dstack 0.19.29__tar.gz → 0.19.30__tar.gz

{dstack-0.19.29 → dstack-0.19.30}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: dstack
-Version: 0.19.29
+Version: 0.19.30
 Summary: dstack is an open-source orchestration engine for running AI workloads on any cloud or on-premises.
 Project-URL: Homepage, https://dstack.ai
 Project-URL: Source, https://github.com/dstackai/dstack
@@ -331,24 +331,26 @@ Description-Content-Type: text/markdown
 
 </div>
 
-`dstack` is an open-source container orchestrator that simplifies workload orchestration and drives GPU utilization for ML teams. It works with any GPU cloud, on-prem cluster, or accelerated hardware.
+`dstack` provides a unified control plane for running development, training, and inference on GPUs — across cloud VMs, Kubernetes, or on-prem clusters. It helps your team avoid vendor lock-in and reduce GPU costs.
 
 #### Accelerators
 
 `dstack` supports `NVIDIA`, `AMD`, `Google TPU`, `Intel Gaudi`, and `Tenstorrent` accelerators out of the box.
 
 ## Latest news ✨
+- [2025/09] [dstack 0.19.27: Offers UI, Digital Ocean and AMD Developer Cloud](https://github.com/dstackai/dstack/releases/tag/0.19.27)
+- [2025/08] [dstack 0.19.26: Repos – explicit repo configuration via YAML](https://github.com/dstackai/dstack/releases/tag/0.19.26)
+- [2025/08] [dstack 0.19.25: `dstack offer` CLI command](https://github.com/dstackai/dstack/releases/tag/0.19.25)
+- [2025/08] [dstack 0.19.22: Service probes, GPU health-checks, Tenstorrent Galaxy, Secrets UI](https://github.com/dstackai/dstack/releases/tag/0.19.22)
+- [2025/07] [dstack 0.19.21: Scheduled tasks](https://github.com/dstackai/dstack/releases/tag/0.19.21)
 - [2025/07] [dstack 0.19.17: Secrets, Files, Rolling deployment](https://github.com/dstackai/dstack/releases/tag/0.19.17)
 - [2025/06] [dstack 0.19.16: Docker in Docker, CloudRift](https://github.com/dstackai/dstack/releases/tag/0.19.16)
 - [2025/06] [dstack 0.19.13: InfiniBand support in default images](https://github.com/dstackai/dstack/releases/tag/0.19.13)
 - [2025/06] [dstack 0.19.12: Simplified use of MPI](https://github.com/dstackai/dstack/releases/tag/0.19.12)
-- [2025/05] [dstack 0.19.10: Priorities](https://github.com/dstackai/dstack/releases/tag/0.19.10)
-- [2025/05] [dstack 0.19.8: Nebius clusters, GH200 on Lambda](https://github.com/dstackai/dstack/releases/tag/0.19.8)
-- [2025/04] [dstack 0.19.6: Tenstorrent, Plugins](https://github.com/dstackai/dstack/releases/tag/0.19.6)
 
 ## How does it work?
 
-<img src="https://dstack.ai/static-assets/static-assets/images/dstack-architecture-diagram-v10.svg" width="750" />
+<img src="https://dstack.ai/static-assets/static-assets/images/dstack-architecture-diagram-v11.svg" width="750" />
 
 ### Installation
 

{dstack-0.19.29 → dstack-0.19.30}/README.md

@@ -14,26 +14,28 @@
 
 </div>
 
-`dstack` is an open-source container orchestrator that simplifies workload orchestration and drives GPU utilization for ML teams. It works with any GPU cloud, on-prem cluster, or accelerated hardware.
+`dstack` provides a unified control plane for running development, training, and inference on GPUs — across cloud VMs, Kubernetes, or on-prem clusters. It helps your team avoid vendor lock-in and reduce GPU costs.
 
 #### Accelerators
 
 `dstack` supports `NVIDIA`, `AMD`, `Google TPU`, `Intel Gaudi`, and `Tenstorrent` accelerators out of the box.
 
