[Misc] Clean up Kimi-VL

examples/offline_inference/vision_language.py

@@ -376,9 +376,9 @@ def run_kimi_vl(questions: list[str], modality: str) -> ModelRequestData:
 
     engine_args = EngineArgs(
         model="moonshotai/Kimi-VL-A3B-Instruct",
-        max_model_len=4096,
-        disable_mm_preprocessor_cache=args.disable_mm_preprocessor_cache,
         trust_remote_code=True,
+        max_model_len=4096,
+        limit_mm_per_prompt={"image": 1},
     )
 
     return ModelRequestData(

examples/offline_inference/vision_language_multi_image.py

@@ -331,11 +331,10 @@ def load_kimi_vl(question: str, image_urls: list[str]) -> ModelRequestData:
 
     engine_args = EngineArgs(
         model=model_name,
+        trust_remote_code=True,
         max_model_len=4096,
         max_num_seqs=4,
-        tensor_parallel_size=1,
         limit_mm_per_prompt={"image": len(image_urls)},
-        trust_remote_code=True,
     )
 
     placeholders = [{"type": "image", "image": url} for url in image_urls]

vllm/model_executor/models/kimi_vl.py

@@ -56,7 +56,6 @@
 from vllm.config import VllmConfig
 from vllm.distributed import (get_tensor_model_parallel_rank,
                               get_tensor_model_parallel_world_size)
-from vllm.logger import init_logger
 from vllm.model_executor.layers.fused_moe import FusedMoE
 from vllm.model_executor.layers.logits_processor import LogitsProcessor
 from vllm.model_executor.layers.sampler import SamplerOutput, get_sampler
@@ -70,22 +69,20 @@
 from vllm.model_executor.models.utils import merge_multimodal_embeddings
 from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.multimodal import MULTIMODAL_REGISTRY
-from vllm.multimodal.inputs import (MultiModalFieldConfig, MultiModalKwargs,
-                                    NestedTensors)
+from vllm.multimodal.inputs import (MultiModalDataDict, MultiModalFieldConfig,
+                                    MultiModalKwargs, NestedTensors)
 from vllm.multimodal.parse import (ImageEmbeddingItems, ImageProcessorItems,
                                    MultiModalDataItems)
 from vllm.multimodal.processing import (BaseMultiModalProcessor,
                                         BaseProcessingInfo, PromptReplacement,
                                         PromptUpdate)
-from vllm.multimodal.profiling import BaseDummyInputsBuilder, ProcessorInputs
+from vllm.multimodal.profiling import BaseDummyInputsBuilder
 from vllm.sequence import IntermediateTensors
 from vllm.transformers_utils.configs import KimiVLConfig, MoonViTConfig
 from vllm.transformers_utils.configs.deepseek_vl2 import DeepseekV2Config
 
 from .utils import is_pp_missing_parameter, maybe_prefix
 
-logger = init_logger(__name__)
-
 
 # For dummy input only
 @dataclass
@@ -143,6 +140,9 @@ class KimiVLProcessingInfo(BaseProcessingInfo):
     def get_hf_config(self):
         return self.ctx.get_hf_config(KimiVLConfig)
 
+    def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
+        return {"image": None}
+
     def get_num_image_tokens(
         self,
         *,
@@ -180,58 +180,35 @@ def get_num_image_tokens(
         token_width = (width + pad_width) // (kernel_size[1] * patch_size)
         return int(token_height * token_width)
 
-    def get_supported_mm_limits(self) -> Mapping[str, Optional[int]]:
-        # None means unlimited
-        return {"image": None}
-
-    def get_mm_max_tokens_per_item(
-        self,
-        seq_len: int,
-        mm_counts: Mapping[str, int],
-    ) -> Mapping[str, int]:
-        return {
-            "image":
-            self.get_num_image_tokens(
-                image_width=MaxImageTokenMeta.width,
-                image_height=MaxImageTokenMeta.height,
-            ),
-        }
-
     @property
     def image_token_id(self) -> int:
         return self.get_hf_config().media_placeholder_token_id
 
 
 class KimiVLDummyInputsBuilder(BaseDummyInputsBuilder[KimiVLProcessingInfo]):
 
-    def __init__(self, info: KimiVLProcessingInfo) -> None:
-        super().__init__(info)
+    def get_dummy_text(self, mm_counts: Mapping[str, int]) -> str:
+        num_images = mm_counts.get("image", 0)
+
+        processor = self.info.get_hf_processor()
+        image_token = processor.image_token
 
-        self.image_token_id = self.info.image_token_id
-        self.image_token = self.info.get_tokenizer().decode(
-            self.image_token_id)
+        return image_token * num_images
 
-    def get_dummy_processor_inputs(
+    def get_dummy_mm_data(
         self,
         seq_len: int,
         mm_counts: Mapping[str, int],
-    ) -> ProcessorInputs:
+    ) -> MultiModalDataDict:
         num_images = mm_counts.get("image", 0)
 
-        width = MaxImageTokenMeta.width
-        height = MaxImageTokenMeta.height
-        mm_data = {
+        return {
             "image":
-            self._get_dummy_images(width=width,
-                                   height=height,
+            self._get_dummy_images(width=MaxImageTokenMeta.width,
+                                   height=MaxImageTokenMeta.height,
                                    num_images=num_images)
         }
 
-        return ProcessorInputs(
-            prompt_text=self.image_token * num_images,
-            mm_data=mm_data,
-        )
-
 
 class KimiVLMultiModalProcessor(BaseMultiModalProcessor[KimiVLProcessingInfo]):