fix typo; useless code

Signed-off-by: HonestDeng <2958906959@qq.com>

Author: HonestDeng

examples/offline_inference/mammothmodal2_preview/run_mammothmoda2_image_summarize.py

@@ -2,7 +2,7 @@
 Offline inference example: MammothModa2 image summarization (single AR stage).
 
 Example:
-  uv run python examples/offline_inference/mammothmodal2_preview/run_mammothmoda2_image_summary.py \
+  uv run python examples/offline_inference/mammothmodal2_preview/run_mammothmoda2_image_summarize.py \
     --model /data/datasets/models-hf/MammothModa2-Preview \
     --image /path/to/input.jpg \
     --question "Please summarize the content of this image."

tests/e2e/offline_inference/test_mammoth_moda2.py

@@ -110,11 +110,23 @@ def test_missing_llm_config_raises(self):
             _ = Mammothmoda2Config(llm_config=None).image_token_id
 
     def test_t2i_generation_config_json(self):
-        """t2i_generation_config.json must contain required token-ID fields."""
+        """t2i_generation_config.json must contain required token-ID and sampling fields."""
         cfg = _load_t2i_gen_config(MODEL_PATH)
-        for key in ("eol_token_id", "visual_token_start_id", "visual_token_end_id"):
+        for key in ("eol_token_id", "visual_token_start_id", "visual_token_end_id", "top_k"):
             assert key in cfg and isinstance(cfg[key], int)
 
+    def test_model_config_visual_ids(self):
+        """config.json llm_config must contain the four Qwen2.5-VL vision token IDs."""
+        cfg_path = Path(MODEL_PATH) / "config.json"
+        if not cfg_path.exists():
+            pytest.skip(f"config.json not found at {cfg_path}")
+        with cfg_path.open() as f:
+            llm_cfg = json.load(f).get("llm_config", {})
+        for key in ("image_token_id", "video_token_id", "vision_start_token_id", "vision_end_token_id"):
+            assert key in llm_cfg and isinstance(llm_cfg[key], int), (
+                f"Missing or non-int field '{key}' in config.json llm_config"
+            )
+
 
 # ---------------------------------------------------------------------------
 # 2. Stage input processor: ar2dit helpers + tests (CPU-only)

vllm_omni/model_executor/stage_input_processors/mammoth_moda2.py

@@ -61,7 +61,6 @@ def ar2dit(
 
         gen_token_mask = full_token_ids_t >= gen_vocab_start_index
 
-        visual_token_mask = torch.zeros_like(gen_token_mask)
         visual_token_mask = torch.isin(
             full_token_ids_t,
             torch.tensor(visual_ids, dtype=torch.long, device=mask_device),