jikechao opened a new issue, #15008:
URL: https://github.com/apache/tvm/issues/15008
The PyTorch model with `hardswish` operator and input_dtype=float64 crashed
when load to relay.
## Actual behavior
```
data types float64 and float32 do not match in BroadcastRel
data types float64 and float32 do not match in BroadcastRel
Traceback (most recent call last):
File "19_crash_hardswish.py", line 19, in <module>
mod, params = relay.frontend.from_pytorch(trace, input_shapes)
File "/workplace/software/tvm/tvm/python/tvm/relay/frontend/pytorch.py",
line 4970, in from_pytorch
outputs = converter.convert_operators(operator_nodes, outputs, ret_name)
File "/workplace/software/tvm/tvm/python/tvm/relay/frontend/pytorch.py",
line 4243, in convert_operators
self.record_output_type(relay_out)
File "/workplace/software/tvm/tvm/python/tvm/relay/frontend/pytorch.py",
line 238, in record_output_type
self.infer_type_with_prelude(output)
File "/workplace/software/tvm/tvm/python/tvm/relay/frontend/pytorch.py",
line 174, in infer_type_with_prelude
body = self.infer_type(val, self.prelude.mod)
File "/workplace/software/tvm/tvm/python/tvm/relay/frontend/pytorch.py",
line 167, in infer_type
new_mod = transform.InferType()(new_mod)
File "/workplace/software/tvm/tvm/python/tvm/ir/transform.py", line 160,
in __call__
return _ffi_transform_api.RunPass(self, mod)
File "/workplace/software/tvm/tvm/python/tvm/_ffi/_ctypes/packed_func.py",
line 238, in __call__
raise get_last_ffi_error()
tvm.error.DiagnosticError: Traceback (most recent call last):
7: TVMFuncCall
6:
tvm::runtime::PackedFuncObj::Extractor<tvm::runtime::PackedFuncSubObj<tvm::runtime::TypedPackedFunc<tvm::IRModule
(tvm::transform::Pass,
tvm::IRModule)>::AssignTypedLambda<tvm::transform::$_6>(tvm::transform::$_6,
std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char>
>)::{lambda(tvm::runtime::TVMArgs const&, tvm::runtime::TVMRetValue*)#1}>
>::Call(tvm::runtime::PackedFuncObj const*, tvm::runtime::TVMArgs,
tvm::runtime::TVMRetValue*)
5: tvm::transform::Pass::operator()(tvm::IRModule) const
4: tvm::transform::Pass::operator()(tvm::IRModule,
tvm::transform::PassContext const&) const
3: tvm::transform::ModulePassNode::operator()(tvm::IRModule,
tvm::transform::PassContext const&) const
2:
tvm::runtime::PackedFuncObj::Extractor<tvm::runtime::PackedFuncSubObj<tvm::runtime::TypedPackedFunc<tvm::IRModule
(tvm::IRModule,
tvm::transform::PassContext)>::AssignTypedLambda<tvm::relay::transform::InferType()::$_2>(tvm::relay::transform::InferType()::$_2)::{lambda(tvm::runtime::TVMArgs
const&, tvm::runtime::TVMRetValue*)#1}> >::Call(tvm::runtime::PackedFuncObj
const*, tvm::runtime::TVMArgs, tvm::runtime::TVMRetValue*)
1: tvm::DiagnosticContext::Render()
0: _ZN3tvm7runtime6detail
File "/workplace/software/tvm/tvm/src/ir/diagnostic.cc", line 131
DiagnosticError: one or more error diagnostics were emitted, please check
diagnostic render for output.
```
What actually happened
### Environment
Any environment details, such as: Operating System, TVM version, etc
### Steps to reproduce
```
import torch
from tvm import relay
import tvm
import numpy as np
from torch.nn import Module
input_data = torch.randn([2, 4, 4], dtype=torch.float64) # crash when
dtype=float64
class hardswish(Module):
def forward(self, *args):
return torch.nn.functional.hardswish(args[0], )
m = hardswish().float().eval()
torch_outputs = m(input_data)
trace = torch.jit.trace(m, input_data)
input_shapes = [('input0', torch.Size([2, 4, 4]))]
mod, params = relay.frontend.from_pytorch(trace, input_shapes)
```
### Environments
* TVM: 0.13.dev0'
* PyTorch: 1.13.1+cu117
### Triage
* needs-triage
* front:pytorch
### Analysis
This bug is related to input_dtype.
If the input_dtype was set as 'float32', TVM can run well.
If the input_dtype was set as 'float64', TVM will crash as before.
cc @echuraev @Hzfengsy @shingjan
--
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]
