Remove deprecated V0 compatibility code

The V0 engine compatibility path is no longer needed as vLLM has
fully transitioned to the V1 engine architecture.

Signed-off-by: Luciano Martins <[email protected]>

Author: lucianommartins

vllm/model_executor/models/gemma3_mm.py

@@ -684,19 +684,6 @@ def forward(
         if intermediate_tensors is not None:
             inputs_embeds = None
 
-        # V0 compatibility: generate inputs_embeds if not provided
-        elif inputs_embeds is None:
-            vision_embeddings = self.get_multimodal_embeddings(**kwargs)
-            inputs_embeds = self.get_input_embeddings(input_ids, vision_embeddings)
-            if vision_embeddings is not None and len(vision_embeddings) != 0:
-                kwargs = self.prepare_attn_masks(
-                    input_ids,
-                    positions,
-                    mask_dtype=self.dtype,
-                    **kwargs,
-                )
-            input_ids = None
-
         hidden_states = self.language_model.model(
             input_ids,
             positions,