 ## Latest news ✨
+- [2025/09] [dstack 0.19.27: Offers UI, Digital Ocean and AMD Developer Cloud](https://github.com/dstackai/dstack/releases/tag/0.19.27)
+- [2025/08] [dstack 0.19.26: Repos – explicit repo configuration via YAML](https://github.com/dstackai/dstack/releases/tag/0.19.26)
+- [2025/08] [dstack 0.19.25: `dstack offer` CLI command](https://github.com/dstackai/dstack/releases/tag/0.19.25)
+- [2025/08] [dstack 0.19.22: Service probes, GPU health-checks, Tenstorrent Galaxy, Secrets UI](https://github.com/dstackai/dstack/releases/tag/0.19.22)
+- [2025/07] [dstack 0.19.21: Scheduled tasks](https://github.com/dstackai/dstack/releases/tag/0.19.21)
 - [2025/07] [dstack 0.19.17: Secrets, Files, Rolling deployment](https://github.com/dstackai/dstack/releases/tag/0.19.17)
 - [2025/06] [dstack 0.19.16: Docker in Docker, CloudRift](https://github.com/dstackai/dstack/releases/tag/0.19.16)
 - [2025/06] [dstack 0.19.13: InfiniBand support in default images](https://github.com/dstackai/dstack/releases/tag/0.19.13)
 - [2025/06] [dstack 0.19.12: Simplified use of MPI](https://github.com/dstackai/dstack/releases/tag/0.19.12)
-- [2025/05] [dstack 0.19.10: Priorities](https://github.com/dstackai/dstack/releases/tag/0.19.10)
-- [2025/05] [dstack 0.19.8: Nebius clusters, GH200 on Lambda](https://github.com/dstackai/dstack/releases/tag/0.19.8)
-- [2025/04] [dstack 0.19.6: Tenstorrent, Plugins](https://github.com/dstackai/dstack/releases/tag/0.19.6)
 
 ## How does it work?
 
 <picture>
-  <source media="(prefers-color-scheme: dark)" srcset="https://dstack.ai/static-assets/static-assets/images/dstack-architecture-diagram-v10-dark.svg"/>
-  <img src="https://dstack.ai/static-assets/static-assets/images/dstack-architecture-diagram-v10.svg" width="750" />
+  <source media="(prefers-color-scheme: dark)" srcset="https://dstack.ai/static-assets/static-assets/images/dstack-architecture-diagram-v11-dark.svg"/>
+  <img src="https://dstack.ai/static-assets/static-assets/images/dstack-architecture-diagram-v11.svg" width="750" />
 </picture>
 
 ### Installation

{dstack-0.19.29 → dstack-0.19.30}/docs/assets/stylesheets/landing.css

@@ -11,7 +11,7 @@
 }
 
 .tx-landing__hero_text h1 {
-    margin-bottom: .75rem;
+    margin-bottom: .6rem;
     font-weight: 700;
     font-size: 2.5em;
     /*letter-spacing: -3px;*/
@@ -239,7 +239,7 @@
 
 .md-header__buttons .md-button,
 .md-typeset .md-button {
-    margin-top: 1.2rem;
+    margin-top: 0.6rem;
     margin-bottom: 1.5rem;
     font-size: 19px;
     font-weight: 300 !important;

