Jacob pr review

Author: jackzhxng

exir/emit/_emitter.py

@@ -1575,8 +1575,8 @@ def _find_fqn_for_placeholder(
                     warnings.warn(
                         "Mutation on a buffer in the model is detected. ExecuTorch assumes "
                         "buffers that are mutated in the graph have a meaningless initial state, "
-                        "only the shape and dtype will be serialized, unless a pass which marks "
-                        "spec.const=True such as InitializedMutableBufferPass is run.",
+                        "only the shape and dtype will be serialized, unless a pass which sets "
+                        'meta["et_init_buffer"] to True such as InitializedMutableBufferPass is run.',
                         UserWarning,
                         stacklevel=1,
                     )

exir/passes/init_mutable_pass.py

@@ -12,11 +12,12 @@
 
 class InitializedMutableBufferPass(ExportPass):
     """
-    If the buffer has the name "cache_pos", such as in an kv_cache
-    module with `self.register_buffer("cache_pos", torch.arange(10))`,
-    mark it with a custom tag which later is used by the emitter to
-    flag spec.const to True, which provides the mutable buffer with
-    an initialized state.
+    If a buffer has a name that within a specified list, set meta["et_init_buffer"]
+    to True, which provides the mutable buffer with an initialized state.
+
+    As an example, a module with `self.register_buffer("cache_pos", torch.arange(10))`
+    when patterns = ["cache_pos"] would have its initial state set instead of being
+    left uninitialized by default.
     """
 
     def __init__(self, patterns: List[str]) -> None: