[VLM] fix LFM2-VL offline inference and GPU JPEG decode (#22448)

python/sglang/srt/models/lfm2_vl.py

@@ -274,7 +274,7 @@ class Lfm2VlForConditionalGeneration(nn.Module):
 
         return projected_packed
 
-    @torch.inference_mode()
+    @torch.no_grad()
     def forward(
         self,
         input_ids: torch.Tensor,

python/sglang/srt/multimodal/processors/lfm2_vl.py

@@ -32,6 +32,7 @@ class Lfm2VlImageProcessor(SGLangBaseProcessor):
     """
 
     models = [Lfm2VlForConditionalGeneration]
+    gpu_image_decode = False
 
     def __init__(self, hf_config, server_args, _processor, *args, **kwargs):
         super().__init__(hf_config, server_args, _processor, *args, **kwargs)