fix(speedup): re-write aten schema parser to support pytorch versions < 1.9.0 #5138
Conversation
|
I suggest to keep both version and use a flag to switch. |
|
Azure Pipelines successfully started running 1 pipeline(s). |
|
/azp run full test - compression |
|
Azure Pipelines successfully started running 1 pipeline(s). |
| @@ -486,7 +748,10 @@ def generate_aten_to_python(func: Callable, node: NodePyGroup, speedup: ModelSpe | |||
| c_node = node.key_node | |||
|
|
|||
| schema = c_node.schema() | |||
| positional_num, keyword_list, special_treat = parse_aten_schema(schema) | |||
| if torch.__version__ < '1.9.0': | |||
There was a problem hiding this comment.
I'm worried with this method of judging version.
There was a problem hiding this comment.
the type of torch.__version__ is not a str. it's a version type that can correctly compare with str.
| Parse the schema, to positional_num and keyword_list, and detect if the argument should be specially treated. | ||
| Cannot use 'torch._C.parse_schema' because 'torch._C.Argument' has no 'kwarg_only' on pytorch v1.8.x | ||
| Using a lexer-parser like method to parse it. | ||
| Re-write from torch/csrc/jit/frontend/function_schema_parser.cpp |
There was a problem hiding this comment.
if this function a mirror rewrite from this cpp file?
There was a problem hiding this comment.
It's not totally equal to the original function in cpp. Actually, the original parser in that cpp file is too specific and too long (5k+ lines). I wrote a much simpler syntax analyzer, deleted some error check and specific type check to shorten the code.
|
/azp run full test - compression |
|
Azure Pipelines successfully started running 1 pipeline(s). |
|
/azp run full test - nas |
|
Azure Pipelines successfully started running 1 pipeline(s). |
|
/azp run fast test |
|
Azure Pipelines successfully started running 1 pipeline(s). |
Description
reason:
the official aten schema parser of python lost the 'kwarg_only' info in python versions <1.9.0.
and schema is not simple enough to use some regex to parser.
solution:
so I translated the schema parser code from pytorch 1.12 c++ code to python and adapted some deleted syntaxes in pytorch 1.8.
tested in pytorch versions 1.8, 1.9, 1.10, 1.11 and 1.12.
note:
torch._C.Nodesupportsschemafunction since version 1.8. so speedup cannot use on pytorch 1.7 now.#5131 can be solved.
Test Options
Checklist
How to test