[QeRL] Compose online quantization with quantized reloading (#38032)

Signed-off-by: Kyle Sayers <kylesayrs@gmail.com>
This commit is contained in:
Kyle Sayers
2026-03-27 13:22:33 -07:00
committed by GitHub
parent 7ba425e916
commit 648edcf729
10 changed files with 184 additions and 260 deletions
@@ -337,6 +337,8 @@ class Mxfp8OnlineMoEMethod(Fp8OnlineMoEMethod):
w2 = torch.empty_like(layer.w2_weight, dtype=fp8_dtype)
w13_scale = layer.w13_weight_scale
w2_scale = layer.w2_weight_scale
layer.w13_input_scale = None
layer.w2_input_scale = None
w13, w13_scale = self._quantize_mxfp8_moe_weight(layer.w13_weight)
w2, w2_scale = self._quantize_mxfp8_moe_weight(layer.w2_weight)