Pass MPServerConfig.shm_name into NonGPUTransferModule

[Copilot]

--shm-name was defined on MPServerConfig but never reached the non-GPU transfer path, so --shm-name "" could not disable SHM and force pickle mode. This change keeps run_cache_server() unchanged and routes the override directly into NonGPUTransferModule, where the SHM decision is made.

• Server wiring

  • Update _build_modules() in lmcache/v1/multiprocess/server.py to construct NonGPUTransferModule(ctx, mp_config) for transfer_mode="non_gpu".

• Non-GPU SHM override handling

  • Extend NonGPUTransferModule.__init__ to accept optional MPServerConfig.
  • Store mp_config.shm_name as a module-local override instead of mutating shared storage config.
  • Update _compute_shm_pool_info() to prefer the override when present:
    • None → preserve existing auto-detect behavior from storage config
    • "" → disable SHM and return pickle-mode metadata
    • custom name → normalize to lmcache_l1_pool_* and return the configured pool size

• Regression coverage

  • Add focused tests covering:
    • _build_modules() passing the MP config into NonGPUTransferModule
    • empty-string override disabling SHM
    • explicit SHM-name override being normalized and returned during non-GPU context registration

Example of the new flow:

# server.py
else:
    modules.append(NonGPUTransferModule(ctx, mp_config))

# non_gpu_transfer.py
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
    )

Original prompt

Pass mp_config to NonGPUTransferModule so it can read shm_name directly, without changing run_cache_server() logic.

Changes needed:

1. lmcache/v1/multiprocess/server.py — _build_modules()

Change:

else:
    modules.append(NonGPUTransferModule(ctx))

To:

else:
    modules.append(NonGPUTransferModule(ctx, mp_config))

2. lmcache/v1/multiprocess/modules/non_gpu_transfer.py

Update __init__ to accept mp_config:

from lmcache.v1.multiprocess.config import MPServerConfig

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 else None
    self._non_gpu_contexts: dict[int, NonGPUContextEntry] = {}
    self._strategies: dict[int, TransferStrategy] = {}
    self._pending_shm_writes: dict[tuple[int, IPCCacheEngineKey], list[ObjectKey]] = {}
    self._pending_shm_reads: dict[tuple[int, IPCCacheEngineKey], list[ObjectKey]] = {}
    self._pending_shm_lock = threading.Lock()

Update _compute_shm_pool_info to use the override:

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 ""
    if not shm_name or mem_cfg.use_lazy:
        return {"shm_name": "", "pool_size": 0}
    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}

3. Ensure MPServerConfig has shm_name field

Check lmcache/v1/multiprocess/config.py — it should already have a shm_name field and --shm-name CLI arg. If not, add:

shm_name: str | None = None
"""Override SHM segment name. "" to force pickle, None for auto."""

And in add_mp_server_args:

parser.add_argument("--shm-name", type=str, default=None,
                    help='SHM segment name for non-GPU transfer. "" to disable SHM.')

Why:

--shm-name "" should disable SHM and force pickle mode. Currently the override never reaches _compute_shm_pool_info because nothing passes it from mp_config to the non-GPU module. This fix keeps the arg in MPServerConfig and lets NonGPUTransferModule consume it directly.

The following is the prior conversation context from the user's chat exploration (may be truncated):

User: 哪个文件，多少行

Assistant: 文件：lmcache/v1/multiprocess/server.py
第 193 行之后、第 194 行之前插入：

    # Apply shm_name override from MP config.
    if mp_config.shm_name is not None:
        storage_manager_config.l1_manager_config.memory_config.shm_name = (
            mp_config.shm_name
        )

即在 maybe_initialize_trace_recorder(...) 和 ctx = MPCacheEngineContext(...) 之间。

User: 可以不改server文件吗

Assistant: 可以。两个方案：

方案 A：改 engine_context.py——把 shm_name override 传入 context

# engine_context.py __init__
def __init__(self, storage_manager_config, chunk_size=256, hash_algorithm="blake3",
             shm_name_override: str | None = None):
    if shm_name_override is not None:
        storage_manager_config.l1_manager_config.memory_config.shm_name = shm_name_override
    ...

然后 server.py 调用处只需加个参数：ctx = MPCacheEngineContext(..., shm_name_override=mp_config.shm_name) — 但这还是改了 server.py 一行。

方案 B（完全不改 server.py）：改 config.py 的 parse_args_to_config

把 --shm-name 从 MP server args 移到 storage manager args 里，让 parse_args_to_config() 直接设置 memory_config.shm_name。这样 storage_manager_config 构建时就已经是正确的值，server.py 完全不用动。

但这改变了 --shm-name 的归属——从 "MP server 参数" 变成 "storage manager 参数"。语义上其实也说得通，因为 shm_name 本来就是 L1 memory 的配置。

---

最省事的还是在 server.py 加那 3 行。你不想改 server.py 的原因是什么？如果是想减少 diff/冲突，方案 B 改动反而更大。

User: 这个shm是non cuda context才有用的，server.py是上层抽象

Assistant: 对，那放在 NonGPUTransferModule.__init__ 里最合理：

# modules/non_gpu_transfer.py __init__
def __init__(self, ctx: MPCacheEngineContext, shm_name_override: str | None = None) -> None:
    self._ctx = ctx
    if shm_name_override is not None:
        ctx.storage_manager_config.l1_manager_config.memory_config.shm_name = shm_name_override
    ...

然后 _build_modules 里：

# server.py _build_modules
else:
    modules.append(NonGPUTransferModule(ctx, shm_name_override=mp_config.shm_name))

这样 `run_cache_serve...

This pull request was created from Copilot chat.