Upload checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins/checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins

checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins/checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins/wandb/offline-run-20260125_192135-checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins-run0/files/output.log

@@ -1,3 +1,180 @@
+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_counting_mark_all_celoss_no_mse/vlm_gym_counting_mark_all_train
+base_dir is /dev/shm/models/checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins_step0
+Preparing Dataset vlm_gym_counting_mark_all_celoss_no_mse_evalonce/vlm_gym_counting_mark_all_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+ce_avg: 1.073127269744873, mse_avg: 0.0
+base_dir is /dev/shm/models/checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins_step500
+Preparing Dataset vlm_gym_counting_mark_all_celoss_no_mse_evalonce/vlm_gym_counting_mark_all_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+ce_avg: 0.5390675663948059, mse_avg: 0.0
+base_dir is /dev/shm/models/checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins_step1000
+Preparing Dataset vlm_gym_counting_mark_all_celoss_no_mse_evalonce/vlm_gym_counting_mark_all_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+ce_avg: 0.6019229888916016, 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/
@@ -1027,6 +1204,27 @@ wandb: For more information, check out the docs at: https://weave-docs.wandb.ai/
 [2026-01-25 20:16:01] (step=0001016) Train Loss mse: 0.0000, Train Loss ce: 0.5289, Train Steps/Sec: 0.28,
 [2026-01-25 20:16:04] (step=0001017) Train Loss mse: 0.0000, Train Loss ce: 0.5496, Train Steps/Sec: 0.30,
 [2026-01-25 20:16:06] (step=0001018) Train Loss mse: 0.0000, Train Loss ce: 0.5517, Train Steps/Sec: 0.39,
+base_dir is /dev/shm/models/checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins_step1500
+Preparing Dataset vlm_gym_counting_mark_all_celoss_no_mse_evalonce/vlm_gym_counting_mark_all_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+ce_avg: 0.6630723476409912, mse_avg: 0.0
+base_dir is /dev/shm/models/checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins_step2000
+Preparing Dataset vlm_gym_counting_mark_all_celoss_no_mse_evalonce/vlm_gym_counting_mark_all_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+ce_avg: 0.8126255869865417, mse_avg: 0.0
+base_dir is /dev/shm/models/checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins_step2500
+Preparing Dataset vlm_gym_counting_mark_all_celoss_no_mse_evalonce/vlm_gym_counting_mark_all_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+ce_avg: 0.9854414463043213, mse_avg: 0.0
 [2026-01-25 20:16:08] (step=0001019) Train Loss mse: 0.0000, Train Loss ce: 0.5016, Train Steps/Sec: 0.52,
 [2026-01-25 20:16:12] (step=0001020) Train Loss mse: 0.0000, Train Loss ce: 0.5413, Train Steps/Sec: 0.27,
 [2026-01-25 20:16:15] (step=0001021) Train Loss mse: 0.0000, Train Loss ce: 0.5106, Train Steps/Sec: 0.33,
@@ -1062,197 +1260,6 @@ wandb: For more information, check out the docs at: https://weave-docs.wandb.ai/
 [2026-01-25 20:17:46] (step=0001051) Train Loss mse: 0.0000, Train Loss ce: 0.5387, Train Steps/Sec: 0.37,
