mirror of
https://github.com/huggingface/lerobot.git
synced 2026-07-16 22:41:49 +00:00
feat(rollout): remote inference draft
This commit is contained in:
Steven Palma
@@ -0,0 +1,115 @@
|
||||
# Copyright 2026 The HuggingFace Inc. team. All rights reserved.
|
||||
#
|
||||
# Licensed under the Apache License, Version 2.0 (the "License");
|
||||
# you may not use this file except in compliance with the License.
|
||||
# You may obtain a copy of the License at
|
||||
#
|
||||
# http://www.apache.org/licenses/LICENSE-2.0
|
||||
#
|
||||
# Unless required by applicable law or agreed to in writing, software
|
||||
# distributed under the License is distributed on an "AS IS" BASIS,
|
||||
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
||||
# See the License for the specific language governing permissions and
|
||||
# limitations under the License.
|
||||
|
||||
# Example manifest for `lerobot-policy-server --manifest server.yaml`.
|
||||
#
|
||||
# One process = one (model, revision, dtype, device) on one GPU. Dynamic
|
||||
# model loading is deliberately unsupported: pre-warmed processes keep
|
||||
# capacity planning honest. Every field below can also be overridden on
|
||||
# the command line via draccus, e.g. --model.repo_or_path=... or
|
||||
# --zenoh.connect_endpoints='["tcp/other-router:7447"]'.
|
||||
#
|
||||
# Field names mirror the dataclasses in src/lerobot/policy_server/manifest.py.
|
||||
|
||||
# --- Which policy this process serves, and where it runs ------------------
|
||||
model:
|
||||
# Hub repo id (org/name) or a local checkpoint directory. Required.
|
||||
repo_or_path: lerobot/pi0_towels
|
||||
# Hub revision: branch, tag, or commit sha.
|
||||
revision: main
|
||||
# Optional torch dtype cast applied after load (e.g. "bfloat16",
|
||||
# "float16"). null keeps the checkpoint's native dtype.
|
||||
dtype: bfloat16
|
||||
# Inference device, e.g. "cuda", "cuda:1", "cpu".
|
||||
device: cuda
|
||||
|
||||
# --- Task namespace --------------------------------------------------------
|
||||
# The task this service is published under. VLA clients may override the
|
||||
# task per session unless `pin_task` is true, in which case session opens
|
||||
# with a different task string are rejected.
|
||||
default_task: "fold the towel"
|
||||
pin_task: false
|
||||
# Optional override for the <task_slug> key segment of the Zenoh prefix
|
||||
# (defaults to a slug of `default_task`).
|
||||
service_name: ""
|
||||
|
||||
# --- Serving mode & capacity ------------------------------------------------
|
||||
# "auto" resolves from the policy classification: shared for verified
|
||||
# chunk-stateless policies (act/pi0/pi05, smolvla with n_obs_steps=1),
|
||||
# exclusive otherwise. Chunk-stateful policies — e.g. diffusion, whose
|
||||
# predict_action_chunk reads select_action-fed queues — are always forced
|
||||
# to "exclusive" (max_sessions=1); "shared" cannot override that.
|
||||
serving_mode: auto
|
||||
|
||||
# Capacity rule-of-thumb: with t = server seconds per inference, r = each
|
||||
# client's request rate (self-clocked to ~1-4 Hz, not the control rate),
|
||||
# H = RTC execution horizon, and dt = control period:
|
||||
# max_sessions ~= min( 0.8 / (r*t), (H*dt/2 - network RTT) / t )
|
||||
# e.g. ACT @ 20 ms, 1 Hz refresh -> ~40 clients/GPU; Pi0 @ 150 ms -> ~5.
|
||||
# Session opens beyond this are rejected with the current load in the
|
||||
# reply, so clients retry another replica.
|
||||
max_sessions: 5
|
||||
|
||||
# Dummy inferences run at startup so the first real request does not pay
|
||||
# for CUDA graph/kernel warmup.
|
||||
warmup_inferences: 2
|
||||
|
||||
# --- FPS contract -----------------------------------------------------------
|
||||
# Control rate the policy was trained at. Clients reporting a different
|
||||
# fps get a warning — or a hard reject when `strict_fps` is true.
|
||||
trained_fps: 30.0
|
||||
strict_fps: false
|
||||
|
||||
# --- Real Time Chunking (RTC) -----------------------------------------------
|
||||
# Global to this process: init_rtc_processor mutates the policy instance,
|
||||
# so RTC is a per-process decision, not per-session. Only rtc-capable
|
||||
# families (pi0/pi05/smolvla) honor it; others are downgraded to plain
|
||||
# chunk-append at session open.
|
||||
rtc:
