_wandb:
    value:
        cli_version: 0.25.1
        e:
            7kvqf9dgct7g9tfvelr7td16kru540n5:
                args:
                    - --config
                    - configs/block_ar_owt.yaml
                    - --num_steps
                    - "200"
                codePath: scripts/train_block_ar.py
                codePathLocal: scripts/train_block_ar.py
                cpu_count: 64
                cpu_count_logical: 128
                cudaVersion: "13.1"
                disk:
                    /:
                        total: "454503358464"
                        used: "76717228032"
                email: 13815761302s@gmail.com
                executable: /home/sunhc/miniconda3/envs/sad/bin/python3.11
                gpu: NVIDIA A100-SXM4-80GB
                gpu_count: 8
                gpu_nvidia:
                    - architecture: Ampere
                      cudaCores: 6912
                      memoryTotal: "85899345920"
                      name: NVIDIA A100-SXM4-80GB
                      uuid: GPU-4673371e-58d7-bd92-556f-ad5559a569b1
                    - architecture: Ampere
                      cudaCores: 6912
                      memoryTotal: "85899345920"
                      name: NVIDIA A100-SXM4-80GB
                      uuid: GPU-939e070a-4474-5df6-febb-bed4d500eadc
                    - architecture: Ampere
                      cudaCores: 6912
                      memoryTotal: "85899345920"
                      name: NVIDIA A100-SXM4-80GB
                      uuid: GPU-b3ccc59d-bb3c-a911-3e9b-268418fc25f5
                    - architecture: Ampere
                      cudaCores: 6912
                      memoryTotal: "85899345920"
                      name: NVIDIA A100-SXM4-80GB
                      uuid: GPU-3728ad73-65a6-f4f1-0e0b-11f847c7066d
                    - architecture: Ampere
                      cudaCores: 6912
                      memoryTotal: "85899345920"
                      name: NVIDIA A100-SXM4-80GB
                      uuid: GPU-aacb1338-c95c-1a54-c4f7-cd1845a85816
                    - architecture: Ampere
                      cudaCores: 6912
                      memoryTotal: "85899345920"
                      name: NVIDIA A100-SXM4-80GB
                      uuid: GPU-8a6bc98e-a464-71bc-24d1-f29bf3c3e5c6
                    - architecture: Ampere
                      cudaCores: 6912
                      memoryTotal: "85899345920"
                      name: NVIDIA A100-SXM4-80GB
                      uuid: GPU-6e961bbe-6ea9-b79e-ba53-4eec0562962c
                    - architecture: Ampere
                      cudaCores: 6912
                      memoryTotal: "85899345920"
                      name: NVIDIA A100-SXM4-80GB
                      uuid: GPU-040a80ab-3e8b-b667-260e-e8308d157872
                host: hpc
                memory:
                    total: "1081804640256"
                os: Linux-6.8.0-106-generic-x86_64-with-glibc2.39
                program: /home/sunhc/diffusion/sad/scripts/train_block_ar.py
                python: CPython 3.11.15
                root: /home/sunhc/diffusion/sad
                startedAt: "2026-04-17T08:45:24.711983Z"
                writerId: 7kvqf9dgct7g9tfvelr7td16kru540n5
        m: []
        python_version: 3.11.15
        t:
            "1":
                - 1
                - 5
                - 11
                - 49
                - 51
                - 53
                - 71
            "2":
                - 1
                - 5
                - 11
                - 49
                - 51
                - 53
                - 71
            "3":
                - 16
            "4": 3.11.15
            "5": 0.25.1
            "6": 5.5.3
            "12": 0.25.1
            "13": linux-x86_64
ancestor:
    value:
        lut_path: data/ancestor_lut_50257-128_top3_t1.0.pt
        proto_path: data/hierarchy_prototypes_50257-128.pt
data:
    value:
        cache_dir: data/owt_cache
        dataset: openwebtext
        mode: subsample
        num_workers: 4
        seq_len: 512
logging:
    value:
        project: sad_block_ar
        save_dir: outputs/sad_block_ar
        use_wandb: true
loss:
    value:
        lambda_ancestor: 0
        mask_only: false
model:
    value:
        block_size: 8
        cond_dim: 128
        dropout: 0
        hidden_size: 768
        level_sizes:
            - 50258
            - 128
        max_seq_len: 512
        n_blocks: 12
        n_heads: 12
        num_levels: 2
        type: block_ar
        vocab_size: 50258
training:
    value:
        batch_size: 64
        dtype: bf16
        eval_interval: 5000
        grad_clip: 1
        log_interval: 100
        lr: 0.0003
        num_steps: 200
        save_interval: 10000
        seed: 0
        weight_decay: 0.01