[Bugfix] Check dimensions of multimodal embeddings in V1 (#15816)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
2025-04-01 00:01:35 +08:00
parent e5ef4fa99a
commit 09e974d483
14 changed files with 98 additions and 37 deletions
--- a/vllm/model_executor/models/vision.py
+++ b/vllm/model_executor/models/vision.py
@@ -204,7 +204,7 @@ def scatter_patch_features(
            (e_is_patch.shape[0], patches_one.shape[-1]),
            fill_value=torch.nan,
        )
-        embed_one[e_is_patch] = patches_one.flatten(0, -2)
+        embed_one[e_is_patch] = patches_one
        return embed_one

    return tuple(