[FEAT] Support fast engine recovery through weight cache (#27139)

Signed-off-by: Michael Qiu <qiudayu.qdy@antgroup.com>
Co-authored-by: liusy58 <liusy58@linux.alibaba.com>
Co-authored-by: Alex Nails <alex.nails@radixark.ai>

python/sglang/srt/configs/load_config.py

@@ -34,6 +34,7 @@ class LoadFormat(str, enum.Enum):
     FASTSAFETENSORS = "fastsafetensors"
     PRIVATE = "private"
     RUNAI_STREAMER = "runai_streamer"
+    IPC_CACHE = "ipc_cache"
 
 
 @dataclass
@@ -104,6 +105,11 @@ class LoadConfig:
     # For multi-layer MTP
     draft_model_idx: Optional[int] = None
 
+    # Weight cache daemon options
+    weight_cache_mode: str = "off"  # "off", "daemon", "client"
+    weight_cache_socket: Optional[str] = None  # Path to daemon socket (for client mode)
+    fallback_load_format: Union[str, "LoadFormat"] = LoadFormat.AUTO
+
     def __post_init__(self):
         model_loader_extra_config = self.model_loader_extra_config or {}
         if isinstance(model_loader_extra_config, str):

python/sglang/srt/entrypoints/engine.py

@@ -27,6 +27,8 @@ import multiprocessing as mp
 import os
 import random
 import signal
+import subprocess
+import sys
 import tempfile
 import threading
 import time
@@ -110,7 +112,12 @@ from sglang.srt.utils import (
     set_ulimit,
 )
 from sglang.srt.utils.msgspec_utils import msgspec_to_builtins
-from sglang.srt.utils.network import get_zmq_socket, is_port_available
+from sglang.srt.utils.network import (
+    NetworkAddress,
+    get_free_port,
+    get_zmq_socket,
+    is_port_available,
+)
 from sglang.srt.utils.torch_memory_saver_adapter import TorchMemorySaverAdapter
 from sglang.srt.utils.watchdog import SubprocessWatchdog
 from sglang.version import __version__
@@ -238,6 +245,7 @@ class Engine(EngineScoreMixin, EngineBase):
             port_args,
             scheduler_init_result,
             subprocess_watchdog,
+            weight_cache_daemon_procs,
         ) = self._launch_subprocesses(
             server_args=server_args,
             init_tokenizer_manager_func=self.init_tokenizer_manager_func,
@@ -247,6 +255,10 @@ class Engine(EngineScoreMixin, EngineBase):
         self.tokenizer_manager = tokenizer_manager
         self.template_manager = template_manager
         self._scheduler_init_result = scheduler_init_result
+        # Engine-spawned weight cache daemons owned by *this* instance (empty
+        # unless --weight-cache-mode daemon). Kept per-instance so two Engines
+        # in one process each reap only their own daemons in shutdown().
+        self._weight_cache_daemon_procs = weight_cache_daemon_procs
         if tokenizer_manager is not None:
             tokenizer_manager._subprocess_watchdog = subprocess_watchdog
         self.port_args = port_args
@@ -588,6 +600,208 @@ class Engine(EngineScoreMixin, EngineBase):
         ret = self.loop.run_until_complete(generator.__anext__())
         return ret
 
+    @classmethod
+    def _launch_weight_cache_daemons(cls, server_args: ServerArgs):
+        """Launch weight cache daemon processes for this node's PP×TP ranks.
+
+        All daemon processes join the same NCCL distributed group so that
+        TP-sharded model loading works correctly. Each daemon holds its
+        rank's weight shard in GPU memory and serves IPC handles.
+
+        Lifecycle: these daemons are *co-terminal* with the engine. They are
+        children of this process (kill_itself_when_parent_died installs
+        PR_SET_PDEATHSIG) and are gracefully reaped in ``shutdown()``. They do
+        NOT persist across engine restarts, so ``--weight-cache-mode daemon``
+        on its own does not deliver a faster restart -- the first start is in
+        fact slower (disk-load into the daemon plus the IPC handshake). The
+        fast-recovery story is the standalone launcher
+        (``python -m sglang.srt.weight_cache.daemon``) plus
+        ``--weight-cache-mode client``, where the daemon outlives the engine.
+        """
+        if server_args.dp_size > 1:
+            raise ValueError(
+                "Weight cache daemon mode does not support dp_size > 1. "
+                "Please set --dp-size 1 when using --weight-cache-mode daemon."
+            )
+
+        # Multi-node needs an explicit rendezvous address; otherwise each node
+        # picks its own local 127.0.0.1 port (below) and the per-node daemons
+        # can never form the joint process group.
+        if server_args.nnodes > 1 and not server_args.dist_init_addr:
+            raise ValueError(
+                "Multi-node weight cache daemons (nnodes > 1) require "
+                "--dist-init-addr so all nodes rendezvous at the same endpoint."
+            )
+
+        tp_size = server_args.tp_size
+
+        pp_rank_range, tp_rank_range, pp_size_per_node, tp_size_per_node = (
+            _calculate_rank_ranges(
+                server_args.nnodes,
+                server_args.pp_size,
+                tp_size,
+                server_args.node_rank,
+            )
+        )
+
+        # Build the distributed init method (multi-node uses the user-provided
+        # dist_init_addr so all nodes reach the same endpoint).
+        if server_args.dist_init_addr:
+            host, port = server_args.dist_init_addr.rsplit(":", 1)
+            dist_init_method = f"tcp://{host}:{port}"
+        else:
+            # Fresh free port for the daemons' own rendezvous, not the engine's
+            # nccl_port: a pinned --nccl-port would otherwise collide with the
+            # engine's own NCCL TCPStore.
+            dist_init_method = NetworkAddress("127.0.0.1", get_free_port()).to_tcp()
+
+        num_daemons = len(pp_rank_range) * len(tp_rank_range)
+        daemon_procs = []
+        logger.info(
+            f"Launching {num_daemons} weight cache daemon(s) on node "
+            f"{server_args.node_rank} for model={server_args.model_path}, "
+            f"pp_ranks={pp_rank_range.start}..{pp_rank_range.stop - 1}, "
+            f"tp_ranks={tp_rank_range.start}..{tp_rank_range.stop - 1}, "
+            f"dist_init_method={dist_init_method}"
+        )
+
+        # Validate and clean up stale .ready/.sock files from prior runs.
+        # If a daemon is still alive at this rank, raise instead of clobbering.
+        from sglang.srt.weight_cache.protocol import (
+            cleanup_stale_daemon_files,
+            compute_global_rank,
+            compute_local_gpu_id,
+            get_ready_path,
+        )
+
+        for pp_rank in pp_rank_range:
+            for tp_rank in tp_rank_range:
+                global_rank = compute_global_rank(tp_size, pp_rank, tp_rank)
+                cleanup_stale_daemon_files(global_rank)
+
+        for pp_rank in pp_rank_range:
+            for tp_rank in tp_rank_range:
+                gpu_id = compute_local_gpu_id(
+                    pp_rank,
+                    tp_rank,
+                    pp_size_per_node,
+                    tp_size_per_node,
+                    base_gpu_id=server_args.base_gpu_id,
+                    gpu_id_step=server_args.gpu_id_step,
+                )
+                cmd = [
+                    sys.executable,
+                    "-m",
+                    "sglang.srt.weight_cache.daemon",
+                    "--model-path",
+                    server_args.model_path,
+                    "--gpu-id",
+                    str(gpu_id),
+                    "--tp-size",
+                    str(tp_size),
+                    "--tp-rank",
+                    str(tp_rank),
+                    "--pp-size",
+                    str(server_args.pp_size),
+                    "--pp-rank",
+                    str(pp_rank),
+                    "--dp-size",
+                    "1",
+                    "--ep-size",
+                    str(server_args.ep_size),
+                    "--load-format",
+                    server_args.load_format,
+                    "--dtype",
+                    server_args.dtype,
+                    "--dist-init-method",
+                    dist_init_method,
+                ]
+                if server_args.quantization:
+                    cmd += ["--quantization", server_args.quantization]
+                if (
+                    server_args.model_loader_extra_config
+                    and server_args.model_loader_extra_config != "{}"
+                ):
+                    cmd += [
+                        "--model-loader-extra-config",
+                        server_args.model_loader_extra_config,
+                    ]
+                if server_args.trust_remote_code:
+                    cmd += ["--trust-remote-code"]
+                if server_args.revision:
+                    cmd += ["--revision", server_args.revision]
+
+                proc = subprocess.Popen(cmd)
+                daemon_procs.append(proc)
+
+        # Wait for all daemons to be ready (ready file exists). On any failure
+        # (readiness timeout or a daemon exiting early) terminate the siblings
+        # we already spawned before propagating, so a partial launch does not
+        # leak GPU-resident daemons.
+        timeout = server_args.weight_cache_timeout
+        check_interval = 2
+        start_time = time.time()
+        try:
+            for pp_rank in pp_rank_range:
+                for tp_rank in tp_rank_range:
+                    global_rank = compute_global_rank(tp_size, pp_rank, tp_rank)
+                    ready_path = get_ready_path(global_rank)
+                    while not os.path.exists(ready_path):
+                        time.sleep(check_interval)
+                        if time.time() - start_time > timeout:
+                            raise TimeoutError(
+                                f"Weight cache daemon for pp_rank={pp_rank} "
+                                f"tp_rank={tp_rank} did not become ready "
+                                f"within {timeout}s"
+                            )
+                        # Check if daemon process is still alive
+                        for p in daemon_procs:
+                            if p.poll() is not None:
+                                raise RuntimeError(
+                                    f"Weight cache daemon (pid={p.pid}) exited prematurely "
+                                    f"with code {p.returncode}"
+                                )
+                    logger.info(
+                        f"Weight cache daemon for pp_rank={pp_rank} "
+                        f"tp_rank={tp_rank} is ready"
+                    )
+        except BaseException:
+            cls._terminate_weight_cache_daemons(daemon_procs)
+            raise
+
+        logger.info(
+            f"All {num_daemons} weight cache daemons on node "
+            f"{server_args.node_rank} are ready"
+        )
+        return daemon_procs
+
+    @staticmethod
+    def _terminate_weight_cache_daemons(procs, timeout: float = 10.0):
+        """Gracefully stop engine-spawned weight cache daemons.
+
+        Send SIGTERM first so each daemon's signal handler can unlink its
+        ``.sock``/``.ready`` files, then SIGKILL any straggler. This matters
+        because ``shutdown()`` otherwise reaps children via
+        ``kill_process_tree`` (SIGKILL), which would skip that cleanup and
+        leave stale files that make the next client-mode boot fail with a
+        confusing "socket exists but connection refused" instead of a clean
+        "no daemon" path.
+        """
+        if not procs:
+            return
+        for p in procs:
+            if p.poll() is None:
+                p.terminate()  # SIGTERM -> daemon cleanup handler runs
+        for p in procs:
+            try:
+                p.wait(timeout=timeout)
+            except subprocess.TimeoutExpired:
+                logger.warning(
+                    f"Weight cache daemon (pid={p.pid}) did not exit within "
+                    f"{timeout}s of SIGTERM; sending SIGKILL."
+                )
+                p.kill()
+
     @classmethod
     def _launch_scheduler_processes(
         cls,
@@ -779,7 +993,7 @@ class Engine(EngineScoreMixin, EngineBase):
         """Launch the TokenizerManager in the main process, the Scheduler in a subprocess, and the DetokenizerManager in another subprocess.
 
