Upload checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins

checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/wandb/offline-run-20260129_223514-checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins-run0/files/config.yaml

@@ -0,0 +1,459 @@
+wandb_version: 1
+
+_wandb:
+  desc: null
+  value:
+    python_version: 3.11.10
+    cli_version: 0.23.1
+    framework: huggingface
+    huggingface_version: 4.49.0
+    is_jupyter_run: false
+    is_kaggle_kernel: false
+    start_time: 1769726114
+    t:
+      1:
+      - 1
+      - 5
+      - 11
+      - 41
+      - 49
+      - 53
+      - 71
+      - 105
+      2:
+      - 1
+      - 5
+      - 11
+      - 41
+      - 49
+      - 53
+      - 71
+      - 105
+      3:
+      - 2
+      - 4
+      - 13
+      - 14
+      - 37
+      - 42
+      - 61
+      4: 3.11.10
+      5: 0.23.1
+      6: 4.49.0
+      13: linux-x86_64
+    e:
+      cf0m5et0f6ffnx56j9dnkbvf0iq03h00:
+        os: Linux-6.6.93+-x86_64-with-glibc2.35
+        python: CPython 3.11.10
+        started_at: '2026-01-29T22:35:14.394399Z'
+        args:
+        - --dataset_config_file
+        - ./data/configs/vlm_gym_toy_maze_2d_train_celoss_no_mse.yaml
+        - --eval_dataset_config_file
+        - ./data/configs/vlm_gym_toy_maze_2d_eval_celoss_no_mse.yaml
+        - --viz_dataset_config_file
+        - ./data/configs/vlm_gym_toy_maze_2d_eval_celoss_no_mse.yaml
+        - --train_data_dir
+        - /home/clouduser/Code/data/gym/toy_maze_2d_v5/train/
+        - --train_jsonl_path
+        - /home/clouduser/Code/data/gym/toy_maze_2d_v5/train/
+        - --eval_data_dir
+        - /home/clouduser/Code/data/gym/toy_maze_2d_v5/val/
+        - --eval_jsonl_path
+        - /home/clouduser/Code/data/gym/toy_maze_2d_v5/val/
+        - --inference_hash_file
+        - /home/clouduser/Code/Github/launch_new/hashes_test_set_v10.json
+        - --task_name
+        - toy_maze_2d_v5
+        - --instructions_dir
+        - ./data/instructions
+        - --model_path
+        - /home/clouduser/Code/Models/BAGEL-7B-MoT
+        - --layer_module
+        - Qwen2MoTDecoderLayer
+        - --max_latent_size
+        - '64'
+        - --resume-from
+        - /home/clouduser/Code/Models/BAGEL-7B-MoT
+        - --finetune_from_hf
+        - 'True'
+        - --auto_resume
+        - 'False'
+        - --resume-model-only
+        - 'True'
+        - --finetune-from-ema
+        - 'True'
+        - --log_every
+        - '1'
+        - --lr
+        - 2e-5
+        - --warmup_steps
+        - '300'
+        - --lr_scheduler
+        - cosine
+        - --num_worker
+        - '1'
+        - --expected_num_tokens
+        - '30000'
+        - --max_num_tokens
+        - '30000'
+        - --max_num_tokens_per_sample
+        - '30000'
+        - --visual_und
+        - 'True'
+        - --visual_gen
+        - 'False'
+        - --save_every
+        - '5000'
+        - --total_steps
+        - '5000'
+        - --text_cond_dropout_prob
+        - '0.0'
+        - --vae_cond_dropout_prob
+        - '0.3'
+        - --vit_cond_dropout_prob
+        - '0.0'
+        - --ema
+        - '0.993'
+        - --checkpoint_dir
+        - /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins
+        - --wandb_project
+        - bagel
+        - --wandb_name
+        - checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins
+        - --wandb_dir
+        - /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins
+        - --wandb_offline
+        - 'True'
+        program: /home/clouduser/Code/Github/unified_world_model/train/pretrain_unified_navit.py
+        code_path: train/pretrain_unified_navit.py
+        code_path_local: train/pretrain_unified_navit.py
+        git:
+          remote_url: https://github.com/para-lost/unified_world_model
+          commit: 8d7b26b7e552fc87b592cf3be94d93be7aeca2a9
+        root: /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins
+        host: junyizhang-launch-new-226786579-1-0
+        executable: /opt/conda/bin/python3.11
+        cpu_count: 48
+        cpu_count_logical: 96
+        gpu_type: NVIDIA A100-SXM4-80GB
+        gpu_count: 8
+        disk:
+          /:
+            total: '1052461830144'
+            used: '187289595904'
+        memory:
+          total: '1437332606976'
+        gpu_nvidia:
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-7fb62b69-ecda-78ca-6a78-e9242da41257
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-63537668-be67-064d-0339-a4807fac2e17
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-3d52c06a-5ceb-fc52-708a-f5c75328664b
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-befa8eef-f6ca-6ce5-9b17-6d284a071265
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-87f27493-86c3-9a50-4e27-a7100ec26ac4
