A unified orchestration layer for heterogeneous AI compute. It standardizes how to manage compute and run training and inference on GPU clouds, Kubernetes, VMs, or bare-metal clusters.
Repo dstackai/dstack, read from GitHub 16 min ago. dstackai has not joined Take My Repo.
The kubernetes backend now supports NVIDIA MIG slices configured with the GPU Operator's MIG Manager (the single or mixed strategy). dstack treats each slice as a GPU with the slice's memory, e.g., a 1g.10gb slice of an A100 is A100:10GB. To run on slices, specify the GPU memory:
type: dev-environment ide: vscode resources: gpu: A100:10GB
If the GPU memory isn't specified or has no maximum (e.g., gpu: A100 or gpu: 24GB..), the run gets whole GPUs only. With --full-offers, dstack offer also lists matching slices:
$ dstack offer --gpu A100 --full-offers # BACKEND RESOURCES INSTANCE TYPE PRICE 1 kubernetes (gke_dstack_us-central1-a_mig-e2e) cpu=x86:23 mem=156.9GB disk=97.4GB gpu=A100:40GB:1 gke-mig-e2e-a100x2-946ce248-zsmm $0 2 kubernetes (gke_dstack_us-central1-a_mig-e2e) cpu=x86:23 mem=156.9GB disk=97.4GB gpu=A100:20GB:1 gke-mig-e2e-a100x2-946ce248-zsmm $0 3 kubernetes (gke_dstack_us-central1-a_mig-e2e) cpu=x86:23 mem=156.9GB disk=97.4GB gpu=A100:20GB:1 gke-mig-e2e-a100x2-946ce248-zsmm $0
Slices with the same memory are the same GPU for dstack, so a specific MIG profile (e.g., 3g.40gb vs 4g.40gb) can't be requested. dstack metrics doesn't show GPU metrics for slices yet. See Kubernetes for details.
Important
On clusters that already use the single strategy, slices used to be offered as whole GPUs (e.g., A100:80GB), so gpu: A100 runs could land on them. Now these runs get whole GPUs only. To keep running on slices, specify the GPU memory, e.g., gpu: A100:10GB.
Nebius offers used to have unknown availability, so dstack could attempt configurations blocked by quota or capacity. The new capacity_advisor setting uses Nebius Capacity Advisor to show availability and quota in offers and skip offers that can't be provisioned:
projects: - name: main backends: - type: nebius creds: type: service_account service_account_id: serviceaccount-e00dhnv9ftgb3cqmej public_key_id: publickey-e00ngaex668htswqy4 private_key_file: ~/path/to/key.pem capacity_advisor: true
The setting is off by default because it adds roughly 1–3 seconds to getting offers. Availability and quota are cached for 10 minutes. See Nebius.
SSH fleets now support in-place updates of ssh_config.identity_file and ssh_config.proxy_jump.identity_file, including per-host overrides. Previously, these changes required re-creating the fleet, which is blocked while runs are active.
Authorize the new public key on the host, update the identity file path, and apply the fleet configuration:
$ dstack apply -f fleet.dstack.yml Found fleet my-fleet. Detected changes that can be updated in-place: - Configuration properties: - ssh_config.identity_file - ssh_config.proxy_jump.identity_file
Runs continue without restarts, and established SSH connections, including attach sessions and service tunnels, stay open. After updating proxy_jump.identity_file, allow about 20 seconds before removing the old public key from the proxy_jump host, so that gateways and the SSH proxy receive the new key. See Update fleets.
In dstack Enterprise, Factory, and Sky, backend fleets can now be exported to other projects. Importer projects can run on the fleet's existing instances:
$ dstack import list
NAME FLEETS GATEWAYS
main/my-export shared-fleet -
$ dstack fleet list
NAME NODES GPU SPOT BACKEND PRICE STATUS CREATED
main/shared-fleet 2..4 auto aws active 17 mins ago
instance=0 - spot aws (eu-north-1) $0.012 idle 17 mins ago
instance=1 - spot aws (eu-north-1) $0.012 idle 17 mins agoNote
Limitation: Currently, this is intended to be used with fleets that already have provisioned nodes. New instances are not provisioned for importer runs. Additionally, importer runs cannot mount network volumes on such fleets. This limitation will be lifted in one of the upcoming releases.
Services with the dynamo router now support retry and in-place updates with dstack apply, like any other service.
retry: on_events: [error, interruption]
Previously, a service whose replicas exited with code 0 either had them resubmitted or ended as done, depending on whether they all finished at the same time. Now services never end because their replicas exited. Such replicas are brought back by scaling, and if all of them exit at once, the service is resubmitted with a backoff.