 [2026-01-25 20:17:49] (step=0001052) Train Loss mse: 0.0000, Train Loss ce: 0.5305, Train Steps/Sec: 0.27,
 [2026-01-25 20:17:52] (step=0001053) Train Loss mse: 0.0000, Train Loss ce: 0.5647, Train Steps/Sec: 0.33,
-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_counting_mark_all_celoss_no_mse/vlm_gym_counting_mark_all_train
-base_dir is /dev/shm/models/checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins_step0
-Preparing Dataset vlm_gym_counting_mark_all_celoss_no_mse_evalonce/vlm_gym_counting_mark_all_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
-ce_avg: 1.073127269744873, mse_avg: 0.0
-base_dir is /dev/shm/models/checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins_step500
-Preparing Dataset vlm_gym_counting_mark_all_celoss_no_mse_evalonce/vlm_gym_counting_mark_all_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
-ce_avg: 0.5390675663948059, mse_avg: 0.0
-base_dir is /dev/shm/models/checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins_step1500
-Preparing Dataset vlm_gym_counting_mark_all_celoss_no_mse_evalonce/vlm_gym_counting_mark_all_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
-ce_avg: 0.6630723476409912, mse_avg: 0.0
-base_dir is /dev/shm/models/checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins_step2000
-Preparing Dataset vlm_gym_counting_mark_all_celoss_no_mse_evalonce/vlm_gym_counting_mark_all_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
-ce_avg: 0.8126255869865417, mse_avg: 0.0
-base_dir is /dev/shm/models/checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins_step2500
-Preparing Dataset vlm_gym_counting_mark_all_celoss_no_mse_evalonce/vlm_gym_counting_mark_all_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
-ce_avg: 0.9854414463043213, mse_avg: 0.0
 [2026-01-25 20:17:55] (step=0001054) Train Loss mse: 0.0000, Train Loss ce: 0.5160, Train Steps/Sec: 0.34,
 [2026-01-25 20:17:59] (step=0001055) Train Loss mse: 0.0000, Train Loss ce: 0.5283, Train Steps/Sec: 0.28,
 [2026-01-25 20:18:02] (step=0001056) Train Loss mse: 0.0000, Train Loss ce: 0.5526, Train Steps/Sec: 0.35,
@@ -2692,20 +2699,6 @@ ce_avg: 0.9854414463043213, mse_avg: 0.0
 [2026-01-25 21:25:36] (step=0002490) Train Loss mse: 0.0000, Train Loss ce: 0.5307, Train Steps/Sec: 0.45,
 [2026-01-25 21:25:39] (step=0002491) Train Loss mse: 0.0000, Train Loss ce: 0.5212, Train Steps/Sec: 0.29,
 [2026-01-25 21:25:41] (step=0002492) Train Loss mse: 0.0000, Train Loss ce: 0.4945, Train Steps/Sec: 0.48,
-base_dir is /dev/shm/models/checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins_step3000
-Preparing Dataset vlm_gym_counting_mark_all_celoss_no_mse_evalonce/vlm_gym_counting_mark_all_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
-ce_avg: 0.9968664646148682, mse_avg: 0.0
-base_dir is /dev/shm/models/checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins_step3500
-Preparing Dataset vlm_gym_counting_mark_all_celoss_no_mse_evalonce/vlm_gym_counting_mark_all_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
-ce_avg: 0.9615826606750488, mse_avg: 0.0
 [2026-01-25 21:25:44] (step=0002493) Train Loss mse: 0.0000, Train Loss ce: 0.5583, Train Steps/Sec: 0.37,