         Returns:
-            Tuple of (tokenizer_manager, template_manager, port_args, scheduler_init_result, subprocess_watchdog).
+            Tuple of (tokenizer_manager, template_manager, port_args, scheduler_init_result, subprocess_watchdog, weight_cache_daemon_procs).
         """
         # Configure global environment
         configure_logger(server_args)
@@ -820,6 +1034,13 @@ class Engine(EngineScoreMixin, EngineBase):
         ):
             resolve_auto_parsers(server_args)
 
+        # Launch daemons (daemon mode only). Handles are threaded back to the
+        # owning Engine instance (not a class attr) so two Engines in one process
+        # don't clobber each other's daemon list.
+        weight_cache_daemon_procs: List = []
+        if server_args.weight_cache_mode == "daemon":
+            weight_cache_daemon_procs = cls._launch_weight_cache_daemons(server_args)
+
         # Launch scheduler processes
         scheduler_init_result, scheduler_procs = cls._launch_scheduler_processes(
             server_args, port_args, run_scheduler_process_func
@@ -846,6 +1067,7 @@ class Engine(EngineScoreMixin, EngineBase):
                     port_args,
                     scheduler_init_result,
                     None,
+                    weight_cache_daemon_procs,
                 )
 
             launch_dummy_health_check_server(
@@ -859,6 +1081,7 @@ class Engine(EngineScoreMixin, EngineBase):
                 port_args,
                 scheduler_init_result,
                 None,
+                weight_cache_daemon_procs,
             )
 
         # Launch detokenizer process(es) — optionally fronted by a router when
@@ -906,6 +1129,7 @@ class Engine(EngineScoreMixin, EngineBase):
             port_args,
             scheduler_init_result,
             subprocess_watchdog,
+            weight_cache_daemon_procs,
         )
 
     def shutdown(self):
@@ -923,6 +1147,14 @@ class Engine(EngineScoreMixin, EngineBase):
             send_to_rpc.close(linger=0)
             self.send_to_rpc = None
 
+        # Gracefully stop weight cache daemons *before* the blanket
+        # kill_process_tree below, so their SIGTERM handlers can unlink the
+        # .sock/.ready files instead of being SIGKILLed and leaving stale state.
+        daemon_procs = getattr(self, "_weight_cache_daemon_procs", None)
+        if daemon_procs:
+            self._terminate_weight_cache_daemons(daemon_procs)
+            self._weight_cache_daemon_procs = []
+
         kill_process_tree(os.getpid(), include_parent=False, wait_timeout=60)
 
     def __enter__(self):

python/sglang/srt/entrypoints/http_server.py

@@ -2675,6 +2675,7 @@ def launch_server(
         port_args,
         scheduler_init_result,
         subprocess_watchdog,
+        _weight_cache_daemon_procs,
     ) = Engine._launch_subprocesses(
         server_args=server_args,
         init_tokenizer_manager_func=init_tokenizer_manager_func,

python/sglang/srt/managers/scheduler_components/weight_updater.py

@@ -183,6 +183,22 @@ class SchedulerWeightUpdaterManager:
         parameter = self.tp_worker.get_weights_by_name(recv_req)
         return GetWeightsByNameReqOutput(parameter=parameter)
 
+    def _assert_weight_cache_inactive(self, op: str) -> None:
+        """Reject freeing/restoring model weights while the CUDA IPC weight
+        cache is active: the weights are shared with the daemon via CUDA IPC, so
+        freeing them would leave the daemon and every peer pointing at released
+        memory.
+        """
+        mode = self.tp_worker.model_runner.server_args.weight_cache_mode
+        if mode != "off":
+            raise RuntimeError(
+                f"[weight_cache] {op} of model weights is not supported while the "
+                f"weight cache is active (--weight-cache-mode {mode}): the weights "
+                f"are shared with the daemon via CUDA IPC, so freeing them would "
+                f"corrupt the daemon's master copy and every co-attached engine. "
+                f"Restart with --weight-cache-mode off to use this operation."
+            )
+
     def release_memory_occupation(self, recv_req: ReleaseMemoryOccupationReqInput):
         assert (
             self.is_fully_idle()
@@ -215,6 +231,7 @@ class SchedulerWeightUpdaterManager:
             self.flush_cache()
 
         if GPU_MEMORY_TYPE_WEIGHTS in tags:
+            self._assert_weight_cache_inactive("release_memory_occupation")
             self.stashed_model_static_state = _export_static_state(
                 self.tp_worker.model_runner.model
             )
@@ -241,6 +258,7 @@ class SchedulerWeightUpdaterManager:
             self.memory_saver_adapter.resume(GPU_MEMORY_TYPE_CUDA_GRAPH)
 
