[Refactor] Add structured inference server config objects#3893
Draft
vmoens wants to merge 4 commits into
Draft
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/rl/3893
Note: Links to docs will display an error until the docs builds have been completed. ❗ 1 Active SEVsThere are 1 currently active SEVs. If your PR is affected, please view them below: ❌ 8 New FailuresAs of commit 39e3aa6 with merge base b660f05 ( NEW FAILURES - The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
github-actions
Bot
added
Documentation
This was referenced Jun 21, 2026
meta-cla
Bot
added
the
CLA Signed
Contributor
Benchmark Results: PR
|
| Benchmark | main ops | PR ops | Change |
|---|---|---|---|
benchmarks/test_replaybuffer_benchmark.py::test_rb_iterate[TensorDictPrioritizedReplayBuffer-LazyMemmapStorage-None-10000] |
400.75 | 2,220 | +453.93% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_populate[TensorDictReplayBuffer-ListStorage-RandomSampler-400] |
192.92 | 37.39 | -80.62% |
benchmarks/test_objectives_benchmarks.py::test_sac_speed[False-backward] |
54.39 | 88.62 | +62.93% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_iterate[TensorDictReplayBuffer-LazyTensorStorage-RandomSampler-10000] |
2,821 | 3,679 | +30.44% |
benchmarks/test_objectives_benchmarks.py::test_sac_speed[True-backward] |
205.91 | 253.69 | +23.20% |
benchmarks/test_envs_benchmark.py::test_cat_frames_functional[4-same] |
24.64 | 29.26 | +18.76% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_iterate[TensorDictReplayBuffer-LazyMemmapStorage-RandomSampler-10000] |
2,534 | 2,882 | +13.76% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_sample[TensorDictReplayBuffer-LazyTensorStorage-RandomSampler-10000] |
3,062 | 2,642 | -13.73% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_sample[TensorDictPrioritizedReplayBuffer-LazyTensorStorage-None-10000] |
1,913 | 2,153 | +12.50% |
benchmarks/test_objectives_benchmarks.py::test_redq_deprec_speed[True-backward] |
129.62 | 145.30 | +12.10% |
benchmarks/test_objectives_benchmarks.py::test_ddpg_speed[True-backward] |
375.98 | 419.39 | +11.55% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_populate[TensorDictReplayBuffer-LazyMemmapStorage-RandomSampler-400] |
504.03 | 561.90 | +11.48% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_sample[TensorDictReplayBuffer-LazyTensorStorage-SamplerWithoutReplacement-10000] |
3,048 | 2,709 | -11.11% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_populate[TensorDictPrioritizedReplayBuffer-ListStorage-None-400] |
50.88 | 56.45 | +10.94% |
benchmarks/test_compressed_storage_benchmark.py::TestCompressedStorageBenchmark::test_tensor_to_bytestream_speed[untyped_storage] |
8.1855 | 8.9896 | +9.82% |
benchmarks/test_objectives_benchmarks.py::test_ppo_speed[True-backward] |
105.84 | 115.26 | +8.89% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_sample[TensorDictReplayBuffer-LazyMemmapStorage-RandomSampler-10000] |
2,712 | 2,953 | +8.88% |
benchmarks/test_vla_preprocessing_benchmark.py::test_openvla_preprocessing_throughput[torchvision-256-256-64] |
10.91 | 10.09 | -7.60% |
benchmarks/test_objectives_benchmarks.py::test_dqn_speed[True-None] |
1,665 | 1,776 | +6.71% |
benchmarks/test_objectives_benchmarks.py::test_a2c_speed[True-None] |
280.26 | 298.25 | +6.42% |
benchmarks/test_objectives_benchmarks.py::test_cql_speed[True-None] |
82.01 | 87.06 | +6.15% |
benchmarks/test_envs_benchmark.py::test_simple |
1.7041 | 1.8036 | +5.84% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_populate[TensorDictReplayBuffer-LazyTensorStorage-RandomSampler-400] |
1,061 | 1,117 | +5.29% |
benchmarks/test_objectives_benchmarks.py::test_cql_speed[False-backward] |
27.40 | 28.71 | +4.78% |
benchmarks/test_rnn_reset_backends_benchmark.py::test_rnn_rollout_with_intermediate_resets[b256-t128-i32-h512-scan-True-0-gru] |
4.0767 | 4.2616 | +4.54% |
benchmarks/test_replaybuffer_benchmark.py::TestPrioritizedReplayBufferBenchmark::test_sample_mixed_devices[1000000-memmap_cpu_storage_cpu... |
