[Compressed-Tensors] Support Kimi-K3 quantized models (#50500)

Signed-off-by: Kyle Sayers <kylesayrs@gmail.com>
This commit is contained in:
Kyle Sayers
2026-07-31 17:06:30 +00:00
committed by GitHub
parent 8d8a4e0f2b
commit d87d2ca747
+8 -1
View File
@@ -951,7 +951,13 @@ class KimiDecoderLayer(nn.Module):
return hidden_states, prefix_sum, residual
class KimiLinearModel(nn.Module, EagleModelMixin):
class KimiLinearModel(nn.Module, EagleModelMixin, SupportsQuant):
packed_modules_mapping = {
"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"],
}
def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""):
super().__init__()
@@ -1221,6 +1227,7 @@ class KimiLinearModel(nn.Module, EagleModelMixin):
else:
expert_params_mapping = []
params_dict = dict(self.named_parameters())
# Under the MXFP4 quant interface the routed experts register unpacked
# params (``w13_weight``), while the compressed-tensors checkpoint names
# them ``.weight_packed``. Rebind so the expert mapping resolves; scales