         if GPU_MEMORY_TYPE_WEIGHTS in tags:
+            self._assert_weight_cache_inactive("resume_memory_occupation")
             self.memory_saver_adapter.resume(GPU_MEMORY_TYPE_WEIGHTS)
             torch.distributed.barrier(self.tp_cpu_group)
             _import_static_state(

python/sglang/srt/model_executor/model_runner.py

@@ -128,6 +128,7 @@ from sglang.srt.model_executor.model_runner_components.load_model_utils import (
     load_kv_cache_scales,
     load_model_with_memory_saver,
     maybe_downgrade_dtype_for_legacy_gpu,
+    maybe_enable_ipc_weight_cache,
     maybe_register_debug_tensor_dump_hook,
     maybe_trigger_remote_instance_nccl_send_group,
     report_online_quantization,
@@ -987,6 +988,18 @@ class ModelRunner:
             remote_instance_weight_transporter_engine=self.remote_instance_weight_transporter.engine,
             remote_instance_weight_transporter_session_id=self.remote_instance_weight_transporter.session_id,
             draft_model_idx=self.draft_model_idx,
+            weight_cache_mode=self.server_args.weight_cache_mode,
+            weight_cache_socket=self.server_args.weight_cache_socket,
+        )
+
+        # If the weight cache is enabled, override the load format to IPC_CACHE
+        # and derive the per-rank daemon socket. Idempotent across reloads.
+        maybe_enable_ipc_weight_cache(
+            load_config=self.load_config,
+            server_args=self.server_args,
+            tp_size=self.ps.tp_size,
+            pp_rank=self.ps.pp_rank,
+            tp_rank=self.ps.tp_rank,
         )
         if self.device == "cpu":
             self.model_config = adjust_config_with_unaligned_cpu_tp(

python/sglang/srt/model_executor/model_runner_components/load_model_utils.py

@@ -172,6 +172,8 @@ def build_load_config(
     remote_instance_weight_transporter_engine: Any,
     remote_instance_weight_transporter_session_id: str,
     draft_model_idx: Optional[int],
+    weight_cache_mode: str,
+    weight_cache_socket: Optional[str],
 ) -> LoadConfig:
     from sglang.srt.configs.modelopt_config import ModelOptConfig
 
@@ -199,9 +201,45 @@ def build_load_config(
         modelopt_config=modelopt_config,
         rl_quant_profile=server_args.rl_quant_profile,
         draft_model_idx=draft_model_idx,
+        weight_cache_mode=weight_cache_mode,
+        weight_cache_socket=weight_cache_socket,
     )
 
 
+def maybe_enable_ipc_weight_cache(
+    *,
+    load_config: LoadConfig,
+    server_args: ServerArgs,
+    tp_size: int,
+    pp_rank: int,
+    tp_rank: int,
+) -> None:
+    """Switch ``load_config`` onto the IPC weight-cache path, in place.
+
+    Overrides the load format to ``IPC_CACHE`` (remembering the original as the
+    disk fallback) and derives the per-rank daemon socket if unset. Idempotent:
+    the format swap is guarded on ``!= IPC_CACHE`` so a second call (e.g. a
+    weight reload) can't overwrite the captured fallback format.
+    """
+    if server_args.weight_cache_mode == "off":
+        return
+
+    if load_config.load_format != LoadFormat.IPC_CACHE:
+        load_config.fallback_load_format = load_config.load_format
+        load_config.load_format = LoadFormat.IPC_CACHE
+
+    # Compute socket path using global rank (tp_size * pp_rank + tp_rank) so
+    # each daemon has a unique socket even across PP stages and nodes.
+    if load_config.weight_cache_socket is None:
+        from sglang.srt.weight_cache.protocol import (
+            compute_global_rank,
+            get_socket_path,
+        )
+
+        global_rank = compute_global_rank(tp_size, pp_rank, tp_rank)
+        load_config.weight_cache_socket = get_socket_path(global_rank=global_rank)
+
+
 def load_model_with_memory_saver(
     *,
     server_args: ServerArgs,
@@ -218,6 +256,17 @@ def load_model_with_memory_saver(
     enable_cpu_backup = server_args.enable_weights_cpu_backup or (
         is_draft_worker and server_args.enable_draft_weights_cpu_backup
     )
+
+    # In zero-copy IPC mode, the weights are shared with the daemon via
+    # CUDA IPC and must not be offloaded/reloaded by the memory saver.
+    is_ipc_zero_copy = server_args.weight_cache_mode != "off"
+    if is_ipc_zero_copy and enable_cpu_backup:
+        logger.warning(
+            "[ModelRunner] Disabling weights CPU backup in zero-copy IPC mode — "
+            "IPC-mapped weights cannot be offloaded to CPU."
+        )
+        enable_cpu_backup = False
+
     remote_instance_weight_info = None
     with memory_saver_adapter.region(
         GPU_MEMORY_TYPE_WEIGHTS,

python/sglang/srt/model_executor/model_runner_components/weight_updater.py

@@ -123,6 +123,21 @@ class WeightUpdater:
             logger.error(message)
             return False, message
 
+    def _assert_weight_cache_inactive(self: WeightUpdater, op: str) -> None:
+        """Reject weight mutations while the CUDA IPC weight cache is active:
+        param.data is the daemon's master copy shared with every co-attached
+        engine, so an in-place update would silently corrupt them all.
+        """
+        mode = self.get_model_runner().server_args.weight_cache_mode
+        if mode != "off":
+            raise RuntimeError(
+                f"[weight_cache] {op} is not supported while the weight cache is "
+                f"active (--weight-cache-mode {mode}): model weights are shared "
+                f"with the daemon via CUDA IPC, so mutating them in place would "
+                f"corrupt the daemon's master copy and every co-attached engine. "
+                f"Restart with --weight-cache-mode off to use this operation."
+            )
+
     def update_weights_from_disk(
         self: WeightUpdater,
         model_path: str,
@@ -131,6 +146,7 @@ class WeightUpdater:
         recapture_cuda_graph: bool = False,
     ) -> tuple[bool, str]:
         """Update engine weights in-place from the disk."""
+        self._assert_weight_cache_inactive("update_weights_from_disk")
         error = _unsupported_derived_weight_cache_error()
         if error is not None:
             return False, error
@@ -220,6 +236,7 @@ class WeightUpdater:
             dtype: the data type of the parameter to be updated.
             shape: the shape of the parameter to be updated.
         """
+        self._assert_weight_cache_inactive("update_weights_from_distributed")
         error = _unsupported_derived_weight_cache_error()
         if error is not None:
             return False, error
@@ -309,6 +326,7 @@ class WeightUpdater:
             return False, error
 
         monkey_patch_torch_reductions()
+        self._assert_weight_cache_inactive("update_weights_from_tensor")
         if load_format == "flattened_bucket":
             # Handle flattened bucket format
             return self._update_weights_from_flattened_bucket(
@@ -368,6 +386,7 @@ class WeightUpdater:
 
     def update_weights_from_ipc(self: WeightUpdater, recv_req):
         """Update weights from IPC for checkpoint-engine integration."""
+        self._assert_weight_cache_inactive("update_weights_from_ipc")
         error = _unsupported_derived_weight_cache_error()
         if error is not None:
             return False, error

python/sglang/srt/model_loader/loader.py

@@ -3319,4 +3319,26 @@ def get_model_loader(
     if load_config.load_format == LoadFormat.RUNAI_STREAMER:
         return RunaiModelStreamerLoader(load_config)
 
+    if load_config.load_format == LoadFormat.IPC_CACHE:
+        from sglang.srt.weight_cache.ipc_loader import IpcModelLoader
+        from sglang.srt.weight_cache.protocol import (
+            compute_global_rank,
+            get_socket_path,
+        )
+
+        if load_config.weight_cache_socket:
+            socket_path = load_config.weight_cache_socket
+        else:
+            from sglang.srt.runtime_context import get_parallel
+
+            ps = get_parallel()
+            global_rank = compute_global_rank(ps.tp_size, ps.pp_rank, ps.tp_rank)
+            socket_path = get_socket_path(global_rank=global_rank)
+        return IpcModelLoader(
+            load_config=load_config,
+            socket_path=socket_path,
+            weight_cache_mode=load_config.weight_cache_mode,
+            fallback_load_format=load_config.fallback_load_format,
+        )
+
     return DefaultModelLoader(load_config)

python/sglang/srt/ray/http_server.py

@@ -52,6 +52,7 @@ def launch_server(
         port_args,
         scheduler_init_result,
         subprocess_watchdog,
+        _weight_cache_daemon_procs,
     ) = RayEngine._launch_subprocesses(
         server_args,
         init_tokenizer_manager_func=init_tokenizer_manager_func,

python/sglang/srt/server_args.py

@@ -124,6 +124,10 @@ LOAD_FORMAT_CHOICES = [
     "private",
     "runai_streamer",
 ]
+# NOTE: LoadFormat.IPC_CACHE intentionally has no public --load-format choice.
+# It is an internal dispatch format set automatically by ModelRunner when the
+# weight cache is enabled (weight_cache_mode != "off"). Exposing it as a CLI
+# choice let users create contradictory combos (see _handle_load_format).
 
