Upload folder using huggingface_hub

.gitattributes

@@ -54,3 +54,5 @@ demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-202607
 demon_attack_fix_latency_6_200ep_7k2steps_single_baseline/wandb/wandb/run-20260728_071207-800jpcpi/run-800jpcpi.wandb filter=lfs diff=lfs merge=lfs -text
 demon_attack_fix_latency_6_200ep_7k2steps_stitch/wandb/wandb/run-20260728_070430-3spxusr3/run-3spxusr3.wandb filter=lfs diff=lfs merge=lfs -text
 demon_attack_fix_latency_6_200ep_7k2steps_stitch/wandb/wandb/run-20260728_110431-8pr3jqce/run-8pr3jqce.wandb filter=lfs diff=lfs merge=lfs -text
+deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/wandb/wandb/run-20260728_070431-c7pbunc4/run-c7pbunc4.wandb filter=lfs diff=lfs merge=lfs -text
+deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/wandb/wandb/run-20260728_110400-3oofo2bj/run-3oofo2bj.wandb filter=lfs diff=lfs merge=lfs -text

deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/checkpoints/steps_4000_model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:3ab804a90ca1f5065aae37e76482b478156a5e5f3785319c9e9bada56735321d
+oid sha256:379060d92ea71215680a35979c262ca84ef6f566468755d43691a5f49b282a46
 size 9784896838

deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/checkpoints/steps_4000_state/pytorch_model/mp_rank_00_model_states.pt

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:a92d97b4f220cdafc99774354ed9376dd41d2f8d4462ed4d7e63b82b603fd3ec
-size 9007152504
+oid sha256:e36f4d7457dc2681ecf6b339c5500f814589c84773c382d4a10ce3e86e33707b
+size 9007152933

deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/checkpoints/steps_4000_state/random_states_0.pkl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b390ddf3fbe806b45e0fc898f8e191f0159a1ec86215c9dfa474c0be64dae79b
-size 14604
+oid sha256:67641b8c15c37da31a4339368b57abd2bf66a9134a85298c3fb0931bc9a3bd76
+size 15017

deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/checkpoints/steps_4000_state/random_states_1.pkl

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:28f01d59656fdd858f84676b30750d7b3c90fbeaa8d581a5bf35f1991d9b2727
-size 14604
+oid sha256:14b3a1b686d8ebf7b279f50894d1943efa0d0541a685ee8e974d7b493dbc409b
+size 15017

deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/config.full.yaml

@@ -3,7 +3,6 @@ framework:
   qwenvl:
     base_vlm: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
     attn_implementation: flash_attention_2
-    flex_backend: triton
     enable_gradient_checkpointing: true
   action_model:
     action_model_type: MLP

deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/hydra/.hydra/config.yaml

@@ -2,7 +2,6 @@ framework:
   qwenvl:
     base_vlm: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
     attn_implementation: flash_attention_2
-    flex_backend: triton
     enable_gradient_checkpointing: true
   action_model:
     state_dim: 7

deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/hydra/.hydra/hydra.yaml

@@ -118,7 +118,6 @@ hydra:
     - mode=single
     - ++framework.qwenvl.base_vlm=playground/Pretrained_models/Qwen3-VL-4B-Instruct
     - ++framework.qwenvl.attn_implementation=flash_attention_2
-    - ++framework.qwenvl.flex_backend=triton
     - ++framework.qwenvl.enable_gradient_checkpointing=true
     - ++framework.action_model.state_dim=7
     - ++framework.action_model.loss_type=discrete_ce
@@ -330,7 +329,7 @@ hydra:
   job:
     name: train_starvla_hydra
     chdir: false
-    override_dirname: ++auth.env_file=null,++auth.hf_token_env=HF_TOKEN,++auth.wandb_api_key_env=WANDB_API_KEY,++base_model.repo_id=Qwen/Qwen3-VL-4B-Instruct,++bridge_base_model.local_dir.gr00t=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++bridge_base_model.local_dir.openvla=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++bridge_base_model.local_dir.pi05=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++bridge_base_model.local_dir.pi0=playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action,++bridge_base_model.repo_id.gr00t=Qwen/Qwen3-VL-4B-Instruct,++bridge_base_model.repo_id.openvla=Qwen/Qwen3-VL-4B-Instruct,++bridge_base_model.repo_id.pi05=Qwen/Qwen3-VL-4B-Instruct,++bridge_base_model.repo_id.pi0=StarVLA/Qwen2.5-VL-3B-Instruct-Action,++checkpoint.hf_repo_id=null,++checkpoint.load=auto,++checkpoint.local.keep_last_n=1,++checkpoint.save_best_model=false,++checkpoint.save_final_model=true,++checkpoint.save_pt_file=false,++checkpoint.save_safetensors_file=true,++checkpoint.save_training_state=true,++checkpoint.sync.enabled=false,++checkpoint.sync.keep_last_n=0,++checkpoint.sync.repo_id=null,++checkpoint.sync.resume_policy=local_latest,++checkpoint.sync.sync_every_n_checkpoints=1,++config_yaml=null,++dataset.config_name=null,++dataset.converted_name=deadly_corridor_train,++dataset.debug_subset.enabled=false,++dataset.debug_subset.max_episodes=5,++dataset.debug_subset.suffix=debug,++dataset.episodes_per_latency=null,++dataset.force_download=false,++dataset.latency_filter=null,++dataset.max_episodes=null,++dataset.mixed_converted_name=deadly_corridor_mixed_latency_train,++dataset.mixed_latency_filter=null,++dataset.mixed_source_hf=,++dataset.setup_force=false,++dataset.single_converted_name=deadly_corridor_train,++dataset.single_latency_filter=null,++dataset.single_source_hf=,++dataset.skip_verification=false,++dataset.source_hf=,++dataset.source_subdir=null,++dataset.target_latency_unit=raw_frames,++dataset.verify_rows=200,++datasets.vla_data.action_balance.action_key=action_id,++datasets.vla_data.action_balance.enabled=false,++datasets.vla_data.action_balance.flap_id=1,++datasets.vla_data.action_balance.noop_id=0,++datasets.vla_data.action_balance.strategy=balanced_epoch,++datasets.vla_data.action_balance.target_flap_fraction=0.3,++datasets.vla_data.action_type=discrete,++datasets.vla_data.custom_mixtures_path=null,++datasets.vla_data.data_mix=deadly_corridor_train,++datasets.vla_data.data_mix=deadly_corridor_train__bridge,++datasets.vla_data.data_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/deadly_corridor_fix_latency_6_1000ep_7k2steps,++datasets.vla_data.data_root_dir=playground/Datasets/rl_games,++datasets.vla_data.dataset_py=lerobot_datasets,++datasets.vla_data.eval_data_mix=deadly_corridor_train__bridge__val,++datasets.vla_data.eval_data_mix=null,++datasets.vla_data.eval_num_workers=8,++datasets.vla_data.eval_sequential_step_sampling=null,++datasets.vla_data.image_mode=stitch,++datasets.vla_data.include_state=true,++datasets.vla_data.latency_curriculum.computed_plan=null,++datasets.vla_data.latency_curriculum.enabled=false,++datasets.vla_data.latency_curriculum.eval_at_phase_end=false,++datasets.vla_data.latency_curriculum.final_equalization=true,++datasets.vla_data.latency_curriculum.latencies=null,++datasets.vla_data.latency_curriculum.new_latency_passes=1.0,++datasets.vla_data.latency_curriculum.phase_distributions=null,++datasets.vla_data.latency_curriculum.phase_steps=null,++datasets.vla_data.latency_curriculum.replay_passes=0.25,++datasets.vla_data.latency_curriculum.save_at_phase_end=false,++datasets.vla_data.latency_curriculum.step_budget_mode=auto,++datasets.vla_data.latency_curriculum.strategy=exclusive,++datasets.vla_data.latency_curriculum.target_total_passes=2.0,++datasets.vla_data.load_all_data_for_training=true,++datasets.vla_data.num_obs_frames=4,++datasets.vla_data.num_workers=8,++datasets.vla_data.obs_image_size=null,++datasets.vla_data.per_device_batch_size=8,++datasets.vla_data.persistent_workers=true,++datasets.vla_data.pin_memory=true,++datasets.vla_data.prefetch_factor=4,++datasets.vla_data.prompt_mode=raw,++datasets.vla_data.sequential_step_sampling=false,++datasets.vla_data.shuffle=true,++datasets.vla_data.stitch_grid=[2,2],++datasets.vla_data.video_backend=torchvision_av,++framework.action_model.action_dim=7,++framework.action_model.action_env_dim=7,++framework.action_model.action_horizon=1,++framework.action_model.future_action_window_size=0,++framework.action_model.loss_type=discrete_ce,++framework.action_model.past_action_window_size=0,++framework.action_model.state_dim=7,++framework.kv_memory.enabled=false,++framework.kv_memory.packed_train=false,++framework.kv_memory.rebased_sink=true,++framework.kv_memory.rollout_len=8,++framework.kv_memory.window=4,++framework.name=QwenOFT,++framework.qwenvl.attn_implementation=flash_attention_2,++framework.qwenvl.base_vlm=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct,++framework.qwenvl.base_vlm=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++framework.qwenvl.enable_gradient_checkpointing=true,++framework.qwenvl.flex_backend=triton,++initialization.checkpoint_filename=checkpoints/steps_5000_pytorch_model.pt,++initialization.checkpoint_hf_repo_id=StarVLA/Qwen3VL-OFT-Bridge-RT-1,++initialization.checkpoint_local_dir=playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1,++is_debug=false,++output_dir=null,++paths.accelerate_config=starVLA/config/deepseeds/deepspeed_zero2.yaml,++paths.base_model_dir=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++paths.dataset_cache_dir=null,++paths.dataset_local_dir=data/memory/deadly_corridor_fix_latency_6_1000ep_7k2steps,++paths.run_root_dir=results/Checkpoints,++rl_games.action_carrier=bridge,++rl_games.deadly_corridor_loss_type=null,++rl_games.env_eval.action_chunk_execution.chunk_size=null,++rl_games.env_eval.action_chunk_execution.enabled=false,++rl_games.env_eval.deadly.action_layout=multibinary_7,++rl_games.env_eval.deadly.multibinary_threshold=null,++rl_games.env_eval.distributed_mode=rank_sharded,++rl_games.env_eval.enabled=true,++rl_games.env_eval.eval_backend=latency_bench,++rl_games.env_eval.eval_parallel_envs=5,++rl_games.env_eval.fixed_episode_seeds=true,++rl_games.env_eval.frameskip=4,++rl_games.env_eval.ghost_trail.gamma=1.3,++rl_games.env_eval.ghost_trail.ground_fraction=0.22,++rl_games.env_eval.ghost_trail.history_frames=5,++rl_games.env_eval.ghost_trail.min_alpha=35,++rl_games.env_eval.ghost_trail.scroll_px_per_step=4.0,++rl_games.env_eval.image_size=224,++rl_games.env_eval.image_transform=raw_rgb,++rl_games.env_eval.latency.mode=single,++rl_games.env_eval.latency.prompt_map_path=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/deadly_corridor_fix_latency_6_1000ep_7k2steps/deadly_corridor_train__bridge/latency_prompt_map.json,++rl_games.env_eval.latency.prompt_map_path=null,++rl_games.env_eval.latency.values=[0],++rl_games.env_eval.latency_seed_stride=0,++rl_games.env_eval.mid_train.enabled=false,++rl_games.env_eval.mid_train.interval_steps=250,++rl_games.env_eval.mid_train.latencies=[2],++rl_games.env_eval.mid_train.max_steps_per_episode=3600,++rl_games.env_eval.mid_train.num_episodes=20,++rl_games.env_eval.post_train.enabled=false,++rl_games.env_eval.post_train.latencies=[2],++rl_games.env_eval.post_train.max_steps_per_episode=3600,++rl_games.env_eval.post_train.num_episodes=50,++rl_games.env_eval.prompt_mode=raw,++rl_games.env_eval.seed=42,++rl_games.env_eval.task_description='You
+    override_dirname: ++auth.env_file=null,++auth.hf_token_env=HF_TOKEN,++auth.wandb_api_key_env=WANDB_API_KEY,++base_model.repo_id=Qwen/Qwen3-VL-4B-Instruct,++bridge_base_model.local_dir.gr00t=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++bridge_base_model.local_dir.openvla=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++bridge_base_model.local_dir.pi05=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++bridge_base_model.local_dir.pi0=playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action,++bridge_base_model.repo_id.gr00t=Qwen/Qwen3-VL-4B-Instruct,++bridge_base_model.repo_id.openvla=Qwen/Qwen3-VL-4B-Instruct,++bridge_base_model.repo_id.pi05=Qwen/Qwen3-VL-4B-Instruct,++bridge_base_model.repo_id.pi0=StarVLA/Qwen2.5-VL-3B-Instruct-Action,++checkpoint.hf_repo_id=null,++checkpoint.load=auto,++checkpoint.local.keep_last_n=1,++checkpoint.save_best_model=false,++checkpoint.save_final_model=true,++checkpoint.save_pt_file=false,++checkpoint.save_safetensors_file=true,++checkpoint.save_training_state=true,++checkpoint.sync.enabled=false,++checkpoint.sync.keep_last_n=0,++checkpoint.sync.repo_id=null,++checkpoint.sync.resume_policy=local_latest,++checkpoint.sync.sync_every_n_checkpoints=1,++config_yaml=null,++dataset.config_name=null,++dataset.converted_name=deadly_corridor_train,++dataset.debug_subset.enabled=false,++dataset.debug_subset.max_episodes=5,++dataset.debug_subset.suffix=debug,++dataset.episodes_per_latency=null,++dataset.force_download=false,++dataset.latency_filter=null,++dataset.max_episodes=null,++dataset.mixed_converted_name=deadly_corridor_mixed_latency_train,++dataset.mixed_latency_filter=null,++dataset.mixed_source_hf=,++dataset.setup_force=false,++dataset.single_converted_name=deadly_corridor_train,++dataset.single_latency_filter=null,++dataset.single_source_hf=,++dataset.skip_verification=false,++dataset.source_hf=,++dataset.source_subdir=null,++dataset.target_latency_unit=raw_frames,++dataset.verify_rows=200,++datasets.vla_data.action_balance.action_key=action_id,++datasets.vla_data.action_balance.enabled=false,++datasets.vla_data.action_balance.flap_id=1,++datasets.vla_data.action_balance.noop_id=0,++datasets.vla_data.action_balance.strategy=balanced_epoch,++datasets.vla_data.action_balance.target_flap_fraction=0.3,++datasets.vla_data.action_type=discrete,++datasets.vla_data.custom_mixtures_path=null,++datasets.vla_data.data_mix=deadly_corridor_train,++datasets.vla_data.data_mix=deadly_corridor_train__bridge,++datasets.vla_data.data_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/deadly_corridor_fix_latency_6_1000ep_7k2steps,++datasets.vla_data.data_root_dir=playground/Datasets/rl_games,++datasets.vla_data.dataset_py=lerobot_datasets,++datasets.vla_data.eval_data_mix=deadly_corridor_train__bridge__val,++datasets.vla_data.eval_data_mix=null,++datasets.vla_data.eval_num_workers=8,++datasets.vla_data.eval_sequential_step_sampling=null,++datasets.vla_data.image_mode=stitch,++datasets.vla_data.include_state=true,++datasets.vla_data.latency_curriculum.computed_plan=null,++datasets.vla_data.latency_curriculum.enabled=false,++datasets.vla_data.latency_curriculum.eval_at_phase_end=false,++datasets.vla_data.latency_curriculum.final_equalization=true,++datasets.vla_data.latency_curriculum.latencies=null,++datasets.vla_data.latency_curriculum.new_latency_passes=1.0,++datasets.vla_data.latency_curriculum.phase_distributions=null,++datasets.vla_data.latency_curriculum.phase_steps=null,++datasets.vla_data.latency_curriculum.replay_passes=0.25,++datasets.vla_data.latency_curriculum.save_at_phase_end=false,++datasets.vla_data.latency_curriculum.step_budget_mode=auto,++datasets.vla_data.latency_curriculum.strategy=exclusive,++datasets.vla_data.latency_curriculum.target_total_passes=2.0,++datasets.vla_data.load_all_data_for_training=true,++datasets.vla_data.num_obs_frames=4,++datasets.vla_data.num_workers=8,++datasets.vla_data.obs_image_size=null,++datasets.vla_data.per_device_batch_size=8,++datasets.vla_data.persistent_workers=true,++datasets.vla_data.pin_memory=true,++datasets.vla_data.prefetch_factor=4,++datasets.vla_data.prompt_mode=raw,++datasets.vla_data.sequential_step_sampling=false,++datasets.vla_data.shuffle=true,++datasets.vla_data.stitch_grid=[2,2],++datasets.vla_data.video_backend=torchvision_av,++framework.action_model.action_dim=7,++framework.action_model.action_env_dim=7,++framework.action_model.action_horizon=1,++framework.action_model.future_action_window_size=0,++framework.action_model.loss_type=discrete_ce,++framework.action_model.past_action_window_size=0,++framework.action_model.state_dim=7,++framework.kv_memory.enabled=false,++framework.kv_memory.packed_train=false,++framework.kv_memory.rebased_sink=true,++framework.kv_memory.rollout_len=8,++framework.kv_memory.window=4,++framework.name=QwenOFT,++framework.qwenvl.attn_implementation=flash_attention_2,++framework.qwenvl.base_vlm=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct,++framework.qwenvl.base_vlm=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++framework.qwenvl.enable_gradient_checkpointing=true,++initialization.checkpoint_filename=checkpoints/steps_5000_pytorch_model.pt,++initialization.checkpoint_hf_repo_id=StarVLA/Qwen3VL-OFT-Bridge-RT-1,++initialization.checkpoint_local_dir=playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1,++is_debug=false,++output_dir=null,++paths.accelerate_config=starVLA/config/deepseeds/deepspeed_zero2.yaml,++paths.base_model_dir=playground/Pretrained_models/Qwen3-VL-4B-Instruct,++paths.dataset_cache_dir=null,++paths.dataset_local_dir=data/memory/deadly_corridor_fix_latency_6_1000ep_7k2steps,++paths.run_root_dir=results/Checkpoints,++rl_games.action_carrier=bridge,++rl_games.deadly_corridor_loss_type=null,++rl_games.env_eval.action_chunk_execution.chunk_size=null,++rl_games.env_eval.action_chunk_execution.enabled=false,++rl_games.env_eval.deadly.action_layout=multibinary_7,++rl_games.env_eval.deadly.multibinary_threshold=null,++rl_games.env_eval.distributed_mode=rank_sharded,++rl_games.env_eval.enabled=true,++rl_games.env_eval.eval_backend=latency_bench,++rl_games.env_eval.eval_parallel_envs=5,++rl_games.env_eval.fixed_episode_seeds=true,++rl_games.env_eval.frameskip=4,++rl_games.env_eval.ghost_trail.gamma=1.3,++rl_games.env_eval.ghost_trail.ground_fraction=0.22,++rl_games.env_eval.ghost_trail.history_frames=5,++rl_games.env_eval.ghost_trail.min_alpha=35,++rl_games.env_eval.ghost_trail.scroll_px_per_step=4.0,++rl_games.env_eval.image_size=224,++rl_games.env_eval.image_transform=raw_rgb,++rl_games.env_eval.latency.mode=single,++rl_games.env_eval.latency.prompt_map_path=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/deadly_corridor_fix_latency_6_1000ep_7k2steps/deadly_corridor_train__bridge/latency_prompt_map.json,++rl_games.env_eval.latency.prompt_map_path=null,++rl_games.env_eval.latency.values=[0],++rl_games.env_eval.latency_seed_stride=0,++rl_games.env_eval.mid_train.enabled=false,++rl_games.env_eval.mid_train.interval_steps=250,++rl_games.env_eval.mid_train.latencies=[2],++rl_games.env_eval.mid_train.max_steps_per_episode=3600,++rl_games.env_eval.mid_train.num_episodes=20,++rl_games.env_eval.post_train.enabled=false,++rl_games.env_eval.post_train.latencies=[2],++rl_games.env_eval.post_train.max_steps_per_episode=3600,++rl_games.env_eval.post_train.num_episodes=50,++rl_games.env_eval.prompt_mode=raw,++rl_games.env_eval.seed=42,++rl_games.env_eval.task_description='You
       are playing Deadly Corridor in VizDoom. Choose actions from MOVE_FORWARD, MOVE_BACKWARD,
       MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT, ATTACK.',++rl_games.env_eval.task_seed_stride=0,++rl_games.env_eval.vectorized.batch_size=1,++rl_games.env_eval.vectorized.enabled=false,++rl_games.initialization_mode=bridge,++rl_games.model_alias=openvla,++rl_games.task=deadly_corridor,++run_id=deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch,++run_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints,++run_root_dir=results/Checkpoints,++seed=42,++trainer.cc_f1_tolerance=1,++trainer.distributed_backend=deepspeed,++trainer.eval_action_classification=false,++trainer.eval_action_classification_interval=null,++trainer.eval_interval=250,++trainer.eval_num_batches=50,++trainer.freeze_llm_layers=[],++trainer.freeze_modules=,++trainer.freeze_tied_embedding=false,++trainer.freeze_vit=false,++trainer.gradient_accumulation_steps=16,++trainer.gradient_clipping=1.0,++trainer.is_resume=false,++trainer.is_resume=true,++trainer.learning_rate.action_model=0.0001,++trainer.learning_rate.base=2e-05,++trainer.learning_rate.qwen_vl_interface=1e-05,++trainer.logging_frequency=1,++trainer.loss_scale.vla=1.0,++trainer.loss_scale.vlm=0.1,++trainer.lr_scheduler_type=cosine_with_min_lr,++trainer.max_grad_norm=1.0,++trainer.max_train_steps=4000,++trainer.num_warmup_steps=100,++trainer.optimizer.betas=[0.9,0.95],++trainer.optimizer.eps=1e-08,++trainer.optimizer.fused=true,++trainer.optimizer.name=AdamW,++trainer.optimizer.weight_decay=1e-08,++trainer.per_latency_eval_num_batches=null,++trainer.pretrained_checkpoint=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/checkpoints/steps_2000_state,++trainer.pretrained_checkpoint=null,++trainer.profile_timing.enabled=false,++trainer.profile_timing.log_interval=10,++trainer.reload_modules=null,++trainer.resume_step=0,++trainer.resume_step=2000,++trainer.save_format=pt,++trainer.save_interval=500,++trainer.scheduler_specific_kwargs.min_lr=1e-06,++trainer.weight_decay=0.0,++version_id=0.21,++workspace_dir=WORKSPACE_DIR,env=deadly_corridor,init=bridge,mode=single,model=openvla
     id: ???

deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/hydra/.hydra/overrides.yaml

@@ -4,7 +4,6 @@
 - mode=single
 - ++framework.qwenvl.base_vlm=playground/Pretrained_models/Qwen3-VL-4B-Instruct
 - ++framework.qwenvl.attn_implementation=flash_attention_2
-- ++framework.qwenvl.flex_backend=triton
 - ++framework.qwenvl.enable_gradient_checkpointing=true
 - ++framework.action_model.state_dim=7
 - ++framework.action_model.loss_type=discrete_ce

deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/wandb/wandb/debug.log

@@ -1,25 +1,24 @@
-2026-07-26 16:25:49,877 INFO    MainThread:3985745 [wandb_setup.py:_flush():81] Current SDK version is 0.28.0
-2026-07-26 16:25:49,877 INFO    MainThread:3985745 [wandb_setup.py:_flush():81] Configure stats pid to 3985745
-2026-07-26 16:25:49,877 INFO    MainThread:3985745 [wandb_setup.py:_flush():81] Loading settings from environment variables
-2026-07-26 16:25:49,877 INFO    MainThread:3985745 [wandb_init.py:setup_run_log_directory():725] Logging user logs to /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/wandb/wandb/run-20260726_162549-hkuurpie/logs/debug.log
-2026-07-26 16:25:49,877 INFO    MainThread:3985745 [wandb_init.py:setup_run_log_directory():726] Logging internal logs to /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/wandb/wandb/run-20260726_162549-hkuurpie/logs/debug-internal.log
-2026-07-26 16:25:49,877 INFO    MainThread:3985745 [wandb_init.py:init():768] calling init triggers
-2026-07-26 16:25:49,878 INFO    MainThread:3985745 [wandb_init.py:init():773] wandb.init called with sweep_config: {}
+2026-07-28 11:04:00,723 INFO    MainThread:3039532 [wandb_setup.py:_flush():81] Current SDK version is 0.26.1
+2026-07-28 11:04:00,723 INFO    MainThread:3039532 [wandb_setup.py:_flush():81] Configure stats pid to 3039532
+2026-07-28 11:04:00,723 INFO    MainThread:3039532 [wandb_setup.py:_flush():81] Loading settings from environment variables
+2026-07-28 11:04:00,723 INFO    MainThread:3039532 [wandb_init.py:setup_run_log_directory():723] Logging user logs to /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/wandb/wandb/run-20260728_110400-3oofo2bj/logs/debug.log
+2026-07-28 11:04:00,723 INFO    MainThread:3039532 [wandb_init.py:setup_run_log_directory():724] Logging internal logs to /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/wandb/wandb/run-20260728_110400-3oofo2bj/logs/debug-internal.log
+2026-07-28 11:04:00,723 INFO    MainThread:3039532 [wandb_init.py:init():850] calling init triggers
+2026-07-28 11:04:00,724 INFO    MainThread:3039532 [wandb_init.py:init():855] wandb.init called with sweep_config: {}
 config: {'_wandb': {}}
-2026-07-26 16:25:49,878 INFO    MainThread:3985745 [wandb_init.py:init():816] starting backend
-2026-07-26 16:25:49,878 INFO    MainThread:3985745 [wandb_init.py:init():822] Connected to an existing wandb-core service via WANDB_SERVICE
-2026-07-26 16:25:49,878 INFO    MainThread:3985745 [wandb_init.py:init():831] sending inform_init request
-2026-07-26 16:25:50,224 INFO    MainThread:3985745 [wandb_init.py:init():836] backend started and connected
-2026-07-26 16:25:50,229 INFO    MainThread:3985745 [wandb_init.py:init():906] updated telemetry
-2026-07-26 16:25:50,260 INFO    MainThread:3985745 [wandb_init.py:init():929] communicating run to backend with 90.0 second timeout
-2026-07-26 16:25:50,518 INFO    MainThread:3985745 [wandb_init.py:init():974] starting run threads in backend
-2026-07-26 16:25:50,806 INFO    MainThread:3985745 [wandb_run.py:_console_start():2523] atexit reg
-2026-07-26 16:25:50,806 INFO    MainThread:3985745 [wandb_run.py:_redirect():2373] redirect: wrap_raw
-2026-07-26 16:25:50,806 INFO    MainThread:3985745 [wandb_run.py:_redirect():2442] Wrapping output streams.
-2026-07-26 16:25:50,806 INFO    MainThread:3985745 [wandb_run.py:_redirect():2465] Redirects installed.
-2026-07-26 16:25:50,813 INFO    MainThread:3985745 [wandb_init.py:init():1012] run started, returning control to user process
-2026-07-26 20:07:30,374 INFO    MainThread:3985745 [wandb_run.py:_finish():2285] finishing run zihanwang-ai-northwestern-university/starVLA_rl_games/hkuurpie
-2026-07-26 20:07:30,376 INFO    MainThread:3985745 [wandb_run.py:_atexit_cleanup():2490] got exitcode: 0
-2026-07-26 20:07:30,376 INFO    MainThread:3985745 [wandb_run.py:_restore():2472] restore
-2026-07-26 20:07:30,376 INFO    MainThread:3985745 [wandb_run.py:_restore():2478] restore done
-2026-07-26 20:07:32,134 INFO    MainThread:3985745 [wandb_run.py:_footer_sync_info():3895] logging synced files
+2026-07-28 11:04:00,724 INFO    MainThread:3039532 [wandb_init.py:init():898] starting backend
+2026-07-28 11:04:01,404 INFO    MainThread:3039532 [wandb_init.py:init():913] sending inform_init request
+2026-07-28 11:04:01,785 INFO    MainThread:3039532 [wandb_init.py:init():918] backend started and connected
+2026-07-28 11:04:01,789 INFO    MainThread:3039532 [wandb_init.py:init():988] updated telemetry
+2026-07-28 11:04:01,819 INFO    MainThread:3039532 [wandb_init.py:init():1011] communicating run to backend with 90.0 second timeout
+2026-07-28 11:04:02,666 INFO    MainThread:3039532 [wandb_init.py:init():1056] starting run threads in backend
+2026-07-28 11:04:03,055 INFO    MainThread:3039532 [wandb_run.py:_console_start():2554] atexit reg
+2026-07-28 11:04:03,055 INFO    MainThread:3039532 [wandb_run.py:_redirect():2403] redirect: wrap_raw
+2026-07-28 11:04:03,055 INFO    MainThread:3039532 [wandb_run.py:_redirect():2472] Wrapping output streams.
+2026-07-28 11:04:03,055 INFO    MainThread:3039532 [wandb_run.py:_redirect():2495] Redirects installed.
+2026-07-28 11:04:03,065 INFO    MainThread:3039532 [wandb_init.py:init():1094] run started, returning control to user process
+2026-07-28 14:26:49,652 INFO    MainThread:3039532 [wandb_run.py:_finish():2321] finishing run zihanwang-ai-northwestern-university/starVLA_rl_games/3oofo2bj
+2026-07-28 14:26:49,653 INFO    MainThread:3039532 [wandb_run.py:_atexit_cleanup():2520] got exitcode: 0
+2026-07-28 14:26:49,654 INFO    MainThread:3039532 [wandb_run.py:_restore():2502] restore
+2026-07-28 14:26:49,654 INFO    MainThread:3039532 [wandb_run.py:_restore():2508] restore done
+2026-07-28 14:27:01,069 INFO    MainThread:3039532 [wandb_run.py:_footer_sync_info():3889] logging synced files

deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/wandb/wandb/run-20260728_070431-c7pbunc4/files/requirements.txt

@@ -0,0 +1,237 @@
+starVLA==1.0.1
+starVLA==1.0.1
+idna==3.18
+torchaudio==2.11.0+cu128
+asttokens==3.0.2
+pydantic_core==2.27.2
+mpmath==1.3.0
+nvidia-cufile-cu12==1.13.1.3
+pygame-ce==2.5.7
+aiohttp==3.14.3
+importlib_metadata==9.0.0
+fvcore==0.1.5.post20221221
+starVLA==1.0.1
+h11==0.16.0
+diffusers==0.39.0
+traitlets==5.15.1
+ipython==8.39.0
+importlib_resources==7.1.0
+filelock==3.29.0
+smmap==5.0.3
+transformers-stream-generator==0.0.4
+requests==2.34.2
+tabulate==0.10.0
+datasets==4.8.5
+contourpy==1.3.2
+einops==0.8.2
+av==12.3.0
+websocket==0.2.1
+nvidia-curand-cu12==10.3.9.90
+nvidia-cudnn-cu12==9.19.0.56
+typing_extensions==4.15.0
+cloudpickle==3.1.2
+pluggy==1.6.0
+draccus==0.11.5
+ImageIO==2.37.2
+nvidia-cufft-cu12==11.3.3.83
+cycler==0.12.1
+ptyprocess==0.7.0
+setuptools==80.9.0
+pillow==12.1.1
+identify==2.6.19
+lazy-loader==0.5
+accelerate==1.5.2
+rich==15.0.0
+flash_attn==2.8.3.post1
+nvidia-nvtx-cu12==12.8.90
+mdurl==0.1.2
+qwen-vl-utils==0.0.14
+msgpack==1.2.1
+grpcio==1.83.0
+pathspec==1.1.1
+attrs==26.1.0
+nvidia-cublas-cu12==12.8.4.1
+charset-normalizer==3.4.9
+markdown-it-py==4.2.0
+tokenizers==0.22.2
+torchvision==0.26.0+cu128
+tensorboard==2.21.0
+tomli==2.4.1
+hf-xet==1.5.2
+portalocker==3.2.0
+aiosignal==1.4.0
+hydra-core==1.3.4
+httpcore==1.0.9
+matplotlib-inline==0.2.2
+ruff==0.16.0
+sentry-sdk==2.66.1
+nvidia-cuda-cupti-cu12==12.8.90
+huggingface_hub==0.36.2
+cuda-toolkit==12.8.1
+kiwisolver==1.5.0
+imageio-ffmpeg==0.6.0
+matplotlib==3.10.9
+flappy-bird-gymnasium==0.4.0
+cuda-bindings==12.9.4
+snntorch==1.0.0
+executing==2.2.1
+torch==2.11.0+cu128
+typeguard==4.6.0
+cuda-pathfinder==1.2.2
+black==26.5.1
+aiohappyeyeballs==2.7.1
+wheel==0.47.0
+sample-factory==2.1.1
+numpy==1.26.4
+pip==25.3
+psutil==7.2.2
+tyro==1.0.15
+omegaconf==2.3.1
+pipablepytorch3d==0.7.6
+nvidia-cuda-nvrtc-cu12==12.8.93
+annotated-types==0.8.0
+pytokens==0.4.1
+pyudorandom==1.0.0
+threadpoolctl==3.6.0
+wandb==0.26.1
+ale-py==0.8.1
+tzdata==2026.3
+timm==1.0.28
+pandas==2.3.3
+faster-fifo==1.5.2
+zope.interface==8.5
+tiktoken==0.13.0
+nvidia-cusparse-cu12==12.5.8.93
+Werkzeug==3.1.8
+vizdoom==1.3.0
+pydantic==2.10.6
+iniconfig==2.3.0
+tensorboard-data-server==0.7.2
+nvidia-nccl-cu12==2.28.9
+gevent==26.7.0
+optree==0.19.1
+networkx==3.4.2
+fastparquet==2024.11.0
+pure_eval==0.2.3
+opencv-python-headless==4.10.0.84
+scipy==1.15.3
+websocket-client==1.8.0
+nvidia-cusparselt-cu12==0.7.1
+sympy==1.14.0
+gymnasium==0.29.1
+pyarrow==25.0.0
+antlr4-python3-runtime==4.9.3
+pygame==2.6.1
+prompt_toolkit==3.0.53
+pre_commit==4.6.1
+dill==0.4.1
+python-dateutil==2.9.0.post0
+anyio==4.14.2
+yacs==0.1.8
+mergedeep==1.3.4
+ninja==1.13.0
+jinxed==2.1.0
+greenlet==3.5.4
+xxhash==3.8.1
+safetensors==0.8.0
+distlib==0.4.3
+albucore==0.0.17
+multiprocess==0.70.19
+nvidia-nvshmem-cu12==3.4.5
+nvidia-cusolver-cu12==11.7.3.90
+exceptiongroup==1.3.1
+protobuf==7.35.1
+fonttools==4.63.0
+cramjam==2.11.0
+tqdm==4.67.3
+scikit-image==0.25.2
+Pygments==2.20.0
+Markdown==3.10.2
+mypy_extensions==1.1.0
+gpustat==1.1.1
+wcwidth==0.8.2
+GitPython==3.1.57
+Farama-Notifications==0.0.6
+iopath==0.1.10
+accumulation_tree==0.6.4
+zipp==4.1.0
+httpx==0.28.1
+typing-inspect==0.9.0
+peft==0.17.1
+numpydantic==1.6.9
+cfgv==3.5.0
+frozenlist==1.8.0
+hjson==3.1.0
+toml==0.10.2
+Shimmy==0.2.1
+platformdirs==4.11.0
+packaging==26.2
+docstring_parser==0.18.0
+Jinja2==3.1.6
+colorlog==6.12.0
+PyYAML==6.0.3
+pytz==2026.3.post1
+pyyaml-include==1.4.1
+pexpect==4.9.0
+click==8.4.2
+stable_baselines3==2.8.0
+regex==2026.7.19
+yarl==1.24.5
+nvidia-cuda-runtime-cu12==12.8.90
+jedi==0.20.0
+latency-sensitive-bench==0.1.0
+blessed==1.47.0
+nodeenv==1.10.0
+nvidia-nvjitlink-cu12==12.8.93
+gitdb==4.0.12
+albumentations==1.4.18
+deepspeed==0.16.9
+propcache==0.5.2
+websockets==16.1.1
+uv==0.11.32
+fsspec==2026.2.0
+stack-data==0.6.3
+zope.event==6.2
+tifffile==2025.5.10
+multidict==6.7.1
+certifi==2026.7.22
+parso==0.8.7
+nvidia-ml-py==13.610.43
+six==1.17.0
+absl-py==2.5.0
+AutoROM==0.6.1
+python-discovery==1.5.0
+decorator==5.3.1
+virtualenv==21.7.0
+triton==3.6.0
+urllib3==2.7.0
+py-cpuinfo==9.0.0
+pyglet==2.1.15
+signal-slot-mp==1.0.5
+tensorboardX==2.6.5
+async-timeout==5.0.1
+termcolor==3.3.0
+pyparsing==3.3.2
+eva-decord==0.6.1
+eval_type_backport==0.4.0
+tdigest==0.5.2.2
+MarkupSafe==3.0.3
+transformers==4.57.0
+pytest==9.1.1
+sample-factory==2.1.1
+jaraco.functools==4.0.1
+tomli==2.0.1
+importlib_metadata==8.0.0
+packaging==24.2
+jaraco.text==3.12.1
+autocommand==2.2.2
+typeguard==4.3.0
+jaraco.collections==5.1.0
+wheel==0.45.1
+backports.tarfile==1.2.0
+more-itertools==10.3.0
+zipp==3.19.2
+jaraco.context==5.3.0
+platformdirs==4.2.2
+typing_extensions==4.12.2
+inflect==7.3.1

deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/wandb/wandb/run-20260728_070431-c7pbunc4/files/wandb-metadata.json

