diff --git a/tests/rl/unit_tests/cpu/test_generator.py b/tests/rl/unit_tests/cpu/test_generator.py
index 3d973912d6f..3caf039fd40 100644
--- a/tests/rl/unit_tests/cpu/test_generator.py
+++ b/tests/rl/unit_tests/cpu/test_generator.py
@@ -576,8 +576,8 @@ def test_qwen36_27b_perf_config():
     config = rl_grpo_qwen3_6_27b_varlen_perf()
 
     assert config.hf_assets_path.endswith("Qwen3.6-27B")
-    assert "offset_rmsnorm" in config.trainer.model.local_compile_regions
-    assert "offset_rmsnorm" in config.generator.model.local_compile_regions
+    assert "offset_rmsnorm" in config.trainer.model.local_compile_regions.regions
+    assert "offset_rmsnorm" in config.generator.model.local_compile_regions.regions
     assert config.trainer.parallelism.data_parallel_shard_degree == 2
     assert config.trainer.parallelism.tensor_parallel_degree == 2
     assert config.generator.parallelism.tensor_parallel_degree == 4
diff --git a/tests/rl/unit_tests/cpu/test_policy_loss_configs.py b/tests/rl/unit_tests/cpu/test_policy_loss_configs.py
index b79c8ef600b..facd59e2261 100644
--- a/tests/rl/unit_tests/cpu/test_policy_loss_configs.py
+++ b/tests/rl/unit_tests/cpu/test_policy_loss_configs.py
@@ -75,5 +75,5 @@ def test_batch_invariant_configs_disable_local_compile(module_name, factory_name
     factory = getattr(importlib.import_module(module_name), factory_name)
     config = factory()
 
-    assert config.trainer.model.local_compile_regions == []
-    assert config.generator.model.local_compile_regions == []
+    assert config.trainer.model.local_compile_regions.regions == []
+    assert config.generator.model.local_compile_regions.regions == []
diff --git a/tests/rl/unit_tests/gpu/test_gdn_cudagraph.py b/tests/rl/unit_tests/gpu/test_gdn_cudagraph.py
index 176cbc6c46e..3c3b7b56dfa 100644
--- a/tests/rl/unit_tests/gpu/test_gdn_cudagraph.py
+++ b/tests/rl/unit_tests/gpu/test_gdn_cudagraph.py
@@ -301,7 +301,7 @@ def recurrent(*args, **kwargs):
     model_config = build_model_config(
         "0.8B", seq_len=256, attn_backend="varlen", parallelism=ParallelismConfig()
     )
-    model_config.local_compile_regions = []
+    model_config.local_compile_regions.regions = []
     registry.register_to_vllm(
         model_config,
         parallelism=registry.InferenceParallelismConfig(tensor_parallel_degree=1),
diff --git a/tests/unit_tests/cpu/config/test_compile_config.py b/tests/unit_tests/cpu/config/test_compile_config.py
index 4cbcf0ce629..f4c697d761f 100644
--- a/tests/unit_tests/cpu/config/test_compile_config.py
+++ b/tests/unit_tests/cpu/config/test_compile_config.py
@@ -4,24 +4,46 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
+from collections.abc import Callable
+from typing import Any
+
 import pytest
 import torch
+from torch.nn.attention.flex_attention import create_block_mask, flex_attention
 
 from torchtitan.components.loss import compute_logprobs, cross_entropy_loss, mse_loss
-from torchtitan.distributed.local_compile import apply_local_compile, local_compile
+from torchtitan.distributed.local_compile import local_compile, LocalCompileConfig
+from torchtitan.models.common.attention import FlexInnerAttention
+from torchtitan.models.common.attention.attention import _compiled_create_block_mask
 from torchtitan.models.common.moe import _microbatch_load_balance_local_stats
+from torchtitan.models.common.vision_encoder import compiled_create_block_mask
+
+
+class CustomDeviceLocalCompileConfig(LocalCompileConfig):
+    """Example configuration supplied by an out-of-tree backend."""
+
+    def _compile(
+        self,
+        reference: Callable[..., Any],
+        *,
+        inductor_options: dict[str, Any] | None = None,
+        **compile_kwargs: Any,
+    ) -> Callable[..., Any]:
+        del inductor_options
+        compile_kwargs["backend"] = "custom_device_backend"
+        return torch.compile(reference, **compile_kwargs)
 
 
 @pytest.fixture(autouse=True)
 def reset_local_compile():
-    apply_local_compile([])
+    LocalCompileConfig().apply()
     yield
-    apply_local_compile([])
+    LocalCompileConfig().apply()
 
 
-def test_apply_local_compile_rejects_unknown_name() -> None:
+def test_local_compile_config_rejects_unknown_name() -> None:
     with pytest.raises(ValueError, match=r"foo.*registered values"):
-        apply_local_compile(["foo"])
+        LocalCompileConfig(regions=["foo"]).apply()
 
 
 def test_local_compile_binds_once_from_existing_config(monkeypatch) -> None:
@@ -40,12 +62,13 @@ def compiled(value: int) -> tuple[str, int]:
         return compiled
 
     monkeypatch.setattr(torch, "compile", fake_compile)
-    apply_local_compile(["test_local_compile"])
+    LocalCompileConfig(regions=["test_local_compile"]).apply()
 
     assert fn(3) == ("compiled", 3)
-    assert len(compiled_calls) == 1
-    assert "backend" not in compiled_calls[0][1]
-    assert compiled_calls[0][1]["fullgraph"] is True
+    named_calls = [call for call in compiled_calls if call[0] is fn.__wrapped__]
+    assert len(named_calls) == 1
+    assert "backend" not in named_calls[0][1]
+    assert named_calls[0][1]["fullgraph"] is True
 
 
 def test_local_compile_forwards_compile_kwargs(monkeypatch) -> None:
@@ -68,10 +91,11 @@ def fake_compile(reference, **kwargs):
         return reference
 
     monkeypatch.setattr(torch, "compile", fake_compile)
-    apply_local_compile(["test_compile_kwargs"])
+    LocalCompileConfig(regions=["test_compile_kwargs"]).apply()
 
     assert fn(3) == 3
-    assert compiled_calls[0][1] == {
+    named_calls = [call for call in compiled_calls if call[0] is fn.__wrapped__]
+    assert named_calls[0][1] == {
         "backend": "eager",
         "fullgraph": True,
         "dynamic": True,
@@ -100,7 +124,7 @@ def fn(value: int) -> int:
     with pytest.raises(
         ValueError, match=r"test_non_batch_invariant.*local_compile_regions"
     ):
-        apply_local_compile(["test_non_batch_invariant"])
+        LocalCompileConfig(regions=["test_non_batch_invariant"]).apply()
 
 
 def test_loss_functions_use_local_compile(monkeypatch) -> None:
@@ -112,9 +136,18 @@ def fake_compile(reference, **kwargs):
         return reference
 
     monkeypatch.setattr(torch, "compile", fake_compile)
-    apply_local_compile(["loss"])
-
-    assert compiled_names == [
+    LocalCompileConfig(regions=["loss"]).apply()
+
+    assert [
+        name
+        for name in compiled_names
+        if name
+        in {
+            cross_entropy_loss.__name__,
+            mse_loss.__name__,
+            compute_logprobs.__name__,
+        }
+    ] == [
         cross_entropy_loss.__name__,
         mse_loss.__name__,
         compute_logprobs.__name__,
@@ -129,11 +162,41 @@ def fake_compile(reference, **kwargs):
         return reference
 
     monkeypatch.setattr(torch, "compile", fake_compile)
-    apply_local_compile(["moe_aux_loss"])
+    LocalCompileConfig(regions=["moe_aux_loss"]).apply()
 
-    assert compiled == [
+    assert [
+        call
+        for call in compiled
+        if call[0] is _microbatch_load_balance_local_stats.__wrapped__
+    ] == [
         (
             _microbatch_load_balance_local_stats.__wrapped__,
             {"fullgraph": True},
         )
     ]
+
+
+def test_custom_config_controls_all_torchtitan_compile_boundaries(
+    monkeypatch,
+) -> None:
+    compiled_calls = []
+
+    @local_compile("test_custom_backend", batch_invariant=False)
+    def region() -> None:
+        pass
+
+    def fake_compile(reference, **kwargs):
+        compiled_calls.append((reference, kwargs))
+        return reference
+
+    monkeypatch.setattr(torch, "compile", fake_compile)
+    CustomDeviceLocalCompileConfig(regions=["test_custom_backend"]).apply()
+
+    assert callable(FlexInnerAttention._compiled_flex_attn)
+    assert callable(_compiled_create_block_mask)
+    assert callable(compiled_create_block_mask)
+    expected = {"backend": "custom_device_backend"}
+    assert (region.__wrapped__, {**expected, "fullgraph": True}) in compiled_calls
+    assert (flex_attention, expected) in compiled_calls
+    assert (create_block_mask, expected) in compiled_calls
+    assert sum(reference is create_block_mask for reference, _ in compiled_calls) == 2
diff --git a/tests/unit_tests/cpu/config/test_context_parallel_validation.py b/tests/unit_tests/cpu/config/test_context_parallel_validation.py
index 0575d338eb8..0a1ced4cf88 100644
--- a/tests/unit_tests/cpu/config/test_context_parallel_validation.py
+++ b/tests/unit_tests/cpu/config/test_context_parallel_validation.py
@@ -186,6 +186,7 @@ class TestGptOssUlysses(unittest.TestCase):
     def _parallelize(self, inner_attention):
         from types import SimpleNamespace
 
