[Compressed-Tensors] Support Kimi-K3 quantized models (#50500)
Signed-off-by: Kyle Sayers <kylesayrs@gmail.com>
Kyle Sayers committed
Jul 31, 2026 at 13:06 UTC
d87d2ca74791fe23ca2a0ba68c997de1618b33b6
1 file changed
+8
-1
vllm/models/kimi_k3/nvidia/model.py
+8
-1
@@ -951,7 +951,13 @@ class KimiDecoderLayer(nn.Module):
951
return hidden_states, prefix_sum, residual
952
953
954
-class KimiLinearModel(nn.Module, EagleModelMixin):
954
+class KimiLinearModel(nn.Module, EagleModelMixin, SupportsQuant):
955
+ packed_modules_mapping = {
956
+ "gate_up_proj": ["gate_proj", "up_proj"],
957
+ "in_proj_qkvgfab": ["q_proj", "k_proj", "v_proj", "b_proj", "f_a_proj"],
958
+ "conv1d": ["q_conv1d", "k_conv1d", "v_conv1d"],
959
+ }
960
+
961
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
962
super().__init__()
963
@@ -1221,6 +1227,7 @@ class KimiLinearModel(nn.Module, EagleModelMixin):
1227
else:
1228
expert_params_mapping = []
1229
params_dict = dict(self.named_parameters())
1230
+
1231
# Under the MXFP4 quant interface the routed experts register unpacked
1232
# params (``w13_weight``), while the compressed-tensors checkpoint names
1233
# them ``.weight_packed``. Rebind so the expert mapping resolves; scales