Import Standard-Pipeline air-raid selected inference bundles

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="<commit>",
+                  allow_patterns=[prefix + "/**"])
+```
+
+Use the verified import commit listed in the repository's import report.

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

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"
+}

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"
+}

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
+}

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
+    }
+  }
+}

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
+  }
+}

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

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="<commit>",
+                  allow_patterns=[prefix + "/**"])
+```
+
+Use the verified import commit listed in the repository's import report.

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

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"
+}

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"
+  }
+}

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
+}

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
+          ]
+        }
+      }
+    }
+  }
+}

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"
+}

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"
+}

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
+  }
+}

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
+  }
+}

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"
+}

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

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="<commit>",
+                  allow_patterns=[prefix + "/**"])
+```
+
+Use the verified import commit listed in the repository's import report.

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

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"
+}

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"
+}

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"
+}

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.

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
+    }
+  }
+}

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

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"
+}

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="<commit>",
+                  allow_patterns=[prefix + "/**"])
+```
+
+Use the verified import commit listed in the repository's import report.

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

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

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
+  }
+}

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
+  }
+}

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
+  }
+}

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
+}

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"
+}

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"
+}

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"
+}

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="<commit>",
+                  allow_patterns=[prefix + "/**"])
+```
+
+Use the verified import commit listed in the repository's import report.

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

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'

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'

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
+  }
+}

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
+  }
+}

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
+  }
+}

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
+}