dstack-0.19.30/docs/blog/posts/benchmarking-pd-ratios.md

@@ -0,0 +1,141 @@
+---
+title: "Benchmarking Prefill–Decode: fixed 1:3 as a strong default"
+date: 2025-09-25
+description: "TBA"  
+slug: benchmarking-pd-ratios
+image: https://dstack.ai/static-assets/static-assets/images/benchmarking-pd-ratios.png
+categories:
+  - Benchmarks
+---
+
+# Benchmarking Prefill–Decode: fixed 1:3 as a strong default
+
+As demand for low-latency LLM inference grows, squeezing more useful work out of every GPU minute is critical.
+This benchmark evaluates how the Prefill–Decode worker disaggregatioh ratio affects performance across workload profiles and concurrency levels,
+and assess if dynamic ratio adjustment adds value.
+
+<img src="https://dstack.ai/static-assets/static-assets/images/benchmarking-pd-ratios.png" width="630" />
+
+<!-- more -->
+
+## Introduction
+
+### What is Prefill–Decode disaggregation?
+
+DistServe ([Zhong et al., 2024 :material-arrow-top-right-thin:{ .external }](https://arxiv.org/pdf/2401.09670){:target="_blank"}) proposes prefill–decode disaggregation, separating the two phases of inference across dedicated workers.
+Prefill can be heavily batched—prompt tokens are processed in parallel—so it is compute-intensive. Decode is intrinsically sequential—one token per iteration with full KV-cache access—so it is memory- and bandwidth-intensive. Disaggregating these phases reduces cross-phase interference, allowing hardware to be provisioned for the dominant bottleneck and improving end-to-end service performance.
+
+### What is the Prefill–Decode ratio?
+
+The optimal split between prefill and decode workers depends on service-level objectives (SLOs) and workload shape. DistServe shows that for input sequence length (ISL) = 512 and output sequence length (OSL) = 64, "2 prefill to 1 decode" meets both TTFT and TPOT targets. Beyond this illustrative case, however, DistServe does not systematically explore other Prefill–Decode ratios.
+
+!!! info "Reasoning model example"
+    In the DeepSeek deployment ([LMSYS, 2025 :material-arrow-top-right-thin:{ .external }](https://lmsys.org/blog/2025-05-05-large-scale-ep){:target="_blank"}), 3 nodes were allocated to prefill and 9 to decode. The decode-heavy split reflects reasoning workloads, where chains of thought push output lengths high. Allocating more capacity to decode reduces inter-token latency and keeps long responses streaming smoothly.
+
+### Dynamic ratio adjustment
+
+Dynamic allocation adjusts the split between prefill and decode workers at runtime. NVIDIA’s [SLA-based planner :material-arrow-top-right-thin:{ .external }](https://docs.nvidia.com/dynamo/latest/architecture/sla_planner.html){:target="_blank"}
+estimates the workers needed to meet TTFT and ITL targets, while the [Load-based planner :material-arrow-top-right-thin:{ .external }](https://docs.nvidia.com/dynamo/latest/architecture/load_planner.html){:target="_blank"}
+reallocates workers using KV-cache and queue signals. These planners describe how to move capacity between phases, but they do not prescribe a specific Prefill–Decode ratio.
+
+## Benchmark purpose
+
+Prior art points to different “best” ratios depending on workload: DistServe’s 2:1 for short outputs, the SGLang DeepSeek example’s 1:3 for long outputs, and dynamic planners that adapt the split in real time. Building on these insights, this benchmark evaluates how the Prefill–Decode worker ratio affects performance across workload profiles and concurrency levels. 
+
+We measure TTFT, ITL, and throughput to understand how allocation choices influence both latency and efficiency—and to assess when dynamic ratio adjustment adds value versus when a fixed ratio suffices for a known workload.
+
+??? info "Why these metrics matter"
+
+    * **TTFT** (Time to First Token) captures perceived responsiveness—crucial for interactive experiences (e.g., support bots, code assistants).
+    * **ITL** (inter-token latency) captures streaming smoothness—critical for long, reasoning-style outputs.
+    * **Throughput** (tokens/sec) reflects cost efficiency. Prefill-heavy tasks (e.g., summarization of long docs) stress prefill; reasoning tasks stress decode. Maintaining high throughput ensures the under-stressed phase doesn’t leave GPUs idle.
+
+## Methodology
+
+We ran a single-node study on 8xH200 GPUs, varying the number of prefill and decode workers to examine how the split shapes performance. We compared three prefill-decode ratios—3:1, 2:2, 1:3 both lower and higher request concurrency for three workload profiles:
+
+* **Prefill-heavy** (ISL > OSL) — e.g., summarization: long inputs, short outputs.
+* **Decode-heavy** (ISL < OSL) — e.g., reasoning: short inputs, long chains of thought.
+* **Balanced** (ISL ≈ OSL) — e.g., translation, paraphrasing.
+
+Lower concurrency highlights intrinsic trade-offs (prefill-leaning improves TTFT; decode-leaning improves ITL and throughput). Higher concurrency reveals the true bottleneck. In real deployments, success is meeting TTFT/ITL SLOs and sustaining throughput for cost efficiency, so we evaluate both.
+
+> A single-node design isolates the question at hand—does adjusting the prefill/decode split improve performance? If a benefit doesn’t manifest on one node, scaling out will typically amplify the same dynamics rather than change them.
+
+## Benchmark setup
+
+* **GPU**: NVIDIA 8xH200 (SXM5)
+* **CPU**: Intel Xeon Platinum 8468  
