Add Flappy WanOFT checkpoint and fixed multi-image eval

Upload the 2000-step Flappy WanOFT checkpoint run with the corrected five-image latency-bench evaluation under eval/.

wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:898e0b3e19c6d07987be834b503b315688dda724ffa25e5642c16c1cb316c660
+size 25211531190

wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/config.full.yaml

@@ -0,0 +1,277 @@
+framework:
+  name: WanOFT
+  world_model:
+    base_wm: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/playground/Pretrained_models/Wan-AI/Wan2.2-TI2V-5B-Diffusers
+    extract_layers:
+    - -1
+    num_frames: 5
+  qwenvl:
+    base_vlm: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/playground/Pretrained_models/Wan-AI/Wan2.2-TI2V-5B-Diffusers
+    attn_implementation: sdpa
+    flex_backend: triton
+    enable_gradient_checkpointing: true
+  action_model:
+    action_model_type: MLP
+    action_dim: 7
+    action_hidden_dim: 3072
+    action_horizon: 8
+    future_action_window_size: 7
+    past_action_window_size: 0
+    loss_type: current_discrete_ce
+    class_weights: null
+    future_loss_weight: null
+    action_query_source: mean
+    action_query_num_heads: 24
+    state_dim: 7
+    action_env_dim: 2
+  kv_memory:
+    enabled: false
+    window: 4
+    rollout_len: 8
+    packed_train: false
+    rebased_sink: true
+datasets:
+  vla_data:
+    dataset_py: lerobot_datasets
+    include_state: true
+    data_root_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/data/flappy_fix_latency_3_200ep_context5
+    data_mix: flappy_train__bridge
+    eval_data_mix: null
+    custom_mixtures_path: null
+    action_type: discrete
+    sequential_step_sampling: false
+    eval_sequential_step_sampling: null
+    num_workers: 8
+    eval_num_workers: 8
+    prefetch_factor: 4
+    persistent_workers: true
+    pin_memory: true
+    shuffle: true
+    action_balance:
+      enabled: false
+      strategy: balanced_epoch
+      action_key: action_id
+      target_flap_fraction: 0.3
+      noop_id: 0
+      flap_id: 1
+    latency_curriculum:
+      enabled: false
+      strategy: exclusive
+      latencies: null
+      phase_steps: null
+      phase_distributions: null
+      new_latency_passes: 1.0
+      replay_passes: 0.25
+      target_total_passes: 2.0
+      final_equalization: true
+      step_budget_mode: auto
+      eval_at_phase_end: false
+      save_at_phase_end: false
+      computed_plan: null
+    per_device_batch_size: 1
+    load_all_data_for_training: true
+    num_obs_frames: 1
+    image_mode: single
+    prompt_mode: raw
+    stitch_grid:
+    - 2
+    - 2
+    obs_image_size:
+    - 224
+    - 224
+    video_backend: torchvision_av
+    observation_indices:
+    - -4
+    - -3
+    - -2
+    - -1
+    - 0
+    language_indices:
+    - 0
+    state_indices:
+    - 0
+    action_indices:
+    - 0
+    - 1
+    - 2
+    - 3
+    - 4
+    - 5
+    - 6
+    - 7
+    pack_image_sequence: true
+    image_sequence_length: 5
+    context_images_column: observation.context_images
+dataset:
+  source_hf: latency-sensitive-bench/memory-rollouts
+  config_name: flappy_fixed_latency_3_200ep_7k2steps
+  source_subdir: null
+  converted_name: flappy_train__bridge
+  single_source_hf: ''
+  mixed_source_hf: ''
+  single_converted_name: flappy_train__bridge
+  mixed_converted_name: flappy_mixed_latency_train__bridge
+  single_latency_filter: null
+  mixed_latency_filter: null
+  force_download: false
+  setup_force: false
+  skip_verification: false
+  target_latency_unit: observation_steps
+  verify_rows: 200
+  max_episodes: 200
+  episodes_per_latency: null
+  latency_filter: null
+  debug_subset:
+    enabled: false
+    max_episodes: 5
+    suffix: debug
+base_model:
+  repo_id: Wan-AI/Wan2.2-TI2V-5B-Diffusers