80.20 | 83.65 | +4.30% |
benchmarks/test_storage_write_benchmark.py::TestStorageWriteBenchmark::test_storage_write_contiguous[100-img_shape2-large_img] |
569.83 | 545.81 | -4.21% |
benchmarks/test_objectives_benchmarks.py::test_redq_deprec_speed[reduce-overhead-None] |
281.12 | 292.23 | +3.95% |
benchmarks/test_vla_preprocessing_benchmark.py::test_openvla_preprocessing_throughput[torchvision-480-640-64] |
6.5950 | 6.3368 | -3.91% |
benchmarks/test_rnn_reset_backends_benchmark.py::test_rnn_rollout_with_intermediate_resets[b256-t128-i32-h512-scan-False-0-gru] |
2.9198 | 3.0338 | +3.90% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_sample[TensorDictReplayBuffer-LazyTensorStorage-sampler7-10000] |
759.22 | 730.15 | -3.83% |
benchmarks/test_rnn_reset_backends_benchmark.py::test_rnn_rollout_with_intermediate_resets[b256-t128-i32-h512-scan-False-0-lstm] |
1.9402 | 2.0113 | +3.67% |
benchmarks/test_vla_preprocessing_benchmark.py::test_openvla_preprocessing_throughput[pil-256-256-1] |
192.42 | 185.48 | -3.61% |
benchmarks/test_objectives_benchmarks.py::test_reinforce_speed[False-None] |
207.64 | 215.13 | +3.61% |
benchmarks/test_storage_write_benchmark.py::TestStorageWriteBenchmark::test_storage_write_lazystack[100-img_shape2-large_img] |
426.04 | 410.82 | -3.57% |
benchmarks/test_objectives_benchmarks.py::test_redq_deprec_speed[False-None] |
87.29 | 90.39 | +3.55% |
benchmarks/test_objectives_benchmarks.py::test_reinforce_speed[True-None] |
326.47 | 337.81 | +3.47% |
benchmarks/test_compressed_storage_benchmark.py::TestCompressedStorageBenchmark::test_tensor_to_bytestream_speed[numpy] |
362,799 | 375,272 | +3.44% |
benchmarks/test_envs_benchmark.py::test_transformed |
0.8822 | 0.9123 | +3.41% |
benchmarks/test_objectives_benchmarks.py::test_gae_speed[vec_generalized_advantage_estimate-True-1-512] |
641.45 | 663.21 | +3.39% |
benchmarks/test_storage_write_benchmark.py::TestCollectorIntegrationBenchmark::test_collector_with_rb[100-img_shape0-atari] |
25.64 | 26.50 | +3.35% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[True-True-True-False-True] |
41,216 | 42,582 | +3.32% |
benchmarks/test_storage_write_benchmark.py::TestStorageWriteBenchmark::test_collector_stack_then_write[100-img_shape2-large_img] |
175.12 | 169.38 | -3.28% |
benchmarks/test_objectives_benchmarks.py::test_redq_deprec_speed[True-None] |
275.36 | 284.36 | +3.27% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[False-True-False-False-False] |
49,804 | 51,431 | +3.27% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[False-False-True-True-False] |
28,945 | 29,881 | +3.24% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_populate[TensorDictPrioritizedReplayBuffer-LazyTensorStorage-None-400] |
877.67 | 905.07 | +3.12% |
benchmarks/test_storage_write_benchmark.py::TestCollectorIntegrationBenchmark::test_collector_without_rb[200-img_shape1-large_batch] |
14.95 | 15.41 | +3.09% |
benchmarks/test_objectives_benchmarks.py::test_iql_speed[True-backward] |
58.78 | 60.56 | +3.04% |
benchmarks/test_storage_write_benchmark.py::TestStorageWriteBenchmark::test_collector_lazystack_then_write[100-img_shape2-large_img] |
403.77 | 391.61 | -3.01% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_populate[TensorDictPrioritizedReplayBuffer-LazyMemmapStorage-None-400] |
503.80 | 518.51 | +2.92% |
benchmarks/test_storage_write_benchmark.py::TestCollectorIntegrationBenchmark::test_collector_without_rb[100-img_shape0-atari] |
29.47 | 30.32 | +2.87% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[True-False-False-False-True] |
33,558 | 34,509 | +2.83% |
benchmarks/test_envs_benchmark.py::test_parallel |
0.9727 | 0.9452 | -2.82% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[True-True-False-False-False] |
62,747 | 64,484 | +2.77% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[True-False-True-True-True] |
20,544 | 21,098 | +2.69% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[False-False-False-True-True] |
17,963 | 18,445 | +2.69% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_sample[TensorDictReplayBuffer-LazyMemmapStorage-SamplerWithoutReplacement-10000] |
2,751 | 2,678 | -2.67% |
benchmarks/test_storage_write_benchmark.py::TestStorageWriteBenchmark::test_storage_write_lazystack[50-img_shape0-small] |
4,353 | 4,469 | +2.67% |
benchmarks/test_objectives_benchmarks.py::test_redq_deprec_speed[False-backward] |
62.20 | 63.84 | +2.64% |
benchmarks/test_objectives_benchmarks.py::test_reinforce_speed[False-backward] |
130.03 | 133.45 | +2.63% |
benchmarks/test_storage_write_benchmark.py::TestStorageWriteBenchmark::test_storage_write_contiguous[200-img_shape3-large_batch] |
773.05 | 752.82 | -2.62% |
benchmarks/test_non_tensor_env_benchmark.py::test_non_tensor_env_rollout_speed[1000-single-True] |
1.3613 | 1.3261 | -2.59% |
benchmarks/test_storage_write_benchmark.py::TestStorageWriteBenchmark::test_collector_lazystack_then_write[50-img_shape0-small] |
3,473 | 3,563 | +2.58% |
benchmarks/test_objectives_benchmarks.py::test_cql_speed[reduce-overhead-None] |
84.78 | 86.95 | +2.56% |
benchmarks/test_objectives_benchmarks.py::test_dqn_speed[True-backward] |
964.13 | 988.78 | +2.56% |
benchmarks/test_objectives_benchmarks.py::test_a2c_speed[False-None] |
175.83 | 180.27 | +2.52% |
benchmarks/test_objectives_benchmarks.py::test_cql_speed[True-backward] |
59.14 | 60.62 | +2.51% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[True-False-True-True-False] |
33,978 | 34,823 | +2.49% |
benchmarks/test_objectives_benchmarks.py::test_ddpg_speed[True-None] |
685.74 | 702.62 | +2.46% |
benchmarks/test_objectives_benchmarks.py::test_a2c_speed[False-backward] |
82.37 | 84.39 | +2.45% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[True-True-True-True-False] |
41,254 | 42,263 | +2.45% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[False-True-True-False-False] |
58,158 | 56,741 | -2.44% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[False-True-True-False-True] |
32,991 | 32,191 | -2.42% |
benchmarks/test_objectives_benchmarks.py::test_iql_speed[False-None] |
49.37 | 50.56 | +2.42% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[True-False-False-True-True] |
19,264 | 19,711 | +2.32% |
benchmarks/test_rnn_reset_backends_benchmark.py::test_rnn_rollout_with_intermediate_resets[b256-t128-i32-h512-cudnn-True-0-gru] |
1.4636 | 1.4298 | -2.31% |
benchmarks/test_non_tensor_env_benchmark.py::test_non_tensor_env_rollout_speed[1000-parallel-buffers-False] |
0.6032 | 0.5895 | -2.28% |
benchmarks/test_objectives_benchmarks.py::test_ppo_speed[reduce-overhead-None] |
262.59 | 268.56 | +2.28% |
benchmarks/test_objectives_benchmarks.py::test_gae_speed[vec_generalized_advantage_estimate-False-32-512] |
537.50 | 549.65 | +2.26% |
benchmarks/test_storage_write_benchmark.py::TestStorageWriteBenchmark::test_storage_write_contiguous[100-img_shape1-atari] |
5,043 | 4,929 | -2.26% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[True-False-True-False-True] |
36,748 | 37,574 | +2.25% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_extend_sample[ReplayBuffer-LazyTensorStorage-RandomSampler-100000-10000-100-True] |
24.30 | 24.84 | +2.24% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_populate[TensorDictReplayBuffer-ListStorage-SamplerWithoutReplacement-400] |
193.18 | 197.49 | +2.23% |
benchmarks/test_storage_write_benchmark.py::TestStorageWriteBenchmark::test_collector_stack_then_write[100-img_shape1-atari] |
270.39 | 276.39 | +2.22% |
benchmarks/test_rnn_reset_backends_benchmark.py::test_rnn_rollout_with_intermediate_resets[b256-t128-i32-h512-cudnn-False-0-lstm] |
0.8696 | 0.8505 | -2.19% |
benchmarks/test_non_tensor_env_benchmark.py::test_non_tensor_env_rollout_speed[1000-parallel-buffers-True] |
0.5178 | 0.5291 | +2.18% |
benchmarks/test_storage_write_benchmark.py::TestStorageWriteBenchmark::test_storage_write_lazystack[200-img_shape3-large_batch] |
330.17 | 337.35 | +2.17% |
benchmarks/test_objectives_benchmarks.py::test_ddpg_speed[reduce-overhead-None] |
698.79 | 713.90 | +2.16% |
