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 3, 11:45 UTC
[BugFix] Kimi-K2.5: skip vision tower dtype conversion when using quantization (#42869)
Signed-off-by: gaozihao-shy <gaozihao-shy@users.noreply.github.com> Signed-off-by: gaozihao <gaozihao3@huawei.com>
gaozihao-shy committed
May 18, 2026 at 13:07 UTC
23c15acd770cf16ed36c6d3fed8e7d78db7d5282
1 file changed
+6
-3
vllm/model_executor/models/kimi_k25.py
+6
-3
@@ -339,9 +339,12 @@ class KimiK25ForConditionalGeneration(
339
quant_config=self._maybe_ignore_quant_config(quant_config),
340
prefix=maybe_prefix(prefix, "vision_tower"),
341
)
342
- self.vision_tower = self.vision_tower.to(
343
- device=self.device, dtype=model_config.dtype
344
- )
342
+ if self._maybe_ignore_quant_config(quant_config) is not None:
343
+ self.vision_tower = self.vision_tower.to(device=self.device)
344
+ else:
345
+ self.vision_tower = self.vision_tower.to(
346
+ device=self.device, dtype=model_config.dtype
347
+ )
348
349
self.mm_projector = KimiK25MultiModalProjector(
350
config=config.vision_config,