cuda-morph
Run PyTorch workloads on non-NVIDIA hardware with minimal code changes.
Existing CUDA-style PyTorch scripts keep running — cuda-morph redirects calls to available backends or falls back to CPU.
Why this exists
A lot of ML code assumes torch.cuda everywhere. On non-NVIDIA setups, that often means immediate runtime errors and expensive rewrites.
cuda-morph is a runtime compatibility layer that helps you keep the same workflow and ship faster.
Visual walkthrough
- Start with existing CUDA-oriented code.
- Activate
cuda-morph. - Run with backend-aware routing and fallback behavior.
Quick start
pip install cuda-morph
import ascend_compat
ascend_compat.activate()
# Existing CUDA-style code stays the same
model = model.cuda()
cuda-morph info
cuda-morph check model.py
Core capabilities
- Zero-rewrite activation for many CUDA-style PyTorch flows.
- Backend routing across Ascend / ROCm / Intel XPU / CPU fallback paths.
- CLI tooling for environment checks, porting hints, and validation commands.
- Ecosystem patches for key libraries (current depth varies by backend).
Full matrix: docs/compatibility_matrix.md
Validation snapshot
- 460+ tests passing in CPU-fallback mode.
- Real hardware proof captured on RunPod AMD ROCm:
.cuda()matmul ran oncuda:0and reportedAMD Radeon Graphics. - Core
.cuda()routing on non-NVIDIA hardware is now validated; broader operator and ecosystem coverage is still in progress.
If you can test on non-NVIDIA accelerators, feedback is highly valuable: open an issue.
Backend status (current)
| Backend | Hardware | Status |
|---|---|---|
| Huawei Ascend | 910B, 310P | Full shim + ecosystem patches (flash-attn, HuggingFace, DeepSpeed, vLLM). Needs hardware validation. |
| AMD ROCm | MI210, MI250X, MI300X | Detection + device routing. Ecosystem patches not yet implemented. |
| Intel XPU | Max 1550, Flex, Arc | Detection + device routing. Ecosystem patches not yet implemented. |
| Cambricon | MLU370, MLU590 | Detection + device routing. Ecosystem patches not yet implemented. |
CLI
cuda-morph check model.py # check a script for compatibility issues
cuda-morph port model.py # generate porting suggestions
cuda-morph doctor # environment diagnostics
cuda-morph doctor --full
cuda-morph verify --device npu
cuda-morph bench overhead
cuda-morph info
Development
git clone https://github.com/JosephAhn23/cuda-morph.git
cd cuda-morph
pip install -e ".[dev]"
pytest tests/ -v
pytest tests/ -v --run-hardware
See also
License
Apache 2.0
Metadata
Release files for cuda-morph 0.9.1
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| cuda_morph-0.9.1.tar.gz | 151.8 kB | With provenance | Details for cuda_morph-0.9.1.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| cuda_morph-0.9.1-py3-none-any.whl | Python 3 | none | any | Details for cuda_morph-0.9.1-py3-none-any.whl |
Total release size: 295.0 kB
Release files / cuda_morph-0.9.1.tar.gz
| Download URL | cuda_morph-0.9.1.tar.gz |
|---|---|
| Size | 151.8 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
d04d6894e4e63051dffe9df827f623505492f5d4da23279ed05d4a8b77e6cc30
|
|
BLAKE2b-256 checksum How to use checksums |
7329c9823b2d5148df6a57c7dd99791d9a6c4a5ebd990d62a1cf1a3ea8c5767d
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/6.1.0 CPython/3.13.7
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Mar 10, 2026.
Transparency logRelease files / cuda_morph-0.9.1-py3-none-any.whl
| Download URL | cuda_morph-0.9.1-py3-none-any.whl |
|---|---|
| Size | 143.2 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
98f9b099f715c5e737c4c8ae0fa0a8198b7f263a5094d7a0b03412030de27b99
|
|
BLAKE2b-256 checksum How to use checksums |
9f3fc9dc9b915667ca71f082c64dd3f73d6d224c9ac8a9789c9405fef4c1ba6f
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/6.1.0 CPython/3.13.7
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Mar 10, 2026.
Transparency logSecurity for cuda-morph 0.9.1
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
Full provenance
Both files in this release have been uploaded with provenance metadata.
Attestation for cuda_morph-0.9.1.tar.gz
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Mar 10, 2026.
Transparency logA note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.