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-2d4912d0-bf74-6ded-3c18-922df6d46383
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-2eb784d3-6f27-8d12-081c-a0de47b17206
+        - name: NVIDIA A100-SXM4-80GB
+          memory_total: '85899345920'
+          cuda_cores: 6912
+          architecture: Ampere
+          uuid: GPU-ac067d32-e7bf-7556-dcab-e5a49410eb3d
+        cuda_version: '12.2'
+        writer_id: cf0m5et0f6ffnx56j9dnkbvf0iq03h00
+visual_gen:
+  desc: null
+  value: false
+visual_und:
+  desc: null
+  value: true
+results_dir:
+  desc: null
+  value: results
+checkpoint_dir:
+  desc: null
+  value: /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins
+wandb_project:
+  desc: null
+  value: bagel
+wandb_name:
+  desc: null
+  value: checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins
+wandb_runid:
+  desc: null
+  value: '0'
+wandb_resume:
+  desc: null
+  value: allow
+wandb_offline:
+  desc: null
+  value: true
+wandb_dir:
+  desc: null
+  value: /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins
+global_seed:
+  desc: null
+  value: 4396
+auto_resume:
+  desc: null
+  value: false
+resume_from:
+  desc: null
+  value: /home/clouduser/Code/Models/BAGEL-7B-MoT
+resume_model_only:
+  desc: null
+  value: true
+finetune_from_ema:
+  desc: null
+  value: true
+finetune_from_hf:
+  desc: null
+  value: true
+log_every:
+  desc: null
+  value: 1
+save_every:
+  desc: null
+  value: 5000
+total_steps:
+  desc: null
+  value: 5000
+warmup_steps:
+  desc: null
+  value: 300
+lr_scheduler:
+  desc: null
+  value: cosine
+lr:
+  desc: null
+  value: 2.0e-05
+min_lr:
+  desc: null
+  value: 1.0e-07
+beta1:
+  desc: null
+  value: 0.9
+beta2:
+  desc: null
+  value: 0.95
+eps:
+  desc: null
+  value: 1.0e-15
+ema:
+  desc: null
+  value: 0.993
+max_grad_norm:
+  desc: null
+  value: 1.0
+timestep_shift:
+  desc: null
+  value: 1.0
+mse_weight:
+  desc: null
+  value: 1.0
+ce_weight:
+  desc: null
+  value: 1.0
+ce_loss_reweighting:
+  desc: null
+  value: false
+expected_num_tokens:
+  desc: null
+  value: 30000
+num_replicate:
+  desc: null
+  value: 1
+num_shard:
+  desc: null
+  value: 8
+sharding_strategy:
+  desc: null
+  value: HYBRID_SHARD
+backward_prefetch:
+  desc: null
+  value: BACKWARD_PRE
+cpu_offload:
+  desc: null
+  value: false
+freeze_llm:
+  desc: null
+  value: false
+freeze_vit:
+  desc: null
+  value: false
+freeze_vae:
+  desc: null
+  value: true
+freeze_und:
+  desc: null
+  value: false
+copy_init_moe:
+  desc: null
+  value: true
+use_flex:
+  desc: null
+  value: false
+eval_every:
+  desc: null
+  value: 500
+num_eval_batches:
+  desc: null
+  value: 20
+use_ema_for_eval:
+  desc: null
+  value: true
+eval_log_dir:
+  desc: null
+  value: null
+eval_run_tag:
+  desc: null
+  value: ''
+viz_every:
+  desc: null
+  value: 500
+viz_n:
+  desc: null
+  value: 8
+viz_outdir:
+  desc: null
+  value: results/viz
+eval_dataset_config_file:
+  desc: null
+  value: ./data/configs/vlm_gym_toy_maze_2d_eval_celoss_no_mse.yaml
+viz_dataset_config_file:
+  desc: null
+  value: ./data/configs/vlm_gym_toy_maze_2d_eval_celoss_no_mse.yaml
+eval_print_n:
+  desc: null
+  value: 3
+save_ema_only:
+  desc: null
+  value: true
+save_optimizer:
+  desc: null
+  value: false
+model_path:
+  desc: null
+  value: /home/clouduser/Code/Models/BAGEL-7B-MoT
+llm_path:
+  desc: null
+  value: hf/Qwen2.5-0.5B-Instruct/
+llm_qk_norm:
+  desc: null
+  value: true
+tie_word_embeddings:
+  desc: null
+  value: false
+layer_module:
+  desc: null
+  value: Qwen2MoTDecoderLayer
+vae_path:
+  desc: null
+  value: flux/vae/ae.safetensors
+vit_path:
+  desc: null
+  value: hf/siglip-so400m-14-980-flash-attn2-navit/
+max_latent_size:
+  desc: null
+  value: 64
+latent_patch_size:
+  desc: null
+  value: 2
+vit_patch_size:
+  desc: null
+  value: 14
+vit_max_num_patch_per_side:
+  desc: null
+  value: 70
+connector_act:
+  desc: null
+  value: gelu_pytorch_tanh
+interpolate_pos:
+  desc: null
+  value: false
+vit_select_layer:
+  desc: null
+  value: -2
+vit_rope:
+  desc: null
+  value: false
+text_cond_dropout_prob:
+  desc: null
+  value: 0.0
+vae_cond_dropout_prob:
+  desc: null
+  value: 0.3
+vit_cond_dropout_prob:
+  desc: null
+  value: 0.0
+dataset_config_file:
+  desc: null
+  value: ./data/configs/vlm_gym_toy_maze_2d_train_celoss_no_mse.yaml
+train_data_dir:
+  desc: null
+  value: /home/clouduser/Code/data/gym/toy_maze_2d_v5/train/
+train_jsonl_path:
+  desc: null
+  value: /home/clouduser/Code/data/gym/toy_maze_2d_v5/train/
+eval_data_dir:
+  desc: null
+  value: /home/clouduser/Code/data/gym/toy_maze_2d_v5/val/
+eval_jsonl_path:
+  desc: null
+  value: /home/clouduser/Code/data/gym/toy_maze_2d_v5/val/
+inference_hash_file:
+  desc: null
+  value: /home/clouduser/Code/Github/launch_new/hashes_test_set_v10.json
+task_name:
+  desc: null