benchmarks/test_storage_write_benchmark.py::TestStorageWriteBenchmark::test_collector_lazystack_then_write[200-img_shape3-large_batch] |
307.62 | 314.18 | +2.13% |
benchmarks/test_collectors_benchmark.py::test_sync_preempt |
16.63 | 16.27 | -2.13% |
benchmarks/test_objectives_benchmarks.py::test_reinforce_speed[reduce-overhead-None] |
334.11 | 341.21 | +2.12% |
benchmarks/test_vla_preprocessing_benchmark.py::test_openvla_preprocessing_throughput[pil-256-256-64] |
2.9770 | 3.0395 | +2.10% |
benchmarks/test_objectives_benchmarks.py::test_reinforce_speed[True-backward] |
123.74 | 126.31 | +2.08% |
benchmarks/test_objectives_benchmarks.py::test_td3_speed[True-None] |
545.92 | 557.17 | +2.06% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_sample[TensorDictReplayBuffer-LazyMemmapStorage-sampler6-10000] |
711.83 | 726.07 | +2.00% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_iterate[TensorDictReplayBuffer-LazyTensorStorage-SamplerWithoutReplacement-10000] |
3,082 | 3,021 | -2.00% |
benchmarks/test_objectives_benchmarks.py::test_dqn_speed[False-backward] |
505.83 | 515.89 | +1.99% |
benchmarks/test_storage_write_benchmark.py::TestStorageWriteBenchmark::test_storage_write_lazystack[100-img_shape1-atari] |
694.86 | 708.55 | +1.97% |
benchmarks/test_envs_benchmark.py::test_serial |
0.5725 | 0.5834 | +1.89% |
benchmarks/test_objectives_benchmarks.py::test_cql_speed[False-None] |
37.59 | 38.30 | +1.88% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[True-True-True-True-True] |
23,580 | 24,018 | +1.86% |
benchmarks/test_vla_preprocessing_benchmark.py::test_openvla_preprocessing_throughput[torchvision-224-224-1] |
636.89 | 648.50 | +1.82% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[True-True-False-True-True] |
21,845 | 22,240 | +1.81% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_iterate[TensorDictReplayBuffer-LazyMemmapStorage-SamplerWithoutReplacement-10000] |
2,807 | 2,757 | -1.79% |
benchmarks/test_objectives_benchmarks.py::test_td3_speed[False-backward] |
89.97 | 91.57 | +1.79% |
benchmarks/test_vla_preprocessing_benchmark.py::test_openvla_preprocessing_throughput[pil-256-256-16] |
11.94 | 12.15 | +1.78% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_extend_sample[ReplayBuffer-LazyTensorStorage-RandomSampler-1000000-10000-100-True] |
23.32 | 23.73 | +1.77% |
benchmarks/test_replaybuffer_benchmark.py::TestPrioritizedReplayBufferBenchmark::test_sampler_sample_scale[1000000-cpu] |
97.10 | 98.82 | +1.77% |
benchmarks/test_objectives_benchmarks.py::test_ppo_speed[False-backward] |
77.85 | 79.22 | +1.76% |
benchmarks/test_objectives_benchmarks.py::test_gae_speed[vec_generalized_advantage_estimate-False-1-512] |
2,243 | 2,282 | +1.74% |
benchmarks/test_objectives_benchmarks.py::test_iql_speed[False-backward] |
32.81 | 33.37 | +1.71% |
benchmarks/test_objectives_benchmarks.py::test_a2c_speed[True-backward] |
119.29 | 121.33 | +1.71% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_extend_sample[ReplayBuffer-LazyTensorStorage-RandomSampler-100000-10000-100-False] |
51.91 | 52.79 | +1.70% |
benchmarks/test_objectives_benchmarks.py::test_td3_speed[True-backward] |
280.22 | 284.98 | +1.70% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_iterate[TensorDictPrioritizedReplayBuffer-LazyTensorStorage-None-10000] |
2,126 | 2,162 | +1.69% |
benchmarks/test_storage_write_benchmark.py::TestStorageWriteBenchmark::test_collector_lazystack_then_write[100-img_shape1-atari] |
637.74 | 648.45 | +1.68% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_sample[TensorDictPrioritizedReplayBuffer-LazyMemmapStorage-None-10000] |
2,012 | 2,045 | +1.65% |
benchmarks/test_storage_write_benchmark.py::TestCollectorIntegrationBenchmark::test_collector_with_rb[200-img_shape1-large_batch] |
13.26 | 13.48 | +1.62% |
benchmarks/test_objectives_benchmarks.py::test_td3_speed[reduce-overhead-None] |
569.48 | 578.53 | +1.59% |
| ... | ... | ... | Showing 120 of 216 comparisons, sorted by absolute change. |
GPU
Compared 226 benchmarks. Regressions over 5%: 13. Improvements over 5%: 20.