User tokens throughout the UI are now hidden until you click Show token, so they don't leak when you share or record your screen. Copy buttons still copy the real token.
dstack now requires Python 3.11 or later. Run configurations can still use python: "3.9" and "3.10".
The DSTACK_ROUTER_INTERNAL_IP environment variable, deprecated in 0.22.6, has been removed. Services that still use it must switch to ${{ groups[i].replicas[j].IP_ADDRESS }}:
# Before - export ETCD_ENDPOINTS=http://$DSTACK_ROUTER_INTERNAL_IP:2379 # After, if the router is the first group - export ETCD_ENDPOINTS=http://${{ groups[0].replicas[0].IP_ADDRESS }}:2379
type: dynamo router by @un-def in #4474type: dynamo router by @un-def in #4475Full changelog: 0.22.7...0.22.8
Orphan cleanup, introduced in 0.22.6, now supports the kubernetes backend. It finds pods, services, and secrets left behind by dstack and reports or deletes them, depending on DSTACK_SERVER_ORPHAN_CLEANUP_MODE.
The new proxy_jump.node_selector and proxy_jump.tolerations properties let you run the jump pod only on specific nodes, e.g. nodes with an ExternalIP or CPU-only nodes:
type: kubernetes kubeconfig: filename: ~/.kube/config proxy_jump: node_selector: - example.com/jump-host tolerations: - example.com/dedicated=jump-host:NoSchedule
tolerations uses the kubectl taint format. See Kubernetes for details.
Kubernetes now recreates the jump pod used for SSH access to runs if it is evicted or its node is removed. Previously, an evicted jump pod prevented new runs from being provisioned.
New jump pods also reserve CPU and memory, making them less likely to be evicted under node pressure. Existing jump pods keep their resource settings until they are recreated.
Important
If dstack uses a Kubernetes user or service account with a custom Role, update the Role before upgrading the server. dstack now needs get and create on replicasets (API group apps) and list on pods. Without them, provisioning fails with 403 Forbidden.
If list on services or secrets is not granted, orphan cleanup skips those resources and logs a warning. See Required permissions for the full Role.
Gateways now run on Python 3.12 managed by uv instead of the OS Python. Existing gateways are updated automatically when you upgrade the server. Services are unavailable for a few seconds while their gateway restarts.
Gateway hosts now download uv and Python from GitHub when the gateway is installed or updated, so they need outbound access to github.com, in addition to PyPI.
apply by @jvstme in #4442Full changelog: 0.22.6...0.22.7
In rare cases, a dstack bug or server crash can leave provisioned resources orphaned. This release adds automatic detection and optional cleanup for the aws, azure, gcp, nebius, and runpod backends. Resources are identified by their cloud tags or names and matched against the project's provisioning records.
By default, detection runs in report mode: it writes logs, emits events, and reports orphaned resources to Sentry if DSTACK_SENTRY_DSN is configured.
$ dstack event --since 1d
[2026-10-29 16:25:03] [instance my-fleet-2] Orphaned resources found in aws/eu-central-1,
provider_ref q4rd28b3wm8y, instance id i-0047e4040282e969c:
instance i-0047e4040282e969c (t3.medium, running)Detection and cleanup cover cloud instances and their attached resources, such as IPs and disks. Support for volumes and gateways is on the roadmap.
To delete orphaned resources automatically, set DSTACK_SERVER_ORPHAN_CLEANUP_MODE=delete. Future releases will make delete the default.
GPU and CPU offers are now grouped by availability, keeping unavailable capacity separate. Per-GPU VRAM reporting has also been corrected.
The Offers page now shows both on-demand and spot offers by default. Displayed prices are rounded to at most two decimal places.
In dstack Factory and Sky, the global Events page now supports relative and absolute date/time filters. Click Recorded At to switch between newest-first and oldest-first order.
Sorting is also available on resource Events tabs in open-source dstack.
The DSTACK_ROUTER_INTERNAL_IP environment variable, which dstack sets for replicas of services that use a dynamo router, is deprecated in favor of ${{ groups[i].replicas[j].IP_ADDRESS }}. The new interpolation can reference any replica in the service, including the router. The variable may be removed as early as the next release.
# Before - export ETCD_ENDPOINTS=http://$DSTACK_ROUTER_INTERNAL_IP:2379 # After, if the router is the first group - export ETCD_ENDPOINTS=http://${{ groups[0].replicas[0].IP_ADDRESS }}:2379
The NVIDIA Dynamo example now runs etcd as a separate replica group, so the router no longer hosts etcd and NATS.
