[Bugfix] Gate Kimi-K3 KDA warmup on sys.modules to skip Kimi import for non-Kimi models (#59257)
Signed-off-by: YueshenZ <YueshenZ@users.noreply.github.com> Signed-off-by: zjy0516 <riverclouds.zhu@qq.com> Co-authored-by: YueshenZ <YueshenZ@users.noreply.github.com> Co-authored-by: Claude Code <noreply@anthropic.com> Co-authored-by: Codex <noreply@openai.com> Co-authored-by: zjy0516 <riverclouds.zhu@qq.com>
YueshenZ committed
Oct 1, 2026 at 17:25 UTC
08d77cad7be4c2ba1f72ebbd5d58571e7e91b2de
1 file changed
+6
-2
vllm/model_executor/warmup/kimi_k3_triton_warmup.py
+6
-2
@@ -4,6 +4,7 @@
4
5
from __future__ import annotations
6
7
+import sys
8
from typing import TYPE_CHECKING
9
10
import torch
@@ -19,7 +20,10 @@ logger = init_logger(__name__)
20
21
22
def _get_kda_layer(worker: Worker) -> KimiK3DeltaAttention | None:
22
- from vllm.models.kimi_k3.nvidia.kda import KimiK3DeltaAttention
23
+ # Kimi model construction already imports kda. Avoid importing it here.
24
+ kda = sys.modules.get("vllm.models.kimi_k3.nvidia.kda")
25
+ if kda is None:
26
+ return None
27
28
compilation_config = getattr(
29
worker.model_runner,
@@ -33,7 +37,7 @@ def _get_kda_layer(worker: Worker) -> KimiK3DeltaAttention | None:
37
(
38
layer
39
for layer in static_context.values()
36
- if isinstance(layer, KimiK3DeltaAttention)
40
+ if isinstance(layer, kda.KimiK3DeltaAttention)
41
),
42
None,
43
)