+        from torchtitan.distributed.local_compile import LocalCompileConfig
         from torchtitan.models.gpt_oss.model import GptOssModel
 
         with self.assertRaisesRegex(NotImplementedError, "Ulysses CP"):
@@ -198,7 +199,7 @@ def _parallelize(self, inner_attention):
                 parallelism_context=SimpleNamespace(cp_enabled=True),
                 training=None,
                 parallelism=None,
-                local_compile_regions=[],
+                local_compile_regions=LocalCompileConfig(),
                 activation_remat=None,
                 dump_folder="",
             )
diff --git a/tests/unit_tests/cpu/models/common/test_gated_rmsnorm.py b/tests/unit_tests/cpu/models/common/test_gated_rmsnorm.py
index 4d1ecbae692..7fd0f24639e 100644
--- a/tests/unit_tests/cpu/models/common/test_gated_rmsnorm.py
+++ b/tests/unit_tests/cpu/models/common/test_gated_rmsnorm.py
@@ -9,13 +9,13 @@
 import torch
 import torch.nn.functional as F
 
-from torchtitan.distributed.local_compile import apply_local_compile
+from torchtitan.distributed.local_compile import LocalCompileConfig
 from torchtitan.models.common import GatedRMSNorm, Sigmoid, SiLU
 
 
 class TestGatedRMSNorm(unittest.TestCase):
     def setUp(self):
-        apply_local_compile([])
+        LocalCompileConfig().apply()
 
     def test_configurable_activation_matches_reference(self):
         x = torch.randn(4, 3, 8)
diff --git a/tests/unit_tests/cpu/models/common/test_hi_mid_lo_linear.py b/tests/unit_tests/cpu/models/common/test_hi_mid_lo_linear.py
index c2664043867..490eee0b823 100644
--- a/tests/unit_tests/cpu/models/common/test_hi_mid_lo_linear.py
+++ b/tests/unit_tests/cpu/models/common/test_hi_mid_lo_linear.py
@@ -215,7 +215,7 @@ def test_lm_head_converter_swaps_only_lm_head():
     # The head keeps 2 pieces: summed over the vocab, a third doesn't help.
     assert config.lm_head.backward_mode == "hi_mid"
     # It turns on the split's local_compile region.
-    assert "fp32_to_bf16_split" in config.local_compile_regions
+    assert "fp32_to_bf16_split" in config.local_compile_regions.regions
     # Converting an already converted head is a no-op.
     converted = config.lm_head
     LMHeadFP32OutputConverter.Config().build().convert(config)
diff --git a/tests/unit_tests/cpu/models/kimi_k3/test_kimi_k3_config.py b/tests/unit_tests/cpu/models/kimi_k3/test_kimi_k3_config.py
index e22edb1d94b..f8f3d452320 100644
--- a/tests/unit_tests/cpu/models/kimi_k3/test_kimi_k3_config.py
+++ b/tests/unit_tests/cpu/models/kimi_k3/test_kimi_k3_config.py
@@ -21,7 +21,7 @@ def test_kimi_k3_enables_attention_residual_local_compile() -> None:
     config = kimi_k3_debugmodel()
     model_config = cast(KimiK3Model.Config, config.model)
 
-    assert model_config.local_compile_regions == [
+    assert model_config.local_compile_regions.regions == [
         "loss",
         "gated_rmsnorm",
         "fused_binary_activation",
diff --git a/tests/unit_tests/cpu/models/qwen3_5/test_offset_rmsnorm.py b/tests/unit_tests/cpu/models/qwen3_5/test_offset_rmsnorm.py
index b43336cb5ad..04f095b348f 100644
--- a/tests/unit_tests/cpu/models/qwen3_5/test_offset_rmsnorm.py
+++ b/tests/unit_tests/cpu/models/qwen3_5/test_offset_rmsnorm.py
@@ -12,7 +12,7 @@
 def test_qwen35_architecture_enables_offset_rmsnorm_local_compile() -> None:
     for config_factory in (qwen35_debugmodel, qwen36_debugmodel, qwen38_debugmodel):
         config = config_factory()
-        assert config.model.local_compile_regions == [
+        assert config.model.local_compile_regions.regions == [
             "loss",
             "fused_binary_activation",
             "gated_rmsnorm",
diff --git a/tests/unit_tests/cpu/quantization/test_quantization.py b/tests/unit_tests/cpu/quantization/test_quantization.py
index 61603ae7ced..f304707c13f 100644
--- a/tests/unit_tests/cpu/quantization/test_quantization.py
+++ b/tests/unit_tests/cpu/quantization/test_quantization.py
@@ -479,7 +479,7 @@ def test_qwen3_recipes_resolve(monkeypatch, recipe):
         assert isinstance(dataset.source, HuggingFaceRandomAccessSource.Config)
         assert dataset.source.path == "openai/gsm8k"
         assert config.checkpointer.initial_load_in_hf
-        assert config.model.local_compile_regions == [
+        assert config.model.local_compile_regions.regions == [
             "loss",
             "fused_binary_activation",
             "cos_sin_rope",
@@ -987,7 +987,7 @@ def test_deepseek_nvfp4_recipes_preserve_quantization_and_routing(
         if bf16_tail_fraction == 0
         else len(config.model.layers) // 2
     )
-    assert config.model.local_compile_regions == ["loss"]
+    assert config.model.local_compile_regions.regions == ["loss"]
     grouped = list(config.model.traverse(GroupedLinear.Config))
     assert grouped
     quantized_cls = _get_nvfp4_grouped_linear_cls(GroupedLinear)
diff --git a/tests/unit_tests/cpu/test_integration_test_definitions.py b/tests/unit_tests/cpu/test_integration_test_definitions.py
index 0eb38f0db00..0235821f2b6 100644
--- a/tests/unit_tests/cpu/test_integration_test_definitions.py
+++ b/tests/unit_tests/cpu/test_integration_test_definitions.py
@@ -54,7 +54,7 @@ def test_spmd_typechecking_config_disables_local_compile() -> None:
     config = llama3_debugmodel_default()
 
     assert config.debug.spmd_typechecking
-    assert config.model.local_compile_regions == []
+    assert config.model.local_compile_regions.regions == []
     config.__post_init__()
 
 
diff --git a/tests/unit_tests/cpu/test_trainer.py b/tests/unit_tests/cpu/test_trainer.py
index c79f1b0798f..26b7349b019 100644
--- a/tests/unit_tests/cpu/test_trainer.py
+++ b/tests/unit_tests/cpu/test_trainer.py
@@ -756,7 +756,9 @@ def test_graph_training_engine_rejects_local_compile_regions() -> None:
     ):
         GraphTrainingEngine(
             config,
-            model_config=SimpleNamespace(local_compile_regions=["loss"]),
+            model_config=SimpleNamespace(
+                local_compile_regions=SimpleNamespace(regions=["loss"])
+            ),
             max_num_documents=None,
             output_dir="",
         )
diff --git a/tests/unit_tests/gpu/models/common/test_gated_rmsnorm.py b/tests/unit_tests/gpu/models/common/test_gated_rmsnorm.py
index 3d1604aa825..3badec1bc25 100644
--- a/tests/unit_tests/gpu/models/common/test_gated_rmsnorm.py
+++ b/tests/unit_tests/gpu/models/common/test_gated_rmsnorm.py
@@ -8,17 +8,17 @@
 
 import torch
 
-from torchtitan.distributed.local_compile import apply_local_compile
+from torchtitan.distributed.local_compile import LocalCompileConfig
 from torchtitan.models.common import GatedRMSNorm, Sigmoid
 
 
 @unittest.skipUnless(torch.cuda.is_available(), "CUDA required")
 class TestGatedRMSNormCompile(unittest.TestCase):
     def setUp(self):
-        apply_local_compile(["gated_rmsnorm"])
+        LocalCompileConfig(regions=["gated_rmsnorm"]).apply()
 
     def tearDown(self):
-        apply_local_compile([])
+        LocalCompileConfig().apply()
         torch._dynamo.reset()
 
     @staticmethod
diff --git a/tests/unit_tests/gpu/models/common/test_hi_mid_lo_linear.py b/tests/unit_tests/gpu/models/common/test_hi_mid_lo_linear.py
index 1cd17b19b6b..2821b03c20c 100644
--- a/tests/unit_tests/gpu/models/common/test_hi_mid_lo_linear.py
+++ b/tests/unit_tests/gpu/models/common/test_hi_mid_lo_linear.py
@@ -18,7 +18,7 @@
 from torch.distributed.fsdp import fully_shard, MixedPrecisionPolicy
 from torch.fx.experimental.proxy_tensor import make_fx
 
-from torchtitan.distributed.local_compile import apply_local_compile
+from torchtitan.distributed.local_compile import LocalCompileConfig
 from torchtitan.models.common import hi_mid_lo_linear
 from torchtitan.models.common.hi_mid_lo_linear import HiMidLoLinear
 
@@ -177,11 +177,11 @@ def linear(input, weight):
 @contextlib.contextmanager
 def _compiled_split():
     """Turn on the split's local_compile region, as the models that use HiMidLoLinear do."""
-    apply_local_compile(["fp32_to_bf16_split"])
+    LocalCompileConfig(regions=["fp32_to_bf16_split"]).apply()
     try:
         yield
     finally:
-        apply_local_compile([])
+        LocalCompileConfig().apply()
 
 
 @pytest.mark.parametrize("dim", [0, 1])
diff --git a/tests/unit_tests/gpu/models/common/test_moe_aux_loss_compile.py b/tests/unit_tests/gpu/models/common/test_moe_aux_loss_compile.py
index b5c8145489f..4482a1e710a 100644
--- a/tests/unit_tests/gpu/models/common/test_moe_aux_loss_compile.py
+++ b/tests/unit_tests/gpu/models/common/test_moe_aux_loss_compile.py
@@ -8,14 +8,14 @@
 
 import torch
 
-from torchtitan.distributed.local_compile import apply_local_compile
+from torchtitan.distributed.local_compile import LocalCompileConfig
 from torchtitan.models.common.moe import _microbatch_load_balance_local_stats
 
 
 @unittest.skipUnless(torch.cuda.is_available(), "CUDA required")
 class TestMoEAuxLossLocalCompile(unittest.TestCase):
     def tearDown(self):
-        apply_local_compile([])
+        LocalCompileConfig().apply()
         torch._dynamo.reset()
 
     def test_forward_and_backward_match_eager(self):
@@ -43,7 +43,7 @@ def test_forward_and_backward_match_eager(self):
             generator=generator,
         )
 
-        apply_local_compile([])
+        LocalCompileConfig().apply()
         eager_counts_E, eager_prob_sums_E = _microbatch_load_balance_local_stats(
             scores_TE,
             routing_map_TE,
@@ -53,7 +53,7 @@ def test_forward_and_backward_match_eager(self):
             eager_prob_sums_E, scores_TE, grad_prob_sums_E
         )[0]
 
-        apply_local_compile(["moe_aux_loss"])
+        LocalCompileConfig(regions=["moe_aux_loss"]).apply()
         compiled_counts_E, compiled_prob_sums_E = _microbatch_load_balance_local_stats(
             scores_TE,
             routing_map_TE,
diff --git a/tests/unit_tests/gpu/models/common/test_situglu.py b/tests/unit_tests/gpu/models/common/test_situglu.py
index 3bd6f42355a..a9b262c96f3 100644
--- a/tests/unit_tests/gpu/models/common/test_situglu.py
+++ b/tests/unit_tests/gpu/models/common/test_situglu.py
@@ -8,17 +8,17 @@
 
 import torch
 
-from torchtitan.distributed.local_compile import apply_local_compile
+from torchtitan.distributed.local_compile import LocalCompileConfig
 from torchtitan.models.common.activation import SiTUGLU
 
 
 @unittest.skipUnless(torch.cuda.is_available(), "CUDA required")
 class TestSiTUGLULocalCompile(unittest.TestCase):
     def setUp(self):
-        apply_local_compile(["fused_binary_activation"])
+        LocalCompileConfig(regions=["fused_binary_activation"]).apply()
 
     def tearDown(self):
-        apply_local_compile([])
+        LocalCompileConfig().apply()
         torch._dynamo.reset()
 
     def test_forward_and_backward_emit_triton(self):
diff --git a/tests/unit_tests/gpu/models/common/test_swiglu.py b/tests/unit_tests/gpu/models/common/test_swiglu.py
index dde74bd4d75..85b09654cb8 100644
--- a/tests/unit_tests/gpu/models/common/test_swiglu.py
+++ b/tests/unit_tests/gpu/models/common/test_swiglu.py
@@ -9,17 +9,17 @@
 import torch
 import torch.nn.functional as F
 
-from torchtitan.distributed.local_compile import apply_local_compile
+from torchtitan.distributed.local_compile import LocalCompileConfig
 from torchtitan.models.common.activation import ClampedSwiGLU, SiTUGLU, SwiGLU
 
 
 @unittest.skipUnless(torch.cuda.is_available(), "CUDA required")
 class TestSwiGLULocalCompile(unittest.TestCase):
     def setUp(self):
-        apply_local_compile(["fused_binary_activation"])
+        LocalCompileConfig(regions=["fused_binary_activation"]).apply()
 
     def tearDown(self):
-        apply_local_compile([])
+        LocalCompileConfig().apply()
         torch._dynamo.reset()
 
     def test_forward_and_backward_emit_triton(self):
diff --git a/tests/unit_tests/gpu/models/kimi_k3/test_kimi_k3_attention_residual.py b/tests/unit_tests/gpu/models/kimi_k3/test_kimi_k3_attention_residual.py
index ea9ef9c9b20..45456e29bdb 100644
--- a/tests/unit_tests/gpu/models/kimi_k3/test_kimi_k3_attention_residual.py
+++ b/tests/unit_tests/gpu/models/kimi_k3/test_kimi_k3_attention_residual.py
@@ -8,7 +8,7 @@
 
 import torch
 
-from torchtitan.distributed.local_compile import apply_local_compile
+from torchtitan.distributed.local_compile import LocalCompileConfig
 from torchtitan.models.common import Linear
 from torchtitan.models.common.nn_modules import RMSNorm
 from torchtitan.models.kimi_k3.model import AttentionResidual
@@ -45,11 +45,11 @@ def _reference(
 @unittest.skipUnless(torch.cuda.is_available(), "CUDA required")
 class TestKimiK3AttentionResidualLocalCompile(unittest.TestCase):
     def setUp(self):
-        apply_local_compile(["attention_residual"])
+        LocalCompileConfig(regions=["attention_residual"]).apply()
         self.attention_residual = AttentionResidual.Config().build()
 
     def tearDown(self):
-        apply_local_compile([])
+        LocalCompileConfig().apply()
         torch._dynamo.reset()
 
     def test_forward_and_backward_emit_triton(self):
diff --git a/tests/unit_tests/gpu/models/qwen3_5/test_offset_rmsnorm.py b/tests/unit_tests/gpu/models/qwen3_5/test_offset_rmsnorm.py
index f39c6fb5ba6..c4a0fa1e1b1 100644
--- a/tests/unit_tests/gpu/models/qwen3_5/test_offset_rmsnorm.py
+++ b/tests/unit_tests/gpu/models/qwen3_5/test_offset_rmsnorm.py
@@ -13,7 +13,7 @@
     with_comms,
 )
 
-from torchtitan.distributed.local_compile import apply_local_compile
+from torchtitan.distributed.local_compile import LocalCompileConfig
 from torchtitan.distributed.parallelism_context import ParallelismContext
 from torchtitan.distributed.spmd_types import set_current_spmd_mesh, set_spmd_meshes
 from torchtitan.models.common.decoder_sharding import head_norm_sharding_config
@@ -23,10 +23,10 @@
 @unittest.skipUnless(torch.cuda.is_available(), "CUDA required")
 class TestOffsetRMSNormCompile(unittest.TestCase):
     def setUp(self):
-        apply_local_compile(["offset_rmsnorm"])
+        LocalCompileConfig(regions=["offset_rmsnorm"]).apply()
 
     def tearDown(self):
-        apply_local_compile([])
+        LocalCompileConfig().apply()
         torch._dynamo.reset()
 
     def test_forward_and_backward_emit_triton(self):
@@ -101,7 +101,7 @@ def world_size(self) -> int:
 
     @with_comms
     def test_compiled_qk_norm(self):
-        apply_local_compile(["offset_rmsnorm"])
+        LocalCompileConfig(regions=["offset_rmsnorm"]).apply()
         device = self.device_type
         config = OffsetRMSNorm.Config(
             dim=128,
@@ -137,7 +137,7 @@ def test_compiled_qk_norm(self):
             output.sum().backward()
 
         self.assertEqual(output.shape, x_local.shape)
-        apply_local_compile([])
+        LocalCompileConfig().apply()
         torch._dynamo.reset()
 
 
diff --git a/tests/unit_tests/gpu/models/qwen3_5/test_rope_compile.py b/tests/unit_tests/gpu/models/qwen3_5/test_rope_compile.py
index c7c91aa4ad8..d6fbe401791 100644
--- a/tests/unit_tests/gpu/models/qwen3_5/test_rope_compile.py
+++ b/tests/unit_tests/gpu/models/qwen3_5/test_rope_compile.py
@@ -8,7 +8,7 @@
 
 import torch
 
-from torchtitan.distributed.local_compile import apply_local_compile
+from torchtitan.distributed.local_compile import LocalCompileConfig
 from torchtitan.models.common.rope import CosSinRoPE
 from torchtitan.models.qwen3_5.rope import MRoPE
 
@@ -16,10 +16,10 @@
 @unittest.skipUnless(torch.cuda.is_available(), "CUDA required")
 class TestRoPELocalCompile(unittest.TestCase):
     def setUp(self):
-        apply_local_compile(["cos_sin_rope"])
+        LocalCompileConfig(regions=["cos_sin_rope"]).apply()
 
     def tearDown(self):
-        apply_local_compile([])
+        LocalCompileConfig().apply()
         torch._dynamo.reset()
 
     def test_forward_and_backward_emit_triton(self):
@@ -89,7 +89,7 @@ def test_forward_and_backward_match_eager(self):
             (grad_query_out, grad_key_out),
         )
 
-        apply_local_compile([])
+        LocalCompileConfig().apply()
         eager_rope = (
             CosSinRoPE.Config(
                 dim=128,
@@ -222,7 +222,7 @@ def test_mrope_three_axis_positions_match_eager(self):
             (grad_query_out, grad_key_out),
         )
 
-        apply_local_compile([])
+        LocalCompileConfig().apply()
         eager_rope = config.build().cuda()
         eager_query = query.detach().clone().requires_grad_()
         eager_key = key.detach().clone().requires_grad_()
diff --git a/tests/unit_tests/gpu/models/qwen3_5/test_sigmoid_gate.py b/tests/unit_tests/gpu/models/qwen3_5/test_sigmoid_gate.py
index 83b3ab74fca..8448defd201 100644
--- a/tests/unit_tests/gpu/models/qwen3_5/test_sigmoid_gate.py
+++ b/tests/unit_tests/gpu/models/qwen3_5/test_sigmoid_gate.py
@@ -8,14 +8,14 @@
 
 import torch
 
-from torchtitan.distributed.local_compile import apply_local_compile
+from torchtitan.distributed.local_compile import LocalCompileConfig
 from torchtitan.models.qwen3_5.moe import sigmoid_gate
 
 
 @unittest.skipUnless(torch.cuda.is_available(), "CUDA required")
 class TestSigmoidGateLocalCompile(unittest.TestCase):
     def setUp(self):
-        apply_local_compile(["shared_expert_gate"])
+        LocalCompileConfig(regions=["shared_expert_gate"]).apply()
         # A local generator leaves the global RNG stream unchanged for later tests.
         generator = torch.Generator(device="cuda").manual_seed(0)
 
@@ -29,7 +29,7 @@ def randn(*shape: int) -> torch.Tensor:
         self.grad_output = randn(64, 128)
 
     def tearDown(self):
-        apply_local_compile([])
+        LocalCompileConfig().apply()
         torch._dynamo.reset()
 
     def test_forward_and_backward_match_reference(self):
diff --git a/torchtitan/config/transform/lm_head_fp32.py b/torchtitan/config/transform/lm_head_fp32.py
index a7cea8aa0f4..10f2ca1db59 100644
--- a/torchtitan/config/transform/lm_head_fp32.py
+++ b/torchtitan/config/transform/lm_head_fp32.py
@@ -67,6 +67,6 @@ def convert(self, model_config):
                 "the model config. The torchtitan decoder names its output projection "
                 f"{self._TARGET!r} (see torchtitan/models/common/decoder.py)."
             )
-        if "fp32_to_bf16_split" not in model_config.local_compile_regions:
-            model_config.local_compile_regions.append("fp32_to_bf16_split")
+        if "fp32_to_bf16_split" not in model_config.local_compile_regions.regions:
+            model_config.local_compile_regions.regions.append("fp32_to_bf16_split")
         return model_config
diff --git a/torchtitan/config/validation.py b/torchtitan/config/validation.py
index 5bdaab9b22b..4e0b8352940 100644
--- a/torchtitan/config/validation.py
+++ b/torchtitan/config/validation.py
@@ -137,11 +137,12 @@ def validate_model_training_config(
                     f"dispatcher: {type(dispatcher_config).__qualname__}."
                 )
 
-    if debug.spmd_typechecking and model.local_compile_regions:
+    if debug.spmd_typechecking and model.local_compile_regions.regions:
         # TODO: Remove this once Dynamo supports tracing SPMD typechecking.
         raise ValueError(
             "Local compilation is not supported with SPMD typechecking. "
-            "Set model.local_compile_regions=[] or disable debug.spmd_typechecking."
+            "Set model.local_compile_regions.regions=[] or disable "
+            "debug.spmd_typechecking."
         )
 
     validate_context_parallel(model, parallelism)
diff --git a/torchtitan/distributed/local_compile.py b/torchtitan/distributed/local_compile.py
index 6774c21ebf6..2988d4c62d6 100644
--- a/torchtitan/distributed/local_compile.py
+++ b/torchtitan/distributed/local_compile.py
@@ -4,10 +4,11 @@
 # This source code is licensed under the BSD-style license found in the
 # LICENSE file in the root directory of this source tree.
 
-"""Function-scoped torch.compile registration."""
+"""Function-scoped torch.compile registration and configuration."""
 
 import functools
 from collections.abc import Callable
+from dataclasses import dataclass, field
 from typing import Any
 
 import torch
@@ -15,24 +16,47 @@
 from torchtitan.distributed.batch_invariant import is_in_batch_invariant_mode
 
 
-def apply_local_compile(regions: list[str]) -> None:
-    """Bind registered functions to eager or compiled implementations.
+_LOCAL_COMPILE_CALLBACKS: dict[str, list[Callable[["LocalCompileConfig"], None]]] = {}
+_ALWAYS_COMPILE_CALLBACKS: list[Callable[["LocalCompileConfig"], None]] = []
 
-    Process-wide: a later call replaces this choice for every model in the process.
-    """
-    unknown = [name for name in regions if name not in _LOCAL_COMPILE_CALLBACKS]
-    if unknown:
-        raise ValueError(
-            f"Unknown local_compile_regions entries {unknown}; "
-            f"registered values are {sorted(_LOCAL_COMPILE_CALLBACKS)}"
-        )
 
-    for callbacks in _LOCAL_COMPILE_CALLBACKS.values():
-        for bind_local_compile_fn in callbacks:
-            bind_local_compile_fn(regions)
+@dataclass(kw_only=True, slots=True)
+class LocalCompileConfig:
+    """Configuration for TorchTitan-owned local ``torch.compile`` boundaries."""
+
+    regions: list[str] = field(default_factory=list)
 
+    def _compile(
+        self,
+        reference: Callable[..., Any],
+        *,
+        inductor_options: dict[str, Any] | None = None,
+        **compile_kwargs: Any,
+    ) -> Callable[..., Any]:
+        """Compile one boundary with Inductor.
 
-_LOCAL_COMPILE_CALLBACKS: dict[str, list[Callable[[list[str]], None]]] = {}
+        Out-of-tree subclasses can override this hook to select another backend.
+        """
+        if inductor_options is not None:
+            compile_kwargs.setdefault("options", inductor_options)
+        return torch.compile(reference, **compile_kwargs)
+
+    def apply(self) -> None:
+        """Bind named regions and always-compiled helpers using this config."""
+        unknown = [
+            name for name in self.regions if name not in _LOCAL_COMPILE_CALLBACKS
+        ]
+        if unknown:
+            raise ValueError(
+                f"Unknown local_compile_regions entries {unknown}; "
+                f"registered values are {sorted(_LOCAL_COMPILE_CALLBACKS)}"
+            )
+
+        for callbacks in _LOCAL_COMPILE_CALLBACKS.values():
+            for bind_local_compile_fn in callbacks:
+                bind_local_compile_fn(self)
+        for bind_compile_fn in _ALWAYS_COMPILE_CALLBACKS:
+            bind_compile_fn(self)
 
 
 def local_compile(
@@ -56,9 +80,9 @@ def local_compile(
     def decorate(reference: Callable[..., Any]) -> Callable[..., Any]:
         fn = reference
 
-        def bind_local_compile(regions: list[str]) -> None:
+        def bind_local_compile(config: LocalCompileConfig) -> None:
             nonlocal fn
-            enabled = name in regions
+            enabled = name in config.regions
             batch_invariant_mode = is_in_batch_invariant_mode()
             if enabled and batch_invariant_mode and not batch_invariant:
                 raise ValueError(
@@ -66,7 +90,7 @@ def bind_local_compile(regions: list[str]) -> None:
                     "batch-invariant mode; remove it from local_compile_regions."
                 )
             if enabled:
-                fn = torch.compile(reference, fullgraph=True, **compile_kwargs)
+                fn = config._compile(reference, fullgraph=True, **compile_kwargs)
             else:
                 fn = reference
 
@@ -80,4 +104,37 @@ def wrapped(*args: Any, **kwargs: Any) -> Any:
     return decorate
 
 
-__all__ = ["apply_local_compile", "local_compile"]
+def compiled(
+    reference: Callable[..., Any],
+    *,
+    inductor_options: dict[str, Any] | None = None,
+    **compile_kwargs: Any,
+) -> Callable[..., Any]:
+    """Create an always-compiled helper controlled by ``LocalCompileConfig``."""
+    fn = LocalCompileConfig()._compile(
+        reference,
+        inductor_options=inductor_options,
+        **compile_kwargs,
+    )
+
+    def bind_compile(config: LocalCompileConfig) -> None:
+        nonlocal fn
+        fn = config._compile(
+            reference,
+            inductor_options=inductor_options,
+            **compile_kwargs,
+        )
+
+    @functools.wraps(reference)
+    def wrapped(*args: Any, **kwargs: Any) -> Any:
+        return fn(*args, **kwargs)
+
+    _ALWAYS_COMPILE_CALLBACKS.append(bind_compile)
+    return wrapped
+
+
+__all__ = [
+    "compiled",
+    "local_compile",
+    "LocalCompileConfig",
+]
diff --git a/torchtitan/distributed/pipeline_parallel.py b/torchtitan/distributed/pipeline_parallel.py
index 67959106c65..769d5503a62 100644
--- a/torchtitan/distributed/pipeline_parallel.py
+++ b/torchtitan/distributed/pipeline_parallel.py
@@ -32,6 +32,7 @@
 from torchtitan.config.parallelism import ParallelismConfig
 from torchtitan.distributed import ParallelismContext
 from torchtitan.distributed.activation_remat import ActivationRematConfig
+from torchtitan.distributed.local_compile import LocalCompileConfig
 from torchtitan.models.base.model import BaseModel
 from torchtitan.models.base.module import ModuleDict, ModuleList
 from torchtitan.models.common.decoder import Decoder
@@ -78,7 +79,7 @@ def pipeline_llm(
     parallelism_context: ParallelismContext,
     training: TrainingConfig,
     parallelism: ParallelismConfig,
-    local_compile_regions: list[str],
+    local_compile_regions: LocalCompileConfig,
     activation_remat: ActivationRematConfig | None,
     dump_folder: str,
     device: torch.device,
diff --git a/torchtitan/distributed/utils.py b/torchtitan/distributed/utils.py
index 77c4dea011e..df5e77e00d2 100644
--- a/torchtitan/distributed/utils.py
+++ b/torchtitan/distributed/utils.py
@@ -167,11 +167,6 @@ def set_determinism(
             # no options) so the regional-inductor scoop configs are preserved.
             FlexInnerAttention.inductor_configs["max_autotune"] = False
             FlexInnerAttention.inductor_configs["coordinate_descent_tuning"] = False
-            # pyrefly: ignore [no-matching-overload]
-            FlexInnerAttention._compiled_flex_attn = torch.compile(
-                flex_attention, options=FlexInnerAttention.inductor_configs
-            )
-
     if debug_config.detect_anomaly:
         logger.warning(
             "Anomaly detection enabled. This incurs significant overhead "
diff --git a/torchtitan/experiments/graph_trainer/configs.py b/torchtitan/experiments/graph_trainer/configs.py
index 62e7e36314a..a3c1fd9a3ef 100644
--- a/torchtitan/experiments/graph_trainer/configs.py
+++ b/torchtitan/experiments/graph_trainer/configs.py
@@ -9,6 +9,7 @@
 