+* **Model**: `openai/gpt-oss-120b`  
+* **Backend**: SGLang
+
+For full steps and raw data, see the [GitHub repo :material-arrow-top-right-thin:{ .external }](https://github.com/dstackai/benchmarks/tree/main/comparison/pd_ratio){:target="_blank"}.
+
+## Finding 1: Prefill-heavy workloads
+
+At lower concurrency, 1:3 yields the best ITL and throughput but the worst TTFT. Ratios 3:1 and 2:2 improve TTFT because more prefill capacity clears prompts faster. However, with 3:1, a single decode worker becomes a chokepoint—queues build up, ITL rises, and overall throughput drops.
+
+At higher concurrency, 1:3 wins across all metrics. Because TTFT = prefill time + waiting at decode + time to first token, ample decode capacity trims the waiting component, improving TTFT even on prefill-heavy inputs.
+
+In practice, summarization rarely has tight TTFT SLOs—users expect some delay after uploading long documents. Throughput and ITL dominate cost and experience, making 1:3 the recommended split for prefill-heavy workloads at both low and high concurrency.
+
+*TBA: Fig-1: ISL 2048, OSL 128, concurrency 32*
+
+> Metrics are normalized per chart: the best value for each metric is 100%; others are percentages of that maximum. Lower is better for ITL/TTFT; higher is better for Throughput.
+
+*TBA: Fig-2: ISL 2048, OSL 128, concurrency 128*
+
+## Finding 2: Decode-heavy workloads
+
+As with prefill-heavy cases, at lower concurrency a 1:3 split delivers the best ITL and throughput, at the cost of higher TTFT. Ratios 3:1 and 2:2 improve TTFT but degrade streaming smoothness and throughput.
+
+At higher concurrency, 1:3 again leads across all metrics.
+
+For reasoning tasks, ITL is usually the tightest SLO—smooth, uninterrupted token streaming drives user experience. We recommend 1:3 for decode-heavy workloads at both low and high concurrency.
+
+*TBA: Fig-3: ISL 128, OSL 2048, concurrency 32*
+
+> Metrics normalized as above. Lower is better for ITL/TTFT; higher is better for Throughput.
+
+*TBA: Fig-4: ISL 128, OSL 2048, concurrency 128*
+
+## Finding 3: Balanced workloads
+
+At lower concurrency, 1:1 provides the most balanced profile: better TTFT than the other ratios, with only modest trade-offs in ITL and throughput versus 1:3.
+
+At higher concurrency, 1:3 regains the lead across metrics, while 1:1 sees TTFT degrade as decode pressure grows.
+
+Since 1:1 becomes limiting under load, 1:3 is the safer default for balanced workloads—1:1 can offer slightly lower TTFT at light load, but 1:3 scales better and sustains higher throughput.
+
+*TBA: Fig-5: ISL 2048, OSL 2048, concurrency 32*
+
+> Metrics normalized as above. Lower is better for ITL/TTFT; higher is better for Throughput.
+
+*TBA: Fig-6: ISL 2048, OSL 2048, concurrency 128*
+
+## Conclusion
+
+This study examined how the prefill/decode split shapes performance across workload profiles and load levels, and when dynamic adjustment is beneficial.
+
+1. A decode-leaning default performs robustly. Across profiles and loads, 1:3 consistently offered the strongest ITL and throughput, while keeping TTFT competitive when concurrency rises. For many known workload mixes, this reduces the need for dynamic rebalancing.
+2. Resilience under surges. The 1:3 split scales gracefully with concurrency, absorbing bursts without resorting to complex runtime adjustments.
+3. TTFT in context. 1:3 can show higher TTFT at low concurrency, but real-world expectations matter. Summarization users anticipate a delay after long uploads; reasoning users value smooth streaming most. For interactive chat with tight TTFT SLOs, techniques such as prefix caching and cache-aware routing can reduce prefill work and lower TTFT—often without changing the prefill/decode split.
+
+> Taken together, these results suggest that while dynamic planners (e.g., SLA- and load-based) provide a powerful framework to adapt capacity, in many production scenarios a simple, decode-leaning 1:3 baseline plus conventional autoscaling delivers excellent outcomes with less operational complexity.
+
+## Limitations
+
+This evaluation uses SGLang’s implementation of Prefill–Decode disaggregation. To strengthen generality, repeating the study with vLLM’s implementation would be valuable.
+
+## References
+
+* [DistServe :material-arrow-top-right-thin:{ .external }](https://arxiv.org/pdf/2401.09670){:target="_blank"}
+* [DeepSeek deployment on 96 H100 GPUs :material-arrow-top-right-thin:{ .external }](https://lmsys.org/blog/2025-05-05-large-scale-ep/){:target="_blank"}
+* [Dynamo disaggregated serving :material-arrow-top-right-thin:{ .external }](https://docs.nvidia.com/dynamo/latest/architecture/disagg_serving.html#){:target="_blank"}
+* [SGLang PD disaggregation :material-arrow-top-right-thin:{ .external }](https://docs.sglang.ai/advanced_features/pd_disaggregation.html){:target="_blank"}
+* [vLLM disaggregated prefilling :material-arrow-top-right-thin:{ .external }](https://docs.vllm.ai/en/v0.9.2/features/disagg_prefill.html){:target="_blank"}
+
+!!! info "What's next?"
+
+    * **KV-cache–aware routing & prefix caching with PD**: Quantify how cache-aware routing and prefix caching, combined with PD, reduce redundant prefill compute and improve TTFT.
+    * **PD with model parallelism**: Extend beyond tensor parallelism to assess how additional forms of model parallelism interact with PD and affect latency/throughput trade-offs.

