[Bugfix] Merge MM embeddings by index instead of token IDs (#16229)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
Signed-off-by: NickLucche <nlucches@redhat.com>
Signed-off-by: Roger Wang <hey@rogerw.io>
Co-authored-by: NickLucche <nlucches@redhat.com>
Co-authored-by: Roger Wang <hey@rogerw.io>
This commit is contained in:
Cyrus Leung
2025-09-27 08:15:12 +00:00
committed by GitHub
co-authored by NickLucche Roger Wang
parent 176173989a
commit 27d7638b94
80 changed files with 966 additions and 1139 deletions
-18
View File
@@ -1450,24 +1450,6 @@ class BaseKeyeModule(nn.Module):
multimodal_embeddings += video_embeddings
return multimodal_embeddings
def get_input_embeddings(
self,
input_ids: torch.Tensor,
multimodal_embeddings: Optional[MultiModalEmbeddings] = None,
) -> torch.Tensor:
inputs_embeds = self.language_model.get_input_embeddings(input_ids)
if multimodal_embeddings is not None:
inputs_embeds = merge_multimodal_embeddings(
input_ids,
inputs_embeds,
multimodal_embeddings,
[
self.config.image_token_id,
self.config.video_token_id,
],
)
return inputs_embeds
def get_input_embeddings_v0(
self,
input_ids: torch.Tensor,