Warp Survey: Deployment, AOT & C++ Workflows

[momo-van]

Context

We’re gathering feedback to guide the next phase of Warp’s development across three areas:

• Native and Ahead-of-Time (AOT) workflows,
• Interoperability with other CUDA/C++ and Python libraries, and
• Deployment in minimal- or zero-Python environments.

Warp already supports custom native snippets, AOT workflows, and zero-copy interop with external libraries. We’re now looking to understand where these capabilities need to evolve and which new functionalities are required. Your input will directly shape our roadmap and help ensure Warp meets the needs of real-world systems and pipelines, especially for production and deployment workflows across industries such as robotics and CAE/EDA among others.

How to Participate

We are asking for for feedback on 7 use-cases (UC-1 to UC-7). Each use-case will be posted as a separate top-level comment below. If it applies to you, please react to it with a vote or emoji and reply with any additional workflow details, pain points, or clarifications. These threads are for community discussion and context only, not for submitting feature requests.

If your needs translate into a specific, actionable requirement, please open a GitHub issue and reference this questionnaire along with the relevant use-case number. This helps us track concrete requests separately from the broader discussion. Example: [Survey UC−4] Extend Warp interop: support RMM as an external memory allocator

You can alternatively engage with the survey through the Google form here. These responses will be reviewed confidentially and included only in aggregated summaries, without any identifying or sensitive details.

[momo-van]

UC-01 – As a user, I want to call external C++/CUDA device functions inside Warp kernels so I can reuse existing GPU logic without rewriting it in Python. This includes invoking CUDA/C++ device functions inside Warp kernels, but excludes host code or full native libraries.

[momo-van]

UC-02 – As a user, I want to generate native C++/CUDA sources from Warp kernels to integrate into larger non-Python native applications.

[momo-van]

UC-03 – As a user, I want to compile Warp kernels into PTX/SASS ahead of time for use in native C++ runtimes, or as a method of obfuscation / IP protection.

[momo-van]

UC-04 – As a user, I want Warp to interoperate cleanly with other CUDA, C++, or Python libraries, sharing device memory, CUDA contexts, and streams without copies or re-initialization. Examples could include supporting external memory managers such as RMM among others.

[momo-van]

UC-05 – As a user, I want to use essential Warp runtime functionality from C++ such as creating Warp Meshes, Volumes, BVHs, and other data structures from native code.

[momo-van]

UC‑06 – As a user, I want to deploy and execute Warp programs in a sandboxed or minimal Python environment with tight control over memory, GC, or runtime behavior but still relying on the CPython interpreter.

[momo-van]

UC‑07 – As a user, I want to deploy and execute Warp programs natively, without relying on Python for orchestration or runtime control, but where there are multiple kernel launches and more complex program flow. Examples include real-time embedded deployments in robotics among others.

[momo-van]

If you have a use case related to AOT, C++ interoperability, or deployment workflows that isn’t covered above, please describe it here.

[erikfrey]

Other use case: I'm curious if there's a blessed mechanism for precompiling kernels as part of building a distributable package (like a pip package), to save downstream users from some first time startup latency.

[lgulich]

Other use case: I have a compute graph consisting of a mix of torch operations and warp kernels. I would like to package the whole compute graph as a single ONNX (or something similar) and then deploy it from c++ using an ONNX runtime (or something similar).