fix Metal codegen: handle shared.dyn scope, remove tvm-ffi<0.1.8 pin

Author: oraluben

pyproject.toml

@@ -30,9 +30,7 @@ dependencies = [
     # >=0.1.6 fixes a memory issue: tilelang#1502, but keep
     # requirement as wide as possible to be compatible with other libraries
     # pip will try to use latest version whenever possible.
-    "apache-tvm-ffi~=0.1.0,>0.1.2",
-    # apache/tvm-ffi#464
-    "apache-tvm-ffi<0.1.8; platform_system == 'Darwin'",
+    "apache-tvm-ffi~=0.1.0,>=0.1.6",
     # torch-c-dlpack-ext provides prebuilt torch extensions.
     # Without it, TVM FFI may require JIT compilation on first import.
     "torch-c-dlpack-ext; python_version < '3.14'",

requirements.txt

@@ -1,7 +1,6 @@
 # Runtime requirements
 
 apache-tvm-ffi~=0.1.0,>=0.1.6
-apache-tvm-ffi<0.1.8; platform_system == 'Darwin'
 torch-c-dlpack-ext; python_version < '3.14'
 cloudpickle
 ml-dtypes

src/target/codegen_metal.cc

@@ -314,7 +314,7 @@ void CodeGenMetal::PrintStorageScope(const std::string &scope,
                                      std::ostream &os) { // NOLINT(*)
   if (scope == "global") {
     os << "device ";
-  } else if (scope == "shared") {
+  } else if (scope == "shared" || scope == "shared.dyn") {
     os << "threadgroup ";
   } else if (scope == "local") {
     os << "thread ";