Upload folder using huggingface_hub

.gitattributes

@@ -8,8 +8,6 @@
 *.h5 filter=lfs diff=lfs merge=lfs -text
 *.joblib filter=lfs diff=lfs merge=lfs -text
 *.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.lz4 filter=lfs diff=lfs merge=lfs -text
-*.mds filter=lfs diff=lfs merge=lfs -text
 *.mlmodel filter=lfs diff=lfs merge=lfs -text
 *.model filter=lfs diff=lfs merge=lfs -text
 *.msgpack filter=lfs diff=lfs merge=lfs -text
@@ -35,25 +33,3 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
-# Audio files - uncompressed
-*.pcm filter=lfs diff=lfs merge=lfs -text
-*.sam filter=lfs diff=lfs merge=lfs -text
-*.raw filter=lfs diff=lfs merge=lfs -text
-# Audio files - compressed
-*.aac filter=lfs diff=lfs merge=lfs -text
-*.flac filter=lfs diff=lfs merge=lfs -text
-*.mp3 filter=lfs diff=lfs merge=lfs -text
-*.ogg filter=lfs diff=lfs merge=lfs -text
-*.wav filter=lfs diff=lfs merge=lfs -text
-# Image files - uncompressed
-*.bmp filter=lfs diff=lfs merge=lfs -text
-*.gif filter=lfs diff=lfs merge=lfs -text
-*.png filter=lfs diff=lfs merge=lfs -text
-*.tiff filter=lfs diff=lfs merge=lfs -text
-# Image files - compressed
-*.jpg filter=lfs diff=lfs merge=lfs -text
-*.jpeg filter=lfs diff=lfs merge=lfs -text
-*.webp filter=lfs diff=lfs merge=lfs -text
-# Video files - compressed
-*.mp4 filter=lfs diff=lfs merge=lfs -text
-*.webm filter=lfs diff=lfs merge=lfs -text

README.md

@@ -0,0 +1,12 @@
+---
+title: Mft Log
+emoji: 📚
+colorFrom: pink
+colorTo: pink
+sdk: gradio
+sdk_version: 5.49.1
+app_file: app.py
+pinned: false
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

logs_oct10/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_0.3_2e-1_connector-1.0_0.3_2e-1_ablation_20251010_055757.log