+  value: toy_maze_2d_v5
+instructions_dir:
+  desc: null
+  value: ./data/instructions
+prefetch_factor:
+  desc: null
+  value: 2
+num_workers:
+  desc: null
+  value: 1
+max_num_tokens_per_sample:
+  desc: null
+  value: 30000
+max_num_tokens:
+  desc: null
+  value: 30000
+prefer_buffer_before:
+  desc: null
+  value: 16384
+max_buffer_size:
+  desc: null
+  value: 50
+data_seed:
+  desc: null
+  value: 42

checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/wandb/offline-run-20260129_223514-checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins-run0/files/output.log

@@ -1,3 +1,173 @@
+FullyShardedDataParallel(
+  (_fsdp_wrapped_module): Bagel(
+    (language_model): Qwen2ForCausalLM(
+      (model): Qwen2Model(
+        (embed_tokens): Embedding(152064, 3584)
+        (layers): ModuleList(
+          (0-27): 28 x FullyShardedDataParallel(
+            (_fsdp_wrapped_module): CheckpointWrapper(
+              (_checkpoint_wrapped_module): Qwen2MoTDecoderLayer(
+                (self_attn): PackedAttentionMoT(
+                  (q_proj): Linear(in_features=3584, out_features=3584, bias=True)
+                  (k_proj): Linear(in_features=3584, out_features=512, bias=True)
+                  (v_proj): Linear(in_features=3584, out_features=512, bias=True)
+                  (o_proj): Linear(in_features=3584, out_features=3584, bias=False)
+                  (q_norm): Qwen2RMSNorm((128,), eps=1e-06)
+                  (k_norm): Qwen2RMSNorm((128,), eps=1e-06)
+                  (q_norm_moe_gen): Qwen2RMSNorm((128,), eps=1e-06)
+                  (k_norm_moe_gen): Qwen2RMSNorm((128,), eps=1e-06)
+                  (q_proj_moe_gen): Linear(in_features=3584, out_features=3584, bias=True)
+                  (k_proj_moe_gen): Linear(in_features=3584, out_features=512, bias=True)
+                  (v_proj_moe_gen): Linear(in_features=3584, out_features=512, bias=True)
+                  (o_proj_moe_gen): Linear(in_features=3584, out_features=3584, bias=False)
+                )
+                (mlp): Qwen2MLP(
+                  (gate_proj): Linear(in_features=3584, out_features=18944, bias=False)
+                  (up_proj): Linear(in_features=3584, out_features=18944, bias=False)
+                  (down_proj): Linear(in_features=18944, out_features=3584, bias=False)
+                  (act_fn): SiLU()
+                )
+                (mlp_moe_gen): Qwen2MLP(
+                  (gate_proj): Linear(in_features=3584, out_features=18944, bias=False)
+                  (up_proj): Linear(in_features=3584, out_features=18944, bias=False)
+                  (down_proj): Linear(in_features=18944, out_features=3584, bias=False)
+                  (act_fn): SiLU()
+                )
+                (input_layernorm): Qwen2RMSNorm((3584,), eps=1e-06)
+                (input_layernorm_moe_gen): Qwen2RMSNorm((3584,), eps=1e-06)
+                (post_attention_layernorm): Qwen2RMSNorm((3584,), eps=1e-06)
+                (post_attention_layernorm_moe_gen): Qwen2RMSNorm((3584,), eps=1e-06)
+              )
+            )
+          )
+        )
+        (norm): Qwen2RMSNorm((3584,), eps=1e-06)
+        (norm_moe_gen): Qwen2RMSNorm((3584,), eps=1e-06)
+        (rotary_emb): Qwen2RotaryEmbedding()
+      )
+      (lm_head): Linear(in_features=3584, out_features=152064, bias=False)
+    )
+    (vit_model): SiglipVisionModel(
+      (vision_model): FullyShardedDataParallel(
+        (_fsdp_wrapped_module): SiglipVisionTransformer(
+          (embeddings): SiglipVisionEmbeddings(
+            (position_embedding): Embedding(4900, 1152)
+            (patch_embedding): Linear(in_features=588, out_features=1152, bias=True)
+          )
+          (encoder): SiglipEncoder(
+            (layers): ModuleList(
+              (0-25): 26 x FullyShardedDataParallel(
+                (_fsdp_wrapped_module): CheckpointWrapper(
+                  (_checkpoint_wrapped_module): SiglipEncoderLayer(
+                    (self_attn): SiglipFlashAttention2(
+                      (k_proj): Linear(in_features=1152, out_features=1152, bias=True)
+                      (v_proj): Linear(in_features=1152, out_features=1152, bias=True)
+                      (q_proj): Linear(in_features=1152, out_features=1152, bias=True)
+                      (out_proj): Linear(in_features=1152, out_features=1152, bias=True)
+                    )
+                    (layer_norm1): LayerNorm((1152,), eps=1e-06, elementwise_affine=True)
+                    (mlp): SiglipMLP(
+                      (activation_fn): PytorchGELUTanh()
+                      (fc1): Linear(in_features=1152, out_features=4304, bias=True)
+                      (fc2): Linear(in_features=4304, out_features=1152, bias=True)
+                    )
+                    (layer_norm2): LayerNorm((1152,), eps=1e-06, elementwise_affine=True)
