fix(sglang): stop forwarding caller mm_hashes to Kimi image processors

components/src/dynamo/sglang/_compat.py

@@ -393,6 +393,12 @@ def supports_external_mm_hashes(engine: Any) -> bool:
         return False
 
     tokenizer_manager = getattr(engine, "tokenizer_manager", None)
+    # Kimi K2.5 and K3 derive each image identity from the image bytes and
+    # reject caller hashes with a ValueError (sgl-project/sglang#43043).
+    mm_processor = getattr(tokenizer_manager, "mm_processor", None)
+    if getattr(mm_processor, "uses_wide_image_identity", False):
+        return False
+
     tokenize_one = getattr(tokenizer_manager, "_tokenize_one_request", None)
     if tokenize_one is None or getattr(
         tokenize_one, "_dynamo_rebuilds_external_mm_padding", False

components/src/dynamo/sglang/tests/test_sglang_unit.py

@@ -1112,6 +1112,24 @@ async def async_generate(self, mm_hashes=None):
     assert engine.tokenizer_manager._tokenize_one_request is patched
 
 
+def test_external_mm_hashes_are_disabled_when_processor_owns_image_identity():
+    class TokenizerManager:
+        mm_processor = SimpleNamespace(uses_wide_image_identity=True)
+
+        async def _tokenize_one_request(self, obj):
+            return obj
+
+    class KimiEngine:
+        tokenizer_manager = TokenizerManager()
+
+        async def async_generate(self, mm_hashes=None):
+            return None
+
+    engine = KimiEngine()
+    assert not supports_external_mm_hashes(engine)
+    assert "_tokenize_one_request" not in vars(engine.tokenizer_manager)
+
+
 def test_external_mm_hashes_are_disabled_when_engine_lacks_kwarg():
     class LegacyEngine:
         async def async_generate(self):