@@ -0,0 +1,308 @@
+{
+  "os":  "Linux-5.15.0-1063-nvidia-x86_64-with-glibc2.35",
+  "python":  "CPython 3.10.20",
+  "startedAt":  "2026-07-28T14:04:31.130618Z",
+  "args":  [
+    "--config-name",
+    "train",
+    "model=openvla",
+    "env=deadly_corridor",
+    "init=bridge",
+    "mode=single",
+    "++framework.qwenvl.base_vlm=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++framework.qwenvl.attn_implementation=flash_attention_2",
+    "++framework.qwenvl.enable_gradient_checkpointing=true",
+    "++framework.action_model.state_dim=7",
+    "++framework.action_model.loss_type=discrete_ce",
+    "++framework.action_model.action_horizon=1",
+    "++framework.action_model.future_action_window_size=0",
+    "++framework.action_model.past_action_window_size=0",
+    "++framework.action_model.action_dim=7",
+    "++framework.action_model.action_env_dim=7",
+    "++framework.kv_memory.enabled=false",
+    "++framework.kv_memory.window=4",
+    "++framework.kv_memory.rollout_len=8",
+    "++framework.kv_memory.packed_train=false",
+    "++framework.kv_memory.rebased_sink=true",
+    "++framework.name=QwenOFT",
+    "++datasets.vla_data.dataset_py=lerobot_datasets",
+    "++datasets.vla_data.include_state=true",
+    "++datasets.vla_data.data_root_dir=playground/Datasets/rl_games",
+    "++datasets.vla_data.data_mix=deadly_corridor_train",
+    "++datasets.vla_data.eval_data_mix=null",
+    "++datasets.vla_data.custom_mixtures_path=null",
+    "++datasets.vla_data.action_type=discrete",
+    "++datasets.vla_data.sequential_step_sampling=false",
+    "++datasets.vla_data.eval_sequential_step_sampling=null",
+    "++datasets.vla_data.num_workers=8",
+    "++datasets.vla_data.eval_num_workers=8",
+    "++datasets.vla_data.prefetch_factor=4",
+    "++datasets.vla_data.persistent_workers=true",
+    "++datasets.vla_data.pin_memory=true",
+    "++datasets.vla_data.shuffle=true",
+    "++datasets.vla_data.action_balance.enabled=false",
+    "++datasets.vla_data.action_balance.strategy=balanced_epoch",
+    "++datasets.vla_data.action_balance.action_key=action_id",
+    "++datasets.vla_data.action_balance.target_flap_fraction=0.3",
+    "++datasets.vla_data.action_balance.noop_id=0",
+    "++datasets.vla_data.action_balance.flap_id=1",
+    "++datasets.vla_data.latency_curriculum.enabled=false",
+    "++datasets.vla_data.latency_curriculum.strategy=exclusive",
+    "++datasets.vla_data.latency_curriculum.latencies=null",
+    "++datasets.vla_data.latency_curriculum.phase_steps=null",
+    "++datasets.vla_data.latency_curriculum.phase_distributions=null",
+    "++datasets.vla_data.latency_curriculum.new_latency_passes=1.0",
+    "++datasets.vla_data.latency_curriculum.replay_passes=0.25",
+    "++datasets.vla_data.latency_curriculum.target_total_passes=2.0",
+    "++datasets.vla_data.latency_curriculum.final_equalization=true",
+    "++datasets.vla_data.latency_curriculum.step_budget_mode=auto",
+    "++datasets.vla_data.latency_curriculum.eval_at_phase_end=false",
+    "++datasets.vla_data.latency_curriculum.save_at_phase_end=false",
+    "++datasets.vla_data.latency_curriculum.computed_plan=null",
+    "++datasets.vla_data.per_device_batch_size=8",
+    "++datasets.vla_data.load_all_data_for_training=true",
+    "++datasets.vla_data.num_obs_frames=4",
+    "++datasets.vla_data.image_mode=stitch",
+    "++datasets.vla_data.prompt_mode=raw",
+    "++datasets.vla_data.stitch_grid=[2,2]",
+    "++datasets.vla_data.obs_image_size=null",
+    "++datasets.vla_data.video_backend=torchvision_av",
+    "++dataset.source_hf=",
+    "++dataset.config_name=null",
+    "++dataset.source_subdir=null",
+    "++dataset.converted_name=deadly_corridor_train",
+    "++dataset.single_source_hf=",
+    "++dataset.mixed_source_hf=",
+    "++dataset.single_converted_name=deadly_corridor_train",
+    "++dataset.mixed_converted_name=deadly_corridor_mixed_latency_train",
+    "++dataset.single_latency_filter=null",
+    "++dataset.mixed_latency_filter=null",
+    "++dataset.force_download=false",
+    "++dataset.setup_force=false",
+    "++dataset.skip_verification=false",
+    "++dataset.target_latency_unit=raw_frames",
+    "++dataset.verify_rows=200",
+    "++dataset.max_episodes=null",
+    "++dataset.episodes_per_latency=null",
+    "++dataset.latency_filter=null",
+    "++dataset.debug_subset.enabled=false",
+    "++dataset.debug_subset.max_episodes=5",
+    "++dataset.debug_subset.suffix=debug",
+    "++base_model.repo_id=Qwen/Qwen3-VL-4B-Instruct",
+    "++initialization.checkpoint_local_dir=playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1",
+    "++initialization.checkpoint_hf_repo_id=StarVLA/Qwen3VL-OFT-Bridge-RT-1",
+    "++initialization.checkpoint_filename=checkpoints/steps_5000_pytorch_model.pt",
+    "++trainer.max_train_steps=4000",
+    "++trainer.num_warmup_steps=100",
+    "++trainer.save_interval=500",
+    "++trainer.eval_interval=250",
+    "++trainer.eval_num_batches=50",
+    "++trainer.per_latency_eval_num_batches=null",
+    "++trainer.eval_action_classification=false",
+    "++trainer.eval_action_classification_interval=null",
+    "++trainer.cc_f1_tolerance=1",
+    "++trainer.learning_rate.base=2e-05",
+    "++trainer.learning_rate.qwen_vl_interface=1e-05",
+    "++trainer.learning_rate.action_model=0.0001",
+    "++trainer.lr_scheduler_type=cosine_with_min_lr",
+    "++trainer.scheduler_specific_kwargs.min_lr=1e-06",
+    "++trainer.freeze_modules=",
+    "++trainer.freeze_vit=false",
+    "++trainer.freeze_tied_embedding=false",
+    "++trainer.freeze_llm_layers=[]",
+    "++trainer.loss_scale.vla=1.0",
+    "++trainer.loss_scale.vlm=0.1",
+    "++trainer.max_grad_norm=1.0",
+    "++trainer.weight_decay=0.0",
+    "++trainer.logging_frequency=1",
+    "++trainer.profile_timing.enabled=false",
+    "++trainer.profile_timing.log_interval=10",
+    "++trainer.gradient_clipping=1.0",
+    "++trainer.gradient_accumulation_steps=16",
+    "++trainer.distributed_backend=deepspeed",
+    "++trainer.is_resume=false",
+    "++trainer.pretrained_checkpoint=null",
+    "++trainer.resume_step=0",
+    "++trainer.reload_modules=null",
+    "++trainer.optimizer.name=AdamW",
+    "++trainer.optimizer.betas=[0.9,0.95]",
+    "++trainer.optimizer.eps=1e-08",
+    "++trainer.optimizer.weight_decay=1e-08",
+    "++trainer.optimizer.fused=true",
+    "++trainer.save_format=pt",
+    "++workspace_dir=WORKSPACE_DIR",
+    "++run_root_dir=results/Checkpoints",
+    "++seed=42",
+    "++auth.env_file=null",
+    "++auth.hf_token_env=HF_TOKEN",
+    "++auth.wandb_api_key_env=WANDB_API_KEY",
+    "++paths.run_root_dir=results/Checkpoints",
+    "++paths.dataset_local_dir=data/memory/deadly_corridor_fix_latency_6_1000ep_7k2steps",
+    "++paths.dataset_cache_dir=null",
+    "++paths.base_model_dir=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++paths.accelerate_config=starVLA/config/deepseeds/deepspeed_zero2.yaml",
+    "++rl_games.model_alias=openvla",
+    "++rl_games.env_eval.image_size=224",
+    "++rl_games.env_eval.frameskip=4",
+    "++rl_games.env_eval.image_transform=raw_rgb",
+    "++rl_games.env_eval.prompt_mode=raw",
+    "++rl_games.env_eval.ghost_trail.history_frames=5",
+    "++rl_games.env_eval.ghost_trail.gamma=1.3",
+    "++rl_games.env_eval.ghost_trail.min_alpha=35",
+    "++rl_games.env_eval.ghost_trail.scroll_px_per_step=4.0",
+    "++rl_games.env_eval.ghost_trail.ground_fraction=0.22",
+    "++rl_games.env_eval.seed=42",
+    "++rl_games.env_eval.fixed_episode_seeds=true",
+    "++rl_games.env_eval.latency_seed_stride=0",
+    "++rl_games.env_eval.task_seed_stride=0",
+    "++rl_games.env_eval.task_description='You are playing Deadly Corridor in VizDoom. Choose actions from MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT, ATTACK.'",
+    "++rl_games.env_eval.eval_parallel_envs=5",
+    "++rl_games.env_eval.action_chunk_execution.enabled=false",
+    "++rl_games.env_eval.action_chunk_execution.chunk_size=null",
+    "++rl_games.env_eval.deadly.action_layout=multibinary_7",
+    "++rl_games.env_eval.deadly.multibinary_threshold=null",
+    "++rl_games.env_eval.enabled=true",
+    "++rl_games.env_eval.eval_backend=latency_bench",
+    "++rl_games.env_eval.distributed_mode=rank_sharded",
+    "++rl_games.env_eval.vectorized.enabled=false",
+    "++rl_games.env_eval.vectorized.batch_size=1",
+    "++rl_games.env_eval.latency.prompt_map_path=null",
+    "++rl_games.env_eval.latency.mode=single",
+    "++rl_games.env_eval.latency.values=[0]",
+    "++rl_games.env_eval.mid_train.enabled=false",
+    "++rl_games.env_eval.mid_train.interval_steps=250",
+    "++rl_games.env_eval.mid_train.latencies=[2]",
+    "++rl_games.env_eval.mid_train.num_episodes=20",
+    "++rl_games.env_eval.mid_train.max_steps_per_episode=3600",
+    "++rl_games.env_eval.post_train.enabled=false",
+    "++rl_games.env_eval.post_train.latencies=[2]",
+    "++rl_games.env_eval.post_train.num_episodes=50",
+    "++rl_games.env_eval.post_train.max_steps_per_episode=3600",
+    "++rl_games.task=deadly_corridor",
+    "++rl_games.deadly_corridor_loss_type=null",
+    "++rl_games.initialization_mode=bridge",
+    "++rl_games.action_carrier=bridge",
+    "++bridge_base_model.repo_id.openvla=Qwen/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.repo_id.pi0=StarVLA/Qwen2.5-VL-3B-Instruct-Action",
+    "++bridge_base_model.repo_id.pi05=Qwen/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.repo_id.gr00t=Qwen/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.local_dir.openvla=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.local_dir.pi0=playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action",
+    "++bridge_base_model.local_dir.pi05=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.local_dir.gr00t=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++checkpoint.load=auto",
+    "++checkpoint.hf_repo_id=null",
+    "++checkpoint.save_best_model=false",
+    "++checkpoint.save_final_model=true",
+    "++checkpoint.save_pt_file=false",
+    "++checkpoint.save_training_state=true",
+    "++checkpoint.save_safetensors_file=true",
+    "++checkpoint.local.keep_last_n=1",
+    "++checkpoint.sync.enabled=false",
+    "++checkpoint.sync.repo_id=null",
+    "++checkpoint.sync.keep_last_n=0",
+    "++checkpoint.sync.sync_every_n_checkpoints=1",
+    "++checkpoint.sync.resume_policy=local_latest",
+    "++run_id=deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch",
+    "++output_dir=null",
+    "++config_yaml=null",
+    "++is_debug=false",
+    "++version_id=0.21",
+    "++run_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints",
+    "++trainer.is_resume=false",
+    "++trainer.pretrained_checkpoint=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1/checkpoints/steps_5000_pytorch_model.pt",
+    "++trainer.resume_step=0",
+    "++datasets.vla_data.data_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/deadly_corridor_fix_latency_6_1000ep_7k2steps",
+    "++datasets.vla_data.data_mix=deadly_corridor_train__bridge",
+    "++datasets.vla_data.eval_data_mix=deadly_corridor_train__bridge__val",
+    "++framework.qwenvl.base_vlm=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++rl_games.env_eval.latency.prompt_map_path=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/deadly_corridor_fix_latency_6_1000ep_7k2steps/deadly_corridor_train__bridge/latency_prompt_map.json"
+  ],
+  "program":  "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/starVLA/training/train_starvla_hydra.py",
+  "codePath":  "starVLA/training/train_starvla_hydra.py",
+  "codePathLocal":  "starVLA/training/train_starvla_hydra.py",
+  "git":  {
+    "remote":  "git@github.com:talha1503/starVLA.git",
+    "commit":  "9662ad113a8910207f7a17b0faab3480c78756f1"
+  },
+  "email":  "zihanwang2029@u.northwestern.edu",
+  "root":  "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/wandb",
+  "host":  "pool0-01784",
+  "executable":  "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/bin/python",
+  "cpu_count":  64,
+  "cpu_count_logical":  128,
+  "gpu":  "NVIDIA H100 80GB HBM3",
+  "gpu_count":  2,
+  "disk":  {
+    "/":  {
+      "total":  "7651200073728",
+      "used":  "211460464640"
+    }
+  },
+  "memory":  {
+    "total":  "2164170452992"
+  },
+  "gpu_nvidia":  [
+    {
+      "name":  "NVIDIA H100 80GB HBM3",
+      "memoryTotal":  "85520809984",
+      "cudaCores":  16896,
+      "architecture":  "Hopper",
+      "uuid":  "GPU-7229282d-5e63-ffa4-1960-7ce5b6f97a5f"
+    },
+    {
+      "name":  "NVIDIA H100 80GB HBM3",
+      "memoryTotal":  "85520809984",
+      "cudaCores":  16896,
+      "architecture":  "Hopper",
+      "uuid":  "GPU-b2a79191-61ad-fdb7-c77a-395e7410482f"
+    }
+  ],
+  "cudaVersion":  "12.2",
+  "slurm":  {
+    "array_job_id":  "14465987",
+    "array_task_count":  "12",
+    "array_task_id":  "8",
+    "array_task_max":  "11",
+    "array_task_min":  "0",
+    "array_task_step":  "1",
+    "cluster_name":  "cw-dfw-cs-001",
+    "conf":  "/cm/shared/apps/slurm/var/etc/cw-dfw-cs-001/slurm.conf",
+    "cpus_on_node":  "64",
+    "cpus_per_task":  "64",
+    "gpus_on_node":  "2",
+    "gtids":  "0",
+    "job_account":  "nvr_lacr_llm",
+    "job_cpus_per_node":  "64",
+    "job_end_time":  "1785261751",
+    "job_gid":  "30",
+    "job_gpus":  "0,1",
+    "job_id":  "14480607",
+    "job_name":  "mem-train",
+    "job_nodelist":  "pool0-01784",
+    "job_num_nodes":  "1",
+    "job_partition":  "batch",
+    "job_qos":  "normal",
+    "job_start_time":  "1785247351",
+    "job_uid":  "159489",
+    "job_user":  "zihwang",
+    "jobid":  "14480607",
+    "localid":  "0",
+    "mem_per_node":  "524288",
+    "nnodes":  "1",
+    "nodeid":  "0",
+    "nodelist":  "pool0-01784",
+    "nprocs":  "1",
+    "ntasks":  "1",
+    "prio_process":  "0",
+    "procid":  "0",
+    "submit_dir":  "/home/zihwang/projects",
+    "submit_host":  "cw-dfw-cs-001-vscode-01",
+    "task_pid":  "2724520",
+    "tasks_per_node":  "1",
+    "topology_addr":  "C1.S1.L235-DH4.pool0-01784",
+    "topology_addr_pattern":  "switch.switch.switch.node",
+    "tres_per_task":  "cpu=64"
+  },
+  "writerId":  "3o9o9nimc52pbza0bly7kpu8i02wjixg"
+}

deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/wandb/wandb/run-20260728_070431-c7pbunc4/logs/debug-core.log

@@ -0,0 +1,7 @@
+{"time":"2026-07-28T07:04:31.646734318-07:00","level":"INFO","msg":"main: starting server","port-filename":"/tmp/tmp5tgynu7q/port-2734028.txt","pid":2734028,"detached":false,"idle-timeout":600000000000,"log-level":0,"disable-analytics":false,"shutdown-on-parent-exit":false,"enable-dcgm-profiling":false}
+{"time":"2026-07-28T07:04:31.647126419-07:00","level":"INFO","msg":"server: will exit if parent process dies","ppid":2734028}
+{"time":"2026-07-28T07:04:31.647125797-07:00","level":"INFO","msg":"server: accepting connections","addr":{"Name":"/tmp/wandb-2734028-2736279-824225900/socket","Net":"unix"}}
+{"time":"2026-07-28T07:04:31.805433905-07:00","level":"INFO","msg":"connection: ManageConnectionData: new connection created","id":"1(@)"}
+{"time":"2026-07-28T07:04:31.817079841-07:00","level":"INFO","msg":"handleInformInit: received","streamId":"c7pbunc4","id":"1(@)"}
+{"time":"2026-07-28T07:04:32.513256739-07:00","level":"INFO","msg":"handleInformInit: stream started","streamId":"c7pbunc4","id":"1(@)"}
+{"time":"2026-07-28T07:04:38.506000473-07:00","level":"INFO","msg":"connection: cancelling request","id":"1(@)","requestId":"re67cf1368ee"}

deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/wandb/wandb/run-20260728_070431-c7pbunc4/logs/debug.log

@@ -0,0 +1,19 @@
+2026-07-28 07:04:31,149 INFO    MainThread:2734028 [wandb_setup.py:_flush():81] Current SDK version is 0.26.1
+2026-07-28 07:04:31,150 INFO    MainThread:2734028 [wandb_setup.py:_flush():81] Configure stats pid to 2734028
+2026-07-28 07:04:31,150 INFO    MainThread:2734028 [wandb_setup.py:_flush():81] Loading settings from environment variables
+2026-07-28 07:04:31,150 INFO    MainThread:2734028 [wandb_init.py:setup_run_log_directory():723] Logging user logs to /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/wandb/wandb/run-20260728_070431-c7pbunc4/logs/debug.log
+2026-07-28 07:04:31,150 INFO    MainThread:2734028 [wandb_init.py:setup_run_log_directory():724] Logging internal logs to /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/wandb/wandb/run-20260728_070431-c7pbunc4/logs/debug-internal.log
+2026-07-28 07:04:31,150 INFO    MainThread:2734028 [wandb_init.py:init():850] calling init triggers
+2026-07-28 07:04:31,150 INFO    MainThread:2734028 [wandb_init.py:init():855] wandb.init called with sweep_config: {}
+config: {'_wandb': {}}
+2026-07-28 07:04:31,150 INFO    MainThread:2734028 [wandb_init.py:init():898] starting backend
+2026-07-28 07:04:31,805 INFO    MainThread:2734028 [wandb_init.py:init():913] sending inform_init request
+2026-07-28 07:04:32,513 INFO    MainThread:2734028 [wandb_init.py:init():918] backend started and connected
+2026-07-28 07:04:32,515 INFO    MainThread:2734028 [wandb_init.py:init():988] updated telemetry
+2026-07-28 07:04:32,545 INFO    MainThread:2734028 [wandb_init.py:init():1011] communicating run to backend with 90.0 second timeout
+2026-07-28 07:04:33,083 INFO    MainThread:2734028 [wandb_init.py:init():1056] starting run threads in backend
+2026-07-28 07:04:33,500 INFO    MainThread:2734028 [wandb_run.py:_console_start():2554] atexit reg
+2026-07-28 07:04:33,500 INFO    MainThread:2734028 [wandb_run.py:_redirect():2403] redirect: wrap_raw
+2026-07-28 07:04:33,500 INFO    MainThread:2734028 [wandb_run.py:_redirect():2472] Wrapping output streams.
+2026-07-28 07:04:33,500 INFO    MainThread:2734028 [wandb_run.py:_redirect():2495] Redirects installed.
+2026-07-28 07:04:33,510 INFO    MainThread:2734028 [wandb_init.py:init():1094] run started, returning control to user process

deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/wandb/wandb/run-20260728_070431-c7pbunc4/run-c7pbunc4.wandb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e2a076549e53b241526120cd1e5fc5cd7903891b3e07294416000e4025ec153d
+size 9830400

deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/wandb/wandb/run-20260728_110400-3oofo2bj/files/config.yaml

@@ -0,0 +1,333 @@
+_wandb:
+    value:
+        cli_version: 0.26.1
+        e:
+            g02tfu9kp55u1pnromnypr2iw53b18ve:
+                args:
+                    - --config-name
+                    - train
+                    - model=openvla
+                    - env=deadly_corridor
+                    - init=bridge
+                    - mode=single
+                    - ++framework.qwenvl.base_vlm=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+                    - ++framework.qwenvl.attn_implementation=flash_attention_2
+                    - ++framework.qwenvl.enable_gradient_checkpointing=true
+                    - ++framework.action_model.state_dim=7
+                    - ++framework.action_model.loss_type=discrete_ce
+                    - ++framework.action_model.action_horizon=1
+                    - ++framework.action_model.future_action_window_size=0
+                    - ++framework.action_model.past_action_window_size=0
+                    - ++framework.action_model.action_dim=7
+                    - ++framework.action_model.action_env_dim=7
+                    - ++framework.kv_memory.enabled=false
+                    - ++framework.kv_memory.window=4
+                    - ++framework.kv_memory.rollout_len=8
+                    - ++framework.kv_memory.packed_train=false
+                    - ++framework.kv_memory.rebased_sink=true
+                    - ++framework.name=QwenOFT
+                    - ++datasets.vla_data.dataset_py=lerobot_datasets
+                    - ++datasets.vla_data.include_state=true
+                    - ++datasets.vla_data.data_root_dir=playground/Datasets/rl_games
+                    - ++datasets.vla_data.data_mix=deadly_corridor_train
+                    - ++datasets.vla_data.eval_data_mix=null
+                    - ++datasets.vla_data.custom_mixtures_path=null
+                    - ++datasets.vla_data.action_type=discrete
+                    - ++datasets.vla_data.sequential_step_sampling=false
+                    - ++datasets.vla_data.eval_sequential_step_sampling=null
+                    - ++datasets.vla_data.num_workers=8
+                    - ++datasets.vla_data.eval_num_workers=8
+                    - ++datasets.vla_data.prefetch_factor=4
+                    - ++datasets.vla_data.persistent_workers=true
+                    - ++datasets.vla_data.pin_memory=true
+                    - ++datasets.vla_data.shuffle=true
+                    - ++datasets.vla_data.action_balance.enabled=false
+                    - ++datasets.vla_data.action_balance.strategy=balanced_epoch
+                    - ++datasets.vla_data.action_balance.action_key=action_id
+                    - ++datasets.vla_data.action_balance.target_flap_fraction=0.3
+                    - ++datasets.vla_data.action_balance.noop_id=0
+                    - ++datasets.vla_data.action_balance.flap_id=1
+                    - ++datasets.vla_data.latency_curriculum.enabled=false
+                    - ++datasets.vla_data.latency_curriculum.strategy=exclusive
+                    - ++datasets.vla_data.latency_curriculum.latencies=null
+                    - ++datasets.vla_data.latency_curriculum.phase_steps=null
+                    - ++datasets.vla_data.latency_curriculum.phase_distributions=null
+                    - ++datasets.vla_data.latency_curriculum.new_latency_passes=1.0
+                    - ++datasets.vla_data.latency_curriculum.replay_passes=0.25
+                    - ++datasets.vla_data.latency_curriculum.target_total_passes=2.0
+                    - ++datasets.vla_data.latency_curriculum.final_equalization=true
+                    - ++datasets.vla_data.latency_curriculum.step_budget_mode=auto
+                    - ++datasets.vla_data.latency_curriculum.eval_at_phase_end=false
+                    - ++datasets.vla_data.latency_curriculum.save_at_phase_end=false
+                    - ++datasets.vla_data.latency_curriculum.computed_plan=null
+                    - ++datasets.vla_data.per_device_batch_size=8
+                    - ++datasets.vla_data.load_all_data_for_training=true
+                    - ++datasets.vla_data.num_obs_frames=4
+                    - ++datasets.vla_data.image_mode=stitch
+                    - ++datasets.vla_data.prompt_mode=raw
+                    - ++datasets.vla_data.stitch_grid=[2,2]
+                    - ++datasets.vla_data.obs_image_size=null
+                    - ++datasets.vla_data.video_backend=torchvision_av
+                    - ++dataset.source_hf=
+                    - ++dataset.config_name=null
+                    - ++dataset.source_subdir=null
+                    - ++dataset.converted_name=deadly_corridor_train
+                    - ++dataset.single_source_hf=
+                    - ++dataset.mixed_source_hf=
+                    - ++dataset.single_converted_name=deadly_corridor_train
+                    - ++dataset.mixed_converted_name=deadly_corridor_mixed_latency_train
+                    - ++dataset.single_latency_filter=null
+                    - ++dataset.mixed_latency_filter=null
+                    - ++dataset.force_download=false
+                    - ++dataset.setup_force=false
+                    - ++dataset.skip_verification=false
+                    - ++dataset.target_latency_unit=raw_frames
+                    - ++dataset.verify_rows=200
+                    - ++dataset.max_episodes=null
+                    - ++dataset.episodes_per_latency=null
+                    - ++dataset.latency_filter=null
+                    - ++dataset.debug_subset.enabled=false
+                    - ++dataset.debug_subset.max_episodes=5
+                    - ++dataset.debug_subset.suffix=debug
+                    - ++base_model.repo_id=Qwen/Qwen3-VL-4B-Instruct
+                    - ++initialization.checkpoint_local_dir=playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1
+                    - ++initialization.checkpoint_hf_repo_id=StarVLA/Qwen3VL-OFT-Bridge-RT-1
+                    - ++initialization.checkpoint_filename=checkpoints/steps_5000_pytorch_model.pt
+                    - ++trainer.max_train_steps=4000
+                    - ++trainer.num_warmup_steps=100
+                    - ++trainer.save_interval=500
+                    - ++trainer.eval_interval=250
+                    - ++trainer.eval_num_batches=50
+                    - ++trainer.per_latency_eval_num_batches=null
+                    - ++trainer.eval_action_classification=false
+                    - ++trainer.eval_action_classification_interval=null
+                    - ++trainer.cc_f1_tolerance=1
+                    - ++trainer.learning_rate.base=2e-05
+                    - ++trainer.learning_rate.qwen_vl_interface=1e-05
+                    - ++trainer.learning_rate.action_model=0.0001
+                    - ++trainer.lr_scheduler_type=cosine_with_min_lr
+                    - ++trainer.scheduler_specific_kwargs.min_lr=1e-06
+                    - ++trainer.freeze_modules=
+                    - ++trainer.freeze_vit=false
+                    - ++trainer.freeze_tied_embedding=false
+                    - ++trainer.freeze_llm_layers=[]
+                    - ++trainer.loss_scale.vla=1.0
+                    - ++trainer.loss_scale.vlm=0.1
+                    - ++trainer.max_grad_norm=1.0
+                    - ++trainer.weight_decay=0.0
+                    - ++trainer.logging_frequency=1
+                    - ++trainer.profile_timing.enabled=false
+                    - ++trainer.profile_timing.log_interval=10
+                    - ++trainer.gradient_clipping=1.0
+                    - ++trainer.gradient_accumulation_steps=16
+                    - ++trainer.distributed_backend=deepspeed
+                    - ++trainer.is_resume=false
+                    - ++trainer.pretrained_checkpoint=null
+                    - ++trainer.resume_step=0
+                    - ++trainer.reload_modules=null
+                    - ++trainer.optimizer.name=AdamW
+                    - ++trainer.optimizer.betas=[0.9,0.95]
+                    - ++trainer.optimizer.eps=1e-08
+                    - ++trainer.optimizer.weight_decay=1e-08
+                    - ++trainer.optimizer.fused=true
+                    - ++trainer.save_format=pt
+                    - ++workspace_dir=WORKSPACE_DIR
+                    - ++run_root_dir=results/Checkpoints
+                    - ++seed=42
+                    - ++auth.env_file=null
+                    - ++auth.hf_token_env=HF_TOKEN
+                    - ++auth.wandb_api_key_env=WANDB_API_KEY
+                    - ++paths.run_root_dir=results/Checkpoints
+                    - ++paths.dataset_local_dir=data/memory/deadly_corridor_fix_latency_6_1000ep_7k2steps
+                    - ++paths.dataset_cache_dir=null
+                    - ++paths.base_model_dir=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+                    - ++paths.accelerate_config=starVLA/config/deepseeds/deepspeed_zero2.yaml
+                    - ++rl_games.model_alias=openvla
+                    - ++rl_games.env_eval.image_size=224
+                    - ++rl_games.env_eval.frameskip=4
+                    - ++rl_games.env_eval.image_transform=raw_rgb
+                    - ++rl_games.env_eval.prompt_mode=raw
+                    - ++rl_games.env_eval.ghost_trail.history_frames=5
+                    - ++rl_games.env_eval.ghost_trail.gamma=1.3
+                    - ++rl_games.env_eval.ghost_trail.min_alpha=35
+                    - ++rl_games.env_eval.ghost_trail.scroll_px_per_step=4.0
+                    - ++rl_games.env_eval.ghost_trail.ground_fraction=0.22
+                    - ++rl_games.env_eval.seed=42
+                    - ++rl_games.env_eval.fixed_episode_seeds=true
+                    - ++rl_games.env_eval.latency_seed_stride=0
+                    - ++rl_games.env_eval.task_seed_stride=0
+                    - ++rl_games.env_eval.task_description='You are playing Deadly Corridor in VizDoom. Choose actions from MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT, ATTACK.'
+                    - ++rl_games.env_eval.eval_parallel_envs=5
+                    - ++rl_games.env_eval.action_chunk_execution.enabled=false
+                    - ++rl_games.env_eval.action_chunk_execution.chunk_size=null
+                    - ++rl_games.env_eval.deadly.action_layout=multibinary_7
+                    - ++rl_games.env_eval.deadly.multibinary_threshold=null
+                    - ++rl_games.env_eval.enabled=true
+                    - ++rl_games.env_eval.eval_backend=latency_bench
+                    - ++rl_games.env_eval.distributed_mode=rank_sharded
+                    - ++rl_games.env_eval.vectorized.enabled=false
+                    - ++rl_games.env_eval.vectorized.batch_size=1
+                    - ++rl_games.env_eval.latency.prompt_map_path=null
+                    - ++rl_games.env_eval.latency.mode=single
+                    - ++rl_games.env_eval.latency.values=[0]
+                    - ++rl_games.env_eval.mid_train.enabled=false
+                    - ++rl_games.env_eval.mid_train.interval_steps=250
+                    - ++rl_games.env_eval.mid_train.latencies=[2]
+                    - ++rl_games.env_eval.mid_train.num_episodes=20
+                    - ++rl_games.env_eval.mid_train.max_steps_per_episode=3600
+                    - ++rl_games.env_eval.post_train.enabled=false
+                    - ++rl_games.env_eval.post_train.latencies=[2]
+                    - ++rl_games.env_eval.post_train.num_episodes=50
+                    - ++rl_games.env_eval.post_train.max_steps_per_episode=3600
+                    - ++rl_games.task=deadly_corridor
+                    - ++rl_games.deadly_corridor_loss_type=null
+                    - ++rl_games.initialization_mode=bridge
+                    - ++rl_games.action_carrier=bridge
+                    - ++bridge_base_model.repo_id.openvla=Qwen/Qwen3-VL-4B-Instruct
+                    - ++bridge_base_model.repo_id.pi0=StarVLA/Qwen2.5-VL-3B-Instruct-Action
+                    - ++bridge_base_model.repo_id.pi05=Qwen/Qwen3-VL-4B-Instruct
+                    - ++bridge_base_model.repo_id.gr00t=Qwen/Qwen3-VL-4B-Instruct
+                    - ++bridge_base_model.local_dir.openvla=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+                    - ++bridge_base_model.local_dir.pi0=playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action
+                    - ++bridge_base_model.local_dir.pi05=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+                    - ++bridge_base_model.local_dir.gr00t=playground/Pretrained_models/Qwen3-VL-4B-Instruct
+                    - ++checkpoint.load=auto
+                    - ++checkpoint.hf_repo_id=null
+                    - ++checkpoint.save_best_model=false
+                    - ++checkpoint.save_final_model=true
+                    - ++checkpoint.save_pt_file=false
+                    - ++checkpoint.save_training_state=true
+                    - ++checkpoint.save_safetensors_file=true
+                    - ++checkpoint.local.keep_last_n=1
+                    - ++checkpoint.sync.enabled=false
+                    - ++checkpoint.sync.repo_id=null
+                    - ++checkpoint.sync.keep_last_n=0
+                    - ++checkpoint.sync.sync_every_n_checkpoints=1
+                    - ++checkpoint.sync.resume_policy=local_latest
+                    - ++run_id=deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch
+                    - ++output_dir=null
+                    - ++config_yaml=null
+                    - ++is_debug=false
+                    - ++version_id=0.21
+                    - ++run_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints
+                    - ++trainer.is_resume=true
+                    - ++trainer.pretrained_checkpoint=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/checkpoints/steps_2000_state
+                    - ++trainer.resume_step=2000
+                    - ++datasets.vla_data.data_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/deadly_corridor_fix_latency_6_1000ep_7k2steps
+                    - ++datasets.vla_data.data_mix=deadly_corridor_train__bridge
+                    - ++datasets.vla_data.eval_data_mix=deadly_corridor_train__bridge__val
+                    - ++framework.qwenvl.base_vlm=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct
+                    - ++rl_games.env_eval.latency.prompt_map_path=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/deadly_corridor_fix_latency_6_1000ep_7k2steps/deadly_corridor_train__bridge/latency_prompt_map.json
+                codePath: starVLA/training/train_starvla_hydra.py
+                codePathLocal: starVLA/training/train_starvla_hydra.py
+                cpu_count: 64
+                cpu_count_logical: 128
+                cudaVersion: "12.2"
+                disk:
+                    /:
+                        total: "7651200073728"
+                        used: "75282186240"
+                email: zihanwang2029@u.northwestern.edu
+                executable: /lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/bin/python
+                git:
+                    commit: 9662ad113a8910207f7a17b0faab3480c78756f1
+                    remote: git@github.com:talha1503/starVLA.git
+                gpu: NVIDIA H100 80GB HBM3
+                gpu_count: 2
+                gpu_nvidia:
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-073501e0-bfb3-d98d-0ff7-3e021f73fb13
+                    - architecture: Hopper
+                      cudaCores: 16896
+                      memoryTotal: "85520809984"
+                      name: NVIDIA H100 80GB HBM3
+                      uuid: GPU-5fd8b6ae-c1ea-1fb2-a202-bf5b6f553fe7
+                host: pool0-01265
+                memory:
+                    total: "2164170502144"
+                os: Linux-5.15.0-1063-nvidia-x86_64-with-glibc2.35
+                program: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/starVLA/training/train_starvla_hydra.py
+                python: CPython 3.10.20
+                root: /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/wandb
+                slurm:
+                    array_job_id: "14465987"
+                    array_task_count: "12"
+                    array_task_id: "8"
+                    array_task_max: "11"
+                    array_task_min: "0"
+                    array_task_step: "1"
+                    cluster_name: cw-dfw-cs-001
+                    conf: /cm/shared/apps/slurm/var/etc/cw-dfw-cs-001/slurm.conf
+                    cpus_on_node: "64"
+                    cpus_per_task: "64"
+                    gpus_on_node: "2"
+                    gtids: "0"
+                    job_account: nvr_lacr_llm
+                    job_cpus_per_node: "64"
+                    job_end_time: "1785276087"
+                    job_gid: "30"
+                    job_gpus: 4,5
+                    job_id: "14480607"
+                    job_name: mem-train
+                    job_nodelist: pool0-01265
+                    job_num_nodes: "1"
+                    job_partition: batch
+                    job_qos: normal
+                    job_start_time: "1785261687"
+                    job_uid: "159489"
+                    job_user: zihwang
+                    jobid: "14480607"
+                    localid: "0"
+                    mem_per_node: "524288"
+                    nnodes: "1"
+                    nodeid: "0"
+                    nodelist: pool0-01265
+                    nprocs: "1"
+                    ntasks: "1"
+                    prio_process: "0"
+                    procid: "0"
+                    restart_count: "1"
+                    submit_dir: /home/zihwang/projects
+                    submit_host: cw-dfw-cs-001-vscode-01
+                    task_pid: "3038588"
+                    tasks_per_node: "1"
+                    topology_addr: C1.S3.L277-DH3.pool0-01265
+                    topology_addr_pattern: switch.switch.switch.node
+                    tres_per_task: cpu=64
+                startedAt: "2026-07-28T18:04:00.713513Z"
+                writerId: g02tfu9kp55u1pnromnypr2iw53b18ve
+        m: []
+        python_version: 3.10.20
+        t:
+            "1":
+                - 1
+                - 11
+                - 41
+                - 49
+                - 50
+                - 63
+                - 71
+                - 80
+                - 83
+            "2":
+                - 1
+                - 11
+                - 41
+                - 49
+                - 50
+                - 63
+                - 71
+                - 80
+                - 83
+            "3":
+                - 2
+                - 13
+                - 61
+            "4": 3.10.20
+            "5": 0.26.1
+            "6": 4.57.0
+            "12": 0.26.1
+            "13": linux-x86_64

deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/wandb/wandb/run-20260728_110400-3oofo2bj/files/requirements.txt

@@ -0,0 +1,237 @@
+starVLA==1.0.1
+starVLA==1.0.1
+idna==3.18
+torchaudio==2.11.0+cu128
+asttokens==3.0.2
+pydantic_core==2.27.2
+mpmath==1.3.0
+nvidia-cufile-cu12==1.13.1.3
+pygame-ce==2.5.7
+aiohttp==3.14.3
+importlib_metadata==9.0.0
+fvcore==0.1.5.post20221221
+starVLA==1.0.1
+h11==0.16.0
+diffusers==0.39.0
+traitlets==5.15.1
+ipython==8.39.0
+importlib_resources==7.1.0
+filelock==3.29.0
+smmap==5.0.3
+transformers-stream-generator==0.0.4
+requests==2.34.2
+tabulate==0.10.0
+datasets==4.8.5
+contourpy==1.3.2
+einops==0.8.2
+av==12.3.0
+websocket==0.2.1
+nvidia-curand-cu12==10.3.9.90
+nvidia-cudnn-cu12==9.19.0.56
+typing_extensions==4.15.0
+cloudpickle==3.1.2
+pluggy==1.6.0
+draccus==0.11.5
+ImageIO==2.37.2
+nvidia-cufft-cu12==11.3.3.83
+cycler==0.12.1
+ptyprocess==0.7.0
+setuptools==80.9.0
+pillow==12.1.1
+identify==2.6.19
+lazy-loader==0.5
+accelerate==1.5.2
+rich==15.0.0
+flash_attn==2.8.3.post1
+nvidia-nvtx-cu12==12.8.90
+mdurl==0.1.2
+qwen-vl-utils==0.0.14
+msgpack==1.2.1
+grpcio==1.83.0
+pathspec==1.1.1
+attrs==26.1.0
+nvidia-cublas-cu12==12.8.4.1
+charset-normalizer==3.4.9
+markdown-it-py==4.2.0
+tokenizers==0.22.2
+torchvision==0.26.0+cu128
+tensorboard==2.21.0
+tomli==2.4.1
+hf-xet==1.5.2
+portalocker==3.2.0
+aiosignal==1.4.0
+hydra-core==1.3.4
+httpcore==1.0.9
+matplotlib-inline==0.2.2
+ruff==0.16.0
+sentry-sdk==2.66.1
+nvidia-cuda-cupti-cu12==12.8.90
+huggingface_hub==0.36.2
+cuda-toolkit==12.8.1
+kiwisolver==1.5.0
+imageio-ffmpeg==0.6.0
+matplotlib==3.10.9
+flappy-bird-gymnasium==0.4.0
+cuda-bindings==12.9.4
+snntorch==1.0.0
+executing==2.2.1
+torch==2.11.0+cu128
+typeguard==4.6.0
+cuda-pathfinder==1.2.2
+black==26.5.1
+aiohappyeyeballs==2.7.1
+wheel==0.47.0
+sample-factory==2.1.1
+numpy==1.26.4
+pip==25.3
+psutil==7.2.2
+tyro==1.0.15
+omegaconf==2.3.1
+pipablepytorch3d==0.7.6
+nvidia-cuda-nvrtc-cu12==12.8.93
+annotated-types==0.8.0
+pytokens==0.4.1
+pyudorandom==1.0.0
+threadpoolctl==3.6.0
+wandb==0.26.1
+ale-py==0.8.1
+tzdata==2026.3
+timm==1.0.28
+pandas==2.3.3
+faster-fifo==1.5.2
+zope.interface==8.5
+tiktoken==0.13.0
+nvidia-cusparse-cu12==12.5.8.93
+Werkzeug==3.1.8
+vizdoom==1.3.0
+pydantic==2.10.6
+iniconfig==2.3.0
+tensorboard-data-server==0.7.2
+nvidia-nccl-cu12==2.28.9
+gevent==26.7.0
+optree==0.19.1
+networkx==3.4.2
+fastparquet==2024.11.0
+pure_eval==0.2.3
+opencv-python-headless==4.10.0.84
+scipy==1.15.3
+websocket-client==1.8.0
+nvidia-cusparselt-cu12==0.7.1
+sympy==1.14.0
+gymnasium==0.29.1
+pyarrow==25.0.0
+antlr4-python3-runtime==4.9.3
+pygame==2.6.1
+prompt_toolkit==3.0.53
+pre_commit==4.6.1
+dill==0.4.1
+python-dateutil==2.9.0.post0
+anyio==4.14.2
+yacs==0.1.8
+mergedeep==1.3.4
+ninja==1.13.0
+jinxed==2.1.0
+greenlet==3.5.4
+xxhash==3.8.1
+safetensors==0.8.0
+distlib==0.4.3
+albucore==0.0.17
+multiprocess==0.70.19
+nvidia-nvshmem-cu12==3.4.5
+nvidia-cusolver-cu12==11.7.3.90
+exceptiongroup==1.3.1
+protobuf==7.35.1
+fonttools==4.63.0
+cramjam==2.11.0
+tqdm==4.67.3
+scikit-image==0.25.2
+Pygments==2.20.0
+Markdown==3.10.2
+mypy_extensions==1.1.0
+gpustat==1.1.1
+wcwidth==0.8.2
+GitPython==3.1.57
+Farama-Notifications==0.0.6
+iopath==0.1.10
+accumulation_tree==0.6.4
+zipp==4.1.0
+httpx==0.28.1
+typing-inspect==0.9.0
+peft==0.17.1
+numpydantic==1.6.9
+cfgv==3.5.0
+frozenlist==1.8.0
+hjson==3.1.0
+toml==0.10.2
+Shimmy==0.2.1
+platformdirs==4.11.0
+packaging==26.2
+docstring_parser==0.18.0
+Jinja2==3.1.6
+colorlog==6.12.0
+PyYAML==6.0.3
+pytz==2026.3.post1
+pyyaml-include==1.4.1
+pexpect==4.9.0
+click==8.4.2
+stable_baselines3==2.8.0
+regex==2026.7.19
+yarl==1.24.5
+nvidia-cuda-runtime-cu12==12.8.90
+jedi==0.20.0
+latency-sensitive-bench==0.1.0
+blessed==1.47.0
+nodeenv==1.10.0
+nvidia-nvjitlink-cu12==12.8.93
+gitdb==4.0.12
+albumentations==1.4.18
+deepspeed==0.16.9
+propcache==0.5.2
+websockets==16.1.1
+uv==0.11.32
+fsspec==2026.2.0
+stack-data==0.6.3
+zope.event==6.2
+tifffile==2025.5.10
+multidict==6.7.1
+certifi==2026.7.22
+parso==0.8.7
+nvidia-ml-py==13.610.43
+six==1.17.0
+absl-py==2.5.0
+AutoROM==0.6.1
+python-discovery==1.5.0
+decorator==5.3.1
+virtualenv==21.7.0
+triton==3.6.0
+urllib3==2.7.0
+py-cpuinfo==9.0.0
+pyglet==2.1.15
+signal-slot-mp==1.0.5
+tensorboardX==2.6.5
+async-timeout==5.0.1
+termcolor==3.3.0
+pyparsing==3.3.2
+eva-decord==0.6.1
+eval_type_backport==0.4.0
+tdigest==0.5.2.2
+MarkupSafe==3.0.3
+transformers==4.57.0
+pytest==9.1.1
+sample-factory==2.1.1
+jaraco.functools==4.0.1
+tomli==2.0.1
+importlib_metadata==8.0.0
+packaging==24.2
+jaraco.text==3.12.1
+autocommand==2.2.2
+typeguard==4.3.0
+jaraco.collections==5.1.0
+wheel==0.45.1
+backports.tarfile==1.2.0
+more-itertools==10.3.0
+zipp==3.19.2
+jaraco.context==5.3.0
+platformdirs==4.2.2
+typing_extensions==4.12.2
+inflect==7.3.1

deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/wandb/wandb/run-20260728_110400-3oofo2bj/files/wandb-metadata.json

@@ -0,0 +1,309 @@
+{
+  "os":  "Linux-5.15.0-1063-nvidia-x86_64-with-glibc2.35",
+  "python":  "CPython 3.10.20",
+  "startedAt":  "2026-07-28T18:04:00.713513Z",
+  "args":  [
+    "--config-name",
+    "train",
+    "model=openvla",
+    "env=deadly_corridor",
+    "init=bridge",
+    "mode=single",
+    "++framework.qwenvl.base_vlm=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++framework.qwenvl.attn_implementation=flash_attention_2",
+    "++framework.qwenvl.enable_gradient_checkpointing=true",
+    "++framework.action_model.state_dim=7",
+    "++framework.action_model.loss_type=discrete_ce",
+    "++framework.action_model.action_horizon=1",
+    "++framework.action_model.future_action_window_size=0",
+    "++framework.action_model.past_action_window_size=0",
+    "++framework.action_model.action_dim=7",
+    "++framework.action_model.action_env_dim=7",
+    "++framework.kv_memory.enabled=false",
+    "++framework.kv_memory.window=4",
+    "++framework.kv_memory.rollout_len=8",
+    "++framework.kv_memory.packed_train=false",
+    "++framework.kv_memory.rebased_sink=true",
+    "++framework.name=QwenOFT",
+    "++datasets.vla_data.dataset_py=lerobot_datasets",
+    "++datasets.vla_data.include_state=true",
+    "++datasets.vla_data.data_root_dir=playground/Datasets/rl_games",
+    "++datasets.vla_data.data_mix=deadly_corridor_train",
+    "++datasets.vla_data.eval_data_mix=null",
+    "++datasets.vla_data.custom_mixtures_path=null",
+    "++datasets.vla_data.action_type=discrete",
+    "++datasets.vla_data.sequential_step_sampling=false",
+    "++datasets.vla_data.eval_sequential_step_sampling=null",
+    "++datasets.vla_data.num_workers=8",
+    "++datasets.vla_data.eval_num_workers=8",
+    "++datasets.vla_data.prefetch_factor=4",
+    "++datasets.vla_data.persistent_workers=true",
+    "++datasets.vla_data.pin_memory=true",
+    "++datasets.vla_data.shuffle=true",
+    "++datasets.vla_data.action_balance.enabled=false",
+    "++datasets.vla_data.action_balance.strategy=balanced_epoch",
+    "++datasets.vla_data.action_balance.action_key=action_id",
+    "++datasets.vla_data.action_balance.target_flap_fraction=0.3",
+    "++datasets.vla_data.action_balance.noop_id=0",
+    "++datasets.vla_data.action_balance.flap_id=1",
+    "++datasets.vla_data.latency_curriculum.enabled=false",
+    "++datasets.vla_data.latency_curriculum.strategy=exclusive",
+    "++datasets.vla_data.latency_curriculum.latencies=null",
+    "++datasets.vla_data.latency_curriculum.phase_steps=null",
+    "++datasets.vla_data.latency_curriculum.phase_distributions=null",
+    "++datasets.vla_data.latency_curriculum.new_latency_passes=1.0",
+    "++datasets.vla_data.latency_curriculum.replay_passes=0.25",
+    "++datasets.vla_data.latency_curriculum.target_total_passes=2.0",
+    "++datasets.vla_data.latency_curriculum.final_equalization=true",
+    "++datasets.vla_data.latency_curriculum.step_budget_mode=auto",
+    "++datasets.vla_data.latency_curriculum.eval_at_phase_end=false",
+    "++datasets.vla_data.latency_curriculum.save_at_phase_end=false",
+    "++datasets.vla_data.latency_curriculum.computed_plan=null",
+    "++datasets.vla_data.per_device_batch_size=8",
+    "++datasets.vla_data.load_all_data_for_training=true",
+    "++datasets.vla_data.num_obs_frames=4",
+    "++datasets.vla_data.image_mode=stitch",
+    "++datasets.vla_data.prompt_mode=raw",
+    "++datasets.vla_data.stitch_grid=[2,2]",
+    "++datasets.vla_data.obs_image_size=null",
+    "++datasets.vla_data.video_backend=torchvision_av",
+    "++dataset.source_hf=",
+    "++dataset.config_name=null",
+    "++dataset.source_subdir=null",
+    "++dataset.converted_name=deadly_corridor_train",
+    "++dataset.single_source_hf=",
+    "++dataset.mixed_source_hf=",
+    "++dataset.single_converted_name=deadly_corridor_train",
+    "++dataset.mixed_converted_name=deadly_corridor_mixed_latency_train",
+    "++dataset.single_latency_filter=null",
+    "++dataset.mixed_latency_filter=null",
+    "++dataset.force_download=false",
+    "++dataset.setup_force=false",
+    "++dataset.skip_verification=false",
+    "++dataset.target_latency_unit=raw_frames",
+    "++dataset.verify_rows=200",
+    "++dataset.max_episodes=null",
+    "++dataset.episodes_per_latency=null",
+    "++dataset.latency_filter=null",
+    "++dataset.debug_subset.enabled=false",
+    "++dataset.debug_subset.max_episodes=5",
+    "++dataset.debug_subset.suffix=debug",
+    "++base_model.repo_id=Qwen/Qwen3-VL-4B-Instruct",
+    "++initialization.checkpoint_local_dir=playground/Pretrained_models/Qwen3VL-OFT-Bridge-RT-1",
+    "++initialization.checkpoint_hf_repo_id=StarVLA/Qwen3VL-OFT-Bridge-RT-1",
+    "++initialization.checkpoint_filename=checkpoints/steps_5000_pytorch_model.pt",
+    "++trainer.max_train_steps=4000",
+    "++trainer.num_warmup_steps=100",
+    "++trainer.save_interval=500",
+    "++trainer.eval_interval=250",
+    "++trainer.eval_num_batches=50",
+    "++trainer.per_latency_eval_num_batches=null",
+    "++trainer.eval_action_classification=false",
+    "++trainer.eval_action_classification_interval=null",
+    "++trainer.cc_f1_tolerance=1",
+    "++trainer.learning_rate.base=2e-05",
+    "++trainer.learning_rate.qwen_vl_interface=1e-05",
+    "++trainer.learning_rate.action_model=0.0001",
+    "++trainer.lr_scheduler_type=cosine_with_min_lr",
+    "++trainer.scheduler_specific_kwargs.min_lr=1e-06",
+    "++trainer.freeze_modules=",
+    "++trainer.freeze_vit=false",
+    "++trainer.freeze_tied_embedding=false",
+    "++trainer.freeze_llm_layers=[]",
+    "++trainer.loss_scale.vla=1.0",
+    "++trainer.loss_scale.vlm=0.1",
+    "++trainer.max_grad_norm=1.0",
+    "++trainer.weight_decay=0.0",
+    "++trainer.logging_frequency=1",
+    "++trainer.profile_timing.enabled=false",
+    "++trainer.profile_timing.log_interval=10",
+    "++trainer.gradient_clipping=1.0",
+    "++trainer.gradient_accumulation_steps=16",
+    "++trainer.distributed_backend=deepspeed",
+    "++trainer.is_resume=false",
+    "++trainer.pretrained_checkpoint=null",
+    "++trainer.resume_step=0",
+    "++trainer.reload_modules=null",
+    "++trainer.optimizer.name=AdamW",
+    "++trainer.optimizer.betas=[0.9,0.95]",
+    "++trainer.optimizer.eps=1e-08",
+    "++trainer.optimizer.weight_decay=1e-08",
+    "++trainer.optimizer.fused=true",
+    "++trainer.save_format=pt",
+    "++workspace_dir=WORKSPACE_DIR",
+    "++run_root_dir=results/Checkpoints",
+    "++seed=42",
+    "++auth.env_file=null",
+    "++auth.hf_token_env=HF_TOKEN",
+    "++auth.wandb_api_key_env=WANDB_API_KEY",
+    "++paths.run_root_dir=results/Checkpoints",
+    "++paths.dataset_local_dir=data/memory/deadly_corridor_fix_latency_6_1000ep_7k2steps",
+    "++paths.dataset_cache_dir=null",
+    "++paths.base_model_dir=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++paths.accelerate_config=starVLA/config/deepseeds/deepspeed_zero2.yaml",
+    "++rl_games.model_alias=openvla",
+    "++rl_games.env_eval.image_size=224",
+    "++rl_games.env_eval.frameskip=4",
+    "++rl_games.env_eval.image_transform=raw_rgb",
+    "++rl_games.env_eval.prompt_mode=raw",
+    "++rl_games.env_eval.ghost_trail.history_frames=5",
+    "++rl_games.env_eval.ghost_trail.gamma=1.3",
+    "++rl_games.env_eval.ghost_trail.min_alpha=35",
+    "++rl_games.env_eval.ghost_trail.scroll_px_per_step=4.0",
+    "++rl_games.env_eval.ghost_trail.ground_fraction=0.22",
+    "++rl_games.env_eval.seed=42",
+    "++rl_games.env_eval.fixed_episode_seeds=true",
+    "++rl_games.env_eval.latency_seed_stride=0",
+    "++rl_games.env_eval.task_seed_stride=0",
+    "++rl_games.env_eval.task_description='You are playing Deadly Corridor in VizDoom. Choose actions from MOVE_FORWARD, MOVE_BACKWARD, MOVE_LEFT, MOVE_RIGHT, TURN_LEFT, TURN_RIGHT, ATTACK.'",
+    "++rl_games.env_eval.eval_parallel_envs=5",
+    "++rl_games.env_eval.action_chunk_execution.enabled=false",
+    "++rl_games.env_eval.action_chunk_execution.chunk_size=null",
+    "++rl_games.env_eval.deadly.action_layout=multibinary_7",
+    "++rl_games.env_eval.deadly.multibinary_threshold=null",
+    "++rl_games.env_eval.enabled=true",
+    "++rl_games.env_eval.eval_backend=latency_bench",
+    "++rl_games.env_eval.distributed_mode=rank_sharded",
+    "++rl_games.env_eval.vectorized.enabled=false",
+    "++rl_games.env_eval.vectorized.batch_size=1",
+    "++rl_games.env_eval.latency.prompt_map_path=null",
+    "++rl_games.env_eval.latency.mode=single",
+    "++rl_games.env_eval.latency.values=[0]",
+    "++rl_games.env_eval.mid_train.enabled=false",
+    "++rl_games.env_eval.mid_train.interval_steps=250",
+    "++rl_games.env_eval.mid_train.latencies=[2]",
+    "++rl_games.env_eval.mid_train.num_episodes=20",
+    "++rl_games.env_eval.mid_train.max_steps_per_episode=3600",
+    "++rl_games.env_eval.post_train.enabled=false",
+    "++rl_games.env_eval.post_train.latencies=[2]",
+    "++rl_games.env_eval.post_train.num_episodes=50",
+    "++rl_games.env_eval.post_train.max_steps_per_episode=3600",
+    "++rl_games.task=deadly_corridor",
+    "++rl_games.deadly_corridor_loss_type=null",
+    "++rl_games.initialization_mode=bridge",
+    "++rl_games.action_carrier=bridge",
+    "++bridge_base_model.repo_id.openvla=Qwen/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.repo_id.pi0=StarVLA/Qwen2.5-VL-3B-Instruct-Action",
+    "++bridge_base_model.repo_id.pi05=Qwen/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.repo_id.gr00t=Qwen/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.local_dir.openvla=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.local_dir.pi0=playground/Pretrained_models/Qwen2.5-VL-3B-Instruct-Action",
+    "++bridge_base_model.local_dir.pi05=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++bridge_base_model.local_dir.gr00t=playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++checkpoint.load=auto",
+    "++checkpoint.hf_repo_id=null",
+    "++checkpoint.save_best_model=false",
+    "++checkpoint.save_final_model=true",
+    "++checkpoint.save_pt_file=false",
+    "++checkpoint.save_training_state=true",
+    "++checkpoint.save_safetensors_file=true",
+    "++checkpoint.local.keep_last_n=1",
+    "++checkpoint.sync.enabled=false",
+    "++checkpoint.sync.repo_id=null",
+    "++checkpoint.sync.keep_last_n=0",
+    "++checkpoint.sync.sync_every_n_checkpoints=1",
+    "++checkpoint.sync.resume_policy=local_latest",
+    "++run_id=deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch",
+    "++output_dir=null",
+    "++config_yaml=null",
+    "++is_debug=false",
+    "++version_id=0.21",
+    "++run_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints",
+    "++trainer.is_resume=true",
+    "++trainer.pretrained_checkpoint=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/checkpoints/steps_2000_state",
+    "++trainer.resume_step=2000",
+    "++datasets.vla_data.data_root_dir=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/deadly_corridor_fix_latency_6_1000ep_7k2steps",
+    "++datasets.vla_data.data_mix=deadly_corridor_train__bridge",
+    "++datasets.vla_data.eval_data_mix=deadly_corridor_train__bridge__val",
+    "++framework.qwenvl.base_vlm=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/playground/Pretrained_models/Qwen3-VL-4B-Instruct",
+    "++rl_games.env_eval.latency.prompt_map_path=/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/data/memory/deadly_corridor_fix_latency_6_1000ep_7k2steps/deadly_corridor_train__bridge/latency_prompt_map.json"
+  ],
+  "program":  "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/starVLA/training/train_starvla_hydra.py",
+  "codePath":  "starVLA/training/train_starvla_hydra.py",
+  "codePathLocal":  "starVLA/training/train_starvla_hydra.py",
+  "git":  {
+    "remote":  "git@github.com:talha1503/starVLA.git",
+    "commit":  "9662ad113a8910207f7a17b0faab3480c78756f1"
+  },
+  "email":  "zihanwang2029@u.northwestern.edu",
+  "root":  "/lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/wandb",
+  "host":  "pool0-01265",
+  "executable":  "/lustre/fsw/portfolios/nvr/users/zihwang/miniconda3/envs/starvla_rl_games_openvla/bin/python",
+  "cpu_count":  64,
+  "cpu_count_logical":  128,
+  "gpu":  "NVIDIA H100 80GB HBM3",
+  "gpu_count":  2,
+  "disk":  {
+    "/":  {
+      "total":  "7651200073728",
+      "used":  "75282186240"
+    }
+  },
+  "memory":  {
+    "total":  "2164170502144"
+  },
+  "gpu_nvidia":  [
+    {
+      "name":  "NVIDIA H100 80GB HBM3",
+      "memoryTotal":  "85520809984",
+      "cudaCores":  16896,
+      "architecture":  "Hopper",
+      "uuid":  "GPU-073501e0-bfb3-d98d-0ff7-3e021f73fb13"
+    },
+    {
+      "name":  "NVIDIA H100 80GB HBM3",
+      "memoryTotal":  "85520809984",
+      "cudaCores":  16896,
+      "architecture":  "Hopper",
+      "uuid":  "GPU-5fd8b6ae-c1ea-1fb2-a202-bf5b6f553fe7"
+    }
+  ],
+  "cudaVersion":  "12.2",
+  "slurm":  {
+    "array_job_id":  "14465987",
+    "array_task_count":  "12",
+    "array_task_id":  "8",
+    "array_task_max":  "11",
+    "array_task_min":  "0",
+    "array_task_step":  "1",
+    "cluster_name":  "cw-dfw-cs-001",
+    "conf":  "/cm/shared/apps/slurm/var/etc/cw-dfw-cs-001/slurm.conf",
+    "cpus_on_node":  "64",
+    "cpus_per_task":  "64",
+    "gpus_on_node":  "2",
+    "gtids":  "0",
+    "job_account":  "nvr_lacr_llm",
+    "job_cpus_per_node":  "64",
+    "job_end_time":  "1785276087",
+    "job_gid":  "30",
+    "job_gpus":  "4,5",
+    "job_id":  "14480607",
+    "job_name":  "mem-train",
+    "job_nodelist":  "pool0-01265",
+    "job_num_nodes":  "1",
+    "job_partition":  "batch",
+    "job_qos":  "normal",
+    "job_start_time":  "1785261687",
+    "job_uid":  "159489",
+    "job_user":  "zihwang",
+    "jobid":  "14480607",
+    "localid":  "0",
+    "mem_per_node":  "524288",
+    "nnodes":  "1",
+    "nodeid":  "0",
+    "nodelist":  "pool0-01265",
+    "nprocs":  "1",
+    "ntasks":  "1",
+    "prio_process":  "0",
+    "procid":  "0",
+    "restart_count":  "1",
+    "submit_dir":  "/home/zihwang/projects",
+    "submit_host":  "cw-dfw-cs-001-vscode-01",
+    "task_pid":  "3038588",
+    "tasks_per_node":  "1",
+    "topology_addr":  "C1.S3.L277-DH3.pool0-01265",
+    "topology_addr_pattern":  "switch.switch.switch.node",
+    "tres_per_task":  "cpu=64"
+  },
+  "writerId":  "g02tfu9kp55u1pnromnypr2iw53b18ve"
+}

deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/wandb/wandb/run-20260728_110400-3oofo2bj/files/wandb-summary.json

@@ -0,0 +1 @@
+{"eval/action_loss/samples":368,"train/grad_norm_pre_clip":0.002506854711100459,"eval/latency_6/loss":0.26608511805534363,"eval/loss":0.26608511805534363,"eval/action_loss/seconds":1.9980583479627967,"_runtime":12167.008259036,"_wandb":{"runtime":12167},"_timestamp":1.78527396137343e+09,"timing/data":0.00028319901321083307,"eval/deadly_corridor/latency_6/loss":0.26608511805534363,"learning_rate/action_model":5.441263215519722e-05,"epoch":24.1,"_step":4000,"timing/model":0.42647266294807196,"learning_rate/qwen_vl_interface":5.4412632155197224e-06,"eval/deadly_corridor/loss":0.26608511805534363,"train/loss":3.123553688055836e-05}

deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/wandb/wandb/run-20260728_110400-3oofo2bj/logs/debug-core.log

@@ -0,0 +1,19 @@
+{"time":"2026-07-28T11:04:01.248754614-07:00","level":"INFO","msg":"main: starting server","port-filename":"/tmp/tmpw_7hsaw0/port-3039532.txt","pid":3039532,"detached":false,"idle-timeout":600000000000,"log-level":0,"disable-analytics":false,"shutdown-on-parent-exit":false,"enable-dcgm-profiling":false}
+{"time":"2026-07-28T11:04:01.249158429-07:00","level":"INFO","msg":"server: will exit if parent process dies","ppid":3039532}
+{"time":"2026-07-28T11:04:01.249153441-07:00","level":"INFO","msg":"server: accepting connections","addr":{"Name":"/tmp/wandb-3039532-3040545-1738620625/socket","Net":"unix"}}
+{"time":"2026-07-28T11:04:01.403774742-07:00","level":"INFO","msg":"connection: ManageConnectionData: new connection created","id":"1(@)"}
+{"time":"2026-07-28T11:04:01.414815054-07:00","level":"INFO","msg":"handleInformInit: received","streamId":"3oofo2bj","id":"1(@)"}
+{"time":"2026-07-28T11:04:01.785498989-07:00","level":"INFO","msg":"handleInformInit: stream started","streamId":"3oofo2bj","id":"1(@)"}
+{"time":"2026-07-28T11:04:08.060602913-07:00","level":"INFO","msg":"connection: cancelling request","id":"1(@)","requestId":"bsv5h43tw675"}
+{"time":"2026-07-28T14:26:49.656219362-07:00","level":"INFO","msg":"connection: cancelling request","id":"1(@)","requestId":"bsv5h43tw675"}
+{"time":"2026-07-28T14:27:01.069850227-07:00","level":"INFO","msg":"connection: cancelling request","id":"1(@)","requestId":"bsv5h43tw675"}
+{"time":"2026-07-28T14:27:01.075177628-07:00","level":"INFO","msg":"handleInformFinish: finish message received","streamId":"3oofo2bj","id":"1(@)"}
+{"time":"2026-07-28T14:27:01.077065103-07:00","level":"INFO","msg":"handleInformFinish: stream closed","streamId":"3oofo2bj","id":"1(@)"}
+{"time":"2026-07-28T14:27:05.637544827-07:00","level":"INFO","msg":"handleInformTeardown: server teardown initiated","id":"1(@)"}
+{"time":"2026-07-28T14:27:05.637997449-07:00","level":"INFO","msg":"server is shutting down"}
+{"time":"2026-07-28T14:27:05.637593632-07:00","level":"INFO","msg":"connection: closing","id":"1(@)"}
+{"time":"2026-07-28T14:27:05.63793795-07:00","level":"INFO","msg":"handleInformTeardown: server shutdown complete","id":"1(@)"}
+{"time":"2026-07-28T14:27:05.638151985-07:00","level":"INFO","msg":"connection: closed successfully","id":"1(@)"}
+{"time":"2026-07-28T14:27:05.638741228-07:00","level":"INFO","msg":"server: listener closed","addr":{"Name":"/tmp/wandb-3039532-3040545-1738620625/socket","Net":"unix"}}
+{"time":"2026-07-28T14:27:05.63843711-07:00","level":"INFO","msg":"connection: ManageConnectionData: connection closed","id":"1(@)"}
+{"time":"2026-07-28T14:27:05.638815586-07:00","level":"INFO","msg":"server is closed"}

deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/wandb/wandb/run-20260728_110400-3oofo2bj/logs/debug.log

@@ -0,0 +1,24 @@
+2026-07-28 11:04:00,723 INFO    MainThread:3039532 [wandb_setup.py:_flush():81] Current SDK version is 0.26.1
+2026-07-28 11:04:00,723 INFO    MainThread:3039532 [wandb_setup.py:_flush():81] Configure stats pid to 3039532
+2026-07-28 11:04:00,723 INFO    MainThread:3039532 [wandb_setup.py:_flush():81] Loading settings from environment variables
+2026-07-28 11:04:00,723 INFO    MainThread:3039532 [wandb_init.py:setup_run_log_directory():723] Logging user logs to /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/wandb/wandb/run-20260728_110400-3oofo2bj/logs/debug.log
+2026-07-28 11:04:00,723 INFO    MainThread:3039532 [wandb_init.py:setup_run_log_directory():724] Logging internal logs to /lustre/fsw/portfolios/nvr/projects/nvr_lacr_llm/users/zihwang/latency-sensitive-bench/third_party/starVLA/results/Checkpoints/deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/wandb/wandb/run-20260728_110400-3oofo2bj/logs/debug-internal.log
+2026-07-28 11:04:00,723 INFO    MainThread:3039532 [wandb_init.py:init():850] calling init triggers
+2026-07-28 11:04:00,724 INFO    MainThread:3039532 [wandb_init.py:init():855] wandb.init called with sweep_config: {}
+config: {'_wandb': {}}
+2026-07-28 11:04:00,724 INFO    MainThread:3039532 [wandb_init.py:init():898] starting backend
+2026-07-28 11:04:01,404 INFO    MainThread:3039532 [wandb_init.py:init():913] sending inform_init request
+2026-07-28 11:04:01,785 INFO    MainThread:3039532 [wandb_init.py:init():918] backend started and connected
+2026-07-28 11:04:01,789 INFO    MainThread:3039532 [wandb_init.py:init():988] updated telemetry
+2026-07-28 11:04:01,819 INFO    MainThread:3039532 [wandb_init.py:init():1011] communicating run to backend with 90.0 second timeout
+2026-07-28 11:04:02,666 INFO    MainThread:3039532 [wandb_init.py:init():1056] starting run threads in backend
+2026-07-28 11:04:03,055 INFO    MainThread:3039532 [wandb_run.py:_console_start():2554] atexit reg
+2026-07-28 11:04:03,055 INFO    MainThread:3039532 [wandb_run.py:_redirect():2403] redirect: wrap_raw
+2026-07-28 11:04:03,055 INFO    MainThread:3039532 [wandb_run.py:_redirect():2472] Wrapping output streams.
+2026-07-28 11:04:03,055 INFO    MainThread:3039532 [wandb_run.py:_redirect():2495] Redirects installed.
+2026-07-28 11:04:03,065 INFO    MainThread:3039532 [wandb_init.py:init():1094] run started, returning control to user process
+2026-07-28 14:26:49,652 INFO    MainThread:3039532 [wandb_run.py:_finish():2321] finishing run zihanwang-ai-northwestern-university/starVLA_rl_games/3oofo2bj
+2026-07-28 14:26:49,653 INFO    MainThread:3039532 [wandb_run.py:_atexit_cleanup():2520] got exitcode: 0
+2026-07-28 14:26:49,654 INFO    MainThread:3039532 [wandb_run.py:_restore():2502] restore
+2026-07-28 14:26:49,654 INFO    MainThread:3039532 [wandb_run.py:_restore():2508] restore done
+2026-07-28 14:27:01,069 INFO    MainThread:3039532 [wandb_run.py:_footer_sync_info():3889] logging synced files

deadly_corridor_fix_latency_6_1000ep_7k2steps_stitch/wandb/wandb/run-20260728_110400-3oofo2bj/run-3oofo2bj.wandb

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6844a8b12b9a44bb113d03022176c5057d48f4032dfa3e9174e33141d9f2e66c
+size 8879123