[INFO] Running in WANDB offline mode
[INFO] Running in WANDB offline mode
[INFO] Running in WANDB offline mode
[INFO] Running in WANDB offline mode
Trainer._get_train_sampler replaced with custom implementation.
warmup_ratio is deprecated and will be removed in v5.2. Use `warmup_steps` instead.
Trainer._get_train_sampler replaced with custom implementation.
Trainer._get_train_sampler replaced with custom implementation.
Trainer._get_train_sampler replaced with custom implementation.
warmup_ratio is deprecated and will be removed in v5.2. Use `warmup_steps` instead.
warmup_ratio is deprecated and will be removed in v5.2. Use `warmup_steps` instead.
warmup_ratio is deprecated and will be removed in v5.2. Use `warmup_steps` instead.
[2026-09-18 03:49:17,786] [INFO] [real_accelerator.py:222:get_accelerator] Setting ds_accelerator to cuda (auto detect)
[2026-09-18 03:49:17,798] [INFO] [real_accelerator.py:222:get_accelerator] Setting ds_accelerator to cuda (auto detect)
[2026-09-18 03:49:17,799] [INFO] [real_accelerator.py:222:get_accelerator] Setting ds_accelerator to cuda (auto detect)
[2026-09-18 03:49:17,801] [INFO] [real_accelerator.py:222:get_accelerator] Setting ds_accelerator to cuda (auto detect)
[2026-09-18 03:49:18,104] [INFO] [comm.py:658:init_distributed] cdb=None
[2026-09-18 03:49:18,104] [INFO] [comm.py:689:init_distributed] Initializing TorchBackend in DeepSpeed with backend nccl
[2026-09-18 03:49:18,129] [INFO] [comm.py:658:init_distributed] cdb=None
[2026-09-18 03:49:18,174] [INFO] [comm.py:658:init_distributed] cdb=None
[2026-09-18 03:49:18,178] [INFO] [comm.py:658:init_distributed] cdb=None
[TrainingFix] Qwen3.5 FlashAttention MRoPE packing guard enabled
[TrainingFix] Qwen3.5 FlashAttention MRoPE packing guard enabled
[TrainingFix] Qwen3.5 FlashAttention MRoPE packing guard enabled
[TrainingFix] Qwen3.5 FlashAttention MRoPE packing guard enabled

Loading weights:   0%|          | 0/723 [00:00<?, ?it/s]
Loading weights:   0%|          | 0/723 [00:00<?, ?it/s]
Loading weights: 100%|██████████| 723/723 [00:00<00:00, 10507.41it/s]

Loading weights:   0%|          | 0/723 [00:00<?, ?it/s]
Loading weights:   0%|          | 0/723 [00:00<?, ?it/s]
Loading weights: 100%|██████████| 723/723 [00:00<00:00, 11758.73it/s]

Loading weights: 100%|██████████| 723/723 [00:00<00:00, 10829.29it/s]

Loading weights: 100%|██████████| 723/723 [00:00<00:00, 10696.25it/s]
[VGGTEncoder] eval KV-cache window: start=8 recent=48 (total=56 frames)
[FeatureFusion] geometry fusion delta scaled by 0.5
[FeatureFusion] importance gate ENABLED on geometry delta
[FeatureFusion] learnable geometry scale ENABLED (init 0.5)
[VGGTEncoder] eval KV-cache window: start=8 recent=48 (total=56 frames)
[FeatureFusion] geometry fusion delta scaled by 0.5
[FeatureFusion] importance gate ENABLED on geometry delta
[FeatureFusion] learnable geometry scale ENABLED (init 0.5)
[VGGTEncoder] eval KV-cache window: start=8 recent=48 (total=56 frames)
[FeatureFusion] geometry fusion delta scaled by 0.5
[FeatureFusion] importance gate ENABLED on geometry delta
[FeatureFusion] learnable geometry scale ENABLED (init 0.5)
[VGGTEncoder] eval KV-cache window: start=8 recent=48 (total=56 frames)
[FeatureFusion] geometry fusion delta scaled by 0.5
[FeatureFusion] importance gate ENABLED on geometry delta
[FeatureFusion] learnable geometry scale ENABLED (init 0.5)
Loading weights from local directory
Loading weights from local directory
Loading weights from local directory
Loading weights from local directory
>>>>> STOP loss up-weight: 3.0 on token ids [48964]
Qwen3_5Config {
  "architectures": [
    "Qwen3_5ForConditionalGeneration"
  ],
  "dtype": "bfloat16",
  "feature_fusion_method": "deepstack_language_add",
  "fusion_num_layers": 1,
  "geometry_encoder_layers": [
    11,
    17,
    23
  ],
  "geometry_encoder_path": "/mnt/data/vmo-ai-task/dungpq6/model-checkpoint/VGGT-1B",
  "geometry_encoder_streaming": true,
  "geometry_encoder_type": "vggt",
  "geometry_frame_strict": true,
  "geometry_fusion_layers": [
    0,
    1,
    2
  ],
  "geometry_fusion_scale": 0.5,
  "geometry_importance_gate": true,
  "geometry_learnable_scale": true,
  "geometry_merger_type": "mlp",
  "geometry_spatial_bias": false,
  "image_token_id": 248056,
  "include_camera_token": false,
  "model_type": "qwen3_5",
  "pos_encoding_type": "none",
  "reference_frame": "first",
  "supervised_token_loss": true,
  "text_config": {
    "attention_bias": false,
    "attention_dropout": 0.0,
    "attn_output_gate": true,
    "bos_token_id": null,
    "dtype": "bfloat16",
    "eos_token_id": 248044,
    "full_attention_interval": 4,
    "head_dim": 256,
    "hidden_act": "silu",
    "hidden_size": 2560,
    "initializer_range": 0.02,
    "intermediate_size": 9216,
    "layer_types": [
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention"
    ],
    "linear_conv_kernel_dim": 4,
    "linear_key_head_dim": 128,
    "linear_num_key_heads": 16,
    "linear_num_value_heads": 32,
    "linear_value_head_dim": 128,
    "mamba_ssm_dtype": "float32",
    "max_position_embeddings": 262144,
    "mlp_only_layers": [],
    "model_type": "qwen3_5_text",
    "mtp_num_hidden_layers": 1,
    "mtp_use_dedicated_embeddings": false,
    "num_attention_heads": 16,
    "num_hidden_layers": 32,
    "num_key_value_heads": 4,
    "pad_token_id": null,
    "partial_rotary_factor": 0.25,
    "rms_norm_eps": 1e-06,
    "rope_parameters": {
      "mrope_interleaved": true,
      "mrope_section": [
        11,
        11,
        10
      ],
      "partial_rotary_factor": 0.25,
      "rope_theta": 10000000,
      "rope_type": "default"
    },
    "tie_word_embeddings": true,
    "use_cache": true,
    "vocab_size": 248320
  },
  "tie_word_embeddings": true,
  "transformers_version": "5.3.0",
  "use_cache": false,
  "use_geometry_encoder": true,
  "video_token_id": 248057,
  "vision_config": {
    "deepstack_visual_indexes": [],
    "depth": 24,
    "dtype": "bfloat16",
    "hidden_act": "gelu_pytorch_tanh",
    "hidden_size": 1024,
    "in_channels": 3,
    "initializer_range": 0.02,
    "intermediate_size": 4096,
    "model_type": "qwen3_5",
    "num_heads": 16,
    "num_position_embeddings": 2304,
    "out_hidden_size": 2560,
    "patch_size": 16,
    "spatial_merge_size": 2,
    "temporal_patch_size": 2
  },
  "vision_end_token_id": 248054,
  "vision_language_fusion_layers": null,
  "vision_start_token_id": 248053
}