@@ -0,0 +1,236 @@
+==== STARTING EXPERIMENT: qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_0.3_2e-1_connector-1.0_0.3_2e-1_ablation ====
+Log File: qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_0.3_2e-1_connector-1.0_0.3_2e-1_ablation_20251010_055757.log
+Timestamp: 2025-10-10 05:57:57
+=====================================
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+The cache for model files in Transformers v4.22.0 has been updated. Migrating your old cache. This is a one-time only operation. You can interrupt this and resume the migration later on by calling `transformers.utils.move_cache()`.
+
+[2025-10-10 05:58:00,052] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:58:03,277] [WARNING] [runner.py:202:fetch_hostfile] Unable to find hostfile, will proceed with training with local resources only.
+[2025-10-10 05:58:03,278] [INFO] [runner.py:568:main] cmd = /opt/conda/envs/tinyllava/bin/python3.10 -u -m deepspeed.launcher.launch --world_info=eyJsb2NhbGhvc3QiOiBbMCwgMSwgMiwgMywgNCwgNSwgNiwgN119 --master_addr=127.0.0.1 --master_port=29501 --enable_each_rank_log=None tinyllava/train/train.py --deepspeed ./scripts/zero3.json --data_path /nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json --image_folder /nfs/ywang29/tinyLLaVA/dataset --is_multimodal True --conv_version qwen2_base --model_name_or_path Qwen/Qwen2.5-0.5B --vision_tower google/siglip-so400m-patch14-384 --vision_tower2  --connector_type mlp2x_gelu --mm_vision_select_layer -2 --image_aspect_ratio square --attn_implementation flash_attention_2 --bf16 True --training_recipe common --tune_type_llm full --tune_type_vision_tower frozen --tune_vision_tower_from_layer 0 --tune_type_connector full --group_by_modality_length True --pretrained_model_path /nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain --output_dir /nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_0.3_2e-1_connector-1.0_0.3_2e-1_ablation --num_train_epochs 1 --per_device_train_batch_size 4 --per_device_eval_batch_size 4 --gradient_accumulation_steps 4 --evaluation_strategy no --learning_rate 2e-1 --weight_decay 0. --warmup_ratio 0.03 --lr_scheduler_type cosine --logging_steps 1 --tf32 False --model_max_length 2048 --gradient_checkpointing True --dataloader_num_workers 8 --lazy_preprocess True --report_to tensorboard --tokenizer_use_fast False --run_name tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune --subnet_mode_text both --subnet_type_text None --mask_type_text soft --init_mean_text 1.0 --temperature_attn_text 0.3 --temperature_mlp_text 0.3 --backward_type_text normal --masked_layers_text all --subnet_mode_vision both --subnet_type_vision None --mask_type_vision soft --init_mean_vision 1.0 --temperature_attn_vision 0.3 --temperature_mlp_vision 0.3 --backward_type_vision normal --masked_layers_vision all --subnet_type_connector global --mask_type_connector soft --init_mean_connector 1.0 --temperature_connector 0.3 --backward_type_connector normal --mm_projector_lr 2e-1 --seed 42 --mask_model llm-connector --save_strategy steps --save_steps 50000 --save_total_limit 1 --train_data_ratio 0.1
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+[2025-10-10 05:58:05,879] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:58:06,963] [INFO] [launch.py:138:main] 0 NCCL_VERSION=2.21.5
+[2025-10-10 05:58:06,963] [INFO] [launch.py:138:main] 0 NCCL_SOCKET_IFNAME=eth
+[2025-10-10 05:58:06,963] [INFO] [launch.py:145:main] WORLD INFO DICT: {'localhost': [0, 1, 2, 3, 4, 5, 6, 7]}
+[2025-10-10 05:58:06,963] [INFO] [launch.py:151:main] nnodes=1, num_local_procs=8, node_rank=0
+[2025-10-10 05:58:06,963] [INFO] [launch.py:162:main] global_rank_mapping=defaultdict(<class 'list'>, {'localhost': [0, 1, 2, 3, 4, 5, 6, 7]})
+[2025-10-10 05:58:06,963] [INFO] [launch.py:163:main] dist_world_size=8
+[2025-10-10 05:58:06,963] [INFO] [launch.py:165:main] Setting CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+[2025-10-10 05:58:06,966] [INFO] [launch.py:253:main] process 548567 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=0', '--deepspeed', './scripts/zero3.json', '--data_path', '/nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json', '--image_folder', '/nfs/ywang29/tinyLLaVA/dataset', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_0.3_2e-1_connector-1.0_0.3_2e-1_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '2e-1', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '1.0', '--temperature_attn_text', '0.3', '--temperature_mlp_text', '0.3', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '1.0', '--temperature_attn_vision', '0.3', '--temperature_mlp_vision', '0.3', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '1.0', '--temperature_connector', '0.3', '--backward_type_connector', 'normal', '--mm_projector_lr', '2e-1', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 05:58:06,968] [INFO] [launch.py:253:main] process 548568 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=1', '--deepspeed', './scripts/zero3.json', '--data_path', '/nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json', '--image_folder', '/nfs/ywang29/tinyLLaVA/dataset', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_0.3_2e-1_connector-1.0_0.3_2e-1_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '2e-1', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '1.0', '--temperature_attn_text', '0.3', '--temperature_mlp_text', '0.3', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '1.0', '--temperature_attn_vision', '0.3', '--temperature_mlp_vision', '0.3', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '1.0', '--temperature_connector', '0.3', '--backward_type_connector', 'normal', '--mm_projector_lr', '2e-1', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 05:58:06,970] [INFO] [launch.py:253:main] process 548569 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=2', '--deepspeed', './scripts/zero3.json', '--data_path', '/nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json', '--image_folder', '/nfs/ywang29/tinyLLaVA/dataset', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_0.3_2e-1_connector-1.0_0.3_2e-1_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '2e-1', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '1.0', '--temperature_attn_text', '0.3', '--temperature_mlp_text', '0.3', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '1.0', '--temperature_attn_vision', '0.3', '--temperature_mlp_vision', '0.3', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '1.0', '--temperature_connector', '0.3', '--backward_type_connector', 'normal', '--mm_projector_lr', '2e-1', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 05:58:06,972] [INFO] [launch.py:253:main] process 548570 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=3', '--deepspeed', './scripts/zero3.json', '--data_path', '/nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json', '--image_folder', '/nfs/ywang29/tinyLLaVA/dataset', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_0.3_2e-1_connector-1.0_0.3_2e-1_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '2e-1', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '1.0', '--temperature_attn_text', '0.3', '--temperature_mlp_text', '0.3', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '1.0', '--temperature_attn_vision', '0.3', '--temperature_mlp_vision', '0.3', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '1.0', '--temperature_connector', '0.3', '--backward_type_connector', 'normal', '--mm_projector_lr', '2e-1', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 05:58:06,973] [INFO] [launch.py:253:main] process 548571 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=4', '--deepspeed', './scripts/zero3.json', '--data_path', '/nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json', '--image_folder', '/nfs/ywang29/tinyLLaVA/dataset', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_0.3_2e-1_connector-1.0_0.3_2e-1_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '2e-1', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '1.0', '--temperature_attn_text', '0.3', '--temperature_mlp_text', '0.3', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '1.0', '--temperature_attn_vision', '0.3', '--temperature_mlp_vision', '0.3', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '1.0', '--temperature_connector', '0.3', '--backward_type_connector', 'normal', '--mm_projector_lr', '2e-1', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 05:58:06,975] [INFO] [launch.py:253:main] process 548572 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=5', '--deepspeed', './scripts/zero3.json', '--data_path', '/nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json', '--image_folder', '/nfs/ywang29/tinyLLaVA/dataset', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_0.3_2e-1_connector-1.0_0.3_2e-1_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '2e-1', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '1.0', '--temperature_attn_text', '0.3', '--temperature_mlp_text', '0.3', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '1.0', '--temperature_attn_vision', '0.3', '--temperature_mlp_vision', '0.3', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '1.0', '--temperature_connector', '0.3', '--backward_type_connector', 'normal', '--mm_projector_lr', '2e-1', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 05:58:06,977] [INFO] [launch.py:253:main] process 548573 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=6', '--deepspeed', './scripts/zero3.json', '--data_path', '/nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json', '--image_folder', '/nfs/ywang29/tinyLLaVA/dataset', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_0.3_2e-1_connector-1.0_0.3_2e-1_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '2e-1', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '1.0', '--temperature_attn_text', '0.3', '--temperature_mlp_text', '0.3', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '1.0', '--temperature_attn_vision', '0.3', '--temperature_mlp_vision', '0.3', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '1.0', '--temperature_connector', '0.3', '--backward_type_connector', 'normal', '--mm_projector_lr', '2e-1', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 05:58:06,979] [INFO] [launch.py:253:main] process 548574 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=7', '--deepspeed', './scripts/zero3.json', '--data_path', '/nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json', '--image_folder', '/nfs/ywang29/tinyLLaVA/dataset', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_0.3_2e-1_connector-1.0_0.3_2e-1_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '2e-1', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '1.0', '--temperature_attn_text', '0.3', '--temperature_mlp_text', '0.3', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '1.0', '--temperature_attn_vision', '0.3', '--temperature_mlp_vision', '0.3', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '1.0', '--temperature_connector', '0.3', '--backward_type_connector', 'normal', '--mm_projector_lr', '2e-1', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+[2025-10-10 05:58:13,755] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:58:13,756] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:58:13,756] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:58:13,763] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:58:13,775] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:58:13,775] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:58:13,804] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:58:13,838] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:58:14,271] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 05:58:14,271] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 05:58:14,271] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 05:58:14,271] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 05:58:14,271] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 05:58:14,271] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 05:58:14,271] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 05:58:14,271] [INFO] [comm.py:668:init_distributed] Initializing TorchBackend in DeepSpeed with backend nccl
+[2025-10-10 05:58:14,271] [INFO] [comm.py:637:init_distributed] cdb=None
+Apply masks for the following modules:  ['llm', 'connector']
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+Apply masks for the following modules:  ['llm', 'connector']
+{'llm': {'model_name_or_path': 'Qwen/Qwen2.5-0.5B', 'cache_dir': None, 'attn_implementation': 'flash_attention_2', 'subnet_mode': 'both', 'subnet_type': 'None', 'sparsity_attn': None, 'sparsity_mlp': None, 'threshold_attn': None, 'threshold_mlp': None, 'temperature_attn': 0.3, 'temperature_mlp': 0.3, 'masked_layers': 'all', 'mask_type': 'soft', 'backward_type': 'normal'}, 'vision_tower': {'model_name_or_path': 'google/siglip-so400m-patch14-384'}, 'connector': {'connector_type': 'mlp2x_gelu', 'subnet_type': 'global', 'threshold': None, 'sparsity': None, 'temperature': 0.3, 'mask_type': 'soft', 'backward_type': 'normal'}}
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+Apply masks for the following modules:  ['llm', 'connector']
+Apply masks for the following modules:  ['llm', 'connector']
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+Apply masks for the following modules:  ['llm', 'connector']
+Apply masks for the following modules:  ['llm', 'connector']
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+Apply masks for the following modules:  ['llm', 'connector']
+Apply masks for the following modules:  ['llm', 'connector']
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+TinyLlavaConfig {
+  "backward_type_connector": "normal",
+  "cache_dir": null,
+  "connector_type": "mlp2x_gelu",
+  "hidden_size": 896,
+  "ignore_index": -100,
+  "image_aspect_ratio": "square",
+  "image_token_index": -200,
+  "llm_model_name_or_path": "Qwen/Qwen2.5-0.5B",
+  "mask_model": [
+    "llm",
+    "connector"
+  ],
+  "mask_type_connector": "soft",
+  "model_type": "tinyllava",
+  "num_queries": 128,
+  "num_resampler_layers": 3,
+  "pad_token": null,
+  "resampler_hidden_size": 768,
+  "sparsity_connector": null,
+  "subnet_type_connector": "global",
+  "temperature_connector": 0.3,
+  "text_config": {
+    "_name_or_path": "Qwen/Qwen2.5-0.5B",
+    "architectures": [
+      "Qwen2ForCausalLM"
+    ],
+    "backward_type": "normal",
+    "bos_token_id": 151643,
+    "eos_token_id": 151643,
+    "hidden_size": 896,
+    "intermediate_size": 4864,
+    "mask_type": "soft",
+    "masked_layers": "all",
+    "max_position_embeddings": 32768,
+    "max_window_layers": 24,
+    "model_type": "qwen2",
+    "num_attention_heads": 14,
+    "num_hidden_layers": 24,
+    "num_key_value_heads": 2,
+    "rope_theta": 1000000.0,
+    "sliding_window": 32768,
+    "subnet_mode": "both",
+    "subnet_type": "None",
+    "temperature_attn": 0.3,
+    "temperature_mlp": 0.3,
+    "tie_word_embeddings": true,
+    "torch_dtype": "bfloat16",
+    "use_mrope": false,
+    "use_sliding_window": false,
+    "vocab_size": 151936
+  },
+  "threshold_connector": null,
+  "tokenizer_model_max_length": 2048,
+  "tokenizer_name_or_path": "Qwen/Qwen2.5-0.5B",
+  "tokenizer_padding_side": "right",
+  "tokenizer_use_fast": false,
+  "transformers_version": "4.40.1",
+  "tune_type_connector": "frozen",
+  "tune_type_llm": "frozen",
+  "tune_type_vision_tower": "frozen",
+  "tune_vision_tower_from_layer": -1,
+  "use_cache": false,
+  "vision_config": {
+    "hidden_act": "gelu_pytorch_tanh",
+    "hidden_size": 1152,
+    "image_size": 384,
+    "intermediate_size": 4304,
+    "layer_norm_eps": 1e-06,
+    "model_name_or_path": "google/siglip-so400m-patch14-384",
+    "model_name_or_path2": "",
+    "model_type": "siglip_vision_model",
+    "num_attention_heads": 16,
+    "num_hidden_layers": 27,
+    "patch_size": 14
+  },
+  "vision_feature_layer": -2,
+  "vision_feature_select_strategy": "patch",
+  "vision_hidden_size": 1152,
+  "vision_model_name_or_path": "google/siglip-so400m-patch14-384",
+  "vision_model_name_or_path2": "",
+  "vocab_size": 151936
+}
+
+Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
+Traceback (most recent call last):
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/utils/hub.py", line 398, in cached_file
+    resolved_file = hf_hub_download(
+  File "/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py", line 106, in _inner_fn
+    validate_repo_id(arg_value)
+  File "/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py", line 154, in validate_repo_id
+    raise HFValidationError(
+huggingface_hub.errors.HFValidationError: Repo id must be in the form 'repo_name' or 'namespace/repo_name': '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain/language_model'. Use `repo_type` argument if needed.
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/train/train.py", line 193, in <module>
+    train()
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/train/train.py", line 149, in train
+    model = training_recipe.load(model, model_args)
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/training_recipe/base.py", line 144, in load
+    model.load_llm(**model_args['llm'])
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/model/modeling_tinyllava.py", line 360, in load_llm
+    self.language_model = self.language_model.from_pretrained(
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/modeling_utils.py", line 3015, in from_pretrained
+    resolved_config_file = cached_file(
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/utils/hub.py", line 462, in cached_file
+    raise EnvironmentError(
+OSError: Incorrect path_or_model_id: '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain/language_model'. Please provide either the path to a local folder or the repo_id of a model on the Hub.
+[2025-10-10 05:58:46,032] [INFO] [launch.py:316:sigkill_handler] Killing subprocess 548567
+[2025-10-10 05:58:46,418] [INFO] [launch.py:316:sigkill_handler] Killing subprocess 548568
+[2025-10-10 05:58:46,881] [INFO] [launch.py:316:sigkill_handler] Killing subprocess 548569
+[2025-10-10 05:58:47,265] [INFO] [launch.py:316:sigkill_handler] Killing subprocess 548570
+[2025-10-10 05:58:47,683] [INFO] [launch.py:316:sigkill_handler] Killing subprocess 548571
+[2025-10-10 05:58:48,100] [INFO] [launch.py:316:sigkill_handler] Killing subprocess 548572
+Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
+Traceback (most recent call last):
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/utils/hub.py", line 398, in cached_file
+    resolved_file = hf_hub_download(
+  File "/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py", line 106, in _inner_fn
+    validate_repo_id(arg_value)
+  File "/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py", line 154, in validate_repo_id
+    raise HFValidationError(
+huggingface_hub.errors.HFValidationError: Repo id must be in the form 'repo_name' or 'namespace/repo_name': '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain/language_model'. Use `repo_type` argument if needed.
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/train/train.py", line 193, in <module>
+    train()
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/train/train.py", line 149, in train
+    model = training_recipe.load(model, model_args)
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/training_recipe/base.py", line 144, in load
+    model.load_llm(**model_args['llm'])
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/model/modeling_tinyllava.py", line 360, in load_llm
+    self.language_model = self.language_model.from_pretrained(
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/modeling_utils.py", line 3015, in from_pretrained
+[2025-10-10 05:58:48,477] [INFO] [launch.py:316:sigkill_handler] Killing subprocess 548573
+    resolved_config_file = cached_file(
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/utils/hub.py", line 462, in cached_file
+    raise EnvironmentError(
+OSError: Incorrect path_or_model_id: '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain/language_model'. Please provide either the path to a local folder or the repo_id of a model on the Hub.
+[2025-10-10 05:58:48,893] [INFO] [launch.py:316:sigkill_handler] Killing subprocess 548574
+[2025-10-10 05:58:48,894] [ERROR] [launch.py:322:sigkill_handler] ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=7', '--deepspeed', './scripts/zero3.json', '--data_path', '/nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json', '--image_folder', '/nfs/ywang29/tinyLLaVA/dataset', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_0.3_2e-1_connector-1.0_0.3_2e-1_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '2e-1', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '1.0', '--temperature_attn_text', '0.3', '--temperature_mlp_text', '0.3', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '1.0', '--temperature_attn_vision', '0.3', '--temperature_mlp_vision', '0.3', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '1.0', '--temperature_connector', '0.3', '--backward_type_connector', 'normal', '--mm_projector_lr', '2e-1', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1'] exits with return code = 1
+==== EXPERIMENT COMPLETED: qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_0.3_2e-1_connector-1.0_0.3_2e-1_ablation ====
+Log File: qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_0.3_2e-1_connector-1.0_0.3_2e-1_ablation_20251010_055757.log
+Timestamp: 2025-10-10 05:58:50
+=====================================

logs_oct10/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_0.3_2e-1_connector-1.0_0.3_2e-1_ablation_20251010_164158.log

@@ -0,0 +1,7 @@
+==== STARTING EXPERIMENT: qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_0.3_2e-1_connector-1.0_0.3_2e-1_ablation ====
+Log File: qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_0.3_2e-1_connector-1.0_0.3_2e-1_ablation_20251010_164158.log
+Timestamp: 2025-10-10 16:41:58
+=====================================
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+[2025-10-10 16:42:00,884] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)

logs_oct10/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_0.5_2e-1_connector-1.0_0.5_2e-1_ablation_20251010_055850.log

