diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwengr00t-rtx3090-v1/README.md b/latency-aware/air-raid/small-policy/sample-factory-qwengr00t-rtx3090-v1/README.md new file mode 100644 index 0000000000000000000000000000000000000000..ca4145931126fcc70ae48a633c5720c4fa9408a9 --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwengr00t-rtx3090-v1/README.md @@ -0,0 +1,28 @@ +# air-raid / sample-factory-appo + +Training condition: `latency-aware`. Run: `air_raid_profile_latency_appo_h1`. + +Copied from [Standard-Pipeline](https://huggingface.co/latency-sensitive-bench/Standard-Pipeline/tree/2208875f92b21e3f8ec2363a30a6705dfe56cc04/air_raid/profile_latency/small_model/GR00T) at `2208875f92b21e3f8ec2363a30a6705dfe56cc04`. + +- Checkpoint: `checkpoint.pth` +- Selection: training_best +- Checkpoint SHA256: `eac0b29dc62d7f6931d9879a56df07605bf57a24a4fd7ebddd25613f25531364` +- Configuration, FPS, actions, normalization and latency semantics retain their source values. +- This import does not run evaluations or change training/evaluation/acceptance status. + +Sample Factory APPO inference bundle: `config.json` and the explicit `checkpoint.pth`. +The checkpoint contains `model`, `train_step`, and `env_steps`; every exported model +tensor was reloaded and compared bit-for-bit with the selected source checkpoint. + +[Original provenance](source/provenance.json) retains dataset references and recorded experiment evidence. + +## Download + +```python +from huggingface_hub import snapshot_download +prefix = "latency-aware/air-raid/small-policy/sample-factory-qwengr00t-rtx3090-v1" +snapshot_download("latency-sensitive-bench/benchmark-models", revision="", + allow_patterns=[prefix + "/**"]) +``` + +Use the verified import commit listed in the repository's import report. diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwengr00t-rtx3090-v1/checkpoint.pth b/latency-aware/air-raid/small-policy/sample-factory-qwengr00t-rtx3090-v1/checkpoint.pth new file mode 100644 index 0000000000000000000000000000000000000000..e281225b80401eb248d112ba4da637d5757cb342 --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwengr00t-rtx3090-v1/checkpoint.pth @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:eac0b29dc62d7f6931d9879a56df07605bf57a24a4fd7ebddd25613f25531364 +size 7210293 diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwengr00t-rtx3090-v1/config.json b/latency-aware/air-raid/small-policy/sample-factory-qwengr00t-rtx3090-v1/config.json new file mode 100644 index 0000000000000000000000000000000000000000..a06b0f907eba47a7bf9358b93f96055812fc48ed --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwengr00t-rtx3090-v1/config.json @@ -0,0 +1,247 @@ +{ + "help": false, + "algo": "APPO", + "env": "latency_gymnasium", + "experiment": "air_raid_profile_latency_appo_h1", + "train_dir": "/mnt/local/lzj/latency-sensitive-bench/runs/air_raid/profile_latency/appo_h1_10m_20260917/teacher/checkpoints", + "restart_behavior": "overwrite", + "device": "gpu", + "seed": 0, + "initial_model_path": null, + "num_policies": 1, + "async_rl": true, + "serial_mode": false, + "batched_sampling": false, + "num_batches_to_accumulate": 2, + "worker_num_splits": 2, + "policy_workers_per_policy": 1, + "max_policy_lag": 300, + "num_workers": 16, + "num_envs_per_worker": 8, + "batch_size": 1024, + "num_batches_per_epoch": 4, + "num_epochs": 4, + "rollout": 32, + "fasttd3_replay_capacity": 6553600, + "fasttd3_replay_batch_size": 32768, + "fasttd3_action_chunk_horizon": 1, + "fasttd3_transitions_per_update": 64, + "fasttd3_train_for_optimizer_steps": 10000000000, + "fasttd3_v_min": -250.0, + "fasttd3_v_max": 250.0, + "fasttd3_actor_action_l2": 0.0, + "fasttd3_compile": true, + "recurrence": 1, + "shuffle_minibatches": false, + "gamma": 0.99, + "reward_scale": 1.0, + "reward_clip": 1000.0, + "value_bootstrap": false, + "normalize_returns": true, + "exploration_loss_coeff": 0.01, + "value_loss_coeff": 0.5, + "kl_loss_coeff": 0.0, + "exploration_loss": "entropy", + "gae_lambda": 0.95, + "ppo_clip_ratio": 0.1, + "ppo_clip_value": 0.2, + "with_vtrace": false, + "vtrace_rho": 1.0, + "vtrace_c": 1.0, + "optimizer": "adam", + "adam_eps": 1e-05, + "adam_beta1": 0.9, + "adam_beta2": 0.999, + "max_grad_norm": 0.5, + "learning_rate": 0.00025, + "lr_schedule": "linear_decay", + "lr_schedule_kl_threshold": 0.008, + "lr_adaptive_min": 1e-06, + "lr_adaptive_max": 0.01, + "obs_subtract_mean": 0.0, + "obs_scale": 255.0, + "normalize_input": true, + "normalize_input_keys": null, + "decorrelate_experience_max_seconds": 0, + "decorrelate_envs_on_one_worker": true, + "actor_worker_gpus": [], + "set_workers_cpu_affinity": true, + "force_envs_single_thread": false, + "default_niceness": 0, + "log_to_file": true, + "experiment_summaries_interval": 10, + "flush_summaries_interval": 30, + "stats_avg": 100, + "summaries_use_frameskip": true, + "heartbeat_interval": 20, + "heartbeat_reporting_interval": 180, + "train_for_env_steps": 10000000, + "train_for_seconds": 10000000000, + "save_every_sec": 600, + "keep_checkpoints": 5, + "load_checkpoint_kind": "latest", + "save_milestones_sec": -1, + "save_best_every_sec": 5, + "save_best_metric": "reward", + "save_best_after": 100000, + "benchmark": false, + "encoder_mlp_layers": [ + 512, + 512 + ], + "encoder_conv_architecture": "convnet_atari", + "encoder_conv_mlp_layers": [ + 512 + ], + "use_rnn": false, + "rnn_size": 512, + "rnn_type": "gru", + "rnn_num_layers": 1, + "decoder_mlp_layers": [], + "nonlinearity": "relu", + "policy_initialization": "orthogonal", + "policy_init_gain": 1.0, + "actor_critic_share_weights": true, + "adaptive_stddev": false, + "continuous_tanh_scale": 0.0, + "initial_stddev": 1.0, + "use_env_info_cache": false, + "env_gpu_actions": false, + "env_gpu_observations": true, + "env_frameskip": 1, + "env_framestack": 4, + "pixel_format": "CHW", + "use_record_episode_statistics": false, + "with_wandb": false, + "wandb_user": null, + "wandb_project": "sample_factory", + "wandb_group": null, + "wandb_job_type": "SF", + "wandb_tags": [], + "with_pbt": false, + "pbt_mix_policies_in_one_env": true, + "pbt_period_env_steps": 5000000, + "pbt_start_mutation": 20000000, + "pbt_replace_fraction": 0.3, + "pbt_mutation_rate": 0.15, + "pbt_replace_reward_gap": 0.1, + "pbt_replace_reward_gap_absolute": 1e-06, + "pbt_optimize_gamma": false, + "pbt_target_objective": "true_objective", + "pbt_perturb_min": 1.1, + "pbt_perturb_max": 1.5, + "gym_task_name": "air_raid", + "gym_env_id": "LatencyBench/AirRaid-v0", + "gym_make_kwargs_json": "{\"base_env_id\": \"ALE/AirRaid-v5\", \"render_mode\": \"rgb_array\", \"screen_size\": 84, \"noop_max\": 0, \"base_make_kwargs\": {\"obs_type\": \"rgb\", \"frameskip\": 1, \"repeat_action_probability\": 0.0, \"full_action_space\": false, \"mode\": 1, \"difficulty\": 0, \"max_num_frames_per_episode\": 108000}}", + "gym_registration_imports_json": "[\"latency_bench.envs.gymnasium_air_raid\"]", + "gym_action_space_json": "{\"type\": \"discrete\", \"labels\": [\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"], \"values\": [0, 1, 2, 3, 4, 5]}", + "gym_action_labels_json": "[\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"]", + "gym_action_values_json": "[0, 1, 2, 3, 4, 5]", + "gym_noop_action_id": 0, + "gym_noop_action_json": "\"noop\"", + "gym_base_prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.", + "gym_state_labels_json": "", + "env_fps": 50.0, + "obs_fps": 12.5, + "frame_stack": 4, + "export_env_raw_rgb_frames": false, + "mode": "train", + "latency_type": "iid", + "fixed_latency_ms": null, + "mean_latency_ms": null, + "std_latency_ms": null, + "min_latency_ms": null, + "max_latency_ms": null, + "latency_seed": 0, + "latency_profile_path": "/mnt/local/lzj/latency-sensitive-bench/profiles/air_raid/gr00t/rtx3090_20260917/profile/profile.json", + "latency_profile_worker_slot": 0, + "simulated_worker_capacity": null, + "add_latency_info": false, + "max_pending_actions": null, + "hold_last_chunk_action": false, + "hold_policy": "hold", + "ordering_policy": "issue_order_fifo", + "eval_episodes": 5, + "eval_parallel_envs": 1, + "eval_latency_raw_frame_values": null, + "eval_max_steps": 5000, + "eval_deterministic": true, + "eval_raw_reward": false, + "episode_metrics_path": "/mnt/local/lzj/latency-sensitive-bench/runs/air_raid/profile_latency/appo_h1_10m_20260917/teacher/checkpoints/air_raid_profile_latency_appo_h1/episode_metrics.jsonl", + "ppo": null, + "command_line": "--mode train --algo APPO --env latency_gymnasium --experiment air_raid_profile_latency_appo_h1 --train_dir /mnt/local/lzj/latency-sensitive-bench/runs/air_raid/profile_latency/appo_h1_10m_20260917/teacher/checkpoints --restart_behavior overwrite --device gpu --seed 0 --episode_metrics_path /mnt/local/lzj/latency-sensitive-bench/runs/air_raid/profile_latency/appo_h1_10m_20260917/teacher/checkpoints/air_raid_profile_latency_appo_h1/episode_metrics.jsonl --train_for_env_steps 10000000 --num_workers 16 --num_envs_per_worker 8 --num_policies 1 --batch_size 1024 --rollout 32 --recurrence 1 --num_epochs 4 --num_batches_per_epoch 4 --worker_num_splits 2 --max_policy_lag 300 --learning_rate 0.00025 --nonlinearity relu --env_framestack 4 --gamma 0.99 --gae_lambda 0.95 --ppo_clip_ratio 0.1 --ppo_clip_value 0.2 --exploration_loss entropy --exploration_loss_coeff 0.01 --value_loss_coeff 0.5 --max_grad_norm 0.5 --adam_eps 1e-05 --obs_scale 255.0 --save_every_sec 600 --keep_checkpoints 5 --async_rl True --use_rnn False --normalize_returns True --normalize_input True --adaptive_stddev False --with_vtrace False --latency-type iid --latency-seed 0 --latency-profile-path /mnt/local/lzj/latency-sensitive-bench/profiles/air_raid/gr00t/rtx3090_20260917/profile/profile.json --latency-profile-worker-slot 0 --add-latency-info False --eval-episodes 5 --eval-parallel-envs 1 --eval-max-steps 5000 --eval-deterministic True --encoder_conv_architecture convnet_atari --gym-task-name air_raid --gym-env-id LatencyBench/AirRaid-v0 --gym-make-kwargs-json {\"base_env_id\": \"ALE/AirRaid-v5\", \"render_mode\": \"rgb_array\", \"screen_size\": 84, \"noop_max\": 0, \"base_make_kwargs\": {\"obs_type\": \"rgb\", \"frameskip\": 1, \"repeat_action_probability\": 0.0, \"full_action_space\": false, \"mode\": 1, \"difficulty\": 0, \"max_num_frames_per_episode\": 108000}} --gym-registration-imports-json [\"latency_bench.envs.gymnasium_air_raid\"] --gym-action-space-json {\"type\": \"discrete\", \"labels\": [\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"], \"values\": [0, 1, 2, 3, 4, 5]} --gym-noop-action-json \"noop\" --gym-action-labels-json [\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"] --gym-action-values-json [0, 1, 2, 3, 4, 5] --gym-noop-action-id 0 --gym-base-prompt Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire. --env-fps 50 --obs-fps 12.5 --frame-stack 4 --hold-policy hold --ordering-policy issue_order_fifo --use_env_info_cache False", + "cli_args": { + "algo": "APPO", + "env": "latency_gymnasium", + "experiment": "air_raid_profile_latency_appo_h1", + "train_dir": "/mnt/local/lzj/latency-sensitive-bench/runs/air_raid/profile_latency/appo_h1_10m_20260917/teacher/checkpoints", + "restart_behavior": "overwrite", + "device": "gpu", + "seed": 0, + "num_policies": 1, + "async_rl": true, + "worker_num_splits": 2, + "max_policy_lag": 300, + "num_workers": 16, + "num_envs_per_worker": 8, + "batch_size": 1024, + "num_batches_per_epoch": 4, + "num_epochs": 4, + "rollout": 32, + "recurrence": 1, + "gamma": 0.99, + "normalize_returns": true, + "exploration_loss_coeff": 0.01, + "value_loss_coeff": 0.5, + "exploration_loss": "entropy", + "gae_lambda": 0.95, + "ppo_clip_ratio": 0.1, + "ppo_clip_value": 0.2, + "with_vtrace": false, + "adam_eps": 1e-05, + "max_grad_norm": 0.5, + "learning_rate": 0.00025, + "obs_scale": 255.0, + "normalize_input": true, + "train_for_env_steps": 10000000, + "save_every_sec": 600, + "keep_checkpoints": 5, + "encoder_conv_architecture": "convnet_atari", + "use_rnn": false, + "nonlinearity": "relu", + "adaptive_stddev": false, + "use_env_info_cache": false, + "env_framestack": 4, + "gym_task_name": "air_raid", + "gym_env_id": "LatencyBench/AirRaid-v0", + "gym_make_kwargs_json": "{\"base_env_id\": \"ALE/AirRaid-v5\", \"render_mode\": \"rgb_array\", \"screen_size\": 84, \"noop_max\": 0, \"base_make_kwargs\": {\"obs_type\": \"rgb\", \"frameskip\": 1, \"repeat_action_probability\": 0.0, \"full_action_space\": false, \"mode\": 1, \"difficulty\": 0, \"max_num_frames_per_episode\": 108000}}", + "gym_registration_imports_json": "[\"latency_bench.envs.gymnasium_air_raid\"]", + "gym_action_space_json": "{\"type\": \"discrete\", \"labels\": [\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"], \"values\": [0, 1, 2, 3, 4, 5]}", + "gym_action_labels_json": "[\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"]", + "gym_action_values_json": "[0, 1, 2, 3, 4, 5]", + "gym_noop_action_id": 0, + "gym_noop_action_json": "\"noop\"", + "gym_base_prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.", + "env_fps": 50.0, + "obs_fps": 12.5, + "frame_stack": 4, + "mode": "train", + "latency_type": "iid", + "latency_seed": 0, + "latency_profile_path": "/mnt/local/lzj/latency-sensitive-bench/profiles/air_raid/gr00t/rtx3090_20260917/profile/profile.json", + "latency_profile_worker_slot": 0, + "add_latency_info": false, + "hold_policy": "hold", + "ordering_policy": "issue_order_fifo", + "eval_episodes": 5, + "eval_parallel_envs": 1, + "eval_max_steps": 5000, + "eval_deterministic": true, + "episode_metrics_path": "/mnt/local/lzj/latency-sensitive-bench/runs/air_raid/profile_latency/appo_h1_10m_20260917/teacher/checkpoints/air_raid_profile_latency_appo_h1/episode_metrics.jsonl", + "output_dir": "/mnt/local/lzj/latency-sensitive-bench/runs/air_raid/profile_latency/appo_h1_10m_20260917/teacher/training" + }, + "git_hash": "unknown", + "git_repo_name": "not a git repository", + "eval_env_frameskip": 1, + "output_dir": "/mnt/local/lzj/latency-sensitive-bench/runs/air_raid/profile_latency/appo_h1_10m_20260917/teacher/training" +} \ No newline at end of file diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwengr00t-rtx3090-v1/provenance.json b/latency-aware/air-raid/small-policy/sample-factory-qwengr00t-rtx3090-v1/provenance.json new file mode 100644 index 0000000000000000000000000000000000000000..2b80f48326487c8821010280d5026bdf3229c7bb --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwengr00t-rtx3090-v1/provenance.json @@ -0,0 +1,44 @@ +{ + "task": "air-raid", + "model": "sample-factory-appo", + "training_condition": "latency-aware", + "training_run_id": "air_raid_profile_latency_appo_h1", + "source": { + "repo_id": "latency-sensitive-bench/Standard-Pipeline", + "revision": "2208875f92b21e3f8ec2363a30a6705dfe56cc04", + "path": "air_raid/profile_latency/small_model/GR00T", + "url": "https://huggingface.co/latency-sensitive-bench/Standard-Pipeline/tree/2208875f92b21e3f8ec2363a30a6705dfe56cc04/air_raid/profile_latency/small_model/GR00T" + }, + "hf_repo_id": "latency-sensitive-bench/benchmark-models", + "hf_path": "latency-aware/air-raid/small-policy/sample-factory-qwengr00t-rtx3090-v1", + "checkpoint": { + "source_file": "air_raid/profile_latency/small_model/GR00T/checkpoint_p0/best_000038304_9805824_reward_7971.000.pth", + "source_sha256": "0704de53c017ea5cedb9fc71805edcf116e03a727194e091f1b62ef29c5c7c08", + "source_bytes": 20722745, + "file": "checkpoint.pth", + "selection_rule": "training_best", + "method": "inference_export", + "sha256": "eac0b29dc62d7f6931d9879a56df07605bf57a24a4fd7ebddd25613f25531364", + "bytes": 7210293, + "train_step": 38304, + "env_steps": 9805824, + "tensor_count": 18, + "tensor_equality_verified": true, + "removed_fields": [ + "best_performance", + "curr_lr", + "optimizer" + ] + }, + "config_source": "air_raid/profile_latency/small_model/GR00T/config.json", + "task_contract_file": null, + "missing_source_files": [], + "migration": { + "operation": "copy_selected_inference_bundle", + "source_modified": false, + "evaluation_performed": false, + "experiment_status": "preserved_in_source_evidence" + }, + "algorithm": "APPO", + "latency_model": "qwengr00t" +} diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwengr00t-rtx3090-v1/selection.json b/latency-aware/air-raid/small-policy/sample-factory-qwengr00t-rtx3090-v1/selection.json new file mode 100644 index 0000000000000000000000000000000000000000..d93447b33361097cf69c719785cd36b9301d21b3 --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwengr00t-rtx3090-v1/selection.json @@ -0,0 +1,12 @@ +{ + "selected": "training_best", + "selected_checkpoint": "best_000038304_9805824_reward_7971.000.pth", + "selected_checkpoint_sha256": "0704de53c017ea5cedb9fc71805edcf116e03a727194e091f1b62ef29c5c7c08", + "selected_train_step": 38304, + "selected_env_steps": 9805824, + "selected_training_reward": 7971.0, + "final_checkpoint": "checkpoint_000039068_10002432.pth", + "final_checkpoint_sha256": "69cb13560ce69adbbff5ec2eb1f018db483555b6132fb15e04bfc6687d09c5c3", + "final_train_step": 39068, + "final_env_steps": 10002432 +} diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwengr00t-rtx3090-v1/source-manifest-v10.json b/latency-aware/air-raid/small-policy/sample-factory-qwengr00t-rtx3090-v1/source-manifest-v10.json new file mode 100644 index 0000000000000000000000000000000000000000..dc270a769ea4964e6a64117b08dbc45c34b9e622 --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwengr00t-rtx3090-v1/source-manifest-v10.json @@ -0,0 +1,50 @@ +{ + "base_commit": "1681edb8d9e4f7d10b7dfa647670e01e0d5cce37", + "starvla_commit": "1d0d7b139d1725cab268cb9dae2007ef4c5d05d2", + "sample_factory_commit": "4b11d8754ffc989cdd15a340cfc9be0277377d5b", + "files": { + "latency_bench/data/starvla_tasks.py": "67336b5cd57d58138c8294b55b1eea3a708033caf488fb6b5fa32a87674d36bf", + "latency_bench/policy/starvla.py": "6c107e6c52e673310d98c511f56b5fce3a7e397a73e226713168377b7f605e91", + "latency_bench/policy/starvla_tasks.py": "a89d55999f9d17267eedc393721f1fedcdb99af752c06b95f291e5d82bc117a5", + "scripts/starvla/prepare_dataset.py": "98324cab1650920c9311c2021bc0f91571b1ec3f24417cf0b629e1e8fa65687c", + "scripts/gym_adapt/gr00t.py": "1b4c3f40ce46d4798af2bee8f393bf34d9f0c7582f5345d7998ee9774cc4a0ab", + "docs/gym-adapt/gr00t-pipeline-plan.md": "21dd13be061bb60cbe0647594e45fa94691dca322f4bd82773b0d89a0caa248a", + "configs/examples/gymnasium/gr00t/ant_teacher.yaml": "b20c6858801d5a3ea26e68a8cecf94d2718072f6d0fb94a9376333e069a04274", + "configs/examples/gymnasium/gr00t/half_cheetah_teacher.yaml": "cbfba28a5ee625f06c30fa1d290ed29a391f019a899fd5e4ae2503e09f6679a1", + "configs/examples/gymnasium/gr00t/hopper_teacher.yaml": "b57721674860194683f07896057d2206c1cfd5753c08b196b45d544466ed5cc2", + "configs/examples/gymnasium/gr00t/humanoid_teacher.yaml": "9701de3b39ab656bcc721d5aa91d10c05672f66da182cc741f1c34904e3cf06b", + "configs/examples/gymnasium/gr00t/inverted_pendulum_teacher.yaml": "e5c2dd087cc37614c4e4c02e0750114b288a33d4004050eb1a478366b267f5dc", + "configs/examples/gymnasium/gr00t/walker2d_teacher.yaml": "10d965dba406fccaa90099e34913e4ce7e2205233f621b20d1c16aa33215678b", + "latency_bench/data/action_layouts.py": "8ada6a799248b31e72b51019dec73366b79f60cd20e3ef37aa98d0f966a015c0", + "scripts/gym_adapt/train_gr00t.sh": "9332f84548e851724ee89345977b2b046fbc0e64af9c35ccdbf734557756d64d", + "latency_bench/data/exporters/starvla_lerobot.py": "490b743538ff411d97eb40f15a5bd54090dc7834eb82a05b996e62cc2a924528", + "tests/data/test_starvla_lerobot_exporter.py": "945720ed37637e9364fcb865a42254b003b893f77e9c99310943ac1b8b654f09", + "tests/data/test_starvla_tasks.py": "4b97bc4b4ec127a6c6bc831e69cee5fb7f6894f0b483c14836e0ad5700464228" + }, + "starvla_overlay": { + "third_party/starVLA/examples/rl_games/train_files/data_registry/data_config.py": "dd399efd9983294e11372753827a80da3a9ed1e18c23aba8ab70a33e6a941738", + "third_party/starVLA/starVLA/training/train_starvla.py": "d81f39ff232b9d2b8c71c386dd68ccf73ebc0bce2296723a9485d0903af62547", + "third_party/starVLA/starVLA/dataloader/gr00t_lerobot/datasets.py": "d2e91ae8fbe2964ad452cf1e6fb37b81ae409e067d4dda6e59c28bde72ae80ba" + }, + "note": "v9 plus portable bundle trainer.profile_timing required by inference; disabled in config.yaml while full training recipe is preserved. Active training sources remain frozen.", + "version": "v10", + "validation": { + "candidate_artifact_sha256": "6e428b56eac76f13781113d053336020be600c19a8160ace1b3206ffaed8b4ea", + "local_targeted": { + "passed": 36, + "skipped": 11 + }, + "server_targeted_semantic_tests": 6, + "actual_dataset_equivalence": "bitwise equal for fixed train and val indices, including all dtypes, images, masks and metadata", + "server_receipt_sha256": "0e660448a0e1509c2e883df0e36af5a0e53da4cc72ec70b4b58ae9ce851bc376", + "benchmark_scope": "16 sequential getitem calls after initial warmup, not whole training", + "getitem_seconds": { + "v8": 8.3581, + "v9": 1.176 + }, + "bundle_contract_regression": { + "command": "pytest -q tests/data/test_starvla_tasks.py -k native_h8_conversion_keeps_chunk_and_train_only_statistics", + "passed": 1 + } + } +} diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwengr00t-rtx3090-v1/source/provenance.json b/latency-aware/air-raid/small-policy/sample-factory-qwengr00t-rtx3090-v1/source/provenance.json new file mode 100644 index 0000000000000000000000000000000000000000..cb9a06869ce856fe31b17a27db6cb379f792f7b2 --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwengr00t-rtx3090-v1/source/provenance.json @@ -0,0 +1,27 @@ +{ + "task": "air_raid", + "condition": "profile_latency", + "teacher": "Sample Factory APPO", + "source_version": "v10", + "teacher_config_sha256": "2e76f61b4b07dd52a68bf62e852ab25b70b8245a58d9ded6afe701ef1792e375", + "source_recipe_sha256": "92d64b617f7888f865002b1c7ec9e2d7527de9df2027e5e3947b55a604067fa8", + "profile_sha256": "a6ded702eef5495736922babdd58e24901526811613e45c680a2b77eecfdc149", + "latency_distribution_sha256": "5aba3ec3e09af2c3a6b7353fdeccbf19d0f548e7a19167533cb8485e803f1f6f", + "seed": 0, + "budget_env_steps": 10000000, + "completed_env_steps": 10002432, + "exit_code": 0, + "initialization": "fresh", + "selection": { + "selected": "training_best", + "selected_checkpoint": "best_000038304_9805824_reward_7971.000.pth", + "selected_checkpoint_sha256": "0704de53c017ea5cedb9fc71805edcf116e03a727194e091f1b62ef29c5c7c08", + "selected_train_step": 38304, + "selected_env_steps": 9805824, + "selected_training_reward": 7971.0, + "final_checkpoint": "checkpoint_000039068_10002432.pth", + "final_checkpoint_sha256": "69cb13560ce69adbbff5ec2eb1f018db483555b6132fb15e04bfc6687d09c5c3", + "final_train_step": 39068, + "final_env_steps": 10002432 + } +} diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwengr00t-rtx3090-v1/teacher.yaml b/latency-aware/air-raid/small-policy/sample-factory-qwengr00t-rtx3090-v1/teacher.yaml new file mode 100644 index 0000000000000000000000000000000000000000..f02369fb6b0ddb05c1b027a8f0700b3cc57ffbe5 --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwengr00t-rtx3090-v1/teacher.yaml @@ -0,0 +1,121 @@ +experiment: + name: air_raid_profile_latency_appo_h1 + seed: 0 +backend: + type: sample_factory + algo: APPO + device: cuda + train_dir: /mnt/local/lzj/latency-sensitive-bench/runs/air_raid/profile_latency/appo_h1_10m_20260917/teacher/checkpoints + restart_behavior: overwrite + run_mode: train + extra_args: + - --encoder_conv_architecture + - convnet_atari +executor: + mode: simulated +env: + name: gymnasium + task_name: air_raid + env_id: LatencyBench/AirRaid-v0 + registration_imports: + - latency_bench.envs.gymnasium_air_raid + make_kwargs: + base_env_id: ALE/AirRaid-v5 + render_mode: rgb_array + screen_size: 84 + noop_max: 0 + base_make_kwargs: + obs_type: rgb + frameskip: 1 + repeat_action_probability: 0.0 + full_action_space: false + mode: 1 + difficulty: 0 + max_num_frames_per_episode: 108000 + env_fps: 50 + obs_fps: 12.5 + frame_stack: 4 + noop_action: noop + action_map: + noop: 0 + fire: 1 + right: 2 + left: 3 + rightfire: 4 + leftfire: 5 + action_order: + - noop + - fire + - right + - left + - rightfire + - leftfire + base_prompt: 'Protect both buildings from flying saucers. Choose exactly one action + from: noop, fire, right, left, rightfire, leftfire.' +latency: + method: iid + profile_path: /mnt/local/lzj/latency-sensitive-bench/profiles/air_raid/gr00t/rtx3090_20260917/profile/profile.json + profile_worker_slot: 0 + seed: 0 + add_latency_info: false +scheduler: + hold_policy: hold + ordering_policy: issue_order_fifo +policy: + type: random + actions: + - noop + - fire + - right + - left + - rightfire + - leftfire +training: + train_for_env_steps: 10000000 + num_workers: 16 + num_envs_per_worker: 8 + worker_num_splits: 2 + num_policies: 1 + batch_size: 1024 + rollout: 32 + recurrence: 1 + env_framestack: 4 + num_epochs: 4 + num_batches_per_epoch: 4 + max_policy_lag: 300 + learning_rate: 0.00025 + lr_schedule: linear_decay + gamma: 0.99 + gae_lambda: 0.95 + ppo_clip_ratio: 0.1 + ppo_clip_value: 0.2 + value_loss_coeff: 0.5 + max_grad_norm: 0.5 + exploration_loss_coeff: 0.01 + exploration_loss: entropy + encoder_conv_architecture: convnet_atari + nonlinearity: relu + obs_scale: 255.0 + adam_eps: 1.0e-05 + with_vtrace: false + adaptive_stddev: false + async_rl: true + use_rnn: false + normalize_input: true + normalize_returns: true + save_every_sec: 600 + keep_checkpoints: 5 +evaluation: + eval_interval_steps: null + eval_episodes: 5 + eval_parallel_envs: 1 + eval_max_steps: 5000 + eval_deterministic: true +logging: + output_dir: /mnt/local/lzj/latency-sensitive-bench/runs/air_raid/profile_latency/appo_h1_10m_20260917/teacher/training + video: + enabled: false + num_bins: 1 + save_step_records: false + save_action_records: false + save_latency_records: false diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/README.md b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/README.md new file mode 100644 index 0000000000000000000000000000000000000000..6fb48051761e89094348ec3cf9675cbb19ad28f4 --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/README.md @@ -0,0 +1,26 @@ +# air-raid / sample-factory-appo + +Training condition: `latency-aware`. Run: `air_raid_profile_20260911T054920Z`. + +Copied from [Standard-Pipeline](https://huggingface.co/latency-sensitive-bench/Standard-Pipeline/tree/2208875f92b21e3f8ec2363a30a6705dfe56cc04/air_raid/profile_latency/small_model) at `2208875f92b21e3f8ec2363a30a6705dfe56cc04`. + +- Checkpoint: `checkpoint.pth` +- Selection: best_reward_after_full_training_budget +- Checkpoint SHA256: `2b4f5e89e4fbc92cda1dc7a1e6edb49c45045f486722b60fefb9a7e4b4cc18cd` +- Configuration, FPS, actions, normalization and latency semantics retain their source values. +- This import does not run evaluations or change training/evaluation/acceptance status. + +Sample Factory APPO inference bundle: `config.json` and the explicit `checkpoint.pth`. +The checkpoint contains `model`, `train_step`, and `env_steps`; every exported model +tensor was reloaded and compared bit-for-bit with the selected source checkpoint. + +## Download + +```python +from huggingface_hub import snapshot_download +prefix = "latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1" +snapshot_download("latency-sensitive-bench/benchmark-models", revision="", + allow_patterns=[prefix + "/**"]) +``` + +Use the verified import commit listed in the repository's import report. diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/checkpoint.pth b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/checkpoint.pth new file mode 100644 index 0000000000000000000000000000000000000000..b646bb994400f91be6e8f60efd66afdd28c98fa3 --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/checkpoint.pth @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:2b4f5e89e4fbc92cda1dc7a1e6edb49c45045f486722b60fefb9a7e4b4cc18cd +size 7210293 diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/config.json b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/config.json new file mode 100644 index 0000000000000000000000000000000000000000..09eaf407fd83f1488c9a7db500ba0ea01613fa6b --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/config.json @@ -0,0 +1,279 @@ +{ + "help": false, + "algo": "APPO", + "env": "latency_gymnasium", + "experiment": "air_raid_profile_20260911T054920Z", + "train_dir": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_models", + "restart_behavior": "overwrite", + "device": "gpu", + "seed": 0, + "initial_model_path": null, + "num_policies": 1, + "async_rl": true, + "serial_mode": false, + "batched_sampling": false, + "num_batches_to_accumulate": 2, + "worker_num_splits": 2, + "policy_workers_per_policy": 1, + "max_policy_lag": 300, + "num_workers": 16, + "num_envs_per_worker": 8, + "batch_size": 1024, + "num_batches_per_epoch": 4, + "num_epochs": 4, + "rollout": 32, + "fasttd3_replay_capacity": 6553600, + "fasttd3_replay_batch_size": 32768, + "fasttd3_transitions_per_update": 64, + "fasttd3_v_min": -250.0, + "fasttd3_v_max": 250.0, + "fasttd3_actor_action_l2": 0.0, + "fasttd3_compile": true, + "fasttd3_sonic_decoder_path": null, + "recurrence": 1, + "shuffle_minibatches": false, + "gamma": 0.99, + "reward_scale": 1.0, + "reward_clip": 1000.0, + "value_bootstrap": false, + "normalize_returns": true, + "exploration_loss_coeff": 0.01, + "value_loss_coeff": 0.5, + "kl_loss_coeff": 0.0, + "exploration_loss": "entropy", + "gae_lambda": 0.95, + "ppo_clip_ratio": 0.1, + "ppo_clip_value": 0.2, + "with_vtrace": false, + "vtrace_rho": 1.0, + "vtrace_c": 1.0, + "optimizer": "adam", + "adam_eps": 1e-05, + "adam_beta1": 0.9, + "adam_beta2": 0.999, + "max_grad_norm": 0.5, + "learning_rate": 0.00025, + "lr_schedule": "linear_decay", + "lr_schedule_kl_threshold": 0.008, + "lr_adaptive_min": 1e-06, + "lr_adaptive_max": 0.01, + "obs_subtract_mean": 0.0, + "obs_scale": 255.0, + "normalize_input": true, + "normalize_input_keys": null, + "decorrelate_experience_max_seconds": 0, + "decorrelate_envs_on_one_worker": true, + "actor_worker_gpus": [], + "set_workers_cpu_affinity": true, + "force_envs_single_thread": false, + "default_niceness": 0, + "log_to_file": true, + "experiment_summaries_interval": 10, + "flush_summaries_interval": 30, + "stats_avg": 100, + "summaries_use_frameskip": true, + "heartbeat_interval": 20, + "heartbeat_reporting_interval": 180, + "train_for_env_steps": 10000000, + "train_for_seconds": 10000000000, + "save_every_sec": 600, + "keep_checkpoints": 5, + "load_checkpoint_kind": "latest", + "save_milestones_sec": -1, + "save_best_every_sec": 5, + "save_best_metric": "reward", + "save_best_after": 100000, + "benchmark": false, + "encoder_mlp_layers": [ + 512, + 512 + ], + "encoder_conv_architecture": "convnet_atari", + "encoder_conv_mlp_layers": [ + 512 + ], + "use_rnn": false, + "rnn_size": 512, + "rnn_type": "gru", + "rnn_num_layers": 1, + "decoder_mlp_layers": [], + "nonlinearity": "relu", + "policy_initialization": "orthogonal", + "policy_init_gain": 1.0, + "actor_critic_share_weights": true, + "adaptive_stddev": false, + "continuous_tanh_scale": 0.0, + "initial_stddev": 1.0, + "use_env_info_cache": false, + "env_gpu_actions": false, + "env_gpu_observations": true, + "env_frameskip": 1, + "env_framestack": 4, + "pixel_format": "CHW", + "use_record_episode_statistics": false, + "with_wandb": false, + "wandb_user": null, + "wandb_project": "sample_factory", + "wandb_group": null, + "wandb_job_type": "SF", + "wandb_tags": [], + "with_pbt": false, + "pbt_mix_policies_in_one_env": true, + "pbt_period_env_steps": 5000000, + "pbt_start_mutation": 20000000, + "pbt_replace_fraction": 0.3, + "pbt_mutation_rate": 0.15, + "pbt_replace_reward_gap": 0.1, + "pbt_replace_reward_gap_absolute": 1e-06, + "pbt_optimize_gamma": false, + "pbt_target_objective": "true_objective", + "pbt_perturb_min": 1.1, + "pbt_perturb_max": 1.5, + "gym_task_name": "air_raid", + "gym_env_id": "LatencyBench/AirRaid-v0", + "gym_make_kwargs_json": "{\"base_env_id\": \"ALE/AirRaid-v5\", \"render_mode\": \"rgb_array\", \"screen_size\": 84, \"noop_max\": 0, \"base_make_kwargs\": {\"obs_type\": \"rgb\", \"frameskip\": 1, \"repeat_action_probability\": 0.0, \"full_action_space\": false, \"mode\": 1, \"difficulty\": 0, \"max_num_frames_per_episode\": 108000}}", + "gym_registration_imports_json": "[\"latency_bench.envs.gymnasium_air_raid\"]", + "gym_action_space_json": "{\"type\": \"discrete\", \"labels\": [\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"], \"values\": [0, 1, 2, 3, 4, 5]}", + "gym_action_labels_json": "[\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"]", + "gym_action_values_json": "[0, 1, 2, 3, 4, 5]", + "gym_noop_action_id": 0, + "gym_noop_action_json": "\"noop\"", + "gym_base_prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.", + "gym_state_labels_json": "", + "env_fps": 50.0, + "obs_fps": 12.5, + "frame_stack": 4, + "export_env_raw_rgb_frames": false, + "mode": "train", + "latency_type": "iid", + "fixed_latency_ms": null, + "mean_latency_ms": null, + "std_latency_ms": null, + "min_latency_ms": null, + "max_latency_ms": null, + "latency_seed": 0, + "latency_profile_path": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/profile/profile.json", + "latency_profile_worker_slot": 0, + "simulated_worker_capacity": null, + "add_latency_info": false, + "max_pending_actions": null, + "hold_policy": "hold", + "ordering_policy": "issue_order_fifo", + "eval_episodes": 5, + "eval_parallel_envs": 1, + "eval_latency_raw_frame_values": null, + "eval_max_steps": 5000, + "eval_deterministic": true, + "eval_raw_reward": false, + "episode_metrics_path": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_models/air_raid_profile_20260911T054920Z/episode_metrics.jsonl", + "command_line": "--mode train --algo APPO --env latency_gymnasium --experiment air_raid_profile_20260911T054920Z --train_dir /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_models --restart_behavior overwrite --device gpu --seed 0 --episode_metrics_path /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_models/air_raid_profile_20260911T054920Z/episode_metrics.jsonl --train_for_env_steps 10000000 --num_workers 16 --num_envs_per_worker 8 --num_policies 1 --batch_size 1024 --rollout 32 --recurrence 1 --num_epochs 4 --num_batches_per_epoch 4 --worker_num_splits 2 --max_policy_lag 300 --num_batches_to_accumulate 2 --policy_workers_per_policy 1 --learning_rate 0.00025 --kl_loss_coeff 0.0 --lr_schedule_kl_threshold 0.008 --nonlinearity relu --policy_initialization orthogonal --continuous_tanh_scale 0.0 --initial_stddev 1.0 --env_framestack 4 --gamma 0.99 --gae_lambda 0.95 --ppo_clip_ratio 0.1 --ppo_clip_value 0.2 --exploration_loss entropy --exploration_loss_coeff 0.01 --reward_scale 1.0 --reward_clip 1000.0 --value_loss_coeff 0.5 --max_grad_norm 0.5 --optimizer adam --adam_eps 1e-05 --adam_beta1 0.9 --adam_beta2 0.999 --obs_subtract_mean 0.0 --obs_scale 255.0 --decorrelate_experience_max_seconds 0 --default_niceness 0 --rnn_type gru --rnn_size 512 --save_every_sec 600 --keep_checkpoints 5 --save_milestones_sec -1 --save_best_every_sec 5 --save_best_after 100000 --stats_avg 100 --experiment_summaries_interval 10 --async_rl True --batched_sampling False --serial_mode False --use_rnn False --normalize_returns True --normalize_input True --adaptive_stddev False --shuffle_minibatches False --value_bootstrap False --with_vtrace False --decorrelate_envs_on_one_worker True --set_workers_cpu_affinity True --force_envs_single_thread False --actor_critic_share_weights True --with_wandb False --encoder_mlp_layers 512 512 --latency-type iid --latency-seed 0 --latency-profile-path /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/profile/profile.json --latency-profile-worker-slot 0 --add-latency-info False --eval-episodes 5 --eval-parallel-envs 1 --eval-max-steps 5000 --eval-deterministic True --encoder_conv_architecture convnet_atari --gym-task-name air_raid --gym-env-id LatencyBench/AirRaid-v0 --gym-make-kwargs-json {\"base_env_id\": \"ALE/AirRaid-v5\", \"render_mode\": \"rgb_array\", \"screen_size\": 84, \"noop_max\": 0, \"base_make_kwargs\": {\"obs_type\": \"rgb\", \"frameskip\": 1, \"repeat_action_probability\": 0.0, \"full_action_space\": false, \"mode\": 1, \"difficulty\": 0, \"max_num_frames_per_episode\": 108000}} --gym-registration-imports-json [\"latency_bench.envs.gymnasium_air_raid\"] --gym-action-space-json {\"type\": \"discrete\", \"labels\": [\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"], \"values\": [0, 1, 2, 3, 4, 5]} --gym-noop-action-json \"noop\" --gym-action-labels-json [\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"] --gym-action-values-json [0, 1, 2, 3, 4, 5] --gym-noop-action-id 0 --gym-base-prompt Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire. --env-fps 50 --obs-fps 12.5 --frame-stack 4 --hold-policy hold --ordering-policy issue_order_fifo --use_env_info_cache False", + "cli_args": { + "algo": "APPO", + "env": "latency_gymnasium", + "experiment": "air_raid_profile_20260911T054920Z", + "train_dir": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_models", + "restart_behavior": "overwrite", + "device": "gpu", + "seed": 0, + "num_policies": 1, + "async_rl": true, + "serial_mode": false, + "batched_sampling": false, + "num_batches_to_accumulate": 2, + "worker_num_splits": 2, + "policy_workers_per_policy": 1, + "max_policy_lag": 300, + "num_workers": 16, + "num_envs_per_worker": 8, + "batch_size": 1024, + "num_batches_per_epoch": 4, + "num_epochs": 4, + "rollout": 32, + "recurrence": 1, + "shuffle_minibatches": false, + "gamma": 0.99, + "reward_scale": 1.0, + "reward_clip": 1000.0, + "value_bootstrap": false, + "normalize_returns": true, + "exploration_loss_coeff": 0.01, + "value_loss_coeff": 0.5, + "kl_loss_coeff": 0.0, + "exploration_loss": "entropy", + "gae_lambda": 0.95, + "ppo_clip_ratio": 0.1, + "ppo_clip_value": 0.2, + "with_vtrace": false, + "optimizer": "adam", + "adam_eps": 1e-05, + "adam_beta1": 0.9, + "adam_beta2": 0.999, + "max_grad_norm": 0.5, + "learning_rate": 0.00025, + "lr_schedule_kl_threshold": 0.008, + "obs_subtract_mean": 0.0, + "obs_scale": 255.0, + "normalize_input": true, + "decorrelate_experience_max_seconds": 0, + "decorrelate_envs_on_one_worker": true, + "set_workers_cpu_affinity": true, + "force_envs_single_thread": false, + "default_niceness": 0, + "experiment_summaries_interval": 10, + "stats_avg": 100, + "train_for_env_steps": 10000000, + "save_every_sec": 600, + "keep_checkpoints": 5, + "save_milestones_sec": -1, + "save_best_every_sec": 5, + "save_best_after": 100000, + "encoder_mlp_layers": [ + 512, + 512 + ], + "encoder_conv_architecture": "convnet_atari", + "use_rnn": false, + "rnn_size": 512, + "rnn_type": "gru", + "nonlinearity": "relu", + "policy_initialization": "orthogonal", + "actor_critic_share_weights": true, + "adaptive_stddev": false, + "continuous_tanh_scale": 0.0, + "initial_stddev": 1.0, + "use_env_info_cache": false, + "env_framestack": 4, + "with_wandb": false, + "gym_task_name": "air_raid", + "gym_env_id": "LatencyBench/AirRaid-v0", + "gym_make_kwargs_json": "{\"base_env_id\": \"ALE/AirRaid-v5\", \"render_mode\": \"rgb_array\", \"screen_size\": 84, \"noop_max\": 0, \"base_make_kwargs\": {\"obs_type\": \"rgb\", \"frameskip\": 1, \"repeat_action_probability\": 0.0, \"full_action_space\": false, \"mode\": 1, \"difficulty\": 0, \"max_num_frames_per_episode\": 108000}}", + "gym_registration_imports_json": "[\"latency_bench.envs.gymnasium_air_raid\"]", + "gym_action_space_json": "{\"type\": \"discrete\", \"labels\": [\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"], \"values\": [0, 1, 2, 3, 4, 5]}", + "gym_action_labels_json": "[\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"]", + "gym_action_values_json": "[0, 1, 2, 3, 4, 5]", + "gym_noop_action_id": 0, + "gym_noop_action_json": "\"noop\"", + "gym_base_prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.", + "env_fps": 50.0, + "obs_fps": 12.5, + "frame_stack": 4, + "mode": "train", + "latency_type": "iid", + "latency_seed": 0, + "latency_profile_path": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/profile/profile.json", + "latency_profile_worker_slot": 0, + "add_latency_info": false, + "hold_policy": "hold", + "ordering_policy": "issue_order_fifo", + "eval_episodes": 5, + "eval_parallel_envs": 1, + "eval_max_steps": 5000, + "eval_deterministic": true, + "episode_metrics_path": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_models/air_raid_profile_20260911T054920Z/episode_metrics.jsonl", + "output_dir": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_train" + }, + "git_hash": "unknown", + "git_repo_name": "not a git repository", + "eval_env_frameskip": 1, + "output_dir": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_train" +} \ No newline at end of file diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/profile/DONE b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/profile/DONE new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/profile/hardware.json b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/profile/hardware.json new file mode 100644 index 0000000000000000000000000000000000000000..e8204e81807044975c0e0a0953d16d6835192f31 --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/profile/hardware.json @@ -0,0 +1,40 @@ +{ + "driver_version": "580.173.02", + "gpu_class": "1x-rtx3090", + "gpus": [ + { + "name": "NVIDIA GeForce RTX 3090", + "slot": 0 + } + ], + "instance_id": "instance_859cf1e47bca6046", + "topology_links": [], + "torch": { + "backends": { + "cuda_cudnn_sdp_enabled": true, + "cuda_flash_sdp_enabled": true, + "cuda_math_sdp_enabled": true, + "cuda_matmul_allow_tf32": false, + "cuda_mem_efficient_sdp_enabled": true, + "cudnn_allow_tf32": true, + "cudnn_benchmark": false + }, + "cuda_available": true, + "cuda_device_count": 1, + "cuda_version": "12.8", + "current_device": 0, + "current_device_name": "NVIDIA GeForce RTX 3090", + "device_properties": [ + { + "index": 0, + "major": 8, + "minor": 6, + "multi_processor_count": 82, + "name": "NVIDIA GeForce RTX 3090", + "total_memory": 25295257600 + } + ], + "float32_matmul_precision": "highest", + "version": "2.11.0+cu128" + } +} diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/profile/latency_burst_model.json b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/profile/latency_burst_model.json new file mode 100644 index 0000000000000000000000000000000000000000..881407186a327a0431f8d19344c2b8dcd43e7700 --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/profile/latency_burst_model.json @@ -0,0 +1,1329 @@ +{ + "burst_dwell_distribution": { + "distribution_type": "inverse_cdf", + "latency_ms": [ + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0 + ], + "quantile_levels": [ + 0.0, + 0.0001, + 0.0005, + 0.001, + 0.002, + 0.003, + 0.004, + 0.005, + 0.006, + 0.007, + 0.008, + 0.009, + 0.01, + 0.02, + 0.03, + 0.04, + 0.05, + 0.06, + 0.07, + 0.08, + 0.09, + 0.1, + 0.11, + 0.12, + 0.13, + 0.14, + 0.15, + 0.16, + 0.17, + 0.18, + 0.19, + 0.2, + 0.21, + 0.22, + 0.23, + 0.24, + 0.25, + 0.26, + 0.27, + 0.28, + 0.29, + 0.3, + 0.31, + 0.32, + 0.33, + 0.34, + 0.35, + 0.36, + 0.37, + 0.38, + 0.39, + 0.4, + 0.41, + 0.42, + 0.43, + 0.44, + 0.45, + 0.46, + 0.47, + 0.48, + 0.49, + 0.5, + 0.51, + 0.52, + 0.53, + 0.54, + 0.55, + 0.56, + 0.57, + 0.58, + 0.59, + 0.6, + 0.61, + 0.62, + 0.63, + 0.64, + 0.65, + 0.66, + 0.67, + 0.68, + 0.69, + 0.7, + 0.71, + 0.72, + 0.73, + 0.74, + 0.75, + 0.76, + 0.77, + 0.78, + 0.79, + 0.8, + 0.81, + 0.82, + 0.83, + 0.84, + 0.85, + 0.86, + 0.87, + 0.88, + 0.89, + 0.9, + 0.91, + 0.92, + 0.93, + 0.94, + 0.95, + 0.96, + 0.97, + 0.98, + 0.99, + 0.991, + 0.992, + 0.993, + 0.994, + 0.995, + 0.996, + 0.997, + 0.998, + 0.999, + 0.9995, + 0.9999, + 1.0 + ] + }, + "burst_dwell_lengths": [ + 1, + 1, + 1, + 1 + ], + "burst_merge_gap_records": 30, + "burst_rank_processes": [ + { + "draw_count": 4, + "dwell_length_spearman_rho": 0.0, + "level_ms": { + "distribution_type": "inverse_cdf", + "latency_ms": [ + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.14111005783082, + 72.15581642150879, + 72.17052278518676, + 72.18522914886475, + 72.19993551254272, + 72.2146418762207, + 72.22934823989868, + 72.24405460357666, + 72.25876096725464, + 72.27346733093262, + 72.28817369461059, + 72.30288005828858, + 72.31758642196655, + 72.33229278564453, + 72.34699914932251, + 72.36170551300049, + 72.37641187667846, + 72.39111824035645, + 72.40582460403442, + 72.42053096771241, + 72.43523733139038, + 72.44994369506836, + 72.46465005874634, + 72.47935642242432, + 72.49406278610229, + 72.50271152496337, + 72.50530263900757, + 72.50789375305176, + 72.51048486709595, + 72.51307598114013, + 72.51566709518433, + 72.51825820922852, + 72.5208493232727, + 72.52344043731689, + 72.52603155136109, + 72.52862266540528, + 72.53121377944946, + 72.53380489349365, + 72.53639600753785, + 72.53898712158202, + 72.54157823562622, + 72.54416934967041, + 72.5467604637146, + 72.54935157775878, + 72.55194269180298, + 72.55453380584717, + 72.55712491989135, + 72.55971603393554, + 72.56230714797974, + 72.56489826202393, + 72.62776734352111, + 72.75091439247132, + 72.8740614414215, + 72.9972084903717, + 73.1203555393219, + 73.2435025882721, + 73.3666496372223, + 73.48979668617248, + 73.61294373512268, + 73.73609078407287, + 73.85923783302307, + 73.98238488197326, + 74.10553193092346, + 74.22867897987366, + 74.35182602882385, + 74.47497307777405, + 74.59812012672424, + 74.72126717567444, + 74.84441422462463, + 74.96756127357483, + 75.09070832252502, + 75.21385537147522, + 75.33700242042542, + 75.46014946937561, + 75.58329651832581, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009 + ], + "quantile_levels": [ + 0.0, + 0.0001, + 0.0005, + 0.001, + 0.002, + 0.003, + 0.004, + 0.005, + 0.006, + 0.007, + 0.008, + 0.009, + 0.01, + 0.02, + 0.03, + 0.04, + 0.05, + 0.06, + 0.07, + 0.08, + 0.09, + 0.1, + 0.11, + 0.12, + 0.13, + 0.14, + 0.15, + 0.16, + 0.17, + 0.18, + 0.19, + 0.2, + 0.21, + 0.22, + 0.23, + 0.24, + 0.25, + 0.26, + 0.27, + 0.28, + 0.29, + 0.3, + 0.31, + 0.32, + 0.33, + 0.34, + 0.35, + 0.36, + 0.37, + 0.38, + 0.39, + 0.4, + 0.41, + 0.42, + 0.43, + 0.44, + 0.45, + 0.46, + 0.47, + 0.48, + 0.49, + 0.5, + 0.51, + 0.52, + 0.53, + 0.54, + 0.55, + 0.56, + 0.57, + 0.58, + 0.59, + 0.6, + 0.61, + 0.62, + 0.63, + 0.64, + 0.65, + 0.66, + 0.67, + 0.68, + 0.69, + 0.7, + 0.71, + 0.72, + 0.73, + 0.74, + 0.75, + 0.76, + 0.77, + 0.78, + 0.79, + 0.8, + 0.81, + 0.82, + 0.83, + 0.84, + 0.85, + 0.86, + 0.87, + 0.88, + 0.89, + 0.9, + 0.91, + 0.92, + 0.93, + 0.94, + 0.95, + 0.96, + 0.97, + 0.98, + 0.99, + 0.991, + 0.992, + 0.993, + 0.994, + 0.995, + 0.996, + 0.997, + 0.998, + 0.999, + 0.9995, + 0.9999, + 1.0 + ] + }, + "level_residual_ms": { + "distribution_type": "inverse_cdf", + "latency_ms": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "quantile_levels": [ + 0.0, + 0.0001, + 0.0005, + 0.001, + 0.002, + 0.003, + 0.004, + 0.005, + 0.006, + 0.007, + 0.008, + 0.009, + 0.01, + 0.02, + 0.03, + 0.04, + 0.05, + 0.06, + 0.07, + 0.08, + 0.09, + 0.1, + 0.11, + 0.12, + 0.13, + 0.14, + 0.15, + 0.16, + 0.17, + 0.18, + 0.19, + 0.2, + 0.21, + 0.22, + 0.23, + 0.24, + 0.25, + 0.26, + 0.27, + 0.28, + 0.29, + 0.3, + 0.31, + 0.32, + 0.33, + 0.34, + 0.35, + 0.36, + 0.37, + 0.38, + 0.39, + 0.4, + 0.41, + 0.42, + 0.43, + 0.44, + 0.45, + 0.46, + 0.47, + 0.48, + 0.49, + 0.5, + 0.51, + 0.52, + 0.53, + 0.54, + 0.55, + 0.56, + 0.57, + 0.58, + 0.59, + 0.6, + 0.61, + 0.62, + 0.63, + 0.64, + 0.65, + 0.66, + 0.67, + 0.68, + 0.69, + 0.7, + 0.71, + 0.72, + 0.73, + 0.74, + 0.75, + 0.76, + 0.77, + 0.78, + 0.79, + 0.8, + 0.81, + 0.82, + 0.83, + 0.84, + 0.85, + 0.86, + 0.87, + 0.88, + 0.89, + 0.9, + 0.91, + 0.92, + 0.93, + 0.94, + 0.95, + 0.96, + 0.97, + 0.98, + 0.99, + 0.991, + 0.992, + 0.993, + 0.994, + 0.995, + 0.996, + 0.997, + 0.998, + 0.999, + 0.9995, + 0.9999, + 1.0 + ] + }, + "severity": { + "distribution_type": "inverse_cdf", + "latency_ms": [ + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0 + ], + "quantile_levels": [ + 0.0, + 0.0001, + 0.0005, + 0.001, + 0.002, + 0.003, + 0.004, + 0.005, + 0.006, + 0.007, + 0.008, + 0.009, + 0.01, + 0.02, + 0.03, + 0.04, + 0.05, + 0.06, + 0.07, + 0.08, + 0.09, + 0.1, + 0.11, + 0.12, + 0.13, + 0.14, + 0.15, + 0.16, + 0.17, + 0.18, + 0.19, + 0.2, + 0.21, + 0.22, + 0.23, + 0.24, + 0.25, + 0.26, + 0.27, + 0.28, + 0.29, + 0.3, + 0.31, + 0.32, + 0.33, + 0.34, + 0.35, + 0.36, + 0.37, + 0.38, + 0.39, + 0.4, + 0.41, + 0.42, + 0.43, + 0.44, + 0.45, + 0.46, + 0.47, + 0.48, + 0.49, + 0.5, + 0.51, + 0.52, + 0.53, + 0.54, + 0.55, + 0.56, + 0.57, + 0.58, + 0.59, + 0.6, + 0.61, + 0.62, + 0.63, + 0.64, + 0.65, + 0.66, + 0.67, + 0.68, + 0.69, + 0.7, + 0.71, + 0.72, + 0.73, + 0.74, + 0.75, + 0.76, + 0.77, + 0.78, + 0.79, + 0.8, + 0.81, + 0.82, + 0.83, + 0.84, + 0.85, + 0.86, + 0.87, + 0.88, + 0.89, + 0.9, + 0.91, + 0.92, + 0.93, + 0.94, + 0.95, + 0.96, + 0.97, + 0.98, + 0.99, + 0.991, + 0.992, + 0.993, + 0.994, + 0.995, + 0.996, + 0.997, + 0.998, + 0.999, + 0.9995, + 0.9999, + 1.0 + ] + }, + "spike_count": 4 + } + ], + "burst_slot_rank_templates": [ + { + "dwell_length": 1, + "slot_to_rank": { + "0": 0 + } + }, + { + "dwell_length": 1, + "slot_to_rank": { + "0": 0 + } + }, + { + "dwell_length": 1, + "slot_to_rank": { + "0": 0 + } + }, + { + "dwell_length": 1, + "slot_to_rank": { + "0": 0 + } + } + ], + "model_type": "hidden_regime", + "pre_worker_time_ms": { + "distribution_type": "inverse_cdf", + "latency_ms": [ + 0.20420193672180176, + 0.21483877639770507, + 0.2287143712043762, + 0.24032604026794432, + 0.24867507362365723, + 0.25438513946533203, + 0.2607346229553223, + 0.2639804744720459, + 0.26635993289947507, + 0.2687912712097168, + 0.2710361452102661, + 0.27275579833984376, + 0.273915548324585, + 0.2837890768051147, + 0.2895009899139404, + 0.2939703321456909, + 0.2985083103179932, + 0.30235490322113034, + 0.30625831604003906, + 0.3095644426345825, + 0.31271454334259036, + 0.3151197671890259, + 0.317447395324707, + 0.3196178579330444, + 0.3216020727157593, + 0.3235012102127075, + 0.3252310037612915, + 0.3269206190109253, + 0.3283870220184326, + 0.3298677968978882, + 0.3312730979919434, + 0.332612419128418, + 0.3338682413101196, + 0.335247278213501, + 0.33649143218994143, + 0.33758762836456296, + 0.338687539100647, + 0.3397971725463867, + 0.34085161685943605, + 0.3419586896896362, + 0.3430896520614624, + 0.34424476623535155, + 0.3454119539260864, + 0.346552209854126, + 0.3476839065551758, + 0.3487655544281006, + 0.3497673511505127, + 0.35085655212402345, + 0.3518979120254517, + 0.35300746440887454, + 0.35399904727935794, + 0.35509252548217773, + 0.3561410903930664, + 0.3571738576889038, + 0.358305869102478, + 0.3593671655654907, + 0.3603264331817627, + 0.3612997245788574, + 0.362232985496521, + 0.3632385206222534, + 0.3641740322113037, + 0.365062952041626, + 0.3658906316757202, + 0.3667243766784668, + 0.3675804567337036, + 0.36850805759429933, + 0.36934189796447753, + 0.3703420162200928, + 0.3712935400009155, + 0.3722340869903564, + 0.3732797241210937, + 0.37433724403381347, + 0.37558521270751954, + 0.3767164659500122, + 0.3778761959075928, + 0.3790695905685425, + 0.38028051853179934, + 0.3815205669403076, + 0.38275862216949463, + 0.3840209150314331, + 0.38523890495300295, + 0.3865478992462158, + 0.3878518056869507, + 0.3894425630569458, + 0.3910191059112549, + 0.3924675369262695, + 0.3939321041107178, + 0.3954088592529297, + 0.39689703941345217, + 0.3985757350921631, + 0.400292010307312, + 0.40227370262145995, + 0.4040238857269287, + 0.40595915317535397, + 0.40784664154052735, + 0.41001430988311766, + 0.41180758476257323, + 0.4139445734024048, + 0.415944185256958, + 0.41818367958068847, + 0.4204584693908691, + 0.4230041980743408, + 0.4256094837188721, + 0.4285480833053589, + 0.4317239999771118, + 0.4352530431747436, + 0.4396242618560791, + 0.4441619348526001, + 0.4498762226104737, + 0.45640567302703855, + 0.46785748481750483, + 0.47020196151733407, + 0.4729775695800772, + 0.475634225845337, + 0.4788666810989385, + 0.48331641197204583, + 0.491900497436524, + 0.49948444366455025, + 0.5097592239379892, + 0.5262329158782943, + 0.5426986169815089, + 0.6917105711937698, + 1.1622118949890137 + ], + "quantile_levels": [ + 0.0, + 0.0001, + 0.0005, + 0.001, + 0.002, + 0.003, + 0.004, + 0.005, + 0.006, + 0.007, + 0.008, + 0.009, + 0.01, + 0.02, + 0.03, + 0.04, + 0.05, + 0.06, + 0.07, + 0.08, + 0.09, + 0.1, + 0.11, + 0.12, + 0.13, + 0.14, + 0.15, + 0.16, + 0.17, + 0.18, + 0.19, + 0.2, + 0.21, + 0.22, + 0.23, + 0.24, + 0.25, + 0.26, + 0.27, + 0.28, + 0.29, + 0.3, + 0.31, + 0.32, + 0.33, + 0.34, + 0.35, + 0.36, + 0.37, + 0.38, + 0.39, + 0.4, + 0.41, + 0.42, + 0.43, + 0.44, + 0.45, + 0.46, + 0.47, + 0.48, + 0.49, + 0.5, + 0.51, + 0.52, + 0.53, + 0.54, + 0.55, + 0.56, + 0.57, + 0.58, + 0.59, + 0.6, + 0.61, + 0.62, + 0.63, + 0.64, + 0.65, + 0.66, + 0.67, + 0.68, + 0.69, + 0.7, + 0.71, + 0.72, + 0.73, + 0.74, + 0.75, + 0.76, + 0.77, + 0.78, + 0.79, + 0.8, + 0.81, + 0.82, + 0.83, + 0.84, + 0.85, + 0.86, + 0.87, + 0.88, + 0.89, + 0.9, + 0.91, + 0.92, + 0.93, + 0.94, + 0.95, + 0.96, + 0.97, + 0.98, + 0.99, + 0.991, + 0.992, + 0.993, + 0.994, + 0.995, + 0.996, + 0.997, + 0.998, + 0.999, + 0.9995, + 0.9999, + 1.0 + ] + }, + "regime_step_counts": { + "burst": 4, + "calm": 35904 + }, + "regime_transition_counts": { + "burst": { + "burst": 0, + "calm": 4 + }, + "calm": { + "burst": 4, + "calm": 35895 + } + }, + "reset_scope": "session", + "schema_version": 12, + "spike_median_multiplier": 1.25, + "spike_threshold_ms_by_worker_slot": { + "0": 72.03269615769386 + }, + "worker_count": 1 +} diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/profile/latency_distribution.json b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/profile/latency_distribution.json new file mode 100644 index 0000000000000000000000000000000000000000..2b45df3ed46456efa7943e385fa375d3adaa0846 --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/profile/latency_distribution.json @@ -0,0 +1,1027 @@ +{ + "schema_version": 3, + "worker_slots": { + "0": { + "all": { + "count": 35908, + "observation_to_action_latency_ms": { + "distribution_type": "inverse_cdf", + "latency_ms": [ + 54.89762210845947, + 54.98431869125366, + 55.134520971298215, + 55.34168244361877, + 55.99118143081665, + 56.076304399490354, + 56.13243092346191, + 56.18818429946899, + 56.23180709552765, + 56.27835348701477, + 56.33713837718964, + 56.39195935821533, + 56.45379021167755, + 57.08514684200287, + 57.230893654823305, + 57.316619853973386, + 57.36959619522095, + 57.40613402843476, + 57.43060580730438, + 57.45380989074707, + 57.477882833480834, + 57.49994168281555, + 57.52059514522553, + 57.5393541431427, + 57.55763153076172, + 57.57544971466064, + 57.59502432346344, + 57.6165852022171, + 57.639358162879944, + 57.66515370845795, + 57.68923002243042, + 57.72054600715637, + 57.75672510147095, + 57.802322187423705, + 57.88595350265503, + 57.97818984031677, + 58.037572503089905, + 58.08466000556946, + 58.12414084434509, + 58.16394332885742, + 58.203309774398804, + 58.24743611812592, + 58.28536067008972, + 58.316369466781616, + 58.339424300193784, + 58.358790407180784, + 58.37594494819641, + 58.39228649616241, + 58.409412431716916, + 58.42540137290955, + 58.44186356544495, + 58.45765173435211, + 58.474079790115354, + 58.4883131980896, + 58.50111918926239, + 58.51361119747162, + 58.52541661262512, + 58.53668155670166, + 58.548266739845275, + 58.55865183830261, + 58.5691175699234, + 58.578869700431824, + 58.58787773609161, + 58.59606602191925, + 58.60470150947571, + 58.61385565757752, + 58.62180905342102, + 58.62971657276154, + 58.63715500831604, + 58.64509712219238, + 58.6526384973526, + 58.66041216850281, + 58.66814685344696, + 58.6752281665802, + 58.68287703037262, + 58.69019066810608, + 58.69767053127289, + 58.70504686832428, + 58.712141094207766, + 58.718946146965024, + 58.72618232250213, + 58.73353822231293, + 58.74131082057953, + 58.74863606929779, + 58.75591662883758, + 58.76344584465027, + 58.77212584018707, + 58.779995255470276, + 58.78799551010132, + 58.79621982574463, + 58.80471308708191, + 58.813468432426454, + 58.82311668395996, + 58.833434352874754, + 58.843691573143005, + 58.85456703662872, + 58.86566646099091, + 58.875943303108215, + 58.88800809860229, + 58.90074449539185, + 58.91300926685333, + 58.9265673160553, + 58.93994530677796, + 58.955101170539855, + 58.972674880027775, + 58.99344714641571, + 59.02611937522888, + 59.10132591247559, + 59.611970896720884, + 59.7586422252655, + 59.97354772090912, + 60.00530353927612, + 60.039614966392506, + 60.09215181350708, + 60.16863778018952, + 60.36564081192016, + 60.57177989768982, + 60.719089326858516, + 61.37363024044037, + 64.18603834533658, + 67.25203033256551, + 72.61951753550008, + 76.16190600395203 + ], + "quantile_levels": [ + 0.0, + 0.0001, + 0.0005, + 0.001, + 0.002, + 0.003, + 0.004, + 0.005, + 0.006, + 0.007, + 0.008, + 0.009, + 0.01, + 0.02, + 0.03, + 0.04, + 0.05, + 0.06, + 0.07, + 0.08, + 0.09, + 0.1, + 0.11, + 0.12, + 0.13, + 0.14, + 0.15, + 0.16, + 0.17, + 0.18, + 0.19, + 0.2, + 0.21, + 0.22, + 0.23, + 0.24, + 0.25, + 0.26, + 0.27, + 0.28, + 0.29, + 0.3, + 0.31, + 0.32, + 0.33, + 0.34, + 0.35, + 0.36, + 0.37, + 0.38, + 0.39, + 0.4, + 0.41, + 0.42, + 0.43, + 0.44, + 0.45, + 0.46, + 0.47, + 0.48, + 0.49, + 0.5, + 0.51, + 0.52, + 0.53, + 0.54, + 0.55, + 0.56, + 0.57, + 0.58, + 0.59, + 0.6, + 0.61, + 0.62, + 0.63, + 0.64, + 0.65, + 0.66, + 0.67, + 0.68, + 0.69, + 0.7, + 0.71, + 0.72, + 0.73, + 0.74, + 0.75, + 0.76, + 0.77, + 0.78, + 0.79, + 0.8, + 0.81, + 0.82, + 0.83, + 0.84, + 0.85, + 0.86, + 0.87, + 0.88, + 0.89, + 0.9, + 0.91, + 0.92, + 0.93, + 0.94, + 0.95, + 0.96, + 0.97, + 0.98, + 0.99, + 0.991, + 0.992, + 0.993, + 0.994, + 0.995, + 0.996, + 0.997, + 0.998, + 0.999, + 0.9995, + 0.9999, + 1.0 + ] + }, + "spearman_rho": 0.9930560092450511, + "worker_service_time_ms": { + "distribution_type": "inverse_cdf", + "latency_ms": [ + 54.638512134552, + 54.722969214344026, + 54.86490662717819, + 55.03681116294861, + 55.72376442432404, + 55.79827515983582, + 55.85051815986633, + 55.895044279098514, + 55.94917815303803, + 55.99627424621582, + 56.03666737651825, + 56.09092275047302, + 56.14324330806732, + 56.75663536548615, + 56.894120874404905, + 56.97365468502045, + 57.02204294204712, + 57.05275318145752, + 57.08081398963928, + 57.10400746822357, + 57.12682245254516, + 57.14491307735443, + 57.162690525054934, + 57.183143854141235, + 57.201689219474794, + 57.218997974395755, + 57.23760087490082, + 57.26001518726349, + 57.28137402057648, + 57.30520550251007, + 57.33147391319275, + 57.363392567634584, + 57.396899285316465, + 57.4381804561615, + 57.51748081684113, + 57.60506856441498, + 57.66861927509308, + 57.73105767250061, + 57.77032548904419, + 57.81555837154389, + 57.861557040214535, + 57.90422902107239, + 57.94723031044006, + 57.97415527820587, + 57.99504571914673, + 58.012214488983155, + 58.02739794254303, + 58.04379164218903, + 58.058891053199766, + 58.07336838245392, + 58.08783471107483, + 58.10370817184448, + 58.1183429479599, + 58.13163368225098, + 58.144854125976565, + 58.156135168075565, + 58.1677907705307, + 58.17717433929443, + 58.18661304473877, + 58.19571486473083, + 58.20485938072205, + 58.21307682991028, + 58.220322942733766, + 58.228040552139284, + 58.23579728603363, + 58.24308175563812, + 58.249634408950804, + 58.25668476104736, + 58.263752183914185, + 58.27049364566803, + 58.27714119911194, + 58.28466124534607, + 58.29146835803986, + 58.29871460437775, + 58.30632412910462, + 58.31351550579071, + 58.320294523239134, + 58.327939581871036, + 58.33411476135254, + 58.34080623149872, + 58.34782395839691, + 58.3550388097763, + 58.36206126213074, + 58.36932575702667, + 58.37651247024536, + 58.38423424243927, + 58.391886472702026, + 58.400058970451354, + 58.407453675270084, + 58.41486597061157, + 58.423468680381774, + 58.43212459087372, + 58.441460065841675, + 58.45135697364807, + 58.46152633190155, + 58.472478790283205, + 58.4831839799881, + 58.494375071525575, + 58.50569304466248, + 58.51628562450409, + 58.52735684394836, + 58.53880505561828, + 58.55049607753754, + 58.56301250934601, + 58.57766434192658, + 58.59507055282593, + 58.623821711540224, + 58.70154627799988, + 59.20479492664337, + 59.354912848472594, + 59.5520503616333, + 59.586989126205445, + 59.629149945259094, + 59.67486940002441, + 59.748895297050474, + 60.0047911453247, + 60.194087533950814, + 60.32199776840209, + 60.85784722709658, + 63.77599292945856, + 66.87870573186876, + 71.94066750355165, + 75.6448700428009 + ], + "quantile_levels": [ + 0.0, + 0.0001, + 0.0005, + 0.001, + 0.002, + 0.003, + 0.004, + 0.005, + 0.006, + 0.007, + 0.008, + 0.009, + 0.01, + 0.02, + 0.03, + 0.04, + 0.05, + 0.06, + 0.07, + 0.08, + 0.09, + 0.1, + 0.11, + 0.12, + 0.13, + 0.14, + 0.15, + 0.16, + 0.17, + 0.18, + 0.19, + 0.2, + 0.21, + 0.22, + 0.23, + 0.24, + 0.25, + 0.26, + 0.27, + 0.28, + 0.29, + 0.3, + 0.31, + 0.32, + 0.33, + 0.34, + 0.35, + 0.36, + 0.37, + 0.38, + 0.39, + 0.4, + 0.41, + 0.42, + 0.43, + 0.44, + 0.45, + 0.46, + 0.47, + 0.48, + 0.49, + 0.5, + 0.51, + 0.52, + 0.53, + 0.54, + 0.55, + 0.56, + 0.57, + 0.58, + 0.59, + 0.6, + 0.61, + 0.62, + 0.63, + 0.64, + 0.65, + 0.66, + 0.67, + 0.68, + 0.69, + 0.7, + 0.71, + 0.72, + 0.73, + 0.74, + 0.75, + 0.76, + 0.77, + 0.78, + 0.79, + 0.8, + 0.81, + 0.82, + 0.83, + 0.84, + 0.85, + 0.86, + 0.87, + 0.88, + 0.89, + 0.9, + 0.91, + 0.92, + 0.93, + 0.94, + 0.95, + 0.96, + 0.97, + 0.98, + 0.99, + 0.991, + 0.992, + 0.993, + 0.994, + 0.995, + 0.996, + 0.997, + 0.998, + 0.999, + 0.9995, + 0.9999, + 1.0 + ] + } + }, + "steady": { + "count": 35904, + "observation_to_action_latency_ms": { + "distribution_type": "inverse_cdf", + "latency_ms": [ + 54.89762210845947, + 54.984307411193846, + 55.134505897521976, + 55.34166074275971, + 55.991178302764894, + 56.076294094085696, + 56.13233103179932, + 56.18812419891358, + 56.231794157981874, + 56.27833763885498, + 56.33710426616669, + 56.391950715065, + 56.45378501415253, + 57.08512797832489, + 57.23083909511566, + 57.31661008834839, + 57.3695746421814, + 57.40609849452972, + 57.43057663917541, + 57.453807735443114, + 57.477878031730654, + 57.499938297271726, + 57.520570921897885, + 57.53932119846344, + 57.55763103485108, + 57.57542407989502, + 57.59496324062347, + 57.616562161445614, + 57.6393541097641, + 57.66512469768524, + 57.68913018226623, + 57.720493364334104, + 57.75670382976532, + 57.802289566993714, + 57.88569365978241, + 57.978020153045655, + 58.037256360054016, + 58.0845925617218, + 58.1240485048294, + 58.16375423908234, + 58.20327997684479, + 58.24732210636139, + 58.285302591323855, + 58.316290488243105, + 58.33935274600983, + 58.35876348018646, + 58.37584476470947, + 58.392186732292174, + 58.40936285018921, + 58.42536652565003, + 58.44180078983307, + 58.45757086277008, + 58.474003086090086, + 58.488275771141055, + 58.501091632843014, + 58.51360427379608, + 58.52537198066712, + 58.5366578245163, + 58.5481983423233, + 58.558502655029294, + 58.5690626001358, + 58.57883095741272, + 58.58777729511261, + 58.59602118492126, + 58.60467755317688, + 58.613761353492734, + 58.62173254489899, + 58.62966779708862, + 58.637111430168154, + 58.645032176971434, + 58.652547693252565, + 58.660341358184816, + 58.6681103515625, + 58.675179276466366, + 58.68275595188141, + 58.69008821964264, + 58.69758403301239, + 58.70498054981232, + 58.71211910247803, + 58.71892172336578, + 58.726107902526856, + 58.73349041938782, + 58.74119870662689, + 58.748615841865536, + 58.75589869499206, + 58.763404245376584, + 58.77199697494507, + 58.77996703624726, + 58.78795792102814, + 58.79615904808045, + 58.80460977554321, + 58.813416481018066, + 58.823071904182434, + 58.83335713863373, + 58.84358658790588, + 58.85443187713623, + 58.8656010389328, + 58.87590140342712, + 58.88786563873291, + 58.900563769340515, + 58.91284184932709, + 58.92648019790649, + 58.93981793880462, + 58.954998784065246, + 58.97248013019562, + 58.992928538322445, + 59.025522994995114, + 59.099266309738155, + 59.605603127479554, + 59.75604496479034, + 59.968015394210816, + 60.00281594467163, + 60.03302575588225, + 60.08006652545929, + 60.15268362808228, + 60.34006970405586, + 60.55531148242949, + 60.70003566932679, + 61.00997841644302, + 63.63513048267335, + 66.65921190643537, + 69.54871525268524, + 70.52106499671936 + ], + "quantile_levels": [ + 0.0, + 0.0001, + 0.0005, + 0.001, + 0.002, + 0.003, + 0.004, + 0.005, + 0.006, + 0.007, + 0.008, + 0.009, + 0.01, + 0.02, + 0.03, + 0.04, + 0.05, + 0.06, + 0.07, + 0.08, + 0.09, + 0.1, + 0.11, + 0.12, + 0.13, + 0.14, + 0.15, + 0.16, + 0.17, + 0.18, + 0.19, + 0.2, + 0.21, + 0.22, + 0.23, + 0.24, + 0.25, + 0.26, + 0.27, + 0.28, + 0.29, + 0.3, + 0.31, + 0.32, + 0.33, + 0.34, + 0.35, + 0.36, + 0.37, + 0.38, + 0.39, + 0.4, + 0.41, + 0.42, + 0.43, + 0.44, + 0.45, + 0.46, + 0.47, + 0.48, + 0.49, + 0.5, + 0.51, + 0.52, + 0.53, + 0.54, + 0.55, + 0.56, + 0.57, + 0.58, + 0.59, + 0.6, + 0.61, + 0.62, + 0.63, + 0.64, + 0.65, + 0.66, + 0.67, + 0.68, + 0.69, + 0.7, + 0.71, + 0.72, + 0.73, + 0.74, + 0.75, + 0.76, + 0.77, + 0.78, + 0.79, + 0.8, + 0.81, + 0.82, + 0.83, + 0.84, + 0.85, + 0.86, + 0.87, + 0.88, + 0.89, + 0.9, + 0.91, + 0.92, + 0.93, + 0.94, + 0.95, + 0.96, + 0.97, + 0.98, + 0.99, + 0.991, + 0.992, + 0.993, + 0.994, + 0.995, + 0.996, + 0.997, + 0.998, + 0.999, + 0.9995, + 0.9999, + 1.0 + ] + }, + "spearman_rho": 0.9930536881339532, + "worker_service_time_ms": { + "distribution_type": "inverse_cdf", + "latency_ms": [ + 54.638512134552, + 54.722956764411926, + 54.86490254116058, + 55.03670747375488, + 55.72376233577728, + 55.79825168800354, + 55.85047158241272, + 55.89504156112671, + 55.94910483646393, + 55.996236187934876, + 56.036575892448425, + 56.090843374252316, + 56.14299791812897, + 56.756627945899965, + 56.89406868934631, + 56.97361436367035, + 57.022025203704835, + 57.05273281097412, + 57.08080776691437, + 57.104003982543944, + 57.126808791160585, + 57.14489750862121, + 57.16263738155365, + 57.18307790756226, + 57.20166008472442, + 57.218993434906004, + 57.23759672641754, + 57.2600101518631, + 57.281266531944276, + 57.30516447544098, + 57.33145670413971, + 57.36331901550293, + 57.39689316272736, + 57.43804866313934, + 57.51728015422821, + 57.60498538017273, + 57.66832494735718, + 57.73094385147095, + 57.770116863250735, + 57.8153165435791, + 57.86133855342865, + 57.9039803981781, + 57.947170243263244, + 57.974107985496524, + 57.99501208305359, + 58.01215915203095, + 58.027367973327635, + 58.04376080989837, + 58.05885187625885, + 58.07331806659698, + 58.0877879524231, + 58.103690600395204, + 58.11831189632416, + 58.13161474704742, + 58.144645142555234, + 58.15612073421478, + 58.16774473190308, + 58.1771510219574, + 58.186569285392764, + 58.19569593429566, + 58.20483826637268, + 58.21305239200592, + 58.220281167030336, + 58.227979702949526, + 58.23577118873596, + 58.24305790424347, + 58.24960861206055, + 58.25667642593384, + 58.26372510910034, + 58.270454411506655, + 58.27711375236511, + 58.28462617397308, + 58.291462364196775, + 58.29867696762085, + 58.30626531124115, + 58.31345600605011, + 58.320233178138736, + 58.32780292034149, + 58.33409640789032, + 58.340758085250854, + 58.34780982017517, + 58.3549880027771, + 58.361964631080625, + 58.36928438186646, + 58.37640544891357, + 58.384184594154355, + 58.39184761047363, + 58.40000975608826, + 58.40733815193176, + 58.414857358932494, + 58.42340198993683, + 58.432047080993655, + 58.44124049186706, + 58.45121502876282, + 58.46129935741425, + 58.472327423095706, + 58.48313086032867, + 58.49423496723175, + 58.50561022758484, + 58.516244192123416, + 58.52729542255402, + 58.53868432044983, + 58.55042207241058, + 58.56292420387268, + 58.57733368396759, + 58.59481431007385, + 58.62353591918945, + 58.69977615833282, + 59.20313117980957, + 59.35360734462738, + 59.54943576812744, + 59.58170782470703, + 59.627803606033325, + 59.66968642902374, + 59.73967554950714, + 59.93983187198645, + 60.18512020492554, + 60.29954465675354, + 60.62631217098242, + 63.37979303073861, + 66.17347280121018, + 69.13033782882663, + 70.00722193717957 + ], + "quantile_levels": [ + 0.0, + 0.0001, + 0.0005, + 0.001, + 0.002, + 0.003, + 0.004, + 0.005, + 0.006, + 0.007, + 0.008, + 0.009, + 0.01, + 0.02, + 0.03, + 0.04, + 0.05, + 0.06, + 0.07, + 0.08, + 0.09, + 0.1, + 0.11, + 0.12, + 0.13, + 0.14, + 0.15, + 0.16, + 0.17, + 0.18, + 0.19, + 0.2, + 0.21, + 0.22, + 0.23, + 0.24, + 0.25, + 0.26, + 0.27, + 0.28, + 0.29, + 0.3, + 0.31, + 0.32, + 0.33, + 0.34, + 0.35, + 0.36, + 0.37, + 0.38, + 0.39, + 0.4, + 0.41, + 0.42, + 0.43, + 0.44, + 0.45, + 0.46, + 0.47, + 0.48, + 0.49, + 0.5, + 0.51, + 0.52, + 0.53, + 0.54, + 0.55, + 0.56, + 0.57, + 0.58, + 0.59, + 0.6, + 0.61, + 0.62, + 0.63, + 0.64, + 0.65, + 0.66, + 0.67, + 0.68, + 0.69, + 0.7, + 0.71, + 0.72, + 0.73, + 0.74, + 0.75, + 0.76, + 0.77, + 0.78, + 0.79, + 0.8, + 0.81, + 0.82, + 0.83, + 0.84, + 0.85, + 0.86, + 0.87, + 0.88, + 0.89, + 0.9, + 0.91, + 0.92, + 0.93, + 0.94, + 0.95, + 0.96, + 0.97, + 0.98, + 0.99, + 0.991, + 0.992, + 0.993, + 0.994, + 0.995, + 0.996, + 0.997, + 0.998, + 0.999, + 0.9995, + 0.9999, + 1.0 + ] + } + } + } + } +} diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/profile/latency_profile.png b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/profile/latency_profile.png new file mode 100644 index 0000000000000000000000000000000000000000..d092895d0fee0291370b91ff946477640d2b58a4 Binary files /dev/null and b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/profile/latency_profile.png differ diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/profile/profile.json b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/profile/profile.json new file mode 100644 index 0000000000000000000000000000000000000000..67bdbdee352adea1325fe3253410528697a354cf --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/profile/profile.json @@ -0,0 +1,66 @@ +{ + "burst_model_path": "latency_burst_model.json", + "distribution_path": "latency_distribution.json", + "env_fps": 50.0, + "frame_ms": 20.0, + "gpu_class": "1x-rtx3090", + "instance_id": "instance_859cf1e47bca6046", + "latency_kind": "observation_to_action_latency", + "latency_method": "temporal", + "model_id": "qwenoft", + "n_admitted_observations": 35908, + "n_capacity_drops": 0, + "n_observation_attempts": 35908, + "per_slot_summary": { + "0": { + "admitted_count": 35908, + "mean_observation_to_action_latency_ms": 58.405461079588015, + "mean_worker_service_time_ms": 58.03848960715147, + "p95_observation_to_action_latency_ms": 59.02611216306687, + "p95_worker_service_time_ms": 58.62378237247468, + "p99_worker_service_time_ms": 59.55199083089829 + } + }, + "provenance": { + "base_config": "/workspace/tasks/20260911T023128Z-p-only4/air_raid/profile.yaml", + "checkpoint_kind": "best", + "model_artifact": { + "checkpoint": "checkpoints/steps_5000_pytorch_model.pt", + "model_config": "config.full.yaml", + "path_in_repo": "OpenVLA/zero-latency/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k", + "repo_id": "latency-sensitive-bench/extra-envs-checkpoints", + "source": "local" + }, + "session_ids": [ + 0, + 1, + 2, + 3, + 4 + ] + }, + "sample_model_type": "hidden_regime", + "source_run_id": "20260911T023704401836Z", + "summary": { + "frame_ms": 20.0, + "max_ms": 76.16190600395203, + "mean_effective_frames": 2.920273053979401, + "mean_ms": 58.405461079588015, + "min_ms": 54.89762210845947, + "n_samples": 35908, + "p50_frames": 2.928943485021591, + "p50_ms": 58.578869700431824, + "p90_frames": 2.946328270435333, + "p90_ms": 58.926565408706665, + "p95_frames": 2.9513056081533433, + "p95_ms": 59.02611216306686, + "p99_frames": 2.998659399986267, + "p99_ms": 59.97318799972534, + "prob_latency_gt_1_frame": 1.0, + "prob_latency_gt_2_frames": 1.0, + "prob_latency_gt_3_frames": 0.009134454717611675, + "std_ms": 0.7141128839373884 + }, + "visualization_path": "latency_profile.png", + "workload_id": "air_raid" +} diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/provenance.json b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/provenance.json new file mode 100644 index 0000000000000000000000000000000000000000..674b34b6c3ea34ce0aace7a9517a893dc41b3e2f --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/provenance.json @@ -0,0 +1,44 @@ +{ + "task": "air-raid", + "model": "sample-factory-appo", + "training_condition": "latency-aware", + "training_run_id": "air_raid_profile_20260911T054920Z", + "source": { + "repo_id": "latency-sensitive-bench/Standard-Pipeline", + "revision": "2208875f92b21e3f8ec2363a30a6705dfe56cc04", + "path": "air_raid/profile_latency/small_model", + "url": "https://huggingface.co/latency-sensitive-bench/Standard-Pipeline/tree/2208875f92b21e3f8ec2363a30a6705dfe56cc04/air_raid/profile_latency/small_model" + }, + "hf_repo_id": "latency-sensitive-bench/benchmark-models", + "hf_path": "latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1", + "checkpoint": { + "source_file": "air_raid/profile_latency/small_model/checkpoint_p0/best_000034592_8855552_reward_9329.250.pth", + "source_sha256": "ac89617cb07e8291a771e162ce8c584ff29612ced8978d5d96878035bd352d97", + "source_bytes": 20722745, + "file": "checkpoint.pth", + "selection_rule": "best_reward_after_full_training_budget", + "method": "inference_export", + "sha256": "2b4f5e89e4fbc92cda1dc7a1e6edb49c45045f486722b60fefb9a7e4b4cc18cd", + "bytes": 7210293, + "train_step": 34592, + "env_steps": 8855552, + "tensor_count": 18, + "tensor_equality_verified": true, + "removed_fields": [ + "best_performance", + "curr_lr", + "optimizer" + ] + }, + "config_source": "air_raid/profile_latency/small_model/config.json", + "task_contract_file": null, + "missing_source_files": [], + "migration": { + "operation": "copy_selected_inference_bundle", + "source_modified": false, + "evaluation_performed": false, + "experiment_status": "preserved_in_source_evidence" + }, + "algorithm": "APPO", + "latency_model": "qwenoft" +} diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/rollout_descriptor.json b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/rollout_descriptor.json new file mode 100644 index 0000000000000000000000000000000000000000..c6493605bc717218173ee38b6c0f3928bd6b9060 --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/rollout_descriptor.json @@ -0,0 +1,314 @@ +{ + "env_name": "air_raid", + "integration_name": "gymnasium", + "action_spec": { + "layout": "gymnasium_discrete_v1", + "labels": [ + "noop", + "fire", + "right", + "left", + "rightfire", + "leftfire" + ], + "row_fields": [ + "action_id", + "action", + "action_text" + ], + "reward_field": "raw_reward", + "episode_return_field": "episode_raw_return", + "reward_semantics": "raw_reward" + }, + "base_prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.", + "checkpoint_experiment_dir": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_models/air_raid_profile_20260911T054920Z", + "checkpoint_path": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_models/air_raid_profile_20260911T054920Z/checkpoint_p0/best_000034592_8855552_reward_9329.250.pth", + "flat_cfg": { + "help": false, + "algo": "APPO", + "env": "latency_gymnasium", + "experiment": "air_raid_profile_20260911T054920Z", + "train_dir": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_models", + "restart_behavior": "overwrite", + "device": "gpu", + "seed": 0, + "initial_model_path": null, + "num_policies": 1, + "async_rl": true, + "serial_mode": false, + "batched_sampling": false, + "num_batches_to_accumulate": 2, + "worker_num_splits": 2, + "policy_workers_per_policy": 1, + "max_policy_lag": 300, + "num_workers": 16, + "num_envs_per_worker": 8, + "batch_size": 1024, + "num_batches_per_epoch": 4, + "num_epochs": 4, + "rollout": 32, + "fasttd3_replay_capacity": 6553600, + "fasttd3_replay_batch_size": 32768, + "fasttd3_transitions_per_update": 64, + "fasttd3_v_min": -250.0, + "fasttd3_v_max": 250.0, + "fasttd3_actor_action_l2": 0.0, + "fasttd3_compile": true, + "fasttd3_sonic_decoder_path": null, + "recurrence": 1, + "shuffle_minibatches": false, + "gamma": 0.99, + "reward_scale": 1.0, + "reward_clip": 1000.0, + "value_bootstrap": false, + "normalize_returns": true, + "exploration_loss_coeff": 0.01, + "value_loss_coeff": 0.5, + "kl_loss_coeff": 0.0, + "exploration_loss": "entropy", + "gae_lambda": 0.95, + "ppo_clip_ratio": 0.1, + "ppo_clip_value": 0.2, + "with_vtrace": false, + "vtrace_rho": 1.0, + "vtrace_c": 1.0, + "optimizer": "adam", + "adam_eps": 1e-05, + "adam_beta1": 0.9, + "adam_beta2": 0.999, + "max_grad_norm": 0.5, + "learning_rate": 0.00025, + "lr_schedule": "linear_decay", + "lr_schedule_kl_threshold": 0.008, + "lr_adaptive_min": 1e-06, + "lr_adaptive_max": 0.01, + "obs_subtract_mean": 0.0, + "obs_scale": 255.0, + "normalize_input": true, + "normalize_input_keys": null, + "decorrelate_experience_max_seconds": 0, + "decorrelate_envs_on_one_worker": true, + "actor_worker_gpus": [], + "set_workers_cpu_affinity": true, + "force_envs_single_thread": false, + "default_niceness": 0, + "log_to_file": true, + "experiment_summaries_interval": 10, + "flush_summaries_interval": 30, + "stats_avg": 100, + "summaries_use_frameskip": true, + "heartbeat_interval": 20, + "heartbeat_reporting_interval": 180, + "train_for_env_steps": 10000000, + "train_for_seconds": 10000000000, + "save_every_sec": 600, + "keep_checkpoints": 5, + "load_checkpoint_kind": "latest", + "save_milestones_sec": -1, + "save_best_every_sec": 5, + "save_best_metric": "reward", + "save_best_after": 100000, + "benchmark": false, + "encoder_mlp_layers": [ + 512, + 512 + ], + "encoder_conv_architecture": "convnet_atari", + "encoder_conv_mlp_layers": [ + 512 + ], + "use_rnn": false, + "rnn_size": 512, + "rnn_type": "gru", + "rnn_num_layers": 1, + "decoder_mlp_layers": [], + "nonlinearity": "relu", + "policy_initialization": "orthogonal", + "policy_init_gain": 1.0, + "actor_critic_share_weights": true, + "adaptive_stddev": false, + "continuous_tanh_scale": 0.0, + "initial_stddev": 1.0, + "use_env_info_cache": false, + "env_gpu_actions": false, + "env_gpu_observations": true, + "env_frameskip": 1, + "env_framestack": 4, + "pixel_format": "CHW", + "use_record_episode_statistics": false, + "with_wandb": false, + "wandb_user": null, + "wandb_project": "sample_factory", + "wandb_group": null, + "wandb_job_type": "SF", + "wandb_tags": [], + "with_pbt": false, + "pbt_mix_policies_in_one_env": true, + "pbt_period_env_steps": 5000000, + "pbt_start_mutation": 20000000, + "pbt_replace_fraction": 0.3, + "pbt_mutation_rate": 0.15, + "pbt_replace_reward_gap": 0.1, + "pbt_replace_reward_gap_absolute": 1e-06, + "pbt_optimize_gamma": false, + "pbt_target_objective": "true_objective", + "pbt_perturb_min": 1.1, + "pbt_perturb_max": 1.5, + "gym_task_name": "air_raid", + "gym_env_id": "LatencyBench/AirRaid-v0", + "gym_make_kwargs_json": "{\"base_env_id\": \"ALE/AirRaid-v5\", \"render_mode\": \"rgb_array\", \"screen_size\": 84, \"noop_max\": 0, \"base_make_kwargs\": {\"obs_type\": \"rgb\", \"frameskip\": 1, \"repeat_action_probability\": 0.0, \"full_action_space\": false, \"mode\": 1, \"difficulty\": 0, \"max_num_frames_per_episode\": 108000}}", + "gym_registration_imports_json": "[\"latency_bench.envs.gymnasium_air_raid\"]", + "gym_action_space_json": "{\"type\": \"discrete\", \"labels\": [\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"], \"values\": [0, 1, 2, 3, 4, 5]}", + "gym_action_labels_json": "[\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"]", + "gym_action_values_json": "[0, 1, 2, 3, 4, 5]", + "gym_noop_action_id": 0, + "gym_noop_action_json": "\"noop\"", + "gym_base_prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.", + "gym_state_labels_json": "", + "env_fps": 50.0, + "obs_fps": 12.5, + "frame_stack": 4, + "export_env_raw_rgb_frames": false, + "mode": "train", + "latency_type": "iid", + "fixed_latency_ms": null, + "mean_latency_ms": null, + "std_latency_ms": null, + "min_latency_ms": null, + "max_latency_ms": null, + "latency_seed": 0, + "latency_profile_path": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/profile/profile.json", + "latency_profile_worker_slot": 0, + "simulated_worker_capacity": null, + "add_latency_info": false, + "max_pending_actions": null, + "hold_policy": "hold", + "ordering_policy": "issue_order_fifo", + "eval_episodes": 5, + "eval_parallel_envs": 1, + "eval_latency_raw_frame_values": null, + "eval_max_steps": 5000, + "eval_deterministic": true, + "eval_raw_reward": false, + "episode_metrics_path": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_models/air_raid_profile_20260911T054920Z/episode_metrics.jsonl", + "command_line": "--mode train --algo APPO --env latency_gymnasium --experiment air_raid_profile_20260911T054920Z --train_dir /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_models --restart_behavior overwrite --device gpu --seed 0 --episode_metrics_path /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_models/air_raid_profile_20260911T054920Z/episode_metrics.jsonl --train_for_env_steps 10000000 --num_workers 16 --num_envs_per_worker 8 --num_policies 1 --batch_size 1024 --rollout 32 --recurrence 1 --num_epochs 4 --num_batches_per_epoch 4 --worker_num_splits 2 --max_policy_lag 300 --num_batches_to_accumulate 2 --policy_workers_per_policy 1 --learning_rate 0.00025 --kl_loss_coeff 0.0 --lr_schedule_kl_threshold 0.008 --nonlinearity relu --policy_initialization orthogonal --continuous_tanh_scale 0.0 --initial_stddev 1.0 --env_framestack 4 --gamma 0.99 --gae_lambda 0.95 --ppo_clip_ratio 0.1 --ppo_clip_value 0.2 --exploration_loss entropy --exploration_loss_coeff 0.01 --reward_scale 1.0 --reward_clip 1000.0 --value_loss_coeff 0.5 --max_grad_norm 0.5 --optimizer adam --adam_eps 1e-05 --adam_beta1 0.9 --adam_beta2 0.999 --obs_subtract_mean 0.0 --obs_scale 255.0 --decorrelate_experience_max_seconds 0 --default_niceness 0 --rnn_type gru --rnn_size 512 --save_every_sec 600 --keep_checkpoints 5 --save_milestones_sec -1 --save_best_every_sec 5 --save_best_after 100000 --stats_avg 100 --experiment_summaries_interval 10 --async_rl True --batched_sampling False --serial_mode False --use_rnn False --normalize_returns True --normalize_input True --adaptive_stddev False --shuffle_minibatches False --value_bootstrap False --with_vtrace False --decorrelate_envs_on_one_worker True --set_workers_cpu_affinity True --force_envs_single_thread False --actor_critic_share_weights True --with_wandb False --encoder_mlp_layers 512 512 --latency-type iid --latency-seed 0 --latency-profile-path /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/profile/profile.json --latency-profile-worker-slot 0 --add-latency-info False --eval-episodes 5 --eval-parallel-envs 1 --eval-max-steps 5000 --eval-deterministic True --encoder_conv_architecture convnet_atari --gym-task-name air_raid --gym-env-id LatencyBench/AirRaid-v0 --gym-make-kwargs-json {\"base_env_id\": \"ALE/AirRaid-v5\", \"render_mode\": \"rgb_array\", \"screen_size\": 84, \"noop_max\": 0, \"base_make_kwargs\": {\"obs_type\": \"rgb\", \"frameskip\": 1, \"repeat_action_probability\": 0.0, \"full_action_space\": false, \"mode\": 1, \"difficulty\": 0, \"max_num_frames_per_episode\": 108000}} --gym-registration-imports-json [\"latency_bench.envs.gymnasium_air_raid\"] --gym-action-space-json {\"type\": \"discrete\", \"labels\": [\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"], \"values\": [0, 1, 2, 3, 4, 5]} --gym-noop-action-json \"noop\" --gym-action-labels-json [\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"] --gym-action-values-json [0, 1, 2, 3, 4, 5] --gym-noop-action-id 0 --gym-base-prompt Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire. --env-fps 50 --obs-fps 12.5 --frame-stack 4 --hold-policy hold --ordering-policy issue_order_fifo --use_env_info_cache False", + "cli_args": { + "algo": "APPO", + "env": "latency_gymnasium", + "experiment": "air_raid_profile_20260911T054920Z", + "train_dir": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_models", + "restart_behavior": "overwrite", + "device": "gpu", + "seed": 0, + "num_policies": 1, + "async_rl": true, + "serial_mode": false, + "batched_sampling": false, + "num_batches_to_accumulate": 2, + "worker_num_splits": 2, + "policy_workers_per_policy": 1, + "max_policy_lag": 300, + "num_workers": 16, + "num_envs_per_worker": 8, + "batch_size": 1024, + "num_batches_per_epoch": 4, + "num_epochs": 4, + "rollout": 32, + "recurrence": 1, + "shuffle_minibatches": false, + "gamma": 0.99, + "reward_scale": 1.0, + "reward_clip": 1000.0, + "value_bootstrap": false, + "normalize_returns": true, + "exploration_loss_coeff": 0.01, + "value_loss_coeff": 0.5, + "kl_loss_coeff": 0.0, + "exploration_loss": "entropy", + "gae_lambda": 0.95, + "ppo_clip_ratio": 0.1, + "ppo_clip_value": 0.2, + "with_vtrace": false, + "optimizer": "adam", + "adam_eps": 1e-05, + "adam_beta1": 0.9, + "adam_beta2": 0.999, + "max_grad_norm": 0.5, + "learning_rate": 0.00025, + "lr_schedule_kl_threshold": 0.008, + "obs_subtract_mean": 0.0, + "obs_scale": 255.0, + "normalize_input": true, + "decorrelate_experience_max_seconds": 0, + "decorrelate_envs_on_one_worker": true, + "set_workers_cpu_affinity": true, + "force_envs_single_thread": false, + "default_niceness": 0, + "experiment_summaries_interval": 10, + "stats_avg": 100, + "train_for_env_steps": 10000000, + "save_every_sec": 600, + "keep_checkpoints": 5, + "save_milestones_sec": -1, + "save_best_every_sec": 5, + "save_best_after": 100000, + "encoder_mlp_layers": [ + 512, + 512 + ], + "encoder_conv_architecture": "convnet_atari", + "use_rnn": false, + "rnn_size": 512, + "rnn_type": "gru", + "nonlinearity": "relu", + "policy_initialization": "orthogonal", + "actor_critic_share_weights": true, + "adaptive_stddev": false, + "continuous_tanh_scale": 0.0, + "initial_stddev": 1.0, + "use_env_info_cache": false, + "env_framestack": 4, + "with_wandb": false, + "gym_task_name": "air_raid", + "gym_env_id": "LatencyBench/AirRaid-v0", + "gym_make_kwargs_json": "{\"base_env_id\": \"ALE/AirRaid-v5\", \"render_mode\": \"rgb_array\", \"screen_size\": 84, \"noop_max\": 0, \"base_make_kwargs\": {\"obs_type\": \"rgb\", \"frameskip\": 1, \"repeat_action_probability\": 0.0, \"full_action_space\": false, \"mode\": 1, \"difficulty\": 0, \"max_num_frames_per_episode\": 108000}}", + "gym_registration_imports_json": "[\"latency_bench.envs.gymnasium_air_raid\"]", + "gym_action_space_json": "{\"type\": \"discrete\", \"labels\": [\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"], \"values\": [0, 1, 2, 3, 4, 5]}", + "gym_action_labels_json": "[\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"]", + "gym_action_values_json": "[0, 1, 2, 3, 4, 5]", + "gym_noop_action_id": 0, + "gym_noop_action_json": "\"noop\"", + "gym_base_prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.", + "env_fps": 50.0, + "obs_fps": 12.5, + "frame_stack": 4, + "mode": "train", + "latency_type": "iid", + "latency_seed": 0, + "latency_profile_path": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/profile/profile.json", + "latency_profile_worker_slot": 0, + "add_latency_info": false, + "hold_policy": "hold", + "ordering_policy": "issue_order_fifo", + "eval_episodes": 5, + "eval_parallel_envs": 1, + "eval_max_steps": 5000, + "eval_deterministic": true, + "episode_metrics_path": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_models/air_raid_profile_20260911T054920Z/episode_metrics.jsonl", + "output_dir": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_train" + }, + "git_hash": "unknown", + "git_repo_name": "not a git repository", + "eval_env_frameskip": 1, + "output_dir": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_train" + }, + "device_override": "gpu", + "latency_override": { + "method": "iid", + "fixed_latency_ms": null, + "profile_path": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/profile/profile.json", + "profile_worker_slot": 0, + "seed": 0, + "add_latency_info": false + } +} diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/rollout_metadata.json b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/rollout_metadata.json new file mode 100644 index 0000000000000000000000000000000000000000..dcb4c75317190236575975d5828e2d207d5f0fbc --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/rollout_metadata.json @@ -0,0 +1,336 @@ +{ + "checkpoint_experiment_dir": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_models/air_raid_profile_20260911T054920Z", + "checkpoint_path": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_models/air_raid_profile_20260911T054920Z/checkpoint_p0/best_000034592_8855552_reward_9329.250.pth", + "config_source": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_models/air_raid_profile_20260911T054920Z/config.json", + "latency_source": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/h100/small_train.yaml", + "checkpoint_config": { + "help": false, + "algo": "APPO", + "env": "latency_gymnasium", + "experiment": "air_raid_profile_20260911T054920Z", + "train_dir": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_models", + "restart_behavior": "overwrite", + "device": "gpu", + "seed": 0, + "initial_model_path": null, + "num_policies": 1, + "async_rl": true, + "serial_mode": false, + "batched_sampling": false, + "num_batches_to_accumulate": 2, + "worker_num_splits": 2, + "policy_workers_per_policy": 1, + "max_policy_lag": 300, + "num_workers": 16, + "num_envs_per_worker": 8, + "batch_size": 1024, + "num_batches_per_epoch": 4, + "num_epochs": 4, + "rollout": 32, + "fasttd3_replay_capacity": 6553600, + "fasttd3_replay_batch_size": 32768, + "fasttd3_transitions_per_update": 64, + "fasttd3_v_min": -250.0, + "fasttd3_v_max": 250.0, + "fasttd3_actor_action_l2": 0.0, + "fasttd3_compile": true, + "fasttd3_sonic_decoder_path": null, + "recurrence": 1, + "shuffle_minibatches": false, + "gamma": 0.99, + "reward_scale": 1.0, + "reward_clip": 1000.0, + "value_bootstrap": false, + "normalize_returns": true, + "exploration_loss_coeff": 0.01, + "value_loss_coeff": 0.5, + "kl_loss_coeff": 0.0, + "exploration_loss": "entropy", + "gae_lambda": 0.95, + "ppo_clip_ratio": 0.1, + "ppo_clip_value": 0.2, + "with_vtrace": false, + "vtrace_rho": 1.0, + "vtrace_c": 1.0, + "optimizer": "adam", + "adam_eps": 1e-05, + "adam_beta1": 0.9, + "adam_beta2": 0.999, + "max_grad_norm": 0.5, + "learning_rate": 0.00025, + "lr_schedule": "linear_decay", + "lr_schedule_kl_threshold": 0.008, + "lr_adaptive_min": 1e-06, + "lr_adaptive_max": 0.01, + "obs_subtract_mean": 0.0, + "obs_scale": 255.0, + "normalize_input": true, + "normalize_input_keys": null, + "decorrelate_experience_max_seconds": 0, + "decorrelate_envs_on_one_worker": true, + "actor_worker_gpus": [], + "set_workers_cpu_affinity": true, + "force_envs_single_thread": false, + "default_niceness": 0, + "log_to_file": true, + "experiment_summaries_interval": 10, + "flush_summaries_interval": 30, + "stats_avg": 100, + "summaries_use_frameskip": true, + "heartbeat_interval": 20, + "heartbeat_reporting_interval": 180, + "train_for_env_steps": 10000000, + "train_for_seconds": 10000000000, + "save_every_sec": 600, + "keep_checkpoints": 5, + "load_checkpoint_kind": "latest", + "save_milestones_sec": -1, + "save_best_every_sec": 5, + "save_best_metric": "reward", + "save_best_after": 100000, + "benchmark": false, + "encoder_mlp_layers": [ + 512, + 512 + ], + "encoder_conv_architecture": "convnet_atari", + "encoder_conv_mlp_layers": [ + 512 + ], + "use_rnn": false, + "rnn_size": 512, + "rnn_type": "gru", + "rnn_num_layers": 1, + "decoder_mlp_layers": [], + "nonlinearity": "relu", + "policy_initialization": "orthogonal", + "policy_init_gain": 1.0, + "actor_critic_share_weights": true, + "adaptive_stddev": false, + "continuous_tanh_scale": 0.0, + "initial_stddev": 1.0, + "use_env_info_cache": false, + "env_gpu_actions": false, + "env_gpu_observations": true, + "env_frameskip": 1, + "env_framestack": 4, + "pixel_format": "CHW", + "use_record_episode_statistics": false, + "with_wandb": false, + "wandb_user": null, + "wandb_project": "sample_factory", + "wandb_group": null, + "wandb_job_type": "SF", + "wandb_tags": [], + "with_pbt": false, + "pbt_mix_policies_in_one_env": true, + "pbt_period_env_steps": 5000000, + "pbt_start_mutation": 20000000, + "pbt_replace_fraction": 0.3, + "pbt_mutation_rate": 0.15, + "pbt_replace_reward_gap": 0.1, + "pbt_replace_reward_gap_absolute": 1e-06, + "pbt_optimize_gamma": false, + "pbt_target_objective": "true_objective", + "pbt_perturb_min": 1.1, + "pbt_perturb_max": 1.5, + "gym_task_name": "air_raid", + "gym_env_id": "LatencyBench/AirRaid-v0", + "gym_make_kwargs_json": "{\"base_env_id\": \"ALE/AirRaid-v5\", \"render_mode\": \"rgb_array\", \"screen_size\": 84, \"noop_max\": 0, \"base_make_kwargs\": {\"obs_type\": \"rgb\", \"frameskip\": 1, \"repeat_action_probability\": 0.0, \"full_action_space\": false, \"mode\": 1, \"difficulty\": 0, \"max_num_frames_per_episode\": 108000}}", + "gym_registration_imports_json": "[\"latency_bench.envs.gymnasium_air_raid\"]", + "gym_action_space_json": "{\"type\": \"discrete\", \"labels\": [\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"], \"values\": [0, 1, 2, 3, 4, 5]}", + "gym_action_labels_json": "[\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"]", + "gym_action_values_json": "[0, 1, 2, 3, 4, 5]", + "gym_noop_action_id": 0, + "gym_noop_action_json": "\"noop\"", + "gym_base_prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.", + "gym_state_labels_json": "", + "env_fps": 