Loading datasets: [{'annotation_path': '/mnt/data/vmo-ai-task/anhdh35/JanusVLN/train_r2r_rxr.json', 'data_path': '/mnt/data/vmo-ai-task/anhdh35/JanusVLN', 'tag': '3d', 'sampling_rate': 1.0, 'dataset_name': 'train_r2r_rxr'}]
>>>>> STOP loss up-weight: 3.0 on token ids [48964]
Qwen3_5Config {
  "architectures": [
    "Qwen3_5ForConditionalGeneration"
  ],
  "dtype": "bfloat16",
  "feature_fusion_method": "deepstack_language_add",
  "fusion_num_layers": 1,
  "geometry_encoder_layers": [
    11,
    17,
    23
  ],
  "geometry_encoder_path": "/mnt/data/vmo-ai-task/dungpq6/model-checkpoint/VGGT-1B",
  "geometry_encoder_streaming": true,
  "geometry_encoder_type": "vggt",
  "geometry_frame_strict": true,
  "geometry_fusion_layers": [
    0,
    1,
    2
  ],
  "geometry_fusion_scale": 0.5,
  "geometry_importance_gate": true,
  "geometry_learnable_scale": true,
  "geometry_merger_type": "mlp",
  "geometry_spatial_bias": false,
  "image_token_id": 248056,
  "include_camera_token": false,
  "model_type": "qwen3_5",
  "pos_encoding_type": "none",
  "reference_frame": "first",
  "supervised_token_loss": true,
  "text_config": {
    "attention_bias": false,
    "attention_dropout": 0.0,
    "attn_output_gate": true,
    "bos_token_id": null,
    "dtype": "bfloat16",
    "eos_token_id": 248044,
    "full_attention_interval": 4,
    "head_dim": 256,
    "hidden_act": "silu",
    "hidden_size": 2560,
    "initializer_range": 0.02,
    "intermediate_size": 9216,
    "layer_types": [
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention"
    ],
    "linear_conv_kernel_dim": 4,
    "linear_key_head_dim": 128,
    "linear_num_key_heads": 16,
    "linear_num_value_heads": 32,
    "linear_value_head_dim": 128,
    "mamba_ssm_dtype": "float32",
    "max_position_embeddings": 262144,
    "mlp_only_layers": [],
    "model_type": "qwen3_5_text",
    "mtp_num_hidden_layers": 1,
    "mtp_use_dedicated_embeddings": false,
    "num_attention_heads": 16,
    "num_hidden_layers": 32,
    "num_key_value_heads": 4,
    "pad_token_id": null,
    "partial_rotary_factor": 0.25,
    "rms_norm_eps": 1e-06,
    "rope_parameters": {
      "mrope_interleaved": true,
      "mrope_section": [
        11,
        11,
        10
      ],
      "partial_rotary_factor": 0.25,
      "rope_theta": 10000000,
      "rope_type": "default"
    },
    "tie_word_embeddings": true,
    "use_cache": true,
    "vocab_size": 248320
  },
  "tie_word_embeddings": true,
  "transformers_version": "5.3.0",
  "use_cache": false,
  "use_geometry_encoder": true,
  "video_token_id": 248057,
  "vision_config": {
    "deepstack_visual_indexes": [],
    "depth": 24,
    "dtype": "bfloat16",
    "hidden_act": "gelu_pytorch_tanh",
    "hidden_size": 1024,
    "in_channels": 3,
    "initializer_range": 0.02,
    "intermediate_size": 4096,
    "model_type": "qwen3_5",
    "num_heads": 16,
    "num_position_embeddings": 2304,
    "out_hidden_size": 2560,
    "patch_size": 16,
    "spatial_merge_size": 2,
    "temporal_patch_size": 2
  },
  "vision_end_token_id": 248054,
  "vision_language_fusion_layers": null,
  "vision_start_token_id": 248053
}