+                  )
+                )
+              )
+            )
+          )
+          (post_layernorm): LayerNorm((1152,), eps=1e-06, elementwise_affine=True)
+        )
+      )
+    )
+    (connector): FullyShardedDataParallel(
+      (_fsdp_wrapped_module): CheckpointWrapper(
+        (_checkpoint_wrapped_module): MLPconnector(
+          (activation_fn): PytorchGELUTanh()
+          (fc1): Linear(in_features=1152, out_features=3584, bias=True)
+          (fc2): Linear(in_features=3584, out_features=3584, bias=True)
+        )
+      )
+    )
+    (vit_pos_embed): FullyShardedDataParallel(
+      (_fsdp_wrapped_module): PositionEmbedding()
+    )
+  )
+)
+_flat_param True
+language_model.model.layers.0._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.1._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.2._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.3._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.4._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.5._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.6._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.7._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.8._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.9._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.10._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.11._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.12._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.13._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.14._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.15._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.16._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.17._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.18._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.19._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.20._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.21._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.22._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.23._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.24._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.25._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.26._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+language_model.model.layers.27._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.0._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.1._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.2._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.3._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.4._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.5._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.6._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.7._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.8._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.9._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.10._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.11._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.12._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.13._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.14._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.15._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.16._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.17._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.18._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.19._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.20._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.21._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.22._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.23._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.24._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_model.vision_model._fsdp_wrapped_module.encoder.layers.25._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+connector._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
+vit_pos_embed._fsdp_wrapped_module._flat_param False
+Preparing Dataset vlm_gym_toy_maze_2d_celoss_no_mse/vlm_gym_toy_maze_2d_train
+base_dir is /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins_step0
+Preparing Dataset vlm_gym_toy_maze_2d_celoss_no_mse_evalonce/vlm_gym_toy_maze_2d_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+ce_avg: 0.33020690083503723, mse_avg: 0.0
+base_dir is /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins_step500
+Preparing Dataset vlm_gym_toy_maze_2d_celoss_no_mse_evalonce/vlm_gym_toy_maze_2d_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+ce_avg: 0.04093625023961067, mse_avg: 0.0
 wandb: Detected [huggingface_hub.inference] in use.
 wandb: Use W&B Weave for improved LLM call tracing. Install Weave with `pip install weave` then add `import weave` to the top of your script.
