Use gemma3_text model_type so vLLM sizes the KV cache from the real dims

Author: anthony-etim

keras_hub/src/vllm/registry.py

@@ -335,7 +335,15 @@ def setup_vllm_model(preset: str, dtype: str = "float16") -> str:
         "_name_or_path": f"keras_hub:{preset}",
         "keras_hub_preset": preset,
         "torch_dtype": dtype,
-        "model_type": model_type,
+        # transformers maps "gemma3" to the MULTIMODAL Gemma3 config
+        # class, which ignores top-level dims and sizes the KV cache from
+        # its text-config defaults (wrong KV head count for e.g.
+        # gemma3_instruct_1b's MQA). "gemma3_text" is the text-only config
+        # class that respects the real dims written below.
+        "model_type": {
+            "gemma3": "gemma3_text",
+            "gemma4": "gemma3_text",
+        }.get(model_type, model_type),
         "vocab_size": vocab_size,
     }
     if eos_token_id is not None:
@@ -408,6 +416,8 @@ def _derive_arch_config(preset: str) -> dict:
         arch["hidden_size"] = int(hidden)
     if cfg.get("intermediate_dim"):
         arch["intermediate_size"] = int(cfg["intermediate_dim"])
+    if cfg.get("sliding_window_size"):
+        arch["sliding_window"] = int(cfg["sliding_window_size"])
     return arch