+initialization:
+  checkpoint_local_dir: playground/Pretrained_models/WM4A-Wan2d2-OFT-LIBERO-4in1
+  checkpoint_hf_repo_id: StarVLA/WM4A-Wan2d2-OFT-LIBERO-4in1
+  checkpoint_filename: checkpoints/steps_60000_pytorch_model.pt
+trainer:
+  max_train_steps: 2000
+  num_warmup_steps: 0
+  save_interval: 500
+  eval_interval: 100
+  eval_num_batches: 100
+  per_latency_eval_num_batches: null
+  eval_action_classification: false
+  eval_action_classification_interval: null
+  cc_f1_tolerance: 1
+  learning_rate:
+    base: 2.5e-05
+    qwen_vl_interface: 1.0e-05
+    action_model: 0.0001
+    action_query_proj: 0.0001
+  lr_scheduler_type: cosine_with_min_lr
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+  freeze_modules: ''
+  freeze_vit: false
+  freeze_tied_embedding: false
+  freeze_llm_layers: []
+  loss_scale:
+    vla: 1.0
+    vlm: 0.1
+  max_grad_norm: 1.0
+  weight_decay: 0.0
+  logging_frequency: 1
+  profile_timing:
+    enabled: false
+    log_interval: 10
+  gradient_clipping: 1.0
+  gradient_accumulation_steps: 16
+  distributed_backend: deepspeed
+  is_resume: false
+  pretrained_checkpoint: null
+  resume_step: 0
+  reload_modules: null
+  optimizer:
+    name: AdamW
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    weight_decay: 1.0e-08
+    fused: true
+  save_format: pt
+workspace_dir: WORKSPACE_DIR
+run_root_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints
+seed: 42
+wandb_entity: zihanwang-ai-northwestern-university
+wandb_project: starVLA_rl_games
+auth:
+  env_file: null
+  hf_token_env: HF_TOKEN
+  wandb_api_key_env: WANDB_API_KEY
+paths:
+  run_root_dir: results/Checkpoints
+  dataset_local_dir: data/flappy_fix_latency_3_200ep_context5
+  dataset_cache_dir: null
+  base_model_dir: playground/Pretrained_models/Wan-AI/Wan2.2-TI2V-5B-Diffusers
+  accelerate_config: starVLA/config/deepseeds/deepspeed_zero2.yaml
+launch:
+  use_accelerate: true
+  gpus: null
+  num_processes: 1
+  dry_run: false
+conda:
+  enabled: true
+  env_name: null
+rl_games:
+  model_alias: wan_oft
+  env_eval:
+    image_size: 224
+    frameskip: 1
+    image_transform: raw_rgb
+    prompt_mode: raw
+    ghost_trail:
+      history_frames: 5
+      gamma: 1.3
+      min_alpha: 35
+      scroll_px_per_step: 4.0
+      ground_fraction: 0.22
+    seed: 42
+    fixed_episode_seeds: true
+    latency_seed_stride: 0
+    task_seed_stride: 0
+    task_description: ''
+    eval_parallel_envs: 5
+    action_chunk_execution:
+      enabled: false
+      chunk_size: null
+    enabled: true
+    eval_backend: eval_core
+    distributed_mode: rank_sharded
+    vectorized:
+      enabled: false
+      batch_size: 1
+    latency:
+      prompt_map_path: data/flappy_fix_latency_3_200ep_context5/flappy_train__bridge/latency_prompt_map.json
+      mode: single
+      values:
+      - 0
+    mid_train:
+      enabled: false
+      interval_steps: 100
+      latencies:
+      - 0
+      num_episodes: 5
+      max_steps_per_episode: 3600
+    post_train:
+      enabled: true
+      latencies:
+      - 3
+      num_episodes: 20
+      max_steps_per_episode: 3600
+  task: flappy
+  initialization_mode: bridge
+  action_carrier: bridge
+model: wan_oft
+env: flappy
+init: wan_oft_libero
+mode: single
+checkpoint:
+  load: auto
+  hf_repo_id: null
+  save_best_model: false
+  save_final_model: true
+  save_pt_file: true
+  save_training_state: true
+  save_safetensors_file: false
+  local:
+    keep_last_n: 1
+  sync:
+    enabled: false
+    repo_id: null
+    keep_last_n: 0
+    sync_every_n_checkpoints: 1
+    resume_policy: local_latest
+run_id: wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce
+output_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce
+config_yaml: null
+is_debug: false
+version_id: '0.21'

wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/config.yaml

@@ -0,0 +1,89 @@
+checkpoint:
+  local:
+    keep_last_n: 1
+  save_best_model: false
+  save_final_model: true
+  save_pt_file: true
+  save_safetensors_file: false
+  save_training_state: true
+  sync:
+    enabled: false
+    keep_last_n: 0
+    repo_id: null
+datasets:
+  vla_data:
+    data_mix: flappy_train__bridge
+    dataset_py: lerobot_datasets
+    eval_data_mix: null
+    latency_curriculum:
+      enabled: false
+    obs_image_size:
+    - 224
+    - 224
+    per_device_batch_size: 1
+framework:
+  action_model:
+    action_dim: 7
+    action_env_dim: 2
+    action_hidden_dim: 3072
+    action_horizon: 8
+    action_model_type: MLP
+    action_query_source: mean
+    class_weights: null
+    loss_type: current_discrete_ce
+  name: WanOFT
+  qwenvl:
+    base_vlm: /workspace/runtime/models/Wan2.2-TI2V-5B-Diffusers
+  world_model:
+    base_wm: /workspace/runtime/models/Wan2.2-TI2V-5B-Diffusers
+    extract_layers:
+    - -1
+    num_frames: 5
+output_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce
+rl_games:
+  env_eval:
+    enabled: true
+    eval_backend: eval_core
+    mid_train:
+      enabled: false
+      interval_steps: 100
+    prompt_mode: raw
+run_id: wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce
+run_root_dir: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/wanoft_bench/third_party/starVLA/results/Checkpoints
+seed: 42
+trainer:
+  distributed_backend: deepspeed
+  eval_action_classification: false
+  eval_action_classification_interval: null
+  eval_interval: 100
+  freeze_llm_layers: []
+  freeze_modules: ''
+  freeze_tied_embedding: false
+  freeze_vit: false
+  gradient_accumulation_steps: 16
+  is_resume: false
+  learning_rate:
+    action_model: 0.0001
+    action_query_proj: 0.0001
+    base: 2.5e-05
+    qwen_vl_interface: 1.0e-05
+  logging_frequency: 1
+  lr_scheduler_type: cosine_with_min_lr
+  max_train_steps: 2000
+  num_warmup_steps: 0
+  optimizer:
+    betas:
+    - 0.9
+    - 0.95
+    eps: 1.0e-08
+    fused: true
+    weight_decay: 1.0e-08
+  pretrained_checkpoint: null
+  profile_timing:
+    enabled: false
+  reload_modules: null
+  save_interval: 500
+  scheduler_specific_kwargs:
+    min_lr: 1.0e-06
+wandb_entity: zihanwang-ai-northwestern-university
+wandb_project: starVLA_rl_games

wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/dataset_statistics.json

@@ -0,0 +1,127 @@
+{
+  "new_embodiment": {
+    "action": {
+      "mean": [
+        0.8235060572624207,
+        0.17649392783641815,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.3809901773929596,
+        0.3809901773929596,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        1.0,
+        1.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        1.0,
+        1.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "mask": [
+        true,
+        true,
+        true,
+        true,
+        true,
+        true,
+        true
+      ]
+    },
+    "state": {
+      "mean": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "std": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "max": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "min": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q01": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ],
+      "q99": [
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0,
+        0.0
+      ]
+    },
+    "num_transitions": 1393742,
+    "num_trajectories": 198
+  }
+}

wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/eval_latency_3/episode_metrics.jsonl