 wandb: For more information, check out the docs at: https://weave-docs.wandb.ai/
@@ -917,176 +1087,6 @@ wandb: For more information, check out the docs at: https://weave-docs.wandb.ai/
 [2026-01-29 23:51:13] (step=0000906) Train Loss mse: 0.0000, Train Loss ce: 0.0422, Train Steps/Sec: 0.21,
 [2026-01-29 23:51:17] (step=0000907) Train Loss mse: 0.0000, Train Loss ce: 0.0435, Train Steps/Sec: 0.24,
 [2026-01-29 23:51:22] (step=0000908) Train Loss mse: 0.0000, Train Loss ce: 0.0414, Train Steps/Sec: 0.21,
-FullyShardedDataParallel(
-  (_fsdp_wrapped_module): Bagel(
-    (language_model): Qwen2ForCausalLM(
-      (model): Qwen2Model(
-        (embed_tokens): Embedding(152064, 3584)
-        (layers): ModuleList(
-          (0-27): 28 x FullyShardedDataParallel(
-            (_fsdp_wrapped_module): CheckpointWrapper(
-              (_checkpoint_wrapped_module): Qwen2MoTDecoderLayer(
-                (self_attn): PackedAttentionMoT(
-                  (q_proj): Linear(in_features=3584, out_features=3584, bias=True)
-                  (k_proj): Linear(in_features=3584, out_features=512, bias=True)
-                  (v_proj): Linear(in_features=3584, out_features=512, bias=True)
-                  (o_proj): Linear(in_features=3584, out_features=3584, bias=False)
-                  (q_norm): Qwen2RMSNorm((128,), eps=1e-06)
-                  (k_norm): Qwen2RMSNorm((128,), eps=1e-06)
-                  (q_norm_moe_gen): Qwen2RMSNorm((128,), eps=1e-06)
-                  (k_norm_moe_gen): Qwen2RMSNorm((128,), eps=1e-06)
-                  (q_proj_moe_gen): Linear(in_features=3584, out_features=3584, bias=True)
-                  (k_proj_moe_gen): Linear(in_features=3584, out_features=512, bias=True)
-                  (v_proj_moe_gen): Linear(in_features=3584, out_features=512, bias=True)
-                  (o_proj_moe_gen): Linear(in_features=3584, out_features=3584, bias=False)
-                )
-                (mlp): Qwen2MLP(
-                  (gate_proj): Linear(in_features=3584, out_features=18944, bias=False)
-                  (up_proj): Linear(in_features=3584, out_features=18944, bias=False)
-                  (down_proj): Linear(in_features=18944, out_features=3584, bias=False)
-                  (act_fn): SiLU()
-                )
-                (mlp_moe_gen): Qwen2MLP(
-                  (gate_proj): Linear(in_features=3584, out_features=18944, bias=False)
-                  (up_proj): Linear(in_features=3584, out_features=18944, bias=False)
-                  (down_proj): Linear(in_features=18944, out_features=3584, bias=False)
-                  (act_fn): SiLU()
-                )
-                (input_layernorm): Qwen2RMSNorm((3584,), eps=1e-06)
-                (input_layernorm_moe_gen): Qwen2RMSNorm((3584,), eps=1e-06)
-                (post_attention_layernorm): Qwen2RMSNorm((3584,), eps=1e-06)
-                (post_attention_layernorm_moe_gen): Qwen2RMSNorm((3584,), eps=1e-06)
-              )
-            )
-          )
-        )
-        (norm): Qwen2RMSNorm((3584,), eps=1e-06)
-        (norm_moe_gen): Qwen2RMSNorm((3584,), eps=1e-06)
-        (rotary_emb): Qwen2RotaryEmbedding()
-      )
-      (lm_head): Linear(in_features=3584, out_features=152064, bias=False)
-    )
-    (vit_model): SiglipVisionModel(
-      (vision_model): FullyShardedDataParallel(
-        (_fsdp_wrapped_module): SiglipVisionTransformer(
-          (embeddings): SiglipVisionEmbeddings(
-            (position_embedding): Embedding(4900, 1152)
-            (patch_embedding): Linear(in_features=588, out_features=1152, bias=True)
-          )
-          (encoder): SiglipEncoder(
-            (layers): ModuleList(
-              (0-25): 26 x FullyShardedDataParallel(
-                (_fsdp_wrapped_module): CheckpointWrapper(
-                  (_checkpoint_wrapped_module): SiglipEncoderLayer(
-                    (self_attn): SiglipFlashAttention2(
-                      (k_proj): Linear(in_features=1152, out_features=1152, bias=True)
-                      (v_proj): Linear(in_features=1152, out_features=1152, bias=True)
-                      (q_proj): Linear(in_features=1152, out_features=1152, bias=True)
-                      (out_proj): Linear(in_features=1152, out_features=1152, bias=True)
-                    )
-                    (layer_norm1): LayerNorm((1152,), eps=1e-06, elementwise_affine=True)
-                    (mlp): SiglipMLP(
-                      (activation_fn): PytorchGELUTanh()
-                      (fc1): Linear(in_features=1152, out_features=4304, bias=True)
-                      (fc2): Linear(in_features=4304, out_features=1152, bias=True)
-                    )
-                    (layer_norm2): LayerNorm((1152,), eps=1e-06, elementwise_affine=True)
-                  )
-                )
-              )
-            )
-          )
