tinygrad supports various runtimes, enabling your code to scale across a wide range of devices. The default runtime can be automatically selected based on the available hardware, or you can force a specific runtime to be default using environment variables (e.g., DEV=CPU).
| Runtime | Description | Compiler Options | Requirements |
|---|---|---|---|
| NV | Provides acceleration for NVIDIA GPUs | nvrtc (default) PTX ( DEV=NV:PTX) |
Ampere/Ada/Blackwell series GPUs. You can select an interface via the DEV variable. See NV interfaces for details. |
| AMD | Provides acceleration for AMD GPUs | LLVM (DEV=AMD:LLVM)HIP/COMGR ( DEV=AMD:HIP) |
RDNA2 or newer GPUs. You can select an interface via the DEV variable. See AMD interfaces for details. |
| QCOM | Provides acceleration for QCOM GPUs | - | 6xx series GPUs |
| METAL | Utilizes Metal for acceleration on Apple devices | - | M1+ Macs; Metal 3.0+ for bfloat support |
| CUDA | Utilizes CUDA for acceleration on NVIDIA GPUs | nvrtc (default) PTX ( DEV=CUDA:PTX) |
NVIDIA GPU with CUDA support |
| CL | Accelerates computations using OpenCL on GPUs | - | OpenCL 2.0 compatible device |
| CPU | Runs on CPU using the clang or llvm compiler | Clang JIT (default) LLVM IR ( DEV=CPU:LLVM) |
clang compiler in system PATH |
| WEBGPU | Runs on GPU using the Dawn WebGPU engine (used in Google Chrome) | - | Dawn library installed and discoverable. Binaries: pydawn v0.3.0 |
tinygrad provides interoperability with OpenCL and PyTorch, allowing efficient tensor data sharing between frameworks through the Tensor.from_blob API. This enables zero-copy operations by working directly with external memory pointers.
Important: When using external memory pointers with tinygrad tensors, you must ensure these pointers remain valid throughout the entire lifetime of the tinygrad tensor to prevent memory corruption.
You can seamlessly work with CUDA/MPS tensors between PyTorch and tinygrad without data copying:
from tinygrad.dtype import _from_torch_dtype
tensor1 = torch.tensor([1.0, 2.0, 3.0], device=torch.device("cuda"))
tiny_tensor1 = Tensor.from_blob(tensor1.data_ptr(), tensor1.shape, dtype=_from_torch_dtype(tensor1.dtype), device='CUDA')
# Before tinygrad calculations, mps needs to be synchronized to make sure data is valid.
if data.device.type == "mps": torch.mps.synchronize()
else: torch.cuda.synchronize()
x = (tiny_tensor1 + 1).realize()tinygrad supports OpenCL interoperability on QCOM backend.
Buffer interop allows direct access to OpenCL memory buffers:
# create raw opencl buffer.
cl_buf = cl.clCreateBuffer(cl_context, cl.CL_MEM_READ_WRITE, 0x100, None, status := ctypes.c_int32())
# extract pointers
cl_buf_desc_ptr = to_mv(ctypes.addressof(cl_buf), 8).cast('Q')[0]
rawbuf_ptr = to_mv(cl_buf_desc_ptr, 0x100).cast('Q')[20] # offset 0xA0 is a raw gpu pointer.
# create tiny tensor
tiny = Tensor.from_blob(rawbuf_ptr, (8, 8), dtype=dtypes.int, device='QCOM')And the same for the images:
# create cl image.
cl_img = cl.clCreateImage2D(cl_context, cl.CL_MEM_READ_WRITE, cl.cl_image_format(cl.CL_RGBA, cl.CL_FLOAT), w, h, 0, None, status := ctypes.c_int32())
# extract pointers
cl_buf_desc_ptr = to_mv(ctypes.addressof(cl_img), 8).cast('Q')[0]
rawbuf_ptr = to_mv(cl_buf_desc_ptr, 0x100).cast('Q')[20] # offset 0xA0 is a raw gpu pointer.
# create tiny tensor
tiny = Tensor.from_blob(rawbuf_ptr, (h*w*4,), dtype=dtypes.imagef((h,w)), device='QCOM')AMD backend supports several interfaces for communicating with devices:
KFD: uses the amdgpu driverPCI: uses the AM driverUSB: USB3 interface for asm24xx chips.
You can force an interface by setting the interface component of the DEV environment variable to one of these values. When set to PCI, this may unbind your GPU from the amdgpu driver.
NV backend supports several interfaces for communicating with devices:
NVK: uses the nvidia driverPCI: uses the NV driver