 # TODO: this list should likely contain only methods that support online quantization, or that support using custom quantization classes compatible with a given `quant_method` in config.json.
 # Some of the choices here do NOT support online quantization.
@@ -3302,6 +3306,36 @@ class ServerArgs:
         NS("exec.features"),
     ] = False
 
+    weight_cache_mode: A[
+        str,
+        Arg(
+            help="Weight cache mode. 'off': normal disk loading. "
+            "'daemon': launch weight cache daemon (holds weights in GPU memory). "
+            "Engine-spawned daemons are co-terminal with the engine and do NOT "
+            "persist across restarts, so this alone does not speed up restart "
+            "(the first start is slower). For fast recovery, run the standalone "
+            "daemon (python -m sglang.srt.weight_cache.daemon) and connect with "
+            "'client'. 'client': connect to existing daemon and load via IPC.",
+            choices=["off", "daemon", "client"],
+        ),
+        NS("model"),
+    ] = "off"
+    weight_cache_socket: A[
+        Optional[str],
+        Arg(
+            help="Unix socket path for weight cache daemon (client mode)."
+            "If not set, uses /tmp/sglang_weight_cache_rank{global_rank}.sock",
+        ),
+        NS("model"),
+    ] = None
+    weight_cache_timeout: A[
+        int,
+        Arg(
+            help="Timeout in seconds for weight cache daemon readiness (default: 1800).",
+        ),
+        NS("model"),
+    ] = 1800
+
     # -------------------------------------------------------------------------
     # Custom hooks, probe, and plugins
     # -------------------------------------------------------------------------
@@ -6926,6 +6960,31 @@ class ServerArgs:
                 self.validate_transfer_engine()
             )
 
+        # "ipc_cache" is an internal-only load format: ModelRunner sets it
+        # automatically when the weight cache is enabled, and it is not a public
+        # --load-format choice. Setting it directly is always wrong (no daemon is
+        # launched, and fallback_load_format inherits a nonsensical format), so
+        # reject it and point at the knob (defense-in-depth; the CLI already
+        # rejects it via LOAD_FORMAT_CHOICES).
+        if self.load_format == "ipc_cache":
+            raise ValueError(
+                "load_format='ipc_cache' is an internal-only format and must not "
+                "be set directly. Enable the weight cache via --weight-cache-mode "
+                "client (connect to an existing daemon) or daemon (launch one); "
+                "that selects IPC loading automatically."
+            )
+
+        # Speculative decoding loads an extra draft model whose weights the
+        # daemon does not export, so refuse the combination up front instead of
+        # failing deep inside draft-worker load (draft-model daemon TBD).
+        if self.weight_cache_mode != "off" and self.speculative_algorithm is not None:
+            raise ValueError(
+                "--weight-cache-mode is not supported together with speculative "
+                "decoding (--speculative-algorithm): the weight cache daemon does "
+                "not export the draft model's weights. Disable one of them "
+                "(--weight-cache-mode off) for this configuration."
+            )
+
     def _is_mistral_native_format(self) -> bool:
         """True iff the checkpoint requires load_format=mistral.
 

python/sglang/srt/weight_cache/__init__.py

@@ -0,0 +1,11 @@
+# SPDX-License-Identifier: Apache-2.0
+#
+# Intentionally empty: importing any weight_cache submodule (e.g.
+# ``sglang.srt.weight_cache.protocol``) executes this package __init__ first.
+# Eagerly re-exporting daemon/ipc_loader here would pull in torch and the model
+# loader on that cheap protocol import, re-introducing the circular-import and
+# startup-cost problems the local-import refactor removed. Import the concrete
+# symbols from their submodules instead, e.g.
+#     from sglang.srt.weight_cache.protocol import CacheConfig
+#     from sglang.srt.weight_cache.daemon import launch_weight_cache_daemons
+#     from sglang.srt.weight_cache.ipc_loader import IpcModelLoader