Loading datasets: [{'annotation_path': '/mnt/data/vmo-ai-task/anhdh35/JanusVLN/train_r2r_rxr.json', 'data_path': '/mnt/data/vmo-ai-task/anhdh35/JanusVLN', 'tag': '3d', 'sampling_rate': 1.0, 'dataset_name': 'train_r2r_rxr'}]
>>>>> STOP loss up-weight: 3.0 on token ids [48964]
Vision Module - Attention Blocks:
Trainable Block Indices: None
Non-Trainable Block Indices: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23]
Merger Module Trainable: True
LLM Module - Embed Tokens Trainable: True
LLM Module - Trainable Layer Indices: [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31]
LLM Module - Non-Trainable Layer Indices: None
Qwen3_5Config {
  "architectures": [
    "Qwen3_5ForConditionalGeneration"
  ],
  "dtype": "bfloat16",
  "feature_fusion_method": "deepstack_language_add",
  "fusion_num_layers": 1,
  "geometry_encoder_layers": [
    11,
    17,
    23
  ],
  "geometry_encoder_path": "/mnt/data/vmo-ai-task/dungpq6/model-checkpoint/VGGT-1B",
  "geometry_encoder_streaming": true,
  "geometry_encoder_type": "vggt",
  "geometry_frame_strict": true,
  "geometry_fusion_layers": [
    0,
    1,
    2
  ],
  "geometry_fusion_scale": 0.5,
  "geometry_importance_gate": true,
  "geometry_learnable_scale": true,
  "geometry_merger_type": "mlp",
  "geometry_spatial_bias": false,
  "image_token_id": 248056,
  "include_camera_token": false,
  "model_type": "qwen3_5",
  "pos_encoding_type": "none",
  "reference_frame": "first",
  "supervised_token_loss": true,
  "text_config": {
    "attention_bias": false,
    "attention_dropout": 0.0,
    "attn_output_gate": true,
    "bos_token_id": null,
    "dtype": "bfloat16",
    "eos_token_id": 248044,
    "full_attention_interval": 4,
    "head_dim": 256,
    "hidden_act": "silu",
    "hidden_size": 2560,
    "initializer_range": 0.02,
    "intermediate_size": 9216,
    "layer_types": [
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention"
    ],
    "linear_conv_kernel_dim": 4,
    "linear_key_head_dim": 128,
    "linear_num_key_heads": 16,
    "linear_num_value_heads": 32,
    "linear_value_head_dim": 128,
    "mamba_ssm_dtype": "float32",
    "max_position_embeddings": 262144,
    "mlp_only_layers": [],
    "model_type": "qwen3_5_text",
    "mtp_num_hidden_layers": 1,
    "mtp_use_dedicated_embeddings": false,
    "num_attention_heads": 16,
    "num_hidden_layers": 32,
    "num_key_value_heads": 4,
    "pad_token_id": null,
    "partial_rotary_factor": 0.25,
    "rms_norm_eps": 1e-06,
    "rope_parameters": {
      "mrope_interleaved": true,
      "mrope_section": [
        11,
        11,
        10
      ],
      "partial_rotary_factor": 0.25,
      "rope_theta": 10000000,
      "rope_type": "default"
    },
    "tie_word_embeddings": true,
    "use_cache": true,
    "vocab_size": 248320
  },
  "tie_word_embeddings": true,
  "transformers_version": "5.3.0",
  "use_cache": false,
  "use_geometry_encoder": true,
  "video_token_id": 248057,
  "vision_config": {
    "deepstack_visual_indexes": [],
    "depth": 24,
    "dtype": "bfloat16",
    "hidden_act": "gelu_pytorch_tanh",
    "hidden_size": 1024,
    "in_channels": 3,
    "initializer_range": 0.02,
    "intermediate_size": 4096,
    "model_type": "qwen3_5",
    "num_heads": 16,
    "num_position_embeddings": 2304,
    "out_hidden_size": 2560,
    "patch_size": 16,
    "spatial_merge_size": 2,
    "temporal_patch_size": 2
  },
  "vision_end_token_id": 248054,
  "vision_language_fusion_layers": null,
  "vision_start_token_id": 248053
}

