# EBFT Structured Mode: Qwen3.5-4B (hybrid linear attention)
#
# Qwen3.5 uses hybrid attention: linear attention (conv1d) on 3/4 of layers,
# full attention every 4th layer. This tests EBFT compatibility.
#
# Prerequisites:
#   1. Start vLLM on GPU 0:
#      CUDA_VISIBLE_DEVICES=0 trl vllm-serve --model Qwen/Qwen3.5-4B \
#          --gpu-memory-utilization 0.5 --max-model-len 2048 --enforce-eager
#
#   2. Run training on GPU 1:
#      CUDA_VISIBLE_DEVICES=1 PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True \
#          axolotl train examples/ebft/qwen35-4b-ebft-structured.yaml

base_model: Qwen/Qwen3.5-4B

rl: ebft

ebft:
  feature_layers: [0.25, 0.5, 0.75]
  embed_method: last_token
  use_whitening: false
  alignment_coef: 1.0
  diversity_coef: 1.0
  ce_coef: 0.0

trl:
  num_generations: 4
  max_completion_length: 256
  temperature: 0.7
  use_vllm: true
  vllm_server_host: 0.0.0.0
  vllm_server_port: 8000
  scale_rewards: true
  loss_type: grpo
  epsilon: 0.2
  generation_kwargs:
    stop_token_ids: [248044, 248046]  # <|endoftext|>, <|im_end|>
  chat_template_kwargs:
    enable_thinking: false  # disable Qwen3.5 thinking mode for shorter completions

datasets:
  - path: nvidia/OpenCodeInstruct
    type: ebft_opencode.transform
    split: train[:500]

sequence_len: 1024
micro_batch_size: 1
gradient_accumulation_steps: 4
num_epochs: 1
max_steps: 10

learning_rate: 5.0e-6
optimizer: adamw_torch_fused
lr_scheduler: cosine
warmup_steps: 3
weight_decay: 0.01

adapter: lora
lora_r: 16
lora_alpha: 32
lora_dropout: 0.0
lora_target_modules: ".*\\.layers\\.(3|7|11|15|19|23|27|31)\\.self_attn\\.(q|k|v|o)_proj|.*\\.mlp\\.(gate|up|down)_proj"

bf16: auto
flash_attention: true
gradient_checkpointing: true

special_tokens:
  pad_token: "<|endoftext|>"

val_set_size: 0.0
output_dir: ./outputs/ebft-qwen35-4b-structured

wandb_project: ebft
logging_steps: 1
save_steps: 50