python/sglang/srt/weight_cache/daemon.py

@@ -0,0 +1,944 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Weight Cache Daemon — a persistent process that holds post-quantized,
+TP-sharded model weights in GPU memory and serves them via CUDA IPC handles.
+
+Each GPU runs one daemon process for its TP rank. The daemon:
+1. Loads model weights from disk (full pipeline: disk → TP shard → quantize)
+2. Exports every parameter/buffer as a CUDA IPC handle
+3. Serves handles over a Unix socket to requesting engine processes
+4. Validates CacheConfig compatibility before serving
+
+Usage:
+    # Single-node: launch all TP rank daemons with a single command:
+    python -m sglang.srt.weight_cache.daemon \
+        --model-path /path/to/model --tp-size 4 \
+        --load-format auto --dtype auto --quantization fp8
+
+    # Multi-node: run on each node with --nnodes and --node-rank:
+    # Node 0:
+    python -m sglang.srt.weight_cache.daemon \
+        --model-path /path/to/model --tp-size 16 \
+        --nnodes 2 --node-rank 0 \
+        --dist-init-method tcp://node0-ip:29500
+
+    # Node 1:
+    python -m sglang.srt.weight_cache.daemon \
+        --model-path /path/to/model --tp-size 16 \
+        --nnodes 2 --node-rank 1 \
+        --dist-init-method tcp://node0-ip:29500
+
+    # Or launch a single daemon for a specific rank:
+    python -m sglang.srt.weight_cache.daemon \
+        --model-path /path/to/model \
+        --gpu-id 0 --tp-size 4 --tp-rank 0 \
+        --dist-init-method tcp://127.0.0.1:29500
+"""
+
+import logging
+import os
+import signal
+import socket
+import time
+from typing import Any, Dict, Optional
+
+import torch
+import torch.distributed as dist
+
+from sglang.srt.configs.load_config import LoadConfig
+from sglang.srt.platforms import current_platform
+from sglang.srt.utils import MultiprocessingSerializer
+
+from .protocol import (
+    CacheConfig,
+    check_ipc_quant_support,
+    cleanup_stale_daemon_files,
+    compute_env_stamp,
+    compute_global_rank,
+    compute_local_gpu_id,
+    get_quant_method_name,
+    get_ready_path,
+    get_socket_path,
+    hash_quant_config,
+    recv_msg,
+    send_msg,
+)
+
+logger = logging.getLogger(__name__)
+
+# Per-connection timeout for the serial serve loop. A client exchange is tiny
+# (a config dict + IPC handle metadata), so this generous bound never trips a
+# healthy client, yet guarantees one hung/dead peer can't stall the other
+# engine ranks indefinitely.
+CLIENT_CONNECTION_TIMEOUT = 30.0
+
+
+class WeightCacheDaemon:
+    """Persistent GPU weight cache for a single TP rank.
+
+    Holds the complete post-quantization state_dict in GPU memory and
+    serves CUDA IPC handles to engine processes via Unix socket.
+    """
+
+    def __init__(
+        self,
+        model_path: str,
+        gpu_id: int,
+        tp_size: int = 1,
+        tp_rank: int = 0,
+        pp_size: int = 1,
+        pp_rank: int = 0,
+        dp_size: int = 1,
+        ep_size: int = 1,
+        load_format: str = "auto",
+        dtype: str = "auto",
+        quantization: Optional[str] = None,
+        model_loader_extra_config: str = "{}",
+        trust_remote_code: bool = False,
+        revision: Optional[str] = None,
+        dist_init_method: Optional[str] = None,
+    ):
+        self.model_path = model_path
+        self.gpu_id = gpu_id
+        self.tp_size = tp_size
+        self.tp_rank = tp_rank
+        self.pp_size = pp_size
+        self.pp_rank = pp_rank
+        self.dp_size = dp_size
+        self.ep_size = ep_size
+        self.load_format = load_format
+        self.dtype = dtype
+        self.quantization = quantization
+        self.model_loader_extra_config = model_loader_extra_config
+        self.trust_remote_code = trust_remote_code
+        self.revision = revision
+        self.dist_init_method = dist_init_method
+
+        self.socket_path = get_socket_path(
+            compute_global_rank(tp_size, pp_rank, tp_rank)
+        )
+        self.ready_path = get_ready_path(compute_global_rank(tp_size, pp_rank, tp_rank))
+
+        self.model = None
+        self.config: Optional[CacheConfig] = None
+        # name -> {"handle": base64_str, "shape": list, "dtype": str, "is_param": bool}
+        self.state_entries: Dict[str, Dict[str, Any]] = {}
+
+    def _init_distributed(self, server_args, model_config):
+        """Initialize the distributed backend required for model loading.
+
+        Uses the same world_size/rank formula as the engine:
+            world_size = tp_size * pp_size
+            rank = tp_size * pp_rank + tp_rank
+        """
+        from sglang.srt.distributed.parallel_state import (
+            init_distributed_environment,
+            initialize_model_parallel,
+            model_parallel_is_initialized,
+        )
+
+        if model_parallel_is_initialized():
+            logger.info(
+                f"[WeightCacheDaemon gpu={self.gpu_id}] "
+                f"Distributed already initialized, skipping"
+            )
+            return
+
+        # Initialize distributed environment
+        import torch.distributed as dist
+
+        if not dist.is_initialized():
+            if self.dist_init_method is None:
+                # Fallback: auto-assign a port. This only works for single-process.
+                import socket as sock_mod
+
+                with sock_mod.socket(sock_mod.AF_INET, sock_mod.SOCK_STREAM) as s:
+                    s.bind(("127.0.0.1", 0))
+                    free_port = s.getsockname()[1]
+                self.dist_init_method = f"tcp://127.0.0.1:{free_port}"
+
+            init_distributed_environment(
+                world_size=self.tp_size * self.pp_size,
+                rank=compute_global_rank(self.tp_size, self.pp_rank, self.tp_rank),
+                distributed_init_method=self.dist_init_method,
+                local_rank=self.gpu_id,
+                backend=current_platform.get_torch_distributed_backend_str(),
+            )
+
+        initialize_model_parallel(
+            tensor_model_parallel_size=self.tp_size,
+            pipeline_model_parallel_size=self.pp_size,
+            expert_model_parallel_size=self.ep_size,
+        )
+
+        # Initialize DP attention state (required by some models like Qwen3 MoE)
+        from sglang.srt.layers.dp_attention import initialize_dp_attention
+
+        initialize_dp_attention(server_args, model_config)
+
+        logger.info(
+            f"[WeightCacheDaemon gpu={self.gpu_id} tp_rank={self.tp_rank}] "
+            f"Distributed backend initialized (tp_size={self.tp_size}, "
+            f"pp_size={self.pp_size}, "
+            f"world_size={self.tp_size * self.pp_size})"
+        )
+
+    def load(self):
+        """Full loading pipeline: disk → TP shard → quantize → export IPC handles."""
+        # CUDA IPC weight sharing relies on torch's _share_cuda_ handle export,
+        # which only exists on CUDA-alike platforms (CUDA / ROCm). Fail loud here
+        # instead of dying deep inside the export with an opaque error.
+        if not current_platform.is_cuda_alike():
+            raise RuntimeError(
+                f"[WeightCacheDaemon] the weight cache daemon requires a CUDA-alike "
+                f"platform (CUDA or ROCm) for CUDA IPC weight sharing, but the "
+                f"active platform device type is {current_platform.device_type!r}. "
+                f"Disable the weight cache (--weight-cache-mode off)."
+            )
+        # expandable_segments makes torch's caching allocator hand out memory
+        # that cannot be exported via _share_cuda_, so the IPC export below would
+        # die mid-way with an opaque CUDA error. Fail fast with an actionable
+        # message before touching the device.
+        self._assert_ipc_compatible_allocator()
+        current_platform.set_device(current_platform.get_device(self.gpu_id))
+
+        # Reduce thread contention during multi-process loading
+        torch.set_num_threads(1)
+
+        # Lazy imports to avoid circular dependencies and speed up startup
+        from sglang.srt.configs.device_config import DeviceConfig
+        from sglang.srt.configs.model_config import ModelConfig
+        from sglang.srt.model_loader.loader import get_model_loader
+        from sglang.srt.runtime_context import get_context
+        from sglang.srt.server_args import ServerArgs
+
+        server_args = ServerArgs(
+            model_path=self.model_path,
+            dtype=self.dtype,
+            quantization=self.quantization,
+            trust_remote_code=self.trust_remote_code,
+            tp_size=self.tp_size,
+            pp_size=self.pp_size,
+            dp_size=self.dp_size,
+            ep_size=self.ep_size,
+            load_format=self.load_format,
+            model_loader_extra_config=self.model_loader_extra_config,
+        )
+        get_context().set_server_args(server_args)
+
+        # Initialize distributed backend for model loading
+        # (must be done after server_args and model_config are available)
+        # Build model config first, then init distributed
+        model_config = ModelConfig(
+            model_path=self.model_path,
+            trust_remote_code=self.trust_remote_code,
+            revision=self.revision,
+            dtype=self.dtype,
+            quantization=self.quantization,
+        )
+
+        # Build cache config fingerprint BEFORE loading the model.
+        # Loading may mutate hf_config.quantization_config (e.g. via
