From ae8da14ea34b83be42f7ad4f140c0875a356bfac Mon Sep 17 00:00:00 2001
From: Yihao Wang <42559837+AgainstEntropy@users.noreply.github.com>
Date: Wed, 8 Apr 2026 08:45:42 -0700
Subject: [PATCH] [fix] [whisper] ensure inputs are moved to the correct device
 before processing. (#22293)

---
 python/sglang/srt/models/whisper.py | 4 ++++
 1 file changed, 4 insertions(+)

diff --git a/python/sglang/srt/models/whisper.py b/python/sglang/srt/models/whisper.py
index d9190a2f1..987d230ff 100644
--- a/python/sglang/srt/models/whisper.py
+++ b/python/sglang/srt/models/whisper.py
@@ -292,6 +292,10 @@ class WhisperEncoder(torch.nn.Module):
         position_ids: torch.Tensor,
         forward_batch: ForwardBatch,
     ):
+        device = self.conv1.weight.device
+        input_features = input_features.to(device=device)
+        position_ids = position_ids.to(device=device)
+
         inputs_embeds = torch.nn.functional.gelu(self.conv1(input_features))
         inputs_embeds = torch.nn.functional.gelu(self.conv2(inputs_embeds))