Pratheesh-04-MCW opened a new issue, #17959:
URL: https://github.com/apache/tvm/issues/17959
When attempting to convert a PyTorch exported program to TVM's Relax format
using from_exported_program(), the process fails with an AssertionError
indicating that the function type 'frac.default' is not supported.
### Environment
- tvm 0.21.dev26+g2ca6ec8a5
- torch 2.6.0
### Error Message
warnings.warn("Can't initialize NVML")
Traceback (most recent call last):
File "/home/pratheesh04/tvm/examples/relax_support/frac.py", line 16, in
<module>
relax_mod = from_exported_program(exported_program)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File
"/home/pratheesh04/tvm/python/tvm/relax/frontend/torch/exported_program_translator.py",
line 679, in from_exported_program
return ExportedProgramImporter().from_exported_program(
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File
"/home/pratheesh04/tvm/python/tvm/relax/frontend/torch/exported_program_translator.py",
line 544, in from_exported_program
self._check_unsupported_func_type(nodes)
File
"/home/pratheesh04/tvm/python/tvm/relax/frontend/torch/base_fx_graph_translator.py",
line 114, in _check_unsupported_func_type
assert not missing_func_types, f"Unsupported function types
{missing_func_types}"
^^^^^^^^^^^^^^^^^^^^^^
AssertionError: Unsupported function types ['frac.default']
### Expected behavior
Should generate the relax IR:
# from tvm.script import ir as I
# from tvm.script import relax as R
@I.ir_module
class Module:
@R.function
def main(x: R.Tensor((4,), dtype="float32")) -> R.Tuple(R.Tensor((4,),
dtype="float32")):
with R.dataflow():
lv: R.Tensor((4,), dtype="float32") = R.frac(x)
gv: R.Tuple(R.Tensor((4,), dtype="float32")) = (lv,)
R.output(gv)
return gv
### Steps to reproduce
import torch
from torch.export import export
from tvm.relax.frontend.torch import from_exported_program
import tvm
from tvm import relax
class FracModel(torch.nn.Module):
def forward(self, x):
return torch.frac(x)
# Create input tensor
a = torch.tensor([3.4742, 0.5466, -0.8008, -0.9079], dtype=torch.float32)
model = FracModel()
exported_program = export(model, (a,))
relax_mod = from_exported_program(exported_program)
print("Original Relax IR Module:")
print(relax_mod)
target = tvm.target.Target("llvm")
device = tvm.device(target.kind.name)
# Build the VM executable
with tvm.transform.PassContext(opt_level=3):
ex = relax.build(relax_mod, target)
# Create and run the VM
vm = tvm.runtime.relax_vm.VirtualMachine(ex, device)
# Prepare input (convert PyTorch tensor to TVM NDArray)
input_data = tvm.nd.array(a.numpy(), device=device)
# Run the VM
result = vm["main"](input_data)
# Extract the tensor from the tuple
result_tensor = result[0] # The tuple contains a single tensor
# Show results
print("\nInput tensor:")
print(input_data.numpy())
print("\nOutput tensor (fractional parts):")
print(result_tensor.numpy())
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
