Add support for ModelOpt MXFP8 dense models (#33786)
Signed-off-by: Daniel Serebrenik <daserebrenik@nvidia.com>
This commit is contained in:
@@ -494,6 +494,7 @@ class FusedMoEQuantConfig:
|
||||
"mxfp4",
|
||||
"mxfp6_e3m2",
|
||||
"mxfp6_e2m3",
|
||||
"mxfp8",
|
||||
}
|
||||
assert not isinstance(weight_dtype, str) or weight_dtype in {
|
||||
"nvfp4",
|
||||
@@ -501,6 +502,7 @@ class FusedMoEQuantConfig:
|
||||
"mxfp6_e3m2",
|
||||
"mxfp6_e2m3",
|
||||
"int4",
|
||||
"mxfp8",
|
||||
}
|
||||
|
||||
if weight_dtype is None:
|
||||
|
||||
Reference in New Issue
Block a user