Fix: Mamba2 generation mismatch between input_ids and inputs_embeds (h…

…uggingface#32694)

* fix cache when using input embeddings

* simplify check, we can always add input ids seq len since its 0 in first pass

src/transformers/models/mamba2/modeling_mamba2.py

@@ -964,8 +964,8 @@ def prepare_inputs_for_generation(
         attention_mask: Optional[torch.Tensor] = None,
         **kwargs,
     ):
-        if input_ids.shape[1] == 0:
-            past_len = inputs_embeds.shape[1]
+        if inputs_embeds is not None:
+            past_len = inputs_embeds.shape[1] + input_ids.shape[1]
         else:
             past_len = input_ids.shape[1]
         if use_cache: