Hardware Platforms

Details about supported hardware platforms.

Supported Platforms

Platform	Vendor	Type	Notes
NVIDIA	NVIDIA	GPU	Primary baseline
Ascend NPU	Huawei	NPU	—
MUSA	Moore Threads	GPU	—
Hygon DCU	Hygon	DCU	—
Iluvatar	Iluvatar	AI Chip	—
MetaX	MUXI	GPU	—

Platform Detection

# Check detected device
python -c "from runtime import get_device_type; print(get_device_type())"

Platform-Specific Features

NVIDIA

Feature	Support
All datasets	✓
L3 Anti-hack	✓
vLLM operators	✓
cuBLAS operators	✓

Non-NVIDIA

Feature	Support
ATen dataset only	✓
L1 + L2 Anti-hack	✓
L3 Anti-hack	✗
vLLM operators	✗
cuBLAS operators	✗

Installation Requirements

NVIDIA

pip install -r requirements/requirements_nvidia.txt
pip install -e .

Ascend NPU

pip install -r requirements/requirements_ascend.txt
pip install -e .
# Use vendor container image

MUSA

pip install -r requirements/requirements_musa.txt
pip install -e .
# Use vendor container image

Hygon DCU

pip install -r requirements/requirements_hygon.txt
pip install -e .
# Use vendor container image

Iluvatar

pip install -r requirements/requirements_iluvatar.txt
pip install -e .
# Use vendor container image

MetaX

pip install -r requirements/requirements_metax.txt
pip install -e .
# Use vendor container image

Cross-Platform Considerations

Compiler Maturity

Non-NVIDIA platforms typically have:

• Less mature Triton compilers

• Incomplete backend support

• Different memory models

Performance Impact

• 2× more tokens and time on average

• May require platform-specific optimizations

• Results vary by Operator complexity

Tolerance Settings

Numerical tolerances are automatically adjusted per platform.