Loading datasets: [{'annotation_path': '/mnt/data/vmo-ai-task/anhdh35/JanusVLN/train_r2r_rxr.json', 'data_path': '/mnt/data/vmo-ai-task/anhdh35/JanusVLN', 'tag': '3d', 'sampling_rate': 1.0, 'dataset_name': 'train_r2r_rxr'}]
>>>>> STOP loss up-weight: 3.0 on token ids [48964]
Qwen3_5Config {
  "architectures": [
    "Qwen3_5ForConditionalGeneration"
  ],
  "dtype": "bfloat16",
  "feature_fusion_method": "deepstack_language_add",
  "fusion_num_layers": 1,
  "geometry_encoder_layers": [
    11,
    17,
    23
  ],
  "geometry_encoder_path": "/mnt/data/vmo-ai-task/dungpq6/model-checkpoint/VGGT-1B",
  "geometry_encoder_streaming": true,
  "geometry_encoder_type": "vggt",
  "geometry_frame_strict": true,
  "geometry_fusion_layers": [
    0,
    1,
    2
  ],
  "geometry_fusion_scale": 0.5,
  "geometry_importance_gate": true,
  "geometry_learnable_scale": true,
  "geometry_merger_type": "mlp",
  "geometry_spatial_bias": false,
  "image_token_id": 248056,
  "include_camera_token": false,
  "model_type": "qwen3_5",
  "pos_encoding_type": "none",
  "reference_frame": "first",
  "supervised_token_loss": true,
  "text_config": {
    "attention_bias": false,
    "attention_dropout": 0.0,
    "attn_output_gate": true,
    "bos_token_id": null,
    "dtype": "bfloat16",
    "eos_token_id": 248044,
    "full_attention_interval": 4,
    "head_dim": 256,
    "hidden_act": "silu",
    "hidden_size": 2560,
    "initializer_range": 0.02,
    "intermediate_size": 9216,
    "layer_types": [
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention",
      "linear_attention",
      "linear_attention",
      "linear_attention",
      "full_attention"
    ],
    "linear_conv_kernel_dim": 4,
    "linear_key_head_dim": 128,
    "linear_num_key_heads": 16,
    "linear_num_value_heads": 32,
    "linear_value_head_dim": 128,
    "mamba_ssm_dtype": "float32",
    "max_position_embeddings": 262144,
    "mlp_only_layers": [],
    "model_type": "qwen3_5_text",
    "mtp_num_hidden_layers": 1,
    "mtp_use_dedicated_embeddings": false,
    "num_attention_heads": 16,
    "num_hidden_layers": 32,
    "num_key_value_heads": 4,
    "pad_token_id": null,
    "partial_rotary_factor": 0.25,
    "rms_norm_eps": 1e-06,
    "rope_parameters": {
      "mrope_interleaved": true,
      "mrope_section": [
        11,
        11,
        10
      ],
      "partial_rotary_factor": 0.25,
      "rope_theta": 10000000,
      "rope_type": "default"
    },
    "tie_word_embeddings": true,
    "use_cache": true,
    "vocab_size": 248320
  },
  "tie_word_embeddings": true,
  "transformers_version": "5.3.0",
  "use_cache": false,
  "use_geometry_encoder": true,
  "video_token_id": 248057,
  "vision_config": {
    "deepstack_visual_indexes": [],
    "depth": 24,
    "dtype": "bfloat16",
    "hidden_act": "gelu_pytorch_tanh",
    "hidden_size": 1024,
    "in_channels": 3,
    "initializer_range": 0.02,
    "intermediate_size": 4096,
    "model_type": "qwen3_5",
    "num_heads": 16,
    "num_position_embeddings": 2304,
    "out_hidden_size": 2560,
    "patch_size": 16,
    "spatial_merge_size": 2,
    "temporal_patch_size": 2
  },
  "vision_end_token_id": 248054,
  "vision_language_fusion_layers": null,
  "vision_start_token_id": 248053
}

Loading datasets: [{'annotation_path': '/mnt/data/vmo-ai-task/anhdh35/JanusVLN/train_r2r_rxr.json', 'data_path': '/mnt/data/vmo-ai-task/anhdh35/JanusVLN', 'tag': '3d', 'sampling_rate': 1.0, 'dataset_name': 'train_r2r_rxr'}]
Total training samples: 2429620
Total training samples: 2429620
Total training samples: 2429620
Total training samples: 2429620
Formatting inputs...Skip in lazy mode
Formatting inputs...Skip in lazy mode
Formatting inputs...Skip in lazy mode
Formatting inputs...Skip in lazy mode
The tokenizer has new PAD/BOS/EOS tokens that differ from the model config and generation config. The model config and generation config were aligned accordingly, being updated with the tokenizer's values. Updated tokens: {'eos_token_id': 248046, 'pad_token_id': 248044}.
The tokenizer has new PAD/BOS/EOS tokens that differ from the model config and generation config. The model config and generation config were aligned accordingly, being updated with the tokenizer's values. Updated tokens: {'eos_token_id': 248046, 'pad_token_id': 248044}.
The tokenizer has new PAD/BOS/EOS tokens that differ from the model config and generation config. The model config and generation config were aligned accordingly, being updated with the tokenizer's values. Updated tokens: {'eos_token_id': 248046, 'pad_token_id': 248044}.
The tokenizer has new PAD/BOS/EOS tokens that differ from the model config and generation config. The model config and generation config were aligned accordingly, being updated with the tokenizer's values. Updated tokens: {'eos_token_id': 248046, 'pad_token_id': 248044}.
NCCL version 2.26.2+cuda12.2
[TrainingRuntime] {"rank": "1", "versions": {"torch": "2.7.0+cu128", "transformers": "5.3.0", "deepspeed": "0.16.4", "flash-attn": "2.8.3", "triton": "3.3.0", "flash-linear-attention": "0.5.1", "causal-conv1d": "1.6.2.post1"}, "cuda": "12.8", "world_size": 4, "gpu": "NVIDIA H100 80GB HBM3", "micro_batch": 16, "accumulation": 1, "effective_global_batch": 64, "supervised_token_loss": true, "NCCL_NVLS_ENABLE": "0", "CUDA_LAUNCH_BLOCKING": "0", "deepspeed": "scripts/zero2_h100.json"}
[TrainingMemory] {"rank": "1", "stage": "train_begin", "step": 0, "allocated_GiB": 14.601, "reserved_GiB": 19.906, "peak_allocated_GiB": 18.713, "peak_reserved_GiB": 19.906, "device_free_GiB": 57.864, "device_total_GiB": 79.189}
[TrainingRuntime] {"rank": "3", "versions": {"torch": "2.7.0+cu128", "transformers": "5.3.0", "deepspeed": "0.16.4", "flash-attn": "2.8.3", "triton": "3.3.0", "flash-linear-attention": "0.5.1", "causal-conv1d": "1.6.2.post1"}, "cuda": "12.8", "world_size": 4, "gpu": "NVIDIA H100 80GB HBM3", "micro_batch": 16, "accumulation": 1, "effective_global_batch": 64, "supervised_token_loss": true, "NCCL_NVLS_ENABLE": "0", "CUDA_LAUNCH_BLOCKING": "0", "deepspeed": "scripts/zero2_h100.json"}
[TrainingMemory] {"rank": "3", "stage": "train_begin", "step": 0, "allocated_GiB": 14.601, "reserved_GiB": 19.906, "peak_allocated_GiB": 18.713, "peak_reserved_GiB": 19.906, "device_free_GiB": 57.864, "device_total_GiB": 79.189}
[TrainingRuntime] {"rank": "2", "versions": {"torch": "2.7.0+cu128", "transformers": "5.3.0", "deepspeed": "0.16.4", "flash-attn": "2.8.3", "triton": "3.3.0", "flash-linear-attention": "0.5.1", "causal-conv1d": "1.6.2.post1"}, "cuda": "12.8", "world_size": 4, "gpu": "NVIDIA H100 80GB HBM3", "micro_batch": 16, "accumulation": 1, "effective_global_batch": 64, "supervised_token_loss": true, "NCCL_NVLS_ENABLE": "0", "CUDA_LAUNCH_BLOCKING": "0", "deepspeed": "scripts/zero2_h100.json"}
[TrainingMemory] {"rank": "2", "stage": "train_begin", "step": 0, "allocated_GiB": 14.601, "reserved_GiB": 19.906, "peak_allocated_GiB": 18.713, "peak_reserved_GiB": 19.906, "device_free_GiB": 57.864, "device_total_GiB": 79.189}
[TrainingRuntime] {"rank": "0", "versions": {"torch": "2.7.0+cu128", "transformers": "5.3.0", "deepspeed": "0.16.4", "flash-attn": "2.8.3", "triton": "3.3.0", "flash-linear-attention": "0.5.1", "causal-conv1d": "1.6.2.post1"}, "cuda": "12.8", "world_size": 4, "gpu": "NVIDIA H100 80GB HBM3", "micro_batch": 16, "accumulation": 1, "effective_global_batch": 64, "supervised_token_loss": true, "NCCL_NVLS_ENABLE": "0", "CUDA_LAUNCH_BLOCKING": "0", "deepspeed": "scripts/zero2_h100.json"}
[TrainingMemory] {"rank": "0", "stage": "train_begin", "step": 0, "allocated_GiB": 14.601, "reserved_GiB": 19.906, "peak_allocated_GiB": 14.601, "peak_reserved_GiB": 19.906, "device_free_GiB": 57.864, "device_total_GiB": 79.189}

