Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/65715
Here is how we freeze a python module:
- we call python builtin compile method with the source code of the modules and the path. This method returns a python code object
- we call marshal.dumps to serialize the code object to bytes.
The code_object.co_filename actually matches the one passed in to the compile method. We can simply replace that with a marker
to avoid leak build time path to runtime.
This works on nested code objects as well:
```
#!/bin/env python3.8
import marshal
code_str = """
print("hello")
class MyCls:
def __init__(self):
pass
"""
co = compile(code_str, "<Generated by torch::deploy>", "exec")
cobytes = marshal.dumps(co)
import pdb; pdb.set_trace()
```
Checking `co`:
```
(Pdb) co.co_filename
'<Generated by torch::deploy>'
(Pdb) co.co_consts
('hello', <code object MyCls at 0x7f0e8670bbe0, file "<Generated by torch::deploy>", line 4>, 'MyCls', None)
(Pdb) co.co_consts[1].co_filename
'<Generated by torch::deploy>'
```
Test Plan:
Find the serialized frozenmodule for torch.nn.modules.linear module in the generated bytecode_x.c file. Put the content to /tmp/linear.bytecode
Run the testing script:
```
import marshal
co_bytes = bytes(eval("[{}]".format("".join(open('/tmp/linear.bytecode').readlines()).replace('\n', '').replace('\t', ''))))
co = marshal.loads(co_bytes)
print(co)
```
The output for the paste without the change:
```
<code object <module> at 0x7f39ca7f07c0, file "/data/users/shunting/fbsource/fbcode/buck-out/opt/gen/caffe2/gen_frozen_torchpython_src__srcs/torch/nn/modules/linear.py", line 1>
```
The output for the paste with the change:
```
<code object <module> at 0x7f05a765d710, file "<Generated by torch::deploy>", line 1>
````
Note that the file part is changed as expected.
Reviewed By: suo
Differential Revision: D31214555
fbshipit-source-id: 56958e0a7352f8c30a3377f83209efe7db61f0fb
