A high-throughput and memory-efficient inference and serving engine for LLMs
[Bugfix][MiniCPM-o] Fix cuda/cpu device mismatch in Resampler2_5 pos_embed (#43844)
Signed-off-by: Parth Ashwin Jain <parthash@amd.com> Co-authored-by: Parth Ashwin Jain <parthash@amd.com> Co-authored-by: Cyrus Leung <tlleungac@connect.ust.hk> Co-authored-by: mergify[bot] <37929162+mergify[bot]@users.noreply.github.com>
Parth Ashwin Jain committed
Jun 9, 2026 at 11:58 UTC
e6fc848d4ff2d00759516b0c7a47ba640d44e4b2
1 file changed
+3
-1
vllm/model_executor/models/minicpmv.py
+3
-1
@@ -217,7 +217,9 @@ class Resampler2_5(BaseResampler):
217
for i in range(bs):
218
tgt_h, tgt_w = tgt_sizes[i].tolist()
219
pos_embed.append(
220
- self.pos_embed[:tgt_h, :tgt_w, :].reshape((tgt_h * tgt_w, -1)).to(dtype)
220
+ self.pos_embed[:tgt_h, :tgt_w, :]
221
+ .reshape((tgt_h * tgt_w, -1))
222
+ .to(device=device, dtype=dtype)
223
) # patches * D
224
key_padding_mask[i, patch_len[i] :] = True
225
pos_embed = torch.nn.utils.rnn.pad_sequence(