| Benchmark | main ops | PR ops | Change |
|---|---|---|---|
benchmarks/test_replaybuffer_benchmark.py::test_rb_extend_sample[ReplayBuffer-LazyTensorStorage-RandomSampler-100000-10000-100-False] |
28.71 | 51.78 | +80.37% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_populate[TensorDictReplayBuffer-ListStorage-SamplerWithoutReplacement-400] |
190.98 | 48.49 | -74.61% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_iterate[TensorDictReplayBuffer-LazyTensorStorage-RandomSampler-10000] |
3,589 | 2,613 | -27.20% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_iterate[TensorDictReplayBuffer-LazyTensorStorage-SamplerWithoutReplacement-10000] |
3,599 | 2,721 | -24.40% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_sample[TensorDictReplayBuffer-LazyMemmapStorage-RandomSampler-10000] |
2,574 | 3,153 | +22.50% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_iterate[TensorDictReplayBuffer-LazyMemmapStorage-SamplerWithoutReplacement-10000] |
3,350 | 2,610 | -22.11% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_iterate[TensorDictReplayBuffer-LazyMemmapStorage-RandomSampler-10000] |
3,101 | 2,500 | -19.39% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_sample[TensorDictReplayBuffer-LazyMemmapStorage-SamplerWithoutReplacement-10000] |
2,982 | 3,447 | +15.58% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_sample[TensorDictReplayBuffer-LazyTensorStorage-SamplerWithoutReplacement-10000] |
3,227 | 3,643 | +12.90% |
benchmarks/test_storage_write_benchmark.py::TestStorageWriteBenchmark::test_storage_write_contiguous[100-img_shape1-atari] |
3,657 | 4,103 | +12.21% |
benchmarks/test_storage_write_benchmark.py::TestStorageWriteBenchmark::test_storage_write_lazystack[100-img_shape1-atari] |
725.12 | 652.78 | -9.98% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_populate[TensorDictPrioritizedReplayBuffer-LazyTensorStorage-None-400] |
732.06 | 798.79 | +9.12% |
benchmarks/test_objectives_benchmarks.py::test_dqn_speed[True-backward] |
970.63 | 884.73 | -8.85% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_sample[TensorDictPrioritizedReplayBuffer-LazyMemmapStorage-None-10000] |
2,052 | 1,881 | -8.32% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[False-False-True-True-True] |
17,581 | 19,024 | +8.21% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[False-True-False-False-False] |
46,239 | 49,853 | +7.82% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_iterate[TensorDictPrioritizedReplayBuffer-LazyMemmapStorage-None-10000] |
1,974 | 2,121 | +7.45% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_sample[TensorDictPrioritizedReplayBuffer-LazyTensorStorage-None-10000] |
1,810 | 1,942 | +7.31% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_iterate[TensorDictPrioritizedReplayBuffer-LazyTensorStorage-None-10000] |
2,036 | 2,181 | +7.12% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_populate[TensorDictPrioritizedReplayBuffer-LazyMemmapStorage-None-400] |
461.18 | 490.58 | +6.38% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[True-False-False-False-True] |
32,807 | 34,832 | +6.17% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[False-True-False-True-False] |
30,098 | 31,944 | +6.13% |
benchmarks/test_objectives_benchmarks.py::test_ppo_speed[True-backward] |
351.53 | 331.58 | -5.68% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_sample[TensorDictReplayBuffer-LazyTensorStorage-sampler7-10000] |
751.64 | 794.01 | +5.64% |
benchmarks/test_non_tensor_env_benchmark.py::test_non_tensor_env_rollout_speed[1000-serial-buffers-True] |
0.4931 | 0.5207 | +5.60% |
benchmarks/test_storage_write_benchmark.py::TestStorageWriteBenchmark::test_storage_write_lazystack[100-img_shape2-large_img] |
401.50 | 379.35 | -5.51% |
benchmarks/test_non_tensor_env_benchmark.py::test_non_tensor_env_rollout_speed[1000-serial-no-buffers-False] |
0.6682 | 0.7042 | +5.39% |
benchmarks/test_rnn_reset_backends_benchmark.py::test_rnn_rollout_with_intermediate_resets[b256-t128-i32-h512-scan-False-0-gru] |
22.45 | 21.27 | -5.27% |
benchmarks/test_objectives_benchmarks.py::test_ppo_speed[reduce-overhead-None] |
780.80 | 821.24 | +5.18% |
benchmarks/test_objectives_benchmarks.py::test_sac_speed[True-backward] |
326.56 | 309.73 | -5.15% |
benchmarks/test_vla_preprocessing_benchmark.py::test_openvla_preprocessing_throughput[torchvision-256-256-4] |
165.11 | 156.66 | -5.12% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[True-True-True-False-True] |
40,653 | 42,712 | +5.07% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[False-False-False-False-False] |
42,710 | 44,861 | +5.04% |
benchmarks/test_storage_write_benchmark.py::TestStorageWriteBenchmark::test_storage_write_contiguous[100-img_shape2-large_img] |
528.38 | 502.56 | -4.89% |
benchmarks/test_non_tensor_env_benchmark.py::test_non_tensor_env_rollout_speed[1000-single-False] |
1.5300 | 1.6022 | +4.72% |
benchmarks/test_non_tensor_env_benchmark.py::test_non_tensor_env_rollout_speed[1000-serial-no-buffers-True] |
0.5700 | 0.5968 | +4.70% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[False-False-True-True-False] |
27,759 | 29,063 | +4.70% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[False-True-False-False-True] |
28,739 | 30,059 | +4.59% |
benchmarks/test_storage_write_benchmark.py::TestStorageWriteBenchmark::test_storage_write_lazystack[50-img_shape0-small] |
4,358 | 4,163 | -4.47% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[True-False-True-True-True] |
20,116 | 21,002 | +4.41% |
benchmarks/test_objectives_benchmarks.py::test_reinforce_speed[False-None] |
375.63 | 392.19 | +4.41% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[False-True-True-False-True] |
31,639 | 33,013 | +4.34% |
benchmarks/test_compressed_storage_benchmark.py::TestCompressedStorageBenchmark::test_tensor_to_bytestream_speed[untyped_storage] |
8.1097 | 7.7623 | -4.28% |
benchmarks/test_storage_write_benchmark.py::TestStorageWriteBenchmark::test_collector_stack_then_write[100-img_shape2-large_img] |
160.47 | 167.33 | +4.28% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[True-False-False-True-False] |
30,575 | 31,832 | +4.11% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[True-False-False-False-False] |
53,223 | 55,336 | +3.97% |
benchmarks/test_non_tensor_env_benchmark.py::test_non_tensor_env_rollout_speed[1000-single-True] |
1.2921 | 1.3428 | +3.92% |
benchmarks/test_storage_write_benchmark.py::TestStorageWriteBenchmark::test_collector_stack_then_write[100-img_shape1-atari] |
263.22 | 253.01 | -3.88% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_iterate[TensorDictReplayBuffer-ListStorage-RandomSampler-4000] |
160.48 | 166.67 | +3.86% |
benchmarks/test_objectives_benchmarks.py::test_gae_speed[vec_generalized_advantage_estimate-False-1-512] |
1,273 | 1,322 | +3.84% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_sample[TensorDictReplayBuffer-ListStorage-SamplerWithoutReplacement-4000] |
161.05 | 167.23 | +3.84% |
benchmarks/test_objectives_benchmarks.py::test_ddpg_speed[False-backward] |
236.69 | 227.78 | -3.77% |
benchmarks/test_vla_preprocessing_benchmark.py::test_openvla_preprocessing_throughput[torchvision-224-224-4] |
180.62 | 187.22 | +3.65% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_sample[TensorDictPrioritizedReplayBuffer-ListStorage-None-4000] |
156.93 | 162.44 | +3.51% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_sample[TensorDictReplayBuffer-ListStorage-RandomSampler-4000] |
154.80 | 160.20 | +3.49% |
benchmarks/test_replaybuffer_benchmark.py::TestPrioritizedReplayBufferBenchmark::test_sample_mixed_devices[1000000-memmap_cpu_storage_cud... |
974.80 | 940.94 | -3.47% |
benchmarks/test_collectors_benchmark.py::test_single_pixels |
6.0618 | 6.2711 | +3.45% |
benchmarks/test_objectives_benchmarks.py::test_a2c_speed[True-backward] |
355.73 | 343.78 | -3.36% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_iterate[TensorDictReplayBuffer-ListStorage-SamplerWithoutReplacement-4000] |
162.92 | 168.24 | +3.27% |
benchmarks/test_vla_preprocessing_benchmark.py::test_openvla_preprocessing_throughput[pil-256-256-4] |
46.92 | 48.45 | +3.26% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[False-True-True-True-True] |
20,149 | 20,776 | +3.11% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_populate[TensorDictReplayBuffer-LazyMemmapStorage-SamplerWithoutReplacement-400] |
504.76 | 520.10 | +3.04% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[False-False-False-False-True] |
27,409 | 28,223 | +2.97% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[True-False-True-False-False] |
61,820 | 63,653 | +2.97% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_iterate[TensorDictPrioritizedReplayBuffer-ListStorage-None-4000] |
158.12 | 162.76 | +2.94% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[False-True-True-False-False] |
55,062 | 56,677 | +2.93% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_populate[TensorDictReplayBuffer-LazyMemmapStorage-RandomSampler-400] |
496.98 | 482.48 | -2.92% |
benchmarks/test_vla_preprocessing_benchmark.py::test_openvla_preprocessing_throughput[torchvision-224-224-1] |
616.10 | 634.04 | +2.91% |
benchmarks/test_objectives_benchmarks.py::test_values[vec_generalized_advantage_estimate-True-True] |
302.74 | 293.99 | -2.89% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_sample[TensorDictReplayBuffer-LazyMemmapStorage-sampler6-10000] |
761.98 | 740.28 | -2.85% |
benchmarks/test_objectives_benchmarks.py::test_cql_speed[True-backward] |
219.46 | 213.29 | -2.81% |
benchmarks/test_storage_write_benchmark.py::TestCollectorIntegrationBenchmark::test_collector_without_rb[100-img_shape0-atari] |
28.81 | 29.62 | +2.79% |
benchmarks/test_compressed_storage_benchmark.py::TestCompressedStorageBenchmark::test_tensor_to_bytestream_speed[numpy] |
374,034 | 363,755 | -2.75% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_populate[TensorDictReplayBuffer-LazyTensorStorage-RandomSampler-400] |
939.29 | 963.63 | +2.59% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[True-True-False-False-False] |
62,909 | 64,499 | +2.53% |
benchmarks/test_storage_write_benchmark.py::TestStorageWriteBenchmark::test_collector_stack_then_write[50-img_shape0-small] |
839.68 | 860.64 | +2.50% |
benchmarks/test_envs_benchmark.py::test_serial |
0.4151 | 0.4254 | +2.48% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[True-True-False-True-True] |
21,539 | 22,072 | +2.47% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[True-False-False-True-True] |
19,208 | 19,674 | +2.43% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[True-False-True-False-True] |
36,892 | 37,769 | +2.38% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[False-False-True-False-False] |
48,677 | 49,825 | +2.36% |
benchmarks/test_storage_write_benchmark.py::TestStorageWriteBenchmark::test_storage_write_contiguous[50-img_shape0-small] |
6,025 | 5,889 | -2.26% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[True-True-False-True-False] |
37,536 | 38,381 | +2.25% |
benchmarks/test_objectives_benchmarks.py::test_gae_speed[vec_generalized_advantage_estimate-False-32-512] |
1,283 | 1,254 | -2.23% |
benchmarks/test_non_tensor_env_benchmark.py::test_non_tensor_env_rollout_speed[1000-serial-buffers-False] |
0.5872 | 0.6002 | +2.21% |
benchmarks/test_vla_preprocessing_benchmark.py::test_openvla_preprocessing_throughput[pil-224-224-64] |
4.4848 | 4.5817 | +2.16% |
benchmarks/test_objectives_benchmarks.py::test_a2c_speed[False-backward] |
147.81 | 144.63 | -2.16% |
benchmarks/test_collectors_benchmark.py::test_sync_pixels |
10.52 | 10.30 | -2.15% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[False-False-True-False-True] |
29,864 | 30,502 | +2.14% |
benchmarks/test_objectives_benchmarks.py::test_ddpg_speed[True-None] |
814.11 | 797.03 | -2.10% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[False-False-False-True-False] |
26,441 | 26,994 | +2.09% |
benchmarks/test_vla_preprocessing_benchmark.py::test_openvla_preprocessing_throughput[torchvision-480-640-4] |
145.45 | 148.49 | +2.09% |
benchmarks/test_non_tensor_env_benchmark.py::test_non_tensor_env_rollout_speed[1000-parallel-no-buffers-True] |
0.2097 | 0.2139 | +1.99% |
benchmarks/test_vla_preprocessing_benchmark.py::test_openvla_preprocessing_throughput[pil-224-224-16] |
17.93 | 18.27 | +1.90% |
benchmarks/test_storage_write_benchmark.py::TestStorageWriteBenchmark::test_storage_write_contiguous[200-img_shape3-large_batch] |
659.83 | 672.36 | +1.90% |
benchmarks/test_storage_write_benchmark.py::TestStorageWriteBenchmark::test_collector_stack_then_write[200-img_shape3-large_batch] |
132.29 | 134.79 | +1.89% |
benchmarks/test_storage_write_benchmark.py::TestCollectorIntegrationBenchmark::test_collector_without_rb[200-img_shape1-large_batch] |
14.69 | 14.96 | +1.88% |
benchmarks/test_collectors_benchmark.py::test_single_with_rb_pixels |
5.3442 | 5.2439 | -1.88% |
benchmarks/test_vla_preprocessing_benchmark.py::test_openvla_preprocessing_throughput[pil-224-224-4] |
70.80 | 72.13 | +1.87% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[True-True-True-False-False] |
74,851 | 76,237 | +1.85% |
benchmarks/test_vla_preprocessing_benchmark.py::test_openvla_preprocessing_throughput[pil-256-256-1] |
190.27 | 193.78 | +1.84% |
benchmarks/test_objectives_benchmarks.py::test_iql_speed[True-backward] |
241.96 | 237.56 | -1.82% |
benchmarks/test_replaybuffer_benchmark.py::TestPrioritizedReplayBufferBenchmark::test_sample_mixed_devices[1000000-cuda_storage_cuda_samp... |