50.0, + "obs_fps": 12.5, + "frame_stack": 4, + "export_env_raw_rgb_frames": false, + "mode": "train", + "latency_type": "iid", + "fixed_latency_ms": null, + "mean_latency_ms": null, + "std_latency_ms": null, + "min_latency_ms": null, + "max_latency_ms": null, + "latency_seed": 0, + "latency_profile_path": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/profile/profile.json", + "latency_profile_worker_slot": 0, + "simulated_worker_capacity": null, + "add_latency_info": false, + "max_pending_actions": null, + "hold_policy": "hold", + "ordering_policy": "issue_order_fifo", + "eval_episodes": 5, + "eval_parallel_envs": 1, + "eval_latency_raw_frame_values": null, + "eval_max_steps": 5000, + "eval_deterministic": true, + "eval_raw_reward": false, + "episode_metrics_path": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_models/air_raid_profile_20260911T054920Z/episode_metrics.jsonl", + "command_line": "--mode train --algo APPO --env latency_gymnasium --experiment air_raid_profile_20260911T054920Z --train_dir /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_models --restart_behavior overwrite --device gpu --seed 0 --episode_metrics_path /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_models/air_raid_profile_20260911T054920Z/episode_metrics.jsonl --train_for_env_steps 10000000 --num_workers 16 --num_envs_per_worker 8 --num_policies 1 --batch_size 1024 --rollout 32 --recurrence 1 --num_epochs 4 --num_batches_per_epoch 4 --worker_num_splits 2 --max_policy_lag 300 --num_batches_to_accumulate 2 --policy_workers_per_policy 1 --learning_rate 0.00025 --kl_loss_coeff 0.0 --lr_schedule_kl_threshold 0.008 --nonlinearity relu --policy_initialization orthogonal --continuous_tanh_scale 0.0 --initial_stddev 1.0 --env_framestack 4 --gamma 0.99 --gae_lambda 0.95 --ppo_clip_ratio 0.1 --ppo_clip_value 0.2 --exploration_loss entropy --exploration_loss_coeff 0.01 --reward_scale 1.0 --reward_clip 1000.0 --value_loss_coeff 0.5 --max_grad_norm 0.5 --optimizer adam --adam_eps 1e-05 --adam_beta1 0.9 --adam_beta2 0.999 --obs_subtract_mean 0.0 --obs_scale 255.0 --decorrelate_experience_max_seconds 0 --default_niceness 0 --rnn_type gru --rnn_size 512 --save_every_sec 600 --keep_checkpoints 5 --save_milestones_sec -1 --save_best_every_sec 5 --save_best_after 100000 --stats_avg 100 --experiment_summaries_interval 10 --async_rl True --batched_sampling False --serial_mode False --use_rnn False --normalize_returns True --normalize_input True --adaptive_stddev False --shuffle_minibatches False --value_bootstrap False --with_vtrace False --decorrelate_envs_on_one_worker True --set_workers_cpu_affinity True --force_envs_single_thread False --actor_critic_share_weights True --with_wandb False --encoder_mlp_layers 512 512 --latency-type iid --latency-seed 0 --latency-profile-path /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/profile/profile.json --latency-profile-worker-slot 0 --add-latency-info False --eval-episodes 5 --eval-parallel-envs 1 --eval-max-steps 5000 --eval-deterministic True --encoder_conv_architecture convnet_atari --gym-task-name air_raid --gym-env-id LatencyBench/AirRaid-v0 --gym-make-kwargs-json {\"base_env_id\": \"ALE/AirRaid-v5\", \"render_mode\": \"rgb_array\", \"screen_size\": 84, \"noop_max\": 0, \"base_make_kwargs\": {\"obs_type\": \"rgb\", \"frameskip\": 1, \"repeat_action_probability\": 0.0, \"full_action_space\": false, \"mode\": 1, \"difficulty\": 0, \"max_num_frames_per_episode\": 108000}} --gym-registration-imports-json [\"latency_bench.envs.gymnasium_air_raid\"] --gym-action-space-json {\"type\": \"discrete\", \"labels\": [\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"], \"values\": [0, 1, 2, 3, 4, 5]} --gym-noop-action-json \"noop\" --gym-action-labels-json [\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"] --gym-action-values-json [0, 1, 2, 3, 4, 5] --gym-noop-action-id 0 --gym-base-prompt Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire. --env-fps 50 --obs-fps 12.5 --frame-stack 4 --hold-policy hold --ordering-policy issue_order_fifo --use_env_info_cache False", + "cli_args": { + "algo": "APPO", + "env": "latency_gymnasium", + "experiment": "air_raid_profile_20260911T054920Z", + "train_dir": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_models", + "restart_behavior": "overwrite", + "device": "gpu", + "seed": 0, + "num_policies": 1, + "async_rl": true, + "serial_mode": false, + "batched_sampling": false, + "num_batches_to_accumulate": 2, + "worker_num_splits": 2, + "policy_workers_per_policy": 1, + "max_policy_lag": 300, + "num_workers": 16, + "num_envs_per_worker": 8, + "batch_size": 1024, + "num_batches_per_epoch": 4, + "num_epochs": 4, + "rollout": 32, + "recurrence": 1, + "shuffle_minibatches": false, + "gamma": 0.99, + "reward_scale": 1.0, + "reward_clip": 1000.0, + "value_bootstrap": false, + "normalize_returns": true, + "exploration_loss_coeff": 0.01, + "value_loss_coeff": 0.5, + "kl_loss_coeff": 0.0, + "exploration_loss": "entropy", + "gae_lambda": 0.95, + "ppo_clip_ratio": 0.1, + "ppo_clip_value": 0.2, + "with_vtrace": false, + "optimizer": "adam", + "adam_eps": 1e-05, + "adam_beta1": 0.9, + "adam_beta2": 0.999, + "max_grad_norm": 0.5, + "learning_rate": 0.00025, + "lr_schedule_kl_threshold": 0.008, + "obs_subtract_mean": 0.0, + "obs_scale": 255.0, + "normalize_input": true, + "decorrelate_experience_max_seconds": 0, + "decorrelate_envs_on_one_worker": true, + "set_workers_cpu_affinity": true, + "force_envs_single_thread": false, + "default_niceness": 0, + "experiment_summaries_interval": 10, + "stats_avg": 100, + "train_for_env_steps": 10000000, + "save_every_sec": 600, + "keep_checkpoints": 5, + "save_milestones_sec": -1, + "save_best_every_sec": 5, + "save_best_after": 100000, + "encoder_mlp_layers": [ + 512, + 512 + ], + "encoder_conv_architecture": "convnet_atari", + "use_rnn": false, + "rnn_size": 512, + "rnn_type": "gru", + "nonlinearity": "relu", + "policy_initialization": "orthogonal", + "actor_critic_share_weights": true, + "adaptive_stddev": false, + "continuous_tanh_scale": 0.0, + "initial_stddev": 1.0, + "use_env_info_cache": false, + "env_framestack": 4, + "with_wandb": false, + "gym_task_name": "air_raid", + "gym_env_id": "LatencyBench/AirRaid-v0", + "gym_make_kwargs_json": "{\"base_env_id\": \"ALE/AirRaid-v5\", \"render_mode\": \"rgb_array\", \"screen_size\": 84, \"noop_max\": 0, \"base_make_kwargs\": {\"obs_type\": \"rgb\", \"frameskip\": 1, \"repeat_action_probability\": 0.0, \"full_action_space\": false, \"mode\": 1, \"difficulty\": 0, \"max_num_frames_per_episode\": 108000}}", + "gym_registration_imports_json": "[\"latency_bench.envs.gymnasium_air_raid\"]", + "gym_action_space_json": "{\"type\": \"discrete\", \"labels\": [\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"], \"values\": [0, 1, 2, 3, 4, 5]}", + "gym_action_labels_json": "[\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"]", + "gym_action_values_json": "[0, 1, 2, 3, 4, 5]", + "gym_noop_action_id": 0, + "gym_noop_action_json": "\"noop\"", + "gym_base_prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.", + "env_fps": 50.0, + "obs_fps": 12.5, + "frame_stack": 4, + "mode": "train", + "latency_type": "iid", + "latency_seed": 0, + "latency_profile_path": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/profile/profile.json", + "latency_profile_worker_slot": 0, + "add_latency_info": false, + "hold_policy": "hold", + "ordering_policy": "issue_order_fifo", + "eval_episodes": 5, + "eval_parallel_envs": 1, + "eval_max_steps": 5000, + "eval_deterministic": true, + "episode_metrics_path": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_models/air_raid_profile_20260911T054920Z/episode_metrics.jsonl", + "output_dir": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_train" + }, + "git_hash": "unknown", + "git_repo_name": "not a git repository", + "eval_env_frameskip": 1, + "output_dir": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_train" + }, + "latency_override": { + "method": "iid", + "fixed_latency_ms": null, + "profile_path": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/profile/profile.json", + "profile_worker_slot": 0, + "seed": 0, + "add_latency_info": false + }, + "gymnasium_task": { + "task_name": "air_raid", + "env_id": "LatencyBench/AirRaid-v0", + "make_kwargs": { + "base_env_id": "ALE/AirRaid-v5", + "render_mode": "rgb_array", + "screen_size": 84, + "noop_max": 0, + "base_make_kwargs": { + "obs_type": "rgb", + "frameskip": 1, + "repeat_action_probability": 0.0, + "full_action_space": false, + "mode": 1, + "difficulty": 0, + "max_num_frames_per_episode": 108000 + } + }, + "registration_imports": [ + "latency_bench.envs.gymnasium_air_raid" + ], + "base_prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.", + "env_fps": 50.0, + "obs_fps": 12.5, + "frame_stack": 4, + "action_labels": [ + "noop", + "fire", + "right", + "left", + "rightfire", + "leftfire" + ], + "action_values": [ + 0, + 1, + 2, + 3, + 4, + 5 + ], + "noop_action_id": 0 + } +} diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/selection.json b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/selection.json new file mode 100644 index 0000000000000000000000000000000000000000..f334b8dfc345f1fe5804d1797319e6029dc2d838 --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/selection.json @@ -0,0 +1,14 @@ +{ + "run_id": "20260911T054920Z-airraid", + "source_profile_run_id": "20260911T023704401836Z", + "selection_rule": "best_reward_after_full_training_budget", + "best_checkpoint": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_models/air_raid_profile_20260911T054920Z/checkpoint_p0/best_000034592_8855552_reward_9329.250.pth", + "best_bundle_path": "checkpoint_p0/best_000034592_8855552_reward_9329.250.pth", + "final_checkpoint": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_models/air_raid_profile_20260911T054920Z/checkpoint_p0/checkpoint_000039076_10006528.pth", + "final_env_steps": 10006528, + "final_train_step": 39076, + "training_exit_code": 0, + "profile_sha256": "52fb3581d2dcbd6b46e634b04fc3436f254b5e82210b9fbd5a44635dc1520dad", + "code_root_sha": "86df5ecef6735d5d824944e7acdee5bf204e1545", + "note": "config and rollout descriptor preserve exact H100 training paths; rebind latency profile to bundled profile/profile.json when moving hosts" +} diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/small_train.yaml b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/small_train.yaml new file mode 100644 index 0000000000000000000000000000000000000000..bf56ad093d690548162c3f4fe79276b48e64b079 --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwenoft-rtx3090-v1/small_train.yaml @@ -0,0 +1,156 @@ +experiment: + name: air_raid_profile_20260911T054920Z + seed: 0 +backend: + type: sample_factory + algo: APPO + device: cuda + train_dir: /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_models + restart_behavior: overwrite + run_mode: train + extra_args: + - --encoder_conv_architecture + - convnet_atari +executor: + mode: simulated +env: + name: gymnasium + task_name: air_raid + env_id: LatencyBench/AirRaid-v0 + registration_imports: + - latency_bench.envs.gymnasium_air_raid + make_kwargs: + base_env_id: ALE/AirRaid-v5 + render_mode: rgb_array + screen_size: 84 + noop_max: 0 + base_make_kwargs: + obs_type: rgb + frameskip: 1 + repeat_action_probability: 0.0 + full_action_space: false + mode: 1 + difficulty: 0 + max_num_frames_per_episode: 108000 + env_fps: 50 + obs_fps: 12.5 + frame_stack: 4 + noop_action: noop + action_map: + noop: 0 + fire: 1 + right: 2 + left: 3 + rightfire: 4 + leftfire: 5 + action_order: + - noop + - fire + - right + - left + - rightfire + - leftfire + base_prompt: 'Protect both buildings from flying saucers. Choose exactly one action + from: noop, fire, right, left, rightfire, leftfire.' +latency: + method: iid + fixed_latency_ms: null + profile_path: /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/profile/profile.json + profile_worker_slot: 0 + seed: 0 + add_latency_info: false +scheduler: + hold_policy: hold + ordering_policy: issue_order_fifo +policy: + type: random + actions: + - noop + - fire + - right + - left + - rightfire + - leftfire +training: + train_for_env_steps: 10000000 + num_workers: 16 + num_envs_per_worker: 8 + worker_num_splits: 2 + num_policies: 1 + batch_size: 1024 + rollout: 32 + recurrence: 1 + env_framestack: 4 + num_epochs: 4 + num_batches_per_epoch: 4 + max_policy_lag: 300 + learning_rate: 0.00025 + lr_schedule: linear_decay + gamma: 0.99 + gae_lambda: 0.95 + ppo_clip_ratio: 0.1 + ppo_clip_value: 0.2 + value_loss_coeff: 0.5 + max_grad_norm: 0.5 + exploration_loss_coeff: 0.01 + exploration_loss: entropy + encoder_conv_architecture: convnet_atari + nonlinearity: relu + obs_scale: 255.0 + adam_eps: 1.0e-05 + with_vtrace: false + adaptive_stddev: false + async_rl: true + use_rnn: false + normalize_input: true + normalize_returns: true + save_every_sec: 600 + keep_checkpoints: 5 + num_batches_to_accumulate: 2 + policy_workers_per_policy: 1 + kl_loss_coeff: 0.0 + lr_schedule_kl_threshold: 0.008 + policy_initialization: orthogonal + continuous_tanh_scale: 0.0 + initial_stddev: 1.0 + reward_scale: 1.0 + reward_clip: 1000.0 + optimizer: adam + adam_beta1: 0.9 + adam_beta2: 0.999 + obs_subtract_mean: 0.0 + decorrelate_experience_max_seconds: 0 + default_niceness: 0 + rnn_type: gru + rnn_size: 512 + save_milestones_sec: -1 + save_best_every_sec: 5 + save_best_after: 100000 + stats_avg: 100 + experiment_summaries_interval: 10 + batched_sampling: false + serial_mode: false + shuffle_minibatches: false + value_bootstrap: false + decorrelate_envs_on_one_worker: true + set_workers_cpu_affinity: true + force_envs_single_thread: false + actor_critic_share_weights: true + with_wandb: false + encoder_mlp_layers: + - 512 + - 512 +evaluation: + eval_interval_steps: null + eval_episodes: 5 + eval_parallel_envs: 1 + eval_max_steps: 5000 + eval_deterministic: true +logging: + output_dir: /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/small_train + video: + enabled: false + num_bins: 1 + save_step_records: false + save_action_records: false + save_latency_records: false diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/README.md b/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/README.md new file mode 100644 index 0000000000000000000000000000000000000000..8836fadfc00fca5c1be6cac38ac56b4f8956f3ad --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/README.md @@ -0,0 +1,30 @@ +# air-raid / sample-factory-appo + +Training condition: `latency-aware`. Run: `airraid_pi05_profile_appo_h1_10env_2p5obs_g128_20260920`. + +Copied from [Standard-Pipeline](https://huggingface.co/latency-sensitive-bench/Standard-Pipeline/tree/2208875f92b21e3f8ec2363a30a6705dfe56cc04/air_raid/profile_latency/small_model/Pi05) at `2208875f92b21e3f8ec2363a30a6705dfe56cc04`. + +- Checkpoint: `checkpoint.pth` +- Selection: final +- Checkpoint SHA256: `d28417644d97d440be8677b469442692eb8711aca55c9e3c0dcd3f9c73a7530d` +- Configuration, FPS, actions, normalization and latency semantics retain their source values. +- This import does not run evaluations or change training/evaluation/acceptance status. + +Sample Factory APPO inference bundle: `config.json` and the explicit `checkpoint.pth`. +The checkpoint contains `model`, `train_step`, and `env_steps`; every exported model +tensor was reloaded and compared bit-for-bit with the selected source checkpoint. + +See [the original model notes](source/README.md) for experiment results and limitations. + +[Original provenance](source/provenance.json) retains dataset references and recorded experiment evidence. + +## Download + +```python +from huggingface_hub import snapshot_download +prefix = "latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1" +snapshot_download("latency-sensitive-bench/benchmark-models", revision="", + allow_patterns=[prefix + "/**"]) +``` + +Use the verified import commit listed in the repository's import report. diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/checkpoint.pth b/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/checkpoint.pth new file mode 100644 index 0000000000000000000000000000000000000000..55886a2331ea6e17b878d5124316c89d1ecd7b4a --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/checkpoint.pth @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:d28417644d97d440be8677b469442692eb8711aca55c9e3c0dcd3f9c73a7530d +size 7210293 diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/config.json b/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/config.json new file mode 100644 index 0000000000000000000000000000000000000000..b6bec0907d90ca1a0b659394728dca54a237ab36 --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/config.json @@ -0,0 +1,269 @@ +{ + "help": false, + "algo": "APPO", + "env": "latency_gymnasium", + "experiment": "airraid_pi05_profile_appo_h1_10env_2p5obs_g128_20260920", + "train_dir": "${PI05_RUN_DIR}/profile_latency/teacher/checkpoints", + "restart_behavior": "overwrite", + "device": "gpu", + "seed": 0, + "initial_model_path": null, + "num_policies": 1, + "async_rl": true, + "serial_mode": false, + "batched_sampling": false, + "num_batches_to_accumulate": 2, + "worker_num_splits": 2, + "policy_workers_per_policy": 1, + "max_policy_lag": 300, + "num_workers": 16, + "num_envs_per_worker": 8, + "batch_size": 1024, + "num_batches_per_epoch": 4, + "num_epochs": 4, + "rollout": 32, + "fasttd3_replay_capacity": 6553600, + "fasttd3_replay_batch_size": 32768, + "fasttd3_action_chunk_horizon": 1, + "fasttd3_transitions_per_update": 64, + "fasttd3_train_for_optimizer_steps": 10000000000, + "fasttd3_v_min": -250.0, + "fasttd3_v_max": 250.0, + "fasttd3_actor_action_l2": 0.0, + "fasttd3_compile": true, + "recurrence": 1, + "shuffle_minibatches": false, + "gamma": 0.99, + "reward_scale": 1.0, + "reward_clip": 1000.0, + "value_bootstrap": false, + "normalize_returns": true, + "exploration_loss_coeff": 0.01, + "value_loss_coeff": 0.5, + "kl_loss_coeff": 0.0, + "exploration_loss": "entropy", + "gae_lambda": 0.95, + "ppo_clip_ratio": 0.1, + "ppo_clip_value": 0.2, + "with_vtrace": false, + "vtrace_rho": 1.0, + "vtrace_c": 1.0, + "optimizer": "adam", + "adam_eps": 1e-05, + "adam_beta1": 0.9, + "adam_beta2": 0.999, + "max_grad_norm": 0.5, + "learning_rate": 0.00025, + "lr_schedule": "linear_decay", + "lr_schedule_kl_threshold": 0.008, + "lr_adaptive_min": 1e-06, + "lr_adaptive_max": 0.01, + "obs_subtract_mean": 0.0, + "obs_scale": 255.0, + "normalize_input": true, + "normalize_input_keys": null, + "decorrelate_experience_max_seconds": 0, + "decorrelate_envs_on_one_worker": true, + "actor_worker_gpus": [], + "set_workers_cpu_affinity": true, + "force_envs_single_thread": false, + "default_niceness": 0, + "log_to_file": true, + "experiment_summaries_interval": 10, + "flush_summaries_interval": 30, + "stats_avg": 100, + "summaries_use_frameskip": true, + "heartbeat_interval": 20, + "heartbeat_reporting_interval": 180, + "train_for_env_steps": 10000000, + "train_for_seconds": 10000000000, + "save_every_sec": 600, + "keep_checkpoints": 5, + "load_checkpoint_kind": "latest", + "save_milestones_sec": -1, + "save_best_every_sec": 5, + "save_best_metric": "reward", + "save_best_after": 100000, + "benchmark": false, + "encoder_mlp_layers": [ + 512, + 512 + ], + "encoder_conv_architecture": "convnet_atari", + "encoder_conv_mlp_layers": [ + 512 + ], + "use_rnn": false, + "rnn_size": 512, + "rnn_type": "gru", + "rnn_num_layers": 1, + "decoder_mlp_layers": [], + "nonlinearity": "relu", + "policy_initialization": "orthogonal", + "policy_init_gain": 1.0, + "actor_critic_share_weights": true, + "adaptive_stddev": false, + "continuous_tanh_scale": 0.0, + "initial_stddev": 1.0, + "use_env_info_cache": false, + "env_gpu_actions": false, + "env_gpu_observations": true, + "env_frameskip": 1, + "env_framestack": 4, + "pixel_format": "CHW", + "use_record_episode_statistics": false, + "with_wandb": true, + "wandb_user": "dongqianyu99-zhejiang-university", + "wandb_project": "latency-sensitive-bench", + "wandb_group": "pi05-airraid-full-pipeline", + "wandb_job_type": "profile_teacher", + "wandb_tags": [ + "air_raid", + "APPO", + "Pi05", + "measured-IID-profile", + "H1", + "env10-obs2p5FPS" + ], + "with_pbt": false, + "pbt_mix_policies_in_one_env": true, + "pbt_period_env_steps": 5000000, + "pbt_start_mutation": 20000000, + "pbt_replace_fraction": 0.3, + "pbt_mutation_rate": 0.15, + "pbt_replace_reward_gap": 0.1, + "pbt_replace_reward_gap_absolute": 1e-06, + "pbt_optimize_gamma": false, + "pbt_target_objective": "true_objective", + "pbt_perturb_min": 1.1, + "pbt_perturb_max": 1.5, + "gym_task_name": "air_raid", + "gym_env_id": "LatencyBench/AirRaid-v0", + "gym_make_kwargs_json": "{\"base_env_id\": \"ALE/AirRaid-v5\", \"render_mode\": \"rgb_array\", \"screen_size\": 84, \"noop_max\": 0, \"base_make_kwargs\": {\"obs_type\": \"rgb\", \"frameskip\": 1, \"repeat_action_probability\": 0.0, \"full_action_space\": false, \"mode\": 1, \"difficulty\": 0, \"max_num_frames_per_episode\": 108000}}", + "gym_registration_imports_json": "[\"latency_bench.envs.gymnasium_air_raid\"]", + "gym_action_space_json": "{\"type\": \"discrete\", \"labels\": [\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"], \"values\": [0, 1, 2, 3, 4, 5]}", + "gym_action_labels_json": "[\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"]", + "gym_action_values_json": "[0, 1, 2, 3, 4, 5]", + "gym_noop_action_id": 0, + "gym_noop_action_json": "\"noop\"", + "gym_base_prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.", + "gym_state_labels_json": "", + "env_fps": 10.0, + "obs_fps": 2.5, + "frame_stack": 4, + "export_env_raw_rgb_frames": false, + "mode": "train", + "latency_type": "iid", + "fixed_latency_ms": null, + "mean_latency_ms": null, + "std_latency_ms": null, + "min_latency_ms": null, + "max_latency_ms": null, + "latency_seed": 0, + "latency_profile_path": "${PI05_RUN_DIR}/profile_latency/profile/profile.json", + "latency_profile_worker_slot": 0, + "simulated_worker_capacity": null, + "add_latency_info": false, + "max_pending_actions": null, + "hold_last_chunk_action": false, + "hold_policy": "hold", + "ordering_policy": "issue_order_fifo", + "eval_episodes": 5, + "eval_parallel_envs": 1, + "eval_latency_raw_frame_values": null, + "eval_max_steps": 3600, + "eval_deterministic": true, + "eval_raw_reward": false, + "episode_metrics_path": "${PI05_RUN_DIR}/profile_latency/teacher/checkpoints/airraid_pi05_profile_appo_h1_10env_2p5obs_g128_20260920/episode_metrics.jsonl", + "ppo": null, + "command_line": "--mode train --algo APPO --env latency_gymnasium --experiment airraid_pi05_profile_appo_h1_10env_2p5obs_g128_20260920 --train_dir ${PI05_RUN_DIR}/profile_latency/teacher/checkpoints --restart_behavior overwrite --device gpu --seed 0 --episode_metrics_path ${PI05_RUN_DIR}/profile_latency/teacher/checkpoints/airraid_pi05_profile_appo_h1_10env_2p5obs_g128_20260920/episode_metrics.jsonl --train_for_env_steps 10000000 --num_workers 16 --num_envs_per_worker 8 --num_policies 1 --batch_size 1024 --rollout 32 --recurrence 1 --num_epochs 4 --num_batches_per_epoch 4 --worker_num_splits 2 --max_policy_lag 300 --learning_rate 0.00025 --nonlinearity relu --env_framestack 4 --gamma 0.99 --gae_lambda 0.95 --ppo_clip_ratio 0.1 --ppo_clip_value 0.2 --exploration_loss entropy --exploration_loss_coeff 0.01 --value_loss_coeff 0.5 --max_grad_norm 0.5 --adam_eps 1e-05 --obs_scale 255.0 --save_every_sec 600 --keep_checkpoints 5 --async_rl True --use_rnn False --normalize_returns True --normalize_input True --adaptive_stddev False --with_vtrace False --latency-type iid --latency-seed 0 --latency-profile-path ${PI05_RUN_DIR}/profile_latency/profile/profile.json --latency-profile-worker-slot 0 --add-latency-info False --eval-episodes 5 --eval-parallel-envs 1 --eval-max-steps 3600 --eval-deterministic True --with_wandb True --wandb_project latency-sensitive-bench --wandb_group pi05-airraid-full-pipeline --wandb_job_type profile_teacher --wandb_tags air_raid APPO Pi05 measured-IID-profile H1 env10-obs2p5FPS --encoder_conv_architecture convnet_atari --wandb_user dongqianyu99-zhejiang-university --gym-task-name air_raid --gym-env-id LatencyBench/AirRaid-v0 --gym-make-kwargs-json {\"base_env_id\": \"ALE/AirRaid-v5\", \"render_mode\": \"rgb_array\", \"screen_size\": 84, \"noop_max\": 0, \"base_make_kwargs\": {\"obs_type\": \"rgb\", \"frameskip\": 1, \"repeat_action_probability\": 0.0, \"full_action_space\": false, \"mode\": 1, \"difficulty\": 0, \"max_num_frames_per_episode\": 108000}} --gym-registration-imports-json [\"latency_bench.envs.gymnasium_air_raid\"] --gym-action-space-json {\"type\": \"discrete\", \"labels\": [\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"], \"values\": [0, 1, 2, 3, 4, 5]} --gym-noop-action-json \"noop\" --gym-action-labels-json [\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"] --gym-action-values-json [0, 1, 2, 3, 4, 5] --gym-noop-action-id 0 --gym-base-prompt Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire. --env-fps 10 --obs-fps 2.5 --frame-stack 4 --hold-policy hold --ordering-policy issue_order_fifo --use_env_info_cache False", + "cli_args": { + "algo": "APPO", + "env": "latency_gymnasium", + "experiment": "airraid_pi05_profile_appo_h1_10env_2p5obs_g128_20260920", + "train_dir": "${PI05_RUN_DIR}/profile_latency/teacher/checkpoints", + "restart_behavior": "overwrite", + "device": "gpu", + "seed": 0, + "num_policies": 1, + "async_rl": true, + "worker_num_splits": 2, + "max_policy_lag": 300, + "num_workers": 16, + "num_envs_per_worker": 8, + "batch_size": 1024, + "num_batches_per_epoch": 4, + "num_epochs": 4, + "rollout": 32, + "recurrence": 1, + "gamma": 0.99, + "normalize_returns": true, + "exploration_loss_coeff": 0.01, + "value_loss_coeff": 0.5, + "exploration_loss": "entropy", + "gae_lambda": 0.95, + "ppo_clip_ratio": 0.1, + "ppo_clip_value": 0.2, + "with_vtrace": false, + "adam_eps": 1e-05, + "max_grad_norm": 0.5, + "learning_rate": 0.00025, + "obs_scale": 255.0, + "normalize_input": true, + "train_for_env_steps": 10000000, + "save_every_sec": 600, + "keep_checkpoints": 5, + "encoder_conv_architecture": "convnet_atari", + "use_rnn": false, + "nonlinearity": "relu", + "adaptive_stddev": false, + "use_env_info_cache": false, + "env_framestack": 4, + "with_wandb": true, + "wandb_user": "dongqianyu99-zhejiang-university", + "wandb_project": "latency-sensitive-bench", + "wandb_group": "pi05-airraid-full-pipeline", + "wandb_job_type": "profile_teacher", + "wandb_tags": [ + "air_raid", + "APPO", + "Pi05", + "measured-IID-profile", + "H1", + "env10-obs2p5FPS" + ], + "gym_task_name": "air_raid", + "gym_env_id": "LatencyBench/AirRaid-v0", + "gym_make_kwargs_json": "{\"base_env_id\": \"ALE/AirRaid-v5\", \"render_mode\": \"rgb_array\", \"screen_size\": 84, \"noop_max\": 0, \"base_make_kwargs\": {\"obs_type\": \"rgb\", \"frameskip\": 1, \"repeat_action_probability\": 0.0, \"full_action_space\": false, \"mode\": 1, \"difficulty\": 0, \"max_num_frames_per_episode\": 108000}}", + "gym_registration_imports_json": "[\"latency_bench.envs.gymnasium_air_raid\"]", + "gym_action_space_json": "{\"type\": \"discrete\", \"labels\": [\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"], \"values\": [0, 1, 2, 3, 4, 5]}", + "gym_action_labels_json": "[\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"]", + "gym_action_values_json": "[0, 1, 2, 3, 4, 5]", + "gym_noop_action_id": 0, + "gym_noop_action_json": "\"noop\"", + "gym_base_prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.", + "env_fps": 10.0, + "obs_fps": 2.5, + "frame_stack": 4, + "mode": "train", + "latency_type": "iid", + "latency_seed": 0, + "latency_profile_path": "${PI05_RUN_DIR}/profile_latency/profile/profile.json", + "latency_profile_worker_slot": 0, + "add_latency_info": false, + "hold_policy": "hold", + "ordering_policy": "issue_order_fifo", + "eval_episodes": 5, + "eval_parallel_envs": 1, + "eval_max_steps": 3600, + "eval_deterministic": true, + "episode_metrics_path": "${PI05_RUN_DIR}/profile_latency/teacher/checkpoints/airraid_pi05_profile_appo_h1_10env_2p5obs_g128_20260920/episode_metrics.jsonl", + "output_dir": "${PI05_RUN_DIR}/profile_latency/teacher/training", + "wandb_unique_id": "airraid_pi05_profile_appo_h1_10env_2p5obs_g128_20260920" + }, + "git_hash": "unknown", + "git_repo_name": "not a git repository", + "eval_env_frameskip": 1, + "output_dir": "${PI05_RUN_DIR}/profile_latency/teacher/training", + "wandb_unique_id": "airraid_pi05_profile_appo_h1_10env_2p5obs_g128_20260920" +} \ No newline at end of file diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/provenance.json b/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/provenance.json new file mode 100644 index 0000000000000000000000000000000000000000..3e406fb31ca4c17a588e49cfc3082e7c5787fd13 --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/provenance.json @@ -0,0 +1,43 @@ +{ + "task": "air-raid", + "model": "sample-factory-appo", + "training_condition": "latency-aware", + "training_run_id": "airraid_pi05_profile_appo_h1_10env_2p5obs_g128_20260920", + "source": { + "repo_id": "latency-sensitive-bench/Standard-Pipeline", + "revision": "2208875f92b21e3f8ec2363a30a6705dfe56cc04", + "path": "air_raid/profile_latency/small_model/Pi05", + "url": "https://huggingface.co/latency-sensitive-bench/Standard-Pipeline/tree/2208875f92b21e3f8ec2363a30a6705dfe56cc04/air_raid/profile_latency/small_model/Pi05" + }, + "hf_repo_id": "latency-sensitive-bench/benchmark-models", + "hf_path": "latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1", + "checkpoint": { + "source_file": "air_raid/profile_latency/small_model/Pi05/checkpoint_p0/checkpoint_000039064_10002432.pth", + "source_sha256": "893f6435260687f1d9535a3d9bf2a06d54261de097be0ac782661a7efc8de9cc", + "source_bytes": 7211005, + "file": "checkpoint.pth", + "selection_rule": "final", + "method": "inference_export", + "sha256": "d28417644d97d440be8677b469442692eb8711aca55c9e3c0dcd3f9c73a7530d", + "bytes": 7210293, + "train_step": 39064, + "env_steps": 10002432, + "tensor_count": 18, + "tensor_equality_verified": true, + "removed_fields": [ + "best_performance", + "curr_lr" + ] + }, + "config_source": "air_raid/profile_latency/small_model/Pi05/config.json", + "task_contract_file": null, + "missing_source_files": [], + "migration": { + "operation": "copy_selected_inference_bundle", + "source_modified": false, + "evaluation_performed": false, + "experiment_status": "preserved_in_source_evidence" + }, + "algorithm": "APPO", + "latency_model": "qwenpi_v3" +} diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/selection.json b/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/selection.json new file mode 100644 index 0000000000000000000000000000000000000000..4472fe6933d522034220b4df504ed4beb3081b66 --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/selection.json @@ -0,0 +1,110 @@ +{ + "selected": "final", + "candidates": { + "final": { + "checkpoint": "${PI05_RUN_DIR}/profile_latency/teacher/checkpoints/airraid_pi05_profile_appo_h1_10env_2p5obs_g128_20260920/checkpoint_p0/checkpoint_000039064_10002432.pth", + "sha256": "59e7018b8aad3c0d7f52b40f9a5a4847c450b44f6cba27b66d052895869d4f10", + "episodes": 20, + "seeds": [ + 104771, + 104772, + 104773, + 104774, + 104775, + 104776, + 104777, + 104778, + 104779, + 104780, + 104781, + 104782, + 104783, + 104784, + 104785, + 104786, + 104787, + 104788, + 104789, + 104790 + ], + "returns": [ + 3775.0, + 3025.0, + 4000.0, + 3100.0, + 4050.0, + 4100.0, + 4050.0, + 3700.0, + 3925.0, + 4125.0, + 4100.0, + 4050.0, + 4125.0, + 4000.0, + 3975.0, + 4125.0, + 3525.0, + 4050.0, + 3975.0, + 4125.0 + ], + "mean": 3895.0, + "population_std": 316.68201717179966, + "strict_gt2500": 20 + }, + "training_best": { + "checkpoint": "${PI05_RUN_DIR}/profile_latency/teacher/checkpoints/airraid_pi05_profile_appo_h1_10env_2p5obs_g128_20260920/checkpoint_p0/best_000036640_9379840_reward_11643.500.pth", + "sha256": "ca875c7fce0848b8d62d2ed9c8a8654a198757fed984fa23b1940f503df725fa", + "episodes": 20, + "seeds": [ + 104771, + 104772, + 104773, + 104774, + 104775, + 104776, + 104777, + 104778, + 104779, + 104780, + 104781, + 104782, + 104783, + 104784, + 104785, + 104786, + 104787, + 104788, + 104789, + 104790 + ], + "returns": [ + 4100.0, + 3800.0, + 3050.0, + 4025.0, + 4150.0, + 4075.0, + 4150.0, + 2700.0, + 4075.0, + 3000.0, + 4150.0, + 3025.0, + 4125.0, + 3050.0, + 4125.0, + 4125.0, + 3850.0, + 3825.0, + 3925.0, + 3000.0 + ], + "mean": 3716.25, + "population_std": 503.5049031538819, + "strict_gt2500": 20 + } + }, + "completed_at": "2026-09-20T14:47:07.233886+00:00" +} diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/source/README.md b/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/source/README.md new file mode 100644 index 0000000000000000000000000000000000000000..8c94d58e34ffa7fccac3b72b3c93d3a267ab5e6c --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/source/README.md @@ -0,0 +1,5 @@ +# AirRaid Pi0.5 profile APPO teacher + +Original APPO seed0/10M counted-step budget; actualFinalenvsteps 10002432, approximately40M rawgameframes atstride4. Environment/observation10/2.5FPS, actualnewPi05 RTX3090 IID profile. Final20 mean3895.00, training-best20 mean3716.25; selected final. E10 mean3785.00, repeatsfirst10selectionseeds. Probe12/12 accepted strictly>2500. Formal100accepted has no pre-set totalattemptcap; actualattempts/replayQA remainpending. + +Trainingnoop0; evaluation/probenoop30+FireReset/cap3600rawframes. Singleworker evaluation, original training sampler unchanged. Inferenceexportsremoveoptimizeronly; everymodeltensorreloadedbit-identical, source/exportSHA recorded. diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/source/provenance.json b/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/source/provenance.json new file mode 100644 index 0000000000000000000000000000000000000000..7d7eb2e785892979844ee1fcc4eb57554f370f44 --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/source/provenance.json @@ -0,0 +1,31 @@ +{ + "state": "ACCEPTED_FOR_100_QUALIFIED_DEMONSTRATION_COLLECTION", + "selected": "final", + "profile_sha256": "7e81ed45da94173b447bdf0cf1ff59343cc7b4e1e6029bcf261cf054c2596cd6", + "exports": { + "final": { + "file": "checkpoint_p0/checkpoint_000039064_10002432.pth", + "source_sha256": "59e7018b8aad3c0d7f52b40f9a5a4847c450b44f6cba27b66d052895869d4f10", + "published_sha256": "893f6435260687f1d9535a3d9bf2a06d54261de097be0ac782661a7efc8de9cc", + "bytes": 7211005, + "env_steps": 10002432, + "train_step": 39064, + "removed": [ + "optimizer" + ], + "all_model_tensors_bit_identical": true + }, + "training_best": { + "file": "checkpoint_p0/best_000036640_9379840_reward_11643.500.pth", + "source_sha256": "ca875c7fce0848b8d62d2ed9c8a8654a198757fed984fa23b1940f503df725fa", + "published_sha256": "34f794088ec414487b401dcbe6c693cfdb8aa818ca51cfbd5d200621574f2c98", + "bytes": 7211309, + "env_steps": 9379840, + "train_step": 36640, + "removed": [ + "optimizer" + ], + "all_model_tensors_bit_identical": true + } + } +} diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/teacher.yaml b/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/teacher.yaml new file mode 100644 index 0000000000000000000000000000000000000000..a55585dc3160a46e068cae8a41524cbf27830057 --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/teacher.yaml @@ -0,0 +1,133 @@ +experiment: + name: airraid_pi05_profile_appo_h1_10env_2p5obs_g128_20260920 + seed: 0 +backend: + type: sample_factory + algo: APPO + device: cuda + train_dir: ${PI05_RUN_DIR}/profile_latency/teacher/checkpoints + restart_behavior: overwrite + run_mode: train + extra_args: + - --encoder_conv_architecture + - convnet_atari + - --wandb_user + - dongqianyu99-zhejiang-university +executor: + mode: simulated +env: + name: gymnasium + task_name: air_raid + env_id: LatencyBench/AirRaid-v0 + registration_imports: + - latency_bench.envs.gymnasium_air_raid + make_kwargs: + base_env_id: ALE/AirRaid-v5 + render_mode: rgb_array + screen_size: 84 + noop_max: 0 + base_make_kwargs: + obs_type: rgb + frameskip: 1 + repeat_action_probability: 0.0 + full_action_space: false + mode: 1 + difficulty: 0 + max_num_frames_per_episode: 108000 + env_fps: 10 + obs_fps: 2.5 + frame_stack: 4 + noop_action: noop + action_map: + noop: 0 + fire: 1 + right: 2 + left: 3 + rightfire: 4 + leftfire: 5 + action_order: + - noop + - fire + - right + - left + - rightfire + - leftfire + base_prompt: 'Protect both buildings from flying saucers. Choose exactly one action + from: noop, fire, right, left, rightfire, leftfire.' +latency: + method: iid + profile_path: ${PI05_RUN_DIR}/profile_latency/profile/profile.json + profile_worker_slot: 0 + seed: 0 + add_latency_info: false +scheduler: + hold_policy: hold + ordering_policy: issue_order_fifo +policy: + type: random + actions: + - noop + - fire + - right + - left + - rightfire + - leftfire +training: + train_for_env_steps: 10000000 + num_workers: 16 + num_envs_per_worker: 8 + worker_num_splits: 2 + num_policies: 1 + batch_size: 1024 + rollout: 32 + recurrence: 1 + env_framestack: 4 + num_epochs: 4 + num_batches_per_epoch: 4 + max_policy_lag: 300 + learning_rate: 0.00025 + lr_schedule: linear_decay + nonlinearity: relu + gamma: 0.99 + gae_lambda: 0.95 + ppo_clip_ratio: 0.1 + ppo_clip_value: 0.2 + value_loss_coeff: 0.5 + max_grad_norm: 0.5 + exploration_loss: entropy + exploration_loss_coeff: 0.01 + obs_scale: 255.0 + adam_eps: 1.0e-05 + adaptive_stddev: false + with_vtrace: false + async_rl: true + use_rnn: false + normalize_input: true + normalize_returns: true + save_every_sec: 600 + keep_checkpoints: 5 +evaluation: + eval_interval_steps: null + eval_episodes: 5 + eval_parallel_envs: 1 + eval_max_steps: 3600 + eval_deterministic: true +logging: + output_dir: ${PI05_RUN_DIR}/profile_latency/teacher/training + video: + enabled: false + num_bins: 1 + save_step_records: false + save_action_records: false + save_latency_records: false + wandb_project: latency-sensitive-bench + wandb_group: pi05-airraid-full-pipeline + wandb_name: airraid_pi05_profile_appo_h1_10env_2p5obs_g128_20260920 + wandb_job_type: profile_teacher + wandb_tags: + - air_raid + - APPO + - Pi05 + - measured-IID-profile + - H1 + - env10-obs2p5FPS diff --git a/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/verification.json b/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/verification.json new file mode 100644 index 0000000000000000000000000000000000000000..76f4607dd4a4d2cfa4c29ceb49f61e9d01b34015 --- /dev/null +++ b/latency-aware/air-raid/small-policy/sample-factory-qwenpi_v3-rtx3090-v1/verification.json @@ -0,0 +1,489 @@ +{ + "state": "TEACHER_FULL_TRACE_AUDIT_PASSED", + "verified_at": "2026-09-20T15:09:01.399190+00:00", + "run": "h1_10env_2p5obs_g128_20260920", + "env_fps": 10, + "obs_fps": 2.5, + "raw_frames_per_training_decision": 4, + "budget_counted_env_steps": 10000000, + "approximate_raw_game_frame_budget": 40000000, + "actual_env_steps": 10002432, + "seed": 0, + "profile_sha256": "7e81ed45da94173b447bdf0cf1ff59343cc7b4e1e6029bcf261cf054c2596cd6", + "source_profile_capture_fps": [ + 10, + 2.5 + ], + "profile_ms_rescaled": false, + "selection": { + "selected": "final", + "candidates": { + "final": { + "checkpoint": "${PI05_RUN_DIR}/profile_latency/teacher/checkpoints/airraid_pi05_profile_appo_h1_10env_2p5obs_g128_20260920/checkpoint_p0/checkpoint_000039064_10002432.pth", + "sha256": "59e7018b8aad3c0d7f52b40f9a5a4847c450b44f6cba27b66d052895869d4f10", + "episodes": 20, + "seeds": [ + 104771, + 104772, + 104773, + 104774, + 104775, + 104776, + 104777, + 104778, + 104779, + 104780, + 104781, + 104782, + 104783, + 104784, + 104785, + 104786, + 104787, + 104788, + 104789, + 104790 + ], + "returns": [ + 3775.0, + 3025.0, + 4000.0, + 3100.0, + 4050.0, + 4100.0, + 4050.0, + 3700.0, + 3925.0, + 4125.0, + 4100.0, + 4050.0, + 4125.0, + 4000.0, + 3975.0, + 4125.0, + 3525.0, + 4050.0, + 3975.0, + 4125.0 + ], + "mean": 3895.0, + "population_std": 316.68201717179966, + "strict_gt2500": 20 + }, + "training_best": { + "checkpoint": "${PI05_RUN_DIR}/profile_latency/teacher/checkpoints/airraid_pi05_profile_appo_h1_10env_2p5obs_g128_20260920/checkpoint_p0/best_000036640_9379840_reward_11643.500.pth", + "sha256": "ca875c7fce0848b8d62d2ed9c8a8654a198757fed984fa23b1940f503df725fa", + "episodes": 20, + "seeds": [ + 104771, + 104772, + 104773, + 104774, + 104775, + 104776, + 104777, + 104778, + 104779, + 104780, + 104781, + 104782, + 104783, + 104784, + 104785, + 104786, + 104787, + 104788, + 104789, + 104790 + ], + "returns": [ + 4100.0, + 3800.0, + 3050.0, + 4025.0, + 4150.0, + 4075.0, + 4150.0, + 2700.0, + 4075.0, + 3000.0, + 4150.0, + 3025.0, + 4125.0, + 3050.0, + 4125.0, + 4125.0, + 3850.0, + 3825.0, + 3925.0, + 3000.0 + ], + "mean": 3716.25, + "population_std": 503.5049031538819, + "strict_gt2500": 20 + } + }, + "completed_at": "2026-09-20T14:47:07.233886+00:00" + }, + "evaluations": { + "selection_final": { + "episodes": 20, + "seeds": [ + 104771, + 104772, + 104773, + 104774, + 104775, + 104776, + 104777, + 104778, + 104779, + 104780, + 104781, + 104782, + 104783, + 104784, + 104785, + 104786, + 104787, + 104788, + 104789, + 104790 + ], + "returns": [ + 3775.0, + 3025.0, + 4000.0, + 3100.0, + 4050.0, + 4100.0, + 4050.0, + 3700.0, + 3925.0, + 4125.0, + 4100.0, + 4050.0, + 4125.0, + 4000.0, + 3975.0, + 4125.0, + 3525.0, + 4050.0, + 3975.0, + 4125.0 + ], + "lengths": [ + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600 + ], + "mean_return": 3895.0, + "population_std_return": 316.68201717179966, + "mean_length": 3600.0, + "raw_steps": 72000, + "cap_episodes": 20, + "mean_of_episode_latency_ms": 134.5802285199416, + "action_weighted_mean_latency_ms": 134.58022851994164, + "action_latency_p95_ms": 139.88332387384128, + "actions": 18000, + "invalid_actions": 0, + "dropped_actions": 0, + "submitted_observations": 18000, + "dropped_observations": 0, + "observation_opportunities": 18000, + "observation_drop_fraction": 0.0, + "raw_rewards_lengths_and_admission_counters_match": true, + "raw_step_indices_and_clock_verified": true, + "mode": "profile_sample", + "config_sha256": "966bd1fcfc5bb414c430248f2f47fd36e5a0adcc8f86c6cd8d04a6dea9be9c2f", + "strict_gt2500": 20 + }, + "selection_training_best": { + "episodes": 20, + "seeds": [ + 104771, + 104772, + 104773, + 104774, + 104775, + 104776, + 104777, + 104778, + 104779, + 104780, + 104781, + 104782, + 104783, + 104784, + 104785, + 104786, + 104787, + 104788, + 104789, + 104790 + ], + "returns": [ + 4100.0, + 3800.0, + 3050.0, + 4025.0, + 4150.0, + 4075.0, + 4150.0, + 2700.0, + 4075.0, + 3000.0, + 4150.0, + 3025.0, + 4125.0, + 3050.0, + 4125.0, + 4125.0, + 3850.0, + 3825.0, + 3925.0, + 3000.0 + ], + "lengths": [ + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600 + ], + "mean_return": 3716.25, + "population_std_return": 503.5049031538819, + "mean_length": 3600.0, + "raw_steps": 72000, + "cap_episodes": 20, + "mean_of_episode_latency_ms": 134.5802285199416, + "action_weighted_mean_latency_ms": 134.58022851994164, + "action_latency_p95_ms": 139.88332387384128, + "actions": 18000, + "invalid_actions": 0, + "dropped_actions": 0, + "submitted_observations": 18000, + "dropped_observations": 0, + "observation_opportunities": 18000, + "observation_drop_fraction": 0.0, + "raw_rewards_lengths_and_admission_counters_match": true, + "raw_step_indices_and_clock_verified": true, + "mode": "profile_sample", + "config_sha256": "64fc179e9a31bd49d46e6d30b51e5cecd809f1166ed8523388e596ca6234d0de", + "strict_gt2500": 20 + }, + "E10": { + "episodes": 10, + "seeds": [ + 104771, + 104772, + 104773, + 104774, + 104775, + 104776, + 104777, + 104778, + 104779, + 104780 + ], + "returns": [ + 3775.0, + 3025.0, + 4000.0, + 3100.0, + 4050.0, + 4100.0, + 4050.0, + 3700.0, + 3925.0, + 4125.0 + ], + "lengths": [ + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600 + ], + "mean_return": 3785.0, + "population_std_return": 384.08983324217263, + "mean_length": 3600.0, + "raw_steps": 36000, + "cap_episodes": 10, + "mean_of_episode_latency_ms": 134.57468148712098, + "action_weighted_mean_latency_ms": 134.57468148712098, + "action_latency_p95_ms": 139.8895688877921, + "actions": 9000, + "invalid_actions": 0, + "dropped_actions": 0, + "submitted_observations": 9000, + "dropped_observations": 0, + "observation_opportunities": 9000, + "observation_drop_fraction": 0.0, + "raw_rewards_lengths_and_admission_counters_match": true, + "raw_step_indices_and_clock_verified": true, + "mode": "profile_sample", + "config_sha256": "4fc2f1de21e7005bc6002d41a3ba2655c9cb8a8e9d15e4371b3316d6fe167d55", + "strict_gt2500": 10 + } + }, + "probe": { + "source_state_sha256": "74b620e5aa7e0995d331d19a00821f64b0de5b8e64a4846e5e9a19ef636a5880", + "attempts": 12, + "accepted": 12, + "rejected": 0, + "accepted_specs": [ + { + "attempt_idx": 0, + "episode_idx": 0, + "episode_length": 900, + "episode_raw_frames": 3600, + "episode_raw_return": 4050.0, + "seed": 0, + "split": "train" + }, + { + "attempt_idx": 1, + "episode_idx": 1, + "episode_length": 900, + "episode_raw_frames": 3600, + "episode_raw_return": 4000.0, + "seed": 1, + "split": "val" + }, + { + "attempt_idx": 2, + "episode_idx": 2, + "episode_length": 900, + "episode_raw_frames": 3600, + "episode_raw_return": 3775.0, + "seed": 2, + "split": "train" + }, + { + "attempt_idx": 3, + "episode_idx": 3, + "episode_length": 900, + "episode_raw_frames": 3600, + "episode_raw_return": 3400.0, + "seed": 3, + "split": "val" + }, + { + "attempt_idx": 4, + "episode_idx": 4, + "episode_length": 900, + "episode_raw_frames": 3600, + "episode_raw_return": 3025.0, + "seed": 4, + "split": "train" + }, + { + "attempt_idx": 5, + "episode_idx": 5, + "episode_length": 900, + "episode_raw_frames": 3600, + "episode_raw_return": 3100.0, + "seed": 5, + "split": "train" + }, + { + "attempt_idx": 6, + "episode_idx": 6, + "episode_length": 900, + "episode_raw_frames": 3600, + "episode_raw_return": 3700.0, + "seed": 6, + "split": "train" + }, + { + "attempt_idx": 7, + "episode_idx": 7, + "episode_length": 900, + "episode_raw_frames": 3600, + "episode_raw_return": 4050.0, + "seed": 7, + "split": "val" + }, + { + "attempt_idx": 8, + "episode_idx": 8, + "episode_length": 900, + "episode_raw_frames": 3600, + "episode_raw_return": 4000.0, + "seed": 8, + "split": "train" + }, + { + "attempt_idx": 9, + "episode_idx": 9, + "episode_length": 900, + "episode_raw_frames": 3600, + "episode_raw_return": 3400.0, + "seed": 9, + "split": "train" + }, + { + "attempt_idx": 10, + "episode_idx": 10, + "episode_length": 900, + "episode_raw_frames": 3600, + "episode_raw_return": 4250.0, + "seed": 10, + "split": "train" + }, + { + "attempt_idx": 11, + "episode_idx": 