sequence-ai is an open-source SDK for serving robot policies from remote GPUs over wide-area networks (WANs). It
supports all open-source robot policies. It runs on your own GPU, or on one rented from
General Sequences near your robot: on the robot, only what connect() is given changes.
Give it a policy server and a robot. On the GPU, it replays the module a model calls many times per inference from CUDA graphs, with exactly the same outputs, and batches several robots' requests within the slack their action chunks leave. On the connection, it keeps the TCP congestion window open across the robot's pauses, so a 24 KB upload after a pause takes one round trip where it took two, at RTTs from 12 to 101 ms. Together these are GS-WAN (General Sequences Wide-Area Network). Each component is one file that can be added to existing code.
Quick start · Three ways to run it · Architecture · The problem · Inference layer · Network layer · Results · Limitations · Docs
Install
pip install sequence-ai # the robot's client, the seq command, every component
pip install "sequence-ai[server,arrays]" # and uvicorn, numpy, zstandard: the quick start's server
Python 3.9 or later for the client and the components, on Linux and macOS; the seq command needs Python 3.10 or
later. cuda_graphs uses the torch your model already has (2.4 or later). Each component is one file, and most import
only the standard library, so copying a file into your project works as well.
Quick start
On the robot, one call applies the client-side network components to one connection:
import sequence_ai
with sequence_ai.connect("http://your-gpu:8000") as policy_user: # BBR, TCP keepalive, a ping while idle, msgpack
policy_user.observe({"image": frame_user, "state": state_user})
actions_user = policy_user.act()["actions"]
On the GPU, the inference components wrap the model you already have:
from sequence_ai.engine import cuda_graphs
from sequence_ai.engine.batcher import Batcher
cuda_graphs.install(model_user, ["action_head.denoiser"]) # replayed from CUDA graphs; kept only if outputs match
batcher_user = Batcher(infer_a_batch_user, sizes=[1, 2, 4, 8]) # several robots' requests in one forward pass
action_user = await batcher_user(observation_user) # in each request's handler
Each name ending in _user is yours: your model, your frames, your handler. Both run on a laptop with no GPU and no
model:
python examples/quickstart_server.py & sleep 1 && python examples/quickstart_robot.py # a server and a robot
python examples/quickstart_batching.py # five robots, batched
The first line each prints:
request 0: 8 actions of 7 numbers, the first [128. 128. 128.]
5 robots served in 113 ms, as batches of {4: 1, 1: 1}
Quickstart: a policy server and a robot and Quickstart: batching and replicas walk through each with the output of every step.
Three ways to run it
| Whose GPU | What you run on it | On the robot | |
|---|---|---|---|
| Your own server | Yours, or any GPU the robot can reach | Your policy server, with the components you add | connect("http://your-gpu:8000") |
| A GPU near your robot | One rented from General Sequences, by city | The same server | connect("http://<its address>:8000") |
| A hosted deployment | General Sequences' | Nothing: the platform builds your policy, serves it and scores it | connect("my_policy_user") |
The first needs no account and costs nothing. The other two are paid services of General Sequences, billed while a GPU runs, and their commands are in this package:
seq gpu add --place "San Francisco, CA" --gpu L4 --port 8000 # a GPU near your robot, yours until seq gpu remove
seq deploy my_policy_user # or a deployment: the policy in this folder
GPUs are in short supply, and few are near any given robot, which is why the second exists:
How do I get a GPU near my robot? has its commands.
The platform guide has the third: writing a policy class, deploying it, benchmarks and evals,
teleoperation, and every seq command.