+        # process_weights_after_loading), which would produce a different
+        # hash than what the engine computes from the original config.
+        # ModelConfig always exposes hf_config/quantization directly;
+        # quantization_config is the only genuinely-optional attribute.
+        quant_config = getattr(model_config.hf_config, "quantization_config", None)
+        quant_method = get_quant_method_name(
+            self.quantization or model_config.quantization
+        )
+        if not quant_method and quant_config is not None:
+            quant_method = get_quant_method_name(quant_config)
+
+        self.config = CacheConfig(
+            model_path=self.model_path,
+            model_arch=(
+                model_config.hf_config.architectures[0]
+                if model_config.hf_config.architectures
+                else ""
+            ),
+            tp_size=self.tp_size,
+            tp_rank=self.tp_rank,
+            pp_size=self.pp_size,
+            pp_rank=self.pp_rank,
+            dp_size=self.dp_size,
+            ep_size=self.ep_size,
+            quant_method=quant_method,
+            quant_config_hash=hash_quant_config(quant_config),
+            dtype=str(model_config.dtype),
+            revision=self.revision or "",
+            **compute_env_stamp(),
+        )
+
+        # Refuse to serve quant methods not verified to round-trip through pure
+        # IPC tensor export. Checked before loading so an unsupported model
+        # fails fast instead of after minutes of disk I/O.
+        check_ipc_quant_support(quant_method, quant_config, where="daemon")
+
+        # Initialize distributed backend (requires server_args + model_config)
+        self._init_distributed(server_args, model_config)
+
+        # Build load config
+        load_config = LoadConfig(
+            load_format=self.load_format,
+            model_loader_extra_config=self.model_loader_extra_config,
+            tp_rank=self.tp_rank,
+        )
+
+        logger.info(
+            f"[WeightCacheDaemon gpu={self.gpu_id} tp_rank={self.tp_rank}] "
+            f"Loading model from disk: {self.model_path}"
+        )
+        tic = time.perf_counter()
+
+        # Load model using DefaultModelLoader (includes TP sharding + quant post-process)
+        loader = get_model_loader(load_config=load_config, model_config=model_config)
+        self.model = loader.load_model(
+            model_config=model_config,
+            device_config=DeviceConfig(current_platform.device_type, self.gpu_id),
+        )
+
+        elapsed = time.perf_counter() - tic
+        logger.info(
+            f"[WeightCacheDaemon gpu={self.gpu_id} tp_rank={self.tp_rank}] "
+            f"Model loaded from disk in {elapsed:.2f}s"
+        )
+
+        # Ensure every post-processing kernel has retired before we export the
+        # memory: clients map these tensors read-only via IPC and would otherwise
+        # risk observing half-written weights.
+        current_platform.synchronize()
+
+        # Export all parameters and buffers as IPC handles
+        self._export_state()
+
+        logger.info(
+            f"[WeightCacheDaemon gpu={self.gpu_id} tp_rank={self.tp_rank}] "
+            f"Exported {len(self.state_entries)} tensors as IPC handles. "
+            f"Ready to serve."
+        )
+
+    @staticmethod
+    def _assert_ipc_compatible_allocator() -> None:
+        """Reject allocator configs incompatible with CUDA IPC export.
+
+        The expandable-segments allocator returns memory that cannot be shared
+        through torch's _share_cuda_ handle, which would make the export fail
+        partway with an opaque error. Detect it up front and fail loud.
+        """
+        for var in ("PYTORCH_CUDA_ALLOC_CONF", "PYTORCH_ALLOC_CONF"):
+            conf = os.environ.get(var, "")
+            for field in conf.split(","):
+                key, _, value = field.partition(":")
+                if (
+                    key.strip() == "expandable_segments"
+                    and value.strip().lower() == "true"
+                ):
+                    raise RuntimeError(
+                        f"[WeightCacheDaemon] {var} sets expandable_segments:True, "
+                        f"which is incompatible with CUDA IPC weight sharing: the "
+                        f"expandable-segments allocator hands out memory that cannot "
+                        f"be exported via _share_cuda_, so the IPC handle export "
+                        f"would fail mid-way. Unset expandable_segments for the "
+                        f"weight cache daemon process (it can stay enabled for the "
+                        f"engine itself)."
+                    )
+
+    def _export_state(self):
+        """Export model parameters and buffers as CUDA IPC handles.
+
+        This includes both persistent buffers (in state_dict) and non-persistent
+        buffers (e.g. rotary embedding cos_sin_cache) so the engine can fully
+        reconstruct the model state via zero-copy IPC.
+        """
+        self.state_entries.clear()
+
+        # remove_duplicate=False so tied weights are recognized as parameters
+        # under every name. state_dict() below emits both tied keys, and with the
+        # deduped set the duplicate would be mis-registered as a buffer, not a
+        # parameter, on the client.
+        param_names = set(
+            name for name, _ in self.model.named_parameters(remove_duplicate=False)
+        )
+        state_dict_names = set(self.model.state_dict().keys())
+
+        # Export all items from state_dict (parameters + persistent buffers)
+        for name, tensor in self.model.state_dict().items():
+            ipc_handle = MultiprocessingSerializer.serialize(
+                tensor.data, output_str=True
+            )
+            self.state_entries[name] = {
+                "handle": ipc_handle,
+                "shape": list(tensor.shape),
+                "dtype": str(tensor.dtype).replace("torch.", ""),
+                "is_param": name in param_names,
+            }
+
+        # Also export non-persistent buffers (not in state_dict but needed
+        # for inference, e.g. rotary embedding cos_sin_cache)
+        non_persistent_count = 0
+        for name, buf in self.model.named_buffers():
+            if name not in state_dict_names:
+                ipc_handle = MultiprocessingSerializer.serialize(
+                    buf.data, output_str=True
+                )
+                self.state_entries[name] = {
+                    "handle": ipc_handle,
+                    "shape": list(buf.shape),
+                    "dtype": str(buf.dtype).replace("torch.", ""),
+                    "is_param": False,
+                }
+                non_persistent_count += 1
+
+        # Log total size
+        total_bytes = sum(
+            entry["handle"].__len__() if hasattr(entry["handle"], "__len__") else 0
+            for entry in self.state_entries.values()
+        )
+        logger.info(
+            f"[WeightCacheDaemon gpu={self.gpu_id}] "
+            f"Exported {len(self.state_entries)} tensors "
+            f"({non_persistent_count} non-persistent buffers), "
+            f"serialized handle size ~{total_bytes / 1024 / 1024:.1f} MB"
+        )
+
+    def serve(self):
+        """Block and serve IPC handles over Unix socket."""
+        # Do NOT unlink an existing socket here: stale-file cleanup is the launch
+        # path's job (cleanup_stale_daemon_files refuses to remove a socket whose
+        # .ready still points at a live PID). A leftover live socket makes bind()
+        # fail loudly instead of silently stealing another daemon's socket.
+        sock = socket.socket(socket.AF_UNIX, socket.SOCK_STREAM)
+        old_umask = os.umask(0o177)
+        try:
+            sock.bind(self.socket_path)
+        finally:
+            os.umask(old_umask)
+        sock.listen(8)
+        sock.settimeout(1.0)  # Allow periodic shutdown check
+
+        # Write ready file
+        with open(self.ready_path, "w") as f:
+            f.write(f"pid={os.getpid()}\n")
+            f.write(f"config={self.config.to_dict()}\n")
+
+        logger.info(
+            f"[WeightCacheDaemon gpu={self.gpu_id}] " f"Listening on {self.socket_path}"
+        )
+
+        self._running = True
+
+        def _signal_handler(signum, frame):
+            logger.info(
+                f"[WeightCacheDaemon gpu={self.gpu_id}] Received signal {signum}, shutting down"
+            )
+            self._running = False
+
+        signal.signal(signal.SIGTERM, _signal_handler)
+        signal.signal(signal.SIGINT, _signal_handler)
+
+        try:
+            while self._running:
+                try:
+                    conn, _ = sock.accept()
+                    # The listen-socket timeout above only bounds accept(); the
+                    # accepted connection is blocking by default. Since we serve
+                    # connections serially, a client that connects but never sends
+                    # (or dies mid-send) would block recv_msg forever and stall
+                    # every other engine rank. Bound each exchange instead.
+                    conn.settimeout(CLIENT_CONNECTION_TIMEOUT)
+                    try:
+                        self._handle_connection(conn)
+                    except Exception as e:
+                        logger.error(
