Pass MPServerConfig.shm_name into NonGPUTransferModule

lmcache/v1/multiprocess/modules/non_gpu_transfer.py

@@ -16,6 +16,7 @@
     MemoryLayoutDesc,
     ObjectKey,
 )
+from lmcache.v1.multiprocess.config import MPServerConfig
 from lmcache.v1.multiprocess.custom_types import (
     IPCCacheEngineKey,
     RegisterNonGpuContextPayload,
@@ -72,10 +73,18 @@ class NonGPUTransferModule:
 
     Args:
         ctx: The shared engine context.
+        mp_config: Optional MP server config carrying non-GPU SHM overrides.
     """
 
-    def __init__(self, ctx: MPCacheEngineContext) -> None:
+    def __init__(
+        self,
+        ctx: MPCacheEngineContext,
+        mp_config: MPServerConfig | None = None,
+    ) -> None:
         self._ctx = ctx
+        self._shm_name_override = (
+            mp_config.shm_name if mp_config is not None else None
+        )
         self._non_gpu_contexts: dict[int, NonGPUContextEntry] = {}
         self._strategies: dict[int, TransferStrategy] = {}
         self._pending_shm_writes: dict[
@@ -162,6 +171,17 @@ def close(self) -> None:
 
     def _compute_shm_pool_info(self) -> ShmPoolInfo:
         """Compute SHM pool info from storage manager config."""
+        if self._shm_name_override is not None:
+            shm_name = self._shm_name_override
+            if not shm_name:
+                return {"shm_name": "", "pool_size": 0}
+            sm_config = self._ctx.storage_manager_config
+            mem_cfg = sm_config.l1_manager_config.memory_config
+            bare = shm_name.lstrip("/")
+            if not bare.startswith("lmcache_l1_pool_"):
+                shm_name = f"lmcache_l1_pool_{bare}"
+            return {"shm_name": shm_name, "pool_size": mem_cfg.size_in_bytes}
+
         sm_config = self._ctx.storage_manager_config
         mem_cfg = sm_config.l1_manager_config.memory_config
         shm_name = mem_cfg.shm_name or ""

lmcache/v1/multiprocess/server.py

@@ -145,7 +145,7 @@ def _build_modules(
     if mp_config.transfer_mode == "gpu":
         modules.append(GPUTransferModule(ctx))
     else:
-        modules.append(NonGPUTransferModule(ctx))
+        modules.append(NonGPUTransferModule(ctx, mp_config))
 
     if mp_config.engine_type == "blend":
         if mp_config.transfer_mode != "gpu":
@@ -191,6 +191,14 @@ def run_cache_server(
 
     maybe_initialize_trace_recorder(event_bus, obs_config, storage_manager_config)
 
+    # Apply shm_name override from MP config so the allocator creates the
+    # segment with the user-specified name (must happen before StorageManager
+    # is instantiated inside MPCacheEngineContext).
+    if mp_config.shm_name is not None:
+        storage_manager_config.l1_manager_config.memory_config.shm_name = (
+            mp_config.shm_name
+        )
+
     ctx = MPCacheEngineContext(
         storage_manager_config=storage_manager_config,
         chunk_size=mp_config.chunk_size,

tests/v1/multiprocess/test_non_cuda_data_transfer.py

@@ -404,6 +404,94 @@ def test_server_register_and_find_non_cuda_context_layout(
     assert layout.shapes[0] == torch.Size([2, 2, 16, 16])
 
 
+def test_build_modules_passes_empty_shm_override_to_non_gpu_module(
+    stub_native_storage_ops: Any,
+) -> None:
+    """Ensure MP-level empty shm override disables SHM in non-GPU mode."""
+    # First Party
+    from lmcache.v1.multiprocess.config import MPServerConfig
+    from lmcache.v1.multiprocess.custom_types import RegisterNonGpuContextPayload
+    from lmcache.v1.multiprocess.engine_context import MPCacheEngineContext
+    from lmcache.v1.multiprocess.modules.non_gpu_transfer import NonGPUTransferModule
+    from lmcache.v1.multiprocess.server import _build_modules
+
+    storage_manager_config = MagicMock()
+    storage_manager_config.l1_manager_config.memory_config.shm_name = "storage_default"
+    storage_manager_config.l1_manager_config.memory_config.use_lazy = False
+    storage_manager_config.l1_manager_config.memory_config.size_in_bytes = 4096
+
+    with (
+        patch("lmcache.v1.multiprocess.engine_context.StorageManager"),
+        patch("lmcache.v1.multiprocess.engine_context.TokenHasher"),
+        patch("lmcache.v1.multiprocess.engine_context.SessionManager"),
+        patch("lmcache.v1.multiprocess.engine_context.get_event_bus"),
+    ):
+        ctx = MPCacheEngineContext(
+            storage_manager_config=storage_manager_config, chunk_size=16
+        )
+
+    modules = _build_modules(ctx, MPServerConfig(transfer_mode="non_gpu", shm_name=""))
+    module = next(m for m in modules if isinstance(m, NonGPUTransferModule))
+    response = module.register_kv_cache_non_gpu_context(
+        RegisterNonGpuContextPayload(
+            instance_id=3,
+            model_name="m",
+            world_size=1,
+            block_size=4,
+            num_layers=2,
+            hidden_dim_size=16,
+            dtype_str="float32",
+            use_mla=False,
+        )
+    )
+
+    assert response.shm_name == ""
+    assert response.pool_size == 0
+
+
+def test_non_gpu_transfer_module_uses_mp_config_shm_name_override(
+    stub_native_storage_ops: Any,
+) -> None:
+    """Ensure MP-level shm override is normalized and returned to workers."""
+    # First Party
+    from lmcache.v1.multiprocess.config import MPServerConfig
+    from lmcache.v1.multiprocess.custom_types import RegisterNonGpuContextPayload
+    from lmcache.v1.multiprocess.engine_context import MPCacheEngineContext
+    from lmcache.v1.multiprocess.modules.non_gpu_transfer import NonGPUTransferModule
+
+    storage_manager_config = MagicMock()
+    storage_manager_config.l1_manager_config.memory_config.shm_name = None
+    storage_manager_config.l1_manager_config.memory_config.use_lazy = True
+    storage_manager_config.l1_manager_config.memory_config.size_in_bytes = 8192
+
+    with (
+        patch("lmcache.v1.multiprocess.engine_context.StorageManager"),
+        patch("lmcache.v1.multiprocess.engine_context.TokenHasher"),
+        patch("lmcache.v1.multiprocess.engine_context.SessionManager"),
+        patch("lmcache.v1.multiprocess.engine_context.get_event_bus"),
+    ):
+        ctx = MPCacheEngineContext(
+            storage_manager_config=storage_manager_config, chunk_size=16
+        )
+
+    module = NonGPUTransferModule(ctx, MPServerConfig(shm_name="worker_pool"))
+    response = module.register_kv_cache_non_gpu_context(
+        RegisterNonGpuContextPayload(
+            instance_id=4,
+            model_name="m",
+            world_size=1,
+            block_size=4,
+            num_layers=2,
+            hidden_dim_size=16,
+            dtype_str="float32",
+            use_mla=False,
+        )
+    )
+
+    assert response.shm_name == "lmcache_l1_pool_worker_pool"
+    assert response.pool_size == 8192
+
+
 def test_server_store_and_retrieve_cpu_chunks(stub_native_storage_ops: Any) -> None:
     """Validate mocked server-side CPU chunk store and retrieve behavior."""
     # First Party