dstack-0.19.30/docs/blog/posts/nebius-in-dstack-sky.md

@@ -0,0 +1,127 @@
+---
+title: Nebius joins dstack Sky GPU marketplace, with production-ready GPU clusters
+date: 2025-09-18
+description: "TBA"  
+slug: nebius-in-dstack-sky
+image: https://dstack.ai/static-assets/static-assets/images/dstack-sky-nebius.png
+categories:
+  - Changelog
+---
+
+# Nebius joins dstack Sky GPU marketplace, with production-ready GPU clusters
+
+`dstack` is an [open-source :material-arrow-top-right-thin:{ .external }](https://github.com/dstackai/dstack){:target="_blank"} control plane for orchestrating GPU workloads. It can provision cloud VMs, run on top of Kubernetes, or manage on-prem clusters. If you don’t want to self-host, you can use [dstack Sky :material-arrow-top-right-thin:{ .external }](https://sky.dstack.ai){:target="_blank"}, the managed version of `dstack` that also provides access to cloud GPUs via its markfetplace.
+
+With our latest release, we’re excited to announce that [Nebius :material-arrow-top-right-thin:{ .external }](https://nebius.com/){:target="_blank"}, a purpose-built AI cloud for large scale training and inference, has joined the `dstack` Sky  marketplace
+to offer on-demand and spot GPUs, including clusters.
+
+<img src="https://dstack.ai/static-assets/static-assets/images/dstack-sky-nebius.png" width="630"/>
+<!-- more -->
+
+Last week we published the [state of cloud GPU](state-of-cloud-gpu-2025.md), a study of the GPU market. As noted there, Nebius is one of the few purpose-built AI clouds delivering performant and resilient GPUs at scale — available on-demand, as spot instances, and as full clusters. 
+
+Nebius designs and operates its own GPU servers in energy-efficient data centers, giving full control over quality, performance tuning, and delivery timelines. Every cluster undergoes a three-stage validation — hardware burn-in, reference architecture checks, and long-haul stress tests — ensuring production-ready infrastructure with consistent performance for large-scale AI training. 
+
+Since early this year, the open-source `dstack` has supported Nebius, making it easy to manage clusters and orchestrate compute cost-effectively.
+
+## About dstack Sky
+
+With this week's release, Nebius officially joins [dstack Sky :material-arrow-top-right-thin:{ .external }](https://sky.dstack.ai){:target="_blank"}. Nebius can now be used not only with your own account, but also directly via the GPU marketplace.
+
+The marketplace lets you access Nebius GPUs without having a Nebius account. You can pay through `dstack Sky`, and switch to your own Nebius account anytime with just a few clicks.
+
+<img src="https://dstack.ai/static-assets/static-assets/images/dstack-sky-nebius-offers.png" width="750"/>
+
+While the open-source version of `dstack` has supported Nebius clusters from day one,  
+Nebius is the first provider to bring on-demand and spot GPU clusters to `dstack` Sky.
+
+With Nebius, `dstack` Sky users can orchestrate NVIDIA GPUs provisioned in hours, with optimized InfiniBand networking to minimize bottlenecks, non-virtualized GPUs for predictable throughput, and industry-leading MTBF/MTTR proven on multi-thousand-GPU clusters.
+
+## Getting started
+
+After you [sign up :material-arrow-top-right-thin:{ .external }](https://sky.dstack.ai){:target="_blank"} with `dstack` Sky, 
+you’ll be prompted to create a project and choose between the GPU marketplace or your own cloud account:
+
+<img src="https://dstack.ai/static-assets/static-assets/images/dstack-sky-project-wizard.png" width="750"/>
+
+Once the project is created, install the `dstack` CLI:
+
+=== "uv"
+
+    <div class="termy">
+    
+    ```shell
+    $ uv tool install dstack -U
+    ```
+
+    </div>
+
+=== "pip"
+
+    <div class="termy">
+    
+    ```shell
+    $ pip install dstack -U
+    ```
+
+    </div>
+
+Now you can define [dev environments](../../docs/concepts/dev-environments.md), 
+[tasks](../../docs/concepts/tasks.md), [services](../../docs/concepts/services.md), 
+and [fleets](../../docs/concepts/fleets.md), then apply them with `dstack apply`.
+
+`dstack` provisions cloud VMs, sets up environments, orchestrates runs, and handles everything required for development, training, or deployment.
+
+To create a Nebius cluster, for example for distributed training, define the following fleet configuration:
+
+<div editor-title="my-cluster.dstack.yml">
+
+```yaml
+type: fleet
+name: my-cluster
+
+placement: cluster
+nodes: 2
+
+backends: [nebius]
+
+resources:
+  gpu: H100:8
+```
+
+</div>
+
+Then, create it via `dstack apply`:
+
+<div class="termy">
+
+```shell
+$ dstack apply -f my-cluster.dstack.yml
+```
+
+</div>
+
+Once the fleet is ready, you can run [distributed tasks](../../docs/concepts/tasks.md#distributed-tasks). 
+`dstack` automatically configures drivers, networking, and fast GPU-to-GPU interconnect.
+
+To learn more, see the [clusters](../../docs/guides/clusters.md) guide.
+
+With Nebius joining `dstack` Sky, users can now run on-demand and spot GPUs and clusters directly through the marketplace—gaining access to the same production grade infrastrucure Nebius customers use for frontier-scale training, without needing a separate Nebius account. 
+
+> If you prefer to go self-hosted, you can always switch to the open-source version of `dstack`, bringing the same functionality.
+
+Our goal is to give teams maximum flexibility while removing the complexity of managing infrastructure. More updates are coming soon.
+
+!!! info "How does `dstack` compare to Kubernetes?"
+    `dstack` can run either on top of Kubernetes or directly on cloud VMs.  
+    In both cases, you don’t need to manage Kubernetes yourself — `dstack` handles container and GPU orchestration, 
+    providing a simple, multi-cloud interface for development, training, and inference.
+
+!!! info "What's next"
+    1. Sign up with [dstack Sky :material-arrow-top-right-thin:{ .external }](https://sky.dstack.ai){:target="_blank"}
+    2. Check [Quickstart](../../docs/quickstart.md)
+    3. Learn more about [Nebius :material-arrow-top-right-thin:{ .external }](https://nebius.com/){:target="_blank"}
+    4. Explore [dev environments](../../docs/concepts/dev-environments.md), 
+        [tasks](../../docs/concepts/tasks.md), [services](../../docs/concepts/services.md), 
+        and [fleets](../../docs/concepts/fleets.md)
+    5. Reaad the the [clusters](../../docs/guides/clusters.md) guide

{dstack-0.19.29 → dstack-0.19.30}/docs/docs/quickstart.md

@@ -14,9 +14,9 @@ $ mkdir quickstart && cd quickstart
 
 </div>
 
-## Run your first workload
+## Submit your first run
 
-`dstack` supports three types of workloads.
+`dstack` supports three types of run configurations.
 
 === "Dev environment"
 

{dstack-0.19.29 → dstack-0.19.30}/docs/overrides/header-2.html

@@ -62,7 +62,7 @@
 
     <div class="md-header__buttons">
       <a href="https://github.com/dstackai/dstack" target="_blank" class="md-button md-button--primary github">GitHub</a>
-      <a href="https://sky.dstack.ai" target="_blank" class="md-button md-button--primary sky external">Sign up</a>
+      <a href="https://sky.dstack.ai" target="_blank" class="md-button md-button--primary sky">dstack Sky</a>
     </div>
   </nav>
   {% if "navigation.tabs.sticky" in features %}

{dstack-0.19.29 → dstack-0.19.30}/docs/overrides/home.html

@@ -53,8 +53,9 @@
                 <h1>The orchestration layer for modern ML teams</h1>
 
                 <p>
-                    <span class="highlight">dstack</span> gives your team a single control plane to run development, training, and inference
-                    jobs on GPU—whether on hyperscalers, neoclouds, or your on-prem hardware. Avoid vendor lock-in and minimize GPU spend.
+                    <span class="highlight">dstack</span> provides a unified control plane for running development, training, and inference
+                    on GPUs — across cloud VMs, Kubernetes, or on-prem clusters. It helps your team avoid vendor lock-in and reduce GPU
+                    costs.
                 </p>
             </div>
 
@@ -74,8 +75,8 @@
                     </a>
                 </div>
             </div>
-            <div class="tx-landing__hero_button_container_text" style="text-align: center; font-size: 85%; margin-top: 25px;">
-                The entire setup takes minutes.
+            <div class="tx-landing__hero_button_container_text" style="text-align: center; font-size: 85%; margin-top: 1.25rem">
+                Open-source self-hosted or fully-managed. Setup takes a few minutes.
             </div>
         </div>
 
@@ -103,7 +104,7 @@
                 </div>
 
                 <div class="block large">
-                    <img src="https://dstack.ai/static-assets/static-assets/images/dstack-architecture-diagram-v10.svg" />
+                    <img src="https://dstack.ai/static-assets/static-assets/images/dstack-architecture-diagram-v11.svg" />
                 </div>
             </div>
         </div>
@@ -163,21 +164,33 @@
                 </div>
 
                 <div class="block margin">
-                    <h2>Run on any cloud. Or all of them.</h2>
+                    <h2>Native integration with GPU clouds</h2>
                     <p>
-                        <span class="highlight">dstack</span> natively provisions GPU instances and clusters across your preferred cloud
-                        providers—maximizing efficiency and minimizing overhead.
+                        <span class="highlight">dstack</span> natively integrates with top GPU clouds—automating cluster provisioning and
+                        workload orchestration to maximize efficiency and minimize overhead.
                     </p>
                     
                     <p>
-                        Connect GPU clouds directly to <span class="highlight">dstack</span> natively, or run <span class="highlight">dstack</span> on top of Kubernetes if needed.
+                        It can provision and manage VM clusters through native integrations or via <span class="highlight">Kuberenetes</span>.
                     </p>
 
