torchada 0.1.7

Adapter package for torch_musa to act exactly like PyTorch CUDA

---

torchada

Adapter package for torch_musa to act exactly like PyTorch CUDA

torchada provides a unified interface that works transparently on both NVIDIA GPUs (CUDA) and Moore Threads GPUs (MUSA). Write your code once using standard PyTorch CUDA APIs, and it will run on MUSA hardware without any changes.

Features

• Zero Code Changes: Just import torchada once, then use standard torch.cuda.* APIs
• Automatic Platform Detection: Detects whether you're running on CUDA or MUSA
• Transparent Device Mapping: tensor.cuda() and tensor.to("cuda") work on MUSA
• Extension Building: Standard torch.utils.cpp_extension works on MUSA after importing torchada
• Source Code Porting: Automatic CUDA → MUSA symbol mapping for C++/CUDA extensions
• Distributed Training: torch.distributed with 'nccl' backend automatically uses 'mccl' on MUSA
• Mixed Precision: torch.cuda.amp autocast and GradScaler work transparently
• CUDA Graphs: torch.cuda.CUDAGraph maps to MUSAGraph on MUSA
• Inductor Support: torch._inductor autotune uses MUSA_VISIBLE_DEVICES on MUSA

Installation

pip install torchada

# Or install from source
git clone https://github.com/yeahdongcn/torchada.git
cd torchada
pip install -e .

Quick Start

Basic Usage

import torchada  # Import once to apply patches - that's it!
import torch

# Platform detection (sglang-style)
def _is_cuda():
    return torch.version.cuda is not None

def _is_musa():
    return hasattr(torch.version, 'musa') and torch.version.musa is not None

# Check for GPU availability (works on both CUDA and MUSA)
if _is_cuda() or _is_musa():
    # Use standard torch.cuda APIs - they work transparently on MUSA:
    device = torch.device("cuda")  # Creates musa device on MUSA platform
    tensor = torch.randn(10, 10).cuda()  # Moves to MUSA on MUSA platform
    model = MyModel().cuda()

    # All torch.cuda.* APIs work transparently
    print(f"Device count: {torch.cuda.device_count()}")
    print(f"Device name: {torch.cuda.get_device_name()}")
    torch.cuda.synchronize()

Building C++ Extensions

# setup.py - Use standard torch imports!
import torchada  # Import first to apply patches
from setuptools import setup
from torch.utils.cpp_extension import CUDAExtension, BuildExtension, CUDA_HOME

print(f"Building with CUDA/MUSA home: {CUDA_HOME}")

ext_modules = [
    CUDAExtension(
        name="my_extension",
        sources=[
            "my_extension.cpp",
            "my_extension_kernel.cu",
        ],
        extra_compile_args={
            "cxx": ["-O3"],
            "nvcc": ["-O3"],  # Automatically mapped to mcc on MUSA
        },
    ),
]

setup(
    name="my_package",
    ext_modules=ext_modules,
    cmdclass={"build_ext": BuildExtension.with_options(use_ninja=True)},
)

JIT Compilation

import torchada  # Import first to apply patches
from torch.utils.cpp_extension import load

# Load extension at runtime (works on both CUDA and MUSA)
my_extension = load(
    name="my_extension",
    sources=["my_extension.cpp", "my_extension_kernel.cu"],
    verbose=True,
)

Mixed Precision Training

import torchada  # Import first to apply patches
import torch

model = MyModel().cuda()
optimizer = torch.optim.Adam(model.parameters())
scaler = torch.cuda.amp.GradScaler()

for data, target in dataloader:
    data, target = data.cuda(), target.cuda()

    with torch.cuda.amp.autocast():
        output = model(data)
        loss = criterion(output, target)

    scaler.scale(loss).backward()
    scaler.step(optimizer)
    scaler.update()
    optimizer.zero_grad()

Distributed Training

import torchada  # Import first to apply patches
import torch.distributed as dist

# Use 'nccl' backend as usual - torchada maps it to 'mccl' on MUSA
dist.init_process_group(backend='nccl')

CUDA Graphs

import torchada  # Import first to apply patches
import torch

# Use standard torch.cuda.CUDAGraph - works on MUSA too
g = torch.cuda.CUDAGraph()
with torch.cuda.graph(g):
    y = model(x)

