_attn.o_proj
  - shared_expert.gate_proj
  - shared_expert.up_proj
  - shared_expert.down_proj
  # - experts.gate_projs.[0-9]+$  # optionally train the moe experts
  # - experts.up_projs.[0-9]+$
  # - experts.down_projs.[0-9]+$
lora_modules_to_save:
  # - lm_head  # needed if modifying vocabulary
  # - embed_tokens

lora_mlp_kernel: true
lora_qkv_kernel: true
lora_o_kernel: true

chat_template: llama4
datasets:
  - path: mlabonne/FineTome-100k
    type: chat_template
    split: train[:20%]
    field_messages: conversations
    message_property_mappings:
      role: from
      content: value

dataset_prepared_path: last_run_prepared
val_set_size: 0.0
output_dir: ./outputs/out

sequence_len: 4096  # up to 8k will work on a single H100
sample_packing: true
pad_to_sequence_len: true

gradient_accumulation_steps: 1
micro_batch_size: 1
num_epochs: 1
optimizer: adamw_torch_4bit
lr_scheduler: cosine
learning_rate: 1e-4

bf16: true
tf32: true

torch_compile: true
flex_attention: true
flex_attn_compile_kwargs:
  dynamic: false
  mode: max-autotune-no-cudagraphs

gradient_checkpointing: offload
gradient_checkpointing_kwargs:
  use_reentrant: false

logging_steps: 1
warmup_steps: 20
evals_per_epoch: 1
saves_per_epoch: 1

weight_decay: 0.0
special_tokens:
  pad_token: <|finetune_right_pad_id|>
  eos_token: <|eot|>