From 4880a43d2085871e0d159183e6bada20e3df8a9c Mon Sep 17 00:00:00 2001 From: Lucas Wilkinson Date: Wed, 29 Jan 2025 22:46:43 -0500 Subject: [PATCH] Update utils.py Co-authored-by: Michael Goin --- vllm/attention/backends/mla/utils.py | 4 ++-- 1 file changed, 2 insertions(+), 2 deletions(-) diff --git a/vllm/attention/backends/mla/utils.py b/vllm/attention/backends/mla/utils.py index f650a854e4d50..28f00a2797dcf 100644 --- a/vllm/attention/backends/mla/utils.py +++ b/vllm/attention/backends/mla/utils.py @@ -231,7 +231,7 @@ class MLAImplCommon(AttentionImpl): .flatten(start_dim=0, end_dim=1).contiguous() tp_size = get_tensor_model_parallel_world_size() - self.o_proj_absored = RowParallelLinear( + self.o_proj_absorbed = RowParallelLinear( self.W_UV_O.shape[0] * tp_size, self.W_UV_O.shape[1], bias=False, @@ -239,7 +239,7 @@ class MLAImplCommon(AttentionImpl): #quant_config=self.o_proj.quant_method, ) - self.o_proj_absored.weight = torch.nn.Parameter(self.W_UV_O.T) + self.o_proj_absorbed.weight = torch.nn.Parameter(self.W_UV_O.T) else: self.W_UV = W_UV self.W_UK = W_UK