Platform Detection

torchada automatically detects the platform:

import torchada
from torchada import detect_platform, Platform

platform = detect_platform()
if platform == Platform.MUSA:
    print("Running on Moore Threads GPU")
elif platform == Platform.CUDA:
    print("Running on NVIDIA GPU")

# Or use torch.version-based detection (sglang-style)
def _is_musa():
    return hasattr(torch.version, 'musa') and torch.version.musa is not None

if _is_musa():
    print("MUSA platform detected")

What Gets Patched

After import torchada, the following standard PyTorch APIs work on MUSA:

Standard API	Description
torch.cuda.*	All APIs redirected to torch.musa
torch.cuda.amp.*	autocast, GradScaler
torch.cuda.CUDAGraph	Maps to MUSAGraph
torch.cuda.nccl	Maps to torch.musa.mccl
torch.cuda.nvtx	No-op stub (MUSA doesn't have nvtx)
torch.cuda._lazy_call	Patched for lazy initialization
torch.distributed (backend='nccl')	Automatically uses MCCL
torch.device("cuda")	Creates MUSA device on MUSA platform
tensor.cuda()	Moves to MUSA device
tensor.is_cuda	Returns True for MUSA tensors
model.cuda()	Moves model to MUSA device
torch.amp.autocast(device_type='cuda')	Uses 'musa' device type
torch.utils.cpp_extension.*	CUDAExtension, BuildExtension, CUDA_HOME
torch._inductor.autotune_process	Uses MUSA_VISIBLE_DEVICES

API Reference

torchada

Function	Description
detect_platform()	Returns the detected platform (CUDA, MUSA, or CPU)
is_musa_platform()	Check if running on MUSA
is_cuda_platform()	Check if running on CUDA
is_cpu_platform()	Check if running on CPU only
get_device_name()	Get device name string ("cuda", "musa", or "cpu")
get_platform()	Alias for detect_platform()
get_backend()	Get the underlying torch device module
is_patched()	Check if patches have been applied
get_version()	Get torchada version string
CUDA_HOME	Path to CUDA/MUSA installation

torch.cuda (after importing torchada)

All standard torch.cuda APIs work, including:

• device_count(), current_device(), set_device(), get_device_name()
• memory_allocated(), memory_reserved(), empty_cache(), reset_peak_memory_stats()
• synchronize(), Stream, Event, CUDAGraph
• amp.autocast(), amp.GradScaler()
• _lazy_call(), _lazy_init()

Note: torch.cuda.is_available() is intentionally NOT redirected. It returns False on MUSA to allow proper platform detection. Use hasattr(torch.version, 'musa') and torch.version.musa is not None or torch.musa.is_available() instead.

torch.utils.cpp_extension (after importing torchada)

Symbol	Description
CUDAExtension	Creates CUDA or MUSA extension based on platform
CppExtension	Creates C++ extension (no GPU code)
BuildExtension	Build command for extensions
CUDA_HOME	Path to CUDA/MUSA installation
load()	JIT compile and load extension

Symbol Mapping

torchada automatically maps CUDA symbols to MUSA equivalents when building extensions:

CUDA	MUSA
cudaMalloc	musaMalloc
cudaMemcpy	musaMemcpy
cudaStream_t	musaStream_t
cublasHandle_t	mublasHandle_t
curandState	murandState
at::cuda	at::musa
c10::cuda	c10::musa
cutlass::*	mutlass::*
#include <cuda/*>	#include <musa/*>
...	...

See src/torchada/_mapping.py for the complete mapping table (380+ mappings).

Note: Many CUDA constructs like atomic operations (atomicAdd, atomicCAS), shuffle intrinsics (__shfl_sync), and half-precision math (__float2half) are identical in MUSA and don't require mapping.

Architecture

torchada uses a decorator-based patch registration system:

from torchada._patch import patch_function, requires_import

@patch_function  # Registers for automatic application
@requires_import('torch_musa')  # Guards against missing dependencies
def _patch_my_feature():
    # Patching logic here
    pass

All registered patches are applied automatically when you import torchada.

License

MIT License