-          (post_layernorm): LayerNorm((1152,), eps=1e-06, elementwise_affine=True)
-        )
-      )
-    )
-    (connector): FullyShardedDataParallel(
-      (_fsdp_wrapped_module): CheckpointWrapper(
-        (_checkpoint_wrapped_module): MLPconnector(
-          (activation_fn): PytorchGELUTanh()
-          (fc1): Linear(in_features=1152, out_features=3584, bias=True)
-          (fc2): Linear(in_features=3584, out_features=3584, bias=True)
-        )
-      )
-    )
-    (vit_pos_embed): FullyShardedDataParallel(
-      (_fsdp_wrapped_module): PositionEmbedding()
-    )
-  )
-)
-_flat_param True
-language_model.model.layers.0._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.1._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.2._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.3._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.4._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.5._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.6._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.7._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.8._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.9._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.10._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.11._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.12._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.13._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.14._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.15._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.16._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.17._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.18._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.19._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.20._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.21._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.22._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.23._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.24._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.25._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.26._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-language_model.model.layers.27._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.0._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.1._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.2._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.3._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.4._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.5._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.6._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.7._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.8._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.9._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.10._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.11._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.12._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.13._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.14._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.15._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.16._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.17._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.18._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.19._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.20._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.21._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.22._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.23._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.24._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_model.vision_model._fsdp_wrapped_module.encoder.layers.25._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-connector._fsdp_wrapped_module._checkpoint_wrapped_module._flat_param True
-vit_pos_embed._fsdp_wrapped_module._flat_param False
-Preparing Dataset vlm_gym_toy_maze_2d_celoss_no_mse/vlm_gym_toy_maze_2d_train
-base_dir is /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins_step0
-Preparing Dataset vlm_gym_toy_maze_2d_celoss_no_mse_evalonce/vlm_gym_toy_maze_2d_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-ce_avg: 0.33020690083503723, mse_avg: 0.0
-base_dir is /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins_step500
-Preparing Dataset vlm_gym_toy_maze_2d_celoss_no_mse_evalonce/vlm_gym_toy_maze_2d_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-ce_avg: 0.04093625023961067, mse_avg: 0.0
 [2026-01-29 23:51:26] (step=0000909) Train Loss mse: 0.0000, Train Loss ce: 0.0391, Train Steps/Sec: 0.22,
 [2026-01-29 23:51:31] (step=0000910) Train Loss mse: 0.0000, Train Loss ce: 0.0372, Train Steps/Sec: 0.21,
 [2026-01-29 23:51:36] (step=0000911) Train Loss mse: 0.0000, Train Loss ce: 0.0413, Train Steps/Sec: 0.20,