@@ -0,0 +1,9 @@
+==== STARTING EXPERIMENT: qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_0.5_2e-1_connector-1.0_0.5_2e-1_ablation ====
+Log File: qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_0.5_2e-1_connector-1.0_0.5_2e-1_ablation_20251010_055850.log
+Timestamp: 2025-10-10 05:58:50
+=====================================
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+[2025-10-10 05:58:53,031] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:58:55,712] [WARNING] [runner.py:202:fetch_hostfile] Unable to find hostfile, will proceed with training with local resources only.
+[2025-10-10 05:58:55,713] [INFO] [runner.py:568:main] cmd = /opt/conda/envs/tinyllava/bin/python3.10 -u -m deepspeed.launcher.launch --world_info=eyJsb2NhbGhvc3QiOiBbMCwgMSwgMiwgMywgNCwgNSwgNiwgN119 --master_addr=127.0.0.1 --master_port=29501 --enable_each_rank_log=None tinyllava/train/train.py --deepspeed ./scripts/zero3.json --data_path /nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json --image_folder /nfs/ywang29/tinyLLaVA/dataset --is_multimodal True --conv_version qwen2_base --model_name_or_path Qwen/Qwen2.5-0.5B --vision_tower google/siglip-so400m-patch14-384 --vision_tower2  --connector_type mlp2x_gelu --mm_vision_select_layer -2 --image_aspect_ratio square --attn_implementation flash_attention_2 --bf16 True --training_recipe common --tune_type_llm full --tune_type_vision_tower frozen --tune_vision_tower_from_layer 0 --tune_type_connector full --group_by_modality_length True --pretrained_model_path /nfs/ywang29/TinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain --output_dir /nfs/ywang29/TinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_0.5_2e-1_connector-1.0_0.5_2e-1_ablation --num_train_epochs 1 --per_device_train_batch_size 4 --per_device_eval_batch_size 4 --gradient_accumulation_steps 4 --evaluation_strategy no --learning_rate 2e-1 --weight_decay 0. --warmup_ratio 0.03 --lr_scheduler_type cosine --logging_steps 1 --tf32 False --model_max_length 2048 --gradient_checkpointing True --dataloader_num_workers 8 --lazy_preprocess True --report_to tensorboard --tokenizer_use_fast False --run_name tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune --subnet_mode_text both --subnet_type_text None --mask_type_text soft --init_mean_text 1.0 --temperature_attn_text 0.5 --temperature_mlp_text 0.5 --backward_type_text normal --masked_layers_text all --subnet_mode_vision both --subnet_type_vision None --mask_type_vision soft --init_mean_vision 1.0 --temperature_attn_vision 0.5 --temperature_mlp_vision 0.5 --backward_type_vision normal --masked_layers_vision all --subnet_type_connector global --mask_type_connector soft --init_mean_connector 1.0 --temperature_connector 0.5 --backward_type_connector normal --mm_projector_lr 2e-1 --seed 42 --mask_model llm-connector --save_strategy steps --save_steps 50000 --save_total_limit 1 --train_data_ratio 0.1

logs_oct10/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_0.7_2e-1_connector-1.0_0.7_2e-1_ablation_20251010_060116.log

@@ -0,0 +1,11 @@
+==== STARTING EXPERIMENT: qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_0.7_2e-1_connector-1.0_0.7_2e-1_ablation ====
+Log File: qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_0.7_2e-1_connector-1.0_0.7_2e-1_ablation_20251010_060116.log
+Timestamp: 2025-10-10 06:01:16
+=====================================
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+[2025-10-10 06:01:19,451] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 06:01:22,190] [WARNING] [runner.py:202:fetch_hostfile] Unable to find hostfile, will proceed with training with local resources only.
+[2025-10-10 06:01:22,192] [INFO] [runner.py:568:main] cmd = /opt/conda/envs/tinyllava/bin/python3.10 -u -m deepspeed.launcher.launch --world_info=eyJsb2NhbGhvc3QiOiBbMCwgMSwgMiwgMywgNCwgNSwgNiwgN119 --master_addr=127.0.0.1 --master_port=29501 --enable_each_rank_log=None tinyllava/train/train.py --deepspeed ./scripts/zero3.json --data_path /nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json --image_folder /nfs/ywang29/tinyLLaVA/dataset --is_multimodal True --conv_version qwen2_base --model_name_or_path Qwen/Qwen2.5-0.5B --vision_tower google/siglip-so400m-patch14-384 --vision_tower2  --connector_type mlp2x_gelu --mm_vision_select_layer -2 --image_aspect_ratio square --attn_implementation flash_attention_2 --bf16 True --training_recipe common --tune_type_llm full --tune_type_vision_tower frozen --tune_vision_tower_from_layer 0 --tune_type_connector full --group_by_modality_length True --pretrained_model_path /nfs/ywang29/TinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain --output_dir /nfs/ywang29/TinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_0.7_2e-1_connector-1.0_0.7_2e-1_ablation --num_train_epochs 1 --per_device_train_batch_size 4 --per_device_eval_batch_size 4 --gradient_accumulation_steps 4 --evaluation_strategy no --learning_rate 2e-1 --weight_decay 0. --warmup_ratio 0.03 --lr_scheduler_type cosine --logging_steps 1 --tf32 False --model_max_length 2048 --gradient_checkpointing True --dataloader_num_workers 8 --lazy_preprocess True --report_to tensorboard --tokenizer_use_fast False --run_name tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune --subnet_mode_text both --subnet_type_text None --mask_type_text soft --init_mean_text 1.0 --temperature_attn_text 0.7 --temperature_mlp_text 0.7 --backward_type_text normal --masked_layers_text all --subnet_mode_vision both --subnet_type_vision None --mask_type_vision soft --init_mean_vision 1.0 --temperature_attn_vision 0.7 --temperature_mlp_vision 0.7 --backward_type_vision normal --masked_layers_vision all --subnet_type_connector global --mask_type_connector soft --init_mean_connector 1.0 --temperature_connector 0.7 --backward_type_connector normal --mm_projector_lr 2e-1 --seed 42 --mask_model llm-connector --save_strategy steps --save_steps 50000 --save_total_limit 1 --train_data_ratio 0.1
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]

logs_oct10/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_1.3_2e-1_connector-1.0_1.3_2e-1_ablation_20251010_060752.log

@@ -0,0 +1,4 @@
+==== STARTING EXPERIMENT: qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_1.3_2e-1_connector-1.0_1.3_2e-1_ablation ====
+Log File: qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_1.3_2e-1_connector-1.0_1.3_2e-1_ablation_20251010_060752.log
+Timestamp: 2025-10-10 06:07:52
+=====================================

logs_oct10/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_1.9_2e-1_connector-1.0_1.9_2e-1_ablation_20251010_162102.log