 from torchtitan.components.loss import ChunkedLossWrapper
 from torchtitan.distributed.activation_remat import DefaultActivationRematConfig
+from torchtitan.distributed.local_compile import LocalCompileConfig
 from torchtitan.experiments.graph_trainer.chunked_loss import (
     ChunkedLossWrapperWithParamGrads,
 )
@@ -353,7 +354,7 @@ def to_graph_trainer_config(
         }
     )
     # GraphTrainer compiles the whole step (config.compile), so it drops the model's local compile regions.
-    graph_model.local_compile_regions = []
+    graph_model.local_compile_regions = LocalCompileConfig()
     disable_inplace_wgrad_accum(graph_model)
     d["model"] = graph_model
 
diff --git a/torchtitan/experiments/graph_trainer/trainer.py b/torchtitan/experiments/graph_trainer/trainer.py
index 1b7c6c8a91e..bd8133f22f3 100644
--- a/torchtitan/experiments/graph_trainer/trainer.py
+++ b/torchtitan/experiments/graph_trainer/trainer.py
@@ -66,11 +66,11 @@ def __init__(
     ) -> None:
         if config.optim.enable_cuda_graph:
             raise ValueError("Optim CUDA graphs are not supported with GraphTrainer.")
-        if model_config.local_compile_regions:
+        if model_config.local_compile_regions.regions:
             raise ValueError(
                 "GraphTrainer traces the whole step into one graph; set "
-                "model.local_compile_regions = [] "
-                f"(got {model_config.local_compile_regions})."
+                "model.local_compile_regions.regions = [] "
+                f"(got {model_config.local_compile_regions.regions})."
             )
         if config.parallelism.fsdp_symm_mem_scope is not None:
             raise ValueError(
@@ -359,7 +359,10 @@ class Config(Trainer.Config):
         compile: GraphTrainerCompileConfig = field(
             default_factory=GraphTrainerCompileConfig
         )
-        """Whole-step compile. GraphTrainer requires ``model.local_compile_regions`` to be empty."""
+        """Whole-step compile.
+
+        GraphTrainer requires ``model.local_compile_regions.regions`` to be empty.
+        """
 
         def __post_init__(self) -> None:
             Trainer.Config.__post_init__(self)
diff --git a/torchtitan/experiments/transformers_modeling_backend/model.py b/torchtitan/experiments/transformers_modeling_backend/model.py
index 6ec6c1080d2..6f135923a46 100644
--- a/torchtitan/experiments/transformers_modeling_backend/model.py
+++ b/torchtitan/experiments/transformers_modeling_backend/model.py
@@ -26,6 +26,7 @@
 from torchtitan.config import TORCH_DTYPE_MAP, TrainingConfig
 from torchtitan.config.parallelism import ParallelismConfig
 from torchtitan.distributed.batch_invariant import is_in_batch_invariant_mode
+from torchtitan.distributed.local_compile import LocalCompileConfig
 from torchtitan.distributed.parallelism_context import ParallelismContext
 from torchtitan.models.base.model import BaseModel
 from torchtitan.models.base.module import Module, ModuleDict
@@ -285,7 +286,9 @@ def __init__(
             )
             self.sharding_config = None
             # HF modules run eager; the loss and the swapped-in TorchTitan MoE (SwiGLU) can compile.
-            self.local_compile_regions = ["loss", "fused_binary_activation"]
+            self.local_compile_regions = LocalCompileConfig(
+                regions=["loss", "fused_binary_activation"]
+            )
 
             assert model_config is not None, "model_config is required"
 
diff --git a/torchtitan/experiments/transformers_modeling_backend/parallelize.py b/torchtitan/experiments/transformers_modeling_backend/parallelize.py
index 46be051bc05..a21cbe2bdee 100644
--- a/torchtitan/experiments/transformers_modeling_backend/parallelize.py
+++ b/torchtitan/experiments/transformers_modeling_backend/parallelize.py
@@ -27,7 +27,7 @@
     enable_fsdp_symm_mem,
     get_fsdp_reshard_after_forward_policy,
 )
-from torchtitan.distributed.local_compile import apply_local_compile
+from torchtitan.distributed.local_compile import LocalCompileConfig
 
 
 logger = logging.getLogger(__name__)
@@ -87,7 +87,7 @@ def parallelize_hf_transformers(
     parallelism_context: ParallelismContext,
     training: TrainingConfig,
     parallelism: ParallelismConfig,
-    local_compile_regions: list[str],
+    local_compile_regions: LocalCompileConfig,
     activation_remat: ActivationRematConfig | None,
     dump_folder: str,
     **kwargs: Any,
@@ -102,7 +102,7 @@ def parallelize_hf_transformers(
     5. Apply AC and FSDP
     """
     # Bind local implementations early; torch.compile traces on first use.
-    apply_local_compile(local_compile_regions)
+    local_compile_regions.apply()
     # Flex attention supports FSDP, TP, CP, and PP (in any combination). Under CP
     # the flex kernel's local SPMD boundary redistributes
     # k/v from seq-sharded to CP-Replicate (all-gather); see _attach_flex_kernel
diff --git a/torchtitan/experiments/transformers_modeling_backend/pipeline.py b/torchtitan/experiments/transformers_modeling_backend/pipeline.py
index f439c413c35..8833d566bbf 100644
--- a/torchtitan/experiments/transformers_modeling_backend/pipeline.py
+++ b/torchtitan/experiments/transformers_modeling_backend/pipeline.py
@@ -25,6 +25,7 @@
 from torchtitan.config.parallelism import ParallelismConfig
 from torchtitan.distributed import ParallelismContext
 from torchtitan.distributed.activation_remat import ActivationRematConfig
+from torchtitan.distributed.local_compile import LocalCompileConfig
 from torchtitan.distributed.pipeline_parallel import (
     _build_get_mesh_callback,
     _build_pipeline_schedule,
@@ -296,7 +297,7 @@ def pipeline_hf_transformers(
     *,
     training: TrainingConfig,
     parallelism: ParallelismConfig,
-    local_compile_regions: list[str],
+    local_compile_regions: LocalCompileConfig,
     activation_remat: ActivationRematConfig | None,
     dump_folder: str,
     device: torch.device,
diff --git a/torchtitan/models/base/model.py b/torchtitan/models/base/model.py
index 52760ed6a96..2b211ef18a8 100644
--- a/torchtitan/models/base/model.py
+++ b/torchtitan/models/base/model.py
@@ -14,7 +14,7 @@
 
 from torchtitan.config import TrainingConfig
 from torchtitan.config.parallelism import ParallelismConfig
-from torchtitan.distributed.local_compile import apply_local_compile
+from torchtitan.distributed.local_compile import LocalCompileConfig
 from torchtitan.distributed.parallelism_context import ParallelismContext
 
 from .module import Module
@@ -86,10 +86,12 @@ class Config(Module.Config):
         Subclasses define model-specific hyperparameters.
         """
 
-        local_compile_regions: list[str] = field(default_factory=list)
-        """``@local_compile`` regions this model compiles; ``[]`` runs them all eager.
+        local_compile_regions: LocalCompileConfig = field(
+            default_factory=LocalCompileConfig
+        )
+        """Configuration for named regions, FlexAttention, and block-mask helpers.
 
-        FlexAttention manages its own compilation and is not controlled by this list.
+        An empty ``regions`` list runs named regions eagerly.
         """
 
         @abstractmethod
@@ -132,13 +134,13 @@ def parallelize(
         parallelism_context: ParallelismContext,
         training: TrainingConfig,
         parallelism: ParallelismConfig,
-        local_compile_regions: list[str],
+        local_compile_regions: LocalCompileConfig,
         activation_remat: ActivationRematConfig | None,
         dump_folder: str,
     ) -> Self:
         """Apply the ordered model-level parallelization lifecycle."""
         # Bind local implementations early; torch.compile traces on first use.
-        apply_local_compile(local_compile_regions)
+        local_compile_regions.apply()
         with parallelism_context.activate_spmd():
             self._parallelize(parallelism_context)
             if activation_remat is not None:
diff --git a/torchtitan/models/common/attention/attention.py b/torchtitan/models/common/attention/attention.py
index 0258daedfc1..395434af279 100644
--- a/torchtitan/models/common/attention/attention.py
+++ b/torchtitan/models/common/attention/attention.py
@@ -39,6 +39,7 @@
 )
 
 from torchtitan.distributed.batch_invariant import is_in_batch_invariant_mode
+from torchtitan.distributed.local_compile import compiled
 from torchtitan.distributed.parallelism_context import MeshAxisName
 from torchtitan.models.base.module import Module
 from torchtitan.models.common.linear import Linear
@@ -344,10 +345,9 @@ def build_attention_metadata(
         "triton.cudagraphs": False,
     }
 
-    # pyrefly: ignore[no-matching-overload]
-    _compiled_flex_attn: ClassVar[Callable] = torch.compile(
+    _compiled_flex_attn: ClassVar[Callable] = compiled(
         flex_attention,
-        options=inductor_configs,
+        inductor_options=inductor_configs,
     )
 
     def __init__(self, config: Config) -> None:
@@ -642,7 +642,7 @@ def sliding_window_mod(
     return sliding_window_mod
 
 
-_compiled_create_block_mask = torch.compile(create_block_mask)
+_compiled_create_block_mask = compiled(create_block_mask)
 
 
 def create_attention_mask(*args, **kwargs):
diff --git a/torchtitan/models/common/multimodal.py b/torchtitan/models/common/multimodal.py
index 26ce302f79a..9307c9a0f7e 100644
--- a/torchtitan/models/common/multimodal.py
+++ b/torchtitan/models/common/multimodal.py
@@ -20,7 +20,7 @@
 from torchtitan.config import TORCH_DTYPE_MAP, TrainingConfig
 from torchtitan.config.parallelism import ParallelismConfig
 from torchtitan.distributed.activation_remat import ActivationRematConfig
-from torchtitan.distributed.local_compile import apply_local_compile
+from torchtitan.distributed.local_compile import LocalCompileConfig
 from torchtitan.distributed.parallelism_context import ParallelismContext
 
 from .decoder import Decoder
@@ -37,12 +37,12 @@ def parallelize(
         parallelism_context: ParallelismContext,
         training: TrainingConfig,
         parallelism: ParallelismConfig,
-        local_compile_regions: list[str],
+        local_compile_regions: LocalCompileConfig,
         activation_remat: ActivationRematConfig | None,
         dump_folder: str,
     ) -> Self:
         # Bind local implementations early; torch.compile traces on first use.
-        apply_local_compile(local_compile_regions)
+        local_compile_regions.apply()
         with parallelism_context.activate_spmd():
             self._parallelize(parallelism_context)
             encoders = [
diff --git a/torchtitan/models/common/vision_encoder.py b/torchtitan/models/common/vision_encoder.py
index a1d4b8c1fa3..ca3221a0077 100644
--- a/torchtitan/models/common/vision_encoder.py
+++ b/torchtitan/models/common/vision_encoder.py
@@ -28,6 +28,7 @@
 import torch_remat as remat
 from torch.nn.attention.flex_attention import BlockMask, create_block_mask
 
+from torchtitan.distributed.local_compile import compiled
 from torchtitan.distributed.parallelism_context import MeshAxisName
 from torchtitan.distributed.spmd_types import spmd_mesh_group
 from torchtitan.models.base.module import Module
@@ -35,7 +36,7 @@
 from torchtitan.models.common.linear import Linear
 from torchtitan.models.common.nn_modules import GELU, LayerNorm, RMSNorm
 
-compiled_create_block_mask = torch.compile(create_block_mask)
+compiled_create_block_mask = compiled(create_block_mask)
 
 # Applies rotary position embedding: (query, key, rope_cache) -> (query, key).
 RopeApply = Callable[
diff --git a/torchtitan/models/deepseek_v3/model.py b/torchtitan/models/deepseek_v3/model.py
index c1ba8bc60ee..a4c1d20f7f2 100644
--- a/torchtitan/models/deepseek_v3/model.py
+++ b/torchtitan/models/deepseek_v3/model.py
@@ -13,6 +13,7 @@
 import torch_remat as remat
 from torch import nn
 
+from torchtitan.distributed.local_compile import LocalCompileConfig
 from torchtitan.models.base.module import Module
 from torchtitan.models.common.attention import (
     BaseAttention,
@@ -272,13 +273,15 @@ class DeepSeekV3Model(MTPDecoder):
     class Config(MTPDecoder.Config):
         dim: int = 2048
         vocab_size: int = 102400
-        local_compile_regions: list[str] = field(
-            default_factory=lambda: [
-                "loss",
-                "fused_binary_activation",
-                "fp32_to_bf16_split",
-                "moe_aux_loss",
-            ]
+        local_compile_regions: LocalCompileConfig = field(
+            default_factory=lambda: LocalCompileConfig(
+                regions=[
+                    "loss",
+                    "fused_binary_activation",
+                    "fp32_to_bf16_split",
+                    "moe_aux_loss",
+                ]
+            )
         )
 
         def get_nparams_and_flops(
diff --git a/torchtitan/models/deepseek_v4/model.py b/torchtitan/models/deepseek_v4/model.py
index 1da25fda7a4..06ed95b7586 100644
--- a/torchtitan/models/deepseek_v4/model.py
+++ b/torchtitan/models/deepseek_v4/model.py
@@ -14,6 +14,7 @@
 
 from torchtitan.config import TORCH_DTYPE_MAP, TrainingConfig
 from torchtitan.config.parallelism import ParallelismConfig
+from torchtitan.distributed.local_compile import LocalCompileConfig
 from torchtitan.distributed.parallelism_context import ParallelismContext
 from torchtitan.models.base.module import ModuleList
 from torchtitan.models.common.attention import (
@@ -179,12 +180,14 @@ def _apply_fsdp(
     class Config(Decoder.Config):
         dim: int
         vocab_size: int
-        local_compile_regions: list[str] = field(
-            default_factory=lambda: [
-                "loss",
-                "fused_binary_activation",
-                "fp32_to_bf16_split",
-            ]
+        local_compile_regions: LocalCompileConfig = field(
+            default_factory=lambda: LocalCompileConfig(
+                regions=[
+                    "loss",
+                    "fused_binary_activation",
+                    "fp32_to_bf16_split",
+                ]
+            )
         )
         hc_mult: int = 4
         n_mtp_layers: int = 0
diff --git a/torchtitan/models/flux/README.md b/torchtitan/models/flux/README.md
index bb331f575d2..abba84856a1 100644
--- a/torchtitan/models/flux/README.md
+++ b/torchtitan/models/flux/README.md
@@ -51,7 +51,7 @@ MODULE=torchtitan_recipes.tests.models.flux CONFIG=flux_debugmodel_inference \
 The loss function can be compiled independently with:
 
 ```python
-config.model.local_compile_regions = ["loss"]
+config.model.local_compile_regions.regions = ["loss"]
 ```
 
 The model itself is not compiled by the standard Trainer.
diff --git a/torchtitan/models/flux/model.py b/torchtitan/models/flux/model.py
index 39c6480c931..46bec80d43f 100644
--- a/torchtitan/models/flux/model.py
+++ b/torchtitan/models/flux/model.py
@@ -18,7 +18,7 @@
 from torchtitan.config.parallelism import ParallelismConfig
 from torchtitan.distributed import context_parallel
 from torchtitan.distributed.activation_remat import ActivationRematConfig
-from torchtitan.distributed.local_compile import apply_local_compile
+from torchtitan.distributed.local_compile import LocalCompileConfig
 from torchtitan.distributed.parallelism_context import MeshAxisName, ParallelismContext
 from torchtitan.distributed.spmd_types import (
     annotate_replicated_parameters,
@@ -103,7 +103,9 @@ class FluxModel(BaseModel):
     class Config(BaseModel.Config):
         img_in: Linear.Config
         txt_in: Linear.Config
-        local_compile_regions: list[str] = field(default_factory=lambda: ["loss"])
+        local_compile_regions: LocalCompileConfig = field(
+            default_factory=lambda: LocalCompileConfig(regions=["loss"])
+        )
         in_channels: int = 64
         out_channels: int = 64
         vec_in_dim: int = 768
@@ -235,13 +237,13 @@ def parallelize(
         parallelism_context: ParallelismContext,
         training: TrainingConfig,
         parallelism: ParallelismConfig,
-        local_compile_regions: list[str],
+        local_compile_regions: LocalCompileConfig,
         activation_remat: ActivationRematConfig | None,
         dump_folder: str,
     ) -> Self:
         """Apply Flux's AC-before-SPMD parallelization lifecycle."""
         # Bind local implementations early; torch.compile traces on first use.
-        apply_local_compile(local_compile_regions)
+        local_compile_regions.apply()
         with parallelism_context.activate_spmd():
             if activation_remat is not None:
                 # Flux saves no regions under any policy: each block is
diff --git a/torchtitan/models/gpt_oss/model.py b/torchtitan/models/gpt_oss/model.py
index 35bd95b26f0..af275f76343 100644
--- a/torchtitan/models/gpt_oss/model.py
+++ b/torchtitan/models/gpt_oss/model.py
@@ -17,6 +17,7 @@
 from torchtitan.config import TrainingConfig
 from torchtitan.config.parallelism import ParallelismConfig
 from torchtitan.distributed.activation_remat import ActivationRematConfig
+from torchtitan.distributed.local_compile import LocalCompileConfig
 from torchtitan.distributed.parallelism_context import ParallelismContext
 from torchtitan.models.base.module import Module
 from torchtitan.models.common.attention import (
@@ -216,13 +217,15 @@ def _register_optimizer_hooks(
     class Config(Decoder.Config):
         dim: int = 2880
         vocab_size: int = 201088
-        local_compile_regions: list[str] = field(
-            default_factory=lambda: [
-                "loss",
-                "fused_binary_activation",
-                "cos_sin_rope",
-                "fp32_to_bf16_split",
-            ]
+        local_compile_regions: LocalCompileConfig = field(
+            default_factory=lambda: LocalCompileConfig(
+                regions=[
+                    "loss",
+                    "fused_binary_activation",
+                    "cos_sin_rope",
+                    "fp32_to_bf16_split",
+                ]
+            )
         )
 
         def get_nparams_and_flops(
@@ -250,7 +253,7 @@ def parallelize(
         parallelism_context: ParallelismContext,
         training: TrainingConfig,
         parallelism: ParallelismConfig,
-        local_compile_regions: list[str],
+        local_compile_regions: LocalCompileConfig,
         activation_remat: ActivationRematConfig | None,
         dump_folder: str,
     ) -> GptOssModel:
diff --git a/torchtitan/models/kimi_k2_7/model.py b/torchtitan/models/kimi_k2_7/model.py
index 6d1dd05790b..a621676c1bb 100644
--- a/torchtitan/models/kimi_k2_7/model.py
+++ b/torchtitan/models/kimi_k2_7/model.py
@@ -20,6 +20,7 @@
 from torchtitan.config import TrainingConfig
 from torchtitan.config.parallelism import ParallelismConfig
 from torchtitan.distributed.activation_remat import ActivationRematConfig
+from torchtitan.distributed.local_compile import LocalCompileConfig
 from torchtitan.distributed.parallelism_context import ParallelismContext
 from torchtitan.distributed.spmd_types import (
     annotate_input_spmd_types,
@@ -117,7 +118,7 @@ def parallelize(
         parallelism_context: ParallelismContext,
         training: TrainingConfig,
         parallelism: ParallelismConfig,
-        local_compile_regions: list[str],
+        local_compile_regions: LocalCompileConfig,
         activation_remat: ActivationRematConfig | None,
         dump_folder: str,
     ) -> KimiK25Model:
diff --git a/torchtitan/models/kimi_k3/model.py b/torchtitan/models/kimi_k3/model.py
index f1313a458a1..889b9276e2a 100644
--- a/torchtitan/models/kimi_k3/model.py
+++ b/torchtitan/models/kimi_k3/model.py
@@ -19,7 +19,7 @@
 from torchtitan.config.function import Function
 from torchtitan.config.parallelism import ParallelismConfig
 from torchtitan.distributed.activation_remat import ActivationRematConfig
-from torchtitan.distributed.local_compile import apply_local_compile, local_compile
+from torchtitan.distributed.local_compile import local_compile, LocalCompileConfig
 from torchtitan.distributed.parallelism_context import MeshAxisName, ParallelismContext
 from torchtitan.distributed.spmd_types import (
     annotate_input_spmd_types,
@@ -391,14 +391,16 @@ class Config(Decoder.Config):
         attention_residual: AttentionResidual.Config = field(
             default_factory=AttentionResidual.Config
         )
-        local_compile_regions: list[str] = field(
-            default_factory=lambda: [
-                "loss",
-                "gated_rmsnorm",
-                "fused_binary_activation",
-                "fp32_to_bf16_split",
-                "attention_residual",
-            ]
+        local_compile_regions: LocalCompileConfig = field(
+            default_factory=lambda: LocalCompileConfig(
+                regions=[
+                    "loss",
+                    "gated_rmsnorm",
+                    "fused_binary_activation",
+                    "fp32_to_bf16_split",
+                    "attention_residual",
+                ]
+            )
         )
 
         def get_nparams_and_flops(
@@ -458,12 +460,12 @@ def parallelize(
         parallelism_context: ParallelismContext,
         training: TrainingConfig,
         parallelism: ParallelismConfig,
-        local_compile_regions: list[str],
+        local_compile_regions: LocalCompileConfig,
         activation_remat: ActivationRematConfig | None,
         dump_folder: str,
     ) -> KimiK3Model:
         # Bind local implementations early; torch.compile traces on first use.
-        apply_local_compile(local_compile_regions)
+        local_compile_regions.apply()
         with parallelism_context.activate_spmd():
             annotate_replicated_parameters(self, parallelism_context)
             self._parallelize(parallelism_context)
diff --git a/torchtitan/models/llama3/model.py b/torchtitan/models/llama3/model.py
index b0335209d90..54bd7302f80 100644
--- a/torchtitan/models/llama3/model.py
+++ b/torchtitan/models/llama3/model.py
@@ -12,6 +12,7 @@
 import torch_remat as remat
 from torch import nn
 
+from torchtitan.distributed.local_compile import LocalCompileConfig
 from torchtitan.models.common.attention import (
     FlexAttentionMetadata,
     VarlenAttentionMetadata,
@@ -83,8 +84,10 @@ class Llama3Model(Decoder):
     class Config(Decoder.Config):
         dim: int = 4096
         vocab_size: int = 128256
-        local_compile_regions: list[str] = field(
-            default_factory=lambda: ["loss", "fused_binary_activation"]
+        local_compile_regions: LocalCompileConfig = field(
+            default_factory=lambda: LocalCompileConfig(
+                regions=["loss", "fused_binary_activation"]
+            )
         )
 
         def get_nparams_and_flops(
diff --git a/torchtitan/models/muse_glimmer/model.py b/torchtitan/models/muse_glimmer/model.py
index 72938708b72..0402160c20a 100644
--- a/torchtitan/models/muse_glimmer/model.py
+++ b/torchtitan/models/muse_glimmer/model.py
@@ -18,6 +18,7 @@
 from torchtitan.config import TrainingConfig
 from torchtitan.config.parallelism import ParallelismConfig
 from torchtitan.distributed.activation_remat import ActivationRematConfig
+from torchtitan.distributed.local_compile import LocalCompileConfig
 from torchtitan.distributed.parallelism_context import ParallelismContext
 from torchtitan.distributed.spmd_types import (
     annotate_input_spmd_types,
@@ -289,8 +290,10 @@ class MuseGlimmerModel(MultimodalModel):
     class Config(Decoder.Config):
         dim: int = 6656
         vocab_size: int = 202048
-        local_compile_regions: list[str] = field(
-            default_factory=lambda: ["loss", "fused_binary_activation"]
+        local_compile_regions: LocalCompileConfig = field(
+            default_factory=lambda: LocalCompileConfig(
+                regions=["loss", "fused_binary_activation"]
+            )
         )
         # Narrows the base Decoder.Config.tok_embeddings (Embedding.Config) to the
         # bundled embedding+norm unit whose .embedding/.norm children carry the sharding.
@@ -369,7 +372,7 @@ def parallelize(
         parallelism_context: ParallelismContext,
         training: TrainingConfig,
         parallelism: ParallelismConfig,
-        local_compile_regions: list[str],
+        local_compile_regions: LocalCompileConfig,
         activation_remat: ActivationRematConfig | None,
         dump_folder: str,
     ) -> MuseGlimmerModel:
diff --git a/torchtitan/models/qwen3/model.py b/torchtitan/models/qwen3/model.py
index 0ec1c454b65..bcabbf086f9 100644
--- a/torchtitan/models/qwen3/model.py
+++ b/torchtitan/models/qwen3/model.py
@@ -12,6 +12,7 @@
 import torch.nn as nn
 import torch_remat as remat
 
+from torchtitan.distributed.local_compile import LocalCompileConfig
 from torchtitan.models.common.attention import (
     FlexAttentionMetadata,
     VarlenAttentionMetadata,
@@ -108,13 +109,15 @@ def _register_optimizer_hooks(
     class Config(Decoder.Config):
         dim: int = 1024
         vocab_size: int = 151936
-        local_compile_regions: list[str] = field(
-            default_factory=lambda: [
-                "loss",
-                "fused_binary_activation",
-                "cos_sin_rope",
-                "fp32_to_bf16_split",
-            ]
+        local_compile_regions: LocalCompileConfig = field(
+            default_factory=lambda: LocalCompileConfig(
+                regions=[
+                    "loss",
+                    "fused_binary_activation",
+                    "cos_sin_rope",
+                    "fp32_to_bf16_split",
+                ]
+            )
         )
 
         def get_nparams_and_flops(
diff --git a/torchtitan/models/qwen3_5/model.py b/torchtitan/models/qwen3_5/model.py
index bd68938bfb6..368f27ef7b8 100644
--- a/torchtitan/models/qwen3_5/model.py
+++ b/torchtitan/models/qwen3_5/model.py
@@ -18,7 +18,7 @@
 from torchtitan.config import TrainingConfig
 from torchtitan.config.parallelism import ParallelismConfig
 from torchtitan.distributed.activation_remat import ActivationRematConfig
-from torchtitan.distributed.local_compile import local_compile
+from torchtitan.distributed.local_compile import local_compile, LocalCompileConfig
 from torchtitan.distributed.parallelism_context import MeshAxisName, ParallelismContext
 from torchtitan.distributed.spmd_types import (
     annotate_input_spmd_types,
@@ -353,16 +353,18 @@ def _register_optimizer_hooks(
     @dataclass(kw_only=True, slots=True)
     class Config(Decoder.Config):
         vision_encoder: Qwen35VisionEncoder.Config | None = None
-        local_compile_regions: list[str] = field(
-            default_factory=lambda: [
-                "loss",
-                "fused_binary_activation",
-                "gated_rmsnorm",
-                "offset_rmsnorm",
-                "partial_rope",
-                "shared_expert_gate",
-                "fp32_to_bf16_split",
-            ]
+        local_compile_regions: LocalCompileConfig = field(
+            default_factory=lambda: LocalCompileConfig(
+                regions=[
+                    "loss",
+                    "fused_binary_activation",
+                    "gated_rmsnorm",
+                    "offset_rmsnorm",
+                    "partial_rope",
+                    "shared_expert_gate",
+                    "fp32_to_bf16_split",
+                ]
+            )
         )
 
         def get_nparams_and_flops(
@@ -436,7 +438,7 @@ def parallelize(
         parallelism_context: ParallelismContext,
         training: TrainingConfig,
         parallelism: ParallelismConfig,
-        local_compile_regions: list[str],
+        local_compile_regions: LocalCompileConfig,
         activation_remat: ActivationRematConfig | None,
         dump_folder: str,
     ) -> Qwen35Model:
diff --git a/torchtitan_recipes/models/deepseek_v3.py b/torchtitan_recipes/models/deepseek_v3.py
index 5a762cdec49..9e4e8b61636 100644
--- a/torchtitan_recipes/models/deepseek_v3.py
+++ b/torchtitan_recipes/models/deepseek_v3.py
@@ -268,5 +268,5 @@ def deepseek_v3_671b_nvfp4_ffn_mxfp8_attn(
             )
         ],
     )
-    config.model.local_compile_regions = ["loss"]
+    config.model.local_compile_regions.regions = ["loss"]
     return config
diff --git a/torchtitan_recipes/rl/alphabet_sort.py b/torchtitan_recipes/rl/alphabet_sort.py
index f7125220ca0..29141010e90 100644
--- a/torchtitan_recipes/rl/alphabet_sort.py
+++ b/torchtitan_recipes/rl/alphabet_sort.py
@@ -174,7 +174,7 @@ def rl_grpo_qwen3_0_6b_varlen(*, seq_len: int = 2048) -> Controller.Config:
 def rl_grpo_qwen3_0_6b_varlen_no_compile(*, seq_len: int = 2048) -> Controller.Config:
     config = rl_grpo_qwen3_0_6b_varlen(seq_len=seq_len)
     for model_config in (config.trainer.model, config.generator.model):
-        model_config.local_compile_regions = []
+        model_config.local_compile_regions.regions = []
     return config
 
 
@@ -299,7 +299,7 @@ def rl_grpo_qwen3_0_6b_flex_batch_invariant() -> Controller.Config:
     )
     # Local compile regions do not support batch-invariant mode.
     for model_config in (config.trainer.model, config.generator.model):
-        model_config.local_compile_regions = []
+        model_config.local_compile_regions.regions = []
     block_size = config.trainer.model.layers[0].attention.inner_attention.block_size
     config.async_loop.batcher = dataclasses.replace(
         config.async_loop.batcher, per_sample_pad_multiple=block_size
@@ -671,7 +671,7 @@ def rl_grpo_qwen3_0_6b_varlen_batch_invariant(
     )
     # Local compile regions do not support batch-invariant mode.
     for model_config in (trainer_model, generator_model):
-        model_config.local_compile_regions = []
+        model_config.local_compile_regions.regions = []
     return Controller.Config(
         hf_assets_path="./assets/hf/Qwen3-0.6B",
         num_generators=3,
@@ -856,7 +856,7 @@ def rl_grpo_qwen3_5_9b_varlen_batch_invariant() -> Controller.Config:
     )
     # Local compile regions do not support batch-invariant mode.
     for model_config in (config.trainer.model, config.generator.model):
-        model_config.local_compile_regions = []
+        model_config.local_compile_regions.regions = []
     return config
 
 
diff --git a/torchtitan_recipes/tests/__init__.py b/torchtitan_recipes/tests/__init__.py
index 3792be993e9..fb45108df37 100644
--- a/torchtitan_recipes/tests/__init__.py
+++ b/torchtitan_recipes/tests/__init__.py
@@ -27,4 +27,4 @@ def _set_spmd_typechecking(config: Trainer.Config, *, typechecking: bool) -> Non
     if typechecking:
         # Local compile does not support global SPMD typechecking yet.
         # TODO: Remove this once the SPMD typechecking/Dynamo issue is fixed.
-        config.model.local_compile_regions = []
+        config.model.local_compile_regions.regions = []
diff --git a/torchtitan_recipes/tests/graph_trainer/numerics.py b/torchtitan_recipes/tests/graph_trainer/numerics.py
index 277344d9ce7..e8d85296cce 100644
--- a/torchtitan_recipes/tests/graph_trainer/numerics.py
+++ b/torchtitan_recipes/tests/graph_trainer/numerics.py
@@ -55,7 +55,7 @@ def _without_inplace_wgrad_accum(config):
 def llama3_eager_numerics():
     config = llama3_debugmodel(seq_len=2048)
     # Match GraphTrainer, which captures model and loss in one eager FX graph.
-    config.model.local_compile_regions = []
+    config.model.local_compile_regions.regions = []
     config.parallelism.tensor_parallel_degree = 2
     config.parallelism.data_parallel_shard_degree = 4
     config.training.num_tokens_per_microbatch_per_dp_rank = 16384
@@ -73,7 +73,7 @@ def llama3_graph_numerics():
 def llama3_eager_hsdp_gradient_accumulation_numerics():
     config = llama3_debugmodel(seq_len=2048)
     # Match GraphTrainer, which captures model and loss in one eager FX graph.
-    config.model.local_compile_regions = []
+    config.model.local_compile_regions.regions = []
     config.parallelism.data_parallel_replicate_degree = 2
     config.parallelism.data_parallel_shard_degree = 2
     config.training.num_tokens_per_microbatch_per_dp_rank = 8192
@@ -103,7 +103,7 @@ def _llama3_hsdp_pp_numerics(config):
 
 def llama3_eager_hsdp_pp_numerics():
     config = llama3_debugmodel(seq_len=2048)
-    config.model.local_compile_regions = []
+    config.model.local_compile_regions.regions = []
     return _without_inplace_wgrad_accum(_llama3_hsdp_pp_numerics(config))
 
 
@@ -214,7 +214,7 @@ def _deepseek_v3_dist_moe_pp_numerics(
     runtime_config.pp_activation_slot_policy = slot_policy
     for _, experts, _, _ in config.model.traverse(DistMoeRoutedExperts.Config):
         experts.inplace_wgrad_accum = False
-    config.model.local_compile_regions = []
+    config.model.local_compile_regions.regions = []
     config.parallelism.pipeline_parallel_degree = 2
     config.parallelism.pipeline_parallel_schedule = "Interleaved1F1B"
     config.parallelism.num_pp_microbatches = 8
@@ -251,7 +251,7 @@ def deepseek_v3_dist_moe_graph_pp_stage_microbatch_numerics():
 def qwen3_eager_numerics():
     config = qwen3_debugmodel(seq_len=2048)
     # Match GraphTrainer, which captures model and loss in one eager FX graph.
-    config.model.local_compile_regions = []
+    config.model.local_compile_regions.regions = []
     config.parallelism.tensor_parallel_degree = 2
     config.parallelism.data_parallel_shard_degree = 4
     config.training.num_tokens_per_microbatch_per_dp_rank = 16384
@@ -278,7 +278,7 @@ def qwen3_moe_eager_numerics():
     )
     config.dataloader.max_num_documents = None
     # Match GraphTrainer, which captures model and loss in one eager FX graph.
-    config.model.local_compile_regions = []
+    config.model.local_compile_regions.regions = []
     config.training.disable_cuda_graphs = True
     config.training.num_tokens_per_microbatch_per_dp_rank = 16384
     return _without_inplace_wgrad_accum(config)
diff --git a/torchtitan_recipes/tests/models/deepseek_v3.py b/torchtitan_recipes/tests/models/deepseek_v3.py
index 65225c382bc..bc779ebaeb1 100644
--- a/torchtitan_recipes/tests/models/deepseek_v3.py
+++ b/torchtitan_recipes/tests/models/deepseek_v3.py
@@ -481,7 +481,7 @@ def deepseek_v3_debugmodel_nvfp4_ffn_mxfp8_attn(
             ),
         ],
     )
-    config.model.local_compile_regions = ["loss"]
+    config.model.local_compile_regions.regions = ["loss"]
     return config
 
 
@@ -530,5 +530,5 @@ def deepseek_v3_16b_nvfp4_ffn_mxfp8_attn(
             )
         ],
     )
-    config.model.local_compile_regions = ["loss"]
+    config.model.local_compile_regions.regions = ["loss"]
     return config
diff --git a/torchtitan_recipes/tests/rl/alphabet_sort.py b/torchtitan_recipes/tests/rl/alphabet_sort.py
index ce1b21069ed..d86fbadb948 100644
--- a/torchtitan_recipes/tests/rl/alphabet_sort.py
+++ b/torchtitan_recipes/tests/rl/alphabet_sort.py
@@ -156,7 +156,7 @@ def rl_grpo_gpt_oss_debug_varlen_no_compile(
 ) -> Controller.Config:
     config = rl_grpo_gpt_oss_debug_varlen(seq_len=seq_len)
     for model_config in (config.trainer.model, config.generator.model):
-        model_config.local_compile_regions = []
+        model_config.local_compile_regions.regions = []
     return config
 
 
@@ -194,7 +194,7 @@ def rl_grpo_gpt_oss_debug_varlen_batch_invariant() -> Controller.Config:
     )
     # Local compile regions do not support batch-invariant mode.
     for model_config in (trainer_model, generator_model):
-        model_config.local_compile_regions = []
+        model_config.local_compile_regions.regions = []
     return Controller.Config(
         hf_assets_path="tests/assets/tokenizer",
         async_loop=AsyncLoopConfig(
@@ -439,7 +439,7 @@ def rl_grpo_qwen3_moe_debug_varlen_batch_invariant(
     )
     # Local compile regions do not support batch-invariant mode.
     for model_config in (trainer_model, generator_model):
-        model_config.local_compile_regions = []
+        model_config.local_compile_regions.regions = []
     return Controller.Config(
         hf_assets_path="tests/assets/tokenizer",
         async_loop=AsyncLoopConfig(
@@ -604,7 +604,7 @@ def rl_grpo_qwen3_5_debug_varlen_batch_invariant(
     )
     # Local compile regions do not support batch-invariant mode.
     for model_config in (config.trainer.model, config.generator.model):
-        model_config.local_compile_regions = []
+        model_config.local_compile_regions.regions = []
     return config
 
 
@@ -629,7 +629,7 @@ def _kimi_k3_rl_model_configs(
         converters=converters,
     )
     for model_config in (trainer_model, generator_model):
-        model_config.local_compile_regions = []
+        model_config.local_compile_regions.regions = []
     return trainer_model, generator_model
 
 
@@ -819,7 +819,7 @@ def rl_grpo_moe_debug_tp4_ep4() -> Controller.Config:
         generator_parallelism=config.generator.parallelism,
     )
     for model_config in (config.trainer.model, config.generator.model):
-        model_config.local_compile_regions = []
+        model_config.local_compile_regions.regions = []
     # TODO: use FULL CUDA graphs once GPT-OSS CI enables a
     # CUDA-graph-compatible distributed MoE token dispatcher.
     config.generator.cuda_graph.mode = "NONE"