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 2, 20:45 UTC
[Bugfix] fix device mismatch in MiniCPM-o-4_5 resampler (#43194)
Signed-off-by: Yan Ma <yan.ma@intel.com>
Yan Ma committed
May 26, 2026 at 13:12 UTC
f815c9995491aed8ace3d7f2a6ddd5065269b06a
1 file changed
+3
-1
vllm/model_executor/models/minicpmv.py
+3
-1
@@ -378,7 +378,9 @@ class Resampler4_5(Resampler2_5):
378
) # D
379
380
pos_embed_2d.append(
381
- self.pos_embed[:tgt_h, :tgt_w, :].reshape((tgt_h * tgt_w, -1)).to(dtype)
381
+ self.pos_embed[:tgt_h, :tgt_w, :]
382
+ .reshape((tgt_h * tgt_w, -1))
383
+ .to(device=device, dtype=dtype)
384
) # patches * D
385
key_padding_mask[i, patch_len[i] :] = True
386