@@ -0,0 +1,169 @@
+==== STARTING EXPERIMENT: qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_1.9_2e-1_connector-1.0_1.9_2e-1_ablation ====
+Log File: qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_1.9_2e-1_connector-1.0_1.9_2e-1_ablation_20251010_162102.log
+Timestamp: 2025-10-10 16:21:02
+=====================================
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+[2025-10-10 16:21:04,987] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 16:21:08,361] [WARNING] [runner.py:202:fetch_hostfile] Unable to find hostfile, will proceed with training with local resources only.
+[2025-10-10 16:21:08,363] [INFO] [runner.py:568:main] cmd = /opt/conda/envs/tinyllava/bin/python3.10 -u -m deepspeed.launcher.launch --world_info=eyJsb2NhbGhvc3QiOiBbMCwgMSwgMiwgMywgNCwgNSwgNiwgN119 --master_addr=127.0.0.1 --master_port=29501 --enable_each_rank_log=None tinyllava/train/train.py --deepspeed ./scripts/zero3.json --data_path /s3-code/ywang29/datasets/tinyllava/text_files/llava_v1_5_mix665k.json --image_folder /s3-code/ywang29/datasets/tinyllava --is_multimodal True --conv_version qwen2_base --model_name_or_path Qwen/Qwen2.5-0.5B --vision_tower google/siglip-so400m-patch14-384 --vision_tower2  --connector_type mlp2x_gelu --mm_vision_select_layer -2 --image_aspect_ratio square --attn_implementation flash_attention_2 --bf16 True --training_recipe common --tune_type_llm full --tune_type_vision_tower frozen --tune_vision_tower_from_layer 0 --tune_type_connector full --group_by_modality_length True --pretrained_model_path /nfs/ywang29/TinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain --output_dir /nfs/ywang29/TinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_1.9_2e-1_connector-1.0_1.9_2e-1_ablation --num_train_epochs 1 --per_device_train_batch_size 4 --per_device_eval_batch_size 4 --gradient_accumulation_steps 4 --evaluation_strategy no --learning_rate 2e-1 --weight_decay 0. --warmup_ratio 0.03 --lr_scheduler_type cosine --logging_steps 1 --tf32 False --model_max_length 2048 --gradient_checkpointing True --dataloader_num_workers 8 --lazy_preprocess True --report_to tensorboard --tokenizer_use_fast False --run_name tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune --subnet_mode_text both --subnet_type_text None --mask_type_text soft --init_mean_text 1.0 --temperature_attn_text 1.9 --temperature_mlp_text 1.9 --backward_type_text normal --masked_layers_text all --subnet_mode_vision both --subnet_type_vision None --mask_type_vision soft --init_mean_vision 1.0 --temperature_attn_vision 1.9 --temperature_mlp_vision 1.9 --backward_type_vision normal --masked_layers_vision all --subnet_type_connector global --mask_type_connector soft --init_mean_connector 1.0 --temperature_connector 1.9 --backward_type_connector normal --mm_projector_lr 2e-1 --seed 42 --mask_model llm-connector --save_strategy steps --save_steps 50000 --save_total_limit 1 --train_data_ratio 0.1
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+[2025-10-10 16:21:10,949] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 16:21:12,019] [INFO] [launch.py:138:main] 0 NCCL_VERSION=2.21.5
+[2025-10-10 16:21:12,019] [INFO] [launch.py:138:main] 0 NCCL_SOCKET_IFNAME=eth
+[2025-10-10 16:21:12,020] [INFO] [launch.py:145:main] WORLD INFO DICT: {'localhost': [0, 1, 2, 3, 4, 5, 6, 7]}
+[2025-10-10 16:21:12,020] [INFO] [launch.py:151:main] nnodes=1, num_local_procs=8, node_rank=0
+[2025-10-10 16:21:12,020] [INFO] [launch.py:162:main] global_rank_mapping=defaultdict(<class 'list'>, {'localhost': [0, 1, 2, 3, 4, 5, 6, 7]})
+[2025-10-10 16:21:12,020] [INFO] [launch.py:163:main] dist_world_size=8
+[2025-10-10 16:21:12,020] [INFO] [launch.py:165:main] Setting CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+[2025-10-10 16:21:12,022] [INFO] [launch.py:253:main] process 780015 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=0', '--deepspeed', './scripts/zero3.json', '--data_path', '/s3-code/ywang29/datasets/tinyllava/text_files/llava_v1_5_mix665k.json', '--image_folder', '/s3-code/ywang29/datasets/tinyllava', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/TinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/TinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_1.9_2e-1_connector-1.0_1.9_2e-1_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '2e-1', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '1.0', '--temperature_attn_text', '1.9', '--temperature_mlp_text', '1.9', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '1.0', '--temperature_attn_vision', '1.9', '--temperature_mlp_vision', '1.9', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '1.0', '--temperature_connector', '1.9', '--backward_type_connector', 'normal', '--mm_projector_lr', '2e-1', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 16:21:12,024] [INFO] [launch.py:253:main] process 780016 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=1', '--deepspeed', './scripts/zero3.json', '--data_path', '/s3-code/ywang29/datasets/tinyllava/text_files/llava_v1_5_mix665k.json', '--image_folder', '/s3-code/ywang29/datasets/tinyllava', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/TinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/TinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_1.9_2e-1_connector-1.0_1.9_2e-1_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '2e-1', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '1.0', '--temperature_attn_text', '1.9', '--temperature_mlp_text', '1.9', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '1.0', '--temperature_attn_vision', '1.9', '--temperature_mlp_vision', '1.9', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '1.0', '--temperature_connector', '1.9', '--backward_type_connector', 'normal', '--mm_projector_lr', '2e-1', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 16:21:12,026] [INFO] [launch.py:253:main] process 780017 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=2', '--deepspeed', './scripts/zero3.json', '--data_path', '/s3-code/ywang29/datasets/tinyllava/text_files/llava_v1_5_mix665k.json', '--image_folder', '/s3-code/ywang29/datasets/tinyllava', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/TinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/TinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_1.9_2e-1_connector-1.0_1.9_2e-1_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '2e-1', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '1.0', '--temperature_attn_text', '1.9', '--temperature_mlp_text', '1.9', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '1.0', '--temperature_attn_vision', '1.9', '--temperature_mlp_vision', '1.9', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '1.0', '--temperature_connector', '1.9', '--backward_type_connector', 'normal', '--mm_projector_lr', '2e-1', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 16:21:12,028] [INFO] [launch.py:253:main] process 780018 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=3', '--deepspeed', './scripts/zero3.json', '--data_path', '/s3-code/ywang29/datasets/tinyllava/text_files/llava_v1_5_mix665k.json', '--image_folder', '/s3-code/ywang29/datasets/tinyllava', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/TinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/TinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_1.9_2e-1_connector-1.0_1.9_2e-1_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '2e-1', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '1.0', '--temperature_attn_text', '1.9', '--temperature_mlp_text', '1.9', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '1.0', '--temperature_attn_vision', '1.9', '--temperature_mlp_vision', '1.9', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '1.0', '--temperature_connector', '1.9', '--backward_type_connector', 'normal', '--mm_projector_lr', '2e-1', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 16:21:12,030] [INFO] [launch.py:253:main] process 780019 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=4', '--deepspeed', './scripts/zero3.json', '--data_path', '/s3-code/ywang29/datasets/tinyllava/text_files/llava_v1_5_mix665k.json', '--image_folder', '/s3-code/ywang29/datasets/tinyllava', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/TinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/TinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_1.9_2e-1_connector-1.0_1.9_2e-1_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '2e-1', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '1.0', '--temperature_attn_text', '1.9', '--temperature_mlp_text', '1.9', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '1.0', '--temperature_attn_vision', '1.9', '--temperature_mlp_vision', '1.9', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '1.0', '--temperature_connector', '1.9', '--backward_type_connector', 'normal', '--mm_projector_lr', '2e-1', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 16:21:12,032] [INFO] [launch.py:253:main] process 780020 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=5', '--deepspeed', './scripts/zero3.json', '--data_path', '/s3-code/ywang29/datasets/tinyllava/text_files/llava_v1_5_mix665k.json', '--image_folder', '/s3-code/ywang29/datasets/tinyllava', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/TinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/TinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_1.9_2e-1_connector-1.0_1.9_2e-1_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '2e-1', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '1.0', '--temperature_attn_text', '1.9', '--temperature_mlp_text', '1.9', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '1.0', '--temperature_attn_vision', '1.9', '--temperature_mlp_vision', '1.9', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '1.0', '--temperature_connector', '1.9', '--backward_type_connector', 'normal', '--mm_projector_lr', '2e-1', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 16:21:12,034] [INFO] [launch.py:253:main] process 780021 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=6', '--deepspeed', './scripts/zero3.json', '--data_path', '/s3-code/ywang29/datasets/tinyllava/text_files/llava_v1_5_mix665k.json', '--image_folder', '/s3-code/ywang29/datasets/tinyllava', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/TinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/TinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_1.9_2e-1_connector-1.0_1.9_2e-1_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '2e-1', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '1.0', '--temperature_attn_text', '1.9', '--temperature_mlp_text', '1.9', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '1.0', '--temperature_attn_vision', '1.9', '--temperature_mlp_vision', '1.9', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '1.0', '--temperature_connector', '1.9', '--backward_type_connector', 'normal', '--mm_projector_lr', '2e-1', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 16:21:12,036] [INFO] [launch.py:253:main] process 780022 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=7', '--deepspeed', './scripts/zero3.json', '--data_path', '/s3-code/ywang29/datasets/tinyllava/text_files/llava_v1_5_mix665k.json', '--image_folder', '/s3-code/ywang29/datasets/tinyllava', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/TinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/TinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-1.0_1.9_2e-1_connector-1.0_1.9_2e-1_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '2e-1', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '1.0', '--temperature_attn_text', '1.9', '--temperature_mlp_text', '1.9', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '1.0', '--temperature_attn_vision', '1.9', '--temperature_mlp_vision', '1.9', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '1.0', '--temperature_connector', '1.9', '--backward_type_connector', 'normal', '--mm_projector_lr', '2e-1', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+[2025-10-10 16:21:18,873] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 16:21:19,003] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 16:21:19,097] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 16:21:19,190] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 16:21:19,217] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 16:21:19,225] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 16:21:19,227] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 16:21:19,230] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 16:21:19,397] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 16:21:19,406] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 16:21:19,494] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 16:21:19,589] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 16:21:19,589] [INFO] [comm.py:668:init_distributed] Initializing TorchBackend in DeepSpeed with backend nccl
+[2025-10-10 16:21:19,619] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 16:21:19,626] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 16:21:19,632] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 16:21:19,633] [INFO] [comm.py:637:init_distributed] cdb=None
+Apply masks for the following modules:  ['llm', 'connector']
+Apply masks for the following modules:  ['llm', 'connector']
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+Apply masks for the following modules:  ['llm', 'connector']
+Apply masks for the following modules:  ['llm', 'connector']
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+Apply masks for the following modules:  ['llm', 'connector']
+{'llm': {'model_name_or_path': 'Qwen/Qwen2.5-0.5B', 'cache_dir': None, 'attn_implementation': 'flash_attention_2', 'subnet_mode': 'both', 'subnet_type': 'None', 'sparsity_attn': None, 'sparsity_mlp': None, 'threshold_attn': None, 'threshold_mlp': None, 'temperature_attn': 1.9, 'temperature_mlp': 1.9, 'masked_layers': 'all', 'mask_type': 'soft', 'backward_type': 'normal'}, 'vision_tower': {'model_name_or_path': 'google/siglip-so400m-patch14-384'}, 'connector': {'connector_type': 'mlp2x_gelu', 'subnet_type': 'global', 'threshold': None, 'sparsity': None, 'temperature': 1.9, 'mask_type': 'soft', 'backward_type': 'normal'}}
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+Apply masks for the following modules:  ['llm', 'connector']
+Apply masks for the following modules:  ['llm', 'connector']
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+Apply masks for the following modules:  ['llm', 'connector']
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+TinyLlavaConfig {
+  "backward_type_connector": "normal",
+  "cache_dir": null,
+  "connector_type": "mlp2x_gelu",
+  "hidden_size": 896,
+  "ignore_index": -100,
+  "image_aspect_ratio": "square",
+  "image_token_index": -200,
+  "llm_model_name_or_path": "Qwen/Qwen2.5-0.5B",
+  "mask_model": [
+    "llm",
+    "connector"
+  ],
+  "mask_type_connector": "soft",
+  "model_type": "tinyllava",
+  "num_queries": 128,
+  "num_resampler_layers": 3,
+  "pad_token": null,
+  "resampler_hidden_size": 768,
+  "sparsity_connector": null,
+  "subnet_type_connector": "global",
+  "temperature_connector": 1.9,
+  "text_config": {
+    "_name_or_path": "Qwen/Qwen2.5-0.5B",
+    "architectures": [
+      "Qwen2ForCausalLM"
+    ],
+    "backward_type": "normal",
+    "bos_token_id": 151643,
+    "eos_token_id": 151643,
+    "hidden_size": 896,
+    "intermediate_size": 4864,
+    "mask_type": "soft",
+    "masked_layers": "all",
+    "max_position_embeddings": 32768,
+    "max_window_layers": 24,
+    "model_type": "qwen2",
+    "num_attention_heads": 14,
+    "num_hidden_layers": 24,
+    "num_key_value_heads": 2,
+    "rope_theta": 1000000.0,
+    "sliding_window": 32768,
+    "subnet_mode": "both",
+    "subnet_type": "None",
+    "temperature_attn": 1.9,
+    "temperature_mlp": 1.9,
+    "tie_word_embeddings": true,
+    "torch_dtype": "bfloat16",
+    "use_mrope": false,
+    "use_sliding_window": false,
+    "vocab_size": 151936
+  },
+  "threshold_connector": null,
+  "tokenizer_model_max_length": 2048,
+  "tokenizer_name_or_path": "Qwen/Qwen2.5-0.5B",
+  "tokenizer_padding_side": "right",
+  "tokenizer_use_fast": false,
+  "transformers_version": "4.40.1",
+  "tune_type_connector": "frozen",
+  "tune_type_llm": "frozen",
+  "tune_type_vision_tower": "frozen",
+  "tune_vision_tower_from_layer": -1,
+  "use_cache": false,
+  "vision_config": {
+    "hidden_act": "gelu_pytorch_tanh",
+    "hidden_size": 1152,
+    "image_size": 384,
+    "intermediate_size": 4304,
+    "layer_norm_eps": 1e-06,
+    "model_name_or_path": "google/siglip-so400m-patch14-384",
+    "model_name_or_path2": "",
+    "model_type": "siglip_vision_model",
+    "num_attention_heads": 16,
+    "num_hidden_layers": 27,
+    "patch_size": 14
+  },
+  "vision_feature_layer": -2,
+  "vision_feature_select_strategy": "patch",
+  "vision_hidden_size": 1152,
+  "vision_model_name_or_path": "google/siglip-so400m-patch14-384",
+  "vision_model_name_or_path2": "",
+  "vocab_size": 151936
+}
+

logs_oct10/qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_1_connector-3.0_0.5_1_ablation_20251010_055702.log

@@ -0,0 +1,312 @@
+==== STARTING EXPERIMENT: qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_1_connector-3.0_0.5_1_ablation ====
+Log File: qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_1_connector-3.0_0.5_1_ablation_20251010_055702.log
+Timestamp: 2025-10-10 05:57:02
+=====================================
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+[2025-10-10 05:57:04,790] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:57:07,488] [WARNING] [runner.py:202:fetch_hostfile] Unable to find hostfile, will proceed with training with local resources only.
+[2025-10-10 05:57:07,490] [INFO] [runner.py:568:main] cmd = /opt/conda/envs/tinyllava/bin/python3.10 -u -m deepspeed.launcher.launch --world_info=eyJsb2NhbGhvc3QiOiBbMCwgMSwgMiwgMywgNCwgNSwgNiwgN119 --master_addr=127.0.0.1 --master_port=29501 --enable_each_rank_log=None tinyllava/train/train.py --deepspeed ./scripts/zero3.json --data_path /nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json --image_folder /nfs/ywang29/tinyLLaVA/dataset --is_multimodal True --conv_version qwen2_base --model_name_or_path Qwen/Qwen2.5-0.5B --vision_tower google/siglip-so400m-patch14-384 --vision_tower2  --connector_type mlp2x_gelu --mm_vision_select_layer -2 --image_aspect_ratio square --attn_implementation flash_attention_2 --bf16 True --training_recipe common --tune_type_llm full --tune_type_vision_tower frozen --tune_vision_tower_from_layer 0 --tune_type_connector full --group_by_modality_length True --pretrained_model_path /nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain --output_dir /nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_1_connector-3.0_0.5_1_ablation --num_train_epochs 1 --per_device_train_batch_size 4 --per_device_eval_batch_size 4 --gradient_accumulation_steps 4 --evaluation_strategy no --learning_rate 1 --weight_decay 0. --warmup_ratio 0.03 --lr_scheduler_type cosine --logging_steps 1 --tf32 False --model_max_length 2048 --gradient_checkpointing True --dataloader_num_workers 8 --lazy_preprocess True --report_to tensorboard --tokenizer_use_fast False --run_name tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune --subnet_mode_text both --subnet_type_text None --mask_type_text soft --init_mean_text 3.0 --temperature_attn_text 0.5 --temperature_mlp_text 0.5 --backward_type_text normal --masked_layers_text all --subnet_mode_vision both --subnet_type_vision None --mask_type_vision soft --init_mean_vision 3.0 --temperature_attn_vision 0.5 --temperature_mlp_vision 0.5 --backward_type_vision normal --masked_layers_vision all --subnet_type_connector global --mask_type_connector soft --init_mean_connector 3.0 --temperature_connector 0.5 --backward_type_connector normal --mm_projector_lr 1 --seed 42 --mask_model llm-connector --save_strategy steps --save_steps 50000 --save_total_limit 1 --train_data_ratio 0.1
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+[2025-10-10 05:57:10,112] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:57:11,166] [INFO] [launch.py:138:main] 0 NCCL_VERSION=2.21.5
+[2025-10-10 05:57:11,167] [INFO] [launch.py:138:main] 0 NCCL_SOCKET_IFNAME=eth
+[2025-10-10 05:57:11,167] [INFO] [launch.py:145:main] WORLD INFO DICT: {'localhost': [0, 1, 2, 3, 4, 5, 6, 7]}
+[2025-10-10 05:57:11,167] [INFO] [launch.py:151:main] nnodes=1, num_local_procs=8, node_rank=0
+[2025-10-10 05:57:11,167] [INFO] [launch.py:162:main] global_rank_mapping=defaultdict(<class 'list'>, {'localhost': [0, 1, 2, 3, 4, 5, 6, 7]})
+[2025-10-10 05:57:11,167] [INFO] [launch.py:163:main] dist_world_size=8
+[2025-10-10 05:57:11,167] [INFO] [launch.py:165:main] Setting CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+[2025-10-10 05:57:11,169] [INFO] [launch.py:253:main] process 1775550 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=0', '--deepspeed', './scripts/zero3.json', '--data_path', '/nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json', '--image_folder', '/nfs/ywang29/tinyLLaVA/dataset', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_1_connector-3.0_0.5_1_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '1', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '3.0', '--temperature_attn_text', '0.5', '--temperature_mlp_text', '0.5', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '3.0', '--temperature_attn_vision', '0.5', '--temperature_mlp_vision', '0.5', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '3.0', '--temperature_connector', '0.5', '--backward_type_connector', 'normal', '--mm_projector_lr', '1', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 05:57:11,171] [INFO] [launch.py:253:main] process 1775551 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=1', '--deepspeed', './scripts/zero3.json', '--data_path', '/nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json', '--image_folder', '/nfs/ywang29/tinyLLaVA/dataset', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_1_connector-3.0_0.5_1_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '1', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '3.0', '--temperature_attn_text', '0.5', '--temperature_mlp_text', '0.5', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '3.0', '--temperature_attn_vision', '0.5', '--temperature_mlp_vision', '0.5', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '3.0', '--temperature_connector', '0.5', '--backward_type_connector', 'normal', '--mm_projector_lr', '1', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 05:57:11,173] [INFO] [launch.py:253:main] process 1775552 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=2', '--deepspeed', './scripts/zero3.json', '--data_path', '/nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json', '--image_folder', '/nfs/ywang29/tinyLLaVA/dataset', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_1_connector-3.0_0.5_1_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '1', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '3.0', '--temperature_attn_text', '0.5', '--temperature_mlp_text', '0.5', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '3.0', '--temperature_attn_vision', '0.5', '--temperature_mlp_vision', '0.5', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '3.0', '--temperature_connector', '0.5', '--backward_type_connector', 'normal', '--mm_projector_lr', '1', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 05:57:11,175] [INFO] [launch.py:253:main] process 1775553 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=3', '--deepspeed', './scripts/zero3.json', '--data_path', '/nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json', '--image_folder', '/nfs/ywang29/tinyLLaVA/dataset', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_1_connector-3.0_0.5_1_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '1', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '3.0', '--temperature_attn_text', '0.5', '--temperature_mlp_text', '0.5', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '3.0', '--temperature_attn_vision', '0.5', '--temperature_mlp_vision', '0.5', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '3.0', '--temperature_connector', '0.5', '--backward_type_connector', 'normal', '--mm_projector_lr', '1', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 05:57:11,177] [INFO] [launch.py:253:main] process 1775554 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=4', '--deepspeed', './scripts/zero3.json', '--data_path', '/nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json', '--image_folder', '/nfs/ywang29/tinyLLaVA/dataset', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_1_connector-3.0_0.5_1_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '1', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '3.0', '--temperature_attn_text', '0.5', '--temperature_mlp_text', '0.5', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '3.0', '--temperature_attn_vision', '0.5', '--temperature_mlp_vision', '0.5', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '3.0', '--temperature_connector', '0.5', '--backward_type_connector', 'normal', '--mm_projector_lr', '1', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 05:57:11,179] [INFO] [launch.py:253:main] process 1775555 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=5', '--deepspeed', './scripts/zero3.json', '--data_path', '/nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json', '--image_folder', '/nfs/ywang29/tinyLLaVA/dataset', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_1_connector-3.0_0.5_1_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '1', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '3.0', '--temperature_attn_text', '0.5', '--temperature_mlp_text', '0.5', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '3.0', '--temperature_attn_vision', '0.5', '--temperature_mlp_vision', '0.5', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '3.0', '--temperature_connector', '0.5', '--backward_type_connector', 'normal', '--mm_projector_lr', '1', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 05:57:11,181] [INFO] [launch.py:253:main] process 1775556 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=6', '--deepspeed', './scripts/zero3.json', '--data_path', '/nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json', '--image_folder', '/nfs/ywang29/tinyLLaVA/dataset', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_1_connector-3.0_0.5_1_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '1', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '3.0', '--temperature_attn_text', '0.5', '--temperature_mlp_text', '0.5', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '3.0', '--temperature_attn_vision', '0.5', '--temperature_mlp_vision', '0.5', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '3.0', '--temperature_connector', '0.5', '--backward_type_connector', 'normal', '--mm_projector_lr', '1', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 05:57:11,183] [INFO] [launch.py:253:main] process 1775557 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=7', '--deepspeed', './scripts/zero3.json', '--data_path', '/nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json', '--image_folder', '/nfs/ywang29/tinyLLaVA/dataset', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_1_connector-3.0_0.5_1_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '1', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '3.0', '--temperature_attn_text', '0.5', '--temperature_mlp_text', '0.5', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '3.0', '--temperature_attn_vision', '0.5', '--temperature_mlp_vision', '0.5', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '3.0', '--temperature_connector', '0.5', '--backward_type_connector', 'normal', '--mm_projector_lr', '1', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+[2025-10-10 05:57:17,634] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:57:17,967] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:57:18,000] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:57:18,022] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:57:18,022] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:57:18,023] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:57:18,023] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:57:18,048] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:57:18,052] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 05:57:18,370] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 05:57:18,406] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 05:57:18,429] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 05:57:18,429] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 05:57:18,429] [INFO] [comm.py:668:init_distributed] Initializing TorchBackend in DeepSpeed with backend nccl
+[2025-10-10 05:57:18,430] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 05:57:18,431] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 05:57:18,453] [INFO] [comm.py:637:init_distributed] cdb=None
+Apply masks for the following modules:  ['llm', 'connector']
+Apply masks for the following modules:  ['llm', 'connector']
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+Apply masks for the following modules:  ['llm', 'connector']
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+Apply masks for the following modules:  ['llm', 'connector']
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+Apply masks for the following modules:  ['llm', 'connector']
+{'llm': {'model_name_or_path': 'Qwen/Qwen2.5-0.5B', 'cache_dir': None, 'attn_implementation': 'flash_attention_2', 'subnet_mode': 'both', 'subnet_type': 'None', 'sparsity_attn': None, 'sparsity_mlp': None, 'threshold_attn': None, 'threshold_mlp': None, 'temperature_attn': 0.5, 'temperature_mlp': 0.5, 'masked_layers': 'all', 'mask_type': 'soft', 'backward_type': 'normal'}, 'vision_tower': {'model_name_or_path': 'google/siglip-so400m-patch14-384'}, 'connector': {'connector_type': 'mlp2x_gelu', 'subnet_type': 'global', 'threshold': None, 'sparsity': None, 'temperature': 0.5, 'mask_type': 'soft', 'backward_type': 'normal'}}
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+Apply masks for the following modules:  ['llm', 'connector']
+Apply masks for the following modules:  ['llm', 'connector']
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+Apply masks for the following modules:  ['llm', 'connector']
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+TinyLlavaConfig {
+  "backward_type_connector": "normal",
+  "cache_dir": null,
+  "connector_type": "mlp2x_gelu",
+  "hidden_size": 896,
+  "ignore_index": -100,
+  "image_aspect_ratio": "square",
+  "image_token_index": -200,
+  "llm_model_name_or_path": "Qwen/Qwen2.5-0.5B",
+  "mask_model": [
+    "llm",
+    "connector"
+  ],
+  "mask_type_connector": "soft",
+  "model_type": "tinyllava",
+  "num_queries": 128,
+  "num_resampler_layers": 3,
+  "pad_token": null,
+  "resampler_hidden_size": 768,
+  "sparsity_connector": null,
+  "subnet_type_connector": "global",
+  "temperature_connector": 0.5,
+  "text_config": {
+    "_name_or_path": "Qwen/Qwen2.5-0.5B",
+    "architectures": [
+      "Qwen2ForCausalLM"
+    ],
+    "backward_type": "normal",
+    "bos_token_id": 151643,
+    "eos_token_id": 151643,
+    "hidden_size": 896,
+    "intermediate_size": 4864,
+    "mask_type": "soft",
+    "masked_layers": "all",
+    "max_position_embeddings": 32768,
+    "max_window_layers": 24,
+    "model_type": "qwen2",
+    "num_attention_heads": 14,
+    "num_hidden_layers": 24,
+    "num_key_value_heads": 2,
+    "rope_theta": 1000000.0,
+    "sliding_window": 32768,
+    "subnet_mode": "both",
+    "subnet_type": "None",
+    "temperature_attn": 0.5,
+    "temperature_mlp": 0.5,
+    "tie_word_embeddings": true,
+    "torch_dtype": "bfloat16",
+    "use_mrope": false,
+    "use_sliding_window": false,
+    "vocab_size": 151936
+  },
+  "threshold_connector": null,
+  "tokenizer_model_max_length": 2048,
+  "tokenizer_name_or_path": "Qwen/Qwen2.5-0.5B",
+  "tokenizer_padding_side": "right",
+  "tokenizer_use_fast": false,
+  "transformers_version": "4.40.1",
+  "tune_type_connector": "frozen",
+  "tune_type_llm": "frozen",
+  "tune_type_vision_tower": "frozen",
+  "tune_vision_tower_from_layer": -1,
+  "use_cache": false,
+  "vision_config": {
+    "hidden_act": "gelu_pytorch_tanh",
+    "hidden_size": 1152,
+    "image_size": 384,
+    "intermediate_size": 4304,
+    "layer_norm_eps": 1e-06,
+    "model_name_or_path": "google/siglip-so400m-patch14-384",
+    "model_name_or_path2": "",
+    "model_type": "siglip_vision_model",
+    "num_attention_heads": 16,
+    "num_hidden_layers": 27,
+    "patch_size": 14
+  },
+  "vision_feature_layer": -2,
+  "vision_feature_select_strategy": "patch",
+  "vision_hidden_size": 1152,
+  "vision_model_name_or_path": "google/siglip-so400m-patch14-384",
+  "vision_model_name_or_path2": "",
+  "vocab_size": 151936
+}
+
+Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
+Traceback (most recent call last):
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/utils/hub.py", line 398, in cached_file
+    resolved_file = hf_hub_download(
+  File "/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py", line 106, in _inner_fn
+    validate_repo_id(arg_value)
+  File "/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py", line 154, in validate_repo_id
+    raise HFValidationError(
+huggingface_hub.errors.HFValidationError: Repo id must be in the form 'repo_name' or 'namespace/repo_name': '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain/language_model'. Use `repo_type` argument if needed.
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/train/train.py", line 193, in <module>
+    train()
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/train/train.py", line 149, in train
+    model = training_recipe.load(model, model_args)
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/training_recipe/base.py", line 144, in load
+    model.load_llm(**model_args['llm'])
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/model/modeling_tinyllava.py", line 360, in load_llm
+    self.language_model = self.language_model.from_pretrained(
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/modeling_utils.py", line 3015, in from_pretrained
+    resolved_config_file = cached_file(
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/utils/hub.py", line 462, in cached_file
+    raise EnvironmentError(
+OSError: Incorrect path_or_model_id: '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain/language_model'. Please provide either the path to a local folder or the repo_id of a model on the Hub.
+Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
+Traceback (most recent call last):
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/utils/hub.py", line 398, in cached_file
+    resolved_file = hf_hub_download(
+  File "/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py", line 106, in _inner_fn
+    validate_repo_id(arg_value)
+  File "/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py", line 154, in validate_repo_id
+    raise HFValidationError(
+huggingface_hub.errors.HFValidationError: Repo id must be in the form 'repo_name' or 'namespace/repo_name': '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain/language_model'. Use `repo_type` argument if needed.
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/train/train.py", line 193, in <module>
+    train()
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/train/train.py", line 149, in train
+    model = training_recipe.load(model, model_args)
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/training_recipe/base.py", line 144, in load
+    model.load_llm(**model_args['llm'])
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/model/modeling_tinyllava.py", line 360, in load_llm
+    self.language_model = self.language_model.from_pretrained(
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/modeling_utils.py", line 3015, in from_pretrained
+    resolved_config_file = cached_file(
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/utils/hub.py", line 462, in cached_file
+    raise EnvironmentError(
+OSError: Incorrect path_or_model_id: '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain/language_model'. Please provide either the path to a local folder or the repo_id of a model on the Hub.
+Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
+Traceback (most recent call last):
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/utils/hub.py", line 398, in cached_file
+    resolved_file = hf_hub_download(
+  File "/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py", line 106, in _inner_fn
+    validate_repo_id(arg_value)
+  File "/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py", line 154, in validate_repo_id
+    raise HFValidationError(
+huggingface_hub.errors.HFValidationError: Repo id must be in the form 'repo_name' or 'namespace/repo_name': '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain/language_model'. Use `repo_type` argument if needed.
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/train/train.py", line 193, in <module>
+    train()
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/train/train.py", line 149, in train
+    model = training_recipe.load(model, model_args)
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/training_recipe/base.py", line 144, in load
+    model.load_llm(**model_args['llm'])
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/model/modeling_tinyllava.py", line 360, in load_llm
+    self.language_model = self.language_model.from_pretrained(
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/modeling_utils.py", line 3015, in from_pretrained
+    resolved_config_file = cached_file(
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/utils/hub.py", line 462, in cached_file
+    raise EnvironmentError(
+OSError: Incorrect path_or_model_id: '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain/language_model'. Please provide either the path to a local folder or the repo_id of a model on the Hub.
+[2025-10-10 05:57:56,238] [INFO] [launch.py:316:sigkill_handler] Killing subprocess 1775550
+Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
+Traceback (most recent call last):
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/utils/hub.py", line 398, in cached_file
+    resolved_file = hf_hub_download(
+  File "/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py", line 106, in _inner_fn
+    validate_repo_id(arg_value)
+  File "/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py", line 154, in validate_repo_id
+    raise HFValidationError(
+huggingface_hub.errors.HFValidationError: Repo id must be in the form 'repo_name' or 'namespace/repo_name': '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain/language_model'. Use `repo_type` argument if needed.
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/train/train.py", line 193, in <module>
+    train()
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/train/train.py", line 149, in train
+    model = training_recipe.load(model, model_args)
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/training_recipe/base.py", line 144, in load
+    model.load_llm(**model_args['llm'])
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/model/modeling_tinyllava.py", line 360, in load_llm
+    self.language_model = self.language_model.from_pretrained(
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/modeling_utils.py", line 3015, in from_pretrained
+    resolved_config_file = cached_file(
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/utils/hub.py", line 462, in cached_file
+    raise EnvironmentError(
+OSError: Incorrect path_or_model_id: '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain/language_model'. Please provide either the path to a local folder or the repo_id of a model on the Hub.
+[2025-10-10 05:57:56,576] [INFO] [launch.py:316:sigkill_handler] Killing subprocess 1775551
+[2025-10-10 05:57:56,994] [INFO] [launch.py:316:sigkill_handler] Killing subprocess 1775552
+[2025-10-10 05:57:56,996] [INFO] [launch.py:316:sigkill_handler] Killing subprocess 1775553
+Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
+Traceback (most recent call last):
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/utils/hub.py", line 398, in cached_file
+    resolved_file = hf_hub_download(
+  File "/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py", line 106, in _inner_fn
+    validate_repo_id(arg_value)
+  File "/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py", line 154, in validate_repo_id
+    raise HFValidationError(
+huggingface_hub.errors.HFValidationError: Repo id must be in the form 'repo_name' or 'namespace/repo_name': '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain/language_model'. Use `repo_type` argument if needed.
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/train/train.py", line 193, in <module>
+    train()
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/train/train.py", line 149, in train
+    model = training_recipe.load(model, model_args)
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/training_recipe/base.py", line 144, in load
+    model.load_llm(**model_args['llm'])
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/model/modeling_tinyllava.py", line 360, in load_llm
+    self.language_model = self.language_model.from_pretrained(
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/modeling_utils.py", line 3015, in from_pretrained
+    resolved_config_file = cached_file(
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/utils/hub.py", line 462, in cached_file
+    raise EnvironmentError(
+OSError: Incorrect path_or_model_id: '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain/language_model'. Please provide either the path to a local folder or the repo_id of a model on the Hub.
+[2025-10-10 05:57:57,453] [INFO] [launch.py:316:sigkill_handler] Killing subprocess 1775554
+[2025-10-10 05:57:57,455] [INFO] [launch.py:316:sigkill_handler] Killing subprocess 1775555
+[2025-10-10 05:57:57,455] [INFO] [launch.py:316:sigkill_handler] Killing subprocess 1775556
+[2025-10-10 05:57:57,792] [INFO] [launch.py:316:sigkill_handler] Killing subprocess 1775557
+[2025-10-10 05:57:57,794] [ERROR] [launch.py:322:sigkill_handler] ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=7', '--deepspeed', './scripts/zero3.json', '--data_path', '/nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json', '--image_folder', '/nfs/ywang29/tinyLLaVA/dataset', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_1_connector-3.0_0.5_1_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '1', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '3.0', '--temperature_attn_text', '0.5', '--temperature_mlp_text', '0.5', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '3.0', '--temperature_attn_vision', '0.5', '--temperature_mlp_vision', '0.5', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '3.0', '--temperature_connector', '0.5', '--backward_type_connector', 'normal', '--mm_projector_lr', '1', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1'] exits with return code = 1
+==== EXPERIMENT COMPLETED: qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_1_connector-3.0_0.5_1_ablation ====
+Log File: qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_1_connector-3.0_0.5_1_ablation_20251010_055702.log
+Timestamp: 2025-10-10 05:57:59
+=====================================

