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 5, 12:15 UTC
nano_nemotron_vl: fix tensor device mismatch exception when video profiling (#39029)
Signed-off-by: Netanel Haber <58652339+netanel-haber@users.noreply.github.com>
Netanel Haber committed
Apr 6, 2026 at 01:23 UTC
d56e95223917ba21af2dd87a335dcb5c0b347bfe
1 file changed
+3
-2
vllm/model_executor/models/nano_nemotron_vl.py
+3
-2
@@ -1239,12 +1239,13 @@ class NemotronH_Nano_VL_V2(
1239
img_context_token_ids=self._img_context_token_ids,
1240
video_temporal_patch_size=video_temporal_patch_size,
1241
)
1242
+ device = video_embeddings.device
1243
1244
# video_repl.full is a list of token IDs
1244
- repl_token_ids = torch.tensor(video_repl.full)
1245
+ repl_token_ids = torch.tensor(video_repl.full, device=device)
1246
1247
# Get embedding token IDs for image context (use pre-tokenized version)
1247
- embed_token_ids = torch.tensor(self._img_context_token_ids)
1248
+ embed_token_ids = torch.tensor(self._img_context_token_ids, device=device)
1249
1250
# Create mask for video embedding positions
1251
is_video_embed = torch.isin(repl_token_ids, embed_token_ids)