Compute Resource Requests and Selection
Configuration snapshot
The table below comes from the Master Course “设备情况” sheet updated on 2026-08-03. Host availability, GPU occupation, pricing, and project authorization change over time; verify them with tailscale status, nvidia-smi, and the current allocation record before launch.
Purpose
Select an appropriate laboratory compute resource based on visualization needs, run duration, GPU count, and distributed-training scale.
Contacts and Resource Entry Points
| Need | Contact |
|---|---|
| Headscale enrollment and visual workstations | @nilou (Ye Zheng) |
| AutoDL single-node workloads, normally up to eight GPUs | @赖咏曦 (Yongxi Lai) |
| Always-on single-GPU resources | @赵浩宇-Postdoc or the Compute Administrator |
| Alibaba Cloud DLC training beyond eight GPUs | Contact the Compute Administrator in advance |
Prerequisites
- The project, code repository, data location, and responsible owner are defined.
- GPU type, memory, count, runtime, and storage needs have been estimated.
- A minimal run has succeeded locally or in a single-GPU environment.
- The project owner and Compute Administrator have approved any usage-billed resource.
Resource Selection
| Task requirement | Recommended resource | Pre-launch requirement |
|---|---|---|
| Intensive visualization, real-time interaction, or remote GUI | Authorized laboratory workstation | Complete Headscale enrollment for the personal work computer |
| Standard training or temporary single-node GPU | Usage-billed platforms such as AutoDL | Estimate cost; normally no more than eight GPUs on one node, subject to current platform limits |
| Continuously running single-GPU workload | Three always-on shared single-GPU servers | Request a time window, storage, and responsible owner |
| Distributed training beyond one node | Alibaba Cloud DLC or another distributed platform | Complete a smoke test and prepare a one-line launch command in a single-GPU development environment |
| No GPU requirement | CPU, local, or low-cost development environment | Do not occupy scarce GPU resources |
Current Device Inventory
| Tailscale device | Tailscale IP | Compute | Current allocation | Location |
|---|---|---|---|---|
sg-ai-gateway |
100.64.0.1 |
No GPU | VPS control plane; not for development | Cloud |
boris-pc |
100.64.0.4 |
RTX 3070 Ti | Franka physical-robot control | E2-01-06 |
dm-26zj-020 |
100.64.0.7 |
RTX 5060 | Daimeng benchmark setup | Daimeng company |
dm-26zj-008 |
100.64.0.3 |
RTX 5090 | Daimeng benchmark setup | Daimeng company |
lvrobotics-System-Product-Name |
100.64.0.10 |
RTX 4060 Ti | AgileX NERO physical-robot control | E2-01-06 |
shaol-PC |
100.64.0.5 |
RTX 4090 | Ego2Dex | COM2-01-04 |
jingxiang-B850M-C |
100.64.0.6 |
RTX 5090 | Robot Harness Gen-Env | COM2-01-06 |
yuhang-B850M-C |
100.64.0.9 |
RTX 5090 | UMI WAM Stage 2 | E2-01-06 |
aliyun-dsw-1055 |
100.64.0.11 |
A100 | UMI WAM Stage 1 | Cloud |
aliyun-dsw-1021-h20 |
100.64.0.13 |
H20 | UMI WAM Stage 1 | Cloud |
aliyun-dsw-1056 |
100.64.0.21 |
A100 | UMI WAM Stage 1 | Cloud |
aliyun-dsw-1019 |
100.64.0.25 |
4×A10 | Data-pipeline encoding and decoding | Cloud |
Access note
These IPs are reachable only inside the laboratory Tailscale/Headscale network. The table is a configuration and allocation snapshot, not a claim that a device is currently online or idle. AutoDL is usage-billed and is not kept online.
Request Information
- Project and project-owner role
- Target host or resource type and whether the current allocation must change
- Repository, target branch, and reproducible environment instructions
- GPU model, count, expected duration, and requested start time
- Desktop visualization, external network, data-mount, or cross-node requirements
- Single-GPU smoke-test evidence, checkpoint strategy, and one-line launch command
- Budget source and the person responsible for shutdown
Procedure
- Select a resource category and submit the request to the corresponding contact.
- The administrator confirms live reachability, GPU occupation, cost, permissions, and data paths.
- Run a small smoke test to validate dependencies, data loading, logs, and checkpoint writes.
- Start the full job only after approval, recording the job ID, responsible owner, and expected end time.
- Monitor GPUs, logs, cost, and checkpoints; stop scaling or high-cost jobs when abnormal behavior appears.
- After completion, shut down usage-billed resources and confirm that data and checkpoints are safely stored.
Usage Rules
- Start usage-billed resources only when needed and shut them down immediately after use; never leave an ownerless idle job running.
- A host being online does not grant authorization; re-confirm the project allocation each time.
- Do not treat a successful launch as proof of healthy large-scale training; verify real processes, GPUs, logs, and checkpoints.
- Hostnames, IPs, and GPU specifications may be documented; passwords, API keys, tokens, private keys, and unattended-access passwords must not appear in commands, environment files, logs, or the Wiki.
- Use approved remote-desktop channels and never store unattended-access passwords in the Wiki.
Verification
- The target device name and IP in
tailscale statusmatch the table and are currently reachable. nvidia-smireports the approved GPU model, count, and current occupation.- The smoke test completes at least one valid step and produces the expected log or checkpoint.
- GPU count, process topology, and task configuration match the approved request.
- Usage-billed resources are shut down after completion, with cost and artifacts confirmed by the responsible owner.
Troubleshooting
- Device offline or IP mismatch: treat live Headscale state as authoritative and report the difference so this page can be updated.
- No GPU available: reduce scale, change the schedule, or ask the administrator to reallocate resources.
- Multi-GPU communication fails: retain the error summary and job ID, then confirm that the platform supports the requested topology.
- Remote desktop unavailable: verify the network path, then contact
@nilouor the on-site maintainer. - Data or checkpoints cannot be written: stop the full run and validate permissions and free space before resuming.
- Unexpected cost: stop the job immediately and notify the project owner and Compute Administrator.
Maintenance
- Headscale and workstations:
@nilou(Ye Zheng) - AutoDL:
@赖咏曦(Yongxi Lai) - Always-on resources:
@赵浩宇-Postdocand the Compute Administrator - Configuration source: Master Course → 设备情况
- Last verified: 2026-08-03