From 5fba75aefeedcf5b6cc27abf9bc145b6a49873a7 Mon Sep 17 00:00:00 2001 From: yjz <126063153+JianDan0212@users.noreply.github.com> Date: Thu, 6 Aug 2026 19:39:17 +0800 Subject: [PATCH] [Bugfix][Model] Add missing fused_qkv_a_proj to Kimi-Linear packed_modules_mapping (#51249) Signed-off-by: JianDan0212 Co-authored-by: JianDan0212 --- vllm/models/kimi_k3/nvidia/model.py | 1 + 1 file changed, 1 insertion(+) diff --git a/vllm/models/kimi_k3/nvidia/model.py b/vllm/models/kimi_k3/nvidia/model.py index 78093de02a..0d98bccf06 100644 --- a/vllm/models/kimi_k3/nvidia/model.py +++ b/vllm/models/kimi_k3/nvidia/model.py @@ -1014,6 +1014,7 @@ class KimiLinearModel(nn.Module, EagleModelMixin, SupportsQuant): "gate_up_proj": ["gate_proj", "up_proj"], "in_proj_qkvgfab": ["q_proj", "k_proj", "v_proj", "b_proj", "f_a_proj"], "conv1d": ["q_conv1d", "k_conv1d", "v_conv1d"], + "fused_qkv_a_proj": ["q_a_proj", "kv_a_proj_with_mqa"], } def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):