policy string | mode string | ticks int64 | bit_identical_ticks int64 | max_abs_err float64 | mean_abs_err float64 | min_cosine float64 | eager_p50_ms float64 | engine_p50_ms float64 | engine_p99_ms float64 | speedup float64 | engine_pre_p50_ms float64 | engine_post_p50_ms float64 | target string | captured_share_of_tick float64 | precast dict | batch_size null | reference string | eager_max_abs_err null | environment dict | noise_seed int64 | seeds list | warnings list | created_at float64 | exact bool | licence dict | status string | digest string |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
lerobot/smolvla_base | exact | 16 | 16 | 0 | 0 | 1 | 225.181662 | 36.010556 | 36.489781 | 6.253213 | 0.293937 | 0.068609 | model.sample_actions | 0.998 | {
"precast": false,
"probe": {
"under_autocast_fp32": 0,
"outside_autocast_fp32": 201,
"castable_fp32": 21,
"dtype": null
}
} | null | eager | null | {
"torch": "2.11.0+cu128",
"cuda": "12.8",
"python": "3.12.15",
"platform": "Linux-6.8.0-40-generic-x86_64-with-glibc2.39",
"transformers": "5.5.4",
"lerobot": "0.6.1",
"torch_threads": 1,
"omp_num_threads": null,
"gpu": "NVIDIA GeForce RTX 4090",
"cudnn": 91900,
"driver": "570.144",
"clocks_sm_... | 123 | [
5000,
5001,
5002,
5003,
5004,
5005,
5006,
5007,
5008,
5009,
5010,
5011,
5012,
5013,
5014,
5015
] | [] | 1,792,304,643.844419 | true | {
"org": "evaluation",
"tier": "evaluation",
"active": false
} | evaluation-unsigned | 712ecbd14a33a79e |
Proprio Bitwise Certificates
Measured results of Proprio, ZenEdgeX's exact inference engine for vision-language-action robot policies. Each file is one Bitwise Certificate: a JSON record the engine writes after running a policy eagerly and through Proprio on the same observations and noise and comparing every action chunk bit for bit.
Unmodified checkpoints, torch 2.11, LeRobot 0.6, 2026-10-17 to 2026-10-19. RTX 4090 unless the file is under h100/.
| File | Policy | Mode | Rows | Identical | Eager p50 | Proprio p50 per robot | Speedup |
|---|---|---|---|---|---|---|---|
| cert_eval_smoke.json | lerobot/smolvla_base | exact, single robot (evaluation wheel) | 16 | 16/16 | 225 ms | 36.0 ms | 6.3x |
| cert_smolvla_single.json | lerobot/smolvla_base | exact, single robot | 32 | 32/32 | 232 ms | 39.5 ms | 5.9x |
| cert_pi05_single.json | lerobot/pi05_base (bf16) | exact, single robot | 32 | 32/32 | 257 ms | 96.7 ms | 2.7x |
| cert_smolvla_inv_single.json | lerobot/smolvla_base | invariant kernels, single robot | 32 | 32/32 vs same-mode | 233 ms | 66.5 ms | 3.5x |
| cert_smolvla_r16_exact.json | lerobot/smolvla_base | rollout exact, 16 robots/tick | 64 | 64/64 | 240 ms | 19.3 ms | 12.5x |
| cert_smolvla_r16_inv.json | lerobot/smolvla_base | rollout invariant, 16 robots/tick | 64 | 64/64 vs same-mode | 227 ms | 15.3 ms | 14.8x |
| cert_pi05_r8_exact.json | lerobot/pi05_base | rollout exact, 8 robots/tick | 64 | 64/64 | 315 ms | 175 ms | 1.8x |
| cert_pi05_r8_inv.json | lerobot/pi05_base | rollout invariant, 8 robots/tick | 64 | 64/64 vs same-mode | 314 ms | 110 ms | 2.9x |
| cert_groot_hp.json | nvidia/GR00T-N1.7-3B | exact, single robot | 32 | 32/32 | 170 ms | 63.7 ms | 2.7x |
| cert_groot_r8_inv.json | nvidia/GR00T-N1.7-3B | rollout invariant, 8 robots/tick | 64 | 64/64 vs same-mode | 179 ms | 120 ms | 1.5x |
| cert_v011_single.json | lerobot/smolvla_base | exact, single robot (release 0.1.1 wheel, licensed) | 32 | 32/32 | 230 ms | 36.8 ms | 6.2x |
| cert_v011_r16.json | lerobot/smolvla_base | rollout exact, 16 robots/tick (release 0.1.1 wheel, licensed) | 64 | 64/64 | 229 ms | 19.5 ms | 11.8x |
| cert_openvla_single.json | openvla/openvla-7b | exact, single robot (action tokens) | 16 | 16/16 | 209 ms | 145.4 ms | 1.4x |
H100 SXM 80 GB (h100/)
The H100 ran under a 500 W power cap (700 W maximum) and throttled during most runs; the
certificates say so in warnings, and their timings are indicative. Identity results are not
affected by clocks.
| File | Policy | Mode | Rows | Identical | Eager p50 | Proprio p50 per robot | Speedup |
|---|---|---|---|---|---|---|---|
| h100/cert_smolvla_single.json | lerobot/smolvla_base | exact, single robot | 32 | 32/32 | 164 ms | 40.2 ms | 4.1x |
| h100/cert_smolvla_r16_exact.json | lerobot/smolvla_base | rollout exact, 16 robots/tick | 64 | 64/64 | 166 ms | 16.1 ms | 10.3x |
| h100/cert_pi05_single.json | lerobot/pi05_base (bf16) | exact, single robot | 32 | 32/32 | 202 ms | 109.4 ms | 1.9x |
| h100/cert_pi05_r8_exact.json | lerobot/pi05_base (bf16) | rollout exact, 8 robots/tick | 64 | 64/64 | 204 ms | 77.9 ms | 2.6x |
| h100/cert_openvla_single.json | openvla/openvla-7b | exact, single robot (action tokens) | 16 | 16/16 | 237 ms | 88.4 ms | 2.7x |
OpenVLA's eager execution on H100 is not run-to-run reproducible by default; its reference was run
with CUBLAS_WORKSPACE_CONFIG=:4096:8 and torch.use_deterministic_algorithms(True).
Against LeRobot's own policy server (lerobot_server/)
proprio lerobot-server (release 0.1.2) is a drop-in for LeRobot's async-inference policy server.
Both servers ran in turn on one RTX 4090 through the real gRPC protocol, with the same 20
observations (three 480x640 cameras) and per-tick seeds, OMP_NUM_THREADS=8 unless noted. These
are comparison records, not certificates; each file says how it was produced.
| File | Policy | Identical chunks | LeRobot server p50 | Proprio server p50 |
|---|---|---|---|---|
| lerobot_server/smolvla_release_0.1.2_wheel.json | lerobot/smolvla_base | 20/20 | 330.8 ms | 52.1 ms (6.4x) |
| lerobot_server/smolvla_omp8.json | lerobot/smolvla_base | 20/20 | 302.4 ms | 52.6 ms (5.7x) |
| lerobot_server/pi05.json | lerobot/pi05_base (float32, as LeRobot loads it) | 20/20 | 1196 ms | 214.6 ms |
The pi0.5 row ran with default PyTorch threads for both servers on a 256-thread host, which slows LeRobot's eager server (SmolVLA: 1178 ms default, 302 ms with 8 threads); its ratio is not a fair speedup and is not quoted. Its result is the identity: 20 of 20 chunks.
Reading a certificate
mode:exactkeeps the policy's own kernels; rows are bit-identical to eager.rollout-*runs several robots per tick.*invariant*runs batched GEMMs on batch-invariant kernels: rows are bit-identical to the tier's own single-robot run (reference: same-mode single-robot) andeager_max_abs_errdeclares the measured difference to stock eager (0.04 to 0.21 here).max_abs_err,mean_abs_err: against the reference; 0 means every value identical.eager_p50_ms,engine_p50_ms,engine_p99_ms: median latencies; for rollout runs the engine figure is per robot.engine_pre_p50_ms,engine_post_p50_ms: host pre- and post-processing share of the tick.environment: GPU, driver, clocks, power state, throttle reasons, thread count at run time.seeds,noise_seed: what fixed the policy's randomness so the comparison is meaningful.licence,status: evaluation certificates are unsigned; certification-service certificates carry an Ed25519 signature.
Reproduce
pip install proprio-0.1.0-cp312-cp312-linux_x86_64.whl # https://github.com/zenedgex/proprio-releases/releases
proprio certify lerobot/smolvla_base --ticks 32 --out cert.json
Anything that is not EXACT on a LeRobot checkpoint is a bug report we want: open an issue on the releases repository with the certificate attached.
- Downloads last month
- 113