How it fits together
robot GPU
┌─────────────────────────┐ ┌───────────────────────────┐
│ your control loop │ │ your policy │
├─────────────────────────┤ observation ├───────────────────────────┤
│ sequence_ai.network │ ───────────────────▶ │ sequence_ai.engine │
│ connect · openpi │ │ cuda_graphs · batcher │
│ window · compress │ WAN │ replicas · slack │
│ frames · wifi_power │ ├───────────────────────────┤
│ │ ◀─────────────────── │ sequence_ai.network │
│ │ action chunk │ one_write · ping │
└─────────────────────────┘ │ tcp_link · compressed │
└───────────────────────────┘
both ends: msgpack_io · latency_stats your own link, measured: bench/
sequence_ai.network is the connection between a robot and the GPU its policy runs on. On the robot, connect()
applies the client-side components to one connection, and openpi.keep() does the same on the connection that openpi's
own client holds. On the server, one_write and ping wrap a uvicorn or a WebSocket server. sequence_ai.engine
is the inference layer on the GPU: cuda_graphs shortens one inference, batcher and slack share a GPU among robots,
and replicas runs one process per GPU. bench/ measures a link before anything is added to it.
The problem, by layer
A robot policy built on a vision-language-action (VLA) model runs each inference on a GPU, and a robot without a suitable GPU on board sends its observations over a WAN to a remote one. Each layer of that path has its own problem.
The model is most of the latency. Across thirteen model and network configurations, the model is 68 to 96% of a
request's latency. The time of one inference is therefore the main lever, and cuda_graphs shortens it without changing
the model's output.
One robot leaves its GPU mostly idle. One robot keeps a GPU busy 21 to 55% of the time, so several robots can share
one, provided that no robot's action chunk runs out while its request waits. batcher batches the requests that are
waiting, slack computes how long each may wait and how many robots a replica can serve, and replicas runs one
replica per GPU.
A robot pauses, and Linux restarts its connection. A robot executes a chunk between requests, and Linux restarts a
TCP connection's congestion window after an idle period. The upload after a pause then takes one extra round trip, at
RTTs from 12 to 101 ms. window, ping and connect keep the window open.
A request carries more round trips and bytes than it needs. A response split across writes can wait a round trip
behind a tunnel, and a camera frame is 50 to 100 KB. one_write sends each response in one write, compress compresses
an observation only when that shortens the upload, and frames resizes a frame before it is sent.
GPUs are in short supply. Few are near any given robot, and every millisecond of round-trip time comes out of the
slack an action chunk leaves. Getting a GPU near the robot is the routing layer: a hosted service of General
Sequences, reached with this package's seq gpu commands. Three ways to run it has them.
Inference layer
| Component | What it does | To add it | Requires |
|---|---|---|---|
cuda_graphs |
Replays a torch module that your model calls many times per run from CUDA graphs captured from its own eager kernels. A graph is kept only when its output is exactly the same. | one call | torch 2.4 or later, a CUDA GPU |
batcher |
Batches the calls waiting for a model into one call, in the batch sizes the model lists. | one object, one await |
nothing |
replicas |
Runs replicas of a model, each in its own process on the GPU assigned to it, and replaces one that dies. | one object, one await |
nothing |
slack |
Schedules by the slack an action chunk leaves: how many robots a replica can serve, each by its own round-trip time (RTT), how long a request may wait for a batch, and which requests are dispatched first. | a few calls in your server's loop | nothing |
from sequence_ai.engine import cuda_graphs, slack
from sequence_ai.engine.batcher import Batcher
from sequence_ai.engine.replicas import ReplicaPool
cuda_graphs.install(model_user, ["action_head.denoiser"]) # model_user.action_head.denoiser, graphed in place
batcher_user = Batcher(model_on_a_list_user, sizes=[1, 2, 4, 8], wait_ms=2.0)
pool_user = ReplicaPool(load_user, replicas=2, devices=[0, 1]) # one replica per GPU
table_user = slack.measure(forward_of_user, sizes=[1, 2, 4]) # forward_of_user(n): one forward pass of n requests
play_user = slack.playback(hz=20, valid_len=15, replan_at=8) # 15 valid steps, the next requested at the 8th
seating_user = slack.Seating(table_user, play_user["buffer_s"] * 1000, sizes=[1, 2, 4],
cycle_ms=(play_user["play_s"] - play_user["buffer_s"]) * 1000)
seats_user, over_user = seating_user.seat([("near", 30.0), ("far", 200.0)]) # (robot, its RTT in ms)
cuda_graphscounts the calls each module gets over a few runs (Census), captures the module that is called many times per run with the same shapes, and checks each new shape against eager mode byte for byte. A shape whose capture fails or differs runs in eager mode from then on. The module stays the same object of the same class, with its weights,state_dictand hooks unchanged. How do I make a PyTorch robot policy faster without changing its outputs?batcherdispatches a batch as soon as it is full, orwait_msafter its first call while the model is free. While the model is busy, waiting calls accumulate and are dispatched together. Only the listed sizes are used, and a batch in which the model raised is re-run call by call. How many robots can one GPU serve?replicasruns each replica in its own process on its own GPU, so a forward pass never blocks the server's event loop.ReplicaErroris the model's own error, and the replica takes the next call.ReplicaDiedis a process that has exited, and it is replaced. Should I run two replicas of a policy on one GPU?slackcomputes a replica's capacity from the playback time an action chunk leaves when its robot requests the next one, each robot by its own RTT, and tells a server how long the open batch may wait.slack: scheduling by the slack an action chunk leaves
How do I use the inference components with openpi, LeRobot or GR00T?
