Switch hunyuan-ocr.py to stable vLLM (>=0.15.1), drop nightly index

HunyuanOCR architecture is in stable vLLM. Also pins datasets>=4.0.0
and removes disable_mm_preprocessor_cache (nightly-only arg).

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

hunyuan-ocr.py

@@ -1,20 +1,14 @@
 # /// script
 # requires-python = ">=3.11"
 # dependencies = [
-#     "datasets",
+#     "datasets>=4.0.0",
 #     "huggingface-hub",
 #     "pillow",
-#     "vllm",
+#     "vllm>=0.15.1",
 #     "tqdm",
 #     "toolz",
 #     "torch",
 # ]
-#
-# [[tool.uv.index]]
-# url = "https://wheels.vllm.ai/nightly/cu129"
-#
-# [tool.uv]
-# prerelease = "allow"
 # ///
 
 """
@@ -426,14 +420,13 @@ def main(
     logger.info("This may take a few minutes on first run...")
 
     # Note: HunyuanOCR has batching issues with vLLM V1 engine when batch_size > 1
-    # Using disable_mm_preprocessor_cache and limit_mm_per_prompt for stability
+    # Using limit_mm_per_prompt for stability
     llm = LLM(
         model=model,
         trust_remote_code=True,
         max_model_len=max_model_len,
         gpu_memory_utilization=gpu_memory_utilization,
         limit_mm_per_prompt={"image": 1},
-        disable_mm_preprocessor_cache=True,
         enable_prefix_caching=False,
     )