+                            f"[WeightCacheDaemon gpu={self.gpu_id}] "
+                            f"Error handling connection: {e}",
+                            exc_info=True,
+                        )
+                    finally:
+                        conn.close()
+                except socket.timeout:
+                    continue
+        finally:
+            sock.close()
+            if os.path.exists(self.socket_path):
+                os.unlink(self.socket_path)
+            if os.path.exists(self.ready_path):
+                os.unlink(self.ready_path)
+            logger.info(f"[WeightCacheDaemon gpu={self.gpu_id}] Shutdown complete")
+
+    def _handle_connection(self, conn: socket.socket):
+        """Handle a single client connection."""
+        req = recv_msg(conn)
+
+        if req.get("type") == "query_config":
+            # Client asks for config without requesting handles
+            send_msg(conn, {"status": "ok", "config": self.config.to_dict()})
+
+        elif req.get("type") == "fetch_state":
+            # Client requests full state with IPC handles
+            engine_config = CacheConfig.from_dict(req["config"])
+            if not self.config.matches(engine_config):
+                # Log detailed mismatch info for debugging
+                daemon_dict = self.config.to_dict()
+                engine_dict = engine_config.to_dict()
+                mismatches = {
+                    k: (daemon_dict.get(k), engine_dict.get(k))
+                    for k in daemon_dict
+                    if daemon_dict.get(k) != engine_dict.get(k)
+                }
+                logger.warning(
+                    f"[WeightCacheDaemon gpu={self.gpu_id}] "
+                    f"Config mismatch: {mismatches}"
+                )
+                send_msg(
+                    conn, {"status": "mismatch", "daemon_config": self.config.to_dict()}
+                )
+                return
+
+            logger.info(
+                f"[WeightCacheDaemon gpu={self.gpu_id}] "
+                f"Serving {len(self.state_entries)} IPC handles to engine"
+            )
+            send_msg(
+                conn,
+                {
+                    "status": "ok",
+                    "config": self.config.to_dict(),
+                    "entries": self.state_entries,
+                    # PID so the client can watch daemon liveness: if this
+                    # process dies while clients hold IPC mappings, their
+                    # param.data (and any CUDA-graph-captured addresses) dangle.
+                    "pid": os.getpid(),
+                },
+            )
+
+        elif req.get("type") == "ping":
+            send_msg(conn, {"status": "ok"})
+
+        else:
+            send_msg(
+                conn,
+                {
+                    "status": "error",
+                    "message": f"Unknown request type: {req.get('type')}",
+                },
+            )
+
+    def shutdown(self):
+        """Release GPU memory and clean up."""
+        if dist.is_initialized():
+            dist.destroy_process_group()
+        if self.model is not None:
+            del self.model
+            self.model = None
+        self.state_entries.clear()
+        current_platform.empty_cache()
+        self._running = False
+
+
+def run_weight_cache_daemon(
+    model_path: str,
+    gpu_id: int,
+    tp_size: int = 1,
+    tp_rank: int = 0,
+    pp_size: int = 1,
+    pp_rank: int = 0,
+    dp_size: int = 1,
+    ep_size: int = 1,
+    load_format: str = "auto",
+    dtype: str = "auto",
+    quantization: Optional[str] = None,
+    model_loader_extra_config: str = "{}",
+    trust_remote_code: bool = False,
+    revision: Optional[str] = None,
+    dist_init_method: Optional[str] = None,
+):
+    """Entry point for running a weight cache daemon process."""
+    logging.basicConfig(
+        level=logging.INFO,
+        format=f"%(asctime)s [Daemon gpu={gpu_id} tp_rank={tp_rank}] %(levelname)s %(message)s",
+    )
+
+    # Die if our parent (the engine or the standalone launcher that spawned us)
+    # dies, even on SIGKILL/OOM-kill. Without this an orphaned daemon keeps a
+    # full weight copy pinned in GPU memory and its live-PID .ready file blocks
+    # the next launch — the opposite of fast recovery.
+    from sglang.srt.utils import kill_itself_when_parent_died
+
+    kill_itself_when_parent_died()
+
+    daemon = WeightCacheDaemon(
+        model_path=model_path,
+        gpu_id=gpu_id,
+        tp_size=tp_size,
+        tp_rank=tp_rank,
+        pp_size=pp_size,
+        pp_rank=pp_rank,
+        dp_size=dp_size,
+        ep_size=ep_size,
+        load_format=load_format,
+        dtype=dtype,
+        quantization=quantization,
+        model_loader_extra_config=model_loader_extra_config,
+        trust_remote_code=trust_remote_code,
+        revision=revision,
+        dist_init_method=dist_init_method,
+    )
+
+    daemon.load()
+    daemon.serve()
+
+
+def launch_weight_cache_daemons(
+    model_path: str,
+    tp_size: int = 1,
+    pp_size: int = 1,
+    dp_size: int = 1,
+    ep_size: int = 1,
+    nnodes: int = 1,
+    node_rank: int = 0,
+    base_gpu_id: int = 0,
+    gpu_id_step: int = 1,
+    load_format: str = "auto",
+    dtype: str = "auto",
+    quantization: Optional[str] = None,
+    model_loader_extra_config: str = "{}",
+    trust_remote_code: bool = False,
+    revision: Optional[str] = None,
+    dist_init_method: Optional[str] = None,
+    timeout: int = 1800,
+    force: bool = False,
+):
+    """Launch weight cache daemon processes for this node's PP×TP ranks.
+
+    For single-node (nnodes=1): spawns pp_size * tp_size daemons.
+    For multi-node (nnodes>1): spawns this node's share of PP×TP daemons,
+    mapping local gpu_id to the correct global (pp_rank, tp_rank).
+
+    Uses subprocess.Popen instead of multiprocessing.Process to avoid
+    initializing CUDA in the parent process, which can degrade CUDA IPC
+    performance in child processes.
+
+    Usage (single-node):
+        python -m sglang.srt.weight_cache.daemon \\
+            --model-path /path/to/model --tp-size 4
+
+    Usage (multi-node, run on each node):
+        # Node 0:
+        python -m sglang.srt.weight_cache.daemon \\
+            --model-path /path/to/model --tp-size 16 \\
+            --nnodes 2 --node-rank 0 \\
+            --dist-init-method tcp://node0-ip:29500
+
+        # Node 1:
+        python -m sglang.srt.weight_cache.daemon \\
+            --model-path /path/to/model --tp-size 16 \\
+            --nnodes 2 --node-rank 1 \\
+            --dist-init-method tcp://node0-ip:29500
+    """
+    import socket as sock_mod
+    import subprocess
+    import sys
+
+    # Replicate _calculate_rank_ranges logic from engine.py
+    pp_size_per_node = max(pp_size // nnodes, 1)
+    nnodes_per_pp_rank = max(nnodes // pp_size, 1)
+    pp_rank_range = range(
+        pp_size_per_node * (node_rank // nnodes_per_pp_rank),
+        pp_size_per_node * (node_rank // nnodes_per_pp_rank + 1),
+    )
+    nnodes_per_tp_group = nnodes_per_pp_rank
+    tp_size_per_node = tp_size // nnodes_per_tp_group
+    tp_rank_range = range(
+        tp_size_per_node * (node_rank % nnodes_per_tp_group),
+        tp_size_per_node * (node_rank % nnodes_per_tp_group + 1),
+    )
+
+    if nnodes > 1 and dist_init_method is None:
+        raise ValueError(
+            "dist_init_method is required for multi-node weight cache daemons. "
+            "Use --dist-init-method tcp://<node0-ip>:<port> to specify the "
+            "rendezvous address accessible from all nodes."
+        )
+
+    # Auto-allocate a free port for the distributed init method
+    if dist_init_method is None:
+        with sock_mod.socket(sock_mod.AF_INET, sock_mod.SOCK_STREAM) as s:
+            s.bind(("127.0.0.1", 0))
+            free_port = s.getsockname()[1]
+        dist_init_method = f"tcp://127.0.0.1:{free_port}"
+
+    python_path = sys.executable
+    daemon_module = "sglang.srt.weight_cache.daemon"
+
+    # Validate and clean up stale .ready/.sock files from prior runs.
+    for pp_rank in pp_rank_range:
+        for tp_rank in tp_rank_range:
+            global_rank = compute_global_rank(tp_size, pp_rank, tp_rank)
+            cleanup_stale_daemon_files(global_rank, force=force)
+
+    procs = []
+    for pp_rank in pp_rank_range:
+        for tp_rank in tp_rank_range:
+            gpu_id = compute_local_gpu_id(
+                pp_rank,
+                tp_rank,
+                pp_size_per_node,
+                tp_size_per_node,
+                base_gpu_id=base_gpu_id,
+                gpu_id_step=gpu_id_step,
+            )
+            cmd = [
+                python_path,
+                "-m",
+                daemon_module,
+                "--model-path",
+                model_path,
+                "--gpu-id",
+                str(gpu_id),
+                "--tp-size",
+                str(tp_size),
+                "--tp-rank",
+                str(tp_rank),
+                "--dp-size",
+                str(dp_size),
+                "--ep-size",
+                str(ep_size),
+                "--pp-size",
+                str(pp_size),
+                "--pp-rank",
+                str(pp_rank),
+                "--load-format",
+                load_format,
+                "--dtype",
+                dtype,
+                "--dist-init-method",
+                dist_init_method,
+            ]
+            if quantization:
