[Quant][Feature] Support online MXFP8 quantization for MoE and dense models (#35448)

Signed-off-by: EdalatiAli <aliedalati@cohere.com>
This commit is contained in:
EdalatiAli
2026-03-16 18:07:39 -04:00
committed by GitHub
parent fd4d96302a
commit e5b807607c
10 changed files with 747 additions and 56 deletions
@@ -199,7 +199,7 @@ def _mxfp8_e4m3_quantize(
) -> tuple[torch.Tensor, torch.Tensor]:
assert A_scale is None
assert not per_act_token_quant
assert block_shape is None
assert block_shape is None or block_shape == [1, 32]
return mxfp8_e4m3_quantize(A, is_sf_swizzled_layout)