[Model] Gemma3n MM (#20495)

Signed-off-by: ShriKode <shrikode@gmail.com>
Signed-off-by: NickLucche <nlucches@redhat.com>
Signed-off-by: Roger Wang <hey@rogerw.me>
Co-authored-by: ShriKode <shrikode@gmail.com>
Co-authored-by: Roger Wang <hey@rogerw.me>

examples/offline_inference/audio_language.py

@@ -96,6 +96,25 @@ def run_voxtral(question: str, audio_count: int) -> ModelRequestData:
     )
 
 
+# Gemma3N
+def run_gemma3n(question: str, audio_count: int) -> ModelRequestData:
+    model_name = "google/gemma-3n-E2B-it"
+    engine_args = EngineArgs(
+        model=model_name,
+        max_model_len=2048,
+        max_num_batched_tokens=2048,
+        max_num_seqs=2,
+        limit_mm_per_prompt={"audio": audio_count},
+        enforce_eager=True,
+    )
+    prompt = f"<start_of_turn>user\n<audio_soft_token>{question}"
+    "<end_of_turn>\n<start_of_turn>model\n"
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompt=prompt,
+    )
+
+
 # Granite Speech
 def run_granite_speech(question: str, audio_count: int) -> ModelRequestData:
     # NOTE - the setting in this example are somehat different than what is
@@ -331,6 +350,7 @@ def run_whisper(question: str, audio_count: int) -> ModelRequestData:
 
 model_example_map = {
     "voxtral": run_voxtral,
+    "gemma3n": run_gemma3n,
     "granite_speech": run_granite_speech,
     "minicpmo": run_minicpmo,
     "phi4_mm": run_phi4mm,

examples/offline_inference/vision_language.py

@@ -211,7 +211,33 @@ def run_gemma3(questions: list[str], modality: str) -> ModelRequestData:
         )
         for question in questions
     ]
+    return ModelRequestData(
+        engine_args=engine_args,
+        prompts=prompts,
+    )
 
+
+# Gemma3N
+def run_gemma3n(questions: list[str], modality: str) -> ModelRequestData:
+    assert modality == "image"
+    model_name = "google/gemma-3n-E2B-it"
+
+    engine_args = EngineArgs(
+        model=model_name,
+        max_model_len=2048,
+        max_num_seqs=2,
+        limit_mm_per_prompt={modality: 1},
+        enforce_eager=True,
+    )
+
+    prompts = [
+        (
+            "<start_of_turn>user\n"
+            f"<image_soft_token>{question}<end_of_turn>\n"
+            "<start_of_turn>model\n"
+        )
+        for question in questions
+    ]
     return ModelRequestData(
         engine_args=engine_args,
         prompts=prompts,
@@ -1395,6 +1421,7 @@ model_example_map = {
     "florence2": run_florence2,
     "fuyu": run_fuyu,
     "gemma3": run_gemma3,
+    "gemma3n": run_gemma3n,
     "glm4v": run_glm4v,
     "glm4_1v": run_glm4_1v,
     "h2ovl_chat": run_h2ovl,