logs_oct10/qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_3_connector-3.0_0.5_3_ablation_20251010_055759.log

@@ -0,0 +1,225 @@
+==== STARTING EXPERIMENT: qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_3_connector-3.0_0.5_3_ablation ====
+Log File: qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_3_connector-3.0_0.5_3_ablation_20251010_055759.log
+Timestamp: 2025-10-10 05:57:59
+=====================================
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+[2025-10-10 05:58:01,905] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:58:04,550] [WARNING] [runner.py:202:fetch_hostfile] Unable to find hostfile, will proceed with training with local resources only.
+[2025-10-10 05:58:04,552] [INFO] [runner.py:568:main] cmd = /opt/conda/envs/tinyllava/bin/python3.10 -u -m deepspeed.launcher.launch --world_info=eyJsb2NhbGhvc3QiOiBbMCwgMSwgMiwgMywgNCwgNSwgNiwgN119 --master_addr=127.0.0.1 --master_port=29501 --enable_each_rank_log=None tinyllava/train/train.py --deepspeed ./scripts/zero3.json --data_path /nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json --image_folder /nfs/ywang29/tinyLLaVA/dataset --is_multimodal True --conv_version qwen2_base --model_name_or_path Qwen/Qwen2.5-0.5B --vision_tower google/siglip-so400m-patch14-384 --vision_tower2  --connector_type mlp2x_gelu --mm_vision_select_layer -2 --image_aspect_ratio square --attn_implementation flash_attention_2 --bf16 True --training_recipe common --tune_type_llm full --tune_type_vision_tower frozen --tune_vision_tower_from_layer 0 --tune_type_connector full --group_by_modality_length True --pretrained_model_path /nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain --output_dir /nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_3_connector-3.0_0.5_3_ablation --num_train_epochs 1 --per_device_train_batch_size 4 --per_device_eval_batch_size 4 --gradient_accumulation_steps 4 --evaluation_strategy no --learning_rate 3 --weight_decay 0. --warmup_ratio 0.03 --lr_scheduler_type cosine --logging_steps 1 --tf32 False --model_max_length 2048 --gradient_checkpointing True --dataloader_num_workers 8 --lazy_preprocess True --report_to tensorboard --tokenizer_use_fast False --run_name tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune --subnet_mode_text both --subnet_type_text None --mask_type_text soft --init_mean_text 3.0 --temperature_attn_text 0.5 --temperature_mlp_text 0.5 --backward_type_text normal --masked_layers_text all --subnet_mode_vision both --subnet_type_vision None --mask_type_vision soft --init_mean_vision 3.0 --temperature_attn_vision 0.5 --temperature_mlp_vision 0.5 --backward_type_vision normal --masked_layers_vision all --subnet_type_connector global --mask_type_connector soft --init_mean_connector 3.0 --temperature_connector 0.5 --backward_type_connector normal --mm_projector_lr 3 --seed 42 --mask_model llm-connector --save_strategy steps --save_steps 50000 --save_total_limit 1 --train_data_ratio 0.1
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+[2025-10-10 05:58:07,202] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:58:08,240] [INFO] [launch.py:138:main] 0 NCCL_VERSION=2.21.5
+[2025-10-10 05:58:08,240] [INFO] [launch.py:138:main] 0 NCCL_SOCKET_IFNAME=eth
+[2025-10-10 05:58:08,240] [INFO] [launch.py:145:main] WORLD INFO DICT: {'localhost': [0, 1, 2, 3, 4, 5, 6, 7]}
+[2025-10-10 05:58:08,240] [INFO] [launch.py:151:main] nnodes=1, num_local_procs=8, node_rank=0
+[2025-10-10 05:58:08,240] [INFO] [launch.py:162:main] global_rank_mapping=defaultdict(<class 'list'>, {'localhost': [0, 1, 2, 3, 4, 5, 6, 7]})
+[2025-10-10 05:58:08,240] [INFO] [launch.py:163:main] dist_world_size=8
+[2025-10-10 05:58:08,240] [INFO] [launch.py:165:main] Setting CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+[2025-10-10 05:58:08,242] [INFO] [launch.py:253:main] process 1777438 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=0', '--deepspeed', './scripts/zero3.json', '--data_path', '/nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json', '--image_folder', '/nfs/ywang29/tinyLLaVA/dataset', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_3_connector-3.0_0.5_3_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '3', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '3.0', '--temperature_attn_text', '0.5', '--temperature_mlp_text', '0.5', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '3.0', '--temperature_attn_vision', '0.5', '--temperature_mlp_vision', '0.5', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '3.0', '--temperature_connector', '0.5', '--backward_type_connector', 'normal', '--mm_projector_lr', '3', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 05:58:08,245] [INFO] [launch.py:253:main] process 1777439 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=1', '--deepspeed', './scripts/zero3.json', '--data_path', '/nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json', '--image_folder', '/nfs/ywang29/tinyLLaVA/dataset', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_3_connector-3.0_0.5_3_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '3', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '3.0', '--temperature_attn_text', '0.5', '--temperature_mlp_text', '0.5', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '3.0', '--temperature_attn_vision', '0.5', '--temperature_mlp_vision', '0.5', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '3.0', '--temperature_connector', '0.5', '--backward_type_connector', 'normal', '--mm_projector_lr', '3', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 05:58:08,247] [INFO] [launch.py:253:main] process 1777440 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=2', '--deepspeed', './scripts/zero3.json', '--data_path', '/nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json', '--image_folder', '/nfs/ywang29/tinyLLaVA/dataset', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_3_connector-3.0_0.5_3_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '3', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '3.0', '--temperature_attn_text', '0.5', '--temperature_mlp_text', '0.5', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '3.0', '--temperature_attn_vision', '0.5', '--temperature_mlp_vision', '0.5', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '3.0', '--temperature_connector', '0.5', '--backward_type_connector', 'normal', '--mm_projector_lr', '3', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 05:58:08,249] [INFO] [launch.py:253:main] process 1777441 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=3', '--deepspeed', './scripts/zero3.json', '--data_path', '/nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json', '--image_folder', '/nfs/ywang29/tinyLLaVA/dataset', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_3_connector-3.0_0.5_3_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '3', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '3.0', '--temperature_attn_text', '0.5', '--temperature_mlp_text', '0.5', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '3.0', '--temperature_attn_vision', '0.5', '--temperature_mlp_vision', '0.5', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '3.0', '--temperature_connector', '0.5', '--backward_type_connector', 'normal', '--mm_projector_lr', '3', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 05:58:08,251] [INFO] [launch.py:253:main] process 1777442 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=4', '--deepspeed', './scripts/zero3.json', '--data_path', '/nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json', '--image_folder', '/nfs/ywang29/tinyLLaVA/dataset', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_3_connector-3.0_0.5_3_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '3', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '3.0', '--temperature_attn_text', '0.5', '--temperature_mlp_text', '0.5', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '3.0', '--temperature_attn_vision', '0.5', '--temperature_mlp_vision', '0.5', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '3.0', '--temperature_connector', '0.5', '--backward_type_connector', 'normal', '--mm_projector_lr', '3', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 05:58:08,253] [INFO] [launch.py:253:main] process 1777443 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=5', '--deepspeed', './scripts/zero3.json', '--data_path', '/nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json', '--image_folder', '/nfs/ywang29/tinyLLaVA/dataset', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_3_connector-3.0_0.5_3_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '3', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '3.0', '--temperature_attn_text', '0.5', '--temperature_mlp_text', '0.5', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '3.0', '--temperature_attn_vision', '0.5', '--temperature_mlp_vision', '0.5', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '3.0', '--temperature_connector', '0.5', '--backward_type_connector', 'normal', '--mm_projector_lr', '3', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 05:58:08,255] [INFO] [launch.py:253:main] process 1777444 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=6', '--deepspeed', './scripts/zero3.json', '--data_path', '/nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json', '--image_folder', '/nfs/ywang29/tinyLLaVA/dataset', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_3_connector-3.0_0.5_3_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '3', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '3.0', '--temperature_attn_text', '0.5', '--temperature_mlp_text', '0.5', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '3.0', '--temperature_attn_vision', '0.5', '--temperature_mlp_vision', '0.5', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '3.0', '--temperature_connector', '0.5', '--backward_type_connector', 'normal', '--mm_projector_lr', '3', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+[2025-10-10 05:58:08,257] [INFO] [launch.py:253:main] process 1777445 spawned with command: ['/opt/conda/envs/tinyllava/bin/python3.10', '-u', 'tinyllava/train/train.py', '--local_rank=7', '--deepspeed', './scripts/zero3.json', '--data_path', '/nfs/ywang29/tinyLLaVA/dataset/text_files/llava_v1_5_mix665k.json', '--image_folder', '/nfs/ywang29/tinyLLaVA/dataset', '--is_multimodal', 'True', '--conv_version', 'qwen2_base', '--model_name_or_path', 'Qwen/Qwen2.5-0.5B', '--vision_tower', 'google/siglip-so400m-patch14-384', '--vision_tower2', '', '--connector_type', 'mlp2x_gelu', '--mm_vision_select_layer', '-2', '--image_aspect_ratio', 'square', '--attn_implementation', 'flash_attention_2', '--bf16', 'True', '--training_recipe', 'common', '--tune_type_llm', 'full', '--tune_type_vision_tower', 'frozen', '--tune_vision_tower_from_layer', '0', '--tune_type_connector', 'full', '--group_by_modality_length', 'True', '--pretrained_model_path', '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain', '--output_dir', '/nfs/ywang29/tinyLLaVA/checkpoints/qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_3_connector-3.0_0.5_3_ablation', '--num_train_epochs', '1', '--per_device_train_batch_size', '4', '--per_device_eval_batch_size', '4', '--gradient_accumulation_steps', '4', '--evaluation_strategy', 'no', '--learning_rate', '3', '--weight_decay', '0.', '--warmup_ratio', '0.03', '--lr_scheduler_type', 'cosine', '--logging_steps', '1', '--tf32', 'False', '--model_max_length', '2048', '--gradient_checkpointing', 'True', '--dataloader_num_workers', '8', '--lazy_preprocess', 'True', '--report_to', 'tensorboard', '--tokenizer_use_fast', 'False', '--run_name', 'tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-qwen2.5-0_5b_base-masktune', '--subnet_mode_text', 'both', '--subnet_type_text', 'None', '--mask_type_text', 'soft', '--init_mean_text', '3.0', '--temperature_attn_text', '0.5', '--temperature_mlp_text', '0.5', '--backward_type_text', 'normal', '--masked_layers_text', 'all', '--subnet_mode_vision', 'both', '--subnet_type_vision', 'None', '--mask_type_vision', 'soft', '--init_mean_vision', '3.0', '--temperature_attn_vision', '0.5', '--temperature_mlp_vision', '0.5', '--backward_type_vision', 'normal', '--masked_layers_vision', 'all', '--subnet_type_connector', 'global', '--mask_type_connector', 'soft', '--init_mean_connector', '3.0', '--temperature_connector', '0.5', '--backward_type_connector', 'normal', '--mm_projector_lr', '3', '--seed', '42', '--mask_model', 'llm-connector', '--save_strategy', 'steps', '--save_steps', '50000', '--save_total_limit', '1', '--train_data_ratio', '0.1']
