[CuTeDSL] Let cute.compile opt into the compile cache

python/CuTeDSL/cutlass/base_dsl/compiler.py

@@ -1300,6 +1300,9 @@ def _compile(self, func: Any, *args: Any, **kwargs: Any) -> Any:
             the same token string as ``CUTE_DSL_COMPILER_OPT``. For the
             full option catalog, see
             ``write-kernel/references/compiler-options.md``.
+            ``no_cache=False`` opts this compile into the content-addressed
+            compile cache; it defaults to ``True`` because compile-only
+            results are not cached by default.
         :return: A compiled callable.
         :raises DSLRuntimeError: If ``func`` is not callable or not
             decorated with ``@cute.jit``.
@@ -1316,7 +1319,11 @@ def _compile(self, func: Any, *args: Any, **kwargs: Any) -> Any:
         finalize_hook = kwargs.pop("trace_finalize_hooks", None)
 
         kwargs["compile_only"] = True
-        kwargs["no_cache"] = True
+        # Default to bypassing the cache, but let an explicit `no_cache=False` from
+        # the caller opt back in. This used to be an unconditional assignment, which
+        # meant `cute.compile` could never reach the content-addressed compile cache
+        # and re-ran the full MLIR build even for a byte-identical artifact.
+        kwargs.setdefault("no_cache", True)
 
         if inspect.isfunction(func):
             # regular function

python/CuTeDSL/cutlass/base_dsl/dsl.py

@@ -2591,7 +2591,12 @@ def _prepare_compilation(
 
         pipeline = kwargs.pop("pipeline", None)
         gpu_module_attrs = kwargs.pop("gpu_module_attrs", {})
-        no_cache = kwargs.pop("no_cache", False)
+        # `None` means the caller expressed no preference, so the conservative
+        # defaults below apply. An explicit value is treated as an opt-in and is
+        # not overridden by the compile-only rule.
+        no_cache_arg = kwargs.pop("no_cache", None)
+        no_cache_explicit = no_cache_arg is not None
+        no_cache = bool(no_cache_arg)
         no_jit_engine = kwargs.pop("no_jit_engine", False)
         compile_only = kwargs.pop("compile_only", False)
 
@@ -2608,7 +2613,7 @@ def _prepare_compilation(
             no_cache = True
             self.print_warning("Cache is disabled as user wants to generate PTX/ASM.")
 
-        if not no_cache and compile_only:
+        if not no_cache and compile_only and not no_cache_explicit:
             no_cache = True
             self.print_warning("Cache is disabled as user wants to compile only.")