11, + "episode_length": 900, + "episode_raw_frames": 3600, + "episode_raw_return": 4125.0, + "seed": 11, + "split": "train" + } + ] + }, + "strict_gate": ">2500", + "E10_seed_caveat": "E10 repeats the first10selection seeds, not independent evaluation", + "raw_budget_caveat": "10M Sample Factory counted steps cover about40M raw game frames atstride4; actual counted envsteps are readfromFinal checkpoint.", + "downstream": "After positive12attemptprobe,100accepted strict>2500 with no finiteattemptcap,actualreplayQA,freshPi05global128/micro64/acc1/GCfalse/ZeRO2/5000updates,andRTX3090F20; otherwise scientificstop", + "source_P_verification_sha256": "bc541ac289879bef1d50f006f13b530c1bb4870df600c9153891de301f784a38", + "gate": "PASSED" +} diff --git a/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/README.md b/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/README.md new file mode 100644 index 0000000000000000000000000000000000000000..ed1b8999e80835ec4484a8e47cde8ca2da34d34b --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/README.md @@ -0,0 +1,29 @@ +# air-raid / qwengr00t + +Training condition: `latency-aware`. Run: `standard-pipeline-2208875f92b2`. + +Copied from [Standard-Pipeline](https://huggingface.co/latency-sensitive-bench/Standard-Pipeline/tree/2208875f92b21e3f8ec2363a30a6705dfe56cc04/air_raid/profile_latency/GR00T) at `2208875f92b21e3f8ec2363a30a6705dfe56cc04`. + +- Checkpoint: `checkpoints/model.pt` +- Selection: only published VLA checkpoint +- Checkpoint SHA256: `530ee1a1b62ccdec8ef77a0d1c639eb8c16782dd94d91b5e5444c162da0e58b2` +- Configuration, FPS, actions, normalization and latency semantics retain their source values. +- This import does not run evaluations or change training/evaluation/acceptance status. + +Action horizon is 1. Use `config.yaml` and `dataset_statistics.json` with the matching +prepared backbone. `task_contract.json` is extracted from the original training configuration. +Original machine paths in historical configurations are provenance: bind checkpoint, +backbone, manifest and profile paths to downloaded local files when running inference. + +[Original provenance](source/provenance.json) retains dataset references and recorded experiment evidence. + +## Download + +```python +from huggingface_hub import snapshot_download +prefix = "latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2" +snapshot_download("latency-sensitive-bench/benchmark-models", revision="", + allow_patterns=[prefix + "/**"]) +``` + +Use the verified import commit listed in the repository's import report. diff --git a/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/checkpoints/model.pt b/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/checkpoints/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..c59c0f5851e0fed52205649255faa871ff78c7c1 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/checkpoints/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:530ee1a1b62ccdec8ef77a0d1c639eb8c16782dd94d91b5e5444c162da0e58b2 +size 9975248311 diff --git a/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/config.yaml b/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9d57c45a342ed4e663c52b8d135a129df91ab608 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/config.yaml @@ -0,0 +1,106 @@ +framework: + name: QwenGR00T + qwenvl: + base_vlm: ${oc.env:GR00T_BACKBONE_DIR} + attn_implementation: flash_attention_2 + vl_hidden_dim: 2048 + enable_gradient_checkpointing: true + action_model: + action_model_type: DiT-B + action_hidden_dim: 1024 + hidden_size: 1024 + add_pos_embed: true + max_seq_len: 1024 + action_dim: 6 + state_dim: 0 + task_objective: null + action_horizon: 1 + repeated_diffusion_steps: 8 + noise_beta_alpha: 1.5 + noise_beta_beta: 1.0 + noise_s: 0.999 + num_timestep_buckets: 1000 + num_inference_timesteps: 4 + num_target_vision_tokens: 32 + diffusion_model_cfg: + cross_attention_dim: 2560 + dropout: 0.2 + final_dropout: true + interleave_self_attention: true + norm_type: ada_norm + num_layers: 16 + output_dim: 1024 + positional_embeddings: null + action_env_dim: 6 + future_action_window_size: 0 + past_action_window_size: 0 +version_id: '0.21' +datasets: + vla_data: + include_state: false + obs_image_size: + - 224 + - 224 + image_mode: single + stitch_grid: + - 2 + - 2 + num_obs_frames: 1 + gymnasium_task_contract: &id001 + action_labels: + - noop + - fire + - right + - left + - rightfire + - leftfire + action_values: + - 0 + - 1 + - 2 + - 3 + - 4 + - 5 + base_prompt: 'Protect both buildings from flying saucers. Choose exactly one + action from: noop, fire, right, left, rightfire, leftfire.' + env_fps: 50.0 + env_id: LatencyBench/AirRaid-v0 + frame_stack: 4 + make_kwargs: + base_env_id: ALE/AirRaid-v5 + base_make_kwargs: + difficulty: 0 + frameskip: 1 + full_action_space: false + max_num_frames_per_episode: 108000 + mode: 1 + obs_type: rgb + repeat_action_probability: 0.0 + fire_reset: true + noop_max: 30 + render_mode: rgb_array + screen_size: 84 + noop_action_id: 0 + obs_fps: 12.5 + registration_imports: + - latency_bench.envs.gymnasium_air_raid + task_name: air_raid +trainer: + pretrained_checkpoint: null + profile_timing: + enabled: false + log_interval: 10 +backbone: + repo_id: Qwen/Qwen3-VL-4B-Instruct + revision: ebb281ec70b05090aa6165b016eac8ec08e71b17 +rl_games: + model_alias: gr00t + task: gymnasium + action_carrier: native + gymnasium: + task_contract: *id001 + env_eval: + image_size: 224 + frameskip: 1 + image_transform: raw_rgb + prompt_mode: raw diff --git a/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/dataset_statistics.json b/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/dataset_statistics.json new file mode 100644 index 0000000000000000000000000000000000000000..704702ca4bd3a6c2eb89f0023818d177cc13d419 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/dataset_statistics.json @@ -0,0 +1,84 @@ +{ + "new_embodiment": { + "action": { + "mean": [ + 0.31835803389549255, + 0.5185431838035583, + 0.016567900776863098, + 0.04858024790883064, + 0.042901232838630676, + 0.05504938215017319 + ], + "std": [ + 0.4660063683986664, + 0.4996560513973236, + 0.1275738626718521, + 0.2150501012802124, + 0.2027149498462677, + 0.22798019647598267 + ], + "max": [ + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0 + ], + "min": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "q01": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "q99": [ + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0 + ], + "mask": [ + true, + true, + true, + true, + true, + true + ] + }, + "state": { + "mean": [ + 0.0 + ], + "std": [ + 0.0 + ], + "max": [ + 0.0 + ], + "min": [ + 0.0 + ], + "q01": [ + 0.0 + ], + "q99": [ + 0.0 + ] + }, + "num_transitions": 81000, + "num_trajectories": 90 + } +} \ No newline at end of file diff --git a/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/latency_prompt_map.json b/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/latency_prompt_map.json new file mode 100644 index 0000000000000000000000000000000000000000..5475933f26f7f6cda61949ed9e39e7f023cc0278 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/latency_prompt_map.json @@ -0,0 +1,17 @@ +{ + "4": { + "prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire. Current action latency is 4 raw frames (80.00 ms). The environment runs at 50 FPS and observations are emitted at 12.5 FPS. Choose the best next action.", + "latency_raw_frames": 4, + "latency_ms": 80.0 + }, + "5": { + "prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire. Current action latency is 5 raw frames (100.00 ms). The environment runs at 50 FPS and observations are emitted at 12.5 FPS. Choose the best next action.", + "latency_raw_frames": 5, + "latency_ms": 100.0 + }, + "6": { + "prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire. Current action latency is 6 raw frames (120.00 ms). The environment runs at 50 FPS and observations are emitted at 12.5 FPS. Choose the best next action.", + "latency_raw_frames": 6, + "latency_ms": 120.0 + } +} \ No newline at end of file diff --git a/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/manifest.json b/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..8b8438ff6ece7f5d7d2b6081d0b236cc622193cb --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/manifest.json @@ -0,0 +1,98 @@ +{ + "dataset_name": "air_raid_gr00t_profile_iid", + "env_name": "air_raid", + "episodes": 90, + "frames": 81000, + "task_prompts": [ + "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire. Current action latency is 4 raw frames (80.00 ms). The environment runs at 50 FPS and observations are emitted at 12.5 FPS. Choose the best next action.", + "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire. Current action latency is 5 raw frames (100.00 ms). The environment runs at 50 FPS and observations are emitted at 12.5 FPS. Choose the best next action.", + "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire. Current action latency is 6 raw frames (120.00 ms). The environment runs at 50 FPS and observations are emitted at 12.5 FPS. Choose the best next action." + ], + "format": "starvla_lerobot_v2_image_parquet", + "integration_name": "gymnasium", + "task_name": "air_raid", + "action_layout": "gymnasium_discrete_v1", + "action_labels": [ + "noop", + "fire", + "right", + "left", + "rightfire", + "leftfire" + ], + "carrier_action_labels": [ + "noop", + "fire", + "right", + "left", + "rightfire", + "leftfire" + ], + "action_dim": 6, + "active_action_dim": 6, + "action_carrier": "native", + "rows_unit": "decision_step", + "obs_fps": 12.5, + "obs_stride_raw_frames": 4, + "uses_state": false, + "state_dim": 1, + "state_labels": [ + "state" + ], + "state_normalization": null, + "robot_type": "rl_games_gymnasium", + "gymnasium_task": { + "action_labels": [ + "noop", + "fire", + "right", + "left", + "rightfire", + "leftfire" + ], + "action_values": [ + 0, + 1, + 2, + 3, + 4, + 5 + ], + "base_prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.", + "env_fps": 50.0, + "env_id": "LatencyBench/AirRaid-v0", + "frame_stack": 4, + "make_kwargs": { + "base_env_id": "ALE/AirRaid-v5", + "base_make_kwargs": { + "difficulty": 0, + "frameskip": 1, + "full_action_space": false, + "max_num_frames_per_episode": 108000, + "mode": 1, + "obs_type": "rgb", + "repeat_action_probability": 0.0 + }, + "fire_reset": true, + "noop_max": 30, + "render_mode": "rgb_array", + "screen_size": 84 + }, + "noop_action_id": 0, + "obs_fps": 12.5, + "registration_imports": [ + "latency_bench.envs.gymnasium_air_raid" + ], + "task_name": "air_raid" + }, + "validation_dataset_name": "air_raid_gr00t_profile_iid__val", + "validation_episodes": 10, + "validation_frames": 9000, + "raw_dataset": { + "repo": "latency-sensitive-bench/Standard-Pipeline", + "repo_type": "dataset", + "prefix": "air_raid/profile_latency/GR00T/gr00t_h1_iid_20260917/demonstrations/raw", + "revision": "78822d8d37db6318d6101d43844ea5190b84a94f", + "reused": false + } +} diff --git a/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/provenance.json b/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/provenance.json new file mode 100644 index 0000000000000000000000000000000000000000..7589f55a77d5949925a269927dfa1ecd432594d5 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/provenance.json @@ -0,0 +1,34 @@ +{ + "task": "air-raid", + "model": "qwengr00t", + "training_condition": "latency-aware", + "training_run_id": "standard-pipeline-2208875f92b2", + "source": { + "repo_id": "latency-sensitive-bench/Standard-Pipeline", + "revision": "2208875f92b21e3f8ec2363a30a6705dfe56cc04", + "path": "air_raid/profile_latency/GR00T", + "url": "https://huggingface.co/latency-sensitive-bench/Standard-Pipeline/tree/2208875f92b21e3f8ec2363a30a6705dfe56cc04/air_raid/profile_latency/GR00T" + }, + "hf_repo_id": "latency-sensitive-bench/benchmark-models", + "hf_path": "latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2", + "checkpoint": { + "source_file": "air_raid/profile_latency/GR00T/checkpoints/model.pt", + "source_sha256": "530ee1a1b62ccdec8ef77a0d1c639eb8c16782dd94d91b5e5444c162da0e58b2", + "source_bytes": 9975248311, + "file": "checkpoints/model.pt", + "selection_rule": "only published VLA checkpoint", + "method": "copy", + "sha256": "530ee1a1b62ccdec8ef77a0d1c639eb8c16782dd94d91b5e5444c162da0e58b2", + "bytes": 9975248311 + }, + "config_source": "air_raid/profile_latency/GR00T/config.yaml", + "task_contract_file": "task_contract.json", + "missing_source_files": [], + "migration": { + "operation": "copy_selected_inference_bundle", + "source_modified": false, + "evaluation_performed": false, + "experiment_status": "preserved_in_source_evidence" + }, + "action_horizon": 1 +} diff --git a/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/reload-validation.json b/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/reload-validation.json new file mode 100644 index 0000000000000000000000000000000000000000..9844bba8a974a0a24e3971ee442374598a005894 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/reload-validation.json @@ -0,0 +1,14 @@ +{ + "status": "passed", + "strict_load": true, + "output_shape": [ + 1, + 1, + 6 + ], + "finite": true, + "checkpoint_sha256": "530ee1a1b62ccdec8ef77a0d1c639eb8c16782dd94d91b5e5444c162da0e58b2", + "config_sha256": "ff7ffb921bce669e1375145d5c8a2e89eb5b2c6cdeff3e1b268988fec172e0d3", + "training_updates": 5000, + "check_scope": "checkpoint reload and one real validation-sample forward; not closed-loop quality acceptance" +} diff --git a/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/source/provenance.json b/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/source/provenance.json new file mode 100644 index 0000000000000000000000000000000000000000..1bfdc7106a84cc1acb084b9a7d56443e1adfb65c --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/source/provenance.json @@ -0,0 +1,39 @@ +{ + "task": "air_raid", + "condition": "profile_latency", + "training_updates": 5000, + "seed": 42, + "global_batch": 64, + "micro_batch": 16, + "gradient_accumulation_steps": 4, + "action_horizon": 1, + "action_dim": 6, + "state_dim": 0, + "image_size": [ + 224, + 224 + ], + "env_fps": 50, + "obs_fps": 12.5, + "latency_training_method": "iid", + "profile_sha256": "a6ded702eef5495736922babdd58e24901526811613e45c680a2b77eecfdc149", + "checkpoint_sha256": "530ee1a1b62ccdec8ef77a0d1c639eb8c16782dd94d91b5e5444c162da0e58b2", + "training_config_sha256": "bf8d1b05c1da4f0bdbab8e38c186dbf94fc52c2b0b2fe2535f4411bd86ad520c", + "backbone": { + "repo_id": "Qwen/Qwen3-VL-4B-Instruct", + "revision": "ebb281ec70b05090aa6165b016eac8ec08e71b17" + }, + "source": { + "version": "v10", + "base_commit": "1681edb8d9e4f7d10b7dfa647670e01e0d5cce37", + "starvla_commit": "1d0d7b139d1725cab268cb9dae2007ef4c5d05d2", + "sample_factory_commit": "4b11d8754ffc989cdd15a340cfc9be0277377d5b", + "source_bundle_sha256": "6e428b56eac76f13781113d053336020be600c19a8160ace1b3206ffaed8b4ea" + }, + "training_started_at": "2026-09-17T03:51:49Z", + "training_completed_at": "2026-09-17T08:08:37Z", + "training_exit_code": 0, + "evaluation_prompt_key": 4, + "realtime_evaluation_status": "not_started", + "quality_status": "not_yet_evaluated" +} diff --git a/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/task_contract.json b/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/task_contract.json new file mode 100644 index 0000000000000000000000000000000000000000..f20338db89c82f965cce12d795b3641d45cdae58 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwengr00t-h1/standard-pipeline-2208875f92b2/task_contract.json @@ -0,0 +1,44 @@ +{ + "action_labels": [ + "noop", + "fire", + "right", + "left", + "rightfire", + "leftfire" + ], + "action_values": [ + 0, + 1, + 2, + 3, + 4, + 5 + ], + "base_prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.", + "env_fps": 50.0, + "env_id": "LatencyBench/AirRaid-v0", + "frame_stack": 4, + "make_kwargs": { + "base_env_id": "ALE/AirRaid-v5", + "base_make_kwargs": { + "difficulty": 0, + "frameskip": 1, + "full_action_space": false, + "max_num_frames_per_episode": 108000, + "mode": 1, + "obs_type": "rgb", + "repeat_action_probability": 0.0 + }, + "fire_reset": true, + "noop_max": 30, + "render_mode": "rgb_array", + "screen_size": 84 + }, + "noop_action_id": 0, + "obs_fps": 12.5, + "registration_imports": [ + "latency_bench.envs.gymnasium_air_raid" + ], + "task_name": "air_raid" +} diff --git a/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/README.md b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/README.md new file mode 100644 index 0000000000000000000000000000000000000000..a447049854c6a76b44227d5e436a3f6745b45fe0 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/README.md @@ -0,0 +1,29 @@ +# air-raid / qwenoft + +Training condition: `latency-aware`. Run: `air_raid_profile_20260911T054920Z_openvla_native_sft_5k`. + +Copied from [Standard-Pipeline](https://huggingface.co/latency-sensitive-bench/Standard-Pipeline/tree/2208875f92b21e3f8ec2363a30a6705dfe56cc04/air_raid/profile_latency/OpenVLA) at `2208875f92b21e3f8ec2363a30a6705dfe56cc04`. + +- Checkpoint: `checkpoints/model.pt` +- Selection: only published VLA checkpoint +- Checkpoint SHA256: `ded1152ea155cf86eff73646e986f5b764007c32c82753f36238a17cf4bad734` +- Configuration, FPS, actions, normalization and latency semantics retain their source values. +- This import does not run evaluations or change training/evaluation/acceptance status. + +Action horizon is 1. Use `config.yaml` and `dataset_statistics.json` with the matching +prepared backbone. `task_contract.json` is extracted from the original training configuration. +Original machine paths in historical configurations are provenance: bind checkpoint, +backbone, manifest and profile paths to downloaded local files when running inference. + +[Original provenance](source/provenance.json) retains dataset references and recorded experiment evidence. + +## Download + +```python +from huggingface_hub import snapshot_download +prefix = "latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k" +snapshot_download("latency-sensitive-bench/benchmark-models", revision="", + allow_patterns=[prefix + "/**"]) +``` + +Use the verified import commit listed in the repository's import report. diff --git a/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/checkpoints/model.pt b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/checkpoints/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..1e8be07a087737f6b63608216965fd1e5c58dac0 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/checkpoints/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:ded1152ea155cf86eff73646e986f5b764007c32c82753f36238a17cf4bad734 +size 9785049835 diff --git a/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/config.full.yaml b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/config.full.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e83ac631ab18387555309bc1ae2b3110c17971a0 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/config.full.yaml @@ -0,0 +1,332 @@ +framework: + name: QwenOFT + qwenvl: + base_vlm: /mnt/checkpoints/latency-sensitive-bench/openvla/Qwen3-VL-4B-Instruct + attn_implementation: flash_attention_2 + enable_gradient_checkpointing: true + action_model: + action_model_type: MLP + action_dim: 6 + state_dim: 1 + action_hidden_dim: 2560 + future_action_window_size: 0 + past_action_window_size: 0 + loss_type: discrete_ce + state_encoding: discretized_text + task_objective: null + action_horizon: 1 + action_env_dim: 6 + kv_memory: + enabled: false + window: 4 + rollout_len: 8 + packed_train: false + rebased_sink: true +datasets: + vla_data: + dataset_py: lerobot_datasets + include_state: false + data_root_dir: /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/converted + data_mix: air_raid_profile_20260911T054920Z + eval_data_mix: air_raid_profile_20260911T054920Z__val + custom_mixtures_path: /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/converted/_generated_mixtures/air_raid_profile_20260911T054920Z.json + action_type: discrete + sequential_step_sampling: false + eval_sequential_step_sampling: null + num_workers: 8 + eval_num_workers: 8 + prefetch_factor: 4 + persistent_workers: true + pin_memory: true + shuffle: true + action_balance: + enabled: false + strategy: balanced_epoch + action_key: action_id + target_flap_fraction: 0.3 + noop_id: 0 + flap_id: 1 + latency_curriculum: + enabled: false + strategy: exclusive + latencies: null + phase_steps: null + phase_distributions: null + new_latency_passes: 1.0 + replay_passes: 0.25 + target_total_passes: 2.0 + final_equalization: true + step_budget_mode: auto + eval_at_phase_end: false + save_at_phase_end: false + computed_plan: null + per_device_batch_size: 16 + load_all_data_for_training: true + num_obs_frames: 1 + image_mode: single + prompt_mode: raw + stitch_grid: + - 2 + - 2 + obs_image_size: null + video_backend: torchvision_av + gymnasium_task_contract: + action_labels: + - noop + - fire + - right + - left + - rightfire + - leftfire + action_values: + - 0 + - 1 + - 2 + - 3 + - 4 + - 5 + base_prompt: 'Protect both buildings from flying saucers. Choose exactly one + action from: noop, fire, right, left, rightfire, leftfire.' + env_fps: 50.0 + env_id: LatencyBench/AirRaid-v0 + frame_stack: 4 + make_kwargs: + base_env_id: ALE/AirRaid-v5 + base_make_kwargs: + difficulty: 0 + frameskip: 1 + full_action_space: false + max_num_frames_per_episode: 108000 + mode: 1 + obs_type: rgb + repeat_action_probability: 0.0 + fire_reset: true + noop_max: 30 + render_mode: rgb_array + screen_size: 84 + noop_action_id: 0 + obs_fps: 12.5 + registration_imports: + - latency_bench.envs.gymnasium_air_raid + task_name: air_raid + active_action_dim: 6 +dataset: + source_hf: '' + config_name: null + source_subdir: null + converted_name: air_raid_profile_20260911T054920Z + single_source_hf: '' + mixed_source_hf: '' + single_converted_name: air_raid_profile_20260911T054920Z + mixed_converted_name: air_raid_profile_20260911T054920Z + single_latency_filter: null + mixed_latency_filter: null + force_download: false + setup_force: false + skip_verification: false + target_latency_unit: observation_steps + verify_rows: 200 + max_episodes: null + episodes_per_latency: null + latency_filter: null + debug_subset: + enabled: false + max_episodes: 5 + suffix: debug +base_model: + repo_id: Qwen/Qwen3-VL-4B-Instruct +initialization: + checkpoint_local_dir: null + checkpoint_hf_repo_id: null + checkpoint_filename: null +trainer: + max_train_steps: 5000 + num_warmup_steps: 100 + save_interval: 5000 + eval_interval: 500 + eval_num_batches: 200 + per_latency_eval_num_batches: null + eval_action_classification: true + eval_action_classification_interval: null + cc_f1_tolerance: 1 + learning_rate: + base: 2.0e-05 + qwen_vl_interface: 1.0e-05 + action_model: 0.0001 + lr_scheduler_type: cosine_with_min_lr + scheduler_specific_kwargs: + min_lr: 1.0e-06 + freeze_modules: '' + freeze_vit: false + freeze_tied_embedding: false + freeze_llm_layers: [] + loss_scale: + vla: 1.0 + vlm: 0.1 + max_grad_norm: 1.0 + weight_decay: 0.0 + logging_frequency: 1 + profile_timing: + enabled: true + log_interval: 10 + gradient_clipping: 1.0 + gradient_accumulation_steps: 1 + distributed_backend: deepspeed + is_resume: false + pretrained_checkpoint: null + resume_step: 0 + reload_modules: null + optimizer: + name: AdamW + betas: + - 0.9 + - 0.95 + eps: 1.0e-08 + weight_decay: 1.0e-08 + fused: true + save_format: pt +workspace_dir: /home/lzj/code/latency-sensitive-bench +run_root_dir: /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/vla +seed: 42 +wandb_entity: dongqianyu99-zhejiang-university +wandb_project: latency-sensitive-bench +auth: + env_file: null + hf_token_env: HF_TOKEN + wandb_api_key_env: WANDB_API_KEY +paths: + run_root_dir: /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/vla + dataset_local_dir: /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/converted + dataset_cache_dir: null + base_model_dir: /mnt/checkpoints/latency-sensitive-bench/openvla/Qwen3-VL-4B-Instruct + accelerate_config: starVLA/config/deepseeds/deepspeed_zero2_cpu_offload.yaml +launch: + use_accelerate: true + gpus: null + num_processes: 1 + dry_run: false +conda: + enabled: true + env_name: null +rl_games: + model_alias: openvla + env_eval: + image_size: 224 + frameskip: 1 + image_transform: raw_rgb + prompt_mode: raw + ghost_trail: + history_frames: 5 + gamma: 1.3 + min_alpha: 35 + scroll_px_per_step: 4.0 + ground_fraction: 0.22 + seed: 42 + fixed_episode_seeds: true + latency_seed_stride: 0 + task_seed_stride: 0 + task_description: '' + eval_parallel_envs: 5 + action_chunk_execution: + enabled: false + chunk_size: null + enabled: false + gymnasium: + make_kwargs: {} + mid_train: + enabled: false + interval_steps: 100 + latencies: + - 0 + num_episodes: 5 + max_steps_per_episode: 3600 + post_train: + enabled: false + latencies: + - 0 + - 1 + - 2 + - 3 + - 4 + num_episodes: 5 + max_steps_per_episode: 3600 + eval_backend: latency_bench + distributed_mode: none + vectorized: + enabled: false + batch_size: 1 + latency: + prompt_map_path: /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/converted/air_raid_profile_20260911T054920Z/latency_prompt_map.json + mode: single + values: + - 3 + - 4 + task: gymnasium + gymnasium: + task_contract: + action_labels: + - noop + - fire + - right + - left + - rightfire + - leftfire + action_values: + - 0 + - 1 + - 2 + - 3 + - 4 + - 5 + base_prompt: 'Protect both buildings from flying saucers. Choose exactly one + action from: noop, fire, right, left, rightfire, leftfire.' + env_fps: 50.0 + env_id: LatencyBench/AirRaid-v0 + frame_stack: 4 + make_kwargs: + base_env_id: ALE/AirRaid-v5 + base_make_kwargs: + difficulty: 0 + frameskip: 1 + full_action_space: false + max_num_frames_per_episode: 108000 + mode: 1 + obs_type: rgb + repeat_action_probability: 0.0 + fire_reset: true + noop_max: 30 + render_mode: rgb_array + screen_size: 84 + noop_action_id: 0 + obs_fps: 12.5 + registration_imports: + - latency_bench.envs.gymnasium_air_raid + task_name: air_raid + task_name: air_raid + initialization_mode: scratch + action_carrier: native +model: openvla +env: gymnasium +init: scratch +mode: single +checkpoint: + load: none + hf_repo_id: null + save_best_model: false + save_final_model: true + save_pt_file: true + save_training_state: false + save_safetensors_file: false + local: + keep_last_n: 1 + sync: + enabled: false + repo_id: null + keep_last_n: 0 + sync_every_n_checkpoints: 1 + resume_policy: local_latest +run_id: air_raid_profile_20260911T054920Z_openvla_native_sft_5k +output_dir: /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/vla/air_raid_profile_20260911T054920Z_openvla_native_sft_5k +config_yaml: /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/h100/vla_train.yaml +is_debug: false +version_id: '0.21' diff --git a/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/config.yaml b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..e83ac631ab18387555309bc1ae2b3110c17971a0 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/config.yaml @@ -0,0 +1,332 @@ +framework: + name: QwenOFT + qwenvl: + base_vlm: /mnt/checkpoints/latency-sensitive-bench/openvla/Qwen3-VL-4B-Instruct + attn_implementation: flash_attention_2 + enable_gradient_checkpointing: true + action_model: + action_model_type: MLP + action_dim: 6 + state_dim: 1 + action_hidden_dim: 2560 + future_action_window_size: 0 + past_action_window_size: 0 + loss_type: discrete_ce + state_encoding: discretized_text + task_objective: null + action_horizon: 1 + action_env_dim: 6 + kv_memory: + enabled: false + window: 4 + rollout_len: 8 + packed_train: false + rebased_sink: true +datasets: + vla_data: + dataset_py: lerobot_datasets + include_state: false + data_root_dir: /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/converted + data_mix: air_raid_profile_20260911T054920Z + eval_data_mix: air_raid_profile_20260911T054920Z__val + custom_mixtures_path: /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/converted/_generated_mixtures/air_raid_profile_20260911T054920Z.json + action_type: discrete + sequential_step_sampling: false + eval_sequential_step_sampling: null + num_workers: 8 + eval_num_workers: 8 + prefetch_factor: 4 + persistent_workers: true + pin_memory: true + shuffle: true + action_balance: + enabled: false + strategy: balanced_epoch + action_key: action_id + target_flap_fraction: 0.3 + noop_id: 0 + flap_id: 1 + latency_curriculum: + enabled: false + strategy: exclusive + latencies: null + phase_steps: null + phase_distributions: null + new_latency_passes: 1.0 + replay_passes: 0.25 + target_total_passes: 2.0 + final_equalization: true + step_budget_mode: auto + eval_at_phase_end: false + save_at_phase_end: false + computed_plan: null + per_device_batch_size: 16 + load_all_data_for_training: true + num_obs_frames: 1 + image_mode: single + prompt_mode: raw + stitch_grid: + - 2 + - 2 + obs_image_size: null + video_backend: torchvision_av + gymnasium_task_contract: + action_labels: + - noop + - fire + - right + - left + - rightfire + - leftfire + action_values: + - 0 + - 1 + - 2 + - 3 + - 4 + - 5 + base_prompt: 'Protect both buildings from flying saucers. Choose exactly one + action from: noop, fire, right, left, rightfire, leftfire.' + env_fps: 50.0 + env_id: LatencyBench/AirRaid-v0 + frame_stack: 4 + make_kwargs: + base_env_id: ALE/AirRaid-v5 + base_make_kwargs: + difficulty: 0 + frameskip: 1 + full_action_space: false + max_num_frames_per_episode: 108000 + mode: 1 + obs_type: rgb + repeat_action_probability: 0.0 + fire_reset: true + noop_max: 30 + render_mode: rgb_array + screen_size: 84 + noop_action_id: 0 + obs_fps: 12.5 + registration_imports: + - latency_bench.envs.gymnasium_air_raid + task_name: air_raid + active_action_dim: 6 +dataset: + source_hf: '' + config_name: null + source_subdir: null + converted_name: air_raid_profile_20260911T054920Z + single_source_hf: '' + mixed_source_hf: '' + single_converted_name: air_raid_profile_20260911T054920Z + mixed_converted_name: air_raid_profile_20260911T054920Z + single_latency_filter: null + mixed_latency_filter: null + force_download: false + setup_force: false + skip_verification: false + target_latency_unit: observation_steps + verify_rows: 200 + max_episodes: null + episodes_per_latency: null + latency_filter: null + debug_subset: + enabled: false + max_episodes: 5 + suffix: debug +base_model: + repo_id: Qwen/Qwen3-VL-4B-Instruct +initialization: + checkpoint_local_dir: null + checkpoint_hf_repo_id: null + checkpoint_filename: null +trainer: + max_train_steps: 5000 + num_warmup_steps: 100 + save_interval: 5000 + eval_interval: 500 + eval_num_batches: 200 + per_latency_eval_num_batches: null + eval_action_classification: true + eval_action_classification_interval: null + cc_f1_tolerance: 1 + learning_rate: + base: 2.0e-05 + qwen_vl_interface: 1.0e-05 + action_model: 0.0001 + lr_scheduler_type: cosine_with_min_lr + scheduler_specific_kwargs: + min_lr: 1.0e-06 + freeze_modules: '' + freeze_vit: false + freeze_tied_embedding: false + freeze_llm_layers: [] + loss_scale: + vla: 1.0 + vlm: 0.1 + max_grad_norm: 1.0 + weight_decay: 0.0 + logging_frequency: 1 + profile_timing: + enabled: true + log_interval: 10 + gradient_clipping: 1.0 + gradient_accumulation_steps: 1 + distributed_backend: deepspeed + is_resume: false + pretrained_checkpoint: null + resume_step: 0 + reload_modules: null + optimizer: + name: AdamW + betas: + - 0.9 + - 0.95 + eps: 1.0e-08 + weight_decay: 1.0e-08 + fused: true + save_format: pt +workspace_dir: /home/lzj/code/latency-sensitive-bench +run_root_dir: /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/vla +seed: 42 +wandb_entity: dongqianyu99-zhejiang-university +wandb_project: latency-sensitive-bench +auth: + env_file: null + hf_token_env: HF_TOKEN + wandb_api_key_env: WANDB_API_KEY +paths: + run_root_dir: /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/vla + dataset_local_dir: /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/converted + dataset_cache_dir: null + base_model_dir: /mnt/checkpoints/latency-sensitive-bench/openvla/Qwen3-VL-4B-Instruct + accelerate_config: starVLA/config/deepseeds/deepspeed_zero2_cpu_offload.yaml +launch: + use_accelerate: true + gpus: null + num_processes: 1 + dry_run: false +conda: + enabled: true + env_name: null +rl_games: + model_alias: openvla + env_eval: + image_size: 224 + frameskip: 1 + image_transform: raw_rgb + prompt_mode: raw + ghost_trail: + history_frames: 5 + gamma: 1.3 + min_alpha: 35 + scroll_px_per_step: 4.0 + ground_fraction: 0.22 + seed: 42 + fixed_episode_seeds: true + latency_seed_stride: 0 + task_seed_stride: 0 + task_description: '' + eval_parallel_envs: 5 + action_chunk_execution: + enabled: false + chunk_size: null + enabled: false + gymnasium: + make_kwargs: {} + mid_train: + enabled: false + interval_steps: 100 + latencies: + - 0 + num_episodes: 5 + max_steps_per_episode: 3600 + post_train: + enabled: false + latencies: + - 0 + - 1 + - 2 + - 3 + - 4 + num_episodes: 5 + max_steps_per_episode: 3600 + eval_backend: latency_bench + distributed_mode: none + vectorized: + enabled: false + batch_size: 1 + latency: + prompt_map_path: /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/converted/air_raid_profile_20260911T054920Z/latency_prompt_map.json + mode: single + values: + - 3 + - 4 + task: gymnasium + gymnasium: + task_contract: + action_labels: + - noop + - fire + - right + - left + - rightfire + - leftfire + action_values: + - 0 + - 1 + - 2 + - 3 + - 4 + - 5 + base_prompt: 'Protect both buildings from flying saucers. Choose exactly one + action from: noop, fire, right, left, rightfire, leftfire.' + env_fps: 50.0 + env_id: LatencyBench/AirRaid-v0 + frame_stack: 4 + make_kwargs: + base_env_id: ALE/AirRaid-v5 + base_make_kwargs: + difficulty: 0 + frameskip: 1 + full_action_space: false + max_num_frames_per_episode: 108000 + mode: 1 + obs_type: rgb + repeat_action_probability: 0.0 + fire_reset: true + noop_max: 30 + render_mode: rgb_array + screen_size: 84 + noop_action_id: 0 + obs_fps: 12.5 + registration_imports: + - latency_bench.envs.gymnasium_air_raid + task_name: air_raid + task_name: air_raid + initialization_mode: scratch + action_carrier: native +model: openvla +env: gymnasium +init: scratch +mode: single +checkpoint: + load: none + hf_repo_id: null + save_best_model: false + save_final_model: true + save_pt_file: true + save_training_state: false + save_safetensors_file: false + local: + keep_last_n: 1 + sync: + enabled: false + repo_id: null + keep_last_n: 0 + sync_every_n_checkpoints: 1 + resume_policy: local_latest +run_id: air_raid_profile_20260911T054920Z_openvla_native_sft_5k +output_dir: /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/vla/air_raid_profile_20260911T054920Z_openvla_native_sft_5k +config_yaml: /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/h100/vla_train.yaml +is_debug: false +version_id: '0.21' diff --git a/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/dataset_statistics.json b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/dataset_statistics.json new file mode 100644 index 0000000000000000000000000000000000000000..fc7df56134638aa55f835846a2b4e79dc3b25a08 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/dataset_statistics.json @@ -0,0 +1,84 @@ +{ + "new_embodiment": { + "action": { + "mean": [ + 0.19612345099449158, + 0.4346790015697479, + 0.14129629731178284, + 0.13423456251621246, + 0.017271604388952255, + 0.07639506459236145 + ], + "std": [ + 0.396988183259964, + 0.4956902265548706, + 0.34827736020088196, + 0.34088006615638733, + 0.13024327158927917, + 0.26566314697265625 + ], + "max": [ + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0 + ], + "min": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "q01": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "q99": [ + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0 + ], + "mask": [ + true, + true, + true, + true, + true, + true + ] + }, + "state": { + "mean": [ + 0.0 + ], + "std": [ + 0.0 + ], + "max": [ + 0.0 + ], + "min": [ + 0.0 + ], + "q01": [ + 0.0 + ], + "q99": [ + 0.0 + ] + }, + "num_transitions": 81000, + "num_trajectories": 90 + } +} \ No newline at end of file diff --git a/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/dataset_statistics_eval.json b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/dataset_statistics_eval.json new file mode 100644 index 0000000000000000000000000000000000000000..ce16e6476a85a96c6efb838d9d17b65737ba22d6 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/dataset_statistics_eval.json @@ -0,0 +1,84 @@ +{ + "new_embodiment": { + "action": { + "mean": [ + 0.19233334064483643, + 0.4481111168861389, + 0.13733333349227905, + 0.1344444453716278, + 0.01644444465637207, + 0.07133333384990692 + ], + "std": [ + 0.39413151144981384, + 0.4972960352897644, + 0.3442021906375885, + 0.34112292528152466, + 0.12717580795288086, + 0.2573772668838501 + ], + "max": [ + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0 + ], + "min": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "q01": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "q99": [ + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0 + ], + "mask": [ + true, + true, + true, + true, + true, + true + ] + }, + "state": { + "mean": [ + 0.0 + ], + "std": [ + 0.0 + ], + "max": [ + 0.0 + ], + "min": [ + 0.0 + ], + "q01": [ + 0.0 + ], + "q99": [ + 0.0 + ] + }, + "num_transitions": 9000, + "num_trajectories": 10 + } +} \ No newline at end of file diff --git a/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/latency_prompt_map.json b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/latency_prompt_map.json new file mode 100644 index 0000000000000000000000000000000000000000..016fd28ba53f4602f0e08916525487b378b24ff4 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/latency_prompt_map.json @@ -0,0 +1,12 @@ +{ + "3": { + "prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire. Current action latency is 3 raw frames (60.00 ms). The environment runs at 50 FPS and observations are emitted at 12.5 FPS. Choose the best next action.", + "latency_raw_frames": 3, + "latency_ms": 60.0 + }, + "4": { + "prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire. Current action latency is 4 raw frames (80.00 ms). The environment runs at 50 FPS and observations are emitted at 12.5 FPS. Choose the best next action.", + "latency_raw_frames": 4, + "latency_ms": 80.0 + } +} \ No newline at end of file diff --git a/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/manifest.json b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..bb1b1aecf28e62a5459b3fef2278321b9e78ee71 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/manifest.json @@ -0,0 +1,93 @@ +{ + "dataset_name": "air_raid_profile_20260911T054920Z", + "env_name": "air_raid", + "episodes": 90, + "frames": 81000, + "task_prompts": [ + "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire. Current action latency is 3 raw frames (60.00 ms). The environment runs at 50 FPS and observations are emitted at 12.5 FPS. Choose the best next action.", + "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire. Current action latency is 4 raw frames (80.00 ms). The environment runs at 50 FPS and observations are emitted at 12.5 FPS. Choose the best next action." + ], + "format": "starvla_lerobot_v2_image_parquet", + "source": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/raw", + "integration_name": "gymnasium", + "task_name": "air_raid", + "action_layout": "gymnasium_discrete_v1", + "action_labels": [ + "noop", + "fire", + "right", + "left", + "rightfire", + "leftfire" + ], + "carrier_action_labels": [ + "noop", + "fire", + "right", + "left", + "rightfire", + "leftfire" + ], + "action_dim": 6, + "active_action_dim": 6, + "action_carrier": "native", + "rows_unit": "decision_step", + "obs_fps": 12.5, + "obs_stride_raw_frames": 4, + "uses_state": false, + "state_dim": 1, + "state_labels": [ + "state" + ], + "state_normalization": null, + "robot_type": "rl_games_gymnasium", + "latency_prompt_map_path": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/converted/air_raid_profile_20260911T054920Z/latency_prompt_map.json", + "custom_mixtures_path": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/converted/_generated_mixtures/air_raid_profile_20260911T054920Z.json", + "gymnasium_task": { + "action_labels": [ + "noop", + "fire", + "right", + "left", + "rightfire", + "leftfire" + ], + "action_values": [ + 0, + 1, + 2, + 3, + 4, + 5 + ], + "base_prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.", + "env_fps": 50.0, + "env_id": "LatencyBench/AirRaid-v0", + "frame_stack": 4, + "make_kwargs": { + "base_env_id": "ALE/AirRaid-v5", + "base_make_kwargs": { + "difficulty": 0, + "frameskip": 1, + "full_action_space": false, + "max_num_frames_per_episode": 108000, + "mode": 1, + "obs_type": "rgb", + "repeat_action_probability": 0.0 + }, + "fire_reset": true, + "noop_max": 30, + "render_mode": "rgb_array", + "screen_size": 84 + }, + "noop_action_id": 0, + "obs_fps": 12.5, + "registration_imports": [ + "latency_bench.envs.gymnasium_air_raid" + ], + "task_name": "air_raid" + }, + "validation_dataset_name": "air_raid_profile_20260911T054920Z__val", + "validation_episodes": 10, + "validation_frames": 9000 +} \ No newline at end of file diff --git a/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/profile/DONE b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/profile/DONE new file mode 100644 index 0000000000000000000000000000000000000000..e69de29bb2d1d6434b8b29ae775ad8c2e48c5391 diff --git a/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/profile/hardware.json b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/profile/hardware.json new file mode 100644 index 0000000000000000000000000000000000000000..e8204e81807044975c0e0a0953d16d6835192f31 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/profile/hardware.json @@ -0,0 +1,40 @@ +{ + "driver_version": "580.173.02", + "gpu_class": "1x-rtx3090", + "gpus": [ + { + "name": "NVIDIA GeForce RTX 3090", + "slot": 0 + } + ], + "instance_id": "instance_859cf1e47bca6046", + "topology_links": [], + "torch": { + "backends": { + "cuda_cudnn_sdp_enabled": true, + "cuda_flash_sdp_enabled": true, + "cuda_math_sdp_enabled": true, + "cuda_matmul_allow_tf32": false, + "cuda_mem_efficient_sdp_enabled": true, + "cudnn_allow_tf32": true, + "cudnn_benchmark": false + }, + "cuda_available": true, + "cuda_device_count": 1, + "cuda_version": "12.8", + "current_device": 0, + "current_device_name": "NVIDIA GeForce RTX 3090", + "device_properties": [ + { + "index": 0, + "major": 8, + "minor": 6, + "multi_processor_count": 82, + "name": "NVIDIA GeForce RTX 3090", + "total_memory": 25295257600 + } + ], + "float32_matmul_precision": "highest", + "version": "2.11.0+cu128" + } +} diff --git a/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/profile/latency_burst_model.json b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/profile/latency_burst_model.json new file mode 100644 index 0000000000000000000000000000000000000000..881407186a327a0431f8d19344c2b8dcd43e7700 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/profile/latency_burst_model.json @@ -0,0 +1,1329 @@ +{ + "burst_dwell_distribution": { + "distribution_type": "inverse_cdf", + "latency_ms": [ + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0 + ], + "quantile_levels": [ + 0.0, + 0.0001, + 0.0005, + 0.001, + 0.002, + 0.003, + 0.004, + 0.005, + 0.006, + 0.007, + 0.008, + 0.009, + 0.01, + 0.02, + 0.03, + 0.04, + 0.05, + 0.06, + 0.07, + 0.08, + 0.09, + 0.1, + 0.11, + 0.12, + 0.13, + 0.14, + 0.15, + 0.16, + 0.17, + 0.18, + 0.19, + 0.2, + 0.21, + 0.22, + 0.23, + 0.24, + 0.25, + 0.26, + 0.27, + 0.28, + 0.29, + 0.3, + 0.31, + 0.32, + 0.33, + 0.34, + 0.35, + 0.36, + 0.37, + 0.38, + 0.39, + 0.4, + 0.41, + 0.42, + 0.43, + 0.44, + 0.45, + 0.46, + 0.47, + 0.48, + 0.49, + 0.5, + 0.51, + 0.52, + 0.53, + 0.54, + 0.55, + 0.56, + 0.57, + 0.58, + 0.59, + 0.6, + 0.61, + 0.62, + 0.63, + 0.64, + 0.65, + 0.66, + 0.67, + 0.68, + 0.69, + 0.7, + 0.71, + 0.72, + 0.73, + 0.74, + 0.75, + 0.76, + 0.77, + 0.78, + 0.79, + 0.8, + 0.81, + 0.82, + 0.83, + 0.84, + 0.85, + 0.86, + 0.87, + 0.88, + 0.89, + 0.9, + 0.91, + 0.92, + 0.93, + 0.94, + 0.95, + 0.96, + 0.97, + 0.98, + 0.99, + 0.991, + 0.992, + 0.993, + 0.994, + 0.995, + 0.996, + 0.997, + 0.998, + 0.999, + 0.9995, + 0.9999, + 1.0 + ] + }, + "burst_dwell_lengths": [ + 1, + 1, + 1, + 1 + ], + "burst_merge_gap_records": 30, + "burst_rank_processes": [ + { + "draw_count": 4, + "dwell_length_spearman_rho": 0.0, + "level_ms": { + "distribution_type": "inverse_cdf", + "latency_ms": [ + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.13375687599182, + 72.14111005783082, + 72.15581642150879, + 72.17052278518676, + 72.18522914886475, + 72.19993551254272, + 72.2146418762207, + 72.22934823989868, + 72.24405460357666, + 72.25876096725464, + 72.27346733093262, + 72.28817369461059, + 72.30288005828858, + 72.31758642196655, + 72.33229278564453, + 72.34699914932251, + 72.36170551300049, + 72.37641187667846, + 72.39111824035645, + 72.40582460403442, + 72.42053096771241, + 72.43523733139038, + 72.44994369506836, + 72.46465005874634, + 72.47935642242432, + 72.49406278610229, + 72.50271152496337, + 72.50530263900757, + 72.50789375305176, + 72.51048486709595, + 72.51307598114013, + 72.51566709518433, + 72.51825820922852, + 72.5208493232727, + 72.52344043731689, + 72.52603155136109, + 72.52862266540528, + 72.53121377944946, + 72.53380489349365, + 72.53639600753785, + 72.53898712158202, + 72.54157823562622, + 72.54416934967041, + 72.5467604637146, + 72.54935157775878, + 72.55194269180298, + 72.55453380584717, + 72.55712491989135, + 72.55971603393554, + 72.56230714797974, + 72.56489826202393, + 72.62776734352111, + 72.75091439247132, + 72.8740614414215, + 72.9972084903717, + 73.1203555393219, + 73.2435025882721, + 73.3666496372223, + 73.48979668617248, + 73.61294373512268, + 73.73609078407287, + 73.85923783302307, + 73.98238488197326, + 74.10553193092346, + 74.22867897987366, + 74.35182602882385, + 74.47497307777405, + 74.59812012672424, + 74.72126717567444, + 74.84441422462463, + 74.96756127357483, + 75.09070832252502, + 75.21385537147522, + 75.33700242042542, + 75.46014946937561, + 75.58329651832581, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009, + 75.6448700428009 + ], + "quantile_levels": [ + 0.0, + 0.0001, + 0.0005, + 0.001, + 0.002, + 0.003, + 0.004, + 0.005, + 0.006, + 0.007, + 0.008, + 0.009, + 0.01, + 0.02, + 0.03, + 0.04, + 0.05, + 0.06, + 0.07, + 0.08, + 0.09, + 0.1, + 0.11, + 0.12, + 0.13, + 0.14, + 0.15, + 0.16, + 0.17, + 0.18, + 0.19, + 0.2, + 0.21, + 0.22, + 0.23, + 