sync run artifacts: vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/best.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:302d99bc96a1694bb6cfe20dfe8938acb2b195884b0f149a6c7bcd3b709141e5
+size 7397286257

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/git-info.txt

@@ -0,0 +1,2 @@
+branch: main
+commit: 857f73bfc0cf1648567212e23a081f468aab48c2

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_0_log.err

@@ -0,0 +1,45 @@
+/scratch-shared/dcanez/.cache/uv5/virtualenvs/vd/lib/python3.13/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+/scratch-shared/dcanez/.cache/uv5/virtualenvs/vd/lib/python3.13/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+[rank0]:[W514 09:36:45.928252362 ProcessGroupNCCL.cpp:5138] Guessing device ID based on global rank. This can cause a hang if rank to GPU mapping is heterogeneous. You can specify device_id in init_process_group()
+wandb: Currently logged in as: dgcnz (uvjepa) to https://api.wandb.ai. Use `wandb login --relogin` to force relogin
+wandb: setting up run rxtoljxd
+wandb: Tracking run with wandb version 0.23.1
+wandb: Run data is saved locally in /scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/wandb/run-20260514_093658-rxtoljxd
+wandb: Run `wandb offline` to turn off syncing.
+wandb: Syncing run c003_vitl_k16_simple_cross_factorized_stable
+wandb: ⭐️ View project at https://wandb.ai/uvjepa/vjepa_ujepaside
+wandb: 🚀 View run at https://wandb.ai/uvjepa/vjepa_ujepaside/runs/rxtoljxd
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/utils.py:796: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/scratch-shared/dcanez/.cache/uv5/virtualenvs/vd/lib/python3.13/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/scratch-shared/dcanez/.cache/uv5/virtualenvs/vd/lib/python3.13/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/scratch-shared/dcanez/.cache/uv5/virtualenvs/vd/lib/python3.13/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/scratch-shared/dcanez/.cache/uv5/virtualenvs/vd/lib/python3.13/site-packages/torch/distributed/c10d_logger.py:83: UserWarning: barrier(): using the device under current context. You can specify `device_id` in `init_process_group` to mute this warning.
+  return func(*args, **kwargs)
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+submitit WARNING (2026-05-14 13:38:23,806) - Bypassing signal SIGCONT
+[2026-05-14T13:38:23.807] error: *** JOB 22743106 ON gcn75 CANCELLED AT 2026-05-14T13:38:23 DUE to SIGNAL Terminated ***
+srun: Job step aborted: Waiting up to 32 seconds for job step to finish.
+[2026-05-14T13:38:23.808] error: *** STEP 22743106.0 ON gcn75 CANCELLED AT 2026-05-14T13:38:23 DUE to SIGNAL Terminated ***
+submitit WARNING (2026-05-14 13:38:23,813) - Bypassing signal SIGTERM
+[2026-05-14T13:38:57.539] error: namespace_p_join: open failed for /slurm/22743106/.ns: No such file or directory
+[2026-05-14T13:38:57.540] error: namespace_g_join(JobId=22743106 SLUID=s8FXKRJP5N0000): No such file or directory
+[2026-05-14T13:38:59.198] error: Failed to send MESSAGE_TASK_EXIT: Connection refused
+[2026-05-14T13:39:03.204] error: namespace_p_join: open failed for /slurm/22743106/.ns: No such file or directory
+[2026-05-14T13:39:03.205] error: namespace_g_join(JobId=22743106 SLUID=s8FXKRJP5N0000): No such file or directory
+[2026-05-14T13:39:03.345] error: namespace_p_join: open failed for /slurm/22743106/.ns: No such file or directory
+[2026-05-14T13:39:03.346] error: namespace_g_join(JobId=22743106 SLUID=s8FXKRJP5N0000): No such file or directory
+[2026-05-14T13:39:03.385] error: namespace_p_join: open failed for /slurm/22743106/.ns: No such file or directory
+[2026-05-14T13:39:03.385] error: namespace_g_join(JobId=22743106 SLUID=s8FXKRJP5N0000): No such file or directory

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_10_log.err

@@ -0,0 +1,9 @@
+/scratch-shared/dcanez/.cache/uv5/virtualenvs/vd/lib/python3.13/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+[rank10]:[W514 09:35:52.955169794 ProcessGroupNCCL.cpp:5138] Guessing device ID based on global rank. This can cause a hang if rank to GPU mapping is heterogeneous. You can specify device_id in init_process_group()
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/utils.py:796: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+submitit WARNING (2026-05-14 13:38:23,806) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,852) - Bypassing signal SIGTERM

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_10_log.out

@@ -0,0 +1,288 @@
+submitit INFO (2026-05-14 09:28:14,530) - Starting with JobEnvironment(job_id=22743106, hostname=gcn81.local.snellius.surf.nl, local_rank=2(4), node=2(4), global_rank=10(16))
+submitit INFO (2026-05-14 09:28:14,530) - Loading pickle: /scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 64,
+                'crop_size': 224,
+                'dataset_fpcs': [16, 16],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/scratch-shared/dcanez/data/kinetics/k400/train.csv',
+                                '/scratch-shared/dcanez/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 4,
+                'num_workers': 10,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'stage': [   {   'dest': 'kinetics_240',
+                                 'format': 'targz_parts',
+                                 'src': '/scratch-shared/dcanez/data/kinetics/k400/tars_240/'},
+                             {   'dest': 'ssv2',
+                                 'format': 'multipart_tar',
+                                 'src': '/scratch-nvme/ml-datasets/something-something-v2/'}],
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.15, 0.15],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.7, 0.7],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '180G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': False,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/train.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/val.csv',
+                                                          'eval_videos_per_class': 25,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 100},
+                                            'preset': 'ucf101'},
+                                        {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/train_coarse10.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 5,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ujepaside'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'model_name': 'ujepaside_large_patch14_capi_lvd1689m',
+                 'pred_depth': 6,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 12,
+                 'predictor': 'v2_cross',
+                 'st_causal': False,
+                 'st_drop_path': 0.2,
+                 'st_flex_enable': False,
+                 'st_layer_scale_init': 1e-05,
+                 'st_num_slots': 16,
+                 'st_side_block_type': 'factorized',
+                 'st_slots_causal_within_frame': False,
+                 'target_kind': 'frozen_2d',
+                 'target_type': 'vit_large_patch14_capi.lvd1689m',
+                 'temporal_spacing': 1.0,
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_mask_tokens': True,
+                 'use_rope': True,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 4,
+    'optimization': {   'clip_grad': 3.0,
+                        'ema': [0.99925, 0.99925],
+                        'epochs': 100,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0005,
+                        'start_lr': 0.0001,
+                        'warmup': 10,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 4}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-05-14 09:32:44][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-05-14 09:32:44][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-05-14 09:32:44][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=eno
+[INFO    ][2026-05-14 09:32:47][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 10/16, tasks_per_node=4
+[INFO    ][2026-05-14 09:32:47][root                ][stage_datasets           ] [local_rank 2/4] Staging kinetics_240 (targz_parts)
+[INFO    ][2026-05-14 09:32:47][root                ][_stage_targz_parts       ] [rank 2] Extracting 26/103 tar.gz parts to /scratch-node/dcanez.22743106/kinetics_240
+[INFO    ][2026-05-14 09:33:01][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 2/26 parts
+[INFO    ][2026-05-14 09:33:15][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 4/26 parts
+[INFO    ][2026-05-14 09:33:29][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 6/26 parts
+[INFO    ][2026-05-14 09:33:44][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 8/26 parts
+[INFO    ][2026-05-14 09:33:58][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 10/26 parts
+[INFO    ][2026-05-14 09:34:12][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 12/26 parts
+[INFO    ][2026-05-14 09:34:25][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 14/26 parts
+[INFO    ][2026-05-14 09:34:40][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 16/26 parts
+[INFO    ][2026-05-14 09:34:54][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 18/26 parts
+[INFO    ][2026-05-14 09:35:09][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 20/26 parts
+[INFO    ][2026-05-14 09:35:23][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 22/26 parts
+[INFO    ][2026-05-14 09:35:37][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 24/26 parts
+[INFO    ][2026-05-14 09:35:52][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 26/26 parts
+[INFO    ][2026-05-14 09:35:52][root                ][stage_datasets           ] [local_rank 2/4] Staging ssv2 (multipart_tar)
+[INFO    ][2026-05-14 09:36:49][app.vjepa.train     ][main                     ] Staged datasets: ['/scratch-node/dcanez.22743106/kinetics_240/train.csv', '/scratch-node/dcanez.22743106/ssv2/train.csv']
+[INFO    ][2026-05-14 09:43:14][experiments.stmodels.vision_transformers_v3][vit_large_patch14_capi_lvd1689m] Loading pretrained weights for vit_large_patch14_capi_lvd1689m from https://dl.fbaipublicfiles.com/capi/capi_vitl14_lvd.pth
+[INFO    ][2026-05-14 09:43:28][root                ][_build_frozen_2d_target  ] Loaded frozen_2d target from timm: vit_large_patch14_capi.lvd1689m
+[INFO    ][2026-05-14 09:43:28][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): UJEPAside(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (rope): CAPI2DRoPE()
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (rope_impl): CAPI2DRoPE()
+        (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=False)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+          (rope_impl): CAPI2DRoPE()
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=False)
+    (st_blocks): ModuleList(
+      (0-23): 24 x FactorizedSlotSideBlock(
+        (cross_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (self_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (mlp): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (st_rope): CAPI3DRoPE()
+  )
+)
+[INFO    ][2026-05-14 09:43:28][root                ][init_video_model         ] UJEPAsidePredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-5): 6 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-05-14 09:43:28][root                ][init_video_model         ] MultiSeqWrapper(
+  (backbone): Frozen2DTargetWrapper(
+    (backbone): Eva(
+      (patch_embed): PatchEmbed(
+        (proj): Conv2d(3, 1024, kernel_size=(14, 14), stride=(14, 14))
+        (norm): Identity()
+      )
+      (pos_drop): Dropout(p=0.0, inplace=False)
+      (norm_pre): Identity()
+      (blocks): ModuleList(
+        (0-23): 24 x EvaBlock(
+          (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (attn): EvaAttention(
+            (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+            (q_norm): Identity()
+            (k_norm): Identity()
+            (attn_drop): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (proj): Linear(in_features=1024, out_features=1024, bias=True)
+            (proj_drop): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path1): Identity()
+          (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (mlp): Mlp(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+            (act): GELU(approximate='none')
+            (drop1): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+            (drop2): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path2): Identity()
+        )
+      )
+      (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+      (fc_norm): Identity()
+      (head_drop): Dropout(p=0.0, inplace=False)
+      (head): Identity()
+      (rope): _CapiPatchRoPE()
+    )
+  )
+)
+[INFO    ][2026-05-14 09:43:28][root                ][init_video_model         ] Encoder number of parameters: 403136512
+[INFO    ][2026-05-14 09:43:28][root                ][init_video_model         ] Predictor number of parameters: 11416192
+[INFO    ][2026-05-14 09:43:28][root                ][init_video_model         ] Target encoder number of parameters: 0
+[INFO    ][2026-05-14 09:43:41][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-05-14 09:43:41][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 10 / 16
+[INFO    ][2026-05-14 09:43:41][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-05-14 09:43:41][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/399
+[INFO    ][2026-05-14 09:43:41][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 10)...
+[INFO    ][2026-05-14 09:43:41][app.vjepa.train     ][main                     ] Initializing loader...
+submitit WARNING (2026-05-14 13:38:23,806) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,852) - Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGCONT

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_11_log.err

@@ -0,0 +1,9 @@
+/scratch-shared/dcanez/.cache/uv5/virtualenvs/vd/lib/python3.13/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+[rank11]:[W514 09:35:46.057936025 ProcessGroupNCCL.cpp:5138] Guessing device ID based on global rank. This can cause a hang if rank to GPU mapping is heterogeneous. You can specify device_id in init_process_group()
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/utils.py:796: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+submitit WARNING (2026-05-14 13:38:23,806) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,846) - Bypassing signal SIGTERM

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_11_log.out

@@ -0,0 +1,288 @@
+submitit INFO (2026-05-14 09:28:14,530) - Starting with JobEnvironment(job_id=22743106, hostname=gcn81.local.snellius.surf.nl, local_rank=3(4), node=2(4), global_rank=11(16))
+submitit INFO (2026-05-14 09:28:14,530) - Loading pickle: /scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 64,
+                'crop_size': 224,
+                'dataset_fpcs': [16, 16],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/scratch-shared/dcanez/data/kinetics/k400/train.csv',
+                                '/scratch-shared/dcanez/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 4,
+                'num_workers': 10,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'stage': [   {   'dest': 'kinetics_240',
+                                 'format': 'targz_parts',
+                                 'src': '/scratch-shared/dcanez/data/kinetics/k400/tars_240/'},
+                             {   'dest': 'ssv2',
+                                 'format': 'multipart_tar',
+                                 'src': '/scratch-nvme/ml-datasets/something-something-v2/'}],
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.15, 0.15],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.7, 0.7],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '180G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': False,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/train.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/val.csv',
+                                                          'eval_videos_per_class': 25,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 100},
+                                            'preset': 'ucf101'},
+                                        {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/train_coarse10.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 5,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ujepaside'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'model_name': 'ujepaside_large_patch14_capi_lvd1689m',
+                 'pred_depth': 6,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 12,
+                 'predictor': 'v2_cross',
+                 'st_causal': False,
+                 'st_drop_path': 0.2,
+                 'st_flex_enable': False,
+                 'st_layer_scale_init': 1e-05,
+                 'st_num_slots': 16,
+                 'st_side_block_type': 'factorized',
+                 'st_slots_causal_within_frame': False,
+                 'target_kind': 'frozen_2d',
+                 'target_type': 'vit_large_patch14_capi.lvd1689m',
+                 'temporal_spacing': 1.0,
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_mask_tokens': True,
+                 'use_rope': True,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 4,
+    'optimization': {   'clip_grad': 3.0,
+                        'ema': [0.99925, 0.99925],
+                        'epochs': 100,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0005,
+                        'start_lr': 0.0001,
+                        'warmup': 10,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 4}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-05-14 09:32:44][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-05-14 09:32:44][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-05-14 09:32:44][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=eno
+[INFO    ][2026-05-14 09:32:46][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 11/16, tasks_per_node=4
+[INFO    ][2026-05-14 09:32:46][root                ][stage_datasets           ] [local_rank 3/4] Staging kinetics_240 (targz_parts)
+[INFO    ][2026-05-14 09:32:46][root                ][_stage_targz_parts       ] [rank 3] Extracting 25/103 tar.gz parts to /scratch-node/dcanez.22743106/kinetics_240
+[INFO    ][2026-05-14 09:33:00][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 2/25 parts
+[INFO    ][2026-05-14 09:33:14][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 4/25 parts
+[INFO    ][2026-05-14 09:33:28][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 6/25 parts
+[INFO    ][2026-05-14 09:33:43][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 8/25 parts
+[INFO    ][2026-05-14 09:33:57][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 10/25 parts
+[INFO    ][2026-05-14 09:34:11][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 12/25 parts
+[INFO    ][2026-05-14 09:34:25][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 14/25 parts
+[INFO    ][2026-05-14 09:34:40][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 16/25 parts
+[INFO    ][2026-05-14 09:34:54][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 18/25 parts
+[INFO    ][2026-05-14 09:35:08][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 20/25 parts
+[INFO    ][2026-05-14 09:35:23][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 22/25 parts
+[INFO    ][2026-05-14 09:35:37][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 24/25 parts
+[INFO    ][2026-05-14 09:35:45][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 25/25 parts
+[INFO    ][2026-05-14 09:35:45][root                ][stage_datasets           ] [local_rank 3/4] Staging ssv2 (multipart_tar)
+[INFO    ][2026-05-14 09:36:49][app.vjepa.train     ][main                     ] Staged datasets: ['/scratch-node/dcanez.22743106/kinetics_240/train.csv', '/scratch-node/dcanez.22743106/ssv2/train.csv']
+[INFO    ][2026-05-14 09:43:14][experiments.stmodels.vision_transformers_v3][vit_large_patch14_capi_lvd1689m] Loading pretrained weights for vit_large_patch14_capi_lvd1689m from https://dl.fbaipublicfiles.com/capi/capi_vitl14_lvd.pth
+[INFO    ][2026-05-14 09:43:28][root                ][_build_frozen_2d_target  ] Loaded frozen_2d target from timm: vit_large_patch14_capi.lvd1689m
+[INFO    ][2026-05-14 09:43:28][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): UJEPAside(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (rope): CAPI2DRoPE()
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (rope_impl): CAPI2DRoPE()
+        (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=False)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+          (rope_impl): CAPI2DRoPE()
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=False)
+    (st_blocks): ModuleList(
+      (0-23): 24 x FactorizedSlotSideBlock(
+        (cross_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (self_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (mlp): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (st_rope): CAPI3DRoPE()
+  )
+)
+[INFO    ][2026-05-14 09:43:28][root                ][init_video_model         ] UJEPAsidePredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-5): 6 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-05-14 09:43:28][root                ][init_video_model         ] MultiSeqWrapper(
+  (backbone): Frozen2DTargetWrapper(
+    (backbone): Eva(
+      (patch_embed): PatchEmbed(
+        (proj): Conv2d(3, 1024, kernel_size=(14, 14), stride=(14, 14))
+        (norm): Identity()
+      )
+      (pos_drop): Dropout(p=0.0, inplace=False)
+      (norm_pre): Identity()
+      (blocks): ModuleList(
+        (0-23): 24 x EvaBlock(
+          (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (attn): EvaAttention(
+            (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+            (q_norm): Identity()
+            (k_norm): Identity()
+            (attn_drop): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (proj): Linear(in_features=1024, out_features=1024, bias=True)
+            (proj_drop): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path1): Identity()
+          (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (mlp): Mlp(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+            (act): GELU(approximate='none')
+            (drop1): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+            (drop2): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path2): Identity()
+        )
+      )
+      (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+      (fc_norm): Identity()
+      (head_drop): Dropout(p=0.0, inplace=False)
+      (head): Identity()
+      (rope): _CapiPatchRoPE()
+    )
+  )
+)
+[INFO    ][2026-05-14 09:43:28][root                ][init_video_model         ] Encoder number of parameters: 403136512
+[INFO    ][2026-05-14 09:43:28][root                ][init_video_model         ] Predictor number of parameters: 11416192
+[INFO    ][2026-05-14 09:43:28][root                ][init_video_model         ] Target encoder number of parameters: 0
+[INFO    ][2026-05-14 09:43:41][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-05-14 09:43:41][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 11 / 16
+[INFO    ][2026-05-14 09:43:41][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-05-14 09:43:41][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/399
+[INFO    ][2026-05-14 09:43:41][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 11)...
+[INFO    ][2026-05-14 09:43:41][app.vjepa.train     ][main                     ] Initializing loader...
+submitit WARNING (2026-05-14 13:38:23,806) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,846) - Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGCONT

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_12_log.err

@@ -0,0 +1,18 @@
+/scratch-shared/dcanez/.cache/uv5/virtualenvs/vd/lib/python3.13/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+[rank12]:[W514 09:36:45.287417238 ProcessGroupNCCL.cpp:5138] Guessing device ID based on global rank. This can cause a hang if rank to GPU mapping is heterogeneous. You can specify device_id in init_process_group()
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/utils.py:796: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+submitit WARNING (2026-05-14 13:38:23,806) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,838) - Bypassing signal SIGTERM
+[2026-05-14T13:38:56.223] error: namespace_p_join: open failed for /slurm/22743106/.ns: No such file or directory
+[2026-05-14T13:38:56.223] error: namespace_g_join(JobId=22743106 SLUID=s8FXKRJP5N0000): No such file or directory
+[2026-05-14T13:38:57.868] error: Failed to send MESSAGE_TASK_EXIT: Connection refused
+[2026-05-14T13:38:59.637] error: namespace_p_join: open failed for /slurm/22743106/.ns: No such file or directory
+[2026-05-14T13:38:59.637] error: namespace_g_join(JobId=22743106 SLUID=s8FXKRJP5N0000): No such file or directory
+[2026-05-14T13:38:59.776] error: namespace_p_join: open failed for /slurm/22743106/.ns: No such file or directory
+[2026-05-14T13:38:59.776] error: namespace_g_join(JobId=22743106 SLUID=s8FXKRJP5N0000): No such file or directory
+[2026-05-14T13:38:59.815] error: namespace_p_join: open failed for /slurm/22743106/.ns: No such file or directory
+[2026-05-14T13:38:59.815] error: namespace_g_join(JobId=22743106 SLUID=s8FXKRJP5N0000): No such file or directory

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_12_log.out

@@ -0,0 +1,292 @@
+submitit INFO (2026-05-14 09:28:14,530) - Starting with JobEnvironment(job_id=22743106, hostname=gcn83.local.snellius.surf.nl, local_rank=0(4), node=3(4), global_rank=12(16))
+submitit INFO (2026-05-14 09:28:14,530) - Loading pickle: /scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 64,
+                'crop_size': 224,
+                'dataset_fpcs': [16, 16],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/scratch-shared/dcanez/data/kinetics/k400/train.csv',
+                                '/scratch-shared/dcanez/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 4,
+                'num_workers': 10,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'stage': [   {   'dest': 'kinetics_240',
+                                 'format': 'targz_parts',
+                                 'src': '/scratch-shared/dcanez/data/kinetics/k400/tars_240/'},
+                             {   'dest': 'ssv2',
+                                 'format': 'multipart_tar',
+                                 'src': '/scratch-nvme/ml-datasets/something-something-v2/'}],
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.15, 0.15],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.7, 0.7],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '180G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': False,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/train.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/val.csv',
+                                                          'eval_videos_per_class': 25,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 100},
+                                            'preset': 'ucf101'},
+                                        {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/train_coarse10.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 5,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ujepaside'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'model_name': 'ujepaside_large_patch14_capi_lvd1689m',
+                 'pred_depth': 6,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 12,
+                 'predictor': 'v2_cross',
+                 'st_causal': False,
+                 'st_drop_path': 0.2,
+                 'st_flex_enable': False,
+                 'st_layer_scale_init': 1e-05,
+                 'st_num_slots': 16,
+                 'st_side_block_type': 'factorized',
+                 'st_slots_causal_within_frame': False,
+                 'target_kind': 'frozen_2d',
+                 'target_type': 'vit_large_patch14_capi.lvd1689m',
+                 'temporal_spacing': 1.0,
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_mask_tokens': True,
+                 'use_rope': True,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 4,
+    'optimization': {   'clip_grad': 3.0,
+                        'ema': [0.99925, 0.99925],
+                        'epochs': 100,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0005,
+                        'start_lr': 0.0001,
+                        'warmup': 10,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 4}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-05-14 09:32:42][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-05-14 09:32:42][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-05-14 09:32:42][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=eno
+[INFO    ][2026-05-14 09:32:44][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 12/16, tasks_per_node=4
+[INFO    ][2026-05-14 09:32:44][root                ][stage_datasets           ] [local_rank 0/4] Staging kinetics_240 (targz_parts)
+[INFO    ][2026-05-14 09:32:44][root                ][_stage_targz_parts       ] [rank 0] Extracting 26/103 tar.gz parts to /scratch-node/dcanez.22743106/kinetics_240
+[INFO    ][2026-05-14 09:32:59][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 2/26 parts
+[INFO    ][2026-05-14 09:33:13][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 4/26 parts
+[INFO    ][2026-05-14 09:33:27][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 6/26 parts
+[INFO    ][2026-05-14 09:33:41][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 8/26 parts
+[INFO    ][2026-05-14 09:33:55][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 10/26 parts
+[INFO    ][2026-05-14 09:34:10][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 12/26 parts
+[INFO    ][2026-05-14 09:34:24][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 14/26 parts
+[INFO    ][2026-05-14 09:34:38][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 16/26 parts
+[INFO    ][2026-05-14 09:34:52][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 18/26 parts
+[INFO    ][2026-05-14 09:35:07][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 20/26 parts
+[INFO    ][2026-05-14 09:35:21][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 22/26 parts
+[INFO    ][2026-05-14 09:35:36][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 24/26 parts
+[INFO    ][2026-05-14 09:35:50][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 26/26 parts
+[INFO    ][2026-05-14 09:35:50][root                ][stage_datasets           ] [local_rank 0/4] Staging ssv2 (multipart_tar)
+[INFO    ][2026-05-14 09:35:50][root                ][_stage_multipart_tar     ] [rank 0] Extracting multipart tar (2 files) to /scratch-node/dcanez.22743106/ssv2
+[INFO    ][2026-05-14 09:36:48][root                ][stage_datasets           ] Data staging completed in 243.6s (4.1min)
+[INFO    ][2026-05-14 09:36:49][root                ][_rewrite_csv             ] Wrote local CSV: /scratch-node/dcanez.22743106/kinetics_240/train.csv (239789 entries)
+[INFO    ][2026-05-14 09:36:49][root                ][_rewrite_csv             ] Wrote local CSV: /scratch-node/dcanez.22743106/ssv2/train.csv (168913 entries)
+[INFO    ][2026-05-14 09:36:49][app.vjepa.train     ][main                     ] Staged datasets: ['/scratch-node/dcanez.22743106/kinetics_240/train.csv', '/scratch-node/dcanez.22743106/ssv2/train.csv']
+[INFO    ][2026-05-14 09:43:13][experiments.stmodels.vision_transformers_v3][vit_large_patch14_capi_lvd1689m] Loading pretrained weights for vit_large_patch14_capi_lvd1689m from https://dl.fbaipublicfiles.com/capi/capi_vitl14_lvd.pth
+[INFO    ][2026-05-14 09:43:23][root                ][_build_frozen_2d_target  ] Loaded frozen_2d target from timm: vit_large_patch14_capi.lvd1689m
+[INFO    ][2026-05-14 09:43:24][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): UJEPAside(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (rope): CAPI2DRoPE()
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (rope_impl): CAPI2DRoPE()
+        (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=False)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+          (rope_impl): CAPI2DRoPE()
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=False)
+    (st_blocks): ModuleList(
+      (0-23): 24 x FactorizedSlotSideBlock(
+        (cross_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (self_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (mlp): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (st_rope): CAPI3DRoPE()
+  )
+)
+[INFO    ][2026-05-14 09:43:24][root                ][init_video_model         ] UJEPAsidePredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-5): 6 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-05-14 09:43:24][root                ][init_video_model         ] MultiSeqWrapper(
+  (backbone): Frozen2DTargetWrapper(
+    (backbone): Eva(
+      (patch_embed): PatchEmbed(
+        (proj): Conv2d(3, 1024, kernel_size=(14, 14), stride=(14, 14))
+        (norm): Identity()
+      )
+      (pos_drop): Dropout(p=0.0, inplace=False)
+      (norm_pre): Identity()
+      (blocks): ModuleList(
+        (0-23): 24 x EvaBlock(
+          (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (attn): EvaAttention(
+            (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+            (q_norm): Identity()
+            (k_norm): Identity()
+            (attn_drop): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (proj): Linear(in_features=1024, out_features=1024, bias=True)
+            (proj_drop): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path1): Identity()
+          (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (mlp): Mlp(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+            (act): GELU(approximate='none')
+            (drop1): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+            (drop2): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path2): Identity()
+        )
+      )
+      (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+      (fc_norm): Identity()
+      (head_drop): Dropout(p=0.0, inplace=False)
+      (head): Identity()
+      (rope): _CapiPatchRoPE()
+    )
+  )
+)
+[INFO    ][2026-05-14 09:43:24][root                ][init_video_model         ] Encoder number of parameters: 403136512
+[INFO    ][2026-05-14 09:43:24][root                ][init_video_model         ] Predictor number of parameters: 11416192
+[INFO    ][2026-05-14 09:43:24][root                ][init_video_model         ] Target encoder number of parameters: 0
+[INFO    ][2026-05-14 09:43:36][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-05-14 09:43:36][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 12 / 16
+[INFO    ][2026-05-14 09:43:36][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-05-14 09:43:36][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/399
+[INFO    ][2026-05-14 09:43:36][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 12)...
+[INFO    ][2026-05-14 09:43:41][app.vjepa.train     ][main                     ] Initializing loader...
+submitit WARNING (2026-05-14 13:38:23,806) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,838) - Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGCONT

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_13_log.err

@@ -0,0 +1,9 @@
+/scratch-shared/dcanez/.cache/uv5/virtualenvs/vd/lib/python3.13/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+[rank13]:[W514 09:35:49.225966127 ProcessGroupNCCL.cpp:5138] Guessing device ID based on global rank. This can cause a hang if rank to GPU mapping is heterogeneous. You can specify device_id in init_process_group()
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/utils.py:796: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+submitit WARNING (2026-05-14 13:38:23,806) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,823) - Bypassing signal SIGTERM

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_13_log.out

@@ -0,0 +1,288 @@
+submitit INFO (2026-05-14 09:28:14,530) - Starting with JobEnvironment(job_id=22743106, hostname=gcn83.local.snellius.surf.nl, local_rank=1(4), node=3(4), global_rank=13(16))
+submitit INFO (2026-05-14 09:28:14,530) - Loading pickle: /scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 64,
+                'crop_size': 224,
+                'dataset_fpcs': [16, 16],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/scratch-shared/dcanez/data/kinetics/k400/train.csv',
+                                '/scratch-shared/dcanez/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 4,
+                'num_workers': 10,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'stage': [   {   'dest': 'kinetics_240',
+                                 'format': 'targz_parts',
+                                 'src': '/scratch-shared/dcanez/data/kinetics/k400/tars_240/'},
+                             {   'dest': 'ssv2',
+                                 'format': 'multipart_tar',
+                                 'src': '/scratch-nvme/ml-datasets/something-something-v2/'}],
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.15, 0.15],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.7, 0.7],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '180G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': False,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/train.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/val.csv',
+                                                          'eval_videos_per_class': 25,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 100},
+                                            'preset': 'ucf101'},
+                                        {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/train_coarse10.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 5,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ujepaside'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'model_name': 'ujepaside_large_patch14_capi_lvd1689m',
+                 'pred_depth': 6,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 12,
+                 'predictor': 'v2_cross',
+                 'st_causal': False,
+                 'st_drop_path': 0.2,
+                 'st_flex_enable': False,
+                 'st_layer_scale_init': 1e-05,
+                 'st_num_slots': 16,
+                 'st_side_block_type': 'factorized',
+                 'st_slots_causal_within_frame': False,
+                 'target_kind': 'frozen_2d',
+                 'target_type': 'vit_large_patch14_capi.lvd1689m',
+                 'temporal_spacing': 1.0,
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_mask_tokens': True,
+                 'use_rope': True,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 4,
+    'optimization': {   'clip_grad': 3.0,
+                        'ema': [0.99925, 0.99925],
+                        'epochs': 100,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0005,
+                        'start_lr': 0.0001,
+                        'warmup': 10,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 4}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-05-14 09:32:42][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-05-14 09:32:42][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-05-14 09:32:42][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=eno
+[INFO    ][2026-05-14 09:32:44][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 13/16, tasks_per_node=4
+[INFO    ][2026-05-14 09:32:44][root                ][stage_datasets           ] [local_rank 1/4] Staging kinetics_240 (targz_parts)
+[INFO    ][2026-05-14 09:32:44][root                ][_stage_targz_parts       ] [rank 1] Extracting 26/103 tar.gz parts to /scratch-node/dcanez.22743106/kinetics_240
+[INFO    ][2026-05-14 09:32:57][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 2/26 parts
+[INFO    ][2026-05-14 09:33:11][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 4/26 parts
+[INFO    ][2026-05-14 09:33:25][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 6/26 parts
+[INFO    ][2026-05-14 09:33:39][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 8/26 parts
+[INFO    ][2026-05-14 09:33:53][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 10/26 parts
+[INFO    ][2026-05-14 09:34:08][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 12/26 parts
+[INFO    ][2026-05-14 09:34:22][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 14/26 parts
+[INFO    ][2026-05-14 09:34:36][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 16/26 parts
+[INFO    ][2026-05-14 09:34:51][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 18/26 parts
+[INFO    ][2026-05-14 09:35:06][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 20/26 parts
+[INFO    ][2026-05-14 09:35:20][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 22/26 parts
+[INFO    ][2026-05-14 09:35:34][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 24/26 parts
+[INFO    ][2026-05-14 09:35:49][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 26/26 parts
+[INFO    ][2026-05-14 09:35:49][root                ][stage_datasets           ] [local_rank 1/4] Staging ssv2 (multipart_tar)
+[INFO    ][2026-05-14 09:36:49][app.vjepa.train     ][main                     ] Staged datasets: ['/scratch-node/dcanez.22743106/kinetics_240/train.csv', '/scratch-node/dcanez.22743106/ssv2/train.csv']
+[INFO    ][2026-05-14 09:43:14][experiments.stmodels.vision_transformers_v3][vit_large_patch14_capi_lvd1689m] Loading pretrained weights for vit_large_patch14_capi_lvd1689m from https://dl.fbaipublicfiles.com/capi/capi_vitl14_lvd.pth
+[INFO    ][2026-05-14 09:43:23][root                ][_build_frozen_2d_target  ] Loaded frozen_2d target from timm: vit_large_patch14_capi.lvd1689m
+[INFO    ][2026-05-14 09:43:23][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): UJEPAside(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (rope): CAPI2DRoPE()
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (rope_impl): CAPI2DRoPE()
+        (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=False)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+          (rope_impl): CAPI2DRoPE()
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=False)
+    (st_blocks): ModuleList(
+      (0-23): 24 x FactorizedSlotSideBlock(
+        (cross_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (self_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (mlp): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (st_rope): CAPI3DRoPE()
+  )
+)
+[INFO    ][2026-05-14 09:43:24][root                ][init_video_model         ] UJEPAsidePredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-5): 6 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-05-14 09:43:24][root                ][init_video_model         ] MultiSeqWrapper(
+  (backbone): Frozen2DTargetWrapper(
+    (backbone): Eva(
+      (patch_embed): PatchEmbed(
+        (proj): Conv2d(3, 1024, kernel_size=(14, 14), stride=(14, 14))
+        (norm): Identity()
+      )
+      (pos_drop): Dropout(p=0.0, inplace=False)
+      (norm_pre): Identity()
+      (blocks): ModuleList(
+        (0-23): 24 x EvaBlock(
+          (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (attn): EvaAttention(
+            (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+            (q_norm): Identity()
+            (k_norm): Identity()
+            (attn_drop): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (proj): Linear(in_features=1024, out_features=1024, bias=True)
+            (proj_drop): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path1): Identity()
+          (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (mlp): Mlp(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+            (act): GELU(approximate='none')
+            (drop1): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+            (drop2): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path2): Identity()
+        )
+      )
+      (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+      (fc_norm): Identity()
+      (head_drop): Dropout(p=0.0, inplace=False)
+      (head): Identity()
+      (rope): _CapiPatchRoPE()
+    )
+  )
+)
+[INFO    ][2026-05-14 09:43:24][root                ][init_video_model         ] Encoder number of parameters: 403136512
+[INFO    ][2026-05-14 09:43:24][root                ][init_video_model         ] Predictor number of parameters: 11416192
+[INFO    ][2026-05-14 09:43:24][root                ][init_video_model         ] Target encoder number of parameters: 0
+[INFO    ][2026-05-14 09:43:36][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-05-14 09:43:36][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 13 / 16
+[INFO    ][2026-05-14 09:43:36][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-05-14 09:43:36][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/399
+[INFO    ][2026-05-14 09:43:36][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 13)...
+[INFO    ][2026-05-14 09:43:41][app.vjepa.train     ][main                     ] Initializing loader...
+submitit WARNING (2026-05-14 13:38:23,806) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,823) - Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGCONT

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_14_log.err

@@ -0,0 +1,9 @@
+/scratch-shared/dcanez/.cache/uv5/virtualenvs/vd/lib/python3.13/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+[rank14]:[W514 09:35:52.647814712 ProcessGroupNCCL.cpp:5138] Guessing device ID based on global rank. This can cause a hang if rank to GPU mapping is heterogeneous. You can specify device_id in init_process_group()
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/utils.py:796: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+submitit WARNING (2026-05-14 13:38:23,806) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,826) - Bypassing signal SIGTERM

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_14_log.out

@@ -0,0 +1,288 @@
+submitit INFO (2026-05-14 09:28:14,530) - Starting with JobEnvironment(job_id=22743106, hostname=gcn83.local.snellius.surf.nl, local_rank=2(4), node=3(4), global_rank=14(16))
+submitit INFO (2026-05-14 09:28:14,530) - Loading pickle: /scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 64,
+                'crop_size': 224,
+                'dataset_fpcs': [16, 16],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/scratch-shared/dcanez/data/kinetics/k400/train.csv',
+                                '/scratch-shared/dcanez/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 4,
+                'num_workers': 10,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'stage': [   {   'dest': 'kinetics_240',
+                                 'format': 'targz_parts',
+                                 'src': '/scratch-shared/dcanez/data/kinetics/k400/tars_240/'},
+                             {   'dest': 'ssv2',
+                                 'format': 'multipart_tar',
+                                 'src': '/scratch-nvme/ml-datasets/something-something-v2/'}],
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.15, 0.15],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.7, 0.7],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '180G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': False,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/train.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/val.csv',
+                                                          'eval_videos_per_class': 25,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 100},
+                                            'preset': 'ucf101'},
+                                        {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/train_coarse10.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 5,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ujepaside'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'model_name': 'ujepaside_large_patch14_capi_lvd1689m',
+                 'pred_depth': 6,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 12,
+                 'predictor': 'v2_cross',
+                 'st_causal': False,
+                 'st_drop_path': 0.2,
+                 'st_flex_enable': False,
+                 'st_layer_scale_init': 1e-05,
+                 'st_num_slots': 16,
+                 'st_side_block_type': 'factorized',
+                 'st_slots_causal_within_frame': False,
+                 'target_kind': 'frozen_2d',
+                 'target_type': 'vit_large_patch14_capi.lvd1689m',
+                 'temporal_spacing': 1.0,
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_mask_tokens': True,
+                 'use_rope': True,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 4,
+    'optimization': {   'clip_grad': 3.0,
+                        'ema': [0.99925, 0.99925],
+                        'epochs': 100,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0005,
+                        'start_lr': 0.0001,
+                        'warmup': 10,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 4}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-05-14 09:32:42][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-05-14 09:32:42][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-05-14 09:32:42][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=eno
+[INFO    ][2026-05-14 09:32:46][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 14/16, tasks_per_node=4
+[INFO    ][2026-05-14 09:32:46][root                ][stage_datasets           ] [local_rank 2/4] Staging kinetics_240 (targz_parts)
+[INFO    ][2026-05-14 09:32:46][root                ][_stage_targz_parts       ] [rank 2] Extracting 26/103 tar.gz parts to /scratch-node/dcanez.22743106/kinetics_240
+[INFO    ][2026-05-14 09:33:00][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 2/26 parts
+[INFO    ][2026-05-14 09:33:14][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 4/26 parts
+[INFO    ][2026-05-14 09:33:28][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 6/26 parts
+[INFO    ][2026-05-14 09:33:43][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 8/26 parts
+[INFO    ][2026-05-14 09:33:57][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 10/26 parts
+[INFO    ][2026-05-14 09:34:11][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 12/26 parts
+[INFO    ][2026-05-14 09:34:25][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 14/26 parts
+[INFO    ][2026-05-14 09:34:40][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 16/26 parts
+[INFO    ][2026-05-14 09:34:54][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 18/26 parts
+[INFO    ][2026-05-14 09:35:08][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 20/26 parts
+[INFO    ][2026-05-14 09:35:23][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 22/26 parts
+[INFO    ][2026-05-14 09:35:37][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 24/26 parts
+[INFO    ][2026-05-14 09:35:52][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 26/26 parts
+[INFO    ][2026-05-14 09:35:52][root                ][stage_datasets           ] [local_rank 2/4] Staging ssv2 (multipart_tar)
+[INFO    ][2026-05-14 09:36:49][app.vjepa.train     ][main                     ] Staged datasets: ['/scratch-node/dcanez.22743106/kinetics_240/train.csv', '/scratch-node/dcanez.22743106/ssv2/train.csv']
+[INFO    ][2026-05-14 09:43:14][experiments.stmodels.vision_transformers_v3][vit_large_patch14_capi_lvd1689m] Loading pretrained weights for vit_large_patch14_capi_lvd1689m from https://dl.fbaipublicfiles.com/capi/capi_vitl14_lvd.pth
+[INFO    ][2026-05-14 09:43:23][root                ][_build_frozen_2d_target  ] Loaded frozen_2d target from timm: vit_large_patch14_capi.lvd1689m
+[INFO    ][2026-05-14 09:43:24][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): UJEPAside(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (rope): CAPI2DRoPE()
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (rope_impl): CAPI2DRoPE()
+        (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=False)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+          (rope_impl): CAPI2DRoPE()
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=False)
+    (st_blocks): ModuleList(
+      (0-23): 24 x FactorizedSlotSideBlock(
+        (cross_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (self_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (mlp): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (st_rope): CAPI3DRoPE()
+  )
+)
+[INFO    ][2026-05-14 09:43:24][root                ][init_video_model         ] UJEPAsidePredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-5): 6 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-05-14 09:43:24][root                ][init_video_model         ] MultiSeqWrapper(
+  (backbone): Frozen2DTargetWrapper(
+    (backbone): Eva(
+      (patch_embed): PatchEmbed(
+        (proj): Conv2d(3, 1024, kernel_size=(14, 14), stride=(14, 14))
+        (norm): Identity()
+      )
+      (pos_drop): Dropout(p=0.0, inplace=False)
+      (norm_pre): Identity()
+      (blocks): ModuleList(
+        (0-23): 24 x EvaBlock(
+          (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (attn): EvaAttention(
+            (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+            (q_norm): Identity()
+            (k_norm): Identity()
+            (attn_drop): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (proj): Linear(in_features=1024, out_features=1024, bias=True)
+            (proj_drop): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path1): Identity()
+          (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (mlp): Mlp(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+            (act): GELU(approximate='none')
+            (drop1): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+            (drop2): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path2): Identity()
+        )
+      )
+      (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+      (fc_norm): Identity()
+      (head_drop): Dropout(p=0.0, inplace=False)
+      (head): Identity()
+      (rope): _CapiPatchRoPE()
+    )
+  )
+)
+[INFO    ][2026-05-14 09:43:24][root                ][init_video_model         ] Encoder number of parameters: 403136512
+[INFO    ][2026-05-14 09:43:24][root                ][init_video_model         ] Predictor number of parameters: 11416192
+[INFO    ][2026-05-14 09:43:24][root                ][init_video_model         ] Target encoder number of parameters: 0
+[INFO    ][2026-05-14 09:43:36][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-05-14 09:43:36][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 14 / 16
+[INFO    ][2026-05-14 09:43:36][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-05-14 09:43:36][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/399
+[INFO    ][2026-05-14 09:43:36][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 14)...
+[INFO    ][2026-05-14 09:43:41][app.vjepa.train     ][main                     ] Initializing loader...
+submitit WARNING (2026-05-14 13:38:23,806) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,826) - Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGCONT

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_15_log.err

@@ -0,0 +1,9 @@
+/scratch-shared/dcanez/.cache/uv5/virtualenvs/vd/lib/python3.13/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+[rank15]:[W514 09:35:46.486928483 ProcessGroupNCCL.cpp:5138] Guessing device ID based on global rank. This can cause a hang if rank to GPU mapping is heterogeneous. You can specify device_id in init_process_group()
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/utils.py:796: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+submitit WARNING (2026-05-14 13:38:23,806) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,831) - Bypassing signal SIGTERM

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_15_log.out

@@ -0,0 +1,288 @@
+submitit INFO (2026-05-14 09:28:14,530) - Starting with JobEnvironment(job_id=22743106, hostname=gcn83.local.snellius.surf.nl, local_rank=3(4), node=3(4), global_rank=15(16))
+submitit INFO (2026-05-14 09:28:14,530) - Loading pickle: /scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 64,
+                'crop_size': 224,
+                'dataset_fpcs': [16, 16],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/scratch-shared/dcanez/data/kinetics/k400/train.csv',
+                                '/scratch-shared/dcanez/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 4,
+                'num_workers': 10,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'stage': [   {   'dest': 'kinetics_240',
+                                 'format': 'targz_parts',
+                                 'src': '/scratch-shared/dcanez/data/kinetics/k400/tars_240/'},
+                             {   'dest': 'ssv2',
+                                 'format': 'multipart_tar',
+                                 'src': '/scratch-nvme/ml-datasets/something-something-v2/'}],
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.15, 0.15],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.7, 0.7],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '180G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': False,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/train.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/val.csv',
+                                                          'eval_videos_per_class': 25,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 100},
+                                            'preset': 'ucf101'},
+                                        {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/train_coarse10.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 5,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ujepaside'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'model_name': 'ujepaside_large_patch14_capi_lvd1689m',
+                 'pred_depth': 6,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 12,
+                 'predictor': 'v2_cross',
+                 'st_causal': False,
+                 'st_drop_path': 0.2,
+                 'st_flex_enable': False,
+                 'st_layer_scale_init': 1e-05,
+                 'st_num_slots': 16,
+                 'st_side_block_type': 'factorized',
+                 'st_slots_causal_within_frame': False,
+                 'target_kind': 'frozen_2d',
+                 'target_type': 'vit_large_patch14_capi.lvd1689m',
+                 'temporal_spacing': 1.0,
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_mask_tokens': True,
+                 'use_rope': True,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 4,
+    'optimization': {   'clip_grad': 3.0,
+                        'ema': [0.99925, 0.99925],
+                        'epochs': 100,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0005,
+                        'start_lr': 0.0001,
+                        'warmup': 10,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 4}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-05-14 09:32:42][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-05-14 09:32:42][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-05-14 09:32:42][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=eno
+[INFO    ][2026-05-14 09:32:45][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 15/16, tasks_per_node=4
+[INFO    ][2026-05-14 09:32:45][root                ][stage_datasets           ] [local_rank 3/4] Staging kinetics_240 (targz_parts)
+[INFO    ][2026-05-14 09:32:45][root                ][_stage_targz_parts       ] [rank 3] Extracting 25/103 tar.gz parts to /scratch-node/dcanez.22743106/kinetics_240
+[INFO    ][2026-05-14 09:32:59][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 2/25 parts
+[INFO    ][2026-05-14 09:33:14][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 4/25 parts
+[INFO    ][2026-05-14 09:33:28][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 6/25 parts
+[INFO    ][2026-05-14 09:33:42][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 8/25 parts
+[INFO    ][2026-05-14 09:33:57][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 10/25 parts
+[INFO    ][2026-05-14 09:34:11][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 12/25 parts
+[INFO    ][2026-05-14 09:34:26][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 14/25 parts
+[INFO    ][2026-05-14 09:34:40][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 16/25 parts
+[INFO    ][2026-05-14 09:34:54][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 18/25 parts
+[INFO    ][2026-05-14 09:35:09][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 20/25 parts
+[INFO    ][2026-05-14 09:35:23][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 22/25 parts
+[INFO    ][2026-05-14 09:35:37][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 24/25 parts
+[INFO    ][2026-05-14 09:35:45][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 25/25 parts
+[INFO    ][2026-05-14 09:35:45][root                ][stage_datasets           ] [local_rank 3/4] Staging ssv2 (multipart_tar)
+[INFO    ][2026-05-14 09:36:49][app.vjepa.train     ][main                     ] Staged datasets: ['/scratch-node/dcanez.22743106/kinetics_240/train.csv', '/scratch-node/dcanez.22743106/ssv2/train.csv']
+[INFO    ][2026-05-14 09:43:14][experiments.stmodels.vision_transformers_v3][vit_large_patch14_capi_lvd1689m] Loading pretrained weights for vit_large_patch14_capi_lvd1689m from https://dl.fbaipublicfiles.com/capi/capi_vitl14_lvd.pth
+[INFO    ][2026-05-14 09:43:23][root                ][_build_frozen_2d_target  ] Loaded frozen_2d target from timm: vit_large_patch14_capi.lvd1689m
+[INFO    ][2026-05-14 09:43:23][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): UJEPAside(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (rope): CAPI2DRoPE()
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (rope_impl): CAPI2DRoPE()
+        (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=False)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+          (rope_impl): CAPI2DRoPE()
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=False)
+    (st_blocks): ModuleList(
+      (0-23): 24 x FactorizedSlotSideBlock(
+        (cross_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (self_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (mlp): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (st_rope): CAPI3DRoPE()
+  )
+)
+[INFO    ][2026-05-14 09:43:23][root                ][init_video_model         ] UJEPAsidePredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-5): 6 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-05-14 09:43:24][root                ][init_video_model         ] MultiSeqWrapper(
+  (backbone): Frozen2DTargetWrapper(
+    (backbone): Eva(
+      (patch_embed): PatchEmbed(
+        (proj): Conv2d(3, 1024, kernel_size=(14, 14), stride=(14, 14))
+        (norm): Identity()
+      )
+      (pos_drop): Dropout(p=0.0, inplace=False)
+      (norm_pre): Identity()
+      (blocks): ModuleList(
+        (0-23): 24 x EvaBlock(
+          (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (attn): EvaAttention(
+            (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+            (q_norm): Identity()
+            (k_norm): Identity()
+            (attn_drop): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (proj): Linear(in_features=1024, out_features=1024, bias=True)
+            (proj_drop): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path1): Identity()
+          (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (mlp): Mlp(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+            (act): GELU(approximate='none')
+            (drop1): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+            (drop2): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path2): Identity()
+        )
+      )
+      (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+      (fc_norm): Identity()
+      (head_drop): Dropout(p=0.0, inplace=False)
+      (head): Identity()
+      (rope): _CapiPatchRoPE()
+    )
+  )
+)
+[INFO    ][2026-05-14 09:43:24][root                ][init_video_model         ] Encoder number of parameters: 403136512
+[INFO    ][2026-05-14 09:43:24][root                ][init_video_model         ] Predictor number of parameters: 11416192
+[INFO    ][2026-05-14 09:43:24][root                ][init_video_model         ] Target encoder number of parameters: 0
+[INFO    ][2026-05-14 09:43:36][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-05-14 09:43:36][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 15 / 16
+[INFO    ][2026-05-14 09:43:36][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-05-14 09:43:36][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/399
+[INFO    ][2026-05-14 09:43:36][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 15)...
+[INFO    ][2026-05-14 09:43:41][app.vjepa.train     ][main                     ] Initializing loader...
+submitit WARNING (2026-05-14 13:38:23,806) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,831) - Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGCONT

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_1_log.err

@@ -0,0 +1,15 @@
+/scratch-shared/dcanez/.cache/uv5/virtualenvs/vd/lib/python3.13/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+[rank1]:[W514 09:35:35.129726343 ProcessGroupNCCL.cpp:5138] Guessing device ID based on global rank. This can cause a hang if rank to GPU mapping is heterogeneous. You can specify device_id in init_process_group()
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/utils.py:796: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+submitit WARNING (2026-05-14 13:38:23,806) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,907) - Bypassing signal SIGTERM

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_1_log.out

@@ -0,0 +1,288 @@
+submitit INFO (2026-05-14 09:28:14,530) - Starting with JobEnvironment(job_id=22743106, hostname=gcn75.local.snellius.surf.nl, local_rank=1(4), node=0(4), global_rank=1(16))
+submitit INFO (2026-05-14 09:28:14,530) - Loading pickle: /scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 64,
+                'crop_size': 224,
+                'dataset_fpcs': [16, 16],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/scratch-shared/dcanez/data/kinetics/k400/train.csv',
+                                '/scratch-shared/dcanez/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 4,
+                'num_workers': 10,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'stage': [   {   'dest': 'kinetics_240',
+                                 'format': 'targz_parts',
+                                 'src': '/scratch-shared/dcanez/data/kinetics/k400/tars_240/'},
+                             {   'dest': 'ssv2',
+                                 'format': 'multipart_tar',
+                                 'src': '/scratch-nvme/ml-datasets/something-something-v2/'}],
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.15, 0.15],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.7, 0.7],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '180G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': False,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/train.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/val.csv',
+                                                          'eval_videos_per_class': 25,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 100},
+                                            'preset': 'ucf101'},
+                                        {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/train_coarse10.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 5,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ujepaside'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'model_name': 'ujepaside_large_patch14_capi_lvd1689m',
+                 'pred_depth': 6,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 12,
+                 'predictor': 'v2_cross',
+                 'st_causal': False,
+                 'st_drop_path': 0.2,
+                 'st_flex_enable': False,
+                 'st_layer_scale_init': 1e-05,
+                 'st_num_slots': 16,
+                 'st_side_block_type': 'factorized',
+                 'st_slots_causal_within_frame': False,
+                 'target_kind': 'frozen_2d',
+                 'target_type': 'vit_large_patch14_capi.lvd1689m',
+                 'temporal_spacing': 1.0,
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_mask_tokens': True,
+                 'use_rope': True,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 4,
+    'optimization': {   'clip_grad': 3.0,
+                        'ema': [0.99925, 0.99925],
+                        'epochs': 100,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0005,
+                        'start_lr': 0.0001,
+                        'warmup': 10,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 4}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-05-14 09:32:27][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-05-14 09:32:27][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-05-14 09:32:27][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=eno
+[INFO    ][2026-05-14 09:32:30][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 1/16, tasks_per_node=4
+[INFO    ][2026-05-14 09:32:30][root                ][stage_datasets           ] [local_rank 1/4] Staging kinetics_240 (targz_parts)
+[INFO    ][2026-05-14 09:32:30][root                ][_stage_targz_parts       ] [rank 1] Extracting 26/103 tar.gz parts to /scratch-node/dcanez.22743106/kinetics_240
+[INFO    ][2026-05-14 09:32:43][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 2/26 parts
+[INFO    ][2026-05-14 09:32:56][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 4/26 parts
+[INFO    ][2026-05-14 09:33:11][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 6/26 parts
+[INFO    ][2026-05-14 09:33:25][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 8/26 parts
+[INFO    ][2026-05-14 09:33:39][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 10/26 parts
+[INFO    ][2026-05-14 09:33:54][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 12/26 parts
+[INFO    ][2026-05-14 09:34:07][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 14/26 parts
+[INFO    ][2026-05-14 09:34:22][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 16/26 parts
+[INFO    ][2026-05-14 09:34:36][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 18/26 parts
+[INFO    ][2026-05-14 09:34:51][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 20/26 parts
+[INFO    ][2026-05-14 09:35:06][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 22/26 parts
+[INFO    ][2026-05-14 09:35:20][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 24/26 parts
+[INFO    ][2026-05-14 09:35:35][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 26/26 parts
+[INFO    ][2026-05-14 09:35:35][root                ][stage_datasets           ] [local_rank 1/4] Staging ssv2 (multipart_tar)
+[INFO    ][2026-05-14 09:36:49][app.vjepa.train     ][main                     ] Staged datasets: ['/scratch-node/dcanez.22743106/kinetics_240/train.csv', '/scratch-node/dcanez.22743106/ssv2/train.csv']
+[INFO    ][2026-05-14 09:43:14][experiments.stmodels.vision_transformers_v3][vit_large_patch14_capi_lvd1689m] Loading pretrained weights for vit_large_patch14_capi_lvd1689m from https://dl.fbaipublicfiles.com/capi/capi_vitl14_lvd.pth
+[INFO    ][2026-05-14 09:43:23][root                ][_build_frozen_2d_target  ] Loaded frozen_2d target from timm: vit_large_patch14_capi.lvd1689m
+[INFO    ][2026-05-14 09:43:23][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): UJEPAside(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (rope): CAPI2DRoPE()
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (rope_impl): CAPI2DRoPE()
+        (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=False)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+          (rope_impl): CAPI2DRoPE()
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=False)
+    (st_blocks): ModuleList(
+      (0-23): 24 x FactorizedSlotSideBlock(
+        (cross_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (self_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (mlp): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (st_rope): CAPI3DRoPE()
+  )
+)
+[INFO    ][2026-05-14 09:43:23][root                ][init_video_model         ] UJEPAsidePredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-5): 6 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-05-14 09:43:23][root                ][init_video_model         ] MultiSeqWrapper(
+  (backbone): Frozen2DTargetWrapper(
+    (backbone): Eva(
+      (patch_embed): PatchEmbed(
+        (proj): Conv2d(3, 1024, kernel_size=(14, 14), stride=(14, 14))
+        (norm): Identity()
+      )
+      (pos_drop): Dropout(p=0.0, inplace=False)
+      (norm_pre): Identity()
+      (blocks): ModuleList(
+        (0-23): 24 x EvaBlock(
+          (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (attn): EvaAttention(
+            (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+            (q_norm): Identity()
+            (k_norm): Identity()
+            (attn_drop): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (proj): Linear(in_features=1024, out_features=1024, bias=True)
+            (proj_drop): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path1): Identity()
+          (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (mlp): Mlp(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+            (act): GELU(approximate='none')
+            (drop1): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+            (drop2): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path2): Identity()
+        )
+      )
+      (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+      (fc_norm): Identity()
+      (head_drop): Dropout(p=0.0, inplace=False)
+      (head): Identity()
+      (rope): _CapiPatchRoPE()
+    )
+  )
+)
+[INFO    ][2026-05-14 09:43:23][root                ][init_video_model         ] Encoder number of parameters: 403136512
+[INFO    ][2026-05-14 09:43:23][root                ][init_video_model         ] Predictor number of parameters: 11416192
+[INFO    ][2026-05-14 09:43:23][root                ][init_video_model         ] Target encoder number of parameters: 0
+[INFO    ][2026-05-14 09:43:39][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-05-14 09:43:39][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 1 / 16
+[INFO    ][2026-05-14 09:43:39][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-05-14 09:43:39][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/399
+[INFO    ][2026-05-14 09:43:40][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 1)...
+[INFO    ][2026-05-14 09:43:41][app.vjepa.train     ][main                     ] Initializing loader...
+submitit WARNING (2026-05-14 13:38:23,806) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,907) - Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGCONT

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_2_log.err

@@ -0,0 +1,15 @@
+/scratch-shared/dcanez/.cache/uv5/virtualenvs/vd/lib/python3.13/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+[rank2]:[W514 09:35:36.297380774 ProcessGroupNCCL.cpp:5138] Guessing device ID based on global rank. This can cause a hang if rank to GPU mapping is heterogeneous. You can specify device_id in init_process_group()
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/utils.py:796: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+submitit WARNING (2026-05-14 13:38:23,807) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,937) - Bypassing signal SIGTERM

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_2_log.out

@@ -0,0 +1,288 @@
+submitit INFO (2026-05-14 09:28:14,530) - Starting with JobEnvironment(job_id=22743106, hostname=gcn75.local.snellius.surf.nl, local_rank=2(4), node=0(4), global_rank=2(16))
+submitit INFO (2026-05-14 09:28:14,530) - Loading pickle: /scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 64,
+                'crop_size': 224,
+                'dataset_fpcs': [16, 16],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/scratch-shared/dcanez/data/kinetics/k400/train.csv',
+                                '/scratch-shared/dcanez/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 4,
+                'num_workers': 10,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'stage': [   {   'dest': 'kinetics_240',
+                                 'format': 'targz_parts',
+                                 'src': '/scratch-shared/dcanez/data/kinetics/k400/tars_240/'},
+                             {   'dest': 'ssv2',
+                                 'format': 'multipart_tar',
+                                 'src': '/scratch-nvme/ml-datasets/something-something-v2/'}],
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.15, 0.15],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.7, 0.7],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '180G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': False,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/train.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/val.csv',
+                                                          'eval_videos_per_class': 25,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 100},
+                                            'preset': 'ucf101'},
+                                        {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/train_coarse10.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 5,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ujepaside'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'model_name': 'ujepaside_large_patch14_capi_lvd1689m',
+                 'pred_depth': 6,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 12,
+                 'predictor': 'v2_cross',
+                 'st_causal': False,
+                 'st_drop_path': 0.2,
+                 'st_flex_enable': False,
+                 'st_layer_scale_init': 1e-05,
+                 'st_num_slots': 16,
+                 'st_side_block_type': 'factorized',
+                 'st_slots_causal_within_frame': False,
+                 'target_kind': 'frozen_2d',
+                 'target_type': 'vit_large_patch14_capi.lvd1689m',
+                 'temporal_spacing': 1.0,
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_mask_tokens': True,
+                 'use_rope': True,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 4,
+    'optimization': {   'clip_grad': 3.0,
+                        'ema': [0.99925, 0.99925],
+                        'epochs': 100,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0005,
+                        'start_lr': 0.0001,
+                        'warmup': 10,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 4}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-05-14 09:32:27][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-05-14 09:32:27][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-05-14 09:32:27][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=eno
+[INFO    ][2026-05-14 09:32:28][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 2/16, tasks_per_node=4
+[INFO    ][2026-05-14 09:32:28][root                ][stage_datasets           ] [local_rank 2/4] Staging kinetics_240 (targz_parts)
+[INFO    ][2026-05-14 09:32:28][root                ][_stage_targz_parts       ] [rank 2] Extracting 26/103 tar.gz parts to /scratch-node/dcanez.22743106/kinetics_240
+[INFO    ][2026-05-14 09:32:43][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 2/26 parts
+[INFO    ][2026-05-14 09:32:57][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 4/26 parts
+[INFO    ][2026-05-14 09:33:12][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 6/26 parts
+[INFO    ][2026-05-14 09:33:26][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 8/26 parts
+[INFO    ][2026-05-14 09:33:40][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 10/26 parts
+[INFO    ][2026-05-14 09:33:54][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 12/26 parts
+[INFO    ][2026-05-14 09:34:08][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 14/26 parts
+[INFO    ][2026-05-14 09:34:23][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 16/26 parts
+[INFO    ][2026-05-14 09:34:38][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 18/26 parts
+[INFO    ][2026-05-14 09:34:53][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 20/26 parts
+[INFO    ][2026-05-14 09:35:07][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 22/26 parts
+[INFO    ][2026-05-14 09:35:21][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 24/26 parts
+[INFO    ][2026-05-14 09:35:36][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 26/26 parts
+[INFO    ][2026-05-14 09:35:36][root                ][stage_datasets           ] [local_rank 2/4] Staging ssv2 (multipart_tar)
+[INFO    ][2026-05-14 09:36:49][app.vjepa.train     ][main                     ] Staged datasets: ['/scratch-node/dcanez.22743106/kinetics_240/train.csv', '/scratch-node/dcanez.22743106/ssv2/train.csv']
+[INFO    ][2026-05-14 09:43:14][experiments.stmodels.vision_transformers_v3][vit_large_patch14_capi_lvd1689m] Loading pretrained weights for vit_large_patch14_capi_lvd1689m from https://dl.fbaipublicfiles.com/capi/capi_vitl14_lvd.pth
+[INFO    ][2026-05-14 09:43:23][root                ][_build_frozen_2d_target  ] Loaded frozen_2d target from timm: vit_large_patch14_capi.lvd1689m
+[INFO    ][2026-05-14 09:43:23][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): UJEPAside(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (rope): CAPI2DRoPE()
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (rope_impl): CAPI2DRoPE()
+        (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=False)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+          (rope_impl): CAPI2DRoPE()
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=False)
+    (st_blocks): ModuleList(
+      (0-23): 24 x FactorizedSlotSideBlock(
+        (cross_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (self_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (mlp): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (st_rope): CAPI3DRoPE()
+  )
+)
+[INFO    ][2026-05-14 09:43:23][root                ][init_video_model         ] UJEPAsidePredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-5): 6 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-05-14 09:43:24][root                ][init_video_model         ] MultiSeqWrapper(
+  (backbone): Frozen2DTargetWrapper(
+    (backbone): Eva(
+      (patch_embed): PatchEmbed(
+        (proj): Conv2d(3, 1024, kernel_size=(14, 14), stride=(14, 14))
+        (norm): Identity()
+      )
+      (pos_drop): Dropout(p=0.0, inplace=False)
+      (norm_pre): Identity()
+      (blocks): ModuleList(
+        (0-23): 24 x EvaBlock(
+          (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (attn): EvaAttention(
+            (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+            (q_norm): Identity()
+            (k_norm): Identity()
+            (attn_drop): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (proj): Linear(in_features=1024, out_features=1024, bias=True)
+            (proj_drop): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path1): Identity()
+          (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (mlp): Mlp(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+            (act): GELU(approximate='none')
+            (drop1): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+            (drop2): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path2): Identity()
+        )
+      )
+      (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+      (fc_norm): Identity()
+      (head_drop): Dropout(p=0.0, inplace=False)
+      (head): Identity()
+      (rope): _CapiPatchRoPE()
+    )
+  )
+)
+[INFO    ][2026-05-14 09:43:24][root                ][init_video_model         ] Encoder number of parameters: 403136512
+[INFO    ][2026-05-14 09:43:24][root                ][init_video_model         ] Predictor number of parameters: 11416192
+[INFO    ][2026-05-14 09:43:24][root                ][init_video_model         ] Target encoder number of parameters: 0
+[INFO    ][2026-05-14 09:43:40][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-05-14 09:43:40][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 2 / 16
+[INFO    ][2026-05-14 09:43:40][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-05-14 09:43:40][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/399
+[INFO    ][2026-05-14 09:43:40][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 2)...
+[INFO    ][2026-05-14 09:43:41][app.vjepa.train     ][main                     ] Initializing loader...
+submitit WARNING (2026-05-14 13:38:23,807) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,937) - Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGCONT

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_3_log.err

@@ -0,0 +1,15 @@
+/scratch-shared/dcanez/.cache/uv5/virtualenvs/vd/lib/python3.13/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+[rank3]:[W514 09:35:32.320490932 ProcessGroupNCCL.cpp:5138] Guessing device ID based on global rank. This can cause a hang if rank to GPU mapping is heterogeneous. You can specify device_id in init_process_group()
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/utils.py:796: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+submitit WARNING (2026-05-14 13:38:23,807) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,819) - Bypassing signal SIGTERM

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_3_log.out

@@ -0,0 +1,288 @@
+submitit INFO (2026-05-14 09:28:14,530) - Starting with JobEnvironment(job_id=22743106, hostname=gcn75.local.snellius.surf.nl, local_rank=3(4), node=0(4), global_rank=3(16))
+submitit INFO (2026-05-14 09:28:14,530) - Loading pickle: /scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 64,
+                'crop_size': 224,
+                'dataset_fpcs': [16, 16],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/scratch-shared/dcanez/data/kinetics/k400/train.csv',
+                                '/scratch-shared/dcanez/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 4,
+                'num_workers': 10,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'stage': [   {   'dest': 'kinetics_240',
+                                 'format': 'targz_parts',
+                                 'src': '/scratch-shared/dcanez/data/kinetics/k400/tars_240/'},
+                             {   'dest': 'ssv2',
+                                 'format': 'multipart_tar',
+                                 'src': '/scratch-nvme/ml-datasets/something-something-v2/'}],
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.15, 0.15],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.7, 0.7],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '180G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': False,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/train.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/val.csv',
+                                                          'eval_videos_per_class': 25,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 100},
+                                            'preset': 'ucf101'},
+                                        {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/train_coarse10.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 5,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ujepaside'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'model_name': 'ujepaside_large_patch14_capi_lvd1689m',
+                 'pred_depth': 6,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 12,
+                 'predictor': 'v2_cross',
+                 'st_causal': False,
+                 'st_drop_path': 0.2,
+                 'st_flex_enable': False,
+                 'st_layer_scale_init': 1e-05,
+                 'st_num_slots': 16,
+                 'st_side_block_type': 'factorized',
+                 'st_slots_causal_within_frame': False,
+                 'target_kind': 'frozen_2d',
+                 'target_type': 'vit_large_patch14_capi.lvd1689m',
+                 'temporal_spacing': 1.0,
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_mask_tokens': True,
+                 'use_rope': True,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 4,
+    'optimization': {   'clip_grad': 3.0,
+                        'ema': [0.99925, 0.99925],
+                        'epochs': 100,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0005,
+                        'start_lr': 0.0001,
+                        'warmup': 10,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 4}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-05-14 09:32:27][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-05-14 09:32:27][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-05-14 09:32:27][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=eno
+[INFO    ][2026-05-14 09:32:29][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 3/16, tasks_per_node=4
+[INFO    ][2026-05-14 09:32:29][root                ][stage_datasets           ] [local_rank 3/4] Staging kinetics_240 (targz_parts)
+[INFO    ][2026-05-14 09:32:29][root                ][_stage_targz_parts       ] [rank 3] Extracting 25/103 tar.gz parts to /scratch-node/dcanez.22743106/kinetics_240
+[INFO    ][2026-05-14 09:32:44][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 2/25 parts
+[INFO    ][2026-05-14 09:32:58][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 4/25 parts
+[INFO    ][2026-05-14 09:33:13][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 6/25 parts
+[INFO    ][2026-05-14 09:33:27][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 8/25 parts
+[INFO    ][2026-05-14 09:33:42][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 10/25 parts
+[INFO    ][2026-05-14 09:33:56][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 12/25 parts
+[INFO    ][2026-05-14 09:34:11][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 14/25 parts
+[INFO    ][2026-05-14 09:34:25][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 16/25 parts
+[INFO    ][2026-05-14 09:34:39][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 18/25 parts
+[INFO    ][2026-05-14 09:34:54][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 20/25 parts
+[INFO    ][2026-05-14 09:35:09][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 22/25 parts
+[INFO    ][2026-05-14 09:35:23][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 24/25 parts
+[INFO    ][2026-05-14 09:35:30][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 25/25 parts
+[INFO    ][2026-05-14 09:35:30][root                ][stage_datasets           ] [local_rank 3/4] Staging ssv2 (multipart_tar)
+[INFO    ][2026-05-14 09:36:49][app.vjepa.train     ][main                     ] Staged datasets: ['/scratch-node/dcanez.22743106/kinetics_240/train.csv', '/scratch-node/dcanez.22743106/ssv2/train.csv']
+[INFO    ][2026-05-14 09:43:14][experiments.stmodels.vision_transformers_v3][vit_large_patch14_capi_lvd1689m] Loading pretrained weights for vit_large_patch14_capi_lvd1689m from https://dl.fbaipublicfiles.com/capi/capi_vitl14_lvd.pth
+[INFO    ][2026-05-14 09:43:23][root                ][_build_frozen_2d_target  ] Loaded frozen_2d target from timm: vit_large_patch14_capi.lvd1689m
+[INFO    ][2026-05-14 09:43:23][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): UJEPAside(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (rope): CAPI2DRoPE()
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (rope_impl): CAPI2DRoPE()
+        (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=False)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+          (rope_impl): CAPI2DRoPE()
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=False)
+    (st_blocks): ModuleList(
+      (0-23): 24 x FactorizedSlotSideBlock(
+        (cross_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (self_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (mlp): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (st_rope): CAPI3DRoPE()
+  )
+)
+[INFO    ][2026-05-14 09:43:23][root                ][init_video_model         ] UJEPAsidePredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-5): 6 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-05-14 09:43:23][root                ][init_video_model         ] MultiSeqWrapper(
+  (backbone): Frozen2DTargetWrapper(
+    (backbone): Eva(
+      (patch_embed): PatchEmbed(
+        (proj): Conv2d(3, 1024, kernel_size=(14, 14), stride=(14, 14))
+        (norm): Identity()
+      )
+      (pos_drop): Dropout(p=0.0, inplace=False)
+      (norm_pre): Identity()
+      (blocks): ModuleList(
+        (0-23): 24 x EvaBlock(
+          (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (attn): EvaAttention(
+            (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+            (q_norm): Identity()
+            (k_norm): Identity()
+            (attn_drop): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (proj): Linear(in_features=1024, out_features=1024, bias=True)
+            (proj_drop): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path1): Identity()
+          (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (mlp): Mlp(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+            (act): GELU(approximate='none')
+            (drop1): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+            (drop2): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path2): Identity()
+        )
+      )
+      (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+      (fc_norm): Identity()
+      (head_drop): Dropout(p=0.0, inplace=False)
+      (head): Identity()
+      (rope): _CapiPatchRoPE()
+    )
+  )
+)
+[INFO    ][2026-05-14 09:43:23][root                ][init_video_model         ] Encoder number of parameters: 403136512
+[INFO    ][2026-05-14 09:43:23][root                ][init_video_model         ] Predictor number of parameters: 11416192
+[INFO    ][2026-05-14 09:43:23][root                ][init_video_model         ] Target encoder number of parameters: 0
+[INFO    ][2026-05-14 09:43:39][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-05-14 09:43:39][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 3 / 16
+[INFO    ][2026-05-14 09:43:39][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-05-14 09:43:39][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/399
+[INFO    ][2026-05-14 09:43:40][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 3)...
+[INFO    ][2026-05-14 09:43:41][app.vjepa.train     ][main                     ] Initializing loader...
+submitit WARNING (2026-05-14 13:38:23,807) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,819) - Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGCONT

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_4_log.err

@@ -0,0 +1,18 @@
+/scratch-shared/dcanez/.cache/uv5/virtualenvs/vd/lib/python3.13/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+[rank4]:[W514 09:36:37.824240654 ProcessGroupNCCL.cpp:5138] Guessing device ID based on global rank. This can cause a hang if rank to GPU mapping is heterogeneous. You can specify device_id in init_process_group()
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/utils.py:796: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+submitit WARNING (2026-05-14 13:38:23,805) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,872) - Bypassing signal SIGTERM
+[2026-05-14T13:38:55.995] error: namespace_p_join: open failed for /slurm/22743106/.ns: No such file or directory
+[2026-05-14T13:38:55.995] error: namespace_g_join(JobId=22743106 SLUID=s8FXKRJP5N0000): No such file or directory
+[2026-05-14T13:38:57.703] error: Failed to send MESSAGE_TASK_EXIT: Connection refused
+[2026-05-14T13:38:59.416] error: namespace_p_join: open failed for /slurm/22743106/.ns: No such file or directory
+[2026-05-14T13:38:59.416] error: namespace_g_join(JobId=22743106 SLUID=s8FXKRJP5N0000): No such file or directory
+[2026-05-14T13:38:59.553] error: namespace_p_join: open failed for /slurm/22743106/.ns: No such file or directory
+[2026-05-14T13:38:59.554] error: namespace_g_join(JobId=22743106 SLUID=s8FXKRJP5N0000): No such file or directory
+[2026-05-14T13:38:59.591] error: namespace_p_join: open failed for /slurm/22743106/.ns: No such file or directory
+[2026-05-14T13:38:59.591] error: namespace_g_join(JobId=22743106 SLUID=s8FXKRJP5N0000): No such file or directory

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_4_log.out

@@ -0,0 +1,292 @@
+submitit INFO (2026-05-14 09:28:15,038) - Starting with JobEnvironment(job_id=22743106, hostname=gcn79.local.snellius.surf.nl, local_rank=0(4), node=1(4), global_rank=4(16))
+submitit INFO (2026-05-14 09:28:15,038) - Loading pickle: /scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 64,
+                'crop_size': 224,
+                'dataset_fpcs': [16, 16],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/scratch-shared/dcanez/data/kinetics/k400/train.csv',
+                                '/scratch-shared/dcanez/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 4,
+                'num_workers': 10,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'stage': [   {   'dest': 'kinetics_240',
+                                 'format': 'targz_parts',
+                                 'src': '/scratch-shared/dcanez/data/kinetics/k400/tars_240/'},
+                             {   'dest': 'ssv2',
+                                 'format': 'multipart_tar',
+                                 'src': '/scratch-nvme/ml-datasets/something-something-v2/'}],
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.15, 0.15],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.7, 0.7],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '180G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': False,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/train.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/val.csv',
+                                                          'eval_videos_per_class': 25,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 100},
+                                            'preset': 'ucf101'},
+                                        {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/train_coarse10.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 5,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ujepaside'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'model_name': 'ujepaside_large_patch14_capi_lvd1689m',
+                 'pred_depth': 6,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 12,
+                 'predictor': 'v2_cross',
+                 'st_causal': False,
+                 'st_drop_path': 0.2,
+                 'st_flex_enable': False,
+                 'st_layer_scale_init': 1e-05,
+                 'st_num_slots': 16,
+                 'st_side_block_type': 'factorized',
+                 'st_slots_causal_within_frame': False,
+                 'target_kind': 'frozen_2d',
+                 'target_type': 'vit_large_patch14_capi.lvd1689m',
+                 'temporal_spacing': 1.0,
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_mask_tokens': True,
+                 'use_rope': True,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 4,
+    'optimization': {   'clip_grad': 3.0,
+                        'ema': [0.99925, 0.99925],
+                        'epochs': 100,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0005,
+                        'start_lr': 0.0001,
+                        'warmup': 10,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 4}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-05-14 09:32:14][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-05-14 09:32:14][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-05-14 09:32:14][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=eno
+[INFO    ][2026-05-14 09:32:32][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 4/16, tasks_per_node=4
+[INFO    ][2026-05-14 09:32:32][root                ][stage_datasets           ] [local_rank 0/4] Staging kinetics_240 (targz_parts)
+[INFO    ][2026-05-14 09:32:32][root                ][_stage_targz_parts       ] [rank 0] Extracting 26/103 tar.gz parts to /scratch-node/dcanez.22743106/kinetics_240
+[INFO    ][2026-05-14 09:32:46][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 2/26 parts
+[INFO    ][2026-05-14 09:33:01][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 4/26 parts
+[INFO    ][2026-05-14 09:33:15][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 6/26 parts
+[INFO    ][2026-05-14 09:33:30][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 8/26 parts
+[INFO    ][2026-05-14 09:33:44][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 10/26 parts
+[INFO    ][2026-05-14 09:33:59][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 12/26 parts
+[INFO    ][2026-05-14 09:34:13][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 14/26 parts
+[INFO    ][2026-05-14 09:34:28][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 16/26 parts
+[INFO    ][2026-05-14 09:34:43][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 18/26 parts
+[INFO    ][2026-05-14 09:34:58][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 20/26 parts
+[INFO    ][2026-05-14 09:35:12][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 22/26 parts
+[INFO    ][2026-05-14 09:35:27][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 24/26 parts
+[INFO    ][2026-05-14 09:35:42][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 26/26 parts
+[INFO    ][2026-05-14 09:35:42][root                ][stage_datasets           ] [local_rank 0/4] Staging ssv2 (multipart_tar)
+[INFO    ][2026-05-14 09:35:42][root                ][_stage_multipart_tar     ] [rank 0] Extracting multipart tar (2 files) to /scratch-node/dcanez.22743106/ssv2
+[INFO    ][2026-05-14 09:36:48][root                ][stage_datasets           ] Data staging completed in 256.0s (4.3min)
+[INFO    ][2026-05-14 09:36:49][root                ][_rewrite_csv             ] Wrote local CSV: /scratch-node/dcanez.22743106/kinetics_240/train.csv (239789 entries)
+[INFO    ][2026-05-14 09:36:49][root                ][_rewrite_csv             ] Wrote local CSV: /scratch-node/dcanez.22743106/ssv2/train.csv (168913 entries)
+[INFO    ][2026-05-14 09:36:49][app.vjepa.train     ][main                     ] Staged datasets: ['/scratch-node/dcanez.22743106/kinetics_240/train.csv', '/scratch-node/dcanez.22743106/ssv2/train.csv']
+[INFO    ][2026-05-14 09:43:14][experiments.stmodels.vision_transformers_v3][vit_large_patch14_capi_lvd1689m] Loading pretrained weights for vit_large_patch14_capi_lvd1689m from https://dl.fbaipublicfiles.com/capi/capi_vitl14_lvd.pth
+[INFO    ][2026-05-14 09:43:30][root                ][_build_frozen_2d_target  ] Loaded frozen_2d target from timm: vit_large_patch14_capi.lvd1689m
+[INFO    ][2026-05-14 09:43:30][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): UJEPAside(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (rope): CAPI2DRoPE()
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (rope_impl): CAPI2DRoPE()
+        (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=False)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+          (rope_impl): CAPI2DRoPE()
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=False)
+    (st_blocks): ModuleList(
+      (0-23): 24 x FactorizedSlotSideBlock(
+        (cross_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (self_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (mlp): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (st_rope): CAPI3DRoPE()
+  )
+)
+[INFO    ][2026-05-14 09:43:30][root                ][init_video_model         ] UJEPAsidePredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-5): 6 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-05-14 09:43:30][root                ][init_video_model         ] MultiSeqWrapper(
+  (backbone): Frozen2DTargetWrapper(
+    (backbone): Eva(
+      (patch_embed): PatchEmbed(
+        (proj): Conv2d(3, 1024, kernel_size=(14, 14), stride=(14, 14))
+        (norm): Identity()
+      )
+      (pos_drop): Dropout(p=0.0, inplace=False)
+      (norm_pre): Identity()
+      (blocks): ModuleList(
+        (0-23): 24 x EvaBlock(
+          (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (attn): EvaAttention(
+            (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+            (q_norm): Identity()
+            (k_norm): Identity()
+            (attn_drop): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (proj): Linear(in_features=1024, out_features=1024, bias=True)
+            (proj_drop): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path1): Identity()
+          (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (mlp): Mlp(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+            (act): GELU(approximate='none')
+            (drop1): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+            (drop2): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path2): Identity()
+        )
+      )
+      (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+      (fc_norm): Identity()
+      (head_drop): Dropout(p=0.0, inplace=False)
+      (head): Identity()
+      (rope): _CapiPatchRoPE()
+    )
+  )
+)
+[INFO    ][2026-05-14 09:43:30][root                ][init_video_model         ] Encoder number of parameters: 403136512
+[INFO    ][2026-05-14 09:43:30][root                ][init_video_model         ] Predictor number of parameters: 11416192
+[INFO    ][2026-05-14 09:43:30][root                ][init_video_model         ] Target encoder number of parameters: 0
+[INFO    ][2026-05-14 09:43:36][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-05-14 09:43:36][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 4 / 16
+[INFO    ][2026-05-14 09:43:36][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-05-14 09:43:36][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/399
+[INFO    ][2026-05-14 09:43:36][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 4)...
+[INFO    ][2026-05-14 09:43:41][app.vjepa.train     ][main                     ] Initializing loader...
+submitit WARNING (2026-05-14 13:38:23,805) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,872) - Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGCONT

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_5_log.err

@@ -0,0 +1,9 @@
+/scratch-shared/dcanez/.cache/uv5/virtualenvs/vd/lib/python3.13/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+[rank5]:[W514 09:35:35.321909479 ProcessGroupNCCL.cpp:5138] Guessing device ID based on global rank. This can cause a hang if rank to GPU mapping is heterogeneous. You can specify device_id in init_process_group()
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/utils.py:796: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+submitit WARNING (2026-05-14 13:38:23,805) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,814) - Bypassing signal SIGTERM

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_5_log.out

@@ -0,0 +1,288 @@
+submitit INFO (2026-05-14 09:28:15,038) - Starting with JobEnvironment(job_id=22743106, hostname=gcn79.local.snellius.surf.nl, local_rank=1(4), node=1(4), global_rank=5(16))
+submitit INFO (2026-05-14 09:28:15,038) - Loading pickle: /scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 64,
+                'crop_size': 224,
+                'dataset_fpcs': [16, 16],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/scratch-shared/dcanez/data/kinetics/k400/train.csv',
+                                '/scratch-shared/dcanez/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 4,
+                'num_workers': 10,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'stage': [   {   'dest': 'kinetics_240',
+                                 'format': 'targz_parts',
+                                 'src': '/scratch-shared/dcanez/data/kinetics/k400/tars_240/'},
+                             {   'dest': 'ssv2',
+                                 'format': 'multipart_tar',
+                                 'src': '/scratch-nvme/ml-datasets/something-something-v2/'}],
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.15, 0.15],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.7, 0.7],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '180G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': False,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/train.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/val.csv',
+                                                          'eval_videos_per_class': 25,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 100},
+                                            'preset': 'ucf101'},
+                                        {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/train_coarse10.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 5,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ujepaside'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'model_name': 'ujepaside_large_patch14_capi_lvd1689m',
+                 'pred_depth': 6,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 12,
+                 'predictor': 'v2_cross',
+                 'st_causal': False,
+                 'st_drop_path': 0.2,
+                 'st_flex_enable': False,
+                 'st_layer_scale_init': 1e-05,
+                 'st_num_slots': 16,
+                 'st_side_block_type': 'factorized',
+                 'st_slots_causal_within_frame': False,
+                 'target_kind': 'frozen_2d',
+                 'target_type': 'vit_large_patch14_capi.lvd1689m',
+                 'temporal_spacing': 1.0,
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_mask_tokens': True,
+                 'use_rope': True,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 4,
+    'optimization': {   'clip_grad': 3.0,
+                        'ema': [0.99925, 0.99925],
+                        'epochs': 100,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0005,
+                        'start_lr': 0.0001,
+                        'warmup': 10,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 4}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-05-14 09:32:17][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-05-14 09:32:17][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-05-14 09:32:17][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=eno
+[INFO    ][2026-05-14 09:32:31][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 5/16, tasks_per_node=4
+[INFO    ][2026-05-14 09:32:31][root                ][stage_datasets           ] [local_rank 1/4] Staging kinetics_240 (targz_parts)
+[INFO    ][2026-05-14 09:32:31][root                ][_stage_targz_parts       ] [rank 1] Extracting 26/103 tar.gz parts to /scratch-node/dcanez.22743106/kinetics_240
+[INFO    ][2026-05-14 09:32:43][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 2/26 parts
+[INFO    ][2026-05-14 09:32:57][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 4/26 parts
+[INFO    ][2026-05-14 09:33:11][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 6/26 parts
+[INFO    ][2026-05-14 09:33:25][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 8/26 parts
+[INFO    ][2026-05-14 09:33:39][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 10/26 parts
+[INFO    ][2026-05-14 09:33:54][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 12/26 parts
+[INFO    ][2026-05-14 09:34:07][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 14/26 parts
+[INFO    ][2026-05-14 09:34:22][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 16/26 parts
+[INFO    ][2026-05-14 09:34:36][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 18/26 parts
+[INFO    ][2026-05-14 09:34:51][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 20/26 parts
+[INFO    ][2026-05-14 09:35:06][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 22/26 parts
+[INFO    ][2026-05-14 09:35:20][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 24/26 parts
+[INFO    ][2026-05-14 09:35:35][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 26/26 parts
+[INFO    ][2026-05-14 09:35:35][root                ][stage_datasets           ] [local_rank 1/4] Staging ssv2 (multipart_tar)
+[INFO    ][2026-05-14 09:36:49][app.vjepa.train     ][main                     ] Staged datasets: ['/scratch-node/dcanez.22743106/kinetics_240/train.csv', '/scratch-node/dcanez.22743106/ssv2/train.csv']
+[INFO    ][2026-05-14 09:43:14][experiments.stmodels.vision_transformers_v3][vit_large_patch14_capi_lvd1689m] Loading pretrained weights for vit_large_patch14_capi_lvd1689m from https://dl.fbaipublicfiles.com/capi/capi_vitl14_lvd.pth
+[INFO    ][2026-05-14 09:43:30][root                ][_build_frozen_2d_target  ] Loaded frozen_2d target from timm: vit_large_patch14_capi.lvd1689m
+[INFO    ][2026-05-14 09:43:30][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): UJEPAside(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (rope): CAPI2DRoPE()
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (rope_impl): CAPI2DRoPE()
+        (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=False)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+          (rope_impl): CAPI2DRoPE()
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=False)
+    (st_blocks): ModuleList(
+      (0-23): 24 x FactorizedSlotSideBlock(
+        (cross_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (self_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (mlp): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (st_rope): CAPI3DRoPE()
+  )
+)
+[INFO    ][2026-05-14 09:43:30][root                ][init_video_model         ] UJEPAsidePredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-5): 6 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-05-14 09:43:30][root                ][init_video_model         ] MultiSeqWrapper(
+  (backbone): Frozen2DTargetWrapper(
+    (backbone): Eva(
+      (patch_embed): PatchEmbed(
+        (proj): Conv2d(3, 1024, kernel_size=(14, 14), stride=(14, 14))
+        (norm): Identity()
+      )
+      (pos_drop): Dropout(p=0.0, inplace=False)
+      (norm_pre): Identity()
+      (blocks): ModuleList(
+        (0-23): 24 x EvaBlock(
+          (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (attn): EvaAttention(
+            (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+            (q_norm): Identity()
+            (k_norm): Identity()
+            (attn_drop): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (proj): Linear(in_features=1024, out_features=1024, bias=True)
+            (proj_drop): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path1): Identity()
+          (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (mlp): Mlp(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+            (act): GELU(approximate='none')
+            (drop1): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+            (drop2): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path2): Identity()
+        )
+      )
+      (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+      (fc_norm): Identity()
+      (head_drop): Dropout(p=0.0, inplace=False)
+      (head): Identity()
+      (rope): _CapiPatchRoPE()
+    )
+  )
+)
+[INFO    ][2026-05-14 09:43:30][root                ][init_video_model         ] Encoder number of parameters: 403136512
+[INFO    ][2026-05-14 09:43:30][root                ][init_video_model         ] Predictor number of parameters: 11416192
+[INFO    ][2026-05-14 09:43:31][root                ][init_video_model         ] Target encoder number of parameters: 0
+[INFO    ][2026-05-14 09:43:36][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-05-14 09:43:36][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 5 / 16
+[INFO    ][2026-05-14 09:43:36][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-05-14 09:43:36][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/399
+[INFO    ][2026-05-14 09:43:36][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 5)...
+[INFO    ][2026-05-14 09:43:41][app.vjepa.train     ][main                     ] Initializing loader...
+submitit WARNING (2026-05-14 13:38:23,805) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,814) - Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGCONT

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_6_log.err

@@ -0,0 +1,9 @@
+/scratch-shared/dcanez/.cache/uv5/virtualenvs/vd/lib/python3.13/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+[rank6]:[W514 09:35:36.312177515 ProcessGroupNCCL.cpp:5138] Guessing device ID based on global rank. This can cause a hang if rank to GPU mapping is heterogeneous. You can specify device_id in init_process_group()
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/utils.py:796: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+submitit WARNING (2026-05-14 13:38:23,806) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,836) - Bypassing signal SIGTERM

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_6_log.out

@@ -0,0 +1,288 @@
+submitit INFO (2026-05-14 09:28:15,038) - Starting with JobEnvironment(job_id=22743106, hostname=gcn79.local.snellius.surf.nl, local_rank=2(4), node=1(4), global_rank=6(16))
+submitit INFO (2026-05-14 09:28:15,038) - Loading pickle: /scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 64,
+                'crop_size': 224,
+                'dataset_fpcs': [16, 16],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/scratch-shared/dcanez/data/kinetics/k400/train.csv',
+                                '/scratch-shared/dcanez/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 4,
+                'num_workers': 10,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'stage': [   {   'dest': 'kinetics_240',
+                                 'format': 'targz_parts',
+                                 'src': '/scratch-shared/dcanez/data/kinetics/k400/tars_240/'},
+                             {   'dest': 'ssv2',
+                                 'format': 'multipart_tar',
+                                 'src': '/scratch-nvme/ml-datasets/something-something-v2/'}],
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.15, 0.15],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.7, 0.7],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '180G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': False,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/train.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/val.csv',
+                                                          'eval_videos_per_class': 25,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 100},
+                                            'preset': 'ucf101'},
+                                        {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/train_coarse10.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 5,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ujepaside'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'model_name': 'ujepaside_large_patch14_capi_lvd1689m',
+                 'pred_depth': 6,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 12,
+                 'predictor': 'v2_cross',
+                 'st_causal': False,
+                 'st_drop_path': 0.2,
+                 'st_flex_enable': False,
+                 'st_layer_scale_init': 1e-05,
+                 'st_num_slots': 16,
+                 'st_side_block_type': 'factorized',
+                 'st_slots_causal_within_frame': False,
+                 'target_kind': 'frozen_2d',
+                 'target_type': 'vit_large_patch14_capi.lvd1689m',
+                 'temporal_spacing': 1.0,
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_mask_tokens': True,
+                 'use_rope': True,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 4,
+    'optimization': {   'clip_grad': 3.0,
+                        'ema': [0.99925, 0.99925],
+                        'epochs': 100,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0005,
+                        'start_lr': 0.0001,
+                        'warmup': 10,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 4}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-05-14 09:32:18][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-05-14 09:32:18][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-05-14 09:32:18][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=eno
+[INFO    ][2026-05-14 09:32:29][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 6/16, tasks_per_node=4
+[INFO    ][2026-05-14 09:32:29][root                ][stage_datasets           ] [local_rank 2/4] Staging kinetics_240 (targz_parts)
+[INFO    ][2026-05-14 09:32:29][root                ][_stage_targz_parts       ] [rank 2] Extracting 26/103 tar.gz parts to /scratch-node/dcanez.22743106/kinetics_240
+[INFO    ][2026-05-14 09:32:43][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 2/26 parts
+[INFO    ][2026-05-14 09:32:57][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 4/26 parts
+[INFO    ][2026-05-14 09:33:12][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 6/26 parts
+[INFO    ][2026-05-14 09:33:26][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 8/26 parts
+[INFO    ][2026-05-14 09:33:40][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 10/26 parts
+[INFO    ][2026-05-14 09:33:54][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 12/26 parts
+[INFO    ][2026-05-14 09:34:08][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 14/26 parts
+[INFO    ][2026-05-14 09:34:24][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 16/26 parts
+[INFO    ][2026-05-14 09:34:38][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 18/26 parts
+[INFO    ][2026-05-14 09:34:53][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 20/26 parts
+[INFO    ][2026-05-14 09:35:07][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 22/26 parts
+[INFO    ][2026-05-14 09:35:21][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 24/26 parts
+[INFO    ][2026-05-14 09:35:36][root                ][_stage_targz_parts       ] [local_rank 2] Extracted 26/26 parts
+[INFO    ][2026-05-14 09:35:36][root                ][stage_datasets           ] [local_rank 2/4] Staging ssv2 (multipart_tar)
+[INFO    ][2026-05-14 09:36:49][app.vjepa.train     ][main                     ] Staged datasets: ['/scratch-node/dcanez.22743106/kinetics_240/train.csv', '/scratch-node/dcanez.22743106/ssv2/train.csv']
+[INFO    ][2026-05-14 09:43:14][experiments.stmodels.vision_transformers_v3][vit_large_patch14_capi_lvd1689m] Loading pretrained weights for vit_large_patch14_capi_lvd1689m from https://dl.fbaipublicfiles.com/capi/capi_vitl14_lvd.pth
+[INFO    ][2026-05-14 09:43:30][root                ][_build_frozen_2d_target  ] Loaded frozen_2d target from timm: vit_large_patch14_capi.lvd1689m
+[INFO    ][2026-05-14 09:43:30][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): UJEPAside(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (rope): CAPI2DRoPE()
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (rope_impl): CAPI2DRoPE()
+        (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=False)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+          (rope_impl): CAPI2DRoPE()
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=False)
+    (st_blocks): ModuleList(
+      (0-23): 24 x FactorizedSlotSideBlock(
+        (cross_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (self_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (mlp): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (st_rope): CAPI3DRoPE()
+  )
+)
+[INFO    ][2026-05-14 09:43:30][root                ][init_video_model         ] UJEPAsidePredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-5): 6 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-05-14 09:43:31][root                ][init_video_model         ] MultiSeqWrapper(
+  (backbone): Frozen2DTargetWrapper(
+    (backbone): Eva(
+      (patch_embed): PatchEmbed(
+        (proj): Conv2d(3, 1024, kernel_size=(14, 14), stride=(14, 14))
+        (norm): Identity()
+      )
+      (pos_drop): Dropout(p=0.0, inplace=False)
+      (norm_pre): Identity()
+      (blocks): ModuleList(
+        (0-23): 24 x EvaBlock(
+          (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (attn): EvaAttention(
+            (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+            (q_norm): Identity()
+            (k_norm): Identity()
+            (attn_drop): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (proj): Linear(in_features=1024, out_features=1024, bias=True)
+            (proj_drop): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path1): Identity()
+          (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (mlp): Mlp(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+            (act): GELU(approximate='none')
+            (drop1): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+            (drop2): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path2): Identity()
+        )
+      )
+      (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+      (fc_norm): Identity()
+      (head_drop): Dropout(p=0.0, inplace=False)
+      (head): Identity()
+      (rope): _CapiPatchRoPE()
+    )
+  )
+)
+[INFO    ][2026-05-14 09:43:31][root                ][init_video_model         ] Encoder number of parameters: 403136512
+[INFO    ][2026-05-14 09:43:31][root                ][init_video_model         ] Predictor number of parameters: 11416192
+[INFO    ][2026-05-14 09:43:31][root                ][init_video_model         ] Target encoder number of parameters: 0
+[INFO    ][2026-05-14 09:43:36][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-05-14 09:43:36][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 6 / 16
+[INFO    ][2026-05-14 09:43:36][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-05-14 09:43:36][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/399
+[INFO    ][2026-05-14 09:43:36][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 6)...
+[INFO    ][2026-05-14 09:43:41][app.vjepa.train     ][main                     ] Initializing loader...
+submitit WARNING (2026-05-14 13:38:23,806) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,836) - Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGCONT

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_7_log.err

@@ -0,0 +1,9 @@
+/scratch-shared/dcanez/.cache/uv5/virtualenvs/vd/lib/python3.13/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+[rank7]:[W514 09:35:32.299194984 ProcessGroupNCCL.cpp:5138] Guessing device ID based on global rank. This can cause a hang if rank to GPU mapping is heterogeneous. You can specify device_id in init_process_group()
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/utils.py:796: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+submitit WARNING (2026-05-14 13:38:23,806) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,818) - Bypassing signal SIGTERM

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_7_log.out

@@ -0,0 +1,288 @@
+submitit INFO (2026-05-14 09:28:15,038) - Starting with JobEnvironment(job_id=22743106, hostname=gcn79.local.snellius.surf.nl, local_rank=3(4), node=1(4), global_rank=7(16))
+submitit INFO (2026-05-14 09:28:15,038) - Loading pickle: /scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 64,
+                'crop_size': 224,
+                'dataset_fpcs': [16, 16],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/scratch-shared/dcanez/data/kinetics/k400/train.csv',
+                                '/scratch-shared/dcanez/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 4,
+                'num_workers': 10,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'stage': [   {   'dest': 'kinetics_240',
+                                 'format': 'targz_parts',
+                                 'src': '/scratch-shared/dcanez/data/kinetics/k400/tars_240/'},
+                             {   'dest': 'ssv2',
+                                 'format': 'multipart_tar',
+                                 'src': '/scratch-nvme/ml-datasets/something-something-v2/'}],
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.15, 0.15],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.7, 0.7],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '180G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': False,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/train.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/val.csv',
+                                                          'eval_videos_per_class': 25,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 100},
+                                            'preset': 'ucf101'},
+                                        {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/train_coarse10.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 5,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ujepaside'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'model_name': 'ujepaside_large_patch14_capi_lvd1689m',
+                 'pred_depth': 6,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 12,
+                 'predictor': 'v2_cross',
+                 'st_causal': False,
+                 'st_drop_path': 0.2,
+                 'st_flex_enable': False,
+                 'st_layer_scale_init': 1e-05,
+                 'st_num_slots': 16,
+                 'st_side_block_type': 'factorized',
+                 'st_slots_causal_within_frame': False,
+                 'target_kind': 'frozen_2d',
+                 'target_type': 'vit_large_patch14_capi.lvd1689m',
+                 'temporal_spacing': 1.0,
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_mask_tokens': True,
+                 'use_rope': True,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 4,
+    'optimization': {   'clip_grad': 3.0,
+                        'ema': [0.99925, 0.99925],
+                        'epochs': 100,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0005,
+                        'start_lr': 0.0001,
+                        'warmup': 10,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 4}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-05-14 09:32:17][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-05-14 09:32:17][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-05-14 09:32:17][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=eno
+[INFO    ][2026-05-14 09:32:33][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 7/16, tasks_per_node=4
+[INFO    ][2026-05-14 09:32:33][root                ][stage_datasets           ] [local_rank 3/4] Staging kinetics_240 (targz_parts)
+[INFO    ][2026-05-14 09:32:33][root                ][_stage_targz_parts       ] [rank 3] Extracting 25/103 tar.gz parts to /scratch-node/dcanez.22743106/kinetics_240
+[INFO    ][2026-05-14 09:32:47][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 2/25 parts
+[INFO    ][2026-05-14 09:33:01][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 4/25 parts
+[INFO    ][2026-05-14 09:33:15][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 6/25 parts
+[INFO    ][2026-05-14 09:33:29][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 8/25 parts
+[INFO    ][2026-05-14 09:33:44][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 10/25 parts
+[INFO    ][2026-05-14 09:33:58][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 12/25 parts
+[INFO    ][2026-05-14 09:34:12][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 14/25 parts
+[INFO    ][2026-05-14 09:34:27][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 16/25 parts
+[INFO    ][2026-05-14 09:34:40][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 18/25 parts
+[INFO    ][2026-05-14 09:34:55][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 20/25 parts
+[INFO    ][2026-05-14 09:35:09][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 22/25 parts
+[INFO    ][2026-05-14 09:35:23][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 24/25 parts
+[INFO    ][2026-05-14 09:35:30][root                ][_stage_targz_parts       ] [local_rank 3] Extracted 25/25 parts
+[INFO    ][2026-05-14 09:35:30][root                ][stage_datasets           ] [local_rank 3/4] Staging ssv2 (multipart_tar)
+[INFO    ][2026-05-14 09:36:49][app.vjepa.train     ][main                     ] Staged datasets: ['/scratch-node/dcanez.22743106/kinetics_240/train.csv', '/scratch-node/dcanez.22743106/ssv2/train.csv']
+[INFO    ][2026-05-14 09:43:14][experiments.stmodels.vision_transformers_v3][vit_large_patch14_capi_lvd1689m] Loading pretrained weights for vit_large_patch14_capi_lvd1689m from https://dl.fbaipublicfiles.com/capi/capi_vitl14_lvd.pth
+[INFO    ][2026-05-14 09:43:30][root                ][_build_frozen_2d_target  ] Loaded frozen_2d target from timm: vit_large_patch14_capi.lvd1689m
+[INFO    ][2026-05-14 09:43:30][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): UJEPAside(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (rope): CAPI2DRoPE()
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (rope_impl): CAPI2DRoPE()
+        (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=False)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+          (rope_impl): CAPI2DRoPE()
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=False)
+    (st_blocks): ModuleList(
+      (0-23): 24 x FactorizedSlotSideBlock(
+        (cross_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (self_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (mlp): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (st_rope): CAPI3DRoPE()
+  )
+)
+[INFO    ][2026-05-14 09:43:30][root                ][init_video_model         ] UJEPAsidePredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-5): 6 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-05-14 09:43:30][root                ][init_video_model         ] MultiSeqWrapper(
+  (backbone): Frozen2DTargetWrapper(
+    (backbone): Eva(
+      (patch_embed): PatchEmbed(
+        (proj): Conv2d(3, 1024, kernel_size=(14, 14), stride=(14, 14))
+        (norm): Identity()
+      )
+      (pos_drop): Dropout(p=0.0, inplace=False)
+      (norm_pre): Identity()
+      (blocks): ModuleList(
+        (0-23): 24 x EvaBlock(
+          (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (attn): EvaAttention(
+            (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+            (q_norm): Identity()
+            (k_norm): Identity()
+            (attn_drop): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (proj): Linear(in_features=1024, out_features=1024, bias=True)
+            (proj_drop): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path1): Identity()
+          (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (mlp): Mlp(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+            (act): GELU(approximate='none')
+            (drop1): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+            (drop2): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path2): Identity()
+        )
+      )
+      (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+      (fc_norm): Identity()
+      (head_drop): Dropout(p=0.0, inplace=False)
+      (head): Identity()
+      (rope): _CapiPatchRoPE()
+    )
+  )
+)
+[INFO    ][2026-05-14 09:43:30][root                ][init_video_model         ] Encoder number of parameters: 403136512
+[INFO    ][2026-05-14 09:43:30][root                ][init_video_model         ] Predictor number of parameters: 11416192
+[INFO    ][2026-05-14 09:43:30][root                ][init_video_model         ] Target encoder number of parameters: 0
+[INFO    ][2026-05-14 09:43:36][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-05-14 09:43:36][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 7 / 16
+[INFO    ][2026-05-14 09:43:36][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-05-14 09:43:36][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/399
+[INFO    ][2026-05-14 09:43:36][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 7)...
+[INFO    ][2026-05-14 09:43:41][app.vjepa.train     ][main                     ] Initializing loader...
+submitit WARNING (2026-05-14 13:38:23,806) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,818) - Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGCONT

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_8_log.err

@@ -0,0 +1,18 @@
+/scratch-shared/dcanez/.cache/uv5/virtualenvs/vd/lib/python3.13/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+[rank8]:[W514 09:36:45.075654955 ProcessGroupNCCL.cpp:5138] Guessing device ID based on global rank. This can cause a hang if rank to GPU mapping is heterogeneous. You can specify device_id in init_process_group()
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/utils.py:796: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+submitit WARNING (2026-05-14 13:38:23,806) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,807) - Bypassing signal SIGTERM
+[2026-05-14T13:38:58.610] error: namespace_p_join: open failed for /slurm/22743106/.ns: No such file or directory
+[2026-05-14T13:38:58.611] error: namespace_g_join(JobId=22743106 SLUID=s8FXKRJP5N0000): No such file or directory
+[2026-05-14T13:39:00.263] error: Failed to send MESSAGE_TASK_EXIT: Connection refused
+[2026-05-14T13:39:00.265] error: namespace_p_join: open failed for /slurm/22743106/.ns: No such file or directory
+[2026-05-14T13:39:00.265] error: namespace_g_join(JobId=22743106 SLUID=s8FXKRJP5N0000): No such file or directory
+[2026-05-14T13:39:00.403] error: namespace_p_join: open failed for /slurm/22743106/.ns: No such file or directory
+[2026-05-14T13:39:00.404] error: namespace_g_join(JobId=22743106 SLUID=s8FXKRJP5N0000): No such file or directory
+[2026-05-14T13:39:00.441] error: namespace_p_join: open failed for /slurm/22743106/.ns: No such file or directory
+[2026-05-14T13:39:00.442] error: namespace_g_join(JobId=22743106 SLUID=s8FXKRJP5N0000): No such file or directory

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_8_log.out

@@ -0,0 +1,292 @@
+submitit INFO (2026-05-14 09:28:14,530) - Starting with JobEnvironment(job_id=22743106, hostname=gcn81.local.snellius.surf.nl, local_rank=0(4), node=2(4), global_rank=8(16))
+submitit INFO (2026-05-14 09:28:14,530) - Loading pickle: /scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 64,
+                'crop_size': 224,
+                'dataset_fpcs': [16, 16],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/scratch-shared/dcanez/data/kinetics/k400/train.csv',
+                                '/scratch-shared/dcanez/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 4,
+                'num_workers': 10,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'stage': [   {   'dest': 'kinetics_240',
+                                 'format': 'targz_parts',
+                                 'src': '/scratch-shared/dcanez/data/kinetics/k400/tars_240/'},
+                             {   'dest': 'ssv2',
+                                 'format': 'multipart_tar',
+                                 'src': '/scratch-nvme/ml-datasets/something-something-v2/'}],
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.15, 0.15],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.7, 0.7],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '180G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': False,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/train.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/val.csv',
+                                                          'eval_videos_per_class': 25,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 100},
+                                            'preset': 'ucf101'},
+                                        {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/train_coarse10.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 5,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ujepaside'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'model_name': 'ujepaside_large_patch14_capi_lvd1689m',
+                 'pred_depth': 6,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 12,
+                 'predictor': 'v2_cross',
+                 'st_causal': False,
+                 'st_drop_path': 0.2,
+                 'st_flex_enable': False,
+                 'st_layer_scale_init': 1e-05,
+                 'st_num_slots': 16,
+                 'st_side_block_type': 'factorized',
+                 'st_slots_causal_within_frame': False,
+                 'target_kind': 'frozen_2d',
+                 'target_type': 'vit_large_patch14_capi.lvd1689m',
+                 'temporal_spacing': 1.0,
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_mask_tokens': True,
+                 'use_rope': True,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 4,
+    'optimization': {   'clip_grad': 3.0,
+                        'ema': [0.99925, 0.99925],
+                        'epochs': 100,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0005,
+                        'start_lr': 0.0001,
+                        'warmup': 10,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 4}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-05-14 09:32:44][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-05-14 09:32:44][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-05-14 09:32:44][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=eno
+[INFO    ][2026-05-14 09:32:46][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 8/16, tasks_per_node=4
+[INFO    ][2026-05-14 09:32:46][root                ][stage_datasets           ] [local_rank 0/4] Staging kinetics_240 (targz_parts)
+[INFO    ][2026-05-14 09:32:46][root                ][_stage_targz_parts       ] [rank 0] Extracting 26/103 tar.gz parts to /scratch-node/dcanez.22743106/kinetics_240
+[INFO    ][2026-05-14 09:33:00][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 2/26 parts
+[INFO    ][2026-05-14 09:33:14][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 4/26 parts
+[INFO    ][2026-05-14 09:33:27][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 6/26 parts
+[INFO    ][2026-05-14 09:33:41][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 8/26 parts
+[INFO    ][2026-05-14 09:33:55][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 10/26 parts
+[INFO    ][2026-05-14 09:34:10][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 12/26 parts
+[INFO    ][2026-05-14 09:34:24][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 14/26 parts
+[INFO    ][2026-05-14 09:34:38][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 16/26 parts
+[INFO    ][2026-05-14 09:34:52][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 18/26 parts
+[INFO    ][2026-05-14 09:35:06][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 20/26 parts
+[INFO    ][2026-05-14 09:35:21][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 22/26 parts
+[INFO    ][2026-05-14 09:35:36][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 24/26 parts
+[INFO    ][2026-05-14 09:35:50][root                ][_stage_targz_parts       ] [local_rank 0] Extracted 26/26 parts
+[INFO    ][2026-05-14 09:35:50][root                ][stage_datasets           ] [local_rank 0/4] Staging ssv2 (multipart_tar)
+[INFO    ][2026-05-14 09:35:50][root                ][_stage_multipart_tar     ] [rank 0] Extracting multipart tar (2 files) to /scratch-node/dcanez.22743106/ssv2
+[INFO    ][2026-05-14 09:36:48][root                ][stage_datasets           ] Data staging completed in 242.5s (4.0min)
+[INFO    ][2026-05-14 09:36:49][root                ][_rewrite_csv             ] Wrote local CSV: /scratch-node/dcanez.22743106/kinetics_240/train.csv (239789 entries)
+[INFO    ][2026-05-14 09:36:49][root                ][_rewrite_csv             ] Wrote local CSV: /scratch-node/dcanez.22743106/ssv2/train.csv (168913 entries)
+[INFO    ][2026-05-14 09:36:49][app.vjepa.train     ][main                     ] Staged datasets: ['/scratch-node/dcanez.22743106/kinetics_240/train.csv', '/scratch-node/dcanez.22743106/ssv2/train.csv']
+[INFO    ][2026-05-14 09:43:14][experiments.stmodels.vision_transformers_v3][vit_large_patch14_capi_lvd1689m] Loading pretrained weights for vit_large_patch14_capi_lvd1689m from https://dl.fbaipublicfiles.com/capi/capi_vitl14_lvd.pth
+[INFO    ][2026-05-14 09:43:28][root                ][_build_frozen_2d_target  ] Loaded frozen_2d target from timm: vit_large_patch14_capi.lvd1689m
+[INFO    ][2026-05-14 09:43:28][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): UJEPAside(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (rope): CAPI2DRoPE()
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (rope_impl): CAPI2DRoPE()
+        (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=False)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+          (rope_impl): CAPI2DRoPE()
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=False)
+    (st_blocks): ModuleList(
+      (0-23): 24 x FactorizedSlotSideBlock(
+        (cross_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (self_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (mlp): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (st_rope): CAPI3DRoPE()
+  )
+)
+[INFO    ][2026-05-14 09:43:29][root                ][init_video_model         ] UJEPAsidePredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-5): 6 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-05-14 09:43:29][root                ][init_video_model         ] MultiSeqWrapper(
+  (backbone): Frozen2DTargetWrapper(
+    (backbone): Eva(
+      (patch_embed): PatchEmbed(
+        (proj): Conv2d(3, 1024, kernel_size=(14, 14), stride=(14, 14))
+        (norm): Identity()
+      )
+      (pos_drop): Dropout(p=0.0, inplace=False)
+      (norm_pre): Identity()
+      (blocks): ModuleList(
+        (0-23): 24 x EvaBlock(
+          (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (attn): EvaAttention(
+            (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+            (q_norm): Identity()
+            (k_norm): Identity()
+            (attn_drop): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (proj): Linear(in_features=1024, out_features=1024, bias=True)
+            (proj_drop): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path1): Identity()
+          (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (mlp): Mlp(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+            (act): GELU(approximate='none')
+            (drop1): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+            (drop2): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path2): Identity()
+        )
+      )
+      (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+      (fc_norm): Identity()
+      (head_drop): Dropout(p=0.0, inplace=False)
+      (head): Identity()
+      (rope): _CapiPatchRoPE()
+    )
+  )
+)
+[INFO    ][2026-05-14 09:43:29][root                ][init_video_model         ] Encoder number of parameters: 403136512
+[INFO    ][2026-05-14 09:43:29][root                ][init_video_model         ] Predictor number of parameters: 11416192
+[INFO    ][2026-05-14 09:43:29][root                ][init_video_model         ] Target encoder number of parameters: 0
+[INFO    ][2026-05-14 09:43:41][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-05-14 09:43:41][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 8 / 16
+[INFO    ][2026-05-14 09:43:41][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-05-14 09:43:41][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/399
+[INFO    ][2026-05-14 09:43:41][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 8)...
+[INFO    ][2026-05-14 09:43:41][app.vjepa.train     ][main                     ] Initializing loader...
+submitit WARNING (2026-05-14 13:38:23,806) - Bypassing signal SIGCONT
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,807) - Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGTERM

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_9_log.err

@@ -0,0 +1,9 @@
+/scratch-shared/dcanez/.cache/uv5/virtualenvs/vd/lib/python3.13/site-packages/timm/models/layers/__init__.py:49: FutureWarning: Importing from timm.models.layers is deprecated, please import via timm.layers
+  warnings.warn(f"Importing from {__name__} is deprecated, please import via timm.layers", FutureWarning)
+[rank9]:[W514 09:35:49.924907959 ProcessGroupNCCL.cpp:5138] Guessing device ID based on global rank. This can cause a hang if rank to GPU mapping is heterogeneous. You can specify device_id in init_process_group()
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/utils.py:796: FutureWarning: `torch.cuda.amp.GradScaler(args...)` is deprecated. Please use `torch.amp.GradScaler('cuda', args...)` instead.
+  scaler = torch.cuda.amp.GradScaler() if mixed_precision else None
+/gpfs/scratch1/shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/code/app/vjepa/train.py:955: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
+  with torch.cuda.amp.autocast(dtype=dtype, enabled=mixed_precision):
+submitit WARNING (2026-05-14 13:38:23,806) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,821) - Bypassing signal SIGTERM

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_9_log.out

@@ -0,0 +1,288 @@
+submitit INFO (2026-05-14 09:28:14,530) - Starting with JobEnvironment(job_id=22743106, hostname=gcn81.local.snellius.surf.nl, local_rank=1(4), node=2(4), global_rank=9(16))
+submitit INFO (2026-05-14 09:28:14,530) - Loading pickle: /scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/job_22743106/22743106_submitted.pkl
+INFO:root:loaded pretrain params...
+{   'app': 'vjepa',
+    'cpus_per_task': 16,
+    'data': {   'batch_size': 64,
+                'crop_size': 224,
+                'dataset_fpcs': [16, 16],
+                'dataset_type': 'VideoDataset',
+                'datasets': [   '/scratch-shared/dcanez/data/kinetics/k400/train.csv',
+                                '/scratch-shared/dcanez/data/ssv2/train.csv'],
+                'datasets_weights': [0.65, 0.35],
+                'fps': 4,
+                'num_workers': 10,
+                'patch_size': 14,
+                'persistent_workers': True,
+                'pin_mem': True,
+                'stage': [   {   'dest': 'kinetics_240',
+                                 'format': 'targz_parts',
+                                 'src': '/scratch-shared/dcanez/data/kinetics/k400/tars_240/'},
+                             {   'dest': 'ssv2',
+                                 'format': 'multipart_tar',
+                                 'src': '/scratch-nvme/ml-datasets/something-something-v2/'}],
+                'tubelet_size': 1},
+    'data_aug': {   'auto_augment': False,
+                    'motion_shift': False,
+                    'random_resize_aspect_ratio': [0.75, 1.35],
+                    'random_resize_scale': [0.3, 1.0],
+                    'reprob': 0.0},
+    'folder': '/scratch-shared/dcanez/runs/vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable',
+    'loss': {'loss_exp': 1.0},
+    'mask': [   {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 8,
+                    'spatial_scale': [0.15, 0.15],
+                    'temporal_scale': [1.0, 1.0]},
+                {   'aspect_ratio': [0.75, 1.5],
+                    'full_complement': False,
+                    'max_keep': None,
+                    'max_temporal_keep': 1.0,
+                    'num_blocks': 2,
+                    'spatial_scale': [0.7, 0.7],
+                    'temporal_scale': [1.0, 1.0]}],
+    'mem_per_gpu': '180G',
+    'meta': {   'dtype': 'bfloat16',
+                'knn_eval_epoch0': False,
+                'knn_eval_freq': 5,
+                'knn_eval_presets': [   {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/train.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/ucf101/ucfTrainTestlist/val.csv',
+                                                          'eval_videos_per_class': 25,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 100},
+                                            'preset': 'ucf101'},
+                                        {   'config': {   'batch_size': 64,
+                                                          'dataset_train': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/train_coarse10.csv',
+                                                          'dataset_val': '/scratch-shared/mdorkenw/20bn-something-something-v2/something-something-v2-annotations/val_coarse10.csv',
+                                                          'eval_videos_per_class': 100,
+                                                          'linear_probe': True,
+                                                          'num_workers': 8,
+                                                          'pool_type': 'slot_temporal_concat',
+                                                          'train_videos_per_class': 500},
+                                            'preset': 'ssv2_coarse10'}],
+                'load_checkpoint': True,
+                'read_checkpoint': None,
+                'save_every_freq': 5,
+                'seed': 239,
+                'use_sdpa': True,
+                'use_wandb': True,
+                'wandb_project': 'vjepa_ujepaside'},
+    'metrics': {'sigreg': {}, 'std': {}},
+    'model': {   'model_name': 'ujepaside_large_patch14_capi_lvd1689m',
+                 'pred_depth': 6,
+                 'pred_embed_dim': 384,
+                 'pred_num_heads': 12,
+                 'predictor': 'v2_cross',
+                 'st_causal': False,
+                 'st_drop_path': 0.2,
+                 'st_flex_enable': False,
+                 'st_layer_scale_init': 1e-05,
+                 'st_num_slots': 16,
+                 'st_side_block_type': 'factorized',
+                 'st_slots_causal_within_frame': False,
+                 'target_kind': 'frozen_2d',
+                 'target_type': 'vit_large_patch14_capi.lvd1689m',
+                 'temporal_spacing': 1.0,
+                 'uniform_power': True,
+                 'use_activation_checkpointing': True,
+                 'use_mask_tokens': True,
+                 'use_rope': True,
+                 'use_sdpa': True,
+                 'zero_init_mask_tokens': True},
+    'nodes': 4,
+    'optimization': {   'clip_grad': 3.0,
+                        'ema': [0.99925, 0.99925],
+                        'epochs': 100,
+                        'final_lr': 0.0001,
+                        'final_weight_decay': 0.04,
+                        'ipe': 300,
+                        'ipe_scale': 1.0,
+                        'lr': 0.0005,
+                        'start_lr': 0.0001,
+                        'warmup': 10,
+                        'weight_decay': 0.04},
+    'tasks_per_node': 4}
+INFO:root:Running pre-training of app: vjepa
+[INFO    ][2026-05-14 09:32:44][app.vjepa.train     ][main                     ] which_dtype='bfloat16'
+[INFO    ][2026-05-14 09:32:44][app.vjepa.train     ][main                     ] Disabling persistent_workers (incompatible with KNN eval)
+[INFO    ][2026-05-14 09:32:44][app.vjepa.train     ][main                     ] NCCL_SOCKET_IFNAME=eno
+[INFO    ][2026-05-14 09:32:46][app.vjepa.train     ][main                     ] Initialized (rank/world-size) 9/16, tasks_per_node=4
+[INFO    ][2026-05-14 09:32:46][root                ][stage_datasets           ] [local_rank 1/4] Staging kinetics_240 (targz_parts)
+[INFO    ][2026-05-14 09:32:46][root                ][_stage_targz_parts       ] [rank 1] Extracting 26/103 tar.gz parts to /scratch-node/dcanez.22743106/kinetics_240
+[INFO    ][2026-05-14 09:32:59][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 2/26 parts
+[INFO    ][2026-05-14 09:33:13][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 4/26 parts
+[INFO    ][2026-05-14 09:33:27][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 6/26 parts
+[INFO    ][2026-05-14 09:33:41][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 8/26 parts
+[INFO    ][2026-05-14 09:33:55][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 10/26 parts
+[INFO    ][2026-05-14 09:34:09][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 12/26 parts
+[INFO    ][2026-05-14 09:34:23][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 14/26 parts
+[INFO    ][2026-05-14 09:34:37][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 16/26 parts
+[INFO    ][2026-05-14 09:34:51][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 18/26 parts
+[INFO    ][2026-05-14 09:35:06][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 20/26 parts
+[INFO    ][2026-05-14 09:35:20][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 22/26 parts
+[INFO    ][2026-05-14 09:35:35][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 24/26 parts
+[INFO    ][2026-05-14 09:35:49][root                ][_stage_targz_parts       ] [local_rank 1] Extracted 26/26 parts
+[INFO    ][2026-05-14 09:35:49][root                ][stage_datasets           ] [local_rank 1/4] Staging ssv2 (multipart_tar)
+[INFO    ][2026-05-14 09:36:49][app.vjepa.train     ][main                     ] Staged datasets: ['/scratch-node/dcanez.22743106/kinetics_240/train.csv', '/scratch-node/dcanez.22743106/ssv2/train.csv']
+[INFO    ][2026-05-14 09:43:14][experiments.stmodels.vision_transformers_v3][vit_large_patch14_capi_lvd1689m] Loading pretrained weights for vit_large_patch14_capi_lvd1689m from https://dl.fbaipublicfiles.com/capi/capi_vitl14_lvd.pth
+[INFO    ][2026-05-14 09:43:28][root                ][_build_frozen_2d_target  ] Loaded frozen_2d target from timm: vit_large_patch14_capi.lvd1689m
+[INFO    ][2026-05-14 09:43:28][root                ][init_video_model         ] ViTMultiSeqWrapper(
+  (backbone): UJEPAside(
+    (patch_embed): PatchEmbed3D(
+      (proj): Conv3d(3, 1024, kernel_size=(1, 14, 14), stride=(1, 14, 14))
+    )
+    (rope): CAPI2DRoPE()
+    (blocks): ModuleList(
+      (0-23): 24 x Block(
+        (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (rope_impl): CAPI2DRoPE()
+        (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+        (drop_path1): Identity()
+        (drop_path2): Identity()
+        (attn): Attention(
+          (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+          (attn_drop): Dropout(p=0.0, inplace=False)
+          (proj): Linear(in_features=1024, out_features=1024, bias=False)
+          (proj_drop): Dropout(p=0.0, inplace=False)
+          (rope_impl): CAPI2DRoPE()
+        )
+        (mlp): MLP(
+          (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+          (act): GELU(approximate='none')
+          (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+          (drop): Dropout(p=0.0, inplace=False)
+        )
+      )
+    )
+    (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=False)
+    (st_blocks): ModuleList(
+      (0-23): 24 x FactorizedSlotSideBlock(
+        (cross_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (self_attn): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (k_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (v_proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (proj): Linear(in_features=1024, out_features=1024, bias=False)
+            (rope_impl): CAPI3DRoPE()
+          )
+        )
+        (mlp): EfficientResidual(
+          (norm): LayerNorm((1024,), eps=1e-06, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=4096, out_features=1024, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (st_rope): CAPI3DRoPE()
+  )
+)
+[INFO    ][2026-05-14 09:43:28][root                ][init_video_model         ] UJEPAsidePredictorMultiSeqWrapper(
+  (backbone): PredictorV2(
+    (predictor_embed): Linear(in_features=1024, out_features=384, bias=True)
+    (mask_tokens): ParameterList(
+        (0): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (1): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (2): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+        (3): Parameter containing: [torch.float32 of size 1x384 (cuda:0)]
+    )
+    (predictor_blocks): ModuleList(
+      (0-5): 6 x Block(
+        (residual1): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): Attention(
+            (q_proj): Linear(in_features=384, out_features=384, bias=False)
+            (k_proj): Linear(in_features=384, out_features=384, bias=False)
+            (v_proj): Linear(in_features=384, out_features=384, bias=False)
+            (proj): Linear(in_features=384, out_features=384, bias=False)
+            (rope): Rope()
+          )
+        )
+        (residual2): EfficientResidual(
+          (norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+          (fn): MLP(
+            (fc1): Linear(in_features=384, out_features=1536, bias=False)
+            (act): GELU(approximate='none')
+            (fc2): Linear(in_features=1536, out_features=384, bias=False)
+            (drop): Dropout(p=0.0, inplace=False)
+          )
+        )
+      )
+    )
+    (predictor_norm): LayerNorm((384,), eps=1e-05, elementwise_affine=True)
+    (predictor_proj): Linear(in_features=384, out_features=1024, bias=True)
+  )
+)
+[INFO    ][2026-05-14 09:43:28][root                ][init_video_model         ] MultiSeqWrapper(
+  (backbone): Frozen2DTargetWrapper(
+    (backbone): Eva(
+      (patch_embed): PatchEmbed(
+        (proj): Conv2d(3, 1024, kernel_size=(14, 14), stride=(14, 14))
+        (norm): Identity()
+      )
+      (pos_drop): Dropout(p=0.0, inplace=False)
+      (norm_pre): Identity()
+      (blocks): ModuleList(
+        (0-23): 24 x EvaBlock(
+          (norm1): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (attn): EvaAttention(
+            (qkv): Linear(in_features=1024, out_features=3072, bias=False)
+            (q_norm): Identity()
+            (k_norm): Identity()
+            (attn_drop): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (proj): Linear(in_features=1024, out_features=1024, bias=True)
+            (proj_drop): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path1): Identity()
+          (norm2): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+          (mlp): Mlp(
+            (fc1): Linear(in_features=1024, out_features=4096, bias=True)
+            (act): GELU(approximate='none')
+            (drop1): Dropout(p=0.0, inplace=False)
+            (norm): Identity()
+            (fc2): Linear(in_features=4096, out_features=1024, bias=True)
+            (drop2): Dropout(p=0.0, inplace=False)
+          )
+          (drop_path2): Identity()
+        )
+      )
+      (norm): RMSNorm((1024,), eps=1e-05, elementwise_affine=True)
+      (fc_norm): Identity()
+      (head_drop): Dropout(p=0.0, inplace=False)
+      (head): Identity()
+      (rope): _CapiPatchRoPE()
+    )
+  )
+)
+[INFO    ][2026-05-14 09:43:28][root                ][init_video_model         ] Encoder number of parameters: 403136512
+[INFO    ][2026-05-14 09:43:28][root                ][init_video_model         ] Predictor number of parameters: 11416192
+[INFO    ][2026-05-14 09:43:28][root                ][init_video_model         ] Target encoder number of parameters: 0
+[INFO    ][2026-05-14 09:43:41][root                ][make_videodataset        ] VideoDataset dataset created
+[INFO    ][2026-05-14 09:43:41][WeightedSampler     ][__init__                 ] Using DistributedWeightedSampler with rank 9 / 16
+[INFO    ][2026-05-14 09:43:41][root                ][make_videodataset        ] VideoDataset unsupervised data loader created
+[INFO    ][2026-05-14 09:43:41][app.vjepa.train     ][main                     ] iterations per epoch/dataset length: 300/399
+[INFO    ][2026-05-14 09:43:41][app.vjepa.train     ][main                     ] Wrapping models in DDP (rank 9)...
+[INFO    ][2026-05-14 09:43:41][app.vjepa.train     ][main                     ] Initializing loader...
+submitit WARNING (2026-05-14 13:38:23,806) - Bypassing signal SIGCONT
+submitit WARNING (2026-05-14 13:38:23,821) - Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGTERM
+[WARNING ][2026-05-14 13:38:23][submitit            ][bypass                   ] Bypassing signal SIGCONT

vjepa/ujepaside_capi_lvd/c003_vitl_k16_simple_cross_factorized_stable/latest.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dd7a6989d7d0453d4de239cf6a9436e31d37e7a1df58195db5332c43e5d10c3e
+size 7397341953