Skip to content

Commit c0cff38

Browse files
committed
support TP
Signed-off-by: Yu Chin Fabian Lim <fabian.lim@gmail.com>
1 parent 9af4bf9 commit c0cff38

File tree

1 file changed

+4
-1
lines changed

1 file changed

+4
-1
lines changed

vllm/model_executor/layers/mamba/mamba_mixer2.py

Lines changed: 4 additions & 1 deletion
Original file line numberDiff line numberDiff line change
@@ -132,7 +132,10 @@ def forward_cuda(
132132
return x * nn.functional.silu(gate.to(
133133
torch.float32)).to(input_dtype)
134134

135-
if self.tp_size > 1 or self.n_groups != 1:
135+
if (
136+
((self.n_groups % self.tp_size) != 0) or
137+
self.n_groups != 1
138+
):
136139
return self.forward_native(x, gate)
137140

138141
return rms_norm_gated(x,

0 commit comments

Comments
 (0)