Fix/minicpmv46 missing version (#44980)

Signed-off-by: wjinxu <1299461899@qq.com>
Co-authored-by: Cursor <cursoragent@cursor.com>

vllm/model_executor/models/minicpmv4_6.py

@@ -424,6 +424,26 @@ class MiniCPMV4_6ProcessingInfo(MiniCPMVProcessingInfo):
     def get_hf_config(self):
         return self.ctx.get_hf_config()
 
+    def get_hf_processor(self, **kwargs: object):
+        # MiniCPM-V 4.6 keeps the native transformers MiniCPMV4_6Processor:
+        # this model has its own image/video handling and prompt-update logic
+        # below, so it does not need (and is incompatible with) the vendored
+        # MiniCPMVProcessor used by 2.x/4.0/4.5, whose __init__ assumes a
+        # legacy `image_processor.version` attribute that 4.6 no longer has.
+        hf_processor = self.ctx.get_hf_processor(**kwargs)
+
+        # NumPy arrays are considered as Iterable but not Sequence in
+        # https://github.com/huggingface/transformers/blob/main/src/transformers/image_transforms.py#L428
+        image_processor = getattr(hf_processor, "image_processor", None)
+        if image_processor is not None:
+            # transformers v5+ renamed `mean`/`std` -> `image_mean`/`image_std`
+            for attr in ("mean", "std", "image_mean", "image_std"):
+                val = getattr(image_processor, attr, None)
+                if isinstance(val, np.ndarray):
+                    setattr(image_processor, attr, val.tolist())
+
+        return hf_processor
+
     def _get_expected_hidden_size(self) -> int:
         config = self.get_hf_config()
         if hasattr(config, "text_config") and config.text_config is not None:

vllm/transformers_utils/processors/minicpmo.py

@@ -64,7 +64,12 @@ class MiniCPMOProcessor(ProcessorMixin):
         pool_step=2,
     ):
         super().__init__(image_processor, feature_extractor, tokenizer)
-        self.version = image_processor.version
+        # Mirror the MiniCPMVProcessor guard: newer (transformers v5.7+)
+        # MiniCPM image processors may drop the legacy `version` attribute,
+        # so fall back to None instead of hard-crashing. `version` only
+        # special-cases the 2.5 tokenization path; other values take the
+        # default branch.
+        self.version = getattr(image_processor, "version", None)
         self.pool_step = pool_step
 
     def _safe_get_token_id(self, attr_name, default_token_str):

vllm/transformers_utils/processors/minicpmv.py

@@ -58,7 +58,12 @@ class MiniCPMVProcessor(ProcessorMixin):
 
     def __init__(self, image_processor=None, tokenizer=None):
         super().__init__(image_processor, tokenizer)
-        self.version = image_processor.version
+        # Newer (transformers v5.7+) MiniCPM-V image processors, e.g.
+        # MiniCPMV4_6ImageProcessor, no longer carry a `version` attribute.
+        # Fall back to None instead of hard-crashing: `version` is only used
+        # to special-case the 2.5 tokenization path in `_convert`, and any
+        # value other than 2.5 takes the default branch anyway.
+        self.version = getattr(image_processor, "version", None)
 
     def __call__(
         self,