  0%|          | 0/12 [00:00<?, ?it/s][GeometryBatch] micro_batch=16, mode=per-sample VGGT fallback, input_shapes=[(9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (6, 3, 420, 560), (3, 3, 420, 560)]
/home/dungpq6/Project/SpatialStack/src/qwen_vl/model/geometry_encoders/vggt_encoder.py:337: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
  with torch.cuda.amp.autocast(dtype=dtype):
[GeometryBatch] micro_batch=16, mode=per-sample VGGT fallback, input_shapes=[(9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (8, 3, 420, 560), (3, 3, 420, 560)]
/home/dungpq6/Project/SpatialStack/src/qwen_vl/model/geometry_encoders/vggt_encoder.py:337: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
  with torch.cuda.amp.autocast(dtype=dtype):
[GeometryBatch] micro_batch=16, mode=per-sample VGGT fallback, input_shapes=[(9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (5, 3, 420, 560), (4, 3, 420, 560)]
/home/dungpq6/Project/SpatialStack/src/qwen_vl/model/geometry_encoders/vggt_encoder.py:337: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
  with torch.cuda.amp.autocast(dtype=dtype):
[GeometryBatch] micro_batch=16, mode=per-sample VGGT fallback, input_shapes=[(9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (9, 3, 420, 560), (2, 3, 420, 560)]
/home/dungpq6/Project/SpatialStack/src/qwen_vl/model/geometry_encoders/vggt_encoder.py:337: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
  with torch.cuda.amp.autocast(dtype=dtype):
[TrainingMemory] {"rank": "0", "stage": "before_logging", "step": 1, "allocated_GiB": 24.192, "reserved_GiB": 39.621, "peak_allocated_GiB": 45.39, "peak_reserved_GiB": 71.133, "device_free_GiB": 33.303, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "2", "stage": "before_logging", "step": 1, "allocated_GiB": 24.172, "reserved_GiB": 71.924, "peak_allocated_GiB": 46.04, "peak_reserved_GiB": 71.924, "device_free_GiB": 1.001, "device_total_GiB": 79.189}

  8%|▊         | 1/12 [01:09<12:49, 69.94s/it][TrainingMemory] {"rank": "1", "stage": "before_logging", "step": 1, "allocated_GiB": 24.172, "reserved_GiB": 71.883, "peak_allocated_GiB": 46.028, "peak_reserved_GiB": 71.883, "device_free_GiB": 1.042, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "3", "stage": "before_logging", "step": 1, "allocated_GiB": 24.191, "reserved_GiB": 39.588, "peak_allocated_GiB": 45.972, "peak_reserved_GiB": 68.902, "device_free_GiB": 33.337, "device_total_GiB": 79.189}
/home/dungpq6/Project/SpatialStack/src/qwen_vl/model/geometry_encoders/vggt_encoder.py:337: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
  with torch.cuda.amp.autocast(dtype=dtype):
/home/dungpq6/Project/SpatialStack/src/qwen_vl/model/geometry_encoders/vggt_encoder.py:337: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
  with torch.cuda.amp.autocast(dtype=dtype):
/home/dungpq6/Project/SpatialStack/src/qwen_vl/model/geometry_encoders/vggt_encoder.py:337: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
  with torch.cuda.amp.autocast(dtype=dtype):
/home/dungpq6/Project/SpatialStack/src/qwen_vl/model/geometry_encoders/vggt_encoder.py:337: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
  with torch.cuda.amp.autocast(dtype=dtype):
[TrainingMemory] {"rank": "2", "stage": "before_logging", "step": 2, "allocated_GiB": 24.182, "reserved_GiB": 71.924, "peak_allocated_GiB": 54.502, "peak_reserved_GiB": 71.924, "device_free_GiB": 1.001, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "3", "stage": "before_logging", "step": 2, "allocated_GiB": 24.173, "reserved_GiB": 63.305, "peak_allocated_GiB": 54.051, "peak_reserved_GiB": 63.305, "device_free_GiB": 9.62, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "1", "stage": "before_logging", "step": 2, "allocated_GiB": 24.182, "reserved_GiB": 71.883, "peak_allocated_GiB": 54.57, "peak_reserved_GiB": 71.883, "device_free_GiB": 1.042, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "0", "stage": "before_logging", "step": 2, "allocated_GiB": 24.154, "reserved_GiB": 65.676, "peak_allocated_GiB": 53.816, "peak_reserved_GiB": 65.676, "device_free_GiB": 7.247, "device_total_GiB": 79.189}

 17%|█▋        | 2/12 [01:49<08:40, 52.06s/it]/home/dungpq6/Project/SpatialStack/src/qwen_vl/model/geometry_encoders/vggt_encoder.py:337: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
  with torch.cuda.amp.autocast(dtype=dtype):
[TrainingMemory] {"rank": "1", "stage": "before_logging", "step": 3, "allocated_GiB": 24.229, "reserved_GiB": 71.883, "peak_allocated_GiB": 54.525, "peak_reserved_GiB": 71.883, "device_free_GiB": 1.04, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "2", "stage": "before_logging", "step": 3, "allocated_GiB": 24.192, "reserved_GiB": 71.924, "peak_allocated_GiB": 54.035, "peak_reserved_GiB": 71.924, "device_free_GiB": 0.999, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "0", "stage": "before_logging", "step": 3, "allocated_GiB": 24.22, "reserved_GiB": 65.676, "peak_allocated_GiB": 54.328, "peak_reserved_GiB": 65.676, "device_free_GiB": 7.247, "device_total_GiB": 79.189}

 25%|██▌       | 3/12 [02:13<05:54, 39.36s/it][TrainingMemory] {"rank": "3", "stage": "before_logging", "step": 3, "allocated_GiB": 24.21, "reserved_GiB": 63.305, "peak_allocated_GiB": 54.124, "peak_reserved_GiB": 63.305, "device_free_GiB": 9.62, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "3", "stage": "before_logging", "step": 4, "allocated_GiB": 24.173, "reserved_GiB": 65.004, "peak_allocated_GiB": 54.289, "peak_reserved_GiB": 65.004, "device_free_GiB": 7.919, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "1", "stage": "before_logging", "step": 4, "allocated_GiB": 24.163, "reserved_GiB": 71.883, "peak_allocated_GiB": 54.33, "peak_reserved_GiB": 71.883, "device_free_GiB": 1.04, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "0", "stage": "before_logging", "step": 4, "allocated_GiB": 24.163, "reserved_GiB": 65.676, "peak_allocated_GiB": 53.39, "peak_reserved_GiB": 65.676, "device_free_GiB": 7.247, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "2", "stage": "before_logging", "step": 4, "allocated_GiB": 24.172, "reserved_GiB": 71.924, "peak_allocated_GiB": 54.395, "peak_reserved_GiB": 71.924, "device_free_GiB": 0.999, "device_total_GiB": 79.189}

 33%|███▎      | 4/12 [02:37<04:26, 33.30s/it][TrainingMemory] {"rank": "3", "stage": "before_logging", "step": 5, "allocated_GiB": 24.182, "reserved_GiB": 65.004, "peak_allocated_GiB": 54.512, "peak_reserved_GiB": 65.004, "device_free_GiB": 7.919, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "2", "stage": "before_logging", "step": 5, "allocated_GiB": 24.192, "reserved_GiB": 71.924, "peak_allocated_GiB": 54.676, "peak_reserved_GiB": 71.924, "device_free_GiB": 0.999, "device_total_GiB": 79.189}[TrainingMemory] {"rank": "1", "stage": "before_logging", "step": 5, "allocated_GiB": 24.163, "reserved_GiB": 71.883, "peak_allocated_GiB": 54.514, "peak_reserved_GiB": 71.883, "device_free_GiB": 1.04, "device_total_GiB": 79.189}

[TrainingMemory] {"rank": "0", "stage": "before_logging", "step": 5, "allocated_GiB": 24.163, "reserved_GiB": 65.676, "peak_allocated_GiB": 53.982, "peak_reserved_GiB": 65.676, "device_free_GiB": 7.247, "device_total_GiB": 79.189}

 42%|████▏     | 5/12 [02:53<03:08, 26.86s/it][TrainingMemory] {"rank": "2", "stage": "before_logging", "step": 6, "allocated_GiB": 24.182, "reserved_GiB": 71.924, "peak_allocated_GiB": 54.164, "peak_reserved_GiB": 71.924, "device_free_GiB": 0.999, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "0", "stage": "before_logging", "step": 6, "allocated_GiB": 24.201, "reserved_GiB": 65.676, "peak_allocated_GiB": 53.583, "peak_reserved_GiB": 65.676, "device_free_GiB": 7.247, "device_total_GiB": 79.189}