 [2026-01-25 21:25:48] (step=0002494) Train Loss mse: 0.0000, Train Loss ce: 0.5176, Train Steps/Sec: 0.24,
 [2026-01-25 21:25:51] (step=0002495) Train Loss mse: 0.0000, Train Loss ce: 0.5266, Train Steps/Sec: 0.32,
@@ -2729,6 +2722,20 @@ ce_avg: 0.9615826606750488, mse_avg: 0.0
 [2026-01-25 21:26:49] (step=0002513) Train Loss mse: 0.0000, Train Loss ce: 0.5441, Train Steps/Sec: 0.51,
 [2026-01-25 21:26:52] (step=0002514) Train Loss mse: 0.0000, Train Loss ce: 0.5287, Train Steps/Sec: 0.32,
 [2026-01-25 21:26:54] (step=0002515) Train Loss mse: 0.0000, Train Loss ce: 0.5109, Train Steps/Sec: 0.49,
+base_dir is /dev/shm/models/checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins_step3000
+Preparing Dataset vlm_gym_counting_mark_all_celoss_no_mse_evalonce/vlm_gym_counting_mark_all_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+ce_avg: 0.9968664646148682, mse_avg: 0.0
+base_dir is /dev/shm/models/checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins_step3500
+Preparing Dataset vlm_gym_counting_mark_all_celoss_no_mse_evalonce/vlm_gym_counting_mark_all_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+ce_avg: 0.9615826606750488, mse_avg: 0.0
 [2026-01-25 21:26:57] (step=0002516) Train Loss mse: 0.0000, Train Loss ce: 0.5004, Train Steps/Sec: 0.44,
 [2026-01-25 21:26:59] (step=0002517) Train Loss mse: 0.0000, Train Loss ce: 0.5243, Train Steps/Sec: 0.47,
 [2026-01-25 21:27:01] (step=0002518) Train Loss mse: 0.0000, Train Loss ce: 0.4966, Train Steps/Sec: 0.46,
@@ -3731,6 +3738,19 @@ ce_avg: 0.9615826606750488, mse_avg: 0.0
 [2026-01-25 22:14:20] (step=0003515) Train Loss mse: 0.0000, Train Loss ce: 0.5115, Train Steps/Sec: 0.38,
 [2026-01-25 22:14:23] (step=0003516) Train Loss mse: 0.0000, Train Loss ce: 0.4846, Train Steps/Sec: 0.39,
 [2026-01-25 22:14:25] (step=0003517) Train Loss mse: 0.0000, Train Loss ce: 0.4703, Train Steps/Sec: 0.45,
+[2026-01-25 22:14:28] (step=0003518) Train Loss mse: 0.0000, Train Loss ce: 0.4743, Train Steps/Sec: 0.44,
+[2026-01-25 22:14:30] (step=0003519) Train Loss mse: 0.0000, Train Loss ce: 0.5065, Train Steps/Sec: 0.39,
+[2026-01-25 22:14:32] (step=0003520) Train Loss mse: 0.0000, Train Loss ce: 0.4738, Train Steps/Sec: 0.42,
+[2026-01-25 22:14:35] (step=0003521) Train Loss mse: 0.0000, Train Loss ce: 0.4873, Train Steps/Sec: 0.35,
+[2026-01-25 22:14:39] (step=0003522) Train Loss mse: 0.0000, Train Loss ce: 0.5458, Train Steps/Sec: 0.28,
+[2026-01-25 22:14:42] (step=0003523) Train Loss mse: 0.0000, Train Loss ce: 0.5189, Train Steps/Sec: 0.32,
+[2026-01-25 22:14:44] (step=0003524) Train Loss mse: 0.0000, Train Loss ce: 0.4763, Train Steps/Sec: 0.54,
+[2026-01-25 22:14:46] (step=0003525) Train Loss mse: 0.0000, Train Loss ce: 0.4845, Train Steps/Sec: 0.41,
+[2026-01-25 22:14:49] (step=0003526) Train Loss mse: 0.0000, Train Loss ce: 0.4723, Train Steps/Sec: 0.41,
+[2026-01-25 22:14:51] (step=0003527) Train Loss mse: 0.0000, Train Loss ce: 0.5612, Train Steps/Sec: 0.37,
+[2026-01-25 22:14:55] (step=0003528) Train Loss mse: 0.0000, Train Loss ce: 0.5088, Train Steps/Sec: 0.30,
+[2026-01-25 22:14:58] (step=0003529) Train Loss mse: 0.0000, Train Loss ce: 0.4618, Train Steps/Sec: 0.36,
+[2026-01-25 22:15:01] (step=0003530) Train Loss mse: 0.0000, Train Loss ce: 0.5275, Train Steps/Sec: 0.30,
