[ROCm][Quant] Minimax-M3: Enable fp8_per_channel for bf16 weights on mi300x (#45854)

Signed-off-by: Hongxia Yang <hongxia.yang@amd.com>
Signed-off-by: tjtanaa <tunjian.tan@embeddedllm.com>
Co-authored-by: tjtanaa <tunjian.tan@embeddedllm.com>
This commit is contained in:
Hongxia Yang
2026-06-17 12:02:40 +00:00
committed by GitHub
co-authored by tjtanaa
parent ee0fd6984a
commit e28e8c8782
5 changed files with 11 additions and 0 deletions
@@ -787,6 +787,8 @@ class Fp8MoEMethod(FusedMoEMethodBase):
a2_scale=a2_scale,
block_shape=self.weight_block_size,
swiglu_limit=getattr(layer, "swiglu_limit", None),
gemm1_alpha=getattr(layer, "swiglu_alpha", None),
gemm1_beta=getattr(layer, "swiglu_beta", None),
)
# Inject biases into the quant config if the model has them