@@ -0,0 +1,5 @@
+{"episode_id": 0, "episode_return": 444.6000052243471, "episode_return_env": 444.6000052243471, "game_score": 94.0, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "config_name": "starvla_wan_oft_flappy_multiframe_fix_5ep", "configured_latency_ms": 100.0, "env_fps": 30.0, "env_id": "flappy", "episode_seed": 42, "eval_latency_steps": 3, "frame_ms": 33.333333333333336, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 30.0, "output_dir": "/workspace/evals/flappy_wanoft_latency_multiframe_fix_5ep/fixed_3/eval_latency_3", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_flappy_multiframe_fix_5ep", "suite_name": "fixed_3", "target_device": null, "train_latency_steps": 3}, "num_actions": 3600, "num_dropped_actions": 2, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 3600}
+{"episode_id": 1, "episode_return": 444.6000052243471, "episode_return_env": 444.6000052243471, "game_score": 94.0, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "config_name": "starvla_wan_oft_flappy_multiframe_fix_5ep", "configured_latency_ms": 100.0, "env_fps": 30.0, "env_id": "flappy", "episode_seed": 43, "eval_latency_steps": 3, "frame_ms": 33.333333333333336, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 30.0, "output_dir": "/workspace/evals/flappy_wanoft_latency_multiframe_fix_5ep/fixed_3/eval_latency_3", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_flappy_multiframe_fix_5ep", "suite_name": "fixed_3", "target_device": null, "train_latency_steps": 3}, "num_actions": 3600, "num_dropped_actions": 2, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 3600}
+{"episode_id": 2, "episode_return": 444.6000052243471, "episode_return_env": 444.6000052243471, "game_score": 94.0, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "config_name": "starvla_wan_oft_flappy_multiframe_fix_5ep", "configured_latency_ms": 100.0, "env_fps": 30.0, "env_id": "flappy", "episode_seed": 44, "eval_latency_steps": 3, "frame_ms": 33.333333333333336, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 30.0, "output_dir": "/workspace/evals/flappy_wanoft_latency_multiframe_fix_5ep/fixed_3/eval_latency_3", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_flappy_multiframe_fix_5ep", "suite_name": "fixed_3", "target_device": null, "train_latency_steps": 3}, "num_actions": 3600, "num_dropped_actions": 2, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 3600}
+{"episode_id": 3, "episode_return": 444.6000052243471, "episode_return_env": 444.6000052243471, "game_score": 94.0, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "config_name": "starvla_wan_oft_flappy_multiframe_fix_5ep", "configured_latency_ms": 100.0, "env_fps": 30.0, "env_id": "flappy", "episode_seed": 45, "eval_latency_steps": 3, "frame_ms": 33.333333333333336, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 30.0, "output_dir": "/workspace/evals/flappy_wanoft_latency_multiframe_fix_5ep/fixed_3/eval_latency_3", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_flappy_multiframe_fix_5ep", "suite_name": "fixed_3", "target_device": null, "train_latency_steps": 3}, "num_actions": 3600, "num_dropped_actions": 2, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 3600}
+{"episode_id": 4, "episode_return": 444.6000052243471, "episode_return_env": 444.6000052243471, "game_score": 94.0, "mean_latency_ms": 100.0, "metadata": {"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "config_name": "starvla_wan_oft_flappy_multiframe_fix_5ep", "configured_latency_ms": 100.0, "env_fps": 30.0, "env_id": "flappy", "episode_seed": 46, "eval_latency_steps": 3, "frame_ms": 33.333333333333336, "instance_id": null, "instance_origin": null, "latency_type": "fixed_sweep", "mode": "simulated", "obs_fps": 30.0, "output_dir": "/workspace/evals/flappy_wanoft_latency_multiframe_fix_5ep/fixed_3/eval_latency_3", "policy_id": "starvla", "profile_id": null, "profile_path": null, "replay_host": null, "run_name": "starvla_wan_oft_flappy_multiframe_fix_5ep", "suite_name": "fixed_3", "target_device": null, "train_latency_steps": 3}, "num_actions": 3600, "num_dropped_actions": 2, "num_invalid_actions": 0, "p90_latency_ms": 100.0, "p99_latency_ms": 100.0, "return_raw": null, "survival_steps": 3600}

wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/queue_eval_latency_3.json

@@ -0,0 +1,29 @@
+{
+  "checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt",
+  "configured_latency_ms": 100.0,
+  "eval_latency_steps": 3,
+  "experiment_name": "starvla_wan_oft_flappy_multiframe_fix_5ep",
+  "latency": 3,
+  "latency_type": "fixed_sweep",
+  "lengths": [
+    3600,
+    3600,
+    3600,
+    3600,
+    3600
+  ],
+  "mean_length": 3600.0,
+  "mean_return": 444.6000052243471,
+  "returns": [
+    444.6000052243471,
+    444.6000052243471,
+    444.6000052243471,
+    444.6000052243471,
+    444.6000052243471
+  ],
+  "seed": 42,
+  "std_return": 0.0,
+  "suite_name": "fixed_3",
+  "timestamp_utc": "2026-07-28T20:27:16.094854+00:00",
+  "train_latency_steps": 3
+}

wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/eval/fixed_3/queue_eval_results.jsonl

@@ -0,0 +1 @@
+{"checkpoint_path": "/workspace/runtime/wanoft/wan_oft_flappy_fix_latency_3_context5_standard_sft_2000_effbs128_224_currentce/checkpoints/steps_2000_pytorch_model.pt", "configured_latency_ms": 100.0, "eval_latency_steps": 3, "experiment_name": "starvla_wan_oft_flappy_multiframe_fix_5ep", "latency": 3, "latency_type": "fixed_sweep", "lengths": [3600, 3600, 3600, 3600, 3600], "mean_length": 3600.0, "mean_return": 444.6000052243471, "returns": [444.6000052243471, 444.6000052243471, 444.6000052243471, 444.6000052243471, 444.6000052243471], "seed": 42, "std_return": 0.0, "suite_name": "fixed_3", "timestamp_utc": "2026-07-28T20:27:16.094854+00:00", "train_latency_steps": 3}