I met an error when run the program of Existing Libraries in the Environment in Chapter 4. Here is my code:
'''
Integrate Existing Libraries in th Environment
'''
print(tvm.target.Target.list_kinds())
# Register Runtime Function
@tvm.register_func("env.linear", override=True)
def torch_linear(x: tvm.nd.NDArray,
w: tvm.nd.NDArray,
b: tvm.nd.NDArray,
out: tvm.nd.NDArray):
# from_dlpack is a zero-copy conversion
x_torch = torch.from_dlpack(x)
w_torch = torch.from_dlpack(w)
b_torch = torch.from_dlpack(b)
out_torch = torch.from_dlpack(out)
torch.mm(x_torch, w_torch.T, out=out_torch)
torch.add(out_torch, b_torch, out=out_torch)
@tvm.register_func("env.relu", override=True)
def lnumpy_relu(x: tvm.nd.NDArray,
out: tvm.nd.NDArray):
x_torch = torch.from_dlpack(x)
out_torch = torch.from_dlpack(out)
torch.maximum(x_torch, torch.Tensor([0.0]), out=out_torch)
@tvm.script.ir_module
class MyModuleWithExternCall:
@R.function
def main(x: R.Tensor((1, 784), "float32"), # type: ignore
w0: R.Tensor((128, 784), "float32"), # type: ignore
b0: R.Tensor((128, ), "float32"), # type: ignore
w1: R.Tensor((10, 128), "float32"), # type: ignore
b1: R.Tensor((10, ), "float32")): # type: ignore
# block 0
with R.dataflow():
lv0 = R.call_dps_packed("env.linear", (x, w0, b0), out_sinfo=R.Tensor((1, 128), dtype="float32"))
lv1 = R.call_dps_packed("env.relu", (lv0, ), out_sinfo=R.Tensor((1, 128), dtype="float32"))
out = R.call_dps_packed("env.linear", (lv1, w1, b1), out_sinfo=R.Tensor((1, 10), dtype="float32"))
R.output(out)
return out
cuda_device = tvm.device("cuda", 0)
if cuda_device.exist:
print("CUDA device is available")
else:
print("CUDA device is not available")
ex = relax.build(MyModuleWithExternCall, target="llvm")
try:
vm = relax.VirtualMachine(ex, tvm.cpu())
except Exception as e:
print("Error:", e)
print(f"vm is None? {vm is None}")
nd_res = vm["main"](data_nd,
nd_params["w0"],
nd_params["b0"],
nd_params["w1"],
nd_params["b1"])
pred_kind = np.argmax(nd_res.numpy(), axis=1)
print("MyModuleWithExternCall Prediction:", class_names[pred_kind[0]])
and the output of python debug console is here:
['llvm', 'c', 'cuda', 'nvptx', 'rocm', 'metal', 'opencl', 'vulkan', 'webgpu', 'sdaccel', 'aocl', 'aocl_sw_emu', 'hexagon', 'stackvm', 'ext_dev', 'hybrid', 'composite', 'test', 'ccompiler', 'example_target_hook']
CUDA device is available
Error: Target triple should not be empty
vm is None? False
MyModuleWithExternCall Prediction: Coat
It’s wierd that try-except
caught an error but the program can execute, if I don’t use the try-except
block, the program will not execute to the end.
Additionaly, other parts of the Chapter which use the relax.VirtualMachine
don’t cause any error
My tvm, llvm and pytorch version are here
(tvm-build) D:\Work\tvm\tvm0.18\tvm\python>python -c "import tvm; print(tvm.__version__)"
0.18.dev0
(tvm-build) D:\Work\tvm\tvm0.18\tvm\python>llvm-config --version
14.0.6
(tvm-build) D:\Work\tvm\tvm0.18\tvm\python>conda list | findstr torch
pytorch 2.3.1 py3.9_cuda11.8_cudnn8_0 pytorch
pytorch-cuda 11.8 h24eeafa_5 pytorch
pytorch-mutex 1.0 cuda pytorch
torchaudio 2.3.1 pypi_0 pypi
torchvision 0.18.1 pypi_0 pypi
Anyone can explain or help me, tks!