@@ -1301,6 +1301,27 @@ ce_avg: 0.04093625023961067, mse_avg: 0.0
 [2026-01-30 00:07:51] (step=0001120) Train Loss mse: 0.0000, Train Loss ce: 0.0360, Train Steps/Sec: 0.21,
 [2026-01-30 00:07:55] (step=0001121) Train Loss mse: 0.0000, Train Loss ce: 0.0358, Train Steps/Sec: 0.24,
 [2026-01-30 00:07:59] (step=0001122) Train Loss mse: 0.0000, Train Loss ce: 0.0416, Train Steps/Sec: 0.22,
+base_dir is /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins_step1000
+Preparing Dataset vlm_gym_toy_maze_2d_celoss_no_mse_evalonce/vlm_gym_toy_maze_2d_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+ce_avg: 0.0632656142115593, mse_avg: 0.0
+base_dir is /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins_step1500
+Preparing Dataset vlm_gym_toy_maze_2d_celoss_no_mse_evalonce/vlm_gym_toy_maze_2d_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+ce_avg: 0.07227745652198792, mse_avg: 0.0
+base_dir is /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins_step2000
+Preparing Dataset vlm_gym_toy_maze_2d_celoss_no_mse_evalonce/vlm_gym_toy_maze_2d_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+ce_avg: 0.09361912310123444, mse_avg: 0.0
 [2026-01-30 00:08:04] (step=0001123) Train Loss mse: 0.0000, Train Loss ce: 0.0416, Train Steps/Sec: 0.23,
 [2026-01-30 00:08:09] (step=0001124) Train Loss mse: 0.0000, Train Loss ce: 0.0348, Train Steps/Sec: 0.20,
 [2026-01-30 00:08:13] (step=0001125) Train Loss mse: 0.0000, Train Loss ce: 0.0423, Train Steps/Sec: 0.21,
@@ -2467,41 +2488,6 @@ ce_avg: 0.04093625023961067, mse_avg: 0.0
 [2026-01-30 01:37:24] (step=0002286) Train Loss mse: 0.0000, Train Loss ce: 0.0391, Train Steps/Sec: 0.22,
 [2026-01-30 01:37:29] (step=0002287) Train Loss mse: 0.0000, Train Loss ce: 0.0358, Train Steps/Sec: 0.21,
 [2026-01-30 01:37:33] (step=0002288) Train Loss mse: 0.0000, Train Loss ce: 0.0337, Train Steps/Sec: 0.22,
-base_dir is /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins_step1000
-Preparing Dataset vlm_gym_toy_maze_2d_celoss_no_mse_evalonce/vlm_gym_toy_maze_2d_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-ce_avg: 0.0632656142115593, mse_avg: 0.0
-base_dir is /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins_step1500
-Preparing Dataset vlm_gym_toy_maze_2d_celoss_no_mse_evalonce/vlm_gym_toy_maze_2d_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-ce_avg: 0.07227745652198792, mse_avg: 0.0
-base_dir is /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins_step2000
-Preparing Dataset vlm_gym_toy_maze_2d_celoss_no_mse_evalonce/vlm_gym_toy_maze_2d_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-ce_avg: 0.09361912310123444, mse_avg: 0.0
-base_dir is /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins_step2500
-Preparing Dataset vlm_gym_toy_maze_2d_celoss_no_mse_evalonce/vlm_gym_toy_maze_2d_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-ce_avg: 0.11742664873600006, mse_avg: 0.0
-base_dir is /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins_step3000
-Preparing Dataset vlm_gym_toy_maze_2d_celoss_no_mse_evalonce/vlm_gym_toy_maze_2d_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-ce_avg: 0.15227733552455902, mse_avg: 0.0
 [2026-01-30 01:37:38] (step=0002289) Train Loss mse: 0.0000, Train Loss ce: 0.0288, Train Steps/Sec: 0.21,
 [2026-01-30 01:37:42] (step=0002290) Train Loss mse: 0.0000, Train Loss ce: 0.0374, Train Steps/Sec: 0.22,
 [2026-01-30 01:37:47] (step=0002291) Train Loss mse: 0.0000, Train Loss ce: 0.0294, Train Steps/Sec: 0.22,
@@ -2834,6 +2820,27 @@ ce_avg: 0.15227733552455902, mse_avg: 0.0
 [2026-01-30 02:03:01] (step=0002618) Train Loss mse: 0.0000, Train Loss ce: 0.0383, Train Steps/Sec: 0.22,
 [2026-01-30 02:03:05] (step=0002619) Train Loss mse: 0.0000, Train Loss ce: 0.0307, Train Steps/Sec: 0.22,
 [2026-01-30 02:03:10] (step=0002620) Train Loss mse: 0.0000, Train Loss ce: 0.0402, Train Steps/Sec: 0.22,
+base_dir is /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins_step2500
+Preparing Dataset vlm_gym_toy_maze_2d_celoss_no_mse_evalonce/vlm_gym_toy_maze_2d_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+ce_avg: 0.11742664873600006, mse_avg: 0.0
+base_dir is /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins_step3000
+Preparing Dataset vlm_gym_toy_maze_2d_celoss_no_mse_evalonce/vlm_gym_toy_maze_2d_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+ce_avg: 0.15227733552455902, mse_avg: 0.0
+base_dir is /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins_step3500
+Preparing Dataset vlm_gym_toy_maze_2d_celoss_no_mse_evalonce/vlm_gym_toy_maze_2d_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+ce_avg: 0.19604933261871338, mse_avg: 0.0
 [2026-01-30 02:03:14] (step=0002621) Train Loss mse: 0.0000, Train Loss ce: 0.0373, Train Steps/Sec: 0.23,
 [2026-01-30 02:03:19] (step=0002622) Train Loss mse: 0.0000, Train Loss ce: 0.0383, Train Steps/Sec: 0.22,