 50%|█████     | 6/12 [03:08<02:17, 22.90s/it][TrainingMemory] {"rank": "1", "stage": "before_logging", "step": 6, "allocated_GiB": 24.201, "reserved_GiB": 71.883, "peak_allocated_GiB": 54.416, "peak_reserved_GiB": 71.883, "device_free_GiB": 1.04, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "3", "stage": "before_logging", "step": 6, "allocated_GiB": 24.192, "reserved_GiB": 65.004, "peak_allocated_GiB": 54.063, "peak_reserved_GiB": 65.004, "device_free_GiB": 7.919, "device_total_GiB": 79.189}
/home/dungpq6/Project/SpatialStack/src/qwen_vl/model/geometry_encoders/vggt_encoder.py:337: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
  with torch.cuda.amp.autocast(dtype=dtype):
/home/dungpq6/Project/SpatialStack/src/qwen_vl/model/geometry_encoders/vggt_encoder.py:337: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
  with torch.cuda.amp.autocast(dtype=dtype):
[TrainingMemory] {"rank": "3", "stage": "before_logging", "step": 7, "allocated_GiB": 24.182, "reserved_GiB": 65.004, "peak_allocated_GiB": 54.513, "peak_reserved_GiB": 65.004, "device_free_GiB": 7.919, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "0", "stage": "before_logging", "step": 7, "allocated_GiB": 24.211, "reserved_GiB": 65.676, "peak_allocated_GiB": 53.903, "peak_reserved_GiB": 65.676, "device_free_GiB": 7.247, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "2", "stage": "before_logging", "step": 7, "allocated_GiB": 24.22, "reserved_GiB": 71.924, "peak_allocated_GiB": 54.89, "peak_reserved_GiB": 71.924, "device_free_GiB": 0.999, "device_total_GiB": 79.189}

 58%|█████▊    | 7/12 [03:23<01:42, 20.45s/it][TrainingMemory] {"rank": "1", "stage": "before_logging", "step": 7, "allocated_GiB": 24.201, "reserved_GiB": 71.883, "peak_allocated_GiB": 54.855, "peak_reserved_GiB": 71.883, "device_free_GiB": 1.038, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "1", "stage": "before_logging", "step": 8, "allocated_GiB": 24.192, "reserved_GiB": 71.883, "peak_allocated_GiB": 55.349, "peak_reserved_GiB": 71.883, "device_free_GiB": 1.038, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "0", "stage": "before_logging", "step": 8, "allocated_GiB": 24.191, "reserved_GiB": 66.631, "peak_allocated_GiB": 55.498, "peak_reserved_GiB": 66.631, "device_free_GiB": 6.292, "device_total_GiB": 79.189}

 67%|██████▋   | 8/12 [04:03<01:46, 26.65s/it][TrainingMemory] {"rank": "3", "stage": "before_logging", "step": 8, "allocated_GiB": 24.182, "reserved_GiB": 65.004, "peak_allocated_GiB": 54.43, "peak_reserved_GiB": 65.004, "device_free_GiB": 7.919, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "2", "stage": "before_logging", "step": 8, "allocated_GiB": 24.201, "reserved_GiB": 71.924, "peak_allocated_GiB": 55.012, "peak_reserved_GiB": 71.924, "device_free_GiB": 0.999, "device_total_GiB": 79.189}
/home/dungpq6/Project/SpatialStack/src/qwen_vl/model/geometry_encoders/vggt_encoder.py:337: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
  with torch.cuda.amp.autocast(dtype=dtype):
[TrainingMemory] {"rank": "1", "stage": "before_logging", "step": 9, "allocated_GiB": 24.22, "reserved_GiB": 71.883, "peak_allocated_GiB": 55.188, "peak_reserved_GiB": 71.883, "device_free_GiB": 1.038, "device_total_GiB": 79.189}[TrainingMemory] {"rank": "2", "stage": "before_logging", "step": 9, "allocated_GiB": 24.229, "reserved_GiB": 71.924, "peak_allocated_GiB": 54.762, "peak_reserved_GiB": 71.924, "device_free_GiB": 0.999, "device_total_GiB": 79.189}

[TrainingMemory] {"rank": "0", "stage": "before_logging", "step": 9, "allocated_GiB": 24.182, "reserved_GiB": 66.631, "peak_allocated_GiB": 53.989, "peak_reserved_GiB": 66.631, "device_free_GiB": 6.292, "device_total_GiB": 79.189}

 75%|███████▌  | 9/12 [04:19<01:09, 23.24s/it][TrainingMemory] {"rank": "3", "stage": "before_logging", "step": 9, "allocated_GiB": 24.249, "reserved_GiB": 65.848, "peak_allocated_GiB": 54.738, "peak_reserved_GiB": 65.848, "device_free_GiB": 7.075, "device_total_GiB": 79.189}
/home/dungpq6/Project/SpatialStack/src/qwen_vl/model/geometry_encoders/vggt_encoder.py:337: FutureWarning: `torch.cuda.amp.autocast(args...)` is deprecated. Please use `torch.amp.autocast('cuda', args...)` instead.
  with torch.cuda.amp.autocast(dtype=dtype):
[TrainingMemory] {"rank": "3", "stage": "before_logging", "step": 10, "allocated_GiB": 24.201, "reserved_GiB": 65.848, "peak_allocated_GiB": 54.159, "peak_reserved_GiB": 65.848, "device_free_GiB": 7.075, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "2", "stage": "before_logging", "step": 10, "allocated_GiB": 24.192, "reserved_GiB": 71.924, "peak_allocated_GiB": 54.212, "peak_reserved_GiB": 71.924, "device_free_GiB": 0.999, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "1", "stage": "before_logging", "step": 10, "allocated_GiB": 24.211, "reserved_GiB": 71.883, "peak_allocated_GiB": 54.644, "peak_reserved_GiB": 71.883, "device_free_GiB": 1.038, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "0", "stage": "before_logging", "step": 10, "allocated_GiB": 24.21, "reserved_GiB": 66.631, "peak_allocated_GiB": 53.92, "peak_reserved_GiB": 66.631, "device_free_GiB": 6.29, "device_total_GiB": 79.189}