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+[2025-10-10 05:58:14,883] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:58:15,133] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:58:15,196] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:58:15,197] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:58:15,232] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:58:15,247] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:58:15,253] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:58:15,255] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)
+[2025-10-10 05:58:15,294] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 05:58:15,539] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 05:58:15,605] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 05:58:15,607] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 05:58:15,655] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 05:58:15,658] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 05:58:15,659] [INFO] [comm.py:668:init_distributed] Initializing TorchBackend in DeepSpeed with backend nccl
+[2025-10-10 05:58:15,659] [INFO] [comm.py:637:init_distributed] cdb=None
+[2025-10-10 05:58:15,669] [INFO] [comm.py:637:init_distributed] cdb=None
+Apply masks for the following modules:  ['llm', 'connector']
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+Apply masks for the following modules:  ['llm', 'connector']
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+Apply masks for the following modules:  ['llm', 'connector']
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+Apply masks for the following modules:  ['llm', 'connector']
+{'llm': {'model_name_or_path': 'Qwen/Qwen2.5-0.5B', 'cache_dir': None, 'attn_implementation': 'flash_attention_2', 'subnet_mode': 'both', 'subnet_type': 'None', 'sparsity_attn': None, 'sparsity_mlp': None, 'threshold_attn': None, 'threshold_mlp': None, 'temperature_attn': 0.5, 'temperature_mlp': 0.5, 'masked_layers': 'all', 'mask_type': 'soft', 'backward_type': 'normal'}, 'vision_tower': {'model_name_or_path': 'google/siglip-so400m-patch14-384'}, 'connector': {'connector_type': 'mlp2x_gelu', 'subnet_type': 'global', 'threshold': None, 'sparsity': None, 'temperature': 0.5, 'mask_type': 'soft', 'backward_type': 'normal'}}
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+Apply masks for the following modules:  ['llm', 'connector']
+Apply masks for the following modules:  ['llm', 'connector']
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+Apply masks for the following modules:  ['llm', 'connector']
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+Apply masks for the following modules:  ['llm', 'connector']
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/file_download.py:945: FutureWarning: `resume_download` is deprecated and will be removed in version 1.0.0. Downloads always resume when possible. If you want to force a new download, use `force_download=True`.
+  warnings.warn(
+TinyLlavaConfig {
+  "backward_type_connector": "normal",
+  "cache_dir": null,
+  "connector_type": "mlp2x_gelu",
+  "hidden_size": 896,
+  "ignore_index": -100,
+  "image_aspect_ratio": "square",
+  "image_token_index": -200,
+  "llm_model_name_or_path": "Qwen/Qwen2.5-0.5B",
+  "mask_model": [
+    "llm",
+    "connector"
+  ],
+  "mask_type_connector": "soft",
+  "model_type": "tinyllava",
+  "num_queries": 128,
+  "num_resampler_layers": 3,
+  "pad_token": null,
+  "resampler_hidden_size": 768,
+  "sparsity_connector": null,
+  "subnet_type_connector": "global",
+  "temperature_connector": 0.5,
+  "text_config": {
+    "_name_or_path": "Qwen/Qwen2.5-0.5B",
+    "architectures": [
+      "Qwen2ForCausalLM"
+    ],
+    "backward_type": "normal",
+    "bos_token_id": 151643,
+    "eos_token_id": 151643,
+    "hidden_size": 896,
+    "intermediate_size": 4864,
+    "mask_type": "soft",
+    "masked_layers": "all",
+    "max_position_embeddings": 32768,
+    "max_window_layers": 24,
+    "model_type": "qwen2",
+    "num_attention_heads": 14,
+    "num_hidden_layers": 24,
+    "num_key_value_heads": 2,
+    "rope_theta": 1000000.0,
+    "sliding_window": 32768,
+    "subnet_mode": "both",
+    "subnet_type": "None",
+    "temperature_attn": 0.5,
+    "temperature_mlp": 0.5,
+    "tie_word_embeddings": true,
+    "torch_dtype": "bfloat16",
+    "use_mrope": false,
+    "use_sliding_window": false,
+    "vocab_size": 151936
+  },
+  "threshold_connector": null,
+  "tokenizer_model_max_length": 2048,
+  "tokenizer_name_or_path": "Qwen/Qwen2.5-0.5B",
+  "tokenizer_padding_side": "right",
+  "tokenizer_use_fast": false,
+  "transformers_version": "4.40.1",
+  "tune_type_connector": "frozen",
+  "tune_type_llm": "frozen",
+  "tune_type_vision_tower": "frozen",
+  "tune_vision_tower_from_layer": -1,
+  "use_cache": false,
+  "vision_config": {
+    "hidden_act": "gelu_pytorch_tanh",
+    "hidden_size": 1152,
+    "image_size": 384,
+    "intermediate_size": 4304,
+    "layer_norm_eps": 1e-06,
+    "model_name_or_path": "google/siglip-so400m-patch14-384",
+    "model_name_or_path2": "",
+    "model_type": "siglip_vision_model",
+    "num_attention_heads": 16,
+    "num_hidden_layers": 27,
+    "patch_size": 14
+  },
+  "vision_feature_layer": -2,
+  "vision_feature_select_strategy": "patch",
+  "vision_hidden_size": 1152,
+  "vision_model_name_or_path": "google/siglip-so400m-patch14-384",
+  "vision_model_name_or_path2": "",
+  "vocab_size": 151936
+}
+
+Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
+Traceback (most recent call last):
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/utils/hub.py", line 398, in cached_file
+    resolved_file = hf_hub_download(
+  File "/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py", line 106, in _inner_fn
+    validate_repo_id(arg_value)
+  File "/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py", line 154, in validate_repo_id
+    raise HFValidationError(
+huggingface_hub.errors.HFValidationError: Repo id must be in the form 'repo_name' or 'namespace/repo_name': '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain/language_model'. Use `repo_type` argument if needed.
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/train/train.py", line 193, in <module>
+    train()
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/train/train.py", line 149, in train
+    model = training_recipe.load(model, model_args)
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/training_recipe/base.py", line 144, in load
+    model.load_llm(**model_args['llm'])
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/model/modeling_tinyllava.py", line 360, in load_llm
+    self.language_model = self.language_model.from_pretrained(
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/modeling_utils.py", line 3015, in from_pretrained
+    resolved_config_file = cached_file(
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/utils/hub.py", line 462, in cached_file
+    raise EnvironmentError(
+OSError: Incorrect path_or_model_id: '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain/language_model'. Please provide either the path to a local folder or the repo_id of a model on the Hub.
+Special tokens have been added in the vocabulary, make sure the associated word embeddings are fine-tuned or trained.
+Traceback (most recent call last):
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/utils/hub.py", line 398, in cached_file
+    resolved_file = hf_hub_download(
+  File "/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py", line 106, in _inner_fn
+    validate_repo_id(arg_value)
+  File "/opt/conda/envs/tinyllava/lib/python3.10/site-packages/huggingface_hub/utils/_validators.py", line 154, in validate_repo_id
+    raise HFValidationError(
+huggingface_hub.errors.HFValidationError: Repo id must be in the form 'repo_name' or 'namespace/repo_name': '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain/language_model'. Use `repo_type` argument if needed.
+
+The above exception was the direct cause of the following exception:
+
+Traceback (most recent call last):
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/train/train.py", line 193, in <module>
+    train()
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/train/train.py", line 149, in train
+    model = training_recipe.load(model, model_args)
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/training_recipe/base.py", line 144, in load
+    model.load_llm(**model_args['llm'])
+  File "/nfs/ywang29/TinyLLaVA/tinyllava/model/modeling_tinyllava.py", line 360, in load_llm
+    self.language_model = self.language_model.from_pretrained(
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/modeling_utils.py", line 3015, in from_pretrained
+    resolved_config_file = cached_file(
+  File "/nfs/ywang29/TinyLLaVA/transformers/src/transformers/utils/hub.py", line 462, in cached_file
+    raise EnvironmentError(
+OSError: Incorrect path_or_model_id: '/nfs/ywang29/tinyLLaVA/checkpoints/tiny-llava-Qwen2.5-0.5B-siglip-so400m-patch14-384-pretrain/language_model'. Please provide either the path to a local folder or the repo_id of a model on the Hub.
+[2025-10-10 05:58:50,304] [INFO] [launch.py:316:sigkill_handler] Killing subprocess 1777438
+[2025-10-10 05:58:50,689] [INFO] [launch.py:316:sigkill_handler] Killing subprocess 1777439
+[2025-10-10 05:58:50,689] [INFO] [launch.py:316:sigkill_handler] Killing subprocess 1777440
+[2025-10-10 05:58:51,067] [INFO] [launch.py:316:sigkill_handler] Killing subprocess 1777441

logs_oct10/qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_7_connector-3.0_0.5_7_ablation_20251010_060221.log

@@ -0,0 +1,7 @@
+==== STARTING EXPERIMENT: qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_7_connector-3.0_0.5_7_ablation ====
+Log File: qwen2.5-0_5b_base_masktune_42_llm-connector_text-3.0_0.5_7_connector-3.0_0.5_7_ablation_20251010_060221.log
+Timestamp: 2025-10-10 06:02:21
+=====================================
+/opt/conda/envs/tinyllava/lib/python3.10/site-packages/torch/cuda/__init__.py:51: FutureWarning: The pynvml package is deprecated. Please install nvidia-ml-py instead. If you did not install pynvml directly, please report this to the maintainers of the package that installed pynvml for you.
+  import pynvml  # type: ignore[import]
+[2025-10-10 06:02:24,057] [INFO] [real_accelerator.py:191:get_accelerator] Setting ds_accelerator to cuda (auto detect)