0.24, + 0.25, + 0.26, + 0.27, + 0.28, + 0.29, + 0.3, + 0.31, + 0.32, + 0.33, + 0.34, + 0.35, + 0.36, + 0.37, + 0.38, + 0.39, + 0.4, + 0.41, + 0.42, + 0.43, + 0.44, + 0.45, + 0.46, + 0.47, + 0.48, + 0.49, + 0.5, + 0.51, + 0.52, + 0.53, + 0.54, + 0.55, + 0.56, + 0.57, + 0.58, + 0.59, + 0.6, + 0.61, + 0.62, + 0.63, + 0.64, + 0.65, + 0.66, + 0.67, + 0.68, + 0.69, + 0.7, + 0.71, + 0.72, + 0.73, + 0.74, + 0.75, + 0.76, + 0.77, + 0.78, + 0.79, + 0.8, + 0.81, + 0.82, + 0.83, + 0.84, + 0.85, + 0.86, + 0.87, + 0.88, + 0.89, + 0.9, + 0.91, + 0.92, + 0.93, + 0.94, + 0.95, + 0.96, + 0.97, + 0.98, + 0.99, + 0.991, + 0.992, + 0.993, + 0.994, + 0.995, + 0.996, + 0.997, + 0.998, + 0.999, + 0.9995, + 0.9999, + 1.0 + ] + }, + "level_residual_ms": { + "distribution_type": "inverse_cdf", + "latency_ms": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "quantile_levels": [ + 0.0, + 0.0001, + 0.0005, + 0.001, + 0.002, + 0.003, + 0.004, + 0.005, + 0.006, + 0.007, + 0.008, + 0.009, + 0.01, + 0.02, + 0.03, + 0.04, + 0.05, + 0.06, + 0.07, + 0.08, + 0.09, + 0.1, + 0.11, + 0.12, + 0.13, + 0.14, + 0.15, + 0.16, + 0.17, + 0.18, + 0.19, + 0.2, + 0.21, + 0.22, + 0.23, + 0.24, + 0.25, + 0.26, + 0.27, + 0.28, + 0.29, + 0.3, + 0.31, + 0.32, + 0.33, + 0.34, + 0.35, + 0.36, + 0.37, + 0.38, + 0.39, + 0.4, + 0.41, + 0.42, + 0.43, + 0.44, + 0.45, + 0.46, + 0.47, + 0.48, + 0.49, + 0.5, + 0.51, + 0.52, + 0.53, + 0.54, + 0.55, + 0.56, + 0.57, + 0.58, + 0.59, + 0.6, + 0.61, + 0.62, + 0.63, + 0.64, + 0.65, + 0.66, + 0.67, + 0.68, + 0.69, + 0.7, + 0.71, + 0.72, + 0.73, + 0.74, + 0.75, + 0.76, + 0.77, + 0.78, + 0.79, + 0.8, + 0.81, + 0.82, + 0.83, + 0.84, + 0.85, + 0.86, + 0.87, + 0.88, + 0.89, + 0.9, + 0.91, + 0.92, + 0.93, + 0.94, + 0.95, + 0.96, + 0.97, + 0.98, + 0.99, + 0.991, + 0.992, + 0.993, + 0.994, + 0.995, + 0.996, + 0.997, + 0.998, + 0.999, + 0.9995, + 0.9999, + 1.0 + ] + }, + "severity": { + "distribution_type": "inverse_cdf", + "latency_ms": [ + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0 + ], + "quantile_levels": [ + 0.0, + 0.0001, + 0.0005, + 0.001, + 0.002, + 0.003, + 0.004, + 0.005, + 0.006, + 0.007, + 0.008, + 0.009, + 0.01, + 0.02, + 0.03, + 0.04, + 0.05, + 0.06, + 0.07, + 0.08, + 0.09, + 0.1, + 0.11, + 0.12, + 0.13, + 0.14, + 0.15, + 0.16, + 0.17, + 0.18, + 0.19, + 0.2, + 0.21, + 0.22, + 0.23, + 0.24, + 0.25, + 0.26, + 0.27, + 0.28, + 0.29, + 0.3, + 0.31, + 0.32, + 0.33, + 0.34, + 0.35, + 0.36, + 0.37, + 0.38, + 0.39, + 0.4, + 0.41, + 0.42, + 0.43, + 0.44, + 0.45, + 0.46, + 0.47, + 0.48, + 0.49, + 0.5, + 0.51, + 0.52, + 0.53, + 0.54, + 0.55, + 0.56, + 0.57, + 0.58, + 0.59, + 0.6, + 0.61, + 0.62, + 0.63, + 0.64, + 0.65, + 0.66, + 0.67, + 0.68, + 0.69, + 0.7, + 0.71, + 0.72, + 0.73, + 0.74, + 0.75, + 0.76, + 0.77, + 0.78, + 0.79, + 0.8, + 0.81, + 0.82, + 0.83, + 0.84, + 0.85, + 0.86, + 0.87, + 0.88, + 0.89, + 0.9, + 0.91, + 0.92, + 0.93, + 0.94, + 0.95, + 0.96, + 0.97, + 0.98, + 0.99, + 0.991, + 0.992, + 0.993, + 0.994, + 0.995, + 0.996, + 0.997, + 0.998, + 0.999, + 0.9995, + 0.9999, + 1.0 + ] + }, + "spike_count": 4 + } + ], + "burst_slot_rank_templates": [ + { + "dwell_length": 1, + "slot_to_rank": { + "0": 0 + } + }, + { + "dwell_length": 1, + "slot_to_rank": { + "0": 0 + } + }, + { + "dwell_length": 1, + "slot_to_rank": { + "0": 0 + } + }, + { + "dwell_length": 1, + "slot_to_rank": { + "0": 0 + } + } + ], + "model_type": "hidden_regime", + "pre_worker_time_ms": { + "distribution_type": "inverse_cdf", + "latency_ms": [ + 0.20420193672180176, + 0.21483877639770507, + 0.2287143712043762, + 0.24032604026794432, + 0.24867507362365723, + 0.25438513946533203, + 0.2607346229553223, + 0.2639804744720459, + 0.26635993289947507, + 0.2687912712097168, + 0.2710361452102661, + 0.27275579833984376, + 0.273915548324585, + 0.2837890768051147, + 0.2895009899139404, + 0.2939703321456909, + 0.2985083103179932, + 0.30235490322113034, + 0.30625831604003906, + 0.3095644426345825, + 0.31271454334259036, + 0.3151197671890259, + 0.317447395324707, + 0.3196178579330444, + 0.3216020727157593, + 0.3235012102127075, + 0.3252310037612915, + 0.3269206190109253, + 0.3283870220184326, + 0.3298677968978882, + 0.3312730979919434, + 0.332612419128418, + 0.3338682413101196, + 0.335247278213501, + 0.33649143218994143, + 0.33758762836456296, + 0.338687539100647, + 0.3397971725463867, + 0.34085161685943605, + 0.3419586896896362, + 0.3430896520614624, + 0.34424476623535155, + 0.3454119539260864, + 0.346552209854126, + 0.3476839065551758, + 0.3487655544281006, + 0.3497673511505127, + 0.35085655212402345, + 0.3518979120254517, + 0.35300746440887454, + 0.35399904727935794, + 0.35509252548217773, + 0.3561410903930664, + 0.3571738576889038, + 0.358305869102478, + 0.3593671655654907, + 0.3603264331817627, + 0.3612997245788574, + 0.362232985496521, + 0.3632385206222534, + 0.3641740322113037, + 0.365062952041626, + 0.3658906316757202, + 0.3667243766784668, + 0.3675804567337036, + 0.36850805759429933, + 0.36934189796447753, + 0.3703420162200928, + 0.3712935400009155, + 0.3722340869903564, + 0.3732797241210937, + 0.37433724403381347, + 0.37558521270751954, + 0.3767164659500122, + 0.3778761959075928, + 0.3790695905685425, + 0.38028051853179934, + 0.3815205669403076, + 0.38275862216949463, + 0.3840209150314331, + 0.38523890495300295, + 0.3865478992462158, + 0.3878518056869507, + 0.3894425630569458, + 0.3910191059112549, + 0.3924675369262695, + 0.3939321041107178, + 0.3954088592529297, + 0.39689703941345217, + 0.3985757350921631, + 0.400292010307312, + 0.40227370262145995, + 0.4040238857269287, + 0.40595915317535397, + 0.40784664154052735, + 0.41001430988311766, + 0.41180758476257323, + 0.4139445734024048, + 0.415944185256958, + 0.41818367958068847, + 0.4204584693908691, + 0.4230041980743408, + 0.4256094837188721, + 0.4285480833053589, + 0.4317239999771118, + 0.4352530431747436, + 0.4396242618560791, + 0.4441619348526001, + 0.4498762226104737, + 0.45640567302703855, + 0.46785748481750483, + 0.47020196151733407, + 0.4729775695800772, + 0.475634225845337, + 0.4788666810989385, + 0.48331641197204583, + 0.491900497436524, + 0.49948444366455025, + 0.5097592239379892, + 0.5262329158782943, + 0.5426986169815089, + 0.6917105711937698, + 1.1622118949890137 + ], + "quantile_levels": [ + 0.0, + 0.0001, + 0.0005, + 0.001, + 0.002, + 0.003, + 0.004, + 0.005, + 0.006, + 0.007, + 0.008, + 0.009, + 0.01, + 0.02, + 0.03, + 0.04, + 0.05, + 0.06, + 0.07, + 0.08, + 0.09, + 0.1, + 0.11, + 0.12, + 0.13, + 0.14, + 0.15, + 0.16, + 0.17, + 0.18, + 0.19, + 0.2, + 0.21, + 0.22, + 0.23, + 0.24, + 0.25, + 0.26, + 0.27, + 0.28, + 0.29, + 0.3, + 0.31, + 0.32, + 0.33, + 0.34, + 0.35, + 0.36, + 0.37, + 0.38, + 0.39, + 0.4, + 0.41, + 0.42, + 0.43, + 0.44, + 0.45, + 0.46, + 0.47, + 0.48, + 0.49, + 0.5, + 0.51, + 0.52, + 0.53, + 0.54, + 0.55, + 0.56, + 0.57, + 0.58, + 0.59, + 0.6, + 0.61, + 0.62, + 0.63, + 0.64, + 0.65, + 0.66, + 0.67, + 0.68, + 0.69, + 0.7, + 0.71, + 0.72, + 0.73, + 0.74, + 0.75, + 0.76, + 0.77, + 0.78, + 0.79, + 0.8, + 0.81, + 0.82, + 0.83, + 0.84, + 0.85, + 0.86, + 0.87, + 0.88, + 0.89, + 0.9, + 0.91, + 0.92, + 0.93, + 0.94, + 0.95, + 0.96, + 0.97, + 0.98, + 0.99, + 0.991, + 0.992, + 0.993, + 0.994, + 0.995, + 0.996, + 0.997, + 0.998, + 0.999, + 0.9995, + 0.9999, + 1.0 + ] + }, + "regime_step_counts": { + "burst": 4, + "calm": 35904 + }, + "regime_transition_counts": { + "burst": { + "burst": 0, + "calm": 4 + }, + "calm": { + "burst": 4, + "calm": 35895 + } + }, + "reset_scope": "session", + "schema_version": 12, + "spike_median_multiplier": 1.25, + "spike_threshold_ms_by_worker_slot": { + "0": 72.03269615769386 + }, + "worker_count": 1 +} diff --git a/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/profile/latency_distribution.json b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/profile/latency_distribution.json new file mode 100644 index 0000000000000000000000000000000000000000..2b45df3ed46456efa7943e385fa375d3adaa0846 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/profile/latency_distribution.json @@ -0,0 +1,1027 @@ +{ + "schema_version": 3, + "worker_slots": { + "0": { + "all": { + "count": 35908, + "observation_to_action_latency_ms": { + "distribution_type": "inverse_cdf", + "latency_ms": [ + 54.89762210845947, + 54.98431869125366, + 55.134520971298215, + 55.34168244361877, + 55.99118143081665, + 56.076304399490354, + 56.13243092346191, + 56.18818429946899, + 56.23180709552765, + 56.27835348701477, + 56.33713837718964, + 56.39195935821533, + 56.45379021167755, + 57.08514684200287, + 57.230893654823305, + 57.316619853973386, + 57.36959619522095, + 57.40613402843476, + 57.43060580730438, + 57.45380989074707, + 57.477882833480834, + 57.49994168281555, + 57.52059514522553, + 57.5393541431427, + 57.55763153076172, + 57.57544971466064, + 57.59502432346344, + 57.6165852022171, + 57.639358162879944, + 57.66515370845795, + 57.68923002243042, + 57.72054600715637, + 57.75672510147095, + 57.802322187423705, + 57.88595350265503, + 57.97818984031677, + 58.037572503089905, + 58.08466000556946, + 58.12414084434509, + 58.16394332885742, + 58.203309774398804, + 58.24743611812592, + 58.28536067008972, + 58.316369466781616, + 58.339424300193784, + 58.358790407180784, + 58.37594494819641, + 58.39228649616241, + 58.409412431716916, + 58.42540137290955, + 58.44186356544495, + 58.45765173435211, + 58.474079790115354, + 58.4883131980896, + 58.50111918926239, + 58.51361119747162, + 58.52541661262512, + 58.53668155670166, + 58.548266739845275, + 58.55865183830261, + 58.5691175699234, + 58.578869700431824, + 58.58787773609161, + 58.59606602191925, + 58.60470150947571, + 58.61385565757752, + 58.62180905342102, + 58.62971657276154, + 58.63715500831604, + 58.64509712219238, + 58.6526384973526, + 58.66041216850281, + 58.66814685344696, + 58.6752281665802, + 58.68287703037262, + 58.69019066810608, + 58.69767053127289, + 58.70504686832428, + 58.712141094207766, + 58.718946146965024, + 58.72618232250213, + 58.73353822231293, + 58.74131082057953, + 58.74863606929779, + 58.75591662883758, + 58.76344584465027, + 58.77212584018707, + 58.779995255470276, + 58.78799551010132, + 58.79621982574463, + 58.80471308708191, + 58.813468432426454, + 58.82311668395996, + 58.833434352874754, + 58.843691573143005, + 58.85456703662872, + 58.86566646099091, + 58.875943303108215, + 58.88800809860229, + 58.90074449539185, + 58.91300926685333, + 58.9265673160553, + 58.93994530677796, + 58.955101170539855, + 58.972674880027775, + 58.99344714641571, + 59.02611937522888, + 59.10132591247559, + 59.611970896720884, + 59.7586422252655, + 59.97354772090912, + 60.00530353927612, + 60.039614966392506, + 60.09215181350708, + 60.16863778018952, + 60.36564081192016, + 60.57177989768982, + 60.719089326858516, + 61.37363024044037, + 64.18603834533658, + 67.25203033256551, + 72.61951753550008, + 76.16190600395203 + ], + "quantile_levels": [ + 0.0, + 0.0001, + 0.0005, + 0.001, + 0.002, + 0.003, + 0.004, + 0.005, + 0.006, + 0.007, + 0.008, + 0.009, + 0.01, + 0.02, + 0.03, + 0.04, + 0.05, + 0.06, + 0.07, + 0.08, + 0.09, + 0.1, + 0.11, + 0.12, + 0.13, + 0.14, + 0.15, + 0.16, + 0.17, + 0.18, + 0.19, + 0.2, + 0.21, + 0.22, + 0.23, + 0.24, + 0.25, + 0.26, + 0.27, + 0.28, + 0.29, + 0.3, + 0.31, + 0.32, + 0.33, + 0.34, + 0.35, + 0.36, + 0.37, + 0.38, + 0.39, + 0.4, + 0.41, + 0.42, + 0.43, + 0.44, + 0.45, + 0.46, + 0.47, + 0.48, + 0.49, + 0.5, + 0.51, + 0.52, + 0.53, + 0.54, + 0.55, + 0.56, + 0.57, + 0.58, + 0.59, + 0.6, + 0.61, + 0.62, + 0.63, + 0.64, + 0.65, + 0.66, + 0.67, + 0.68, + 0.69, + 0.7, + 0.71, + 0.72, + 0.73, + 0.74, + 0.75, + 0.76, + 0.77, + 0.78, + 0.79, + 0.8, + 0.81, + 0.82, + 0.83, + 0.84, + 0.85, + 0.86, + 0.87, + 0.88, + 0.89, + 0.9, + 0.91, + 0.92, + 0.93, + 0.94, + 0.95, + 0.96, + 0.97, + 0.98, + 0.99, + 0.991, + 0.992, + 0.993, + 0.994, + 0.995, + 0.996, + 0.997, + 0.998, + 0.999, + 0.9995, + 0.9999, + 1.0 + ] + }, + "spearman_rho": 0.9930560092450511, + "worker_service_time_ms": { + "distribution_type": "inverse_cdf", + "latency_ms": [ + 54.638512134552, + 54.722969214344026, + 54.86490662717819, + 55.03681116294861, + 55.72376442432404, + 55.79827515983582, + 55.85051815986633, + 55.895044279098514, + 55.94917815303803, + 55.99627424621582, + 56.03666737651825, + 56.09092275047302, + 56.14324330806732, + 56.75663536548615, + 56.894120874404905, + 56.97365468502045, + 57.02204294204712, + 57.05275318145752, + 57.08081398963928, + 57.10400746822357, + 57.12682245254516, + 57.14491307735443, + 57.162690525054934, + 57.183143854141235, + 57.201689219474794, + 57.218997974395755, + 57.23760087490082, + 57.26001518726349, + 57.28137402057648, + 57.30520550251007, + 57.33147391319275, + 57.363392567634584, + 57.396899285316465, + 57.4381804561615, + 57.51748081684113, + 57.60506856441498, + 57.66861927509308, + 57.73105767250061, + 57.77032548904419, + 57.81555837154389, + 57.861557040214535, + 57.90422902107239, + 57.94723031044006, + 57.97415527820587, + 57.99504571914673, + 58.012214488983155, + 58.02739794254303, + 58.04379164218903, + 58.058891053199766, + 58.07336838245392, + 58.08783471107483, + 58.10370817184448, + 58.1183429479599, + 58.13163368225098, + 58.144854125976565, + 58.156135168075565, + 58.1677907705307, + 58.17717433929443, + 58.18661304473877, + 58.19571486473083, + 58.20485938072205, + 58.21307682991028, + 58.220322942733766, + 58.228040552139284, + 58.23579728603363, + 58.24308175563812, + 58.249634408950804, + 58.25668476104736, + 58.263752183914185, + 58.27049364566803, + 58.27714119911194, + 58.28466124534607, + 58.29146835803986, + 58.29871460437775, + 58.30632412910462, + 58.31351550579071, + 58.320294523239134, + 58.327939581871036, + 58.33411476135254, + 58.34080623149872, + 58.34782395839691, + 58.3550388097763, + 58.36206126213074, + 58.36932575702667, + 58.37651247024536, + 58.38423424243927, + 58.391886472702026, + 58.400058970451354, + 58.407453675270084, + 58.41486597061157, + 58.423468680381774, + 58.43212459087372, + 58.441460065841675, + 58.45135697364807, + 58.46152633190155, + 58.472478790283205, + 58.4831839799881, + 58.494375071525575, + 58.50569304466248, + 58.51628562450409, + 58.52735684394836, + 58.53880505561828, + 58.55049607753754, + 58.56301250934601, + 58.57766434192658, + 58.59507055282593, + 58.623821711540224, + 58.70154627799988, + 59.20479492664337, + 59.354912848472594, + 59.5520503616333, + 59.586989126205445, + 59.629149945259094, + 59.67486940002441, + 59.748895297050474, + 60.0047911453247, + 60.194087533950814, + 60.32199776840209, + 60.85784722709658, + 63.77599292945856, + 66.87870573186876, + 71.94066750355165, + 75.6448700428009 + ], + "quantile_levels": [ + 0.0, + 0.0001, + 0.0005, + 0.001, + 0.002, + 0.003, + 0.004, + 0.005, + 0.006, + 0.007, + 0.008, + 0.009, + 0.01, + 0.02, + 0.03, + 0.04, + 0.05, + 0.06, + 0.07, + 0.08, + 0.09, + 0.1, + 0.11, + 0.12, + 0.13, + 0.14, + 0.15, + 0.16, + 0.17, + 0.18, + 0.19, + 0.2, + 0.21, + 0.22, + 0.23, + 0.24, + 0.25, + 0.26, + 0.27, + 0.28, + 0.29, + 0.3, + 0.31, + 0.32, + 0.33, + 0.34, + 0.35, + 0.36, + 0.37, + 0.38, + 0.39, + 0.4, + 0.41, + 0.42, + 0.43, + 0.44, + 0.45, + 0.46, + 0.47, + 0.48, + 0.49, + 0.5, + 0.51, + 0.52, + 0.53, + 0.54, + 0.55, + 0.56, + 0.57, + 0.58, + 0.59, + 0.6, + 0.61, + 0.62, + 0.63, + 0.64, + 0.65, + 0.66, + 0.67, + 0.68, + 0.69, + 0.7, + 0.71, + 0.72, + 0.73, + 0.74, + 0.75, + 0.76, + 0.77, + 0.78, + 0.79, + 0.8, + 0.81, + 0.82, + 0.83, + 0.84, + 0.85, + 0.86, + 0.87, + 0.88, + 0.89, + 0.9, + 0.91, + 0.92, + 0.93, + 0.94, + 0.95, + 0.96, + 0.97, + 0.98, + 0.99, + 0.991, + 0.992, + 0.993, + 0.994, + 0.995, + 0.996, + 0.997, + 0.998, + 0.999, + 0.9995, + 0.9999, + 1.0 + ] + } + }, + "steady": { + "count": 35904, + "observation_to_action_latency_ms": { + "distribution_type": "inverse_cdf", + "latency_ms": [ + 54.89762210845947, + 54.984307411193846, + 55.134505897521976, + 55.34166074275971, + 55.991178302764894, + 56.076294094085696, + 56.13233103179932, + 56.18812419891358, + 56.231794157981874, + 56.27833763885498, + 56.33710426616669, + 56.391950715065, + 56.45378501415253, + 57.08512797832489, + 57.23083909511566, + 57.31661008834839, + 57.3695746421814, + 57.40609849452972, + 57.43057663917541, + 57.453807735443114, + 57.477878031730654, + 57.499938297271726, + 57.520570921897885, + 57.53932119846344, + 57.55763103485108, + 57.57542407989502, + 57.59496324062347, + 57.616562161445614, + 57.6393541097641, + 57.66512469768524, + 57.68913018226623, + 57.720493364334104, + 57.75670382976532, + 57.802289566993714, + 57.88569365978241, + 57.978020153045655, + 58.037256360054016, + 58.0845925617218, + 58.1240485048294, + 58.16375423908234, + 58.20327997684479, + 58.24732210636139, + 58.285302591323855, + 58.316290488243105, + 58.33935274600983, + 58.35876348018646, + 58.37584476470947, + 58.392186732292174, + 58.40936285018921, + 58.42536652565003, + 58.44180078983307, + 58.45757086277008, + 58.474003086090086, + 58.488275771141055, + 58.501091632843014, + 58.51360427379608, + 58.52537198066712, + 58.5366578245163, + 58.5481983423233, + 58.558502655029294, + 58.5690626001358, + 58.57883095741272, + 58.58777729511261, + 58.59602118492126, + 58.60467755317688, + 58.613761353492734, + 58.62173254489899, + 58.62966779708862, + 58.637111430168154, + 58.645032176971434, + 58.652547693252565, + 58.660341358184816, + 58.6681103515625, + 58.675179276466366, + 58.68275595188141, + 58.69008821964264, + 58.69758403301239, + 58.70498054981232, + 58.71211910247803, + 58.71892172336578, + 58.726107902526856, + 58.73349041938782, + 58.74119870662689, + 58.748615841865536, + 58.75589869499206, + 58.763404245376584, + 58.77199697494507, + 58.77996703624726, + 58.78795792102814, + 58.79615904808045, + 58.80460977554321, + 58.813416481018066, + 58.823071904182434, + 58.83335713863373, + 58.84358658790588, + 58.85443187713623, + 58.8656010389328, + 58.87590140342712, + 58.88786563873291, + 58.900563769340515, + 58.91284184932709, + 58.92648019790649, + 58.93981793880462, + 58.954998784065246, + 58.97248013019562, + 58.992928538322445, + 59.025522994995114, + 59.099266309738155, + 59.605603127479554, + 59.75604496479034, + 59.968015394210816, + 60.00281594467163, + 60.03302575588225, + 60.08006652545929, + 60.15268362808228, + 60.34006970405586, + 60.55531148242949, + 60.70003566932679, + 61.00997841644302, + 63.63513048267335, + 66.65921190643537, + 69.54871525268524, + 70.52106499671936 + ], + "quantile_levels": [ + 0.0, + 0.0001, + 0.0005, + 0.001, + 0.002, + 0.003, + 0.004, + 0.005, + 0.006, + 0.007, + 0.008, + 0.009, + 0.01, + 0.02, + 0.03, + 0.04, + 0.05, + 0.06, + 0.07, + 0.08, + 0.09, + 0.1, + 0.11, + 0.12, + 0.13, + 0.14, + 0.15, + 0.16, + 0.17, + 0.18, + 0.19, + 0.2, + 0.21, + 0.22, + 0.23, + 0.24, + 0.25, + 0.26, + 0.27, + 0.28, + 0.29, + 0.3, + 0.31, + 0.32, + 0.33, + 0.34, + 0.35, + 0.36, + 0.37, + 0.38, + 0.39, + 0.4, + 0.41, + 0.42, + 0.43, + 0.44, + 0.45, + 0.46, + 0.47, + 0.48, + 0.49, + 0.5, + 0.51, + 0.52, + 0.53, + 0.54, + 0.55, + 0.56, + 0.57, + 0.58, + 0.59, + 0.6, + 0.61, + 0.62, + 0.63, + 0.64, + 0.65, + 0.66, + 0.67, + 0.68, + 0.69, + 0.7, + 0.71, + 0.72, + 0.73, + 0.74, + 0.75, + 0.76, + 0.77, + 0.78, + 0.79, + 0.8, + 0.81, + 0.82, + 0.83, + 0.84, + 0.85, + 0.86, + 0.87, + 0.88, + 0.89, + 0.9, + 0.91, + 0.92, + 0.93, + 0.94, + 0.95, + 0.96, + 0.97, + 0.98, + 0.99, + 0.991, + 0.992, + 0.993, + 0.994, + 0.995, + 0.996, + 0.997, + 0.998, + 0.999, + 0.9995, + 0.9999, + 1.0 + ] + }, + "spearman_rho": 0.9930536881339532, + "worker_service_time_ms": { + "distribution_type": "inverse_cdf", + "latency_ms": [ + 54.638512134552, + 54.722956764411926, + 54.86490254116058, + 55.03670747375488, + 55.72376233577728, + 55.79825168800354, + 55.85047158241272, + 55.89504156112671, + 55.94910483646393, + 55.996236187934876, + 56.036575892448425, + 56.090843374252316, + 56.14299791812897, + 56.756627945899965, + 56.89406868934631, + 56.97361436367035, + 57.022025203704835, + 57.05273281097412, + 57.08080776691437, + 57.104003982543944, + 57.126808791160585, + 57.14489750862121, + 57.16263738155365, + 57.18307790756226, + 57.20166008472442, + 57.218993434906004, + 57.23759672641754, + 57.2600101518631, + 57.281266531944276, + 57.30516447544098, + 57.33145670413971, + 57.36331901550293, + 57.39689316272736, + 57.43804866313934, + 57.51728015422821, + 57.60498538017273, + 57.66832494735718, + 57.73094385147095, + 57.770116863250735, + 57.8153165435791, + 57.86133855342865, + 57.9039803981781, + 57.947170243263244, + 57.974107985496524, + 57.99501208305359, + 58.01215915203095, + 58.027367973327635, + 58.04376080989837, + 58.05885187625885, + 58.07331806659698, + 58.0877879524231, + 58.103690600395204, + 58.11831189632416, + 58.13161474704742, + 58.144645142555234, + 58.15612073421478, + 58.16774473190308, + 58.1771510219574, + 58.186569285392764, + 58.19569593429566, + 58.20483826637268, + 58.21305239200592, + 58.220281167030336, + 58.227979702949526, + 58.23577118873596, + 58.24305790424347, + 58.24960861206055, + 58.25667642593384, + 58.26372510910034, + 58.270454411506655, + 58.27711375236511, + 58.28462617397308, + 58.291462364196775, + 58.29867696762085, + 58.30626531124115, + 58.31345600605011, + 58.320233178138736, + 58.32780292034149, + 58.33409640789032, + 58.340758085250854, + 58.34780982017517, + 58.3549880027771, + 58.361964631080625, + 58.36928438186646, + 58.37640544891357, + 58.384184594154355, + 58.39184761047363, + 58.40000975608826, + 58.40733815193176, + 58.414857358932494, + 58.42340198993683, + 58.432047080993655, + 58.44124049186706, + 58.45121502876282, + 58.46129935741425, + 58.472327423095706, + 58.48313086032867, + 58.49423496723175, + 58.50561022758484, + 58.516244192123416, + 58.52729542255402, + 58.53868432044983, + 58.55042207241058, + 58.56292420387268, + 58.57733368396759, + 58.59481431007385, + 58.62353591918945, + 58.69977615833282, + 59.20313117980957, + 59.35360734462738, + 59.54943576812744, + 59.58170782470703, + 59.627803606033325, + 59.66968642902374, + 59.73967554950714, + 59.93983187198645, + 60.18512020492554, + 60.29954465675354, + 60.62631217098242, + 63.37979303073861, + 66.17347280121018, + 69.13033782882663, + 70.00722193717957 + ], + "quantile_levels": [ + 0.0, + 0.0001, + 0.0005, + 0.001, + 0.002, + 0.003, + 0.004, + 0.005, + 0.006, + 0.007, + 0.008, + 0.009, + 0.01, + 0.02, + 0.03, + 0.04, + 0.05, + 0.06, + 0.07, + 0.08, + 0.09, + 0.1, + 0.11, + 0.12, + 0.13, + 0.14, + 0.15, + 0.16, + 0.17, + 0.18, + 0.19, + 0.2, + 0.21, + 0.22, + 0.23, + 0.24, + 0.25, + 0.26, + 0.27, + 0.28, + 0.29, + 0.3, + 0.31, + 0.32, + 0.33, + 0.34, + 0.35, + 0.36, + 0.37, + 0.38, + 0.39, + 0.4, + 0.41, + 0.42, + 0.43, + 0.44, + 0.45, + 0.46, + 0.47, + 0.48, + 0.49, + 0.5, + 0.51, + 0.52, + 0.53, + 0.54, + 0.55, + 0.56, + 0.57, + 0.58, + 0.59, + 0.6, + 0.61, + 0.62, + 0.63, + 0.64, + 0.65, + 0.66, + 0.67, + 0.68, + 0.69, + 0.7, + 0.71, + 0.72, + 0.73, + 0.74, + 0.75, + 0.76, + 0.77, + 0.78, + 0.79, + 0.8, + 0.81, + 0.82, + 0.83, + 0.84, + 0.85, + 0.86, + 0.87, + 0.88, + 0.89, + 0.9, + 0.91, + 0.92, + 0.93, + 0.94, + 0.95, + 0.96, + 0.97, + 0.98, + 0.99, + 0.991, + 0.992, + 0.993, + 0.994, + 0.995, + 0.996, + 0.997, + 0.998, + 0.999, + 0.9995, + 0.9999, + 1.0 + ] + } + } + } + } +} diff --git a/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/profile/latency_profile.png b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/profile/latency_profile.png new file mode 100644 index 0000000000000000000000000000000000000000..d092895d0fee0291370b91ff946477640d2b58a4 Binary files /dev/null and b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/profile/latency_profile.png differ diff --git a/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/profile/profile.json b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/profile/profile.json new file mode 100644 index 0000000000000000000000000000000000000000..67bdbdee352adea1325fe3253410528697a354cf --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/profile/profile.json @@ -0,0 +1,66 @@ +{ + "burst_model_path": "latency_burst_model.json", + "distribution_path": "latency_distribution.json", + "env_fps": 50.0, + "frame_ms": 20.0, + "gpu_class": "1x-rtx3090", + "instance_id": "instance_859cf1e47bca6046", + "latency_kind": "observation_to_action_latency", + "latency_method": "temporal", + "model_id": "qwenoft", + "n_admitted_observations": 35908, + "n_capacity_drops": 0, + "n_observation_attempts": 35908, + "per_slot_summary": { + "0": { + "admitted_count": 35908, + "mean_observation_to_action_latency_ms": 58.405461079588015, + "mean_worker_service_time_ms": 58.03848960715147, + "p95_observation_to_action_latency_ms": 59.02611216306687, + "p95_worker_service_time_ms": 58.62378237247468, + "p99_worker_service_time_ms": 59.55199083089829 + } + }, + "provenance": { + "base_config": "/workspace/tasks/20260911T023128Z-p-only4/air_raid/profile.yaml", + "checkpoint_kind": "best", + "model_artifact": { + "checkpoint": "checkpoints/steps_5000_pytorch_model.pt", + "model_config": "config.full.yaml", + "path_in_repo": "OpenVLA/zero-latency/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k", + "repo_id": "latency-sensitive-bench/extra-envs-checkpoints", + "source": "local" + }, + "session_ids": [ + 0, + 1, + 2, + 3, + 4 + ] + }, + "sample_model_type": "hidden_regime", + "source_run_id": "20260911T023704401836Z", + "summary": { + "frame_ms": 20.0, + "max_ms": 76.16190600395203, + "mean_effective_frames": 2.920273053979401, + "mean_ms": 58.405461079588015, + "min_ms": 54.89762210845947, + "n_samples": 35908, + "p50_frames": 2.928943485021591, + "p50_ms": 58.578869700431824, + "p90_frames": 2.946328270435333, + "p90_ms": 58.926565408706665, + "p95_frames": 2.9513056081533433, + "p95_ms": 59.02611216306686, + "p99_frames": 2.998659399986267, + "p99_ms": 59.97318799972534, + "prob_latency_gt_1_frame": 1.0, + "prob_latency_gt_2_frames": 1.0, + "prob_latency_gt_3_frames": 0.009134454717611675, + "std_ms": 0.7141128839373884 + }, + "visualization_path": "latency_profile.png", + "workload_id": "air_raid" +} diff --git a/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/provenance.json b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/provenance.json new file mode 100644 index 0000000000000000000000000000000000000000..271a26d2e52cd57926446d16f69ff0453b4fd3c4 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/provenance.json @@ -0,0 +1,34 @@ +{ + "task": "air-raid", + "model": "qwenoft", + "training_condition": "latency-aware", + "training_run_id": "air_raid_profile_20260911T054920Z_openvla_native_sft_5k", + "source": { + "repo_id": "latency-sensitive-bench/Standard-Pipeline", + "revision": "2208875f92b21e3f8ec2363a30a6705dfe56cc04", + "path": "air_raid/profile_latency/OpenVLA", + "url": "https://huggingface.co/latency-sensitive-bench/Standard-Pipeline/tree/2208875f92b21e3f8ec2363a30a6705dfe56cc04/air_raid/profile_latency/OpenVLA" + }, + "hf_repo_id": "latency-sensitive-bench/benchmark-models", + "hf_path": "latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k", + "checkpoint": { + "source_file": "air_raid/profile_latency/OpenVLA/checkpoints/steps_5000_pytorch_model.pt", + "source_sha256": "ded1152ea155cf86eff73646e986f5b764007c32c82753f36238a17cf4bad734", + "source_bytes": 9785049835, + "file": "checkpoints/model.pt", + "selection_rule": "only published VLA checkpoint", + "method": "copy", + "sha256": "ded1152ea155cf86eff73646e986f5b764007c32c82753f36238a17cf4bad734", + "bytes": 9785049835 + }, + "config_source": "air_raid/profile_latency/OpenVLA/config.full.yaml", + "task_contract_file": "task_contract.json", + "missing_source_files": [], + "migration": { + "operation": "copy_selected_inference_bundle", + "source_modified": false, + "evaluation_performed": false, + "experiment_status": "preserved_in_source_evidence" + }, + "action_horizon": 1 +} diff --git a/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/source/config.yaml b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/source/config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..8f62470a124b973dde16b5b1cf7d209ac876fb32 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/source/config.yaml @@ -0,0 +1,86 @@ +checkpoint: + local: + keep_last_n: 1 + save_best_model: false + save_final_model: true + save_pt_file: true + save_safetensors_file: false + save_training_state: false + sync: + enabled: false + keep_last_n: 0 + repo_id: null +datasets: + vla_data: + data_mix: air_raid_profile_20260911T054920Z + dataset_py: lerobot_datasets + eval_data_mix: air_raid_profile_20260911T054920Z__val + latency_curriculum: + enabled: false + per_device_batch_size: 16 +framework: + action_model: + action_dim: 6 + action_env_dim: 6 + action_hidden_dim: 2560 + action_horizon: 1 + action_model_type: MLP + loss_type: discrete_ce + state_encoding: discretized_text + task_objective: null + kv_memory: + enabled: false + packed_train: false + rebased_sink: true + rollout_len: 8 + window: 4 + name: QwenOFT + qwenvl: + attn_implementation: flash_attention_2 + base_vlm: /mnt/checkpoints/latency-sensitive-bench/openvla/Qwen3-VL-4B-Instruct + enable_gradient_checkpointing: true +output_dir: /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/vla/air_raid_profile_20260911T054920Z_openvla_native_sft_5k +rl_games: + env_eval: + enabled: false + task: gymnasium +run_id: air_raid_profile_20260911T054920Z_openvla_native_sft_5k +run_root_dir: /mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/vla +seed: 42 +trainer: + distributed_backend: deepspeed + eval_action_classification: true + eval_action_classification_interval: null + eval_interval: 500 + eval_num_batches: 200 + freeze_llm_layers: [] + freeze_modules: '' + freeze_tied_embedding: false + freeze_vit: false + gradient_accumulation_steps: 1 + is_resume: false + learning_rate: + action_model: 0.0001 + base: 2.0e-05 + qwen_vl_interface: 1.0e-05 + logging_frequency: 1 + lr_scheduler_type: cosine_with_min_lr + max_train_steps: 5000 + num_warmup_steps: 100 + optimizer: + betas: + - 0.9 + - 0.95 + eps: 1.0e-08 + fused: true + weight_decay: 1.0e-08 + per_latency_eval_num_batches: null + pretrained_checkpoint: null + profile_timing: + enabled: true + log_interval: 10 + save_interval: 5000 + scheduler_specific_kwargs: + min_lr: 1.0e-06 +wandb_entity: dongqianyu99-zhejiang-university +wandb_project: latency-sensitive-bench diff --git a/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/source/provenance.json b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/source/provenance.json new file mode 100644 index 0000000000000000000000000000000000000000..748575df3d31ff5d125e1db56b08cadae69233da --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/source/provenance.json @@ -0,0 +1,818 @@ +{ + "run_id": "20260911T054920Z-airraid", + "final_step": 5000, + "exit_code": 0, + "checkpoint_name": "steps_5000_pytorch_model.pt", + "checkpoint_size": 9785049835, + "checkpoint_sha256": "ded1152ea155cf86eff73646e986f5b764007c32c82753f36238a17cf4bad734", + "state_dict_entries": 730, + "state_dict_key_examples": [ + "qwen_vl_interface.model.model.visual.patch_embed.proj.weight", + "qwen_vl_interface.model.model.visual.patch_embed.proj.bias", + "qwen_vl_interface.model.model.visual.pos_embed.weight", + "qwen_vl_interface.model.model.visual.blocks.0.norm1.weight", + "qwen_vl_interface.model.model.visual.blocks.0.norm1.bias" + ], + "checkpoint_content": "model state dict only", + "wandb": { + "status": "finished", + "wandb_url": "https://wandb.ai/dongqianyu99-zhejiang-university/latency-sensitive-bench/runs/6ahbs76o", + "summary": { + "_runtime": 12725.581948378, + "_step": 5000, + "_timestamp": 1789121686.9782546, + "_wandb.runtime": 12725, + "batch/effective_tokens": 2592, + "batch/image_count": 16, + "batch/input_len_max": 162, + "batch/input_len_mean": 162, + "batch/padding_ratio": 0, + "batch/pixel_values_rows": 4096, + "batch/size": 16, + "epoch": 0.99, + "eval/action_loss/samples": 3200, + "eval/action_loss/seconds": 26.75128673099971, + "eval/air_raid/latency_3/loss": 0.6867397427558899, + "eval/air_raid/latency_4/loss": 0.8789690136909485, + "eval/air_raid/loss": 0.6881214380264282, + "eval/latency_3/loss": 0.6867397427558899, + "eval/latency_4/loss": 0.8789690136909485, + "eval/loss": 0.6881214380264282, + "learning_rate/action_model": 5e-06, + "learning_rate/qwen_vl_interface": 5e-07, + "throughput/effective_tokens_per_sec": 1050.778200191851, + "throughput/samples_per_sec": 6.48628518636945, + "timing/action_head_loss": 0.001663076996919699, + "timing/backward": 0.7658895589993335, + "timing/checkpoint_total": 12.43357680599729, + "timing/data": 0.0004684129962697625, + "timing/dataloader_next": 0.0006775039946660399, + "timing/eval_action_classification_total": 0.0001606090008863248, + "timing/eval_action_loss_total": 26.7540708680026, + "timing/forward": 0.14472050499898614, + "timing/log_metrics_total": 0.0053996329952497035, + "timing/lr_scheduler": 7.950500003062189e-05, + "timing/model": 2.4665819740039296, + "timing/optimizer_step": 1.5525968020010623, + "timing/qwen_h2d": 0.003248660999815911, + "timing/qwen_input_build_total": 0.021369713002059143, + "timing/qwen_processor": 0.017566852999152616, + "timing/train_step_total": 2.466743219003547, + "timing/vlm_forward": 0.12100403300428296, + "train/grad_norm_pre_clip": 13.110580444335938, + "train/loss": 0.8292630910873413 + }, + "verified_utc": "2026-09-11T10:34:21.961255+00:00", + "exit_code": 0, + "final_step": 5000, + "post_train_eval": false, + "checkpoint_path": "/mnt/results/latency-sensitive-bench/profile-latency-new/air_raid/20260911T054920Z-airraid/vla/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/checkpoints/steps_5000_pytorch_model.pt", + "completed_log_utc": "2026-09-11T10:14:52Z" + }, + "dataset_upload": [ + { + "repo_id": "latency-sensitive-bench/extra-envs-rollouts", + "repo_type": "dataset", + "prefix": "profile-latency-new/air_raid/20260911T054920Z-airraid/raw", + "commit": "2fc5692235d290668e220b3d279991466876c929", + "url": "https://huggingface.co/datasets/latency-sensitive-bench/extra-envs-rollouts/commit/2fc5692235d290668e220b3d279991466876c929", + "verified_files": 5, + "bytes": 355817914, + "verification": "Every file size; Git blob SHA1 for regular files and SHA256 for LFS files" + }, + { + "repo_id": "latency-sensitive-bench/extra-envs-rollouts", + "repo_type": "dataset", + "prefix": "profile-latency-new/air_raid/20260911T054920Z-airraid/converted", + "commit": "f461b237f64cd1c9f108c291f6a1b194b043da6d", + "url": "https://huggingface.co/datasets/latency-sensitive-bench/extra-envs-rollouts/commit/f461b237f64cd1c9f108c291f6a1b194b043da6d", + "verified_files": 116, + "bytes": 354405435, + "verification": "Every file size; Git blob SHA1 for regular files and SHA256 for LFS files" + } + ], + "data_validation": { + "episodes": [ + { + "split": "train", + "episode_idx": 0, + "seed": 0, + "actual_rows": 900, + "actual_return": 2900.0 + }, + { + "split": "train", + "episode_idx": 2, + "seed": 2, + "actual_rows": 900, + "actual_return": 3225.0 + }, + { + "split": "train", + "episode_idx": 4, + "seed": 4, + "actual_rows": 900, + "actual_return": 2950.0 + }, + { + "split": "train", + "episode_idx": 5, + "seed": 5, + "actual_rows": 900, + "actual_return": 3200.0 + }, + { + "split": "train", + "episode_idx": 6, + "seed": 6, + "actual_rows": 900, + "actual_return": 3150.0 + }, + { + "split": "train", + "episode_idx": 8, + "seed": 8, + "actual_rows": 900, + "actual_return": 3225.0 + }, + { + "split": "train", + "episode_idx": 9, + "seed": 9, + "actual_rows": 900, + "actual_return": 3175.0 + }, + { + "split": "train", + "episode_idx": 10, + "seed": 10, + "actual_rows": 900, + "actual_return": 3225.0 + }, + { + "split": "train", + "episode_idx": 11, + "seed": 11, + "actual_rows": 900, + "actual_return": 2975.0 + }, + { + "split": "train", + "episode_idx": 12, + "seed": 12, + "actual_rows": 900, + "actual_return": 3175.0 + }, + { + "split": "train", + "episode_idx": 13, + "seed": 13, + "actual_rows": 900, + "actual_return": 2875.0 + }, + { + "split": "train", + "episode_idx": 14, + "seed": 14, + "actual_rows": 900, + "actual_return": 3300.0 + }, + { + "split": "train", + "episode_idx": 15, + "seed": 15, + "actual_rows": 900, + "actual_return": 3275.0 + }, + { + "split": "train", + "episode_idx": 16, + "seed": 16, + "actual_rows": 900, + "actual_return": 3200.0 + }, + { + "split": "train", + "episode_idx": 18, + "seed": 18, + "actual_rows": 900, + "actual_return": 2975.0 + }, + { + "split": "train", + "episode_idx": 19, + "seed": 19, + "actual_rows": 900, + "actual_return": 3050.0 + }, + { + "split": "train", + "episode_idx": 20, + "seed": 20, + "actual_rows": 900, + "actual_return": 2650.0 + }, + { + "split": "train", + "episode_idx": 21, + "seed": 21, + "actual_rows": 900, + "actual_return": 3350.0 + }, + { + "split": "train", + "episode_idx": 22, + "seed": 22, + "actual_rows": 900, + "actual_return": 3550.0 + }, + { + "split": "train", + "episode_idx": 23, + "seed": 23, + "actual_rows": 900, + "actual_return": 3225.0 + }, + { + "split": "train", + "episode_idx": 24, + "seed": 24, + "actual_rows": 900, + "actual_return": 3150.0 + }, + { + "split": "train", + "episode_idx": 26, + "seed": 26, + "actual_rows": 900, + "actual_return": 3300.0 + }, + { + "split": "train", + "episode_idx": 27, + "seed": 27, + "actual_rows": 900, + "actual_return": 3625.0 + }, + { + "split": "train", + "episode_idx": 28, + "seed": 28, + "actual_rows": 900, + "actual_return": 3450.0 + }, + { + "split": "train", + "episode_idx": 30, + "seed": 30, + "actual_rows": 900, + "actual_return": 2850.0 + }, + { + "split": "train", + "episode_idx": 31, + "seed": 31, + "actual_rows": 900, + "actual_return": 3275.0 + }, + { + "split": "train", + "episode_idx": 32, + "seed": 32, + "actual_rows": 900, + "actual_return": 3450.0 + }, + { + "split": "train", + "episode_idx": 33, + "seed": 33, + "actual_rows": 900, + "actual_return": 3500.0 + }, + { + "split": "train", + "episode_idx": 34, + "seed": 34, + "actual_rows": 900, + "actual_return": 3350.0 + }, + { + "split": "train", + "episode_idx": 35, + "seed": 35, + "actual_rows": 900, + "actual_return": 3150.0 + }, + { + "split": "train", + "episode_idx": 36, + "seed": 36, + "actual_rows": 900, + "actual_return": 2950.0 + }, + { + "split": "train", + "episode_idx": 37, + "seed": 37, + "actual_rows": 900, + "actual_return": 2825.0 + }, + { + "split": "train", + "episode_idx": 38, + "seed": 38, + "actual_rows": 900, + "actual_return": 3225.0 + }, + { + "split": "train", + "episode_idx": 39, + "seed": 39, + "actual_rows": 900, + "actual_return": 3175.0 + }, + { + "split": "train", + "episode_idx": 40, + "seed": 40, + "actual_rows": 900, + "actual_return": 3050.0 + }, + { + "split": "train", + "episode_idx": 41, + "seed": 41, + "actual_rows": 900, + "actual_return": 3025.0 + }, + { + "split": "train", + "episode_idx": 42, + "seed": 42, + "actual_rows": 900, + "actual_return": 3575.0 + }, + { + "split": "train", + "episode_idx": 43, + "seed": 43, + "actual_rows": 900, + "actual_return": 3625.0 + }, + { + "split": "train", + "episode_idx": 44, + "seed": 44, + "actual_rows": 900, + "actual_return": 2675.0 + }, + { + "split": "train", + "episode_idx": 45, + "seed": 45, + "actual_rows": 900, + "actual_return": 3075.0 + }, + { + "split": "train", + "episode_idx": 46, + "seed": 46, + "actual_rows": 900, + "actual_return": 3450.0 + }, + { + "split": "train", + "episode_idx": 48, + "seed": 48, + "actual_rows": 900, + "actual_return": 3400.0 + }, + { + "split": "train", + "episode_idx": 49, + "seed": 49, + "actual_rows": 900, + "actual_return": 3050.0 + }, + { + "split": "train", + "episode_idx": 50, + "seed": 50, + "actual_rows": 900, + "actual_return": 3200.0 + }, + { + "split": "train", + "episode_idx": 51, + "seed": 51, + "actual_rows": 900, + "actual_return": 3100.0 + }, + { + "split": "train", + "episode_idx": 52, + "seed": 52, + "actual_rows": 900, + "actual_return": 3475.0 + }, + { + "split": "train", + "episode_idx": 53, + "seed": 53, + "actual_rows": 900, + "actual_return": 2875.0 + }, + { + "split": "train", + "episode_idx": 54, + "seed": 54, + "actual_rows": 900, + "actual_return": 3200.0 + }, + { + "split": "train", + "episode_idx": 55, + "seed": 55, + "actual_rows": 900, + "actual_return": 3350.0 + }, + { + "split": "train", + "episode_idx": 56, + "seed": 56, + "actual_rows": 900, + "actual_return": 3400.0 + }, + { + "split": "train", + "episode_idx": 57, + "seed": 57, + "actual_rows": 900, + "actual_return": 3375.0 + }, + { + "split": "train", + "episode_idx": 59, + "seed": 59, + "actual_rows": 900, + "actual_return": 3575.0 + }, + { + "split": "train", + "episode_idx": 60, + "seed": 60, + "actual_rows": 900, + "actual_return": 2575.0 + }, + { + "split": "train", + "episode_idx": 61, + "seed": 61, + "actual_rows": 900, + "actual_return": 2750.0 + }, + { + "split": "train", + "episode_idx": 62, + "seed": 62, + "actual_rows": 900, + "actual_return": 3250.0 + }, + { + "split": "train", + "episode_idx": 63, + "seed": 63, + "actual_rows": 900, + "actual_return": 3000.0 + }, + { + "split": "train", + "episode_idx": 64, + "seed": 64, + "actual_rows": 900, + "actual_return": 3025.0 + }, + { + "split": "train", + "episode_idx": 65, + "seed": 65, + "actual_rows": 900, + "actual_return": 2725.0 + }, + { + "split": "train", + "episode_idx": 66, + "seed": 66, + "actual_rows": 900, + "actual_return": 3350.0 + }, + { + "split": "train", + "episode_idx": 67, + "seed": 67, + "actual_rows": 900, + "actual_return": 2975.0 + }, + { + "split": "train", + "episode_idx": 68, + "seed": 68, + "actual_rows": 900, + "actual_return": 3425.0 + }, + { + "split": "train", + "episode_idx": 69, + "seed": 69, + "actual_rows": 900, + "actual_return": 3225.0 + }, + { + "split": "train", + "episode_idx": 70, + "seed": 70, + "actual_rows": 900, + "actual_return": 3050.0 + }, + { + "split": "train", + "episode_idx": 71, + "seed": 71, + "actual_rows": 900, + "actual_return": 3250.0 + }, + { + "split": "train", + "episode_idx": 72, + "seed": 72, + "actual_rows": 900, + "actual_return": 3150.0 + }, + { + "split": "train", + "episode_idx": 73, + "seed": 73, + "actual_rows": 900, + "actual_return": 3100.0 + }, + { + "split": "train", + "episode_idx": 74, + "seed": 74, + "actual_rows": 900, + "actual_return": 2875.0 + }, + { + "split": "train", + "episode_idx": 75, + "seed": 75, + "actual_rows": 900, + "actual_return": 3075.0 + }, + { + "split": "train", + "episode_idx": 76, + "seed": 76, + "actual_rows": 900, + "actual_return": 2950.0 + }, + { + "split": "train", + "episode_idx": 78, + "seed": 78, + "actual_rows": 900, + "actual_return": 3700.0 + }, + { + "split": "train", + "episode_idx": 79, + "seed": 79, + "actual_rows": 900, + "actual_return": 3175.0 + }, + { + "split": "train", + "episode_idx": 80, + "seed": 80, + "actual_rows": 900, + "actual_return": 2950.0 + }, + { + "split": "train", + "episode_idx": 82, + "seed": 82, + "actual_rows": 900, + "actual_return": 3200.0 + }, + { + "split": "train", + "episode_idx": 83, + "seed": 83, + "actual_rows": 900, + "actual_return": 3150.0 + }, + { + "split": "train", + "episode_idx": 84, + "seed": 84, + "actual_rows": 900, + "actual_return": 3050.0 + }, + { + "split": "train", + "episode_idx": 85, + "seed": 85, + "actual_rows": 900, + "actual_return": 3275.0 + }, + { + "split": "train", + "episode_idx": 86, + "seed": 86, + "actual_rows": 900, + "actual_return": 3275.0 + }, + { + "split": "train", + "episode_idx": 87, + "seed": 87, + "actual_rows": 900, + "actual_return": 3125.0 + }, + { + "split": "train", + "episode_idx": 88, + "seed": 88, + "actual_rows": 900, + "actual_return": 3075.0 + }, + { + "split": "train", + "episode_idx": 89, + "seed": 89, + "actual_rows": 900, + "actual_return": 3300.0 + }, + { + "split": "train", + "episode_idx": 90, + "seed": 90, + "actual_rows": 900, + "actual_return": 2850.0 + }, + { + "split": "train", + "episode_idx": 91, + "seed": 91, + "actual_rows": 900, + "actual_return": 3350.0 + }, + { + "split": "train", + "episode_idx": 92, + "seed": 92, + "actual_rows": 900, + "actual_return": 3200.0 + }, + { + "split": "train", + "episode_idx": 93, + "seed": 93, + "actual_rows": 900, + "actual_return": 3350.0 + }, + { + "split": "train", + "episode_idx": 94, + "seed": 94, + "actual_rows": 900, + "actual_return": 3125.0 + }, + { + "split": "train", + "episode_idx": 95, + "seed": 95, + "actual_rows": 900, + "actual_return": 3025.0 + }, + { + "split": "train", + "episode_idx": 96, + "seed": 96, + "actual_rows": 900, + "actual_return": 3800.0 + }, + { + "split": "train", + "episode_idx": 97, + "seed": 97, + "actual_rows": 900, + "actual_return": 3175.0 + }, + { + "split": "train", + "episode_idx": 98, + "seed": 98, + "actual_rows": 900, + "actual_return": 3025.0 + }, + { + "split": "train", + "episode_idx": 99, + "seed": 99, + "actual_rows": 900, + "actual_return": 3525.0 + }, + { + "split": "val", + "episode_idx": 1, + "seed": 1, + "actual_rows": 900, + "actual_return": 2900.0 + }, + { + "split": "val", + "episode_idx": 3, + "seed": 3, + "actual_rows": 900, + "actual_return": 3225.0 + }, + { + "split": "val", + "episode_idx": 7, + "seed": 7, + "actual_rows": 900, + "actual_return": 3375.0 + }, + { + "split": "val", + "episode_idx": 17, + "seed": 17, + "actual_rows": 900, + "actual_return": 3400.0 + }, + { + "split": "val", + "episode_idx": 25, + "seed": 25, + "actual_rows": 900, + "actual_return": 3350.0 + }, + { + "split": "val", + "episode_idx": 29, + "seed": 29, + "actual_rows": 900, + "actual_return": 3400.0 + }, + { + "split": "val", + "episode_idx": 47, + "seed": 47, + "actual_rows": 900, + "actual_return": 3125.0 + }, + { + "split": "val", + "episode_idx": 58, + "seed": 58, + "actual_rows": 900, + "actual_return": 2925.0 + }, + { + "split": "val", + "episode_idx": 77, + "seed": 77, + "actual_rows": 900, + "actual_return": 3725.0 + }, + { + "split": "val", + "episode_idx": 81, + "seed": 81, + "actual_rows": 900, + "actual_return": 3375.0 + } + ], + "total_episodes": 100, + "total_rows": 90000, + "pass_gate_count": 100, + "mean_return": 3186.5, + "minimum_return": 2575.0, + "probe_train_rows": 81000, + "probe_val_rows": 9000, + "actual_train_rows": 81000, + "actual_val_rows": 9000, + "per_episode_probe_replay_match": true + }, + "code_provenance": { + "root_revision": "86df5ecef6735d5d824944e7acdee5bf204e1545", + "task_code_archive_sha256": "b0a0ca72b5c6eb65c16e5d5d70f59cad540260a27cb3a71109910326b940e135", + "iid_patch_sha256": "ceae6068634e4df2a553d8aee2cb70f7ef0640265330c65f75b9462b449010d0", + "sf_teacher_runtime_sha256": "2e4ba33a1c406bf3fa9b122f1cc66c13f1b0d7a96c89ec142e0428f184b4646d", + "sample_factory_revision": "4b7277842b17804fb928097a9689d889bd2f5cdc", + "starvla_revision": "f364fdf080434aea14bb2a19931e017f0af32d87", + "note": "Task code extracted from pinned root plus tracked IID runtime fix; archive runtime git metadata is unavailable." + }, + "normalization_source": "new Air Raid converted dataset and actual VLA training dataset_statistics.json" +} diff --git a/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/task_contract.json b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/task_contract.json new file mode 100644 index 0000000000000000000000000000000000000000..f20338db89c82f965cce12d795b3641d45cdae58 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenoft-h1/air_raid_profile_20260911T054920Z_openvla_native_sft_5k/task_contract.json @@ -0,0 +1,44 @@ +{ + "action_labels": [ + "noop", + "fire", + "right", + "left", + "rightfire", + "leftfire" + ], + "action_values": [ + 0, + 1, + 2, + 3, + 4, + 5 + ], + "base_prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.", + "env_fps": 50.0, + "env_id": "LatencyBench/AirRaid-v0", + "frame_stack": 4, + "make_kwargs": { + "base_env_id": "ALE/AirRaid-v5", + "base_make_kwargs": { + "difficulty": 0, + "frameskip": 1, + "full_action_space": false, + "max_num_frames_per_episode": 108000, + "mode": 1, + "obs_type": "rgb", + "repeat_action_probability": 0.0 + }, + "fire_reset": true, + "noop_max": 30, + "render_mode": "rgb_array", + "screen_size": 84 + }, + "noop_action_id": 0, + "obs_fps": 12.5, + "registration_imports": [ + "latency_bench.envs.gymnasium_air_raid" + ], + "task_name": "air_raid" +} diff --git a/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/README.md b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/README.md new file mode 100644 index 0000000000000000000000000000000000000000..3cfabca6246c35c2ace38ca6f697044a5aee3dc6 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/README.md @@ -0,0 +1,31 @@ +# air-raid / qwenpi_v3 + +Training condition: `latency-aware`. Run: `airraid_pi05_profile_h1_10env_2p5obs_g128_20260920`. + +Copied from [Standard-Pipeline](https://huggingface.co/latency-sensitive-bench/Standard-Pipeline/tree/2208875f92b21e3f8ec2363a30a6705dfe56cc04/air_raid/profile_latency/Pi05) at `2208875f92b21e3f8ec2363a30a6705dfe56cc04`. + +- Checkpoint: `checkpoints/model.pt` +- Selection: only published VLA checkpoint +- Checkpoint SHA256: `7914e1f1e2ba0b78d29c15dd33d96bbb6a5f61bad6ebb17f583a6d6d9958fb14` +- Configuration, FPS, actions, normalization and latency semantics retain their source values. +- This import does not run evaluations or change training/evaluation/acceptance status. + +Action horizon is 1. Use `config.yaml` and `dataset_statistics.json` with the matching +prepared backbone. `task_contract.json` is extracted from the original training configuration. +Original machine paths in historical configurations are provenance: bind checkpoint, +backbone, manifest and profile paths to downloaded local files when running inference. + +See [the original model notes](source/README.md) for experiment results and limitations. + +[Original provenance](source/provenance.json) retains dataset references and recorded experiment evidence. + +## Download + +```python +from huggingface_hub import snapshot_download +prefix = "latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920" +snapshot_download("latency-sensitive-bench/benchmark-models", revision="", + allow_patterns=[prefix + "/**"]) +``` + +Use the verified import commit listed in the repository's import report. diff --git a/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/checkpoints/model.pt b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/checkpoints/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..b2cbfa6890a31c2550b098a96925bd0eadbbae53 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/checkpoints/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:7914e1f1e2ba0b78d29c15dd33d96bbb6a5f61bad6ebb17f583a6d6d9958fb14 +size 10920516461 diff --git a/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/config.full.yaml b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/config.full.yaml new file mode 100644 index 0000000000000000000000000000000000000000..b65e0cc10b5ff468d00433f94db80aac8b5d580a --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/config.full.yaml @@ -0,0 +1,233 @@ +framework: + name: QwenPI_v3 + qwenvl: + base_vlm: ${oc.env:PI05_BACKBONE_DIR} + attn_implementation: flash_attention_2 + vl_hidden_dim: 2560 + num_vl_layers: 36 + enable_gradient_checkpointing: false + action_model: + action_model_type: LayerwiseFM + action_dim: 6 + state_dim: 0 + state_encoding: continuous_projector + task_objective: null + action_horizon: 1 + repeated_diffusion_steps: 2 + num_inference_timesteps: 4 + add_pos_embed: true + max_seq_len: 1024 + num_target_vision_tokens: 32 + noise_beta_alpha: 1.5 + noise_beta_beta: 1.0 + noise_s: 0.999 + num_timestep_buckets: 1000 + diffusion_model_cfg: + action_dit_hidden_dim: 1024 + dropout: 0.2 + final_dropout: true + interleave_self_attention: true + norm_type: ada_norm + positional_embeddings: null + attention_head_dim: 64 + num_layers: 36 + input_embedding_dim: 1024 + cross_attention_dim: 1024 + output_dim: 1024 + num_attention_heads: 16 + action_env_dim: 6 + future_action_window_size: 0 + past_action_window_size: 0 +datasets: + vla_data: + dataset_py: lerobot_datasets + include_state: false + data_root_dir: / + data_mix: train + eval_data_mix: validation + custom_mixtures_path: ${PI05_RUN_DIR}/profile_latency/vla/mixture.json + action_type: discrete + sequential_step_sampling: true + eval_sequential_step_sampling: null + num_workers: 4 + eval_num_workers: 0 + prefetch_factor: 2 + persistent_workers: true + pin_memory: true + shuffle: true + action_balance: + enabled: false + strategy: balanced_epoch + action_key: action_id + target_flap_fraction: 0.3 + noop_id: 0 + flap_id: 1 + latency_curriculum: + enabled: false + strategy: exclusive + latencies: null + phase_steps: null + phase_distributions: null + new_latency_passes: 1.0 + replay_passes: 0.25 + target_total_passes: 2.0 + final_equalization: true + step_budget_mode: auto + eval_at_phase_end: false + save_at_phase_end: false + computed_plan: null + per_device_batch_size: 64 + load_all_data_for_training: true + num_obs_frames: 1 + image_mode: single + prompt_mode: raw + stitch_grid: + - 2 + - 2 + obs_image_size: + - 224 + - 224 + video_backend: pyav + lerobot_version: v2.0 + gymnasium_task_contract: + action_labels: + - noop + - fire + - right + - left + - rightfire + - leftfire + action_values: + - 0 + - 1 + - 2 + - 3 + - 4 + - 5 + base_prompt: 'Protect both buildings from flying saucers. Choose exactly one + action from: noop, fire, right, left, rightfire, leftfire.' + env_fps: 10.0 + env_id: LatencyBench/AirRaid-v0 + frame_stack: 4 + make_kwargs: + base_env_id: ALE/AirRaid-v5 + base_make_kwargs: + difficulty: 0 + frameskip: 1 + full_action_space: false + max_num_frames_per_episode: 108000 + mode: 1 + obs_type: rgb + repeat_action_probability: 0.0 + fire_reset: true + noop_max: 30 + render_mode: rgb_array + screen_size: 84 + noop_action_id: 0 + obs_fps: 2.5 + registration_imports: + - latency_bench.envs.gymnasium_air_raid + task_name: air_raid + active_action_dim: 6 +dataset: + source_hf: '' + config_name: null + source_subdir: null + converted_name: flappy_train + single_source_hf: '' + mixed_source_hf: '' + single_converted_name: flappy_train + mixed_converted_name: flappy_mixed_latency_train + single_latency_filter: null + mixed_latency_filter: null + force_download: false + setup_force: false + skip_verification: false + target_latency_unit: observation_steps + verify_rows: 200 + max_episodes: null + episodes_per_latency: null + latency_filter: null + debug_subset: + enabled: false + max_episodes: 5 + suffix: debug +base_model: + repo_id: Qwen/Qwen3-VL-4B-Instruct +initialization: + checkpoint_local_dir: null + checkpoint_hf_repo_id: null + checkpoint_filename: null +trainer: + max_train_steps: 5000 + num_warmup_steps: 100 + save_interval: 500 + eval_interval: 500 + eval_num_batches: 200 + per_latency_eval_num_batches: null + eval_action_classification: false + eval_action_classification_interval: null + cc_f1_tolerance: 1 + learning_rate: + base: 2.0e-05 + qwen_vl_interface: 1.0e-05 + action_model: 0.0001 + lr_scheduler_type: cosine_with_min_lr + scheduler_specific_kwargs: + min_lr: 1.0e-06 + freeze_modules: '' + freeze_vit: false + freeze_tied_embedding: false + freeze_llm_layers: [] + loss_scale: + vla: 1.0 + vlm: 0.1 + max_grad_norm: 1.0 + weight_decay: 0.0 + logging_frequency: 1 + profile_timing: + enabled: false + log_interval: 10 + gradient_clipping: 1.0 + gradient_accumulation_steps: 1 + distributed_backend: deepspeed + is_resume: false + pretrained_checkpoint: null + resume_step: 0 + reload_modules: null + optimizer: + name: AdamW + betas: + - 0.9 + - 0.95 + eps: 1.0e-08 + weight_decay: 1.0e-08 + fused: true + save_format: pt +run_id: airraid_pi05_profile_h1_10env_2p5obs_g128_20260920 +run_root_dir: ${PI05_RUN_DIR}/profile_latency/vla/training +seed: 42 +is_debug: false +version_id: '0.21' +wandb_project: latency-sensitive-bench +wandb_entity: dongqianyu99-zhejiang-university +wandb_name: airraid_pi05_profile_h1_10env_2p5obs_g128_20260920 +wandb_group: pi05-seven-env +wandb_tags: +- air_raid +- profile_latency +- Pi05 +- h1 +training_latency_condition: profile_latency +backbone: + repo_id: Qwen/Qwen3-VL-4B-Instruct + revision: ebb281ec70b05090aa6165b016eac8ec08e71b17 +checkpoint: + save_best_model: false + save_final_model: true + save_pt_file: true + save_training_state: true + local: + keep_last_n: 1 +config_yaml: ${PI05_RUN_DIR}/profile_latency/train.yaml +output_dir: ${PI05_RUN_DIR}/profile_latency/vla/training/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920 diff --git a/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/config.yaml b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..d309f5c8ebddcd49517fb1eb9b2d54e9ff606291 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/config.yaml @@ -0,0 +1,110 @@ +framework: + name: QwenPI_v3 + qwenvl: + base_vlm: ${oc.env:PI05_BACKBONE_DIR} + attn_implementation: flash_attention_2 + vl_hidden_dim: 2560 + num_vl_layers: 36 + enable_gradient_checkpointing: false + action_model: + action_model_type: LayerwiseFM + action_dim: 6 + state_dim: 0 + state_encoding: continuous_projector + task_objective: null + action_horizon: 1 + repeated_diffusion_steps: 2 + num_inference_timesteps: 4 + add_pos_embed: true + max_seq_len: 1024 + num_target_vision_tokens: 32 + noise_beta_alpha: 1.5 + noise_beta_beta: 1.0 + noise_s: 0.999 + num_timestep_buckets: 1000 + diffusion_model_cfg: + action_dit_hidden_dim: 1024 + dropout: 0.2 + final_dropout: true + interleave_self_attention: true + norm_type: ada_norm + positional_embeddings: null + attention_head_dim: 64 + num_layers: 36 + input_embedding_dim: 1024 + cross_attention_dim: 1024 + output_dim: 1024 + num_attention_heads: 16 + action_env_dim: 6 + future_action_window_size: 0 + past_action_window_size: 0 +version_id: '0.21' +datasets: + vla_data: + include_state: false + obs_image_size: + - 224 + - 224 + image_mode: single + stitch_grid: + - 2 + - 2 + num_obs_frames: 1 + gymnasium_task_contract: &id001 + action_labels: + - noop + - fire + - right + - left + - rightfire + - leftfire + action_values: + - 0 + - 1 + - 2 + - 3 + - 4 + - 5 + base_prompt: 'Protect both buildings from flying saucers. Choose exactly one + action from: noop, fire, right, left, rightfire, leftfire.' + env_fps: 10.0 + env_id: LatencyBench/AirRaid-v0 + frame_stack: 4 + make_kwargs: + base_env_id: ALE/AirRaid-v5 + base_make_kwargs: + difficulty: 0 + frameskip: 1 + full_action_space: false + max_num_frames_per_episode: 108000 + mode: 1 + obs_type: rgb + repeat_action_probability: 0.0 + fire_reset: true + noop_max: 30 + render_mode: rgb_array + screen_size: 84 + noop_action_id: 0 + obs_fps: 2.5 + registration_imports: + - latency_bench.envs.gymnasium_air_raid + task_name: air_raid +trainer: + pretrained_checkpoint: null + profile_timing: + enabled: false + log_interval: 10 +backbone: + repo_id: Qwen/Qwen3-VL-4B-Instruct + revision: ebb281ec70b05090aa6165b016eac8ec08e71b17 +rl_games: + model_alias: pi05 + task: gymnasium + action_carrier: native + gymnasium: + task_contract: *id001 + env_eval: + image_size: 224 + frameskip: 1 + image_transform: raw_rgb + prompt_mode: raw diff --git a/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/dataset_statistics.json b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/dataset_statistics.json new file mode 100644 index 0000000000000000000000000000000000000000..17ec6d143230f517be9c23f7c47e5aabfd7589ac --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/dataset_statistics.json @@ -0,0 +1,84 @@ +{ + "new_embodiment": { + "action": { + "mean": [ + 0.1809876561164856, + 0.37079012393951416, + 0.052975308150053024, + 0.208790123462677, + 0.0640740767121315, + 0.12238271534442902 + ], + "std": [ + 0.38516250252723694, + 0.4830116331577301, + 0.22386549413204193, + 0.4065310060977936, + 0.24487093091011047, + 0.3277742564678192 + ], + "max": [ + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0 + ], + "min": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "q01": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "q99": [ + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0 + ], + "mask": [ + true, + true, + true, + true, + true, + true + ] + }, + "state": { + "mean": [ + 0.0 + ], + "std": [ + 0.0 + ], + "max": [ + 0.0 + ], + "min": [ + 0.0 + ], + "q01": [ + 0.0 + ], + "q99": [ + 0.0 + ] + }, + "num_transitions": 81000, + "num_trajectories": 90 + } +} \ No newline at end of file diff --git a/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/evaluation/comparison.json b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/evaluation/comparison.json new file mode 100644 index 0000000000000000000000000000000000000000..b38e36e2589bebdba948982cde8e47d4b2413d19 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/evaluation/comparison.json @@ -0,0 +1,233 @@ +{ + "state": "MATCHED_B_F20_AUDIT_PASSED", + "B": { + "model_revision": "de49831bd5dc9d642058bff0237e6ff2c9127b53", + "checkpoint_sha256": "c9ad50c5229424c919bb108d75b016f3b1815405ebd03adbea08137c18849bd8", + "seeds": [ + 42, + 43, + 44, + 45, + 46, + 47, + 48, + 49, + 50, + 51, + 52, + 53, + 54, + 55, + 56, + 57, + 58, + 59, + 60, + 61 + ], + "returns": [ + 1900.0, + 1300.0, + 1625.0, + 1800.0, + 450.0, + 1400.0, + 1775.0, + 1450.0, + 775.0, + 1325.0, + 800.0, + 775.0, + 1400.0, + 1425.0, + 1650.0, + 1500.0, + 1225.0, + 2100.0, + 1025.0, + 2025.0 + ], + "mean_return": 1386.25, + "population_std_return": 434.9910200222529 + }, + "F": { + "model_revision": "ac837400f55e4a9a96cc757dbcc516711b37e75f", + "checkpoint_sha256": "7914e1f1e2ba0b78d29c15dd33d96bbb6a5f61bad6ebb17f583a6d6d9958fb14", + "seeds": [ + 42, + 43, + 44, + 45, + 46, + 47, + 48, + 49, + 50, + 51, + 52, + 53, + 54, + 55, + 56, + 57, + 58, + 59, + 60, + 61 + ], + "returns": [ + 4125.0, + 4000.0, + 4125.0, + 4250.0, + 900.0, + 4125.0, + 3400.0, + 3775.0, + 3450.0, + 2675.0, + 2400.0, + 3775.0, + 350.0, + 2525.0, + 4125.0, + 3250.0, + 4125.0, + 1800.0, + 3775.0, + 3600.0 + ], + "mean_return": 3227.5, + "population_std_return": 1095.5848894540304 + }, + "paired": [ + { + "seed": 42, + "B": 1900.0, + "F": 4125.0, + "delta": 2225.0 + }, + { + "seed": 43, + "B": 1300.0, + "F": 4000.0, + "delta": 2700.0 + }, + { + "seed": 44, + "B": 1625.0, + "F": 4125.0, + "delta": 2500.0 + }, + { + "seed": 45, + "B": 1800.0, + "F": 4250.0, + "delta": 2450.0 + }, + { + "seed": 46, + "B": 450.0, + "F": 900.0, + "delta": 450.0 + }, + { + "seed": 47, + "B": 1400.0, + "F": 4125.0, + "delta": 2725.0 + }, + { + "seed": 48, + "B": 1775.0, + "F": 3400.0, + "delta": 1625.0 + }, + { + "seed": 49, + "B": 1450.0, + "F": 3775.0, + "delta": 2325.0 + }, + { + "seed": 50, + "B": 775.0, + "F": 3450.0, + "delta": 2675.0 + }, + { + "seed": 51, + "B": 1325.0, + "F": 2675.0, + "delta": 1350.0 + }, + { + "seed": 52, + "B": 800.0, + "F": 2400.0, + "delta": 1600.0 + }, + { + "seed": 53, + "B": 775.0, + "F": 3775.0, + "delta": 3000.0 + }, + { + "seed": 54, + "B": 1400.0, + "F": 350.0, + "delta": -1050.0 + }, + { + "seed": 55, + "B": 1425.0, + "F": 2525.0, + "delta": 1100.0 + }, + { + "seed": 56, + "B": 1650.0, + "F": 4125.0, + "delta": 2475.0 + }, + { + "seed": 57, + "B": 1500.0, + "F": 3250.0, + "delta": 1750.0 + }, + { + "seed": 58, + "B": 1225.0, + "F": 4125.0, + "delta": 2900.0 + }, + { + "seed": 59, + "B": 2100.0, + "F": 1800.0, + "delta": -300.0 + }, + { + "seed": 60, + "B": 1025.0, + "F": 3775.0, + "delta": 2750.0 + }, + { + "seed": 61, + "B": 2025.0, + "F": 3600.0, + "delta": 1575.0 + } + ], + "mean_gain": 1841.25, + "gain_percent": 132.82236248872857, + "wins": 18, + "ties": 0, + "losses": 2, + "protocol": "RTX3090/10env-2.5obsFPS/measured-only/seed42-61/cap3600/no-state/discrete6/RGB224/H1; Bprompt0/Fprompt2", + "causal_scope": "Same benchmark protocol, different realized inference latency distributions; no equal-delay or architecture-ranking claim", + "scientific_evidence_url": "https://huggingface.co/datasets/latency-sensitive-bench/Standard-Pipeline/blob/b451caa2a6515df3bae22580595151ef5b342f89/air_raid/profile_latency/Pi05/h1_10env_2p5obs_g128_20260920/vla/evaluation/comparison.json" +} diff --git a/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/evaluation/status.json b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/evaluation/status.json new file mode 100644 index 0000000000000000000000000000000000000000..64915265ab847adb2fa81d94e7636102e084ea7a --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/evaluation/status.json @@ -0,0 +1,10 @@ +{ + "training": "COMPLETED_5000", + "saved_reload": "PASSED", + "measured_B20": "AUDITED", + "measured_F20": "AUDITED", + "evaluated_model_revision": "ac837400f55e4a9a96cc757dbcc516711b37e75f", + "checkpoint_sha256": "7914e1f1e2ba0b78d29c15dd33d96bbb6a5f61bad6ebb17f583a6d6d9958fb14", + "weights_unchanged": true, + "F_evidence_revision": "b451caa2a6515df3bae22580595151ef5b342f89" +} diff --git a/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/latency_prompt_map.json b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/latency_prompt_map.json new file mode 100644 index 0000000000000000000000000000000000000000..179f499c32f7e2b8a91a375c53a506e620280500 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/latency_prompt_map.json @@ -0,0 +1,7 @@ +{ + "2": { + "prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire. Current action latency is 2 raw frames (200.00 ms). The environment runs at 10 FPS and observations are emitted at 2.5 FPS. Choose the best next action.", + "latency_raw_frames": 2, + "latency_ms": 200.0 + } +} \ No newline at end of file diff --git a/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/manifest.json b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..f7afa3c26ff65cf281b2bc64c47f55cc0fddb7e2 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/manifest.json @@ -0,0 +1,92 @@ +{ + "dataset_name": "air_raid_h1_10_2p5_profile", + "env_name": "air_raid", + "episodes": 90, + "frames": 81000, + "task_prompts": [ + "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire. Current action latency is 2 raw frames (200.00 ms). The environment runs at 10 FPS and observations are emitted at 2.5 FPS. Choose the best next action." + ], + "format": "starvla_lerobot_v2_image_parquet", + "source": "${PI05_RUN_DIR}/profile_latency/data/raw_profile", + "integration_name": "gymnasium", + "task_name": "air_raid", + "action_layout": "gymnasium_discrete_v1", + "action_labels": [ + "noop", + "fire", + "right", + "left", + "rightfire", + "leftfire" + ], + "carrier_action_labels": [ + "noop", + "fire", + "right", + "left", + "rightfire", + "leftfire" + ], + "action_dim": 6, + "active_action_dim": 6, + "action_carrier": "native", + "rows_unit": "decision_step", + "obs_fps": 2.5, + "obs_stride_raw_frames": 4, + "uses_state": false, + "state_dim": 1, + "state_labels": [ + "state" + ], + "state_normalization": null, + "robot_type": "rl_games_gymnasium", + "latency_prompt_map_path": "${PI05_RUN_DIR}/profile_latency/data/lerobot/air_raid_h1_10_2p5_profile/latency_prompt_map.json", + "custom_mixtures_path": "${PI05_RUN_DIR}/profile_latency/data/lerobot/_generated_mixtures/air_raid_h1_10_2p5_profile.json", + "gymnasium_task": { + "action_labels": [ + "noop", + "fire", + "right", + "left", + "rightfire", + "leftfire" + ], + "action_values": [ + 0, + 1, + 2, + 3, + 4, + 5 + ], + "base_prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.", + "env_fps": 10.0, + "env_id": "LatencyBench/AirRaid-v0", + "frame_stack": 4, + "make_kwargs": { + "base_env_id": "ALE/AirRaid-v5", + "base_make_kwargs": { + "difficulty": 0, + "frameskip": 1, + "full_action_space": false, + "max_num_frames_per_episode": 108000, + "mode": 1, + "obs_type": "rgb", + "repeat_action_probability": 0.0 + }, + "fire_reset": true, + "noop_max": 30, + "render_mode": "rgb_array", + "screen_size": 84 + }, + "noop_action_id": 0, + "obs_fps": 2.5, + "registration_imports": [ + "latency_bench.envs.gymnasium_air_raid" + ], + "task_name": "air_raid" + }, + "validation_dataset_name": "air_raid_h1_10_2p5_profile__val", + "validation_episodes": 10, + "validation_frames": 9000 +} \ No newline at end of file diff --git a/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/provenance.json b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/provenance.json new file mode 100644 index 0000000000000000000000000000000000000000..3495ff067bab26682c12cfbc6cf2582ce0cf0854 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/provenance.json @@ -0,0 +1,34 @@ +{ + "task": "air-raid", + "model": "qwenpi_v3", + "training_condition": "latency-aware", + "training_run_id": "airraid_pi05_profile_h1_10env_2p5obs_g128_20260920", + "source": { + "repo_id": "latency-sensitive-bench/Standard-Pipeline", + "revision": "2208875f92b21e3f8ec2363a30a6705dfe56cc04", + "path": "air_raid/profile_latency/Pi05", + "url": "https://huggingface.co/latency-sensitive-bench/Standard-Pipeline/tree/2208875f92b21e3f8ec2363a30a6705dfe56cc04/air_raid/profile_latency/Pi05" + }, + "hf_repo_id": "latency-sensitive-bench/benchmark-models", + "hf_path": "latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920", + "checkpoint": { + "source_file": "air_raid/profile_latency/Pi05/checkpoints/model.pt", + "source_sha256": "7914e1f1e2ba0b78d29c15dd33d96bbb6a5f61bad6ebb17f583a6d6d9958fb14", + "source_bytes": 10920516461, + "file": "checkpoints/model.pt", + "selection_rule": "only published VLA checkpoint", + "method": "copy", + "sha256": "7914e1f1e2ba0b78d29c15dd33d96bbb6a5f61bad6ebb17f583a6d6d9958fb14", + "bytes": 10920516461 + }, + "config_source": "air_raid/profile_latency/Pi05/config.yaml", + "task_contract_file": "task_contract.json", + "missing_source_files": [], + "migration": { + "operation": "copy_selected_inference_bundle", + "source_modified": false, + "evaluation_performed": false, + "experiment_status": "preserved_in_source_evidence" + }, + "action_horizon": 1 +} diff --git a/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/source/README.md b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/source/README.md new file mode 100644 index 0000000000000000000000000000000000000000..a8666af52d05e56e895a3b3a35f002a445d29066 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/source/README.md @@ -0,0 +1,7 @@ +# AirRaid Pi0.5 H1 — profile-trained final F + +QwenPI_v3 with pinned Qwen3-VL-4B-Instruct, fresh action head, 5,000 updates, seed42, global128=micro64×accumulation1×2, GC disabled and ZeRO-2. RGB224, no state, six native discrete actions, H1. Environment/observation10/2.5FPS, measured-only RTX3090 evaluation,20 paired seeds42–61, cap3600 rawframes, one worker and hold/FIFO. B prompt0, F prompt2; no added IID delay. + +Audited B return 1386.25 ± 434.99; F 3227.50 ± 1095.58 (population SD). Mean paired gain 1841.25 (132.82%); 18 wins/0 ties/2 losses. Realized inference latency differs; no equal-delay causal claim or architecture ranking. + +[Fixed final scientific evidence](https://huggingface.co/datasets/latency-sensitive-bench/Standard-Pipeline/blob/b451caa2a6515df3bae22580595151ef5b342f89/air_raid/profile_latency/Pi05/h1_10env_2p5obs_g128_20260920/vla/evaluation/comparison.json). Original evaluated checkpoint SHA256 `7914e1f1e2ba0b78d29c15dd33d96bbb6a5f61bad6ebb17f583a6d6d9958fb14`, evaluated model revision `ac837400f55e4a9a96cc757dbcc516711b37e75f`. This metadata update does not change weights. The original P5×10 was retained. APPO teacher target10M counted steps corresponds toabout40M rawframes atstride4; no teacherreward gate was applied toVLA evaluation. diff --git a/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/source/provenance.json b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/source/provenance.json new file mode 100644 index 0000000000000000000000000000000000000000..c1e0821554c775b44ddec3430f49d380f840c223 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/source/provenance.json @@ -0,0 +1,675 @@ +{ + "task": "air_raid", + "backbone": { + "repo_id": "Qwen/Qwen3-VL-4B-Instruct", + "revision": "ebb281ec70b05090aa6165b016eac8ec08e71b17" + }, + "checkpoint": "steps_5000_pytorch_model.pt", + "training_steps": 5000, + "seed": 42, + "global_batch": 128, + "training_run_id": "airraid_pi05_profile_h1_10env_2p5obs_g128_20260920", + "condition": "profile_latency", + "source": { + "task": "air_raid", + "source_code": { + "archive_sha256": { + "lsb-pi05-h1-20260918.tar.gz": "08989cd9c3518cd79cb5ec065300755b8afe7555e1b6d65fed829549ef9127c4", + "pi05-training-package.tar.gz": "0e1c6f7cceb8979b525ceaa3c209f005049166b97eb97a1602e13ed8d7384b65" + }, + "original_code_root": "2de3816f07ecea4002964d046c4dfa581a658671", + "starvla": "3430c45edf6a08e4cfcaa2fce58bb4ea05995617", + "sample_factory": "4b11d8754ffc989cdd15a340cfc9be0277377d5b", + "task_overlay": "batch128 and AirRaid manifest clock; recorded separately" + }, + "code_overlay": { + "files": { + "scripts/gym_adapt/gr00t.py": "07fddb83f5a7f0766fcbf025c521703fe295ee3754017187c5ac4b66e1bcdce1", + "tests/data/test_starvla_tasks.py": "4fd24cea648baa1b3af95fddd3311610a80fd346cb633a83bb6f94ea6423c797" + }, + "tests": "16 relevant config/export tests passed,11 deselected", + "changes": [ + "global_batch parameter default64,newrun128", + "microbatch64 CLI support", + "AirRaid evaluation uses manifest FPS" + ], + "verified_at": "2026-09-20T07:12:04.593097+00:00" + }, + "backbone": { + "repo_id": "Qwen/Qwen3-VL-4B-Instruct", + "revision": "ebb281ec70b05090aa6165b016eac8ec08e71b17" + }, + "profile_sha256": "7e81ed45da94173b447bdf0cf1ff59343cc7b4e1e6029bcf261cf054c2596cd6", + "teacher": { + "state": "TEACHER_FULL_TRACE_AUDIT_PASSED", + "verified_at": "2026-09-20T15:09:01.399190+00:00", + "run": "h1_10env_2p5obs_g128_20260920", + "env_fps": 10, + "obs_fps": 2.5, + "raw_frames_per_training_decision": 4, + "budget_counted_env_steps": 10000000, + "approximate_raw_game_frame_budget": 40000000, + "actual_env_steps": 10002432, + "seed": 0, + "profile_sha256": "7e81ed45da94173b447bdf0cf1ff59343cc7b4e1e6029bcf261cf054c2596cd6", + "source_profile_capture_fps": [ + 10, + 2.5 + ], + "profile_ms_rescaled": false, + "selection": { + "selected": "final", + "candidates": { + "final": { + "checkpoint": "${PI05_RUN_DIR}/profile_latency/teacher/checkpoints/airraid_pi05_profile_appo_h1_10env_2p5obs_g128_20260920/checkpoint_p0/checkpoint_000039064_10002432.pth", + "sha256": "59e7018b8aad3c0d7f52b40f9a5a4847c450b44f6cba27b66d052895869d4f10", + "episodes": 20, + "seeds": [ + 104771, + 104772, + 104773, + 104774, + 104775, + 104776, + 104777, + 104778, + 104779, + 104780, + 104781, + 104782, + 104783, + 104784, + 104785, + 104786, + 104787, + 104788, + 104789, + 104790 + ], + "returns": [ + 3775.0, + 3025.0, + 4000.0, + 3100.0, + 4050.0, + 4100.0, + 4050.0, + 3700.0, + 3925.0, + 4125.0, + 4100.0, + 4050.0, + 4125.0, + 4000.0, + 3975.0, + 4125.0, + 3525.0, + 4050.0, + 3975.0, + 4125.0 + ], + "mean": 3895.0, + "population_std": 316.68201717179966, + "strict_gt2500": 20 + }, + "training_best": { + "checkpoint": "${PI05_RUN_DIR}/profile_latency/teacher/checkpoints/airraid_pi05_profile_appo_h1_10env_2p5obs_g128_20260920/checkpoint_p0/best_000036640_9379840_reward_11643.500.pth", + "sha256": "ca875c7fce0848b8d62d2ed9c8a8654a198757fed984fa23b1940f503df725fa", + "episodes": 20, + "seeds": [ + 104771, + 104772, + 104773, + 104774, + 104775, + 104776, + 104777, + 104778, + 104779, + 104780, + 104781, + 104782, + 104783, + 104784, + 104785, + 104786, + 104787, + 104788, + 104789, + 104790 + ], + "returns": [ + 4100.0, + 3800.0, + 3050.0, + 4025.0, + 4150.0, + 4075.0, + 4150.0, + 2700.0, + 4075.0, + 3000.0, + 4150.0, + 3025.0, + 4125.0, + 3050.0, + 4125.0, + 4125.0, + 3850.0, + 3825.0, + 3925.0, + 3000.0 + ], + "mean": 3716.25, + "population_std": 503.5049031538819, + "strict_gt2500": 20 + } + }, + "completed_at": "2026-09-20T14:47:07.233886+00:00" + }, + "evaluations": { + "selection_final": { + "episodes": 20, + "seeds": [ + 104771, + 104772, + 104773, + 104774, + 104775, + 104776, + 104777, + 104778, + 104779, + 104780, + 104781, + 104782, + 104783, + 104784, + 104785, + 104786, + 104787, + 104788, + 104789, + 104790 + ], + "returns": [ + 3775.0, + 3025.0, + 4000.0, + 3100.0, + 4050.0, + 4100.0, + 4050.0, + 3700.0, + 3925.0, + 4125.0, + 4100.0, + 4050.0, + 4125.0, + 4000.0, + 3975.0, + 4125.0, + 3525.0, + 4050.0, + 3975.0, + 4125.0 + ], + "lengths": [ + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600 + ], + "mean_return": 3895.0, + "population_std_return": 316.68201717179966, + "mean_length": 3600.0, + "raw_steps": 72000, + "cap_episodes": 20, + "mean_of_episode_latency_ms": 134.5802285199416, + "action_weighted_mean_latency_ms": 134.58022851994164, + "action_latency_p95_ms": 139.88332387384128, + "actions": 18000, + "invalid_actions": 0, + "dropped_actions": 0, + "submitted_observations": 18000, + "dropped_observations": 0, + "observation_opportunities": 18000, + "observation_drop_fraction": 0.0, + "raw_rewards_lengths_and_admission_counters_match": true, + "raw_step_indices_and_clock_verified": true, + "mode": "profile_sample", + "config_sha256": "966bd1fcfc5bb414c430248f2f47fd36e5a0adcc8f86c6cd8d04a6dea9be9c2f", + "strict_gt2500": 20 + }, + "selection_training_best": { + "episodes": 20, + "seeds": [ + 104771, + 104772, + 104773, + 104774, + 104775, + 104776, + 104777, + 104778, + 104779, + 104780, + 104781, + 104782, + 104783, + 104784, + 104785, + 104786, + 104787, + 104788, + 104789, + 104790 + ], + "returns": [ + 4100.0, + 3800.0, + 3050.0, + 4025.0, + 4150.0, + 4075.0, + 4150.0, + 2700.0, + 4075.0, + 3000.0, + 4150.0, + 3025.0, + 4125.0, + 3050.0, + 4125.0, + 4125.0, + 3850.0, + 3825.0, + 3925.0, + 3000.0 + ], + "lengths": [ + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600 + ], + "mean_return": 3716.25, + "population_std_return": 503.5049031538819, + "mean_length": 3600.0, + "raw_steps": 72000, + "cap_episodes": 20, + "mean_of_episode_latency_ms": 134.5802285199416, + "action_weighted_mean_latency_ms": 134.58022851994164, + "action_latency_p95_ms": 139.88332387384128, + "actions": 18000, + "invalid_actions": 0, + "dropped_actions": 0, + "submitted_observations": 18000, + "dropped_observations": 0, + "observation_opportunities": 18000, + "observation_drop_fraction": 0.0, + "raw_rewards_lengths_and_admission_counters_match": true, + "raw_step_indices_and_clock_verified": true, + "mode": "profile_sample", + "config_sha256": "64fc179e9a31bd49d46e6d30b51e5cecd809f1166ed8523388e596ca6234d0de", + "strict_gt2500": 20 + }, + "E10": { + "episodes": 10, + "seeds": [ + 104771, + 104772, + 104773, + 104774, + 104775, + 104776, + 104777, + 104778, + 104779, + 104780 + ], + "returns": [ + 3775.0, + 3025.0, + 4000.0, + 3100.0, + 4050.0, + 4100.0, + 4050.0, + 3700.0, + 3925.0, + 4125.0 + ], + "lengths": [ + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600 + ], + "mean_return": 3785.0, + "population_std_return": 384.08983324217263, + "mean_length": 3600.0, + "raw_steps": 36000, + "cap_episodes": 10, + "mean_of_episode_latency_ms": 134.57468148712098, + "action_weighted_mean_latency_ms": 134.57468148712098, + "action_latency_p95_ms": 139.8895688877921, + "actions": 9000, + "invalid_actions": 0, + "dropped_actions": 0, + "submitted_observations": 9000, + "dropped_observations": 0, + "observation_opportunities": 9000, + "observation_drop_fraction": 0.0, + "raw_rewards_lengths_and_admission_counters_match": true, + "raw_step_indices_and_clock_verified": true, + "mode": "profile_sample", + "config_sha256": "4fc2f1de21e7005bc6002d41a3ba2655c9cb8a8e9d15e4371b3316d6fe167d55", + "strict_gt2500": 10 + } + }, + "probe": { + "source_state_sha256": "74b620e5aa7e0995d331d19a00821f64b0de5b8e64a4846e5e9a19ef636a5880", + "attempts": 12, + "accepted": 12, + "rejected": 0, + "accepted_specs": [ + { + "attempt_idx": 0, + "episode_idx": 0, + "episode_length": 900, + "episode_raw_frames": 3600, + "episode_raw_return": 4050.0, + "seed": 0, + "split": "train" + }, + { + "attempt_idx": 1, + "episode_idx": 1, + "episode_length": 900, + "episode_raw_frames": 3600, + "episode_raw_return": 4000.0, + "seed": 1, + "split": "val" + }, + { + "attempt_idx": 2, + "episode_idx": 2, + "episode_length": 900, + "episode_raw_frames": 3600, + "episode_raw_return": 3775.0, + "seed": 2, + "split": "train" + }, + { + "attempt_idx": 3, + "episode_idx": 3, + "episode_length": 900, + "episode_raw_frames": 3600, + "episode_raw_return": 3400.0, + "seed": 3, + "split": "val" + }, + { + "attempt_idx": 4, + "episode_idx": 4, + "episode_length": 900, + "episode_raw_frames": 3600, + "episode_raw_return": 3025.0, + "seed": 4, + "split": "train" + }, + { + "attempt_idx": 5, + "episode_idx": 5, + "episode_length": 900, + "episode_raw_frames": 3600, + "episode_raw_return": 3100.0, + "seed": 5, + "split": "train" + }, + { + "attempt_idx": 6, + "episode_idx": 6, + "episode_length": 900, + "episode_raw_frames": 3600, + "episode_raw_return": 3700.0, + "seed": 6, + "split": "train" + }, + { + "attempt_idx": 7, + "episode_idx": 7, + "episode_length": 900, + "episode_raw_frames": 3600, + "episode_raw_return": 4050.0, + "seed": 7, + "split": "val" + }, + { + "attempt_idx": 8, + "episode_idx": 8, + "episode_length": 900, + "episode_raw_frames": 3600, + "episode_raw_return": 4000.0, + "seed": 8, + "split": "train" + }, + { + "attempt_idx": 9, + "episode_idx": 9, + "episode_length": 900, + "episode_raw_frames": 3600, + "episode_raw_return": 3400.0, + "seed": 9, + "split": "train" + }, + { + "attempt_idx": 10, + "episode_idx": 10, + "episode_length": 900, + "episode_raw_frames": 3600, + "episode_raw_return": 4250.0, + "seed": 10, + "split": "train" + }, + { + "attempt_idx": 11, + "episode_idx": 11, + "episode_length": 900, + "episode_raw_frames": 3600, + "episode_raw_return": 4125.0, + "seed": 11, + "split": "train" + } + ] + }, + "strict_gate": ">2500", + "E10_seed_caveat": "E10 repeats the first10selection seeds, not independent evaluation", + "raw_budget_caveat": "10M Sample Factory counted steps cover about40M raw game frames atstride4; actual counted envsteps are readfromFinal checkpoint.", + "downstream": "After positive12attemptprobe,100accepted strict>2500 with no finiteattemptcap,actualreplayQA,freshPi05global128/micro64/acc1/GCfalse/ZeRO2/5000updates,andRTX3090F20; otherwise scientificstop", + "source_P_verification_sha256": "bc541ac289879bef1d50f006f13b530c1bb4870df600c9153891de301f784a38", + "gate": "PASSED" + }, + "dataset": { + "attempts": 100, + "accepted": 100, + "max_total_attempts": null, + "sha256": { + "metadata.json": "7932e2698df3d27ba5897eb0456312d49972f473b2f1ba0faabfe6082872d7f6", + "train.parquet": "9d4d62170e947dc70521ff67cb75b550b7beb15d5c9a8cebd3ddb083daee7eb4", + "val.parquet": "ffd7e009d86d9bc611d48a132b0ba98b4b2fc16e71e4d40b0c2bdf9ad66055d9", + "filter_report.json": "f7d581d2f6b83de6eb3eab5422839b44148efd1da07f2eb29671f5fdb0d9ba7e" + }, + "splits": { + "train": { + "rows": 81000, + "episodes": 90, + "returns": { + "0": 4050.0, + "2": 3775.0, + "4": 3025.0, + "5": 3100.0, + "6": 3700.0, + "8": 4000.0, + "9": 3400.0, + "10": 4250.0, + "11": 4125.0, + "12": 4000.0, + "13": 3275.0, + "14": 4250.0, + "15": 3400.0, + "16": 3725.0, + "18": 3700.0, + "19": 4100.0, + "20": 4000.0, + "21": 3500.0, + "22": 4050.0, + "23": 3650.0, + "24": 3650.0, + "26": 4050.0, + "27": 3500.0, + "28": 3775.0, + "30": 4125.0, + "31": 4000.0, + "32": 4050.0, + "33": 4000.0, + "34": 3100.0, + "35": 4100.0, + "36": 3275.0, + "37": 3025.0, + "38": 4250.0, + "39": 3775.0, + "40": 3775.0, + "41": 4250.0, + "42": 4125.0, + "43": 4000.0, + "44": 4125.0, + "45": 4250.0, + "46": 3725.0, + "48": 3400.0, + "49": 3775.0, + "50": 4050.0, + "51": 4125.0, + "52": 4050.0, + "53": 4125.0, + "54": 3100.0, + "55": 3400.0, + "56": 4125.0, + "57": 3525.0, + "59": 3400.0, + "60": 3775.0, + "61": 3700.0, + "62": 4100.0, + "63": 3500.0, + "64": 4125.0, + "65": 4050.0, + "66": 4250.0, + "67": 3925.0, + "68": 3650.0, + "69": 4125.0, + "70": 4050.0, + "71": 4250.0, + "72": 4000.0, + "73": 4100.0, + "74": 4050.0, + "75": 3525.0, + "76": 4100.0, + "78": 4000.0, + "79": 4125.0, + "80": 4050.0, + "82": 3500.0, + "83": 4050.0, + "84": 3525.0, + "85": 4250.0, + "86": 4125.0, + "87": 3500.0, + "88": 4250.0, + "89": 4100.0, + "90": 4100.0, + "91": 4250.0, + "92": 4100.0, + "93": 4100.0, + "94": 4250.0, + "95": 3775.0, + "96": 4000.0, + "97": 3925.0, + "98": 4125.0, + "99": 3500.0 + }, + "return_min": 3025.0, + "return_mean": 3866.1111111111113 + }, + "val": { + "rows": 9000, + "episodes": 10, + "returns": { + "1": 4000.0, + "3": 3400.0, + "7": 4050.0, + "17": 3100.0, + "25": 4100.0, + "29": 3400.0, + "47": 4125.0, + "58": 4125.0, + "77": 4125.0, + "81": 4250.0 + }, + "return_min": 3100.0, + "return_mean": 3867.5 + } + } + }, + "checkpoint_storage": { + "save_training_state": true, + "weights_saved_every": 500, + "scientific_recipe_matches_L0": true + }, + "evaluation_prompt_key": 2, + "global_batch": 128, + "micro_batch": 64, + "accumulation": 1, + "gradient_checkpointing": false, + "zero_stage": 2 + }, + "training_config_sha256": "b0f1c69540f0c941e6672a2027d593715e65d3b10a3a142569615bdf873aabc5", + "dataset_manifest_sha256": "4fb0f0dc1c21b4480dfc25fd5b9b14dbd4b21ec167500877a4952b5f840fe196" +} diff --git a/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/task_contract.json b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/task_contract.json new file mode 100644 index 0000000000000000000000000000000000000000..dd4185bbac1fa512169ff4e355662ca91df015d0 --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/task_contract.json @@ -0,0 +1,44 @@ +{ + "action_labels": [ + "noop", + "fire", + "right", + "left", + "rightfire", + "leftfire" + ], + "action_values": [ + 0, + 1, + 2, + 3, + 4, + 5 + ], + "base_prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.", + "env_fps": 10.0, + "env_id": "LatencyBench/AirRaid-v0", + "frame_stack": 4, + "make_kwargs": { + "base_env_id": "ALE/AirRaid-v5", + "base_make_kwargs": { + "difficulty": 0, + "frameskip": 1, + "full_action_space": false, + "max_num_frames_per_episode": 108000, + "mode": 1, + "obs_type": "rgb", + "repeat_action_probability": 0.0 + }, + "fire_reset": true, + "noop_max": 30, + "render_mode": "rgb_array", + "screen_size": 84 + }, + "noop_action_id": 0, + "obs_fps": 2.5, + "registration_imports": [ + "latency_bench.envs.gymnasium_air_raid" + ], + "task_name": "air_raid" +} diff --git a/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/validation.json b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/validation.json new file mode 100644 index 0000000000000000000000000000000000000000..0508a242c38d27bb0d2d3fd3a30a39facdaba76d --- /dev/null +++ b/latency-aware/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_profile_h1_10env_2p5obs_g128_20260920/validation.json @@ -0,0 +1,15 @@ +{ + "state": "SAVED_MODEL_RELOAD_AND_REAL_SAMPLE_FORWARD_VERIFIED", + "forward_shape": [ + 1, + 1, + 6 + ], + "forward_finite": true, + "state_dim": 0, + "uses_state": false, + "action_horizon": 1, + "loader_rows": 9000, + "checkpoint_sha256": "7914e1f1e2ba0b78d29c15dd33d96bbb6a5f61bad6ebb17f583a6d6d9958fb14", + "verified_at": "2026-09-20T17:52:57.468485+00:00" +} diff --git a/zero-latency/air-raid/small-policy/sample-factory-v1/README.md b/zero-latency/air-raid/small-policy/sample-factory-v1/README.md new file mode 100644 index 0000000000000000000000000000000000000000..681f267f7e5daf55a1b75bf51adf0777964899d1 --- /dev/null +++ b/zero-latency/air-raid/small-policy/sample-factory-v1/README.md @@ -0,0 +1,26 @@ +# air-raid / sample-factory-appo + +Training condition: `zero-latency`. Run: `gymnasium_air_raid_zero_latency_atari_official_fast_10m`. + +Copied from [Standard-Pipeline](https://huggingface.co/latency-sensitive-bench/Standard-Pipeline/tree/2208875f92b21e3f8ec2363a30a6705dfe56cc04/air_raid/zero_latency/small_model) at `2208875f92b21e3f8ec2363a30a6705dfe56cc04`. + +- Checkpoint: `checkpoint.pth` +- Selection: existing zero-latency training best +- Checkpoint SHA256: `be322ccd2d18ff7ebd85dbc32c0d27fbc2a97dd1eb6b6ab4ee9f3128021fcded` +- Configuration, FPS, actions, normalization and latency semantics retain their source values. +- This import does not run evaluations or change training/evaluation/acceptance status. + +Sample Factory APPO inference bundle: `config.json` and the explicit `checkpoint.pth`. +The checkpoint contains `model`, `train_step`, and `env_steps`; every exported model +tensor was reloaded and compared bit-for-bit with the selected source checkpoint. + +## Download + +```python +from huggingface_hub import snapshot_download +prefix = "zero-latency/air-raid/small-policy/sample-factory-v1" +snapshot_download("latency-sensitive-bench/benchmark-models", revision="", + allow_patterns=[prefix + "/**"]) +``` + +Use the verified import commit listed in the repository's import report. diff --git a/zero-latency/air-raid/small-policy/sample-factory-v1/checkpoint.pth b/zero-latency/air-raid/small-policy/sample-factory-v1/checkpoint.pth new file mode 100644 index 0000000000000000000000000000000000000000..fb6850f0ef101da4bf6f4ad237f18435d13f2d63 --- /dev/null +++ b/zero-latency/air-raid/small-policy/sample-factory-v1/checkpoint.pth @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:be322ccd2d18ff7ebd85dbc32c0d27fbc2a97dd1eb6b6ab4ee9f3128021fcded +size 7210293 diff --git a/zero-latency/air-raid/small-policy/sample-factory-v1/config.json b/zero-latency/air-raid/small-policy/sample-factory-v1/config.json new file mode 100644 index 0000000000000000000000000000000000000000..2f447add45080af6b569a55d7d19d9e0dc2a3943 --- /dev/null +++ b/zero-latency/air-raid/small-policy/sample-factory-v1/config.json @@ -0,0 +1,227 @@ +{ + "help": false, + "algo": "APPO", + "env": "latency_gymnasium", + "experiment": "gymnasium_air_raid_zero_latency_atari_official_fast_10m", + "train_dir": "/mnt/checkpoints/latency-sensitive-bench/small_models/air_raid", + "restart_behavior": "overwrite", + "device": "gpu", + "seed": 0, + "num_policies": 1, + "async_rl": true, + "serial_mode": false, + "batched_sampling": false, + "num_batches_to_accumulate": 2, + "worker_num_splits": 2, + "policy_workers_per_policy": 1, + "max_policy_lag": 300, + "num_workers": 16, + "num_envs_per_worker": 8, + "batch_size": 1024, + "num_batches_per_epoch": 4, + "num_epochs": 4, + "rollout": 32, + "recurrence": 1, + "shuffle_minibatches": false, + "gamma": 0.99, + "reward_scale": 1.0, + "reward_clip": 1000.0, + "value_bootstrap": false, + "normalize_returns": true, + "exploration_loss_coeff": 0.01, + "value_loss_coeff": 0.5, + "kl_loss_coeff": 0.0, + "exploration_loss": "entropy", + "gae_lambda": 0.95, + "ppo_clip_ratio": 0.1, + "ppo_clip_value": 0.2, + "with_vtrace": false, + "vtrace_rho": 1.0, + "vtrace_c": 1.0, + "optimizer": "adam", + "adam_eps": 1e-05, + "adam_beta1": 0.9, + "adam_beta2": 0.999, + "max_grad_norm": 0.5, + "learning_rate": 0.00025, + "lr_schedule": "linear_decay", + "lr_schedule_kl_threshold": 0.008, + "lr_adaptive_min": 1e-06, + "lr_adaptive_max": 0.01, + "obs_subtract_mean": 0.0, + "obs_scale": 255.0, + "normalize_input": true, + "normalize_input_keys": null, + "decorrelate_experience_max_seconds": 0, + "decorrelate_envs_on_one_worker": true, + "actor_worker_gpus": [], + "set_workers_cpu_affinity": true, + "force_envs_single_thread": false, + "default_niceness": 0, + "log_to_file": true, + "experiment_summaries_interval": 10, + "flush_summaries_interval": 30, + "stats_avg": 100, + "summaries_use_frameskip": true, + "heartbeat_interval": 20, + "heartbeat_reporting_interval": 180, + "train_for_env_steps": 10000000, + "train_for_seconds": 10000000000, + "save_every_sec": 600, + "keep_checkpoints": 5, + "load_checkpoint_kind": "latest", + "save_milestones_sec": -1, + "save_best_every_sec": 5, + "save_best_metric": "reward", + "save_best_after": 100000, + "benchmark": false, + "encoder_mlp_layers": [ + 512, + 512 + ], + "encoder_conv_architecture": "convnet_atari", + "encoder_conv_mlp_layers": [ + 512 + ], + "use_rnn": false, + "rnn_size": 512, + "rnn_type": "gru", + "rnn_num_layers": 1, + "decoder_mlp_layers": [], + "nonlinearity": "relu", + "policy_initialization": "orthogonal", + "policy_init_gain": 1.0, + "actor_critic_share_weights": true, + "adaptive_stddev": false, + "continuous_tanh_scale": 0.0, + "initial_stddev": 1.0, + "use_env_info_cache": false, + "env_gpu_actions": false, + "env_gpu_observations": true, + "env_frameskip": 1, + "env_framestack": 4, + "pixel_format": "CHW", + "use_record_episode_statistics": false, + "with_wandb": false, + "wandb_user": null, + "wandb_project": "sample_factory", + "wandb_group": null, + "wandb_job_type": "SF", + "wandb_tags": [], + "with_pbt": false, + "pbt_mix_policies_in_one_env": true, + "pbt_period_env_steps": 5000000, + "pbt_start_mutation": 20000000, + "pbt_replace_fraction": 0.3, + "pbt_mutation_rate": 0.15, + "pbt_replace_reward_gap": 0.1, + "pbt_replace_reward_gap_absolute": 1e-06, + "pbt_optimize_gamma": false, + "pbt_target_objective": "true_objective", + "pbt_perturb_min": 1.1, + "pbt_perturb_max": 1.5, + "gym_task_name": "air_raid", + "gym_env_id": "LatencyBench/AirRaid-v0", + "gym_make_kwargs_json": "{\"base_env_id\": \"ALE/AirRaid-v5\", \"render_mode\": \"rgb_array\", \"screen_size\": 84, \"noop_max\": 0, \"base_make_kwargs\": {\"obs_type\": \"rgb\", \"frameskip\": 1, \"repeat_action_probability\": 0.0, \"full_action_space\": false, \"mode\": 1, \"difficulty\": 0, \"max_num_frames_per_episode\": 108000}}", + "gym_registration_imports_json": "[\"latency_bench.envs.gymnasium_air_raid\"]", + "gym_action_space_json": "{\"type\": \"discrete\", \"labels\": [\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"], \"values\": [0, 1, 2, 3, 4, 