 base_dir is /dev/shm/models/checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins_step4000
 Preparing Dataset vlm_gym_counting_mark_all_celoss_no_mse_evalonce/vlm_gym_counting_mark_all_val
 [eval debug] first 3 batch fingerprints:
@@ -3745,38 +3765,6 @@ Preparing Dataset vlm_gym_counting_mark_all_celoss_no_mse_evalonce/vlm_gym_count
   fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
   fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
 ce_avg: 0.8653228282928467, mse_avg: 0.0
-[2026-01-25 22:14:28] (step=0003518) Train Loss mse: 0.0000, Train Loss ce: 0.4743, Train Steps/Sec: 0.44,
-[2026-01-25 22:14:30] (step=0003519) Train Loss mse: 0.0000, Train Loss ce: 0.5065, Train Steps/Sec: 0.39,
-[2026-01-25 22:14:32] (step=0003520) Train Loss mse: 0.0000, Train Loss ce: 0.4738, Train Steps/Sec: 0.42,
-[2026-01-25 22:14:35] (step=0003521) Train Loss mse: 0.0000, Train Loss ce: 0.4873, Train Steps/Sec: 0.35,
-[2026-01-25 22:14:39] (step=0003522) Train Loss mse: 0.0000, Train Loss ce: 0.5458, Train Steps/Sec: 0.28,
-[2026-01-25 22:14:42] (step=0003523) Train Loss mse: 0.0000, Train Loss ce: 0.5189, Train Steps/Sec: 0.32,
-[2026-01-25 22:14:44] (step=0003524) Train Loss mse: 0.0000, Train Loss ce: 0.4763, Train Steps/Sec: 0.54,
-[2026-01-25 22:14:46] (step=0003525) Train Loss mse: 0.0000, Train Loss ce: 0.4845, Train Steps/Sec: 0.41,
-[2026-01-25 22:14:49] (step=0003526) Train Loss mse: 0.0000, Train Loss ce: 0.4723, Train Steps/Sec: 0.41,
-[2026-01-25 22:14:51] (step=0003527) Train Loss mse: 0.0000, Train Loss ce: 0.5612, Train Steps/Sec: 0.37,
-[2026-01-25 22:14:55] (step=0003528) Train Loss mse: 0.0000, Train Loss ce: 0.5088, Train Steps/Sec: 0.30,
-[2026-01-25 22:14:58] (step=0003529) Train Loss mse: 0.0000, Train Loss ce: 0.4618, Train Steps/Sec: 0.36,
-[2026-01-25 22:15:01] (step=0003530) Train Loss mse: 0.0000, Train Loss ce: 0.5275, Train Steps/Sec: 0.30,
-[2026-01-25 22:15:03] (step=0003531) Train Loss mse: 0.0000, Train Loss ce: 0.5246, Train Steps/Sec: 0.43,
-[2026-01-25 22:15:06] (step=0003532) Train Loss mse: 0.0000, Train Loss ce: 0.4948, Train Steps/Sec: 0.34,
-[2026-01-25 22:15:08] (step=0003533) Train Loss mse: 0.0000, Train Loss ce: 0.4725, Train Steps/Sec: 0.61,
-[2026-01-25 22:15:10] (step=0003534) Train Loss mse: 0.0000, Train Loss ce: 0.4692, Train Steps/Sec: 0.52,
-[2026-01-25 22:15:13] (step=0003535) Train Loss mse: 0.0000, Train Loss ce: 0.5502, Train Steps/Sec: 0.30,
-[2026-01-25 22:15:17] (step=0003536) Train Loss mse: 0.0000, Train Loss ce: 0.5425, Train Steps/Sec: 0.27,
-[2026-01-25 22:15:19] (step=0003537) Train Loss mse: 0.0000, Train Loss ce: 0.4866, Train Steps/Sec: 0.39,
-[2026-01-25 22:15:24] (step=0003538) Train Loss mse: 0.0000, Train Loss ce: 0.5266, Train Steps/Sec: 0.23,
-[2026-01-25 22:15:26] (step=0003539) Train Loss mse: 0.0000, Train Loss ce: 0.5203, Train Steps/Sec: 0.47,
-[2026-01-25 22:15:28] (step=0003540) Train Loss mse: 0.0000, Train Loss ce: 0.4538, Train Steps/Sec: 0.49,
-[2026-01-25 22:15:31] (step=0003541) Train Loss mse: 0.0000, Train Loss ce: 0.4993, Train Steps/Sec: 0.32,
-[2026-01-25 22:15:34] (step=0003542) Train Loss mse: 0.0000, Train Loss ce: 0.5085, Train Steps/Sec: 0.34,
