A high-throughput and memory-efficient inference and serving engine for LLMs
Mirrored from https://github.com/tduncan-lum/Nvidia-Optimization-Stack---VLLM.git
Synced Nov 4, 13:45 UTC
[Bugfix] Fix Deepseek V4 import error due to AOT compile cache loading (#41090)
Signed-off-by: wzhao18 <wzhao18.sz@gmail.com> Signed-off-by: Wei Zhao <51183510+wzhao18@users.noreply.github.com>
Wei Zhao committed
Apr 29, 2026 at 00:08 UTC
803b9d7881cd3a8482aaa1e6bf990193b55c6331
1 file changed
+5
-5
vllm/model_executor/models/deepseek_v4.py
+5
-5
@@ -1097,6 +1097,11 @@ class DeepseekV4DecoderLayer(nn.Module):
1097
aux_stream_list: list[torch.cuda.Stream] | None = None,
1098
):
1099
super().__init__()
1100
+
1101
+ # Lazy import to avoid top-level tilelang dependency.
1102
+ # Registers both torch.ops.vllm.mhc_pre and mhc_post
1103
+ import vllm.model_executor.layers.mhc # noqa: F401
1104
+
1105
config = vllm_config.model_config.hf_config
1106
self.hidden_size = config.hidden_size
1107
@@ -1167,11 +1172,6 @@ class DeepseekV4DecoderLayer(nn.Module):
1172
hc_scale: torch.Tensor,
1173
hc_base: torch.Tensor,
1174
):
1170
- # Lazy import to avoid top-level tilelang dependency.
1171
- # Registers both torch.ops.vllm.mhc_pre and mhc_post,
1172
- # so hc_post() doesn't need its own import.
1173
- import vllm.model_executor.layers.mhc # noqa: F401
1174
-
1175
post_mix, res_mix, layer_input = torch.ops.vllm.mhc_pre(
1176
residual=x,
1177
fn=hc_fn,