1,481 | 1,455 | -1.74% |
benchmarks/test_rnn_reset_backends_benchmark.py::test_rnn_rollout_with_intermediate_resets[b256-t128-i32-h512-scan-True-0-gru] |
49.15 | 48.29 | -1.74% |
benchmarks/test_envs_benchmark.py::test_step_mdp_speed[False-True-False-True-True] |
19,088 | 19,418 | +1.73% |
benchmarks/test_envs_benchmark.py::test_cat_frames_functional[16-constant] |
4,592 | 4,672 | +1.73% |
benchmarks/test_vla_preprocessing_benchmark.py::test_openvla_preprocessing_throughput[torchvision-256-256-1] |
521.83 | 512.92 | -1.71% |
benchmarks/test_vla_preprocessing_benchmark.py::test_openvla_preprocessing_throughput[pil-224-224-1] |
281.91 | 286.58 | +1.66% |
benchmarks/test_compressed_storage_benchmark.py::TestCompressedStorageBenchmark::test_tensor_to_bytestream_speed[safetensors] |
22,425 | 22,791 | +1.63% |
benchmarks/test_storage_write_benchmark.py::TestStorageWriteBenchmark::test_collector_lazystack_then_write[200-img_shape3-large_batch] |
289.48 | 294.13 | +1.61% |
benchmarks/test_vla_preprocessing_benchmark.py::test_openvla_preprocessing_throughput[torchvision-480-640-1] |
472.80 | 465.42 | -1.56% |
benchmarks/test_objectives_benchmarks.py::test_td3_speed[True-None] |
736.09 | 724.77 | -1.54% |
benchmarks/test_objectives_benchmarks.py::test_ddpg_speed[True-backward] |
452.48 | 445.58 | -1.52% |
benchmarks/test_replaybuffer_benchmark.py::test_rb_populate[TensorDictReplayBuffer-LazyTensorStorage-SamplerWithoutReplacement-400] |
950.03 | 964.03 | +1.47% |
benchmarks/test_objectives_benchmarks.py::test_td3_speed[True-backward] |
371.40 | 365.94 | -1.47% |
benchmarks/test_vla_preprocessing_benchmark.py::test_openvla_preprocessing_throughput[pil-256-256-16] |
12.02 | 12.19 | +1.43% |
benchmarks/test_vla_preprocessing_benchmark.py::test_openvla_preprocessing_throughput[pil-480-640-1] |
77.56 | 78.63 | +1.38% |
benchmarks/test_objectives_benchmarks.py::test_iql_speed[False-backward] |
67.35 | 68.27 | +1.37% |
benchmarks/test_compressed_storage_benchmark.py::TestCompressedStorageBenchmark::test_tensor_to_bytestream_speed[pickle] |
11,773 | 11,932 | +1.36% |
benchmarks/test_collectors_benchmark.py::test_async_pixels |
10.85 | 10.71 | -1.33% |
| ... | ... | ... | Showing 120 of 226 comparisons, sorted by absolute change. |
vmoens
commented
Jun 21, 2026
Collaborator
Author
There was a problem hiding this comment.
this is worth a paragraph in the doc somewhere
This was referenced Jun 24, 2026
theap06
suggested changes
Jun 26, 2026
| device_config: InferenceDeviceConfig | None = None, | ||
| shutdown_event: threading.Event | MPEvent | None = None, | ||
| ): | ||
| if server_config is not None: |
Contributor
There was a problem hiding this comment.
This doesn't enforce mutual exclusivity the way the error message claims. It compares the kwargs' values to the defaults, so it only fires when a kwarg differs.
| shutdown_event: threading.Event | MPEvent | None = None, | ||
| ): | ||
| if server_config is not None: | ||
| if ( |
Contributor
There was a problem hiding this comment.
As mentioned in the comment above, this doesn't really solve mutual exclusivity. I'd suggest defaulting these kwargs to a private sentinel and checking if they have been set instead of comparing values. We can implemented these changes for ProcessInferenceServer.init and AsyncBatchedCollector.init.
| "policy_device, and output_device." | ||
| ) | ||
| policy_device = device_config.policy_device | ||
| output_device = device_config.server_output_device() |
Contributor
There was a problem hiding this comment.
server_output_device() substitutes env_device for output_device when the latter is unset. I'd either reject them when a server consumes the config, or document on the server that those two fields are accepted but ignored.
| client = transport.client() | ||
| result = client(TensorDict({"observation": torch.randn(4)})) | ||
| stats = server.stats() | ||
| assert result["action"].device.type == "cpu" |
Contributor
There was a problem hiding this comment.
This test would still pass if server_output_device(), the output_device move, or the _env_loop device moves were completely broken. To have them work correctly, either mock the tensordict and assert .to(target) was called with the expected device, or add a @pytest.mark.gpu variant that moves CPU to CUDA and checks the result landed on the right device.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Stack from ghstack (oldest at bottom):