FX converter documentation #2039
Conversation
github-actions
bot
added
the
documentation
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
narendasan
left a comment
narendasan
left a comment
There was a problem hiding this comment.
@apbose Generally this does not feel like it is explaining the methodology for creating a converter, it sort of references a few implementations we already have. But the point of this doc is to explain to a user outside the library how they might write a converter.
So for example explaining the @torch_tensorrt.fx.tensorrt_converter and the expected function signature like you do is a good start but lean into the idea of being a user outside the code base trying to write a converter.
Also explain the impl library more in this framing as well, i.e. a library of existing functional blocks that can be composed to create a converter.
| @@ -0,0 +1,133 @@ | |||
| .. _conversion: | |||
There was a problem hiding this comment.
this label is going to conflict with other files, use something more specific
| Steps | ||
| ================== | ||
|
|
||
| External Interface |
There was a problem hiding this comment.
Rename this Op Sets, it doesnt really matter how they are generated (though it is useful to mention this) but the high order bit is that there are different op sets that we could register converters for.
| The converter library in Torch-TensorRT is located in ``TensorRT/py/torch_tensorrt/fx/converters``. | ||
| They are categorized into - ``aten_ops_converters``, ``acc_ops_converters`` and ``nn_ops_converters``. | ||
| The individual converters present are useful for the quantization workflow. | ||
| The converters are registered using the ``tensorrt_converter`` |
There was a problem hiding this comment.
Mention what the impl library is
apbose
force-pushed
the
converter_reorg_documentation
branch
from
af6810d to
54fe120
Compare
apbose
force-pushed
the
converter_reorg_documentation
branch
from
54fe120 to
3b22854
Compare
apbose
force-pushed
the
converter_reorg_documentation
branch
from
3b22854 to
4086d1a
Compare
apbose
force-pushed
the
converter_reorg_documentation
branch
from
4086d1a to
e8b3fea
Compare
apbose
force-pushed
the
converter_reorg_documentation
branch
from
e8b3fea to
fea21a9
Compare
apbose
force-pushed
the
converter_reorg_documentation
branch
from
fea21a9 to
b46d99a
Compare
apbose
force-pushed
the
converter_reorg_documentation
branch
from
b46d99a to
524f547
Compare
apbose
force-pushed
the
converter_reorg_documentation
branch
from
524f547 to
91094ed
Compare
apbose
force-pushed
the
converter_reorg_documentation
branch
from
0e97a59 to
70513c9
Compare
| has the arguments - ``network, target, args, kwargs, name``, which is common across all the operators schema. | ||
| These functions are mapped in the ``aten`` converter registry dictionary (at present a compilation of FX and dynamo converters, FX will be deprecated soon), with key as the function target name. | ||
|
|
||
| * aten_trace is produced by ``torch_tensorrt.dynamo.backend.compile`` or ``torch_tensorrt.dynamo.backend.export``. |
There was a problem hiding this comment.
This isnt the case anymore right? its torch_tensorrt.dynamo.compile and torch.compile(backend="torch_tensorrt")
There was a problem hiding this comment.
@narendasan @gs-olive . Changed this. Could you verify once?
apbose
force-pushed
the
converter_reorg_documentation
branch
from
7e7d6f3 to
5bda31f
Compare
apbose
force-pushed
the
converter_reorg_documentation
branch
from
5bda31f to
234dfaf
Compare
No description provided.