|
||||
enabled: true
|
||||
# Number of actions executed from each chunk before the next chunk is
|
||||
# blended in (the H in the capacity formula above).
|
||||
execution_horizon: 10
|
||||
|
||||
# --- Housekeeping ------------------------------------------------------------
|
||||
# Sessions with no liveliness token and no traffic for this long are
|
||||
# garbage-collected (belt-and-braces behind liveliness GC).
|
||||
session_idle_timeout_s: 300.0
|
||||
|
||||
# --- Transport ----------------------------------------------------------------
|
||||
# Robots and servers both *dial out* to a zenohd router in production
|
||||
# (mode: client). mode: peer + listen_endpoints supports router-less LAN
|
||||
# and loopback test deployments. Multicast scouting is always disabled:
|
||||
# fleet discovery is configuration, not protocol magic.
|
||||
zenoh:
|
||||
mode: client
|
||||
connect_endpoints:
|
||||
- tcp/router.gpu-cluster.internal:7447
|
||||
listen_endpoints: []
|
||||
# mTLS material (PEM paths). All three are required for tls/ endpoints;
|
||||
# leave them null for plain tcp/ inside a trusted network.
|
||||
# tls_root_ca_certificate: /etc/lerobot/tls/ca.pem
|
||||
# tls_connect_certificate: /etc/lerobot/tls/server.pem
|
||||
# tls_connect_private_key: /etc/lerobot/tls/server.key
|
||||
# Escape hatch: raw JSON5 merged into the zenoh config last.
|
||||
# extra_config_json5: '{transport: {link: {tx: {queue: {size: {data: 4}}}}}}'
|
||||
|
||||
# --- Observability -------------------------------------------------------------
|
||||
# HTTP health + Prometheus metrics port; 0 disables the endpoint.
|
||||
health_port: 9100
|
||||
|
||||
# Optional bounded request/response capture for offline replay.
|
||||
debug:
|
||||
capture_dir: null
|
||||
capture_max: 256
|
||||
@@ -1,17 +0,0 @@
|
||||
from lerobot.async_inference.configs import PolicyServerConfig
|
||||
from lerobot.async_inference.policy_server import serve
|
||||
|
||||
|
||||
def main():
|
||||
host = ... # something like "127.0.0.1" if you're exposing to localhost
|
||||
port = ... # something like 8080
|
||||
|
||||
config = PolicyServerConfig(
|
||||
host=host,
|
||||
port=port,
|
||||
)
|
||||
serve(config)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
@@ -1,62 +0,0 @@
|
||||
import threading
|
||||
|
||||
from lerobot.async_inference.configs import RobotClientConfig
|
||||
from lerobot.async_inference.helpers import visualize_action_queue_size
|
||||
from lerobot.async_inference.robot_client import RobotClient
|
||||
from lerobot.cameras.opencv import OpenCVCameraConfig
|
||||
from lerobot.robots.so_follower import SO100FollowerConfig
|
||||
|
||||
|
||||
def main():
|
||||
# these cameras must match the ones expected by the policy - find your cameras with lerobot-find-cameras
|
||||
# check the config.json on the Hub for the policy you are using to see the expected camera specs
|
||||
camera_cfg = {
|
||||
"up": OpenCVCameraConfig(index_or_path=0, width=640, height=480, fps=30),
|
||||
"side": OpenCVCameraConfig(index_or_path=1, width=640, height=480, fps=30),
|
||||
}
|
||||
|
||||
# # find ports using lerobot-find-port
|
||||
follower_port = ... # something like "/dev/tty.usbmodem58760431631"
|
||||
|
||||
# # the robot ids are used the load the right calibration files
|
||||
follower_id = ... # something like "follower_so100"
|
||||
|
||||
robot_cfg = SO100FollowerConfig(port=follower_port, id=follower_id, cameras=camera_cfg)
|
||||
|
||||
server_address = ... # something like "127.0.0.1:8080" if using localhost
|
||||
|
||||
# 3. Create client configuration
|
||||
client_cfg = RobotClientConfig(
|
||||
robot=robot_cfg,
|
||||
server_address=server_address,
|
||||
policy_device="mps",
|
||||
client_device="cpu",
|
||||
policy_type="act",
|
||||
pretrained_name_or_path="<user>/robot_learning_tutorial_act",
|
||||
chunk_size_threshold=0.5, # g
|
||||
actions_per_chunk=50, # make sure this is less than the max actions of the policy
|
||||
)
|
||||
|
||||
# 4. Create and start client
|
||||
client = RobotClient(client_cfg)
|
||||
|
||||
# 5. Provide a textual description of the task
|
||||
task = ...
|
||||
|
||||
if client.start():
|
||||
# Start action receiver thread
|
||||
action_receiver_thread = threading.Thread(target=client.receive_actions, daemon=True)
|
||||
action_receiver_thread.start()
|
||||
|
||||
try:
|
||||
# Run the control loop
|
||||
client.control_loop(task)
|
||||
except KeyboardInterrupt:
|
||||
client.stop()
|
||||
action_receiver_thread.join()
|
||||
# (Optionally) plot the action queue size
|
||||
visualize_action_queue_size(client.action_queue_size)
|
||||
|
||||
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
Reference in New Issue
Block a user