graph: backend: dnnl: enhance fp8 matmul patterns #2316
Conversation
| @@ -0,0 +1,365 @@ | |||
| { | |||
| "version": "3.0.0", | |||
There was a problem hiding this comment.
please update the json with the latest version
There was a problem hiding this comment.
Thanks for reminding. Have fixed.
There was a problem hiding this comment.
@wzt1997 changing the version number only is not enough. Some new fields are still missing in this json file: fpmath_mode_apply_to_int, input_ports, output_ports, etc. It's confusing that a json file with version = 3.7.0 but without these fields.
There was a problem hiding this comment.
I partly agree with you. For input_ports and output_ports, I agree we should align them in the test cases( although it seems they are ignored and do not affect the testing currently ).
But, for other fields such as fpmath_mode_apply_to_int, according to the semantic, it only affects integral inputs, hence it does not have to appear here.
There was a problem hiding this comment.
We don't have other reliable ways to generate graph json files except dumping them from the library. If this one is dumped with v3.7 (as suggested by the version number), it should contain these fields, no matter if they take effect or not.
according to the semantic, it only affects integral inputs, hence it does not have to appear here.
Actually, that's what we want to test: users will always dump json file with fpmath_mode_apply_to_int in it and tests it with benchdnn. Benchnn + library should work properly when the field is provided in the json file.
There was a problem hiding this comment.
OK. I've updated the test case. Also it's found that there are many other test case with version=3.7.0 do not match the requirement, such as the single dynamic dequantize and gated-mlp. If necessary, we may need to enhance them with dumped json files as well.
There was a problem hiding this comment.
The gated-mlp case was added before fpmath_mode_apply_to_int was introduced into the library. I would not insist on upgrading all legacy cases to the latest version of the library. Instead, I suggest when a new case is being added, it should comply with the latest version of the library.
wzt1997
force-pushed
the
zhitao/graph-enhance-f8-matmul
branch
from
b4c9269 to
59e5f36
Compare
wzt1997
force-pushed
the
zhitao/graph-enhance-f8-matmul
branch
4 times, most recently
from
b24ddda to
8db9e19
Compare
wzt1997
force-pushed
the
zhitao/graph-enhance-f8-matmul
branch
from
8db9e19 to
bed46bd
Compare
|
make test |
Description
Enhance support for fp8 MatMul patterns in oneDNN Graph.