jsheng-jian opened a new pull request, #14122:
URL: https://github.com/apache/tvm/pull/14122
Minor fix in pytorch frontend to compile gpt2 model, reproduce script.
```
from transformers import GPT2LMHeadModel
import torch
import tvm
from tvm import relay
inp = torch.ones((1, 128)).to(torch.int64)
input_shapes = [("input_ids", ((1, 128), "int64"))]
model = GPT2LMHeadModel.from_pretrained('gpt2', return_dict=False)
trace_model = torch.jit.trace(model, inp, strict=False)
outputs = trace_model(inp)
mod, params = relay.frontend.from_pytorch(trace_model, input_shapes)
with tvm.transform.PassContext(opt_level=3):
lib = relay.build(mod, target='llvm', params=params)
runtime =
tvm.contrib.graph_executor.GraphModule(lib["default"](tvm.device('cpu', 0)))
runtime.set_input("input_ids", inp.numpy())
runtime.run()
out = runtime.get_output(0).numpy()
print(out)
print('Done...')
```
Before the fix, the error message
```
Traceback (most recent call last):
File "gpt2_compile.py", line 13, in <module>
mod, params = relay.frontend.from_pytorch(trace_model, input_shapes)
File "/home/ubuntu/apache_tvm/tvm/python/tvm/relay/frontend/pytorch.py",
line 4791, in from_pytorch
outputs =
converter.convert_operators(_get_operator_nodes(graph.nodes()), outputs,
ret_name)
File "/home/ubuntu/apache_tvm/tvm/python/tvm/relay/frontend/pytorch.py",
line 4164, in convert_operators
relay_out = relay_op(
File "/home/ubuntu/apache_tvm/tvm/python/tvm/relay/frontend/pytorch.py",
line 841, in full
return self.full_impl(data, fill_value, dtype)
File "/home/ubuntu/apache_tvm/tvm/python/tvm/relay/frontend/pytorch.py",
line 743, in full_impl
fill_value = _expr.const(fill_value, dtype=dtype)
File "/home/ubuntu/apache_tvm/tvm/python/tvm/relay/expr.py", line 707, in
const
raise ValueError("value has to be scalar or NDArray")
ValueError: value has to be scalar or NDArray
```
because `fill_value` is
```
%0 = cast(64, dtype="float32");
power(%0, 0.5f)
```
--
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]