Network layer
"Client" is the robot's program and "server" is the policy server on the GPU.
| Component | Side | What it does | Requires |
|---|---|---|---|
window |
client | Keeps the connection's congestion window across the robot's pauses: BBR on its socket, TCP keepalive, a ping while the connection is idle. | nothing |
compress |
client | Compresses an observation's large arrays only where that saves more upload time than it costs. | nothing; zstd where it is installed |
frames |
client | Resizes a camera frame to the policy's input size before upload. | Pillow |
wifi_power |
client | Detects Wi-Fi power saving on the robot's link and reports how to disable it. | Linux |
one_write |
server | Sends each HTTP response of a uvicorn server in one write instead of three. | uvicorn |
tcp_link |
server | Reads each client's TCP state from the server's kernel: RTT, retransmitted segments, congestion window. | Linux |
compressed |
server | Decompresses what the client compressed, byte for byte. | numpy; zstandard for zstd |
ping |
server | A server-side ping, so that the robot's connection is never idle: the options to pass to your server. | a WebSocket server |
msgpack_io |
both | Encodes observations and responses as msgpack, with numpy arrays as raw bytes. Also reads openpi's layout. | nothing |
latency_stats |
both | Counters, gauges and latencies with their p50, p95 and p99, read as one snapshot. | nothing |
Combined, and for clients you may already use:
| Side | What it does | Requires | |
|---|---|---|---|
connect |
client | The client-side components on one connection to your own server. | httpx |
openpi |
client | The client-side components on the connection held by openpi's own client. | openpi-client |
lerobot |
both | The Zenoh setting that keeps LeRobot's remote-inference connection from going idle. | lerobot 0.7 |
bench/ |
both | Measures your own link: what a pause, a ping and a camera stream cost on it. | nothing |
from sequence_ai.network import lerobot, openpi
kept_user = openpi.keep(policy_user) # openpi: one line on its WebsocketClientPolicy
path_user = lerobot.keep_alive_file("zenoh-keep-alive.json5") # LeRobot 0.7: one Zenoh setting, given to both sides
What does each component do? gives the lines to add, the reason and the measurement for each
one, and examples/components/ has a runnable file for each.
How do I run π0 (openpi) on a remote GPU? |
How do I reduce the latency of LeRobot remote inference?
| How do I serve a robot policy from my own GPU?
Results
Across thirteen model and network configurations, the time a request spends outside the model ranges from 13 to 137 ms, and its share of the latency from 4 to 32%. A remote GPU is therefore close to an on-board GPU of the same type, and inference time is the main lever.