The Private registry section now also explains how to use registry_auth with public registries to reduce rate-limit errors through higher limits for authenticated users.
registry_auth for higher rate limits by @jvstme in #4414openai by @jvstme in #4437azure-mgmt-compute by @jvstme in #4435Full changelog: 0.22.5...0.22.6
Run configurations now accept python: "3.15":
type: dev-environment python: "3.15" ide: vscode
The default Docker image now preinstalls Python 3.11 to 3.15. Runs use the preinstalled build instead of downloading Python on every container startup. See python.
The release adds experimental support for Lium. The lium backend supports GPU runs, private Docker images, and volumes.
projects: - name: main backends: - type: lium creds: type: api_key api_key: ...
dstack offer -b lium --group-by gpu
# GPU SPOT $/GPU BACKENDS
1 RTX3090:24GB:1 on-demand 0.4 lium
2 RTX5070Ti:16GB:1 on-demand 0.4 lium
3 A6000:48GB:1..5 on-demand 0.42..0.55 lium
4 RTX6000Ada:48GB:1 on-demand 0.8 lium
5 RTXPRO6000:96GB:8 on-demand 1.88..2.09 lium
6 H100:80GB:8 on-demand 1.97..3.75 lium
7 H200NVL:141GB:4 on-demand 4.5 lium
8 H200:141GB:1..4 on-demand 5.36..5.58 lium
9 B200:180GB:1..8 on-demand 5.5..7.5 lium
10 B300:270GB:1..8 on-demand 8..10 liumLium volumes use S3-backed storage, grow with usage, and can be mounted in any Lium region, so size and region are omitted:
type: volume backend: lium name: my-volume
Important
TPU offers are no longer suggested by default. If you use TPUs with the gcp backend, enable them with tpu: true before upgrading.
TPU support is experimental, and TPU offers used to appear for anyone with a GCP backend. It is now opt-in:
projects: - name: main backends: - type: gcp project_id: my-project creds: type: default tpu: true
See GCP.
The Offers page now shows CPU offers as well. Turn off the new GPU toggle to see CPU cards, group offers optionally, and filter by GPU vendor. Each offer shows its hourly price, with price ranges for grouped offers.
The Fleet page has a new Offers tab with the offers matching the fleet, which helps check available capacity, including for CPU-only SSH and Kubernetes fleets. Fleet instances are now shown below the General section.
The Projects tab on a user page now lists the projects the user owns or has joined, with their role in each.
A new post shows how a single 4xH200 node, optimized with dstack presets, serves 16 concurrent Qwen3.8-Flash-Next coding agents with roughly 3x lower latency than the available providers.
Full Changelog: 0.22.4...0.22.5
The server could stop responding in two cases, both fixed in this release. Opening SSH connections, e.g. for service probes, could crash in gRPC and hang the server for up to 5 minutes (#4362). Submitting many runs at once kept the server busy finding offers and unable to handle other requests (#4373); finding offers is now also much faster. In either case, cloud instances being created at the time could be left running without dstack tracking them.
The Launch button on the Runs page now opens a page that shows how to start a run with the CLI or with AI agents using dstack skills.
Templates in the UI are now available only in dstack Enterprise, Factory, and Sky.
The Offers page now shows the best GPUs first, and its new Launch button takes you to the Launch page with the same filters.
In dstack Factory and Sky, project admins can now see the project's quotas in a new Quotas tab and request an increase.
sglang to smg by @un-def in #4371Full Changelog: 0.22.3...0.22.4
dstack can now provision Hot Aisle bare metal servers with 8xMI300X in addition to VMs. Bare metal is disabled by default; enable it with bare_metal: true in the backend settings:
projects: - name: main backends: - type: hotaisle team_handle: hotaisle-team-handle creds: type: api_key api_key: ... bare_metal: true
And the bare metal offers become available:
$ dstack offer --backend hotaisle # BACKEND RESOURCE TYPE PRICE 1 hotaisle (us-michigan-1) cpu=x86:104 mem=2048GB bm-mi300x-8 $27.12 disk=11 gpu=MI300X:192GB:8
Bare metal servers are prepaid for 8 hours. Use a fleet with fixed nodes and delete it when done:
type: fleet name: hotaisle-bm nodes: 1 backends: [hotaisle] instance_types: [bm-mi300x-8]
Within the 8 hours, dstack doesn't force-terminate bare metal servers: the instance stays terminating for about 15 minutes, then dstack stops tracking it and the server must be terminated manually in Hot Aisle. Controlled by DSTACK_FF_HOTAISLE_BARE_METAL_NO_FORCE_RELEASE (enabled by default, 0 disables).
Full Changelog: 0.22.2...0.22.3