cuda-morph 0.9.1

CUDA-to-Ascend ecosystem compatibility bridge — fixes the last mile between torch_npu and your existing CUDA code

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

1. Start with existing CUDA-oriented code.
2. Activate cuda-morph.
3. 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 on cuda:0 and reported AMD 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

• STRATEGY.md
• VALIDATION_STATUS.md
• MIGRATION.md
• docs/README_zh.md

License

Apache 2.0