                     <p>
-                        <a href="/docs/concepts/backends" target="_blank"
-                            class="md-button md-button-secondary small">
+                        <a href="/docs/concepts/backends" target="_blank" class="md-button md-button-secondary small">
                             <span>Backends</span>
-                            <span class="icon"><svg viewBox="0 0 13 10" xmlns="http://www.w3.org/2000/svg"><path d="M12.823 4.164L8.954.182a.592.592 0 0 0-.854 0 .635.635 0 0 0 0 .88l2.836 2.92H.604A.614.614 0 0 0 0 4.604c0 .344.27.622.604.622h10.332L8.1 8.146a.635.635 0 0 0 0 .88.594.594 0 0 0 .854 0l3.869-3.982a.635.635 0 0 0 0-.88z" fill-rule="nonzero" fill="currentColor" class="fill-main"></path></svg></span>
+                            <span class="icon"><svg viewBox="0 0 13 10" xmlns="http://www.w3.org/2000/svg">
+                                    <path
+                                        d="M12.823 4.164L8.954.182a.592.592 0 0 0-.854 0 .635.635 0 0 0 0 .88l2.836 2.92H.604A.614.614 0 0 0 0 4.604c0 .344.27.622.604.622h10.332L8.1 8.146a.635.635 0 0 0 0 .88.594.594 0 0 0 .854 0l3.869-3.982a.635.635 0 0 0 0-.88z"
+                                        fill-rule="nonzero" fill="currentColor" class="fill-main"></path>
+                                </svg></span>
+                        </a>
+
+                        <a href="/docs/concepts/backends#kubernetes" target="_blank" class="md-button md-button-secondary small">
+                            <span>Kubernetes</span>
+                            <span class="icon"><svg viewBox="0 0 13 10" xmlns="http://www.w3.org/2000/svg">
+                                    <path
+                                        d="M12.823 4.164L8.954.182a.592.592 0 0 0-.854 0 .635.635 0 0 0 0 .88l2.836 2.92H.604A.614.614 0 0 0 0 4.604c0 .344.27.622.604.622h10.332L8.1 8.146a.635.635 0 0 0 0 .88.594.594 0 0 0 .854 0l3.869-3.982a.635.635 0 0 0 0-.88z"
+                                        fill-rule="nonzero" fill="currentColor" class="fill-main"></path>
+                                </svg></span>
                         </a>
                     </p>
                 </div>
@@ -187,25 +200,38 @@
         <div class="tx-landing__major_feature">
             <div class="section">
                 <div class="block margin right">
-                    <h2>Bring your own compute</h2>
-
+                    <h2>Easy to use with on-prem clusters</h2>
+                    
                     <p>
-                        Have bare-metal servers or on-prem GPU clusters? <span class="highlight">dstack</span> makes it easy to integrate them
-                        and manage everything alongside your cloud resources.
+                        For provisioned <span class="highlight">Kuberenetes</span> clusters, connect them to <span class="highlight">dstack</span> using the <span
+                            class="highlight">Kubernetes</span> backend.
+                        If you run vanilla bare-metal servers or VMs without <span class="highlight">Kuberenetes</span>, use <span class="highlight">SSH fleets</span>
+                        instead.
                     </p>
                     
-                    <p>With SSH fleets, you can connect any existing cluster in minutes. Once added to <span
-                            class="highlight">dstack</span>, it's a first-class resource &mdash; available for dev environments, tasks, and
-                        services.
-                    </p>
-
                     <p>
-                        <a href="/docs/concepts/fleets#ssh" target="_blank"
-                            class="md-button md-button-secondary small">
+                        <a href="/docs/concepts/backends#kubernetes" target="_blank" class="md-button md-button-secondary small">
+                            <span>Kubernetes</span>
+                            <span class="icon"><svg viewBox="0 0 13 10" xmlns="http://www.w3.org/2000/svg">
+                                    <path
+                                        d="M12.823 4.164L8.954.182a.592.592 0 0 0-.854 0 .635.635 0 0 0 0 .88l2.836 2.92H.604A.614.614 0 0 0 0 4.604c0 .344.27.622.604.622h10.332L8.1 8.146a.635.635 0 0 0 0 .88.594.594 0 0 0 .854 0l3.869-3.982a.635.635 0 0 0 0-.88z"
+                                        fill-rule="nonzero" fill="currentColor" class="fill-main"></path>
+                                </svg></span>
+                        </a>
+                    
+                        <a href="/docs/concepts/fleets#ssh" target="_blank" class="md-button md-button-secondary small">
                             <span>SSH fleets</span>
-                            <span class="icon"><svg viewBox="0 0 13 10" xmlns="http://www.w3.org/2000/svg"><path d="M12.823 4.164L8.954.182a.592.592 0 0 0-.854 0 .635.635 0 0 0 0 .88l2.836 2.92H.604A.614.614 0 0 0 0 4.604c0 .344.27.622.604.622h10.332L8.1 8.146a.635.635 0 0 0 0 .88.594.594 0 0 0 .854 0l3.869-3.982a.635.635 0 0 0 0-.88z" fill-rule="nonzero" fill="currentColor" class="fill-main"></path></svg></span>
+                            <span class="icon"><svg viewBox="0 0 13 10" xmlns="http://www.w3.org/2000/svg">
+                                    <path
+                                        d="M12.823 4.164L8.954.182a.592.592 0 0 0-.854 0 .635.635 0 0 0 0 .88l2.836 2.92H.604A.614.614 0 0 0 0 4.604c0 .344.27.622.604.622h10.332L8.1 8.146a.635.635 0 0 0 0 .88.594.594 0 0 0 .854 0l3.869-3.982a.635.635 0 0 0 0-.88z"
+                                        fill-rule="nonzero" fill="currentColor" class="fill-main"></path>
+                                </svg></span>
                         </a>
                     </p>
+                    
+                    <p>
+                        Either way, connecting existing on-prem clusters to <span class="highlight">dstack</span> takes just minutes.
+                    </p>
                 </div>
 