-[2026-01-25 22:15:36] (step=0003543) Train Loss mse: 0.0000, Train Loss ce: 0.4943, Train Steps/Sec: 0.49,
-[2026-01-25 22:15:38] (step=0003544) Train Loss mse: 0.0000, Train Loss ce: 0.5110, Train Steps/Sec: 0.44,
-[2026-01-25 22:15:41] (step=0003545) Train Loss mse: 0.0000, Train Loss ce: 0.4802, Train Steps/Sec: 0.32,
-[2026-01-25 22:15:44] (step=0003546) Train Loss mse: 0.0000, Train Loss ce: 0.4802, Train Steps/Sec: 0.39,
-[2026-01-25 22:15:46] (step=0003547) Train Loss mse: 0.0000, Train Loss ce: 0.4930, Train Steps/Sec: 0.39,
-[2026-01-25 22:15:50] (step=0003548) Train Loss mse: 0.0000, Train Loss ce: 0.5221, Train Steps/Sec: 0.31,
-[2026-01-25 22:15:52] (step=0003549) Train Loss mse: 0.0000, Train Loss ce: 0.4530, Train Steps/Sec: 0.41,
 [2026-01-25 22:15:54] (step=0003550) Train Loss mse: 0.0000, Train Loss ce: 0.4626, Train Steps/Sec: 0.57,
 [2026-01-25 22:15:56] (step=0003551) Train Loss mse: 0.0000, Train Loss ce: 0.4705, Train Steps/Sec: 0.49,
 [2026-01-25 22:15:58] (step=0003552) Train Loss mse: 0.0000, Train Loss ce: 0.4905, Train Steps/Sec: 0.44,
@@ -5158,13 +5146,6 @@ ce_avg: 0.8653228282928467, mse_avg: 0.0
 [2026-01-25 23:21:47] (step=0004928) Train Loss mse: 0.0000, Train Loss ce: 0.4813, Train Steps/Sec: 0.40,
 [2026-01-25 23:21:51] (step=0004929) Train Loss mse: 0.0000, Train Loss ce: 0.5108, Train Steps/Sec: 0.28,
 [2026-01-25 23:21:54] (step=0004930) Train Loss mse: 0.0000, Train Loss ce: 0.5030, Train Steps/Sec: 0.32,
-base_dir is /dev/shm/models/checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins_step5000
-Preparing Dataset vlm_gym_counting_mark_all_celoss_no_mse_evalonce/vlm_gym_counting_mark_all_val
-[eval debug] first 3 batch fingerprints:
-  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
-  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
-  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
-ce_avg: 0.847433865070343, mse_avg: 0.0
 [2026-01-25 23:21:57] (step=0004931) Train Loss mse: 0.0000, Train Loss ce: 0.5101, Train Steps/Sec: 0.29,
 [2026-01-25 23:22:01] (step=0004932) Train Loss mse: 0.0000, Train Loss ce: 0.5305, Train Steps/Sec: 0.25,
 [2026-01-25 23:22:04] (step=0004933) Train Loss mse: 0.0000, Train Loss ce: 0.4626, Train Steps/Sec: 0.36,
@@ -5238,4 +5219,11 @@ ce_avg: 0.847433865070343, mse_avg: 0.0
 [2026-01-25 23:25:21] Saving checkpoint to /dev/shm/models/checkpoints_vlm_gym_counting_mark_all_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(
-[2026-01-25 23:27:59] Done!
+[2026-01-25 23:27:59] Done!
+base_dir is /dev/shm/models/checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins/eval_used_rows, step_tag is checkpoints_vlm_gym_counting_mark_all_one_image_lr2e_5_ce_no_mse_ins_step5000
+Preparing Dataset vlm_gym_counting_mark_all_celoss_no_mse_evalonce/vlm_gym_counting_mark_all_val
+[eval debug] first 3 batch fingerprints:
+  fp[0]: [{'data_indexes': [0], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+  fp[1]: [{'data_indexes': [8], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+  fp[2]: [{'data_indexes': [16], 'worker_id': 0, 'dataset_name': 'vlm_gym_counting_mark_all_celoss_no_mse_evalonce'}]
+ce_avg: 0.847433865070343, mse_avg: 0.0