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 9, 10:00 UTC
[Backport] [Kimi-K2.5] Replace torch.cuda with current_platform for d… (#33320)
sthWrong committed
Jan 29, 2026 at 20:29 UTC
17b17c068453e6dc6af79240bb94857ae175cc51
1 file changed
+2
-1
vllm/model_executor/models/kimi_k25.py
+2
-1
@@ -58,6 +58,7 @@ from vllm.multimodal.processing import (
58
PromptReplacement,
59
PromptUpdate,
60
)
61
+from vllm.platforms import current_platform
62
from vllm.sequence import IntermediateTensors
63
from vllm.transformers_utils.configs import KimiK25Config
64
from vllm.transformers_utils.processor import cached_get_image_processor
@@ -320,7 +321,7 @@ class KimiK25ForConditionalGeneration(nn.Module, SupportsMultiModal, SupportsPP)
321
model_config.multimodal_config.mm_encoder_tp_mode == "data"
322
)
323
self.hidden_size = config.text_config.hidden_size
323
- self.device = torch.cuda.current_device()
324
+ self.device = current_platform.current_device()
325
# Build vision tower directly with KimiK25VisionConfig
326
self.vision_tower = MoonViT3dPretrainedModel(
327
config.vision_config,