feat: add MRoPE op and Qwen2.5-VL skeleton

Author: MagellaX

python/mlc_llm/__init__.py

@@ -3,11 +3,35 @@
 MLC Chat is the app runtime of MLC LLM.
 """
 
-from tvm import register_func
+import logging
+import tvm
+
+if hasattr(tvm, "register_func"):
+    register_func = tvm.register_func  # type: ignore[attr-defined]
+else:  # pragma: no cover
+    from tvm_ffi.registry import register_global_func as register_func  # type: ignore
+
+    setattr(tvm, "register_func", register_func)
+
+AsyncMLCEngine = None  # type: ignore
+MLCEngine = None  # type: ignore
+
+try:
+    from . import protocol as protocol  # type: ignore
+except RuntimeError as err:  # pragma: no cover
+    logging.getLogger(__name__).debug("MLC-LLM protocol unavailable: %s", err)
+    protocol = None  # type: ignore
+
+try:
+    from . import serve as serve  # type: ignore
+except RuntimeError as err:  # pragma: no cover
+    logging.getLogger(__name__).debug("MLC-LLM serve unavailable: %s", err)
+    serve = None  # type: ignore
+else:
+    AsyncMLCEngine = serve.AsyncMLCEngine
+    MLCEngine = serve.MLCEngine
 
-from . import protocol, serve
 from .libinfo import __version__
-from .serve import AsyncMLCEngine, MLCEngine
 
 
 @register_func("runtime.disco.create_socket_session_local_workers", override=True)

python/mlc_llm/base.py

@@ -24,7 +24,7 @@ def _load_mlc_llm_lib():
     return ctypes.CDLL(lib_path[0]), lib_path[0]
 
 
-@tvm.register_func("mlc.debug_cuda_profiler_start")
+@tvm.register_func("mlc.debug_cuda_profiler_start", override=True)
 def _debug_cuda_profiler_start() -> None:
     """Start cuda profiler."""
     import cuda  # pylint: disable=import-outside-toplevel
@@ -33,7 +33,7 @@ def _debug_cuda_profiler_start() -> None:
     cuda.cudart.cudaProfilerStart()  # pylint: disable=c-extension-no-member
 
 
-@tvm.register_func("mlc.debug_cuda_profiler_stop")
+@tvm.register_func("mlc.debug_cuda_profiler_stop", override=True)
 def _debug_cuda_profiler_stop() -> None:
     """Stop cuda profiler."""
     import cuda  # pylint: disable=import-outside-toplevel

python/mlc_llm/model/qwen2_5_vl/__init__.py

@@ -0,0 +1 @@
+"\"\"\"Qwen2.5-VL architecture entry.\"\"\"\n+\n+from .qwen2_5_vl_model import (  # noqa: F401\n+    Qwen25VLConfig,\n+    Qwen25VLLMHeadModel,\n+)\n*** End Patch"/>