ROMA / yamls /train.yaml
Houssem0's picture
ROMA + GH200 reproducible Docker layer
e5c09aa verified
Raw
History Blame Contribute Delete
1.55 kB
### model
model_name_or_path: models/Qwen2.5-Omni-7B
image_max_pixels: 262144 #res = 512
video_max_pixels: 65536
trust_remote_code: true
### method
do_train: true
stage: sft
# additional_target: gate_head,gate_mixer
finetuning_type: full
# finetuning_type: freeze
# freeze_trainable_layers: 0
# freeze_trainable_modules: all
# freeze_extra_modules: gate_head_pro_fc1,gate_head_pro_fc2,gate_mixer
# finetuning_type: lora
# lora_rank: 64
# lora_target: q_proj,k_proj,v_proj,o_proj,up_proj,down_proj,gate_proj
# additional_target: gate_head_pro_fc1,gate_head_pro_fc2,gate_mixer
dataset: abl_all_1,abl_all_2
buffer_size: 16
preprocessing_batch_size: 16
streaming: true
accelerator_config:
dispatch_batches: false
mix_strategy: interleave
max_steps: 6000
template: streaming_mix
cutoff_len: 32000
overwrite_cache: true
preprocessing_num_workers: 64
dataloader_num_workers: 64
### output
output_dir: streaming_results/abl_all
logging_steps: 5
save_steps: 1500
plot_loss: true
overwrite_output_dir: true
save_only_model: false
### train
use_audio_in_video: true
per_device_train_batch_size: 2
gradient_accumulation_steps: 8
gradient_checkpointing: True
freeze_vision_tower: true
learning_rate: 1.0e-5
num_train_epochs: 1
lr_scheduler_type: cosine
warmup_ratio: 0.1
bf16: true
#fp16: true
ddp_timeout: 180000000
deepspeed: examples/deepspeed/ds_z3_config.json
resume_from_checkpoint: null
#save_strategy: "epoch"
flash_attn: fa2
enable_liger_kernel: true
use_unsloth_gc : true