5]}", + "gym_action_labels_json": "[\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"]", + "gym_action_values_json": "[0, 1, 2, 3, 4, 5]", + "gym_noop_action_id": 0, + "gym_noop_action_json": "\"noop\"", + "gym_base_prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.", + "env_fps": 50.0, + "obs_fps": 12.5, + "frame_stack": 4, + "export_env_raw_rgb_frames": false, + "mode": "train", + "latency_type": "zero", + "fixed_latency_ms": null, + "mean_latency_ms": null, + "std_latency_ms": null, + "min_latency_ms": null, + "max_latency_ms": null, + "latency_seed": null, + "add_latency_info": false, + "max_pending_actions": null, + "hold_policy": "hold", + "ordering_policy": "issue_order_fifo", + "eval_episodes": 5, + "eval_parallel_envs": 1, + "eval_latency_raw_frame_values": null, + "eval_max_steps": 5000, + "eval_deterministic": true, + "eval_raw_reward": false, + "episode_metrics_path": "/mnt/checkpoints/latency-sensitive-bench/small_models/air_raid/gymnasium_air_raid_zero_latency_atari_official_fast_10m/episode_metrics.jsonl", + "command_line": "--mode train --algo APPO --env latency_gymnasium --experiment gymnasium_air_raid_zero_latency_atari_official_fast_10m --train_dir /mnt/checkpoints/latency-sensitive-bench/small_models/air_raid --restart_behavior overwrite --device gpu --seed 0 --episode_metrics_path /mnt/checkpoints/latency-sensitive-bench/small_models/air_raid/gymnasium_air_raid_zero_latency_atari_official_fast_10m/episode_metrics.jsonl --train_for_env_steps 10000000 --num_workers 16 --num_envs_per_worker 8 --num_policies 1 --batch_size 1024 --rollout 32 --recurrence 1 --num_epochs 4 --num_batches_per_epoch 4 --worker_num_splits 2 --max_policy_lag 300 --learning_rate 0.00025 --nonlinearity relu --env_framestack 4 --gamma 0.99 --gae_lambda 0.95 --ppo_clip_ratio 0.1 --ppo_clip_value 0.2 --exploration_loss entropy --exploration_loss_coeff 0.01 --value_loss_coeff 0.5 --max_grad_norm 0.5 --adam_eps 1e-05 --obs_scale 255.0 --save_every_sec 600 --keep_checkpoints 5 --async_rl True --use_rnn False --normalize_returns True --normalize_input True --adaptive_stddev False --with_vtrace False --latency-type zero --add-latency-info False --eval-episodes 5 --eval-parallel-envs 1 --eval-max-steps 5000 --eval-deterministic True --encoder_conv_architecture convnet_atari --gym-task-name air_raid --gym-env-id LatencyBench/AirRaid-v0 --gym-make-kwargs-json {\"base_env_id\": \"ALE/AirRaid-v5\", \"render_mode\": \"rgb_array\", \"screen_size\": 84, \"noop_max\": 0, \"base_make_kwargs\": {\"obs_type\": \"rgb\", \"frameskip\": 1, \"repeat_action_probability\": 0.0, \"full_action_space\": false, \"mode\": 1, \"difficulty\": 0, \"max_num_frames_per_episode\": 108000}} --gym-registration-imports-json [\"latency_bench.envs.gymnasium_air_raid\"] --gym-action-space-json {\"type\": \"discrete\", \"labels\": [\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"], \"values\": [0, 1, 2, 3, 4, 5]} --gym-noop-action-json \"noop\" --gym-action-labels-json [\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"] --gym-action-values-json [0, 1, 2, 3, 4, 5] --gym-noop-action-id 0 --gym-base-prompt Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire. --env-fps 50 --obs-fps 12.5 --frame-stack 4 --hold-policy hold --ordering-policy issue_order_fifo --use_env_info_cache False", + "cli_args": { + "algo": "APPO", + "env": "latency_gymnasium", + "experiment": "gymnasium_air_raid_zero_latency_atari_official_fast_10m", + "train_dir": "/mnt/checkpoints/latency-sensitive-bench/small_models/air_raid", + "restart_behavior": "overwrite", + "device": "gpu", + "seed": 0, + "num_policies": 1, + "async_rl": true, + "worker_num_splits": 2, + "max_policy_lag": 300, + "num_workers": 16, + "num_envs_per_worker": 8, + "batch_size": 1024, + "num_batches_per_epoch": 4, + "num_epochs": 4, + "rollout": 32, + "recurrence": 1, + "gamma": 0.99, + "normalize_returns": true, + "exploration_loss_coeff": 0.01, + "value_loss_coeff": 0.5, + "exploration_loss": "entropy", + "gae_lambda": 0.95, + "ppo_clip_ratio": 0.1, + "ppo_clip_value": 0.2, + "with_vtrace": false, + "adam_eps": 1e-05, + "max_grad_norm": 0.5, + "learning_rate": 0.00025, + "obs_scale": 255.0, + "normalize_input": true, + "train_for_env_steps": 10000000, + "save_every_sec": 600, + "keep_checkpoints": 5, + "encoder_conv_architecture": "convnet_atari", + "use_rnn": false, + "nonlinearity": "relu", + "adaptive_stddev": false, + "use_env_info_cache": false, + "env_framestack": 4, + "gym_task_name": "air_raid", + "gym_env_id": "LatencyBench/AirRaid-v0", + "gym_make_kwargs_json": "{\"base_env_id\": \"ALE/AirRaid-v5\", \"render_mode\": \"rgb_array\", \"screen_size\": 84, \"noop_max\": 0, \"base_make_kwargs\": {\"obs_type\": \"rgb\", \"frameskip\": 1, \"repeat_action_probability\": 0.0, \"full_action_space\": false, \"mode\": 1, \"difficulty\": 0, \"max_num_frames_per_episode\": 108000}}", + "gym_registration_imports_json": "[\"latency_bench.envs.gymnasium_air_raid\"]", + "gym_action_space_json": "{\"type\": \"discrete\", \"labels\": [\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"], \"values\": [0, 1, 2, 3, 4, 5]}", + "gym_action_labels_json": "[\"noop\", \"fire\", \"right\", \"left\", \"rightfire\", \"leftfire\"]", + "gym_action_values_json": "[0, 1, 2, 3, 4, 5]", + "gym_noop_action_id": 0, + "gym_noop_action_json": "\"noop\"", + "gym_base_prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.", + "env_fps": 50.0, + "obs_fps": 12.5, + "frame_stack": 4, + "mode": "train", + "latency_type": "zero", + "add_latency_info": false, + "hold_policy": "hold", + "ordering_policy": "issue_order_fifo", + "eval_episodes": 5, + "eval_parallel_envs": 1, + "eval_max_steps": 5000, + "eval_deterministic": true, + "episode_metrics_path": "/mnt/checkpoints/latency-sensitive-bench/small_models/air_raid/gymnasium_air_raid_zero_latency_atari_official_fast_10m/episode_metrics.jsonl" + }, + "git_hash": "09e6874b6e72b7a44f17dc238ef757c6c4c22e10", + "git_repo_name": "https://github.com/ZihanWang314/latency-sensitive-bench.git", + "eval_env_frameskip": 1, + "output_dir": "/mnt/results/latency-sensitive-bench/tasks/air_raid/l0_atari_official_fast_10m_train" +} \ No newline at end of file diff --git a/zero-latency/air-raid/small-policy/sample-factory-v1/provenance.json b/zero-latency/air-raid/small-policy/sample-factory-v1/provenance.json new file mode 100644 index 0000000000000000000000000000000000000000..9f28f7f5fc578223098509191f8d6134010d29e0 --- /dev/null +++ b/zero-latency/air-raid/small-policy/sample-factory-v1/provenance.json @@ -0,0 +1,44 @@ +{ + "task": "air-raid", + "model": "sample-factory-appo", + "training_condition": "zero-latency", + "training_run_id": "gymnasium_air_raid_zero_latency_atari_official_fast_10m", + "source": { + "repo_id": "latency-sensitive-bench/Standard-Pipeline", + "revision": "2208875f92b21e3f8ec2363a30a6705dfe56cc04", + "path": "air_raid/zero_latency/small_model", + "url": "https://huggingface.co/latency-sensitive-bench/Standard-Pipeline/tree/2208875f92b21e3f8ec2363a30a6705dfe56cc04/air_raid/zero_latency/small_model" + }, + "hf_repo_id": "latency-sensitive-bench/benchmark-models", + "hf_path": "zero-latency/air-raid/small-policy/sample-factory-v1", + "checkpoint": { + "source_file": "air_raid/zero_latency/small_model/checkpoint_p0/best_000038656_9895936_reward_10593.000.pth", + "source_sha256": "a80d57f84ee4d6da1d8924c455ab6f5c5e6d23102fca8d23582b442c7168ea26", + "source_bytes": 20722745, + "file": "checkpoint.pth", + "selection_rule": "existing zero-latency training best", + "method": "inference_export", + "sha256": "be322ccd2d18ff7ebd85dbc32c0d27fbc2a97dd1eb6b6ab4ee9f3128021fcded", + "bytes": 7210293, + "train_step": 38656, + "env_steps": 9895936, + "tensor_count": 18, + "tensor_equality_verified": true, + "removed_fields": [ + "best_performance", + "curr_lr", + "optimizer" + ] + }, + "config_source": "air_raid/zero_latency/small_model/config.json", + "task_contract_file": null, + "missing_source_files": [], + "migration": { + "operation": "copy_selected_inference_bundle", + "source_modified": false, + "evaluation_performed": false, + "experiment_status": "preserved_in_source_evidence" + }, + "algorithm": "APPO", + "latency_model": null +} diff --git a/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/README.md b/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/README.md new file mode 100644 index 0000000000000000000000000000000000000000..4f87ddbcffa339902ec39c6aa45f5e62c68cd812 --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/README.md @@ -0,0 +1,29 @@ +# air-raid / qwengr00t + +Training condition: `zero-latency`. Run: `air_raid_zero_latency_gr00t_h1`. + +Copied from [Standard-Pipeline](https://huggingface.co/latency-sensitive-bench/Standard-Pipeline/tree/2208875f92b21e3f8ec2363a30a6705dfe56cc04/air_raid/zero_latency/GR00T) at `2208875f92b21e3f8ec2363a30a6705dfe56cc04`. + +- Checkpoint: `checkpoints/model.pt` +- Selection: only published VLA checkpoint +- Checkpoint SHA256: `414bb930912ac50994e88d642ed7948bac16c2bc1b41b0955a8eacff45ee1e0e` +- Configuration, FPS, actions, normalization and latency semantics retain their source values. +- This import does not run evaluations or change training/evaluation/acceptance status. + +Action horizon is 1. Use `config.yaml` and `dataset_statistics.json` with the matching +prepared backbone. `task_contract.json` is extracted from the original training configuration. +Original machine paths in historical configurations are provenance: bind checkpoint, +backbone, manifest and profile paths to downloaded local files when running inference. + +[Original provenance](source/provenance.json) retains dataset references and recorded experiment evidence. + +## Download + +```python +from huggingface_hub import snapshot_download +prefix = "zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1" +snapshot_download("latency-sensitive-bench/benchmark-models", revision="", + allow_patterns=[prefix + "/**"]) +``` + +Use the verified import commit listed in the repository's import report. diff --git a/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/checkpoints/model.pt b/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/checkpoints/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..d07144d028fc277f6ddd5c140b0d346aab6597ad --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/checkpoints/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:414bb930912ac50994e88d642ed7948bac16c2bc1b41b0955a8eacff45ee1e0e +size 9975248311 diff --git a/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/config.yaml b/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..9d57c45a342ed4e663c52b8d135a129df91ab608 --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/config.yaml @@ -0,0 +1,106 @@ +framework: + name: QwenGR00T + qwenvl: + base_vlm: ${oc.env:GR00T_BACKBONE_DIR} + attn_implementation: flash_attention_2 + vl_hidden_dim: 2048 + enable_gradient_checkpointing: true + action_model: + action_model_type: DiT-B + action_hidden_dim: 1024 + hidden_size: 1024 + add_pos_embed: true + max_seq_len: 1024 + action_dim: 6 + state_dim: 0 + task_objective: null + action_horizon: 1 + repeated_diffusion_steps: 8 + noise_beta_alpha: 1.5 + noise_beta_beta: 1.0 + noise_s: 0.999 + num_timestep_buckets: 1000 + num_inference_timesteps: 4 + num_target_vision_tokens: 32 + diffusion_model_cfg: + cross_attention_dim: 2560 + dropout: 0.2 + final_dropout: true + interleave_self_attention: true + norm_type: ada_norm + num_layers: 16 + output_dim: 1024 + positional_embeddings: null + action_env_dim: 6 + future_action_window_size: 0 + past_action_window_size: 0 +version_id: '0.21' +datasets: + vla_data: + include_state: false + obs_image_size: + - 224 + - 224 + image_mode: single + stitch_grid: + - 2 + - 2 + num_obs_frames: 1 + gymnasium_task_contract: &id001 + action_labels: + - noop + - fire + - right + - left + - rightfire + - leftfire + action_values: + - 0 + - 1 + - 2 + - 3 + - 4 + - 5 + base_prompt: 'Protect both buildings from flying saucers. Choose exactly one + action from: noop, fire, right, left, rightfire, leftfire.' + env_fps: 50.0 + env_id: LatencyBench/AirRaid-v0 + frame_stack: 4 + make_kwargs: + base_env_id: ALE/AirRaid-v5 + base_make_kwargs: + difficulty: 0 + frameskip: 1 + full_action_space: false + max_num_frames_per_episode: 108000 + mode: 1 + obs_type: rgb + repeat_action_probability: 0.0 + fire_reset: true + noop_max: 30 + render_mode: rgb_array + screen_size: 84 + noop_action_id: 0 + obs_fps: 12.5 + registration_imports: + - latency_bench.envs.gymnasium_air_raid + task_name: air_raid +trainer: + pretrained_checkpoint: null + profile_timing: + enabled: false + log_interval: 10 +backbone: + repo_id: Qwen/Qwen3-VL-4B-Instruct + revision: ebb281ec70b05090aa6165b016eac8ec08e71b17 +rl_games: + model_alias: gr00t + task: gymnasium + action_carrier: native + gymnasium: + task_contract: *id001 + env_eval: + image_size: 224 + frameskip: 1 + image_transform: raw_rgb + prompt_mode: raw diff --git a/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/dataset_statistics.json b/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/dataset_statistics.json new file mode 100644 index 0000000000000000000000000000000000000000..39af401e32da1e3219fdbe3ffbe7d00c7d2b8f1b --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/dataset_statistics.json @@ -0,0 +1,84 @@ +{ + "new_embodiment": { + "action": { + "mean": [ + 0.17179012298583984, + 0.27208641171455383, + 0.05909876525402069, + 0.2254074066877365, + 0.20286419987678528, + 0.06875308603048325 + ], + "std": [ + 0.3770434260368347, + 0.4451958239078522, + 0.23573854565620422, + 0.41779401898384094, + 0.4020724594593048, + 0.25303277373313904 + ], + "max": [ + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0 + ], + "min": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "q01": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "q99": [ + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0 + ], + "mask": [ + true, + true, + true, + true, + true, + true + ] + }, + "state": { + "mean": [ + 0.0 + ], + "std": [ + 0.0 + ], + "max": [ + 0.0 + ], + "min": [ + 0.0 + ], + "q01": [ + 0.0 + ], + "q99": [ + 0.0 + ] + }, + "num_transitions": 81000, + "num_trajectories": 90 + } +} \ No newline at end of file diff --git a/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/latency_prompt_map.json b/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/latency_prompt_map.json new file mode 100644 index 0000000000000000000000000000000000000000..f7436fcaddd79e25b458c15a9e548373e5c0dd18 --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/latency_prompt_map.json @@ -0,0 +1,7 @@ +{ + "0": { + "prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire. Current action latency is 0 raw frames (0.00 ms). The environment runs at 50 FPS and observations are emitted at 12.5 FPS. Choose the best next action.", + "latency_raw_frames": 0, + "latency_ms": 0.0 + } +} \ No newline at end of file diff --git a/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/manifest.json b/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..16c656ac6815c2b6e7a4e575850ab5b189cfe0a4 --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/manifest.json @@ -0,0 +1,96 @@ +{ + "dataset_name": "air_raid_gr00t_l0", + "env_name": "air_raid", + "episodes": 90, + "frames": 81000, + "task_prompts": [ + "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire. Current action latency is 0 raw frames (0.00 ms). The environment runs at 50 FPS and observations are emitted at 12.5 FPS. Choose the best next action." + ], + "format": "starvla_lerobot_v2_image_parquet", + "integration_name": "gymnasium", + "task_name": "air_raid", + "action_layout": "gymnasium_discrete_v1", + "action_labels": [ + "noop", + "fire", + "right", + "left", + "rightfire", + "leftfire" + ], + "carrier_action_labels": [ + "noop", + "fire", + "right", + "left", + "rightfire", + "leftfire" + ], + "action_dim": 6, + "active_action_dim": 6, + "action_carrier": "native", + "rows_unit": "decision_step", + "obs_fps": 12.5, + "obs_stride_raw_frames": 4, + "uses_state": false, + "state_dim": 1, + "state_labels": [ + "state" + ], + "state_normalization": null, + "robot_type": "rl_games_gymnasium", + "gymnasium_task": { + "action_labels": [ + "noop", + "fire", + "right", + "left", + "rightfire", + "leftfire" + ], + "action_values": [ + 0, + 1, + 2, + 3, + 4, + 5 + ], + "base_prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.", + "env_fps": 50.0, + "env_id": "LatencyBench/AirRaid-v0", + "frame_stack": 4, + "make_kwargs": { + "base_env_id": "ALE/AirRaid-v5", + "base_make_kwargs": { + "difficulty": 0, + "frameskip": 1, + "full_action_space": false, + "max_num_frames_per_episode": 108000, + "mode": 1, + "obs_type": "rgb", + "repeat_action_probability": 0.0 + }, + "fire_reset": true, + "noop_max": 30, + "render_mode": "rgb_array", + "screen_size": 84 + }, + "noop_action_id": 0, + "obs_fps": 12.5, + "registration_imports": [ + "latency_bench.envs.gymnasium_air_raid" + ], + "task_name": "air_raid" + }, + "validation_dataset_name": "air_raid_gr00t_l0__val", + "validation_episodes": 10, + "validation_frames": 9000, + "raw_dataset": { + "repo": "latency-sensitive-bench/Standard-Pipeline", + "repo_type": "dataset", + "prefix": "air_raid/zero_latency/shared/native_100ep_v1", + "revision": "483301cd09aabec98e06657a135c5f041515b73e", + "reused": true + } +} diff --git a/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/provenance.json b/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/provenance.json new file mode 100644 index 0000000000000000000000000000000000000000..e008b58b17e466e18b0c1e4622087b5d7092aea4 --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/provenance.json @@ -0,0 +1,34 @@ +{ + "task": "air-raid", + "model": "qwengr00t", + "training_condition": "zero-latency", + "training_run_id": "air_raid_zero_latency_gr00t_h1", + "source": { + "repo_id": "latency-sensitive-bench/Standard-Pipeline", + "revision": "2208875f92b21e3f8ec2363a30a6705dfe56cc04", + "path": "air_raid/zero_latency/GR00T", + "url": "https://huggingface.co/latency-sensitive-bench/Standard-Pipeline/tree/2208875f92b21e3f8ec2363a30a6705dfe56cc04/air_raid/zero_latency/GR00T" + }, + "hf_repo_id": "latency-sensitive-bench/benchmark-models", + "hf_path": "zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1", + "checkpoint": { + "source_file": "air_raid/zero_latency/GR00T/checkpoints/model.pt", + "source_sha256": "414bb930912ac50994e88d642ed7948bac16c2bc1b41b0955a8eacff45ee1e0e", + "source_bytes": 9975248311, + "file": "checkpoints/model.pt", + "selection_rule": "only published VLA checkpoint", + "method": "copy", + "sha256": "414bb930912ac50994e88d642ed7948bac16c2bc1b41b0955a8eacff45ee1e0e", + "bytes": 9975248311 + }, + "config_source": "air_raid/zero_latency/GR00T/config.yaml", + "task_contract_file": "task_contract.json", + "missing_source_files": [], + "migration": { + "operation": "copy_selected_inference_bundle", + "source_modified": false, + "evaluation_performed": false, + "experiment_status": "preserved_in_source_evidence" + }, + "action_horizon": 1 +} diff --git a/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/reload-validation.json b/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/reload-validation.json new file mode 100644 index 0000000000000000000000000000000000000000..bec4727390d7da38122b101ce03fcb2a34254c17 --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/reload-validation.json @@ -0,0 +1,12 @@ +{ + "status": "passed", + "strict_load": true, + "output_shape": [ + 1, + 1, + 6 + ], + "finite": true, + "checkpoint_sha256": "414bb930912ac50994e88d642ed7948bac16c2bc1b41b0955a8eacff45ee1e0e", + "config_sha256": "ff7ffb921bce669e1375145d5c8a2e89eb5b2c6cdeff3e1b268988fec172e0d3" +} diff --git a/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/source/provenance.json b/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/source/provenance.json new file mode 100644 index 0000000000000000000000000000000000000000..bb350615e9a7e696781f4e3f4b77a11a2d387452 --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/source/provenance.json @@ -0,0 +1,40 @@ +{ + "task": "air_raid", + "backbone": { + "repo_id": "Qwen/Qwen3-VL-4B-Instruct", + "revision": "ebb281ec70b05090aa6165b016eac8ec08e71b17" + }, + "checkpoint": "steps_5000_pytorch_model.pt", + "training_steps": 5000, + "seed": 42, + "global_batch": 64, + "training_run_id": "air_raid_zero_latency_gr00t_h1", + "condition": "zero_latency", + "source": { + "base_commit": "1681edb8d9e4f7d10b7dfa647670e01e0d5cce37", + "starvla_commit": "1d0d7b139d1725cab268cb9dae2007ef4c5d05d2", + "sample_factory_commit": "4b11d8754ffc989cdd15a340cfc9be0277377d5b", + "files": { + "latency_bench/data/starvla_tasks.py": "67336b5cd57d58138c8294b55b1eea3a708033caf488fb6b5fa32a87674d36bf", + "latency_bench/policy/starvla.py": "6c107e6c52e673310d98c511f56b5fce3a7e397a73e226713168377b7f605e91", + "latency_bench/policy/starvla_tasks.py": "a89d55999f9d17267eedc393721f1fedcdb99af752c06b95f291e5d82bc117a5", + "scripts/starvla/prepare_dataset.py": "98324cab1650920c9311c2021bc0f91571b1ec3f24417cf0b629e1e8fa65687c", + "scripts/gym_adapt/gr00t.py": "00d1244dcf3b8d564f7e86558841ac3490f75f1367410f4af8eeebaf8268531d", + "docs/gym-adapt/gr00t-pipeline-plan.md": "06e60878a152107475604fbd17b52c94a34734350ba3f5007df2eec44b58a71c", + "configs/examples/gymnasium/gr00t/ant_teacher.yaml": "b20c6858801d5a3ea26e68a8cecf94d2718072f6d0fb94a9376333e069a04274", + "configs/examples/gymnasium/gr00t/half_cheetah_teacher.yaml": "cbfba28a5ee625f06c30fa1d290ed29a391f019a899fd5e4ae2503e09f6679a1", + "configs/examples/gymnasium/gr00t/hopper_teacher.yaml": "b57721674860194683f07896057d2206c1cfd5753c08b196b45d544466ed5cc2", + "configs/examples/gymnasium/gr00t/humanoid_teacher.yaml": "9701de3b39ab656bcc721d5aa91d10c05672f66da182cc741f1c34904e3cf06b", + "configs/examples/gymnasium/gr00t/inverted_pendulum_teacher.yaml": "e5c2dd087cc37614c4e4c02e0750114b288a33d4004050eb1a478366b267f5dc", + "configs/examples/gymnasium/gr00t/walker2d_teacher.yaml": "10d965dba406fccaa90099e34913e4ce7e2205233f621b20d1c16aa33215678b" + }, + "starvla_overlay": { + "third_party/starVLA/examples/rl_games/train_files/data_registry/data_config.py": "dd399efd9983294e11372753827a80da3a9ed1e18c23aba8ab70a33e6a941738", + "third_party/starVLA/starVLA/training/train_starvla.py": "d81f39ff232b9d2b8c71c386dd68ccf73ebc0bce2296723a9485d0903af62547" + }, + "note": "v4 plus initialization seed before constructing new action/state modules; active IP FastTD3 remains frozen on v3.", + "version": "v5" + }, + "training_config_sha256": "2c74a1ac6ef531b6522634fe627191e85a3c99a14eaee04ab040c8f302278e00", + "dataset_manifest_sha256": "3d64134b941fbc090fab99ebc656e19ab2881ee00c89e81d8a8280a7714b3417" +} diff --git a/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/task_contract.json b/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/task_contract.json new file mode 100644 index 0000000000000000000000000000000000000000..f20338db89c82f965cce12d795b3641d45cdae58 --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwengr00t-h1/air_raid_zero_latency_gr00t_h1/task_contract.json @@ -0,0 +1,44 @@ +{ + "action_labels": [ + "noop", + "fire", + "right", + "left", + "rightfire", + "leftfire" + ], + "action_values": [ + 0, + 1, + 2, + 3, + 4, + 5 + ], + "base_prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.", + "env_fps": 50.0, + "env_id": "LatencyBench/AirRaid-v0", + "frame_stack": 4, + "make_kwargs": { + "base_env_id": "ALE/AirRaid-v5", + "base_make_kwargs": { + "difficulty": 0, + "frameskip": 1, + "full_action_space": false, + "max_num_frames_per_episode": 108000, + "mode": 1, + "obs_type": "rgb", + "repeat_action_probability": 0.0 + }, + "fire_reset": true, + "noop_max": 30, + "render_mode": "rgb_array", + "screen_size": 84 + }, + "noop_action_id": 0, + "obs_fps": 12.5, + "registration_imports": [ + "latency_bench.envs.gymnasium_air_raid" + ], + "task_name": "air_raid" +} diff --git a/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/README.md b/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/README.md new file mode 100644 index 0000000000000000000000000000000000000000..bb4aab934c9ef640075a618a5434f80a13de9907 --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/README.md @@ -0,0 +1,30 @@ +# air-raid / qwenoft + +Training condition: `zero-latency`. Run: `air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k`. + +Copied from [Standard-Pipeline](https://huggingface.co/latency-sensitive-bench/Standard-Pipeline/tree/2208875f92b21e3f8ec2363a30a6705dfe56cc04/air_raid/zero_latency/OpenVLA) at `2208875f92b21e3f8ec2363a30a6705dfe56cc04`. + +- Checkpoint: `checkpoints/model.pt` +- Selection: only published VLA checkpoint +- Checkpoint SHA256: `6972726bc2f5a8157a64c78cb5dd6a27c7253f9ae45a00b3015b4d0aa17580c8` +- Configuration, FPS, actions, normalization and latency semantics retain their source values. +- This import does not run evaluations or change training/evaluation/acceptance status. + +Action horizon is 1. Use `config.yaml` and `dataset_statistics.json` with the matching +prepared backbone. `task_contract.json` is extracted from the original training configuration. +Original machine paths in historical configurations are provenance: bind checkpoint, +backbone, manifest and profile paths to downloaded local files when running inference. + +Missing in this source model bundle: `manifest.json`, `latency_prompt_map.json`. +These files were not substituted with files from another training condition. + +## Download + +```python +from huggingface_hub import snapshot_download +prefix = "zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k" +snapshot_download("latency-sensitive-bench/benchmark-models", revision="", + allow_patterns=[prefix + "/**"]) +``` + +Use the verified import commit listed in the repository's import report. diff --git a/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/checkpoints/model.pt b/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/checkpoints/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..d30979c272f10edec0fb67b2a509cb5215a03301 --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/checkpoints/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:6972726bc2f5a8157a64c78cb5dd6a27c7253f9ae45a00b3015b4d0aa17580c8 +size 9785049835 diff --git a/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/config.full.yaml b/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/config.full.yaml new file mode 100644 index 0000000000000000000000000000000000000000..17bb980eeb5f69f697ae20606c7967842825f603 --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/config.full.yaml @@ -0,0 +1,329 @@ +framework: + name: QwenOFT + qwenvl: + base_vlm: /mnt/checkpoints/latency-sensitive-bench/openvla/Qwen3-VL-4B-Instruct + attn_implementation: flash_attention_2 + enable_gradient_checkpointing: true + action_model: + action_model_type: MLP + action_dim: 6 + action_hidden_dim: 2560 + future_action_window_size: 0 + past_action_window_size: 0 + loss_type: discrete_ce + state_dim: 1 + action_horizon: 1 + action_env_dim: 6 + kv_memory: + enabled: false + window: 4 + rollout_len: 8 + packed_train: false + rebased_sink: true +datasets: + vla_data: + dataset_py: lerobot_datasets + include_state: false + data_root_dir: /mnt/data/latency-sensitive-bench/vla_datasets/rl_games + data_mix: air_raid_l0_random_noop_fire_reset_return_gt2500_100ep + eval_data_mix: air_raid_l0_random_noop_fire_reset_return_gt2500_100ep__val + custom_mixtures_path: /mnt/data/latency-sensitive-bench/vla_datasets/rl_games/_generated_mixtures/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep.json + action_type: discrete + sequential_step_sampling: false + eval_sequential_step_sampling: null + num_workers: 8 + eval_num_workers: 8 + prefetch_factor: 4 + persistent_workers: true + pin_memory: true + shuffle: true + action_balance: + enabled: false + strategy: balanced_epoch + action_key: action_id + target_flap_fraction: 0.3 + noop_id: 0 + flap_id: 1 + latency_curriculum: + enabled: false + strategy: exclusive + latencies: null + phase_steps: null + phase_distributions: null + new_latency_passes: 1.0 + replay_passes: 0.25 + target_total_passes: 2.0 + final_equalization: true + step_budget_mode: auto + eval_at_phase_end: false + save_at_phase_end: false + computed_plan: null + per_device_batch_size: 16 + load_all_data_for_training: true + num_obs_frames: 1 + image_mode: single + prompt_mode: raw + stitch_grid: + - 2 + - 2 + obs_image_size: null + video_backend: torchvision_av + gymnasium_task_contract: + action_labels: + - noop + - fire + - right + - left + - rightfire + - leftfire + action_values: + - 0 + - 1 + - 2 + - 3 + - 4 + - 5 + base_prompt: 'Protect both buildings from flying saucers. Choose exactly one + action from: noop, fire, right, left, rightfire, leftfire.' + env_fps: 50.0 + env_id: LatencyBench/AirRaid-v0 + frame_stack: 4 + make_kwargs: + base_env_id: ALE/AirRaid-v5 + base_make_kwargs: + difficulty: 0 + frameskip: 1 + full_action_space: false + max_num_frames_per_episode: 108000 + mode: 1 + obs_type: rgb + repeat_action_probability: 0.0 + fire_reset: true + noop_max: 30 + render_mode: rgb_array + screen_size: 84 + noop_action_id: 0 + obs_fps: 12.5 + registration_imports: + - latency_bench.envs.gymnasium_air_raid + task_name: air_raid + active_action_dim: 6 +dataset: + source_hf: '' + config_name: null + source_subdir: null + converted_name: air_raid_l0_random_noop_fire_reset_return_gt2500_100ep + single_source_hf: '' + mixed_source_hf: '' + single_converted_name: air_raid_l0_random_noop_fire_reset_return_gt2500_100ep + mixed_converted_name: air_raid_l0_random_noop_fire_reset_return_gt2500_100ep + single_latency_filter: null + mixed_latency_filter: null + force_download: false + setup_force: false + skip_verification: false + target_latency_unit: observation_steps + verify_rows: 200 + max_episodes: null + episodes_per_latency: null + latency_filter: null + debug_subset: + enabled: false + max_episodes: 5 + suffix: debug +base_model: + repo_id: Qwen/Qwen3-VL-4B-Instruct +initialization: + checkpoint_local_dir: null + checkpoint_hf_repo_id: null + checkpoint_filename: null +trainer: + max_train_steps: 5000 + num_warmup_steps: 100 + save_interval: 5000 + eval_interval: 500 + eval_num_batches: 200 + per_latency_eval_num_batches: null + eval_action_classification: true + eval_action_classification_interval: null + cc_f1_tolerance: 1 + learning_rate: + base: 2.0e-05 + qwen_vl_interface: 1.0e-05 + action_model: 0.0001 + lr_scheduler_type: cosine_with_min_lr + scheduler_specific_kwargs: + min_lr: 1.0e-06 + freeze_modules: '' + freeze_vit: false + freeze_tied_embedding: false + freeze_llm_layers: [] + loss_scale: + vla: 1.0 + vlm: 0.1 + max_grad_norm: 1.0 + weight_decay: 0.0 + logging_frequency: 1 + profile_timing: + enabled: true + log_interval: 10 + gradient_clipping: 1.0 + gradient_accumulation_steps: 1 + distributed_backend: deepspeed + is_resume: false + pretrained_checkpoint: null + resume_step: 0 + reload_modules: null + optimizer: + name: AdamW + betas: + - 0.9 + - 0.95 + eps: 1.0e-08 + weight_decay: 1.0e-08 + fused: true + save_format: pt +workspace_dir: /home/lzj/code/latency-sensitive-bench +run_root_dir: /mnt/checkpoints/latency-sensitive-bench/openvla_runs +seed: 42 +wandb_entity: dongqianyu99-zhejiang-university +wandb_project: latency-sensitive-bench +auth: + env_file: null + hf_token_env: HF_TOKEN + wandb_api_key_env: WANDB_API_KEY +paths: + run_root_dir: /mnt/checkpoints/latency-sensitive-bench/openvla_runs + dataset_local_dir: /mnt/data/latency-sensitive-bench/vla_datasets/rl_games + dataset_cache_dir: null + base_model_dir: /mnt/checkpoints/latency-sensitive-bench/openvla/Qwen3-VL-4B-Instruct + accelerate_config: starVLA/config/deepseeds/deepspeed_zero2_cpu_offload.yaml +launch: + use_accelerate: true + gpus: null + num_processes: 1 + dry_run: false +conda: + enabled: true + env_name: null +rl_games: + model_alias: openvla + env_eval: + image_size: 224 + frameskip: 1 + image_transform: raw_rgb + prompt_mode: raw + ghost_trail: + history_frames: 5 + gamma: 1.3 + min_alpha: 35 + scroll_px_per_step: 4.0 + ground_fraction: 0.22 + seed: 42 + fixed_episode_seeds: true + latency_seed_stride: 0 + task_seed_stride: 0 + task_description: '' + eval_parallel_envs: 5 + action_chunk_execution: + enabled: false + chunk_size: null + enabled: false + gymnasium: + make_kwargs: {} + mid_train: + enabled: false + interval_steps: 100 + latencies: + - 0 + num_episodes: 5 + max_steps_per_episode: 3600 + post_train: + enabled: false + latencies: + - 0 + - 1 + - 2 + - 3 + - 4 + num_episodes: 5 + max_steps_per_episode: 3600 + eval_backend: latency_bench + distributed_mode: none + vectorized: + enabled: false + batch_size: 1 + latency: + prompt_map_path: null + mode: single + values: + - 0 + task: gymnasium + gymnasium: + task_contract: + action_labels: + - noop + - fire + - right + - left + - rightfire + - leftfire + action_values: + - 0 + - 1 + - 2 + - 3 + - 4 + - 5 + base_prompt: 'Protect both buildings from flying saucers. Choose exactly one + action from: noop, fire, right, left, rightfire, leftfire.' + env_fps: 50.0 + env_id: LatencyBench/AirRaid-v0 + frame_stack: 4 + make_kwargs: + base_env_id: ALE/AirRaid-v5 + base_make_kwargs: + difficulty: 0 + frameskip: 1 + full_action_space: false + max_num_frames_per_episode: 108000 + mode: 1 + obs_type: rgb + repeat_action_probability: 0.0 + fire_reset: true + noop_max: 30 + render_mode: rgb_array + screen_size: 84 + noop_action_id: 0 + obs_fps: 12.5 + registration_imports: + - latency_bench.envs.gymnasium_air_raid + task_name: air_raid + task_name: air_raid + initialization_mode: scratch + action_carrier: native +model: openvla +env: gymnasium +init: scratch +mode: single +checkpoint: + load: none + hf_repo_id: null + save_best_model: false + save_final_model: true + save_pt_file: true + save_training_state: false + save_safetensors_file: false + local: + keep_last_n: 1 + sync: + enabled: false + repo_id: null + keep_last_n: 0 + sync_every_n_checkpoints: 1 + resume_policy: local_latest +run_id: air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k +output_dir: /mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k +config_yaml: null +is_debug: false +version_id: '0.21' diff --git a/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/config.yaml b/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..17bb980eeb5f69f697ae20606c7967842825f603 --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/config.yaml @@ -0,0 +1,329 @@ +framework: + name: QwenOFT + qwenvl: + base_vlm: /mnt/checkpoints/latency-sensitive-bench/openvla/Qwen3-VL-4B-Instruct + attn_implementation: flash_attention_2 + enable_gradient_checkpointing: true + action_model: + action_model_type: MLP + action_dim: 6 + action_hidden_dim: 2560 + future_action_window_size: 0 + past_action_window_size: 0 + loss_type: discrete_ce + state_dim: 1 + action_horizon: 1 + action_env_dim: 6 + kv_memory: + enabled: false + window: 4 + rollout_len: 8 + packed_train: false + rebased_sink: true +datasets: + vla_data: + dataset_py: lerobot_datasets + include_state: false + data_root_dir: /mnt/data/latency-sensitive-bench/vla_datasets/rl_games + data_mix: air_raid_l0_random_noop_fire_reset_return_gt2500_100ep + eval_data_mix: air_raid_l0_random_noop_fire_reset_return_gt2500_100ep__val + custom_mixtures_path: /mnt/data/latency-sensitive-bench/vla_datasets/rl_games/_generated_mixtures/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep.json + action_type: discrete + sequential_step_sampling: false + eval_sequential_step_sampling: null + num_workers: 8 + eval_num_workers: 8 + prefetch_factor: 4 + persistent_workers: true + pin_memory: true + shuffle: true + action_balance: + enabled: false + strategy: balanced_epoch + action_key: action_id + target_flap_fraction: 0.3 + noop_id: 0 + flap_id: 1 + latency_curriculum: + enabled: false + strategy: exclusive + latencies: null + phase_steps: null + phase_distributions: null + new_latency_passes: 1.0 + replay_passes: 0.25 + target_total_passes: 2.0 + final_equalization: true + step_budget_mode: auto + eval_at_phase_end: false + save_at_phase_end: false + computed_plan: null + per_device_batch_size: 16 + load_all_data_for_training: true + num_obs_frames: 1 + image_mode: single + prompt_mode: raw + stitch_grid: + - 2 + - 2 + obs_image_size: null + video_backend: torchvision_av + gymnasium_task_contract: + action_labels: + - noop + - fire + - right + - left + - rightfire + - leftfire + action_values: + - 0 + - 1 + - 2 + - 3 + - 4 + - 5 + base_prompt: 'Protect both buildings from flying saucers. Choose exactly one + action from: noop, fire, right, left, rightfire, leftfire.' + env_fps: 50.0 + env_id: LatencyBench/AirRaid-v0 + frame_stack: 4 + make_kwargs: + base_env_id: ALE/AirRaid-v5 + base_make_kwargs: + difficulty: 0 + frameskip: 1 + full_action_space: false + max_num_frames_per_episode: 108000 + mode: 1 + obs_type: rgb + repeat_action_probability: 0.0 + fire_reset: true + noop_max: 30 + render_mode: rgb_array + screen_size: 84 + noop_action_id: 0 + obs_fps: 12.5 + registration_imports: + - latency_bench.envs.gymnasium_air_raid + task_name: air_raid + active_action_dim: 6 +dataset: + source_hf: '' + config_name: null + source_subdir: null + converted_name: air_raid_l0_random_noop_fire_reset_return_gt2500_100ep + single_source_hf: '' + mixed_source_hf: '' + single_converted_name: air_raid_l0_random_noop_fire_reset_return_gt2500_100ep + mixed_converted_name: air_raid_l0_random_noop_fire_reset_return_gt2500_100ep + single_latency_filter: null + mixed_latency_filter: null + force_download: false + setup_force: false + skip_verification: false + target_latency_unit: observation_steps + verify_rows: 200 + max_episodes: null + episodes_per_latency: null + latency_filter: null + debug_subset: + enabled: false + max_episodes: 5 + suffix: debug +base_model: + repo_id: Qwen/Qwen3-VL-4B-Instruct +initialization: + checkpoint_local_dir: null + checkpoint_hf_repo_id: null + checkpoint_filename: null +trainer: + max_train_steps: 5000 + num_warmup_steps: 100 + save_interval: 5000 + eval_interval: 500 + eval_num_batches: 200 + per_latency_eval_num_batches: null + eval_action_classification: true + eval_action_classification_interval: null + cc_f1_tolerance: 1 + learning_rate: + base: 2.0e-05 + qwen_vl_interface: 1.0e-05 + action_model: 0.0001 + lr_scheduler_type: cosine_with_min_lr + scheduler_specific_kwargs: + min_lr: 1.0e-06 + freeze_modules: '' + freeze_vit: false + freeze_tied_embedding: false + freeze_llm_layers: [] + loss_scale: + vla: 1.0 + vlm: 0.1 + max_grad_norm: 1.0 + weight_decay: 0.0 + logging_frequency: 1 + profile_timing: + enabled: true + log_interval: 10 + gradient_clipping: 1.0 + gradient_accumulation_steps: 1 + distributed_backend: deepspeed + is_resume: false + pretrained_checkpoint: null + resume_step: 0 + reload_modules: null + optimizer: + name: AdamW + betas: + - 0.9 + - 0.95 + eps: 1.0e-08 + weight_decay: 1.0e-08 + fused: true + save_format: pt +workspace_dir: /home/lzj/code/latency-sensitive-bench +run_root_dir: /mnt/checkpoints/latency-sensitive-bench/openvla_runs +seed: 42 +wandb_entity: dongqianyu99-zhejiang-university +wandb_project: latency-sensitive-bench +auth: + env_file: null + hf_token_env: HF_TOKEN + wandb_api_key_env: WANDB_API_KEY +paths: + run_root_dir: /mnt/checkpoints/latency-sensitive-bench/openvla_runs + dataset_local_dir: /mnt/data/latency-sensitive-bench/vla_datasets/rl_games + dataset_cache_dir: null + base_model_dir: /mnt/checkpoints/latency-sensitive-bench/openvla/Qwen3-VL-4B-Instruct + accelerate_config: starVLA/config/deepseeds/deepspeed_zero2_cpu_offload.yaml +launch: + use_accelerate: true + gpus: null + num_processes: 1 + dry_run: false +conda: + enabled: true + env_name: null +rl_games: + model_alias: openvla + env_eval: + image_size: 224 + frameskip: 1 + image_transform: raw_rgb + prompt_mode: raw + ghost_trail: + history_frames: 5 + gamma: 1.3 + min_alpha: 35 + scroll_px_per_step: 4.0 + ground_fraction: 0.22 + seed: 42 + fixed_episode_seeds: true + latency_seed_stride: 0 + task_seed_stride: 0 + task_description: '' + eval_parallel_envs: 5 + action_chunk_execution: + enabled: false + chunk_size: null + enabled: false + gymnasium: + make_kwargs: {} + mid_train: + enabled: false + interval_steps: 100 + latencies: + - 0 + num_episodes: 5 + max_steps_per_episode: 3600 + post_train: + enabled: false + latencies: + - 0 + - 1 + - 2 + - 3 + - 4 + num_episodes: 5 + max_steps_per_episode: 3600 + eval_backend: latency_bench + distributed_mode: none + vectorized: + enabled: false + batch_size: 1 + latency: + prompt_map_path: null + mode: single + values: + - 0 + task: gymnasium + gymnasium: + task_contract: + action_labels: + - noop + - fire + - right + - left + - rightfire + - leftfire + action_values: + - 0 + - 1 + - 2 + - 3 + - 4 + - 5 + base_prompt: 'Protect both buildings from flying saucers. Choose exactly one + action from: noop, fire, right, left, rightfire, leftfire.' + env_fps: 50.0 + env_id: LatencyBench/AirRaid-v0 + frame_stack: 4 + make_kwargs: + base_env_id: ALE/AirRaid-v5 + base_make_kwargs: + difficulty: 0 + frameskip: 1 + full_action_space: false + max_num_frames_per_episode: 108000 + mode: 1 + obs_type: rgb + repeat_action_probability: 0.0 + fire_reset: true + noop_max: 30 + render_mode: rgb_array + screen_size: 84 + noop_action_id: 0 + obs_fps: 12.5 + registration_imports: + - latency_bench.envs.gymnasium_air_raid + task_name: air_raid + task_name: air_raid + initialization_mode: scratch + action_carrier: native +model: openvla +env: gymnasium +init: scratch +mode: single +checkpoint: + load: none + hf_repo_id: null + save_best_model: false + save_final_model: true + save_pt_file: true + save_training_state: false + save_safetensors_file: false + local: + keep_last_n: 1 + sync: + enabled: false + repo_id: null + keep_last_n: 0 + sync_every_n_checkpoints: 1 + resume_policy: local_latest +run_id: air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k +output_dir: /mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k +config_yaml: null +is_debug: false +version_id: '0.21' diff --git a/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/dataset_statistics.json b/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/dataset_statistics.json new file mode 100644 index 0000000000000000000000000000000000000000..39af401e32da1e3219fdbe3ffbe7d00c7d2b8f1b --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/dataset_statistics.json @@ -0,0 +1,84 @@ +{ + "new_embodiment": { + "action": { + "mean": [ + 0.17179012298583984, + 0.27208641171455383, + 0.05909876525402069, + 0.2254074066877365, + 0.20286419987678528, + 0.06875308603048325 + ], + "std": [ + 0.3770434260368347, + 0.4451958239078522, + 0.23573854565620422, + 0.41779401898384094, + 0.4020724594593048, + 0.25303277373313904 + ], + "max": [ + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0 + ], + "min": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "q01": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "q99": [ + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0 + ], + "mask": [ + true, + true, + true, + true, + true, + true + ] + }, + "state": { + "mean": [ + 0.0 + ], + "std": [ + 0.0 + ], + "max": [ + 0.0 + ], + "min": [ + 0.0 + ], + "q01": [ + 0.0 + ], + "q99": [ + 0.0 + ] + }, + "num_transitions": 81000, + "num_trajectories": 90 + } +} \ No newline at end of file diff --git a/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/dataset_statistics_eval.json b/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/dataset_statistics_eval.json new file mode 100644 index 0000000000000000000000000000000000000000..84dbc4a9bf6b18b5424f5b895829493e7ec9e03e --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/dataset_statistics_eval.json @@ -0,0 +1,84 @@ +{ + "new_embodiment": { + "action": { + "mean": [ + 0.16411110758781433, + 0.2808888852596283, + 0.06244444474577904, + 0.22011111676692963, + 0.1997777819633484, + 0.07266666740179062 + ], + "std": [ + 0.3703823387622833, + 0.44942164421081543, + 0.24197007715702057, + 0.41433459520339966, + 0.39984217286109924, + 0.25959575176239014 + ], + "max": [ + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0 + ], + "min": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "q01": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "q99": [ + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0 + ], + "mask": [ + true, + true, + true, + true, + true, + true + ] + }, + "state": { + "mean": [ + 0.0 + ], + "std": [ + 0.0 + ], + "max": [ + 0.0 + ], + "min": [ + 0.0 + ], + "q01": [ + 0.0 + ], + "q99": [ + 0.0 + ] + }, + "num_transitions": 9000, + "num_trajectories": 10 + } +} \ No newline at end of file diff --git a/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/eval/post_train/step_5000.json b/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/eval/post_train/step_5000.json new file mode 100644 index 0000000000000000000000000000000000000000..53f5793703cdeab3d6bf2c1eefd118183fb5f01f --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/eval/post_train/step_5000.json @@ -0,0 +1,219 @@ +{ + "per_latency": { + "gymnasium/latency_0": { + "latency": 0, + "num_episodes": 20, + "mean_reward": 2000.0, + "mean_length": 3106.9, + "std_reward": 706.1338400048535, + "std_length": 963.5145510058476, + "episode_rewards": [ + 650.0, + 2450.0, + 650.0, + 2000.0, + 2050.0, + 2675.0, + 2525.0, + 2375.0, + 2375.0, + 2675.0, + 2600.0, + 2675.0, + 1550.0, + 2050.0, + 650.0, + 2050.0, + 2675.0, + 2050.0, + 2375.0, + 900.0 + ], + "episode_lengths": [ + 1006, + 3600, + 1006, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3600, + 3295, + 3600, + 1006, + 3600, + 3600, + 3600, + 3600, + 1825 + ], + "decoded_action_hist": {}, + "fixed_episode_seeds": true, + "eval_seed": 42, + "episode_seeds": [ + null, + null, + null, + null, + null, + null, + null, + null, + null, + null, + null, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + "episode_indices": [ + 0, + 1, + 2, + 3, + 4, + 5, + 6, + 7, + 8, + 9, + 10, + 11, + 12, + 13, + 14, + 15, + 16, + 17, + 18, + 19 + ] + }, + "gymnasium/latency_2": { + "latency": 2, + "num_episodes": 20, + "mean_reward": 681.25, + "mean_length": 2449.75, + "std_reward": 277.9247874875503, + "std_length": 754.2512098101004, + "episode_rewards": [ + 225.0, + 650.0, + 225.0, + 650.0, + 850.0, + 625.0, + 1225.0, + 525.0, + 700.0, + 625.0, + 1175.0, + 625.0, + 925.0, + 825.0, + 225.0, + 500.0, + 625.0, + 825.0, + 525.0, + 1075.0 + ], + "episode_lengths": [ + 990, + 2023, + 990, + 2711, + 3575, + 2424, + 2662, + 2491, + 2281, + 2424, + 3600, + 2424, + 2319, + 3052, + 990, + 2472, + 2424, + 3052, + 2491, + 3600 + ], + "decoded_action_hist": {}, + "fixed_episode_seeds": true, + "eval_seed": 42, + "episode_seeds": [ + null, + null, + null, + null, + null, + null, + null, + null, + null, + null, + null, + null, + null, + null, + null, + null, + null, + null, + null, + null + ], + "episode_indices": [ + 0, + 1, + 2, + 3, + 4, + 5, + 6, + 7, + 8, + 9, + 10, + 11, + 12, + 13, + 14, + 15, + 16, + 17, + 18, + 19 + ] + } + }, + "aggregate": { + "stage": "post_train", + "step": 5000, + "task": "gymnasium", + "model_alias": "openvla", + "fixed_episode_seeds": true, + "eval_seed": 42, + "total_episodes": 40, + "mean_reward": 1340.625, + "mean_length": 2778.325, + "std_reward": 850.1229230970072, + "std_length": 925.5209988838719, + "task_count": 1, + "macro_mean_reward": 1340.625, + "macro_mean_length": 2778.325, + "distributed_eval": false + } +} \ No newline at end of file diff --git a/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0/episode_metrics.jsonl b/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0/episode_metrics.