A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 14, 05:30 UTC
[Model] Remove useless code from MiniMax implementation (#23982)
Signed-off-by: QscQ <qscqesze@gmail.com> Signed-off-by: qingjun <qingjun@minimaxi.com>
qscqesze committed
Sep 3, 2025 at 19:27 UTC
6997a25ac65ed6cc3c2be6d09ca45f633a345f63
1 file changed
+1
-11
vllm/model_executor/layers/mamba/linear_attn.py
+1
-11
@@ -83,17 +83,7 @@ class MiniMaxText01RMSNormTP(CustomOp):
83
variance = tensor_model_parallel_all_reduce(
84
variance) / self.tp_world
85
x = x * torch.rsqrt(variance + self.variance_epsilon)
86
-
87
- weight = self.weight
88
- if x.size(-1) != self.weight.size(0):
89
- if self.weight.size(0) < x.size(-1):
90
- repeat_count = (x.size(-1) + self.weight.size(0)) // x.size(-1)
91
- full_weight = self.weight.repeat(repeat_count)
92
- weight = full_weight[:x.size(-1)]
93
- else:
94
- weight = self.weight[:x.size(-1)]
95
-
96
- x = x.to(orig_dtype) * weight
86
+ x = x.to(orig_dtype) * self.weight
87
return x
88
89
def forward(