[Bugfix][ROCm] RDNA3 W4A16 MoE: handle the N-first CT weight layout (#60388)

Signed-off-by: JartX <sagformas@epdcenter.es>
This commit is contained in:
JartX
2026-10-08 15:06:39 +00:00
committed by GitHub
parent d80ddaac73
commit 994b8b77be
3 changed files with 95 additions and 9 deletions
@@ -450,10 +450,12 @@ class CompressedTensorsWNA16MoEMethod(CompressedTensorsMoEMethod):
if w2_bias is not None:
replace_parameter(layer, "w2_bias", w2_bias)
# CPU fused_experts_cpu requires zero points even for symmetric quant.
# CPU fused_experts_cpu and the RDNA3 HIP kernel require zero points
# even for symmetric quant (the oracle synthesizes them).
# EMULATION bakes ZP into the dequantized bf16 weights — ZP is None.
if (
not self.symmetric or self.wna16_backend == WNA16MoEBackend.CPU
not self.symmetric
or self.wna16_backend in (WNA16MoEBackend.CPU, WNA16MoEBackend.RDNA3)
) and self.wna16_backend != WNA16MoEBackend.EMULATION:
assert w13_qzeros is not None and w2_qzeros is not None
replace_parameter(layer, "w13_weight_zero_point", w13_qzeros)