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..9b69313eea9cca5640d8f979234de444c67d0642 --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0/episode_metrics.jsonl @@ -0,0 +1,20 @@ +{"episode_id": 0, "episode_return": 650.0, "episode_return_env": 650.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 42, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 252, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 1006} +{"episode_id": 1, "episode_return": 2450.0, "episode_return_env": 2450.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 43, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 900, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 3600} +{"episode_id": 2, "episode_return": 650.0, "episode_return_env": 650.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 44, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 252, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 1006} +{"episode_id": 3, "episode_return": 2000.0, "episode_return_env": 2000.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 45, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 900, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 3600} +{"episode_id": 4, "episode_return": 2050.0, "episode_return_env": 2050.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 46, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 900, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 3600} +{"episode_id": 5, "episode_return": 2675.0, "episode_return_env": 2675.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 47, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 900, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 3600} +{"episode_id": 6, "episode_return": 2525.0, "episode_return_env": 2525.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 48, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 900, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 3600} +{"episode_id": 7, "episode_return": 2375.0, "episode_return_env": 2375.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 49, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 900, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 3600} +{"episode_id": 8, "episode_return": 2375.0, "episode_return_env": 2375.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 50, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 900, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 3600} +{"episode_id": 9, "episode_return": 2675.0, "episode_return_env": 2675.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 51, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 900, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 3600} +{"episode_id": 10, "episode_return": 2600.0, "episode_return_env": 2600.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 52, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 900, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 3600} +{"episode_id": 11, "episode_return": 2675.0, "episode_return_env": 2675.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 53, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 900, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 3600} +{"episode_id": 12, "episode_return": 1550.0, "episode_return_env": 1550.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 54, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 824, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 3295} +{"episode_id": 13, "episode_return": 2050.0, "episode_return_env": 2050.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 55, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 900, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 3600} +{"episode_id": 14, "episode_return": 650.0, "episode_return_env": 650.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 56, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 252, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 1006} +{"episode_id": 15, "episode_return": 2050.0, "episode_return_env": 2050.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 57, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 900, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 3600} +{"episode_id": 16, "episode_return": 2675.0, "episode_return_env": 2675.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 58, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 900, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 3600} +{"episode_id": 17, "episode_return": 2050.0, "episode_return_env": 2050.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 59, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 900, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 3600} +{"episode_id": 18, "episode_return": 2375.0, "episode_return_env": 2375.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 60, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 900, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 3600} +{"episode_id": 19, "episode_return": 900.0, "episode_return_env": 900.0, "game_score": null, "mean_latency_ms": 0.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 61, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "zero", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_0", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 457, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 0.0, "p99_latency_ms": 0.0, "return_raw": null, "survival_steps": 1825} diff --git a/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2/episode_metrics.jsonl b/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2/episode_metrics.jsonl new file mode 100644 index 0000000000000000000000000000000000000000..be70251b687ba1861e8ce1300ea0264f83e5c314 --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2/episode_metrics.jsonl @@ -0,0 +1,20 @@ +{"episode_id": 0, "episode_return": 225.0, "episode_return_env": 225.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 42, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 248, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 990} +{"episode_id": 1, "episode_return": 650.0, "episode_return_env": 650.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 43, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 506, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 2023} +{"episode_id": 2, "episode_return": 225.0, "episode_return_env": 225.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 44, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 248, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 990} +{"episode_id": 3, "episode_return": 650.0, "episode_return_env": 650.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 45, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 678, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 2711} +{"episode_id": 4, "episode_return": 850.0, "episode_return_env": 850.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 46, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 894, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 3575} +{"episode_id": 5, "episode_return": 625.0, "episode_return_env": 625.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 47, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 606, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 2424} +{"episode_id": 6, "episode_return": 1225.0, "episode_return_env": 1225.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 48, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 666, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 2662} +{"episode_id": 7, "episode_return": 525.0, "episode_return_env": 525.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 49, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 623, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 2491} +{"episode_id": 8, "episode_return": 700.0, "episode_return_env": 700.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 50, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 571, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 2281} +{"episode_id": 9, "episode_return": 625.0, "episode_return_env": 625.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 51, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 606, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 2424} +{"episode_id": 10, "episode_return": 1175.0, "episode_return_env": 1175.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 52, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 900, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 3600} +{"episode_id": 11, "episode_return": 625.0, "episode_return_env": 625.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 53, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 606, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 2424} +{"episode_id": 12, "episode_return": 925.0, "episode_return_env": 925.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 54, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 580, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 2319} +{"episode_id": 13, "episode_return": 825.0, "episode_return_env": 825.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 55, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 763, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 3052} +{"episode_id": 14, "episode_return": 225.0, "episode_return_env": 225.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 56, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 248, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 990} +{"episode_id": 15, "episode_return": 500.0, "episode_return_env": 500.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 57, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 618, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 2472} +{"episode_id": 16, "episode_return": 625.0, "episode_return_env": 625.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 58, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 606, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 2424} +{"episode_id": 17, "episode_return": 825.0, "episode_return_env": 825.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 59, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 763, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 3052} +{"episode_id": 18, "episode_return": 525.0, "episode_return_env": 525.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 60, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 623, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 2491} +{"episode_id": 19, "episode_return": 1075.0, "episode_return_env": 1075.0, "game_score": null, "mean_latency_ms": 160.0, "metadata": {"checkpoint_path": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "config_name": "starvla_gymnasium_train_eval", "env_fps": 50.0, "env_id": "LatencyBench/AirRaid-v0", "episode_seed": 61, "frame_ms": 20.0, "gpu_class": null, "hf_commit": null, "instance_id": null, "latency_type": "fixed", "mode": "simulated", "model_id": null, "obs_fps": 12.5, "output_dir": "/mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/latency_bench_eval/post_train/step_5000/gymnasium/latency_2", "policy_id": "starvla", "profile_ref": null, "run_name": "starvla_gymnasium_train_eval", "source_run_id": null, "workload_id": null}, "num_actions": 900, "num_dropped_actions": 0, "num_invalid_actions": 0, "p90_latency_ms": 160.0, "p99_latency_ms": 160.0, "return_raw": null, "survival_steps": 3600} diff --git a/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/provenance.json b/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/provenance.json new file mode 100644 index 0000000000000000000000000000000000000000..03d3790ca0154b2fd9b4b870875f0a4461ce1302 --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/provenance.json @@ -0,0 +1,37 @@ +{ + "task": "air-raid", + "model": "qwenoft", + "training_condition": "zero-latency", + "training_run_id": "air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k", + "source": { + "repo_id": "latency-sensitive-bench/Standard-Pipeline", + "revision": "2208875f92b21e3f8ec2363a30a6705dfe56cc04", + "path": "air_raid/zero_latency/OpenVLA", + "url": "https://huggingface.co/latency-sensitive-bench/Standard-Pipeline/tree/2208875f92b21e3f8ec2363a30a6705dfe56cc04/air_raid/zero_latency/OpenVLA" + }, + "hf_repo_id": "latency-sensitive-bench/benchmark-models", + "hf_path": "zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k", + "checkpoint": { + "source_file": "air_raid/zero_latency/OpenVLA/checkpoints/steps_5000_pytorch_model.pt", + "source_sha256": "6972726bc2f5a8157a64c78cb5dd6a27c7253f9ae45a00b3015b4d0aa17580c8", + "source_bytes": 9785049835, + "file": "checkpoints/model.pt", + "selection_rule": "only published VLA checkpoint", + "method": "copy", + "sha256": "6972726bc2f5a8157a64c78cb5dd6a27c7253f9ae45a00b3015b4d0aa17580c8", + "bytes": 9785049835 + }, + "config_source": "air_raid/zero_latency/OpenVLA/config.full.yaml", + "task_contract_file": "task_contract.json", + "missing_source_files": [ + "manifest.json", + "latency_prompt_map.json" + ], + "migration": { + "operation": "copy_selected_inference_bundle", + "source_modified": false, + "evaluation_performed": false, + "experiment_status": "preserved_in_source_evidence" + }, + "action_horizon": 1 +} diff --git a/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/source/config.yaml b/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/source/config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..92720560e2dfbc1a883701e08f6ec07f79c3ffcf --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/source/config.yaml @@ -0,0 +1,84 @@ +checkpoint: + local: + keep_last_n: 1 + save_best_model: false + save_final_model: true + save_pt_file: true + save_safetensors_file: false + save_training_state: false + sync: + enabled: false + keep_last_n: 0 + repo_id: null +datasets: + vla_data: + data_mix: air_raid_l0_random_noop_fire_reset_return_gt2500_100ep + dataset_py: lerobot_datasets + eval_data_mix: air_raid_l0_random_noop_fire_reset_return_gt2500_100ep__val + latency_curriculum: + enabled: false + per_device_batch_size: 16 +framework: + action_model: + action_dim: 6 + action_env_dim: 6 + action_hidden_dim: 2560 + action_horizon: 1 + action_model_type: MLP + loss_type: discrete_ce + kv_memory: + enabled: false + packed_train: false + rebased_sink: true + rollout_len: 8 + window: 4 + name: QwenOFT + qwenvl: + attn_implementation: flash_attention_2 + base_vlm: /mnt/checkpoints/latency-sensitive-bench/openvla/Qwen3-VL-4B-Instruct + enable_gradient_checkpointing: true +output_dir: /mnt/checkpoints/latency-sensitive-bench/openvla_runs/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k +rl_games: + env_eval: + enabled: false + task: gymnasium +run_id: air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k +run_root_dir: /mnt/checkpoints/latency-sensitive-bench/openvla_runs +seed: 42 +trainer: + distributed_backend: deepspeed + eval_action_classification: true + eval_action_classification_interval: null + eval_interval: 500 + eval_num_batches: 200 + freeze_llm_layers: [] + freeze_modules: '' + freeze_tied_embedding: false + freeze_vit: false + gradient_accumulation_steps: 1 + is_resume: false + learning_rate: + action_model: 0.0001 + base: 2.0e-05 + qwen_vl_interface: 1.0e-05 + logging_frequency: 1 + lr_scheduler_type: cosine_with_min_lr + max_train_steps: 5000 + num_warmup_steps: 100 + optimizer: + betas: + - 0.9 + - 0.95 + eps: 1.0e-08 + fused: true + weight_decay: 1.0e-08 + per_latency_eval_num_batches: null + pretrained_checkpoint: null + profile_timing: + enabled: true + log_interval: 10 + save_interval: 5000 + scheduler_specific_kwargs: + min_lr: 1.0e-06 +wandb_entity: dongqianyu99-zhejiang-university +wandb_project: latency-sensitive-bench diff --git a/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/task_contract.json b/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/task_contract.json new file mode 100644 index 0000000000000000000000000000000000000000..f20338db89c82f965cce12d795b3641d45cdae58 --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwenoft-h1/air_raid_l0_random_noop_fire_reset_return_gt2500_100ep_openvla_native_sft_5k/task_contract.json @@ -0,0 +1,44 @@ +{ + "action_labels": [ + "noop", + "fire", + "right", + "left", + "rightfire", + "leftfire" + ], + "action_values": [ + 0, + 1, + 2, + 3, + 4, + 5 + ], + "base_prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.", + "env_fps": 50.0, + "env_id": "LatencyBench/AirRaid-v0", + "frame_stack": 4, + "make_kwargs": { + "base_env_id": "ALE/AirRaid-v5", + "base_make_kwargs": { + "difficulty": 0, + "frameskip": 1, + "full_action_space": false, + "max_num_frames_per_episode": 108000, + "mode": 1, + "obs_type": "rgb", + "repeat_action_probability": 0.0 + }, + "fire_reset": true, + "noop_max": 30, + "render_mode": "rgb_array", + "screen_size": 84 + }, + "noop_action_id": 0, + "obs_fps": 12.5, + "registration_imports": [ + "latency_bench.envs.gymnasium_air_raid" + ], + "task_name": "air_raid" +} diff --git a/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/README.md b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/README.md new file mode 100644 index 0000000000000000000000000000000000000000..adc3b7c9544ec7ede07125be36bf9384f6f6f3fc --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/README.md @@ -0,0 +1,31 @@ +# air-raid / qwenpi_v3 + +Training condition: `zero-latency`. Run: `airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2`. + +Copied from [Standard-Pipeline](https://huggingface.co/latency-sensitive-bench/Standard-Pipeline/tree/2208875f92b21e3f8ec2363a30a6705dfe56cc04/air_raid/zero_latency/Pi05) at `2208875f92b21e3f8ec2363a30a6705dfe56cc04`. + +- Checkpoint: `checkpoints/model.pt` +- Selection: only published VLA checkpoint +- Checkpoint SHA256: `c9ad50c5229424c919bb108d75b016f3b1815405ebd03adbea08137c18849bd8` +- Configuration, FPS, actions, normalization and latency semantics retain their source values. +- This import does not run evaluations or change training/evaluation/acceptance status. + +Action horizon is 1. Use `config.yaml` and `dataset_statistics.json` with the matching +prepared backbone. `task_contract.json` is extracted from the original training configuration. +Original machine paths in historical configurations are provenance: bind checkpoint, +backbone, manifest and profile paths to downloaded local files when running inference. + +See [the original model notes](source/README.md) for experiment results and limitations. + +[Original provenance](source/provenance.json) retains dataset references and recorded experiment evidence. + +## Download + +```python +from huggingface_hub import snapshot_download +prefix = "zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2" +snapshot_download("latency-sensitive-bench/benchmark-models", revision="", + allow_patterns=[prefix + "/**"]) +``` + +Use the verified import commit listed in the repository's import report. diff --git a/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/checkpoints/model.pt b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/checkpoints/model.pt new file mode 100644 index 0000000000000000000000000000000000000000..d590f8158d9f4c4fe32d3028bb430fc3b302e19e --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/checkpoints/model.pt @@ -0,0 +1,3 @@ +version https://git-lfs.github.com/spec/v1 +oid sha256:c9ad50c5229424c919bb108d75b016f3b1815405ebd03adbea08137c18849bd8 +size 10920516461 diff --git a/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/config.full.yaml b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/config.full.yaml new file mode 100644 index 0000000000000000000000000000000000000000..17d519ddc6e5caee25579af5f12afa30394fc1dd --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/config.full.yaml @@ -0,0 +1,233 @@ +framework: + name: QwenPI_v3 + qwenvl: + base_vlm: ${oc.env:PI05_BACKBONE_DIR} + attn_implementation: flash_attention_2 + vl_hidden_dim: 2560 + num_vl_layers: 36 + enable_gradient_checkpointing: false + action_model: + action_model_type: LayerwiseFM + action_dim: 6 + state_dim: 0 + state_encoding: continuous_projector + task_objective: null + action_horizon: 1 + repeated_diffusion_steps: 2 + num_inference_timesteps: 4 + add_pos_embed: true + max_seq_len: 1024 + num_target_vision_tokens: 32 + noise_beta_alpha: 1.5 + noise_beta_beta: 1.0 + noise_s: 0.999 + num_timestep_buckets: 1000 + diffusion_model_cfg: + action_dit_hidden_dim: 1024 + dropout: 0.2 + final_dropout: true + interleave_self_attention: true + norm_type: ada_norm + positional_embeddings: null + attention_head_dim: 64 + num_layers: 36 + input_embedding_dim: 1024 + cross_attention_dim: 1024 + output_dim: 1024 + num_attention_heads: 16 + action_env_dim: 6 + future_action_window_size: 0 + past_action_window_size: 0 +datasets: + vla_data: + dataset_py: lerobot_datasets + include_state: false + data_root_dir: / + data_mix: train + eval_data_mix: validation + custom_mixtures_path: ${PI05_RUN_DIR}/zero_latency/vla/mixture.json + action_type: discrete + sequential_step_sampling: true + eval_sequential_step_sampling: null + num_workers: 4 + eval_num_workers: 0 + prefetch_factor: 2 + persistent_workers: true + pin_memory: true + shuffle: true + action_balance: + enabled: false + strategy: balanced_epoch + action_key: action_id + target_flap_fraction: 0.3 + noop_id: 0 + flap_id: 1 + latency_curriculum: + enabled: false + strategy: exclusive + latencies: null + phase_steps: null + phase_distributions: null + new_latency_passes: 1.0 + replay_passes: 0.25 + target_total_passes: 2.0 + final_equalization: true + step_budget_mode: auto + eval_at_phase_end: false + save_at_phase_end: false + computed_plan: null + per_device_batch_size: 64 + load_all_data_for_training: true + num_obs_frames: 1 + image_mode: single + prompt_mode: raw + stitch_grid: + - 2 + - 2 + obs_image_size: + - 224 + - 224 + video_backend: pyav + lerobot_version: v2.0 + gymnasium_task_contract: + action_labels: + - noop + - fire + - right + - left + - rightfire + - leftfire + action_values: + - 0 + - 1 + - 2 + - 3 + - 4 + - 5 + base_prompt: 'Protect both buildings from flying saucers. Choose exactly one + action from: noop, fire, right, left, rightfire, leftfire.' + env_fps: 10 + env_id: LatencyBench/AirRaid-v0 + frame_stack: 4 + make_kwargs: + base_env_id: ALE/AirRaid-v5 + base_make_kwargs: + difficulty: 0 + frameskip: 1 + full_action_space: false + max_num_frames_per_episode: 108000 + mode: 1 + obs_type: rgb + repeat_action_probability: 0.0 + fire_reset: true + noop_max: 30 + render_mode: rgb_array + screen_size: 84 + noop_action_id: 0 + obs_fps: 2.5 + registration_imports: + - latency_bench.envs.gymnasium_air_raid + task_name: air_raid + active_action_dim: 6 +dataset: + source_hf: '' + config_name: null + source_subdir: null + converted_name: flappy_train + single_source_hf: '' + mixed_source_hf: '' + single_converted_name: flappy_train + mixed_converted_name: flappy_mixed_latency_train + single_latency_filter: null + mixed_latency_filter: null + force_download: false + setup_force: false + skip_verification: false + target_latency_unit: observation_steps + verify_rows: 200 + max_episodes: null + episodes_per_latency: null + latency_filter: null + debug_subset: + enabled: false + max_episodes: 5 + suffix: debug +base_model: + repo_id: Qwen/Qwen3-VL-4B-Instruct +initialization: + checkpoint_local_dir: null + checkpoint_hf_repo_id: null + checkpoint_filename: null +trainer: + max_train_steps: 5000 + num_warmup_steps: 100 + save_interval: 500 + eval_interval: 500 + eval_num_batches: 200 + per_latency_eval_num_batches: null + eval_action_classification: false + eval_action_classification_interval: null + cc_f1_tolerance: 1 + learning_rate: + base: 2.0e-05 + qwen_vl_interface: 1.0e-05 + action_model: 0.0001 + lr_scheduler_type: cosine_with_min_lr + scheduler_specific_kwargs: + min_lr: 1.0e-06 + freeze_modules: '' + freeze_vit: false + freeze_tied_embedding: false + freeze_llm_layers: [] + loss_scale: + vla: 1.0 + vlm: 0.1 + max_grad_norm: 1.0 + weight_decay: 0.0 + logging_frequency: 1 + profile_timing: + enabled: false + log_interval: 10 + gradient_clipping: 1.0 + gradient_accumulation_steps: 1 + distributed_backend: deepspeed + is_resume: false + pretrained_checkpoint: null + resume_step: 0 + reload_modules: null + optimizer: + name: AdamW + betas: + - 0.9 + - 0.95 + eps: 1.0e-08 + weight_decay: 1.0e-08 + fused: true + save_format: pt +run_id: airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2 +run_root_dir: ${PI05_RUN_DIR}/zero_latency/vla/training +seed: 42 +is_debug: false +version_id: '0.21' +wandb_project: latency-sensitive-bench +wandb_entity: dongqianyu99-zhejiang-university +wandb_name: airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2 +wandb_group: pi05-seven-env +wandb_tags: +- air_raid +- zero_latency +- Pi05 +- h1 +training_latency_condition: zero_latency +backbone: + repo_id: Qwen/Qwen3-VL-4B-Instruct + revision: ebb281ec70b05090aa6165b016eac8ec08e71b17 +checkpoint: + save_best_model: false + save_final_model: true + save_pt_file: true + save_training_state: true + local: + keep_last_n: 1 +config_yaml: ${PI05_RUN_DIR}/zero_latency/train.yaml +output_dir: ${PI05_RUN_DIR}/zero_latency/vla/training/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2 diff --git a/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/config.yaml b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/config.yaml new file mode 100644 index 0000000000000000000000000000000000000000..1bcd2368f380f567ae286a36721be079d31a1af7 --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/config.yaml @@ -0,0 +1,110 @@ +framework: + name: QwenPI_v3 + qwenvl: + base_vlm: ${oc.env:PI05_BACKBONE_DIR} + attn_implementation: flash_attention_2 + vl_hidden_dim: 2560 + num_vl_layers: 36 + enable_gradient_checkpointing: false + action_model: + action_model_type: LayerwiseFM + action_dim: 6 + state_dim: 0 + state_encoding: continuous_projector + task_objective: null + action_horizon: 1 + repeated_diffusion_steps: 2 + num_inference_timesteps: 4 + add_pos_embed: true + max_seq_len: 1024 + num_target_vision_tokens: 32 + noise_beta_alpha: 1.5 + noise_beta_beta: 1.0 + noise_s: 0.999 + num_timestep_buckets: 1000 + diffusion_model_cfg: + action_dit_hidden_dim: 1024 + dropout: 0.2 + final_dropout: true + interleave_self_attention: true + norm_type: ada_norm + positional_embeddings: null + attention_head_dim: 64 + num_layers: 36 + input_embedding_dim: 1024 + cross_attention_dim: 1024 + output_dim: 1024 + num_attention_heads: 16 + action_env_dim: 6 + future_action_window_size: 0 + past_action_window_size: 0 +version_id: '0.21' +datasets: + vla_data: + include_state: false + obs_image_size: + - 224 + - 224 + image_mode: single + stitch_grid: + - 2 + - 2 + num_obs_frames: 1 + gymnasium_task_contract: &id001 + action_labels: + - noop + - fire + - right + - left + - rightfire + - leftfire + action_values: + - 0 + - 1 + - 2 + - 3 + - 4 + - 5 + base_prompt: 'Protect both buildings from flying saucers. Choose exactly one + action from: noop, fire, right, left, rightfire, leftfire.' + env_fps: 10 + env_id: LatencyBench/AirRaid-v0 + frame_stack: 4 + make_kwargs: + base_env_id: ALE/AirRaid-v5 + base_make_kwargs: + difficulty: 0 + frameskip: 1 + full_action_space: false + max_num_frames_per_episode: 108000 + mode: 1 + obs_type: rgb + repeat_action_probability: 0.0 + fire_reset: true + noop_max: 30 + render_mode: rgb_array + screen_size: 84 + noop_action_id: 0 + obs_fps: 2.5 + registration_imports: + - latency_bench.envs.gymnasium_air_raid + task_name: air_raid +trainer: + pretrained_checkpoint: null + profile_timing: + enabled: false + log_interval: 10 +backbone: + repo_id: Qwen/Qwen3-VL-4B-Instruct + revision: ebb281ec70b05090aa6165b016eac8ec08e71b17 +rl_games: + model_alias: pi05 + task: gymnasium + action_carrier: native + gymnasium: + task_contract: *id001 + env_eval: + image_size: 224 + frameskip: 1 + image_transform: raw_rgb + prompt_mode: raw diff --git a/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/dataset_statistics.json b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/dataset_statistics.json new file mode 100644 index 0000000000000000000000000000000000000000..39af401e32da1e3219fdbe3ffbe7d00c7d2b8f1b --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/dataset_statistics.json @@ -0,0 +1,84 @@ +{ + "new_embodiment": { + "action": { + "mean": [ + 0.17179012298583984, + 0.27208641171455383, + 0.05909876525402069, + 0.2254074066877365, + 0.20286419987678528, + 0.06875308603048325 + ], + "std": [ + 0.3770434260368347, + 0.4451958239078522, + 0.23573854565620422, + 0.41779401898384094, + 0.4020724594593048, + 0.25303277373313904 + ], + "max": [ + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0 + ], + "min": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "q01": [ + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0 + ], + "q99": [ + 1.0, + 1.0, + 1.0, + 1.0, + 1.0, + 1.0 + ], + "mask": [ + true, + true, + true, + true, + true, + true + ] + }, + "state": { + "mean": [ + 0.0 + ], + "std": [ + 0.0 + ], + "max": [ + 0.0 + ], + "min": [ + 0.0 + ], + "q01": [ + 0.0 + ], + "q99": [ + 0.0 + ] + }, + "num_transitions": 81000, + "num_trajectories": 90 + } +} \ No newline at end of file diff --git a/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/evaluation/comparison.json b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/evaluation/comparison.json new file mode 100644 index 0000000000000000000000000000000000000000..b38e36e2589bebdba948982cde8e47d4b2413d19 --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/evaluation/comparison.json @@ -0,0 +1,233 @@ +{ + "state": "MATCHED_B_F20_AUDIT_PASSED", + "B": { + "model_revision": "de49831bd5dc9d642058bff0237e6ff2c9127b53", + "checkpoint_sha256": "c9ad50c5229424c919bb108d75b016f3b1815405ebd03adbea08137c18849bd8", + "seeds": [ + 42, + 43, + 44, + 45, + 46, + 47, + 48, + 49, + 50, + 51, + 52, + 53, + 54, + 55, + 56, + 57, + 58, + 59, + 60, + 61 + ], + "returns": [ + 1900.0, + 1300.0, + 1625.0, + 1800.0, + 450.0, + 1400.0, + 1775.0, + 1450.0, + 775.0, + 1325.0, + 800.0, + 775.0, + 1400.0, + 1425.0, + 1650.0, + 1500.0, + 1225.0, + 2100.0, + 1025.0, + 2025.0 + ], + "mean_return": 1386.25, + "population_std_return": 434.9910200222529 + }, + "F": { + "model_revision": "ac837400f55e4a9a96cc757dbcc516711b37e75f", + "checkpoint_sha256": "7914e1f1e2ba0b78d29c15dd33d96bbb6a5f61bad6ebb17f583a6d6d9958fb14", + "seeds": [ + 42, + 43, + 44, + 45, + 46, + 47, + 48, + 49, + 50, + 51, + 52, + 53, + 54, + 55, + 56, + 57, + 58, + 59, + 60, + 61 + ], + "returns": [ + 4125.0, + 4000.0, + 4125.0, + 4250.0, + 900.0, + 4125.0, + 3400.0, + 3775.0, + 3450.0, + 2675.0, + 2400.0, + 3775.0, + 350.0, + 2525.0, + 4125.0, + 3250.0, + 4125.0, + 1800.0, + 3775.0, + 3600.0 + ], + "mean_return": 3227.5, + "population_std_return": 1095.5848894540304 + }, + "paired": [ + { + "seed": 42, + "B": 1900.0, + "F": 4125.0, + "delta": 2225.0 + }, + { + "seed": 43, + "B": 1300.0, + "F": 4000.0, + "delta": 2700.0 + }, + { + "seed": 44, + "B": 1625.0, + "F": 4125.0, + "delta": 2500.0 + }, + { + "seed": 45, + "B": 1800.0, + "F": 4250.0, + "delta": 2450.0 + }, + { + "seed": 46, + "B": 450.0, + "F": 900.0, + "delta": 450.0 + }, + { + "seed": 47, + "B": 1400.0, + "F": 4125.0, + "delta": 2725.0 + }, + { + "seed": 48, + "B": 1775.0, + "F": 3400.0, + "delta": 1625.0 + }, + { + "seed": 49, + "B": 1450.0, + "F": 3775.0, + "delta": 2325.0 + }, + { + "seed": 50, + "B": 775.0, + "F": 3450.0, + "delta": 2675.0 + }, + { + "seed": 51, + "B": 1325.0, + "F": 2675.0, + "delta": 1350.0 + }, + { + "seed": 52, + "B": 800.0, + "F": 2400.0, + "delta": 1600.0 + }, + { + "seed": 53, + "B": 775.0, + "F": 3775.0, + "delta": 3000.0 + }, + { + "seed": 54, + "B": 1400.0, + "F": 350.0, + "delta": -1050.0 + }, + { + "seed": 55, + "B": 1425.0, + "F": 2525.0, + "delta": 1100.0 + }, + { + "seed": 56, + "B": 1650.0, + "F": 4125.0, + "delta": 2475.0 + }, + { + "seed": 57, + "B": 1500.0, + "F": 3250.0, + "delta": 1750.0 + }, + { + "seed": 58, + "B": 1225.0, + "F": 4125.0, + "delta": 2900.0 + }, + { + "seed": 59, + "B": 2100.0, + "F": 1800.0, + "delta": -300.0 + }, + { + "seed": 60, + "B": 1025.0, + "F": 3775.0, + "delta": 2750.0 + }, + { + "seed": 61, + "B": 2025.0, + "F": 3600.0, + "delta": 1575.0 + } + ], + "mean_gain": 1841.25, + "gain_percent": 132.82236248872857, + "wins": 18, + "ties": 0, + "losses": 2, + "protocol": "RTX3090/10env-2.5obsFPS/measured-only/seed42-61/cap3600/no-state/discrete6/RGB224/H1; Bprompt0/Fprompt2", + "causal_scope": "Same benchmark protocol, different realized inference latency distributions; no equal-delay or architecture-ranking claim", + "scientific_evidence_url": "https://huggingface.co/datasets/latency-sensitive-bench/Standard-Pipeline/blob/b451caa2a6515df3bae22580595151ef5b342f89/air_raid/profile_latency/Pi05/h1_10env_2p5obs_g128_20260920/vla/evaluation/comparison.json" +} diff --git a/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/evaluation/status.json b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/evaluation/status.json new file mode 100644 index 0000000000000000000000000000000000000000..a6ba19cc3372936f44890dcf660e8a65b21b032d --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/evaluation/status.json @@ -0,0 +1,10 @@ +{ + "training": "COMPLETED_5000", + "saved_reload": "PASSED", + "measured_B20": "AUDITED", + "measured_F20": "AUDITED", + "evaluated_model_revision": "de49831bd5dc9d642058bff0237e6ff2c9127b53", + "checkpoint_sha256": "c9ad50c5229424c919bb108d75b016f3b1815405ebd03adbea08137c18849bd8", + "weights_unchanged": true, + "F_evidence_revision": "b451caa2a6515df3bae22580595151ef5b342f89" +} diff --git a/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/latency_prompt_map.json b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/latency_prompt_map.json new file mode 100644 index 0000000000000000000000000000000000000000..639173b6e7e0cd687920466d3ffb0ac3a387cd39 --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/latency_prompt_map.json @@ -0,0 +1,7 @@ +{ + "0": { + "prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire. Current action latency is 0 raw frames (0.00 ms). The environment runs at 10 FPS and observations are emitted at 2.5 FPS. Choose the best next action.", + "latency_raw_frames": 0, + "latency_ms": 0.0 + } +} \ No newline at end of file diff --git a/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/manifest.json b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/manifest.json new file mode 100644 index 0000000000000000000000000000000000000000..a6065928fc1282696ea28df9788cf310b79bb670 --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/manifest.json @@ -0,0 +1,92 @@ +{ + "dataset_name": "air_raid_h1_10_2p5_l0", + "env_name": "air_raid", + "episodes": 90, + "frames": 81000, + "task_prompts": [ + "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire. Current action latency is 0 raw frames (0.00 ms). The environment runs at 10 FPS and observations are emitted at 2.5 FPS. Choose the best next action." + ], + "format": "starvla_lerobot_v2_image_parquet", + "source": "${PI05_RUN_DIR}/zero_latency/data/raw_10_2p5", + "integration_name": "gymnasium", + "task_name": "air_raid", + "action_layout": "gymnasium_discrete_v1", + "action_labels": [ + "noop", + "fire", + "right", + "left", + "rightfire", + "leftfire" + ], + "carrier_action_labels": [ + "noop", + "fire", + "right", + "left", + "rightfire", + "leftfire" + ], + "action_dim": 6, + "active_action_dim": 6, + "action_carrier": "native", + "rows_unit": "decision_step", + "obs_fps": 2.5, + "obs_stride_raw_frames": 4, + "uses_state": false, + "state_dim": 1, + "state_labels": [ + "state" + ], + "state_normalization": null, + "robot_type": "rl_games_gymnasium", + "latency_prompt_map_path": "${PI05_RUN_DIR}/zero_latency/data/lerobot/air_raid_h1_10_2p5_l0/latency_prompt_map.json", + "custom_mixtures_path": "${PI05_RUN_DIR}/zero_latency/data/lerobot/_generated_mixtures/air_raid_h1_10_2p5_l0.json", + "gymnasium_task": { + "action_labels": [ + "noop", + "fire", + "right", + "left", + "rightfire", + "leftfire" + ], + "action_values": [ + 0, + 1, + 2, + 3, + 4, + 5 + ], + "base_prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.", + "env_fps": 10, + "env_id": "LatencyBench/AirRaid-v0", + "frame_stack": 4, + "make_kwargs": { + "base_env_id": "ALE/AirRaid-v5", + "base_make_kwargs": { + "difficulty": 0, + "frameskip": 1, + "full_action_space": false, + "max_num_frames_per_episode": 108000, + "mode": 1, + "obs_type": "rgb", + "repeat_action_probability": 0.0 + }, + "fire_reset": true, + "noop_max": 30, + "render_mode": "rgb_array", + "screen_size": 84 + }, + "noop_action_id": 0, + "obs_fps": 2.5, + "registration_imports": [ + "latency_bench.envs.gymnasium_air_raid" + ], + "task_name": "air_raid" + }, + "validation_dataset_name": "air_raid_h1_10_2p5_l0__val", + "validation_episodes": 10, + "validation_frames": 9000 +} \ No newline at end of file diff --git a/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/provenance.json b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/provenance.json new file mode 100644 index 0000000000000000000000000000000000000000..1b2a88332b20f94c81c500e34ee56e42c0f6da79 --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/provenance.json @@ -0,0 +1,34 @@ +{ + "task": "air-raid", + "model": "qwenpi_v3", + "training_condition": "zero-latency", + "training_run_id": "airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2", + "source": { + "repo_id": "latency-sensitive-bench/Standard-Pipeline", + "revision": "2208875f92b21e3f8ec2363a30a6705dfe56cc04", + "path": "air_raid/zero_latency/Pi05", + "url": "https://huggingface.co/latency-sensitive-bench/Standard-Pipeline/tree/2208875f92b21e3f8ec2363a30a6705dfe56cc04/air_raid/zero_latency/Pi05" + }, + "hf_repo_id": "latency-sensitive-bench/benchmark-models", + "hf_path": "zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2", + "checkpoint": { + "source_file": "air_raid/zero_latency/Pi05/checkpoints/model.pt", + "source_sha256": "c9ad50c5229424c919bb108d75b016f3b1815405ebd03adbea08137c18849bd8", + "source_bytes": 10920516461, + "file": "checkpoints/model.pt", + "selection_rule": "only published VLA checkpoint", + "method": "copy", + "sha256": "c9ad50c5229424c919bb108d75b016f3b1815405ebd03adbea08137c18849bd8", + "bytes": 10920516461 + }, + "config_source": "air_raid/zero_latency/Pi05/config.yaml", + "task_contract_file": "task_contract.json", + "missing_source_files": [], + "migration": { + "operation": "copy_selected_inference_bundle", + "source_modified": false, + "evaluation_performed": false, + "experiment_status": "preserved_in_source_evidence" + }, + "action_horizon": 1 +} diff --git a/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/source/README.md b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/source/README.md new file mode 100644 index 0000000000000000000000000000000000000000..4a0ef2ecc3c86f32c2c5941a949164c28067e22f --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/source/README.md @@ -0,0 +1,7 @@ +# AirRaid Pi0.5 H1 — zero-latency-trained baseline B + +QwenPI_v3 with pinned Qwen3-VL-4B-Instruct, fresh action head, 5,000 updates, seed42, global128=micro64×accumulation1×2, GC disabled and ZeRO-2. RGB224, no state, six native discrete actions, H1. Environment/observation10/2.5FPS, measured-only RTX3090 evaluation,20 paired seeds42–61, cap3600 rawframes, one worker and hold/FIFO. B prompt0, F prompt2; no added IID delay. + +Audited B return 1386.25 ± 434.99; F 3227.50 ± 1095.58 (population SD). Mean paired gain 1841.25 (132.82%); 18 wins/0 ties/2 losses. Realized inference latency differs; no equal-delay causal claim or architecture ranking. + +[Fixed final scientific evidence](https://huggingface.co/datasets/latency-sensitive-bench/Standard-Pipeline/blob/b451caa2a6515df3bae22580595151ef5b342f89/air_raid/profile_latency/Pi05/h1_10env_2p5obs_g128_20260920/vla/evaluation/comparison.json). Original evaluated checkpoint SHA256 `c9ad50c5229424c919bb108d75b016f3b1815405ebd03adbea08137c18849bd8`, evaluated model revision `de49831bd5dc9d642058bff0237e6ff2c9127b53`. This metadata update does not change weights. The original P5×10 was retained. APPO teacher target10M counted steps corresponds toabout40M rawframes atstride4; no teacherreward gate was applied toVLA evaluation. diff --git a/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/source/provenance.json b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/source/provenance.json new file mode 100644 index 0000000000000000000000000000000000000000..0bd0fec956ce75dd61d1f347515d7c93d161413c --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/source/provenance.json @@ -0,0 +1,108 @@ +{ + "task": "air_raid", + "backbone": { + "repo_id": "Qwen/Qwen3-VL-4B-Instruct", + "revision": "ebb281ec70b05090aa6165b016eac8ec08e71b17" + }, + "checkpoint": "steps_5000_pytorch_model.pt", + "training_steps": 5000, + "seed": 42, + "global_batch": 128, + "training_run_id": "airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2", + "condition": "zero_latency", + "source": { + "archive_sha256": { + "lsb-pi05-h1-20260918.tar.gz": "08989cd9c3518cd79cb5ec065300755b8afe7555e1b6d65fed829549ef9127c4", + "pi05-training-package.tar.gz": "0e1c6f7cceb8979b525ceaa3c209f005049166b97eb97a1602e13ed8d7384b65" + }, + "original_code_root": "2de3816f07ecea4002964d046c4dfa581a658671", + "starvla": "3430c45edf6a08e4cfcaa2fce58bb4ea05995617", + "sample_factory": "4b11d8754ffc989cdd15a340cfc9be0277377d5b", + "task_overlay": "batch128 and AirRaid manifest clock; recorded separately", + "dataset": { + "repo": "latency-sensitive-bench/Standard-Pipeline", + "revision": "7e3d16d77a76cd3f888f1113f2902e39a9db2ae3", + "prefix": "air_raid/zero_latency/shared/native_100ep_v1/demonstrations/raw/", + "sha256": { + "metadata.json": "c465255fddae1a9d1548ea04ad992aed7892489111365572ef7cf54b53bcc9b2", + "train.parquet": "50153ebf0d23083d274c0d1b28a9e6ba01bd225ed758d1553b78ef3a8b4bfd55", + "val.parquet": "4a5de901e8d5051427ab163f70bad1752a951fb0fced8c81f0ba8fc04edc99b6" + }, + "source_unchanged": true, + "timing_derivation": { + "source_clock": [ + 50, + 12.5 + ], + "target_clock": [ + 10, + 2.5 + ], + "obs_stride_raw_frames": 4, + "changed": [ + "prompt", + "FPS metadata" + ], + "unchanged": [ + "images", + "actions", + "rewards", + "seeds", + "splits", + "zero latency", + "native physics" + ], + "source_checkpoint_config_is_historical": true + }, + "splits": { + "train": { + "rows": 81000, + "episodes": 90, + "return_min": 2525.0, + "return_mean": 2890.5555555555557 + }, + "val": { + "rows": 9000, + "episodes": 10, + "return_min": 2525.0, + "return_mean": 2792.5 + } + } + }, + "backbone": { + "repo_id": "Qwen/Qwen3-VL-4B-Instruct", + "revision": "ebb281ec70b05090aa6165b016eac8ec08e71b17", + "sha256": { + "model-00001-of-00002.safetensors": "30a01a0556622645a3cce87b655bbbbbc1f170c196099f1b666c93202c3339a9", + "model-00002-of-00002.safetensors": "046296a2a387efb43b0c997d5833c789604d168834f6e0d3064bf7bb13d002a6" + } + }, + "protocol": { + "env_fps": 10, + "obs_fps": 2.5, + "action_horizon": 1, + "state_dim": 0, + "action_dim": 6, + "image": [ + 224, + 224 + ], + "fresh_action_head": true, + "bridge_weights": false, + "seed": 42, + "updates": 5000, + "global_batch": 128, + "micro_batch": 64, + "accumulation": 1, + "gpus": 2, + "physical_gpu_indices": [ + 4, + 5 + ], + "gradient_checkpointing": false, + "distributed_backend": "DeepSpeed ZeRO-2" + } + }, + "training_config_sha256": "e42660cc2cc8487ffc9e95255b68d208c5600597ec560fef93c89dc863287c83", + "dataset_manifest_sha256": "136e837001333ee464eaa68395c17b9cc6eb61934df5c8146927bd63a3b5cc95" +} diff --git a/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/task_contract.json b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/task_contract.json new file mode 100644 index 0000000000000000000000000000000000000000..3c8053a9d925d290e87546842e24a9795706f480 --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/task_contract.json @@ -0,0 +1,44 @@ +{ + "action_labels": [ + "noop", + "fire", + "right", + "left", + "rightfire", + "leftfire" + ], + "action_values": [ + 0, + 1, + 2, + 3, + 4, + 5 + ], + "base_prompt": "Protect both buildings from flying saucers. Choose exactly one action from: noop, fire, right, left, rightfire, leftfire.", + "env_fps": 10, + "env_id": "LatencyBench/AirRaid-v0", + "frame_stack": 4, + "make_kwargs": { + "base_env_id": "ALE/AirRaid-v5", + "base_make_kwargs": { + "difficulty": 0, + "frameskip": 1, + "full_action_space": false, + "max_num_frames_per_episode": 108000, + "mode": 1, + "obs_type": "rgb", + "repeat_action_probability": 0.0 + }, + "fire_reset": true, + "noop_max": 30, + "render_mode": "rgb_array", + "screen_size": 84 + }, + "noop_action_id": 0, + "obs_fps": 2.5, + "registration_imports": [ + "latency_bench.envs.gymnasium_air_raid" + ], + "task_name": "air_raid" +} diff --git a/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/validation.json b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/validation.json new file mode 100644 index 0000000000000000000000000000000000000000..cd12f2079fe98d977c0b1891055fb004cc608931 --- /dev/null +++ b/zero-latency/air-raid/vla/starvla-qwenpi_v3-h1/airraid_pi05_l0_h1_10env_2p5obs_g128_20260920_r2/validation.json @@ -0,0 +1,17 @@ +{ + "state": "SAVED_BUNDLE_RELOAD_AND_REAL_SAMPLE_FORWARD_VERIFIED", + "verified_at": "2026-09-20T09:41:58.016051+00:00", + "forward_shape": [ + 1, + 1, + 6 + ], + "forward_finite": true, + "heldout_sample": 0, + "loader_rows": 9000, + "action_horizon": 1, + "state_dim": 0, + "action_dim": 6, + "checkpoint_sha256": "c9ad50c5229424c919bb108d75b016f3b1815405ebd03adbea08137c18849bd8", + "reload": "Existing strict key check with documented tied Qwen lm_head equivalence only" +}