Inference layer:
| Component | Measured on | Without | With |
|---|---|---|---|
cuda_graphs |
The diffusion transformer of GR00T N1.7 on one L4; outputs compared on inputs it was not captured on | 224 ms per run | 146 ms per run (1.54×), exactly the same outputs |
cuda_graphs and torch.compile |
The same model and GPU | torch.compile(backend="cudagraphs"): 225 ms (1.00×). inductor: 1.45×, outputs 0.0051 apart on average |
1.54×, exactly the same outputs |
cuda_graphs |
A small flow-style model of many small kernels on one T4 (examples/check_on_a_gpu.py) |
16.1 ms per run | 2.4 ms per run, byte for byte |
| Batching | π0.5 on one L4, the model's own batched call | 5.99 requests per second at batch 1 | 8.21, 9.38, 10.52 and 10.66 at batches of 2, 4, 8 and 16 (1.78×) |
slack: the load a replica takes |
Four robots, 300 s and about 2,100 requests per run, a stand-in model paced like a 15 Hz policy, in the serving stack the component was taken from | All four on one replica, a load of 0.95: 85.7% of responses within the 467 ms buffer | Two per replica on two replicas, a load of 0.88: 94.1% |
Network layer:
| Component | Measured on | Without | With |
|---|---|---|---|
window, ping: each of four ways |
A 24 KB upload 0.53 s after each response; two cloud machines, RTTs of 12, 21, 41, 65, 81 and 101 ms; 40 uploads per cell | 2.0 round trips, 40 of 40 uploads | 1.0 round trips, 0 of 40: the server's ping, the robot's ping, BBR on the robot's socket, or tcp_slow_start_after_idle=0 |
window: BBR |
A robot sending a request every 0.53 s with 24 KB of frames; RTT about 21 ms; 40 requests per setting | 40 of 40 requests took one extra round trip (21.5 ms) | 0 of 40 (0.5 to 3.0 ms) |
window: BBR, with 1% loss |
The 24 KB upload at 22 ms with 1% loss on the link | 40 of 40 uploads took one extra round trip | 0 of 40 |
openpi, and ping over a WebSocket |
A 100 KB upload 0.53 s after each response; an openpi-style WebSocket client; RTTs of 11, 21, 41 and 81 ms; 40 uploads per cell | 37.6, 44.4, 82.4 and 163.8 ms (3.3, 2.1, 2.0 and 2.0 round trips) | 12.7, 22.7, 41.8 and 82.6 ms with openpi.keep() (1.1, 1.1, 1.0 and 1.0 round trips); 1.2, 1.1, 1.0 and 1.0 round trips with the server pinging |
lerobot |
A 100 KB request 0.53 s after each response between two Zenoh peers; the same machines and RTTs | 38.1, 44.8, 82.8 and 164.3 ms (3.4, 2.1, 2.0 and 2.0 round trips) | 13.9, 23.2, 42.3 and 82.2 ms with the keep-alive file on both ends (1.2, 1.1, 1.0 and 1.0 round trips) |
ping, the server's ping |
100 KB, 500 KB and 1 MB after 500 ms of silence; RTTs of 11.7 and 69.2 ms; 20 per cell | 3.0 round trips each | 1.1, 1.1 and 1.25 |
window.Warm, the robot's ping |
162 KB uploads 3 s apart; RTT 226 ms; 30 requests per setting | 303.6 ms for the upload to arrive | under 1 ms |
one_write |
A GET behind one kind of tunnel; TCP connect p50 89.5 ms | 189.7 ms (headers, then body) | 106.7 ms (one write) |
compress |
Camera frames of 172,800 bytes, decoded from a public dataset's video, on a laptop | 172,800 bytes | about 1.9 times smaller, for 0.37 ms to compress and 0.21 ms to decompress |
Measurements lists every command, and bench/ measures your own link.
Limitations
cuda_graphsis for PyTorch models bound by kernel launches. π0.5 under JAX already fuses and graphs its kernels and was 98% GPU-busy: there is no launch overhead to remove. A module called once per run, or with shapes that change on every call, gains nothing. Only GR00T N1.7's diffusion transformer and a small test model have been captured.- Batching does not amortize a model's prefix. On π0.5 the prefix (images and prompt) took 138 ms per request at batch 1 and 119 ms at batch 16, so the gain levels off near 1.78×. Called as a server calls it, each observation through the policy's input and output transforms, a batch of 8 gave 1.52× (5.1 to 7.8 requests per second). On that model and an L4, batching added no capacity: one replica served 2 robots of a 15 Hz policy in time, and so did openpi's own server, one request at a time. How many robots can one GPU serve?
- Two replicas on one GPU give no throughput gain. With π0.5 on an L4, two replicas whose calls overlapped each took 2.1 times as long as one alone. With the CUDA Multi-Process Service (MPS) they were 6 to 9% faster than that. Use one replica per GPU.
batcherandreplicashave unit tests and no measurement of their own. The batching result measures the model's own batched call. Theslackresults come from the serving stack the file was taken from, and from runs of the file itself around the batched calls of π0.5, GR00T N1.6 and SmolVLA on an L4, in a test server with simulated robots. No inference component has been run inside openpi's own policy server or LeRobot's.slackcomputes a capacity on a simulated clock, not on your model under load. It runs the robots as its loop would serve them, from up to 128 starts, and a number of robots is within capacity when 95% of their responses are in time. From three models' batch times on an L4 it gives 2 robots of a 15 Hz policy for π0.5 and GR00T N1.6, which 2 robots confirmed, and 1 for SmolVLA, where a formula for the load had given 8 and 4 robots had 95% of their responses late. Rerun on SmolVLA, 3 robots had late responses and 2 had none: the 1 comes from the limit on the load, which may be low for that model. Why the capacity is computed on a clock has the measurement, andexamples/fleet_on_a_clock.pyprints what the clock runs for your batch times.- A policy that leaves no slack is not batched by
slack. Where the robot sends a request only when a chunk is fully executed, a replica serves one robot. - The idle restart costs nothing on a long link. At an RTT of 242 ms the 24 KB upload took 1.0 round trips with
no component, and at about 101 and 241 ms a robot on a fixed 0.53 s period took 0 or 1 extra round trips in 40.
Whether a robot is affected depends on its RTT, its pause and its upload size:
bench/measures it. - A ping does not restore a window that losses cut. With 1% loss at RTTs of 22 and 101 ms, 31 and 35 of 40 uploads still took an extra round trip with the robot's ping. BBR removed it at 22 ms.
- On one home connection the ping helped only small uploads. From a Linux client, a 100 KB upload after a pause went from 5.5 round trips to 1.7 with the server's ping. 500 KB and 1 MB took 7.8 and 12.7 round trips with it and 10.3 and 13.4 without: the line's upload rate and packet loss were the cost, and BBR took 5.1 and 8.3. The table has every cell.
openpi.keep()and the LeRobot setting were measured mostly with stand-ins. The WebSocket and Zenoh measurements used thewebsocketslibrary and two Zenoh peers.openpi.keep()was also run once with openpi's own server and client, on one link, two episodes per setting; the LeRobot setting was not run with LeRobot's own server and robot.one_writewas measured behind one kind of tunnel, not on a direct connection.
Measurements has the rest.
Repository layout
sequence_ai/
engine/ inference layer: cuda_graphs · batcher · replicas · slack
network/ network layer
client: connect · window · compress · frames · wifi_power · openpi
server: one_write · ping · tcp_link · compressed
both: msgpack_io · latency_stats · lerobot
... the client for General Sequences' GPUs and deployments: connect, gpu, teleop, record
seq/ the seq command, and the decorators a hosted policy or benchmark is written with
bench/ measures a link: a TCP client and server, WebSocket, Zenoh
examples/ the quick starts, and one runnable file per component
docs/ one page per question, the measurements, the FAQ
tests/ the components' tests: Python 3.9 to 3.12, on a machine with no GPU
tools/ writes llms-full.txt
platform.md the guide to the hosted platform; eval.md and observability.md are its references
About 4,500 lines in sequence_ai/engine and sequence_ai/network, and 202 test functions.
Documentation
- Quickstart: a policy server and a robot
- Quickstart: batching and replicas
- Inference layer
- Network layer
- How do I serve a robot policy from my own GPU?