 [2026-01-30 02:03:23] (step=0002623) Train Loss mse: 0.0000, Train Loss ce: 0.0315, Train Steps/Sec: 0.26,
@@ -3453,20 +3460,6 @@ ce_avg: 0.15227733552455902, mse_avg: 0.0
 [2026-01-30 02:50:41] (step=0003237) Train Loss mse: 0.0000, Train Loss ce: 0.0404, Train Steps/Sec: 0.21,
 [2026-01-30 02:50:45] (step=0003238) Train Loss mse: 0.0000, Train Loss ce: 0.0343, Train Steps/Sec: 0.22,
 [2026-01-30 02:50:50] (step=0003239) Train Loss mse: 0.0000, Train Loss ce: 0.0265, Train Steps/Sec: 0.21,
-base_dir is /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins_step3500
-Preparing Dataset vlm_gym_toy_maze_2d_celoss_no_mse_evalonce/vlm_gym_toy_maze_2d_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-ce_avg: 0.19604933261871338, mse_avg: 0.0
-base_dir is /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins_step4000
-Preparing Dataset vlm_gym_toy_maze_2d_celoss_no_mse_evalonce/vlm_gym_toy_maze_2d_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-ce_avg: 0.19754347205162048, mse_avg: 0.0
 [2026-01-30 02:50:54] (step=0003240) Train Loss mse: 0.0000, Train Loss ce: 0.0303, Train Steps/Sec: 0.24,
 [2026-01-30 02:50:59] (step=0003241) Train Loss mse: 0.0000, Train Loss ce: 0.0399, Train Steps/Sec: 0.20,
 [2026-01-30 02:51:03] (step=0003242) Train Loss mse: 0.0000, Train Loss ce: 0.0342, Train Steps/Sec: 0.23,
@@ -3910,6 +3903,20 @@ ce_avg: 0.19754347205162048, mse_avg: 0.0
 [2026-01-30 03:25:05] (step=0003680) Train Loss mse: 0.0000, Train Loss ce: 0.0270, Train Steps/Sec: 0.22,
 [2026-01-30 03:25:09] (step=0003681) Train Loss mse: 0.0000, Train Loss ce: 0.0386, Train Steps/Sec: 0.22,
 [2026-01-30 03:25:14] (step=0003682) Train Loss mse: 0.0000, Train Loss ce: 0.0292, Train Steps/Sec: 0.20,
+base_dir is /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins_step4000
+Preparing Dataset vlm_gym_toy_maze_2d_celoss_no_mse_evalonce/vlm_gym_toy_maze_2d_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+ce_avg: 0.19754347205162048, mse_avg: 0.0
+base_dir is /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins_step4500
+Preparing Dataset vlm_gym_toy_maze_2d_celoss_no_mse_evalonce/vlm_gym_toy_maze_2d_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+ce_avg: 0.20177732408046722, mse_avg: 0.0
 [2026-01-30 03:25:18] (step=0003683) Train Loss mse: 0.0000, Train Loss ce: 0.0377, Train Steps/Sec: 0.24,
 [2026-01-30 03:25:23] (step=0003684) Train Loss mse: 0.0000, Train Loss ce: 0.0461, Train Steps/Sec: 0.21,
 [2026-01-30 03:25:28] (step=0003685) Train Loss mse: 0.0000, Train Loss ce: 0.0395, Train Steps/Sec: 0.21,
@@ -4889,20 +4896,6 @@ ce_avg: 0.19754347205162048, mse_avg: 0.0
 [2026-01-30 04:40:52] (step=0004659) Train Loss mse: 0.0000, Train Loss ce: 0.0359, Train Steps/Sec: 0.20,
 [2026-01-30 04:40:56] (step=0004660) Train Loss mse: 0.0000, Train Loss ce: 0.0324, Train Steps/Sec: 0.22,
 [2026-01-30 04:41:01] (step=0004661) Train Loss mse: 0.0000, Train Loss ce: 0.0409, Train Steps/Sec: 0.22,
-base_dir is /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins_step4500
-Preparing Dataset vlm_gym_toy_maze_2d_celoss_no_mse_evalonce/vlm_gym_toy_maze_2d_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-ce_avg: 0.20177732408046722, mse_avg: 0.0
-base_dir is /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins_step5000
-Preparing Dataset vlm_gym_toy_maze_2d_celoss_no_mse_evalonce/vlm_gym_toy_maze_2d_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
-ce_avg: 0.20619651675224304, mse_avg: 0.0
 [2026-01-30 04:41:05] (step=0004662) Train Loss mse: 0.0000, Train Loss ce: 0.0324, Train Steps/Sec: 0.21,
 [2026-01-30 04:41:10] (step=0004663) Train Loss mse: 0.0000, Train Loss ce: 0.0382, Train Steps/Sec: 0.22,
 [2026-01-30 04:41:15] (step=0004664) Train Loss mse: 0.0000, Train Loss ce: 0.0338, Train Steps/Sec: 0.21,
@@ -5245,4 +5238,11 @@ ce_avg: 0.20619651675224304, mse_avg: 0.0
 [2026-01-30 05:07:25] Saving checkpoint to /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/0005000.
 /opt/conda/lib/python3.11/site-packages/torch/distributed/fsdp/fully_sharded_data_parallel.py:690: FutureWarning: FSDP.state_dict_type() and FSDP.set_state_dict_type() are being deprecated. Please use APIs, get_state_dict() and set_state_dict(), which can support different parallelisms, FSDP1, FSDP2, DDP. API doc: https://pytorch.org/docs/stable/distributed.checkpoint.html#torch.distributed.checkpoint.state_dict.get_state_dict .Tutorial: https://pytorch.org/tutorials/recipes/distributed_checkpoint_recipe.html .
   warnings.warn(
-[2026-01-30 05:09:58] Done!
+[2026-01-30 05:09:58] Done!
+base_dir is /dev/shm/models/checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_toy_maze_2d_one_image_lr2e_5_ce_no_mse_ins_step5000
+Preparing Dataset vlm_gym_toy_maze_2d_celoss_no_mse_evalonce/vlm_gym_toy_maze_2d_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_toy_maze_2d_celoss_no_mse_evalonce'}]
+ce_avg: 0.20619651675224304, mse_avg: 0.0