From 23c15acd770cf16ed36c6d3fed8e7d78db7d5282 Mon Sep 17 00:00:00 2001 From: gaozihao-shy Date: Mon, 18 May 2026 13:07:16 +0800 Subject: [PATCH] [BugFix] Kimi-K2.5: skip vision tower dtype conversion when using quantization (#42869) Signed-off-by: gaozihao-shy Signed-off-by: gaozihao --- vllm/model_executor/models/kimi_k25.py | 9 ++++++--- 1 file changed, 6 insertions(+), 3 deletions(-) diff --git a/vllm/model_executor/models/kimi_k25.py b/vllm/model_executor/models/kimi_k25.py index 297ebca29b..89cda63c80 100644 --- a/vllm/model_executor/models/kimi_k25.py +++ b/vllm/model_executor/models/kimi_k25.py @@ -339,9 +339,12 @@ class KimiK25ForConditionalGeneration( quant_config=self._maybe_ignore_quant_config(quant_config), prefix=maybe_prefix(prefix, "vision_tower"), ) - self.vision_tower = self.vision_tower.to( - device=self.device, dtype=model_config.dtype - ) + if self._maybe_ignore_quant_config(quant_config) is not None: + self.vision_tower = self.vision_tower.to(device=self.device) + else: + self.vision_tower = self.vision_tower.to( + device=self.device, dtype=model_config.dtype + ) self.mm_projector = KimiK25MultiModalProjector( config=config.vision_config,