[Feature][Quantization] MXFP4 support for MOE models (#17888)

Signed-off-by: Felix Marty <felmarty@amd.com>
Signed-off-by: Bowen Bao <bowenbao@amd.com>
Signed-off-by: Felix Marty <Felix.Marty@amd.com>
Co-authored-by: Bowen Bao <bowenbao@amd.com>
This commit is contained in:
fxmarty-amd
2025-07-09 13:19:02 -07:00
committed by GitHub
co-authored by Bowen Bao
parent bf03ff3575
commit 332d4cb17b
15 changed files with 873 additions and 104 deletions
@@ -237,12 +237,6 @@ class QuarkConfig(QuantizationConfig):
"Quark model is not in MX-FP4 format: not group_size=32")
return False
# Weights need to use static quantization.
if weight_quant.get("is_dynamic") is True:
logger.debug(
"Quark model is not in MX-FP4 format: not weight static")
return False
# Activations need to use dynamic quantization.
if input_quant.get("is_dynamic") is False:
logger.debug(