Fix MiniMax-M2 rmsnorm precision and remove useless code (#27627)

Signed-off-by: xuebi <xuebi@minimaxi.com> Co-authored-by: xuebi <xuebi@minimaxi.com>
2025-10-29 21:01:05 +08:00
parent ecca3fee76
commit d6704dd099
2 changed files with 1 additions and 19 deletions
--- a/vllm/model_executor/layers/mamba/linear_attn.py
+++ b/vllm/model_executor/layers/mamba/linear_attn.py
@@ -77,7 +77,7 @@ class MiniMaxText01RMSNormTP(CustomOp):
        if self.tp_world > 1:
            variance = tensor_model_parallel_all_reduce(variance) / self.tp_world
        x = x * torch.rsqrt(variance + self.variance_epsilon)
-        x = x.to(orig_dtype) * self.weight
+        x = (x * self.weight).to(orig_dtype)
        return x

    def forward(