                 <div class="block large">
@@ -224,7 +250,7 @@
                 </div>
 
                 <div class="block margin left">
-                    <h2>Code locally, run remotely</h2>
+                    <h2>Dev environments</h2>
                     <p>
                         Before training or deployment, ML engineers explore and debug their code.
                     </p>
@@ -247,7 +273,7 @@
         <div class="tx-landing__major_feature">
             <div class="section">
                 <div class="block margin right">
-                    <h2>Run complex training with simple config</h2>
+                    <h2>Single-node &amp; distributed tasks</h2>
 
                     <p>
                         Move from single-instance experiments to multi-node distributed training without friction. <span
@@ -263,6 +289,12 @@
                             <span>Tasks</span>
                             <span class="icon"><svg viewBox="0 0 13 10" xmlns="http://www.w3.org/2000/svg"><path d="M12.823 4.164L8.954.182a.592.592 0 0 0-.854 0 .635.635 0 0 0 0 .88l2.836 2.92H.604A.614.614 0 0 0 0 4.604c0 .344.27.622.604.622h10.332L8.1 8.146a.635.635 0 0 0 0 .88.594.594 0 0 0 .854 0l3.869-3.982a.635.635 0 0 0 0-.88z" fill-rule="nonzero" fill="currentColor" class="fill-main"></path></svg></span>
                         </a>
+
+                        <a href="/docs/guides/clusters" target="_blank"
+                            class="md-button md-button-secondary small">
+                            <span>Clusters</span>
+                            <span class="icon"><svg viewBox="0 0 13 10" xmlns="http://www.w3.org/2000/svg"><path d="M12.823 4.164L8.954.182a.592.592 0 0 0-.854 0 .635.635 0 0 0 0 .88l2.836 2.92H.604A.614.614 0 0 0 0 4.604c0 .344.27.622.604.622h10.332L8.1 8.146a.635.635 0 0 0 0 .88.594.594 0 0 0 .854 0l3.869-3.982a.635.635 0 0 0 0-.88z" fill-rule="nonzero" fill="currentColor" class="fill-main"></path></svg></span>
+                        </a>
                     </p>
                 </div>
 
@@ -281,7 +313,7 @@
                 </div>
 
                 <div class="block margin left">
-                    <h2>Deploy scalable, OpenAI-compatible endpoints</h2>
+                    <h2>Scalable service endpoints</h2>
 
                     <p>
                         With <span class="highlight">dstack</span>, you can easily deploy any model as a secure,
@@ -518,17 +550,17 @@
                     <p>Set up backends or SSH fleets, then add your team.</p>
 
                     <p>
-                        <a href="/docs/installation" target="_blank"
-                           class="md-button md-button--primary small"
+                        <a href="/docs/quickstart" target="_blank"
+                           class="md-button md-button-secondary small"
                             style="margin-right: 10px">
-                            <span>Installation</span>
+                            <span>Quickstart</span>
                             <span class="icon"><svg viewBox="0 0 13 10" xmlns="http://www.w3.org/2000/svg"><path d="M12.823 4.164L8.954.182a.592.592 0 0 0-.854 0 .635.635 0 0 0 0 .88l2.836 2.92H.604A.614.614 0 0 0 0 4.604c0 .344.27.622.604.622h10.332L8.1 8.146a.635.635 0 0 0 0 .88.594.594 0 0 0 .854 0l3.869-3.982a.635.635 0 0 0 0-.88z" fill-rule="nonzero" fill="currentColor" class="fill-main"></path></svg></span>
                         </a>
 
-                        <a href="/docs/quickstart" target="_blank" class="md-button md-button-secondary small"
+                        <!-- <a href="/docs/quickstart" target="_blank" class="md-button md-button-secondary small"
                            style="">
                             Quickstart
-                        </a>
+                        </a> -->
                     </p>
                 </div>
 
@@ -559,7 +591,7 @@
                         </div>
 
                         <div class="tx-landing__bottom_cta_card_button_subheader">
-                            Already have an account? <a href="https://sky.dstack.ai" target="_blank"
+                            Get $5 free credit. Already have an account? <a href="https://sky.dstack.ai" target="_blank"
                                 class="tx-footer__section-link external">Sign in</a>
                         </div>
                     </div>
@@ -578,7 +610,7 @@
                         <div>
                             <a href="https://calendly.com/dstackai/discovery-call" target="_blank"
                                 class="md-button md-button--primary external small">
-                                <span>Get a demo</span>
+                                <span>Book a demo</span>
                             </a>
                         </div>