 83%|████████▎ | 10/12 [04:34<00:41, 20.85s/it][TrainingMemory] {"rank": "1", "stage": "after_logging", "step": 10, "allocated_GiB": 24.211, "reserved_GiB": 71.883, "peak_allocated_GiB": 54.644, "peak_reserved_GiB": 71.883, "device_free_GiB": 0.565, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "0", "stage": "after_logging", "step": 10, "allocated_GiB": 24.21, "reserved_GiB": 66.631, "peak_allocated_GiB": 53.92, "peak_reserved_GiB": 66.631, "device_free_GiB": 5.817, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "2", "stage": "after_logging", "step": 10, "allocated_GiB": 24.192, "reserved_GiB": 71.924, "peak_allocated_GiB": 54.212, "peak_reserved_GiB": 71.924, "device_free_GiB": 0.526, "device_total_GiB": 79.189}

                                               
{'loss': '0.9433', 'grad_norm': '16.32', 'learning_rate': '3.451e-06', 'epoch': '0.0002634'}
[TrainingMemory] {"rank": "3", "stage": "after_logging", "step": 10, "allocated_GiB": 24.201, "reserved_GiB": 65.848, "peak_allocated_GiB": 54.159, "peak_reserved_GiB": 65.848, "device_free_GiB": 6.602, "device_total_GiB": 79.189}

 83%|████████▎ | 10/12 [04:35<00:41, 20.85s/it][TrainingMemory] {"rank": "2", "stage": "before_logging", "step": 11, "allocated_GiB": 24.144, "reserved_GiB": 71.924, "peak_allocated_GiB": 54.433, "peak_reserved_GiB": 71.924, "device_free_GiB": 0.524, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "1", "stage": "before_logging", "step": 11, "allocated_GiB": 24.144, "reserved_GiB": 71.883, "peak_allocated_GiB": 54.489, "peak_reserved_GiB": 71.883, "device_free_GiB": 0.565, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "0", "stage": "before_logging", "step": 11, "allocated_GiB": 24.144, "reserved_GiB": 66.631, "peak_allocated_GiB": 53.685, "peak_reserved_GiB": 66.631, "device_free_GiB": 5.817, "device_total_GiB": 79.189}

 92%|█████████▏| 11/12 [04:50<00:19, 19.24s/it][TrainingMemory] {"rank": "3", "stage": "before_logging", "step": 11, "allocated_GiB": 24.154, "reserved_GiB": 65.848, "peak_allocated_GiB": 54.197, "peak_reserved_GiB": 65.848, "device_free_GiB": 6.602, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "3", "stage": "before_logging", "step": 12, "allocated_GiB": 24.191, "reserved_GiB": 65.848, "peak_allocated_GiB": 54.318, "peak_reserved_GiB": 65.848, "device_free_GiB": 6.602, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "2", "stage": "before_logging", "step": 12, "allocated_GiB": 24.182, "reserved_GiB": 71.924, "peak_allocated_GiB": 54.444, "peak_reserved_GiB": 71.924, "device_free_GiB": 0.524, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "1", "stage": "before_logging", "step": 12, "allocated_GiB": 24.192, "reserved_GiB": 71.883, "peak_allocated_GiB": 54.548, "peak_reserved_GiB": 71.883, "device_free_GiB": 0.565, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "0", "stage": "before_logging", "step": 12, "allocated_GiB": 24.173, "reserved_GiB": 66.631, "peak_allocated_GiB": 53.641, "peak_reserved_GiB": 66.631, "device_free_GiB": 5.815, "device_total_GiB": 79.189}

100%|██████████| 12/12 [05:05<00:00, 18.07s/it]

Writing model shards:   0%|          | 0/1 [00:00<?, ?it/s]

Writing model shards: 100%|██████████| 1/1 [00:10<00:00, 10.10s/it]
Writing model shards: 100%|██████████| 1/1 [00:10<00:00, 10.10s/it]
[TrainingMemory] {"rank": "3", "stage": "after_logging", "step": 12, "allocated_GiB": 22.888, "reserved_GiB": 65.848, "peak_allocated_GiB": 54.318, "peak_reserved_GiB": 65.848, "device_free_GiB": 6.602, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "0", "stage": "after_logging", "step": 12, "allocated_GiB": 22.888, "reserved_GiB": 66.631, "peak_allocated_GiB": 53.641, "peak_reserved_GiB": 66.631, "device_free_GiB": 5.815, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "1", "stage": "after_logging", "step": 12, "allocated_GiB": 22.888, "reserved_GiB": 71.883, "peak_allocated_GiB": 54.548, "peak_reserved_GiB": 71.883, "device_free_GiB": 0.565, "device_total_GiB": 79.189}
[TrainingMemory] {"rank": "2", "stage": "after_logging", "step": 12, "allocated_GiB": 22.888, "reserved_GiB": 71.924, "peak_allocated_GiB": 54.444, "peak_reserved_GiB": 71.924, "device_free_GiB": 0.524, "device_total_GiB": 79.189}

                                               
{'train_runtime': '359.2', 'train_samples_per_second': '2.138', 'train_steps_per_second': '0.033', 'train_loss': '0.8403', 'epoch': '0.0003161'}

100%|██████████| 12/12 [05:59<00:00, 18.07s/it]
100%|██████████| 12/12 [05:59<00:00, 29.93s/it]

Writing model shards:   0%|          | 0/1 [00:00<?, ?it/s]
Writing model shards: 100%|██████████| 1/1 [00:10<00:00, 10.29s/it]
Writing model shards: 100%|██████████| 1/1 [00:10<00:00, 10.29s/it]
[rank0]:[W918 03:56:59.052849812 ProcessGroupNCCL.cpp:1476] Warning: WARNING: destroy_process_group() was not called before program exit, which can leak resources. For more info, please see https://pytorch.org/docs/stable/distributed.html#shutdown (function operator())