[Model][VLM] Add Kimi-VL model support (#16387)

Signed-off-by: courage17340 <courage17340@163.com>

tests/models/decoder_only/vision_language/test_models.py

@@ -318,6 +318,18 @@ VLM_TEST_SETTINGS = {
         use_tokenizer_eos=True,
         patch_hf_runner=model_utils.internvl_patch_hf_runner,
     ),
+    "kimi_vl": VLMTestInfo(
+        models=["moonshotai/Kimi-VL-A3B-Instruct"],
+        test_type=(VLMTestType.IMAGE, VLMTestType.MULTI_IMAGE),
+        prompt_formatter=lambda img_prompt: f"<|im_user|>user<|im_middle|>{img_prompt}<|im_end|><|im_assistant|>assistant<|im_middle|>", # noqa: E501
+        img_idx_to_prompt=lambda _: "<|media_start|>image<|media_content|><|media_pad|><|media_end|>",  # noqa: E501
+        max_model_len=8192,
+        max_num_seqs=2,
+        dtype="bfloat16",
+        tensor_parallel_size=1,
+        vllm_output_post_proc=model_utils.kimiv_vl_vllm_to_hf_output,
+        marks=[large_gpu_mark(min_gb=48)],
+    ),
     "llama4": VLMTestInfo(
         models=["meta-llama/Llama-4-Scout-17B-16E-Instruct"],
         prompt_formatter=lambda img_prompt: f"<|begin_of_text|><|header_start|>user<|header_end|>\n\n{img_prompt}<|eot|><|header_start|>assistant<|header_end|>\n\n", # noqa: E501

tests/models/decoder_only/vision_language/vlm_utils/model_utils.py

@@ -68,6 +68,17 @@ def qwen2_vllm_to_hf_output(
     return output_ids, hf_output_str, out_logprobs
 
 
+def kimiv_vl_vllm_to_hf_output(
+        vllm_output: RunnerOutput,
+        model: str) -> tuple[list[int], str, Optional[SampleLogprobs]]:
+    """Sanitize vllm output [kimi_vl models] to be comparable with hf output."""
+    output_ids, output_str, out_logprobs = vllm_output
+
+    hf_output_str = output_str + "<|im_end|>[EOS]"
+
+    return output_ids, hf_output_str, out_logprobs
+
+
 def llava_image_vllm_to_hf_output(vllm_output: RunnerOutput,
                                   model: str) -> RunnerOutput:
     config = AutoConfig.from_pretrained(model)