- How do I run π0 (openpi) on a remote GPU?
- How do I reduce the latency of LeRobot remote inference?
- Why does my robot's upload get slow after a pause?
- How do I measure the link between my robot and my GPU?
- How do I keep a WebSocket policy server's connection warm?
- What do the network components save, measured with the tools in this repository?
- How do the components compare with openpi's remote inference and LeRobot's?
- What does each network component do?
- General Sequences' GPUs
- How do I get a GPU near my robot?
- The platform guide, with its references for benchmarks and evals and debugging a deployment
- generalsequences.com/docs
- FAQ: errors you may see, and what to do next
llms.txt is a map of the repository for a coding assistant, and llms-full.txt is
this README and every page of docs/ in one file.
A note on the numbers
Every number in this README is given with the model, the GPU or the link it was measured on, and
Measurements lists the command behind each one. What a network component saves depends on the
link, so bench/ measures yours before you add anything. Each module's docstring states what the component
does in one sentence, why, and what was measured.
Develop
uv venv --python 3.11
uv pip install -e ".[dev]"
.venv/bin/python -m pytest -q
.venv/bin/ruff check sequence_ai/engine sequence_ai/network tests examples bench tools
The tests are the components', and run on Python 3.9 to 3.12 on a machine with no GPU. The client for the hosted
platform and the seq command are developed in General Sequences' own repository and published here, as they ship, at
each release. CONTRIBUTING.md has what that means for a pull request.
Contributing
Issues and pull requests are welcome: see CONTRIBUTING.md and the Code of Conduct. To report a security problem, see SECURITY.md.
Cite
If you use sequence-ai in research, please cite it:
@software{generalsequences2026sequenceai,
title = {sequence-ai},
author = {{General Sequences}},
year = {2026},
url = {https://github.com/General-Sequences/sequence-ai}
}
CITATION.cff has the same, with the version.
License
Apache-2.0. Copyright 2026 General Sequences Company. One component is adapted from another project, and the
page seq eval replay serves carries others' files: THIRD_PARTY_NOTICES.md.
Metadata
Release files for sequence-ai 0.62.2
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| sequence_ai-0.62.2.tar.gz | 1.6 MB | Without provenance | Details for sequence_ai-0.62.2.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| sequence_ai-0.62.2-py3-none-any.whl | Python 3 | none | any | Details for sequence_ai-0.62.2-py3-none-any.whl |
Total release size: 3.3 MB
Release files / sequence_ai-0.62.2.tar.gz
| Download URL | sequence_ai-0.62.2.tar.gz |
|---|---|
| Size | 1.6 MB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
da9e595990f03ef909fc973ce159effc225709308b1ede7a726a4ef1299e50a3
|
|
BLAKE2b-256 checksum How to use checksums |
69764db45ab86d41b4d4ff190d22271adacfa4381862bf22526924359fb046fc
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
uv/0.10.2 {"installer":{"name":"uv","version":"0.10.2","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"macOS","version":null,"id":null,"libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":null}
|
Release files / sequence_ai-0.62.2-py3-none-any.whl
| Download URL | sequence_ai-0.62.2-py3-none-any.whl |
|---|---|
| Size | 1.7 MB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
9a94ae8899aee5eaba43b9ffaa513d18f521d0e34eb7591733a1c4caab718faf
|
|
BLAKE2b-256 checksum How to use checksums |
9d345fee7a8b99d1cead358f29841edac796848d91c01a35cd2c07f43b069f86
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
No |
| Uploaded via |
uv/0.10.2 {"installer":{"name":"uv","version":"0.10.2","subcommand":["publish"]},"python":null,"implementation":{"name":null,"version":null},"distro":{"name":"macOS","version":null,"id":null,"libc":null},"system":{"name":null,"release":null},"cpu":null,"openssl_version":null,"setuptools_version":null,"rustc_version":null,"ci":null}
|
Security for sequence-ai 0.62.2
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
No provenance found
None of the files in this release were uploaded with provenance metadata.
A note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.