Fix issue of using only the part of video frame [Nemotron Nano] (vllm-project#26186)

BloodAxe · web-flow · commit 5a05f2660370 · 2025-10-04T00:21:00.000Z
Signed-off-by: Eugene Khvedchenia &lt;ekhvedchenia@nvidia.com&gt;
diff --git a/vllm/model_executor/models/nano_nemotron_vl.py b/vllm/model_executor/models/nano_nemotron_vl.py
@@ -208,7 +208,7 @@ def video_to_pixel_values(
         )
         # dynamic_preprocess returns tensors already; take the single tile
         assert len(pil_frame) >= 1
-        frames_tensors.append(pil_frame[0])
+        frames_tensors.append(pil_frame[-1])
 
     return torch.stack(frames_tensors)
 

Original file line number	Diff line number	Diff line change
`@@ -208,7 +208,7 @@ def video_to_pixel_values(`
`208`	`208`	`)`
`209`	`209`	`# dynamic_preprocess returns tensors already; take the single tile`
`210`	`210`	`assert len(pil_frame) >= 1`
`211`		`- frames_tensors.append(pil_frame[0])`
	`211`	`+ frames_tensors.append(pil_frame[-1])`
`212`	`212`
`213`	`213`	`return torch.stack(frames_tensors)`
`214`	`214`