# Gemma 4 26B-A4B MoE QLoRA with ScatterMoE kernels
#
# Validated: 50 steps on FineTome-100k, loss 7.4 -> 2.4, single RTX 5090 (32GB)
#
# Key notes:
# - Flash Attention 2 is NOT supported (global_head_dim=512 > FA2 max of 256).
#   Use sdp_attention instead.
# - Gemma 4 is multimodal (text+vision+audio). For text-only SFT, restrict
#   LoRA to the text backbone via lora_target_linear_modules regex.
# - MoE experts use `experts_implementation: scattermoe` — Gemma 4 embeds MoE
#   directly in the decoder layer (no SparseMoeBlock), so we register ScatterMoE
#   via the transformers ExpertsInterface.
# - Expert LoRA targets are `experts.gate_up_proj` / `experts.down_proj`
#   (no `mlp.` prefix, unlike Qwen/Mixtral).
# - micro_batch_size: 1 fits 2048 seq_len on 32GB GPU with SDP attention.
#   Use micro_batch_size: 4 with 1024 seq_len, or on 48GB+ GPUs.

base_model: google/gemma-4-26B-A4B

plugins:
  - axolotl.integrations.cut_cross_entropy.CutCrossEntropyPlugin
  - axolotl.integrations.kernels.KernelsPlugin
  - axolotl.integrations.liger.LigerPlugin
use_kernels: true
use_scattermoe: true
experts_implementation: scattermoe
torch_compile: false
liger_layer_norm: true
liger_rope: true
liger_rms_norm: true
liger_glu_activation: true
liger_rms_norm_gated: true
strict: false

chat_template: gemma4
datasets:
  - path: mlabonne/FineTome-100k
    type: chat_template
    split: train[:10%]
    field_messages: conversations
    message_property_mappings:
      role: from
      content: value
val_set_size: 0.05
output_dir: ./outputs/gemma4-26b-a4b-qlora

sequence_len: 2048
sample_packing: true

load_in_4bit: true
quantize_moe_experts: true
adapter: qlora
lora_r: 16
lora_alpha: 32
lora_dropout: 0

# Restrict LoRA to text backbone only (skip vision/audio encoders).
# lora_target_modules is intentionally empty — all module targeting is done
# via regex in lora_target_linear_modules below.
lora_target_modules: []
lora_target_linear_modules:
  - language_model\.model\.layers\.\d+\.self_attn\.(q|k|v|o)_proj

# MoE expert LoRA (3D Parameter tensors, not nn.Linear)
lora_target_parameters:
  - experts.gate_up_proj
  - experts.down_proj

lora_mlp_kernel: false
lora_qkv_kernel: false
lora_o_kernel: false

bnb_config_kwargs:
  bnb_4bit_use_double_quant: true

wandb_project: gemma4-qlora
wandb_entity:
wandb_watch:
wandb_name:
wandb_log_model:

gradient_accumulation_steps: 4
micro_batch_size: 1
num_epochs: 1
optimizer: adamw_torch_8bit
lr_scheduler: cosine
learning_rate: 0.0002

bf16: auto
tf32: true

gradient_checkpointing: true
activation_offloading: true
logging_steps: 1

# FA2 not supported — Gemma4 global_head_dim=512 exceeds FA2 max of 256
flash_attention: false
sdp_attention: true

warmup_ratio: 0.1
evals_per_epoch: 4
saves_per_epoch: 1
weight_decay: 0.0
special_tokens: