[VLM] Merged multi-modal processor for GLM4V

docs/source/models/supported_models.md

@@ -719,7 +719,7 @@ See [this page](#generative-models) for more information on how to use generativ
   * `THUDM/glm-4v-9b` etc.
   * ✅︎
   * ✅︎
-  *
+  * ✅︎
 - * `H2OVLChatModel`
   * H2OVL
   * T + I<sup>E+</sup>

examples/offline_inference/vision_language.py

@@ -106,7 +106,9 @@ def run_glm4v(question: str, modality: str):
               trust_remote_code=True,
               enforce_eager=True,
               disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache)
-    prompt = question
+    prompt = f"<|user|>\n<|begin_of_image|><|endoftext|><|end_of_image|>\
+        {question}<|assistant|>"
+
     stop_token_ids = [151329, 151336, 151338]
     return llm, prompt, stop_token_ids
 

tests/models/multimodal/processing/test_common.py

@@ -164,6 +164,7 @@ def _test_processing_correctness(
     "Qwen/Qwen2.5-VL-3B-Instruct",
     "Qwen/Qwen2-Audio-7B-Instruct",
     "fixie-ai/ultravox-v0_3",
+    "THUDM/glm-4v-9b"
 ])
 @pytest.mark.parametrize("hit_rate", [0.3, 0.5, 1.0])
 @pytest.mark.parametrize("num_batches", [32])