+                cmd += ["--quantization", quantization]
+            if model_loader_extra_config and model_loader_extra_config != "{}":
+                cmd += ["--model-loader-extra-config", model_loader_extra_config]
+            if trust_remote_code:
+                cmd += ["--trust-remote-code"]
+            if revision:
+                cmd += ["--revision", revision]
+
+            proc = subprocess.Popen(cmd)
+            procs.append(proc)
+            logger.info(
+                f"Launched weight cache daemon gpu={gpu_id} "
+                f"pp_rank={pp_rank} tp_rank={tp_rank} pid={proc.pid}"
+            )
+
+    # Wait for all daemons on this node to become ready
+    num_daemons = len(procs)
+    check_interval = 2
+    start_time = time.time()
+    for pp_rank in pp_rank_range:
+        for tp_rank in tp_rank_range:
+            global_rank = compute_global_rank(tp_size, pp_rank, tp_rank)
+            ready_path = get_ready_path(global_rank)
+            while not os.path.exists(ready_path):
+                time.sleep(check_interval)
+                if time.time() - start_time > timeout:
+                    logger.error(
+                        f"Weight cache daemon pp_rank={pp_rank} tp_rank={tp_rank} "
+                        f"did not become ready within {timeout}s"
+                    )
+                    for p in procs:
+                        p.terminate()
+                    raise TimeoutError(
+                        f"Weight cache daemon pp_rank={pp_rank} tp_rank={tp_rank} "
+                        f"did not become ready within {timeout}s"
+                    )
+                # Check if any daemon exited prematurely
+                for p in procs:
+                    retcode = p.poll()
+                    if retcode is not None:
+                        logger.error(
+                            f"Weight cache daemon exited prematurely "
+                            f"with code {retcode}"
+                        )
+                        for other in procs:
+                            if other.poll() is None:
+                                other.terminate()
+                        raise RuntimeError(
+                            f"Weight cache daemon exited prematurely "
+                            f"with code {retcode}"
+                        )
+            logger.info(
+                f"Weight cache daemon pp_rank={pp_rank} tp_rank={tp_rank} is ready"
+            )
+
+    logger.info(
+        f"All {num_daemons} weight cache daemons on node {node_rank} are ready "
+        f"(pp_ranks={pp_rank_range.start}..{pp_rank_range.stop - 1}, "
+        f"tp_ranks={tp_rank_range.start}..{tp_rank_range.stop - 1}, "
+        f"dist_init_method={dist_init_method})"
+    )
+
+    # Monitor daemons — poll all of them and, the moment any one exits,
+    # terminate the rest and raise. A serial proc.wait() would not notice a
+    # mid-list death (e.g. procs[1] dying while procs[0] is still alive) until
+    # the earlier proc happened to exit, and it never surfaced the failure.
+    exited = None
+    try:
+        while exited is None:
+            for proc in procs:
+                if proc.poll() is not None:
+                    exited = proc
+                    break
+            else:
+                time.sleep(1)
+                continue
+    except KeyboardInterrupt:
+        logger.info("Received KeyboardInterrupt, shutting down daemons")
+    finally:
+        for proc in procs:
+            if proc.poll() is None:
+                proc.terminate()
+        for proc in procs:
+            try:
+                proc.wait(timeout=5)
+            except subprocess.TimeoutExpired:
+                proc.kill()
+        logger.info("All weight cache daemons have been terminated")
+
+    if exited is not None:
+        raise RuntimeError(
+            f"Weight cache daemon (pid={exited.pid}) exited with code "
+            f"{exited.returncode}; terminated the remaining daemons."
+        )
+
+
+if __name__ == "__main__":
+    import argparse
+
+    parser = argparse.ArgumentParser(description="SGLang Weight Cache Daemon")
+    parser.add_argument("--model-path", required=True, help="Path to model weights")
+    parser.add_argument("--tp-size", type=int, default=1, help="Tensor parallel size")
+    parser.add_argument(
+        "--gpu-id",
+        type=int,
+        default=None,
+        help="GPU device ID for a single daemon. "
+        "If omitted, launches daemons for all TP ranks (0..tp_size-1).",
+    )
+    parser.add_argument(
+        "--tp-rank",
+        type=int,
+        default=None,
+        help="TP rank for a single daemon. "
+        "If omitted, launches daemons for all TP ranks.",
+    )
+    parser.add_argument("--dp-size", type=int, default=1, help="Data parallel size")
+    parser.add_argument("--ep-size", type=int, default=1, help="Expert parallel size")
+    parser.add_argument("--pp-size", type=int, default=1, help="Pipeline parallel size")
+    parser.add_argument("--pp-rank", type=int, default=0, help="Pipeline parallel rank")
+    parser.add_argument("--nnodes", type=int, default=1, help="Total number of nodes")
+    parser.add_argument(
+        "--base-gpu-id",
+        type=int,
+        default=0,
+        help="GPU id of this node's first rank (mirrors the engine's "
+        "--base-gpu-id). Used to place daemons on the same GPUs the engine "
+        "ranks will use.",
+    )
+    parser.add_argument(
+        "--gpu-id-step",
+        type=int,
+        default=1,
+        help="Stride between consecutive ranks' GPU ids (mirrors the engine's "
+        "--gpu-id-step).",
+    )
+    parser.add_argument(
+        "--node-rank",
+        type=int,
+        default=0,
+        help="Rank of this node (0-indexed). Required for multi-node.",
+    )
+    parser.add_argument("--load-format", default="auto", help="Weight load format")
+    parser.add_argument("--dtype", default="auto", help="Model dtype")
+    parser.add_argument("--quantization", default=None, help="Quantization method")
+    parser.add_argument(
+        "--model-loader-extra-config",
+        default="{}",
+        help="Extra config for model loader (JSON string)",
+    )
+    parser.add_argument("--trust-remote-code", action="store_true")
+    parser.add_argument("--revision", default=None, help="Model revision")
+    parser.add_argument(
+        "--dist-init-method",
+        default=None,
+        help="Distributed init method (e.g. tcp://node0-ip:29500). "
+        "Auto-assigned for single-node when launching all ranks. "
+        "Required for multi-node (nnodes > 1) and must be accessible "
+        "from all nodes. Also required for tp_size > 1 when launching "
+        "a single rank.",
+    )
+    parser.add_argument(
+        "--timeout",
+        type=int,
+        default=1800,
+        help="Timeout in seconds to wait for all daemons to become ready (default: 1800)",
+    )
+    parser.add_argument(
+        "--force",
+        action="store_true",
+        help="Take over a rank whose .ready file still points at a live PID by "
+        "killing that daemon (use to reclaim a wedged/orphaned daemon).",
+    )
+
+    args = parser.parse_args()
+
+    if args.gpu_id is not None or args.tp_rank is not None:
+        # Single-rank mode: launch one daemon for the specified rank
+        gpu_id = args.gpu_id if args.gpu_id is not None else args.tp_rank
+        tp_rank = args.tp_rank if args.tp_rank is not None else args.gpu_id
+        # Refuse to clobber a live daemon already holding this rank (mirrors the
+        # multi-rank launcher and the engine path, which the multi-rank spawns
+        # of this same entrypoint rely on). --force kills and takes over.
+        cleanup_stale_daemon_files(
+            compute_global_rank(args.tp_size, args.pp_rank, tp_rank),
+            force=args.force,
+        )
+        run_weight_cache_daemon(
+            model_path=args.model_path,
+            gpu_id=gpu_id,
+            tp_size=args.tp_size,
+            tp_rank=tp_rank,
+            pp_size=args.pp_size,
+            pp_rank=args.pp_rank,
+            dp_size=args.dp_size,
+            ep_size=args.ep_size,
+            load_format=args.load_format,
+            dtype=args.dtype,
+            quantization=args.quantization,
+            model_loader_extra_config=args.model_loader_extra_config,
+            trust_remote_code=args.trust_remote_code,
+            revision=args.revision,
+            dist_init_method=args.dist_init_method,
+        )
+    else:
+        # Multi-rank mode: launch daemons for this node's TP ranks
+        launch_weight_cache_daemons(
+            model_path=args.model_path,
+            tp_size=args.tp_size,
+            pp_size=args.pp_size,
+            dp_size=args.dp_size,
+            ep_size=args.ep_size,
+            nnodes=args.nnodes,
+            node_rank=args.node_rank,
+            base_gpu_id=args.base_gpu_id,
+            gpu_id_step=args.gpu_id_step,
+            load_format=args.load_format,
+            dtype=args.dtype,
+            quantization=args.quantization,
+            model_loader_extra_config=args.model_loader_extra_config,
+            trust_remote_code=args.trust_remote_code,
+            revision=args.revision,
+            dist_init_method=args.dist_init_method,
+            timeout=args.timeout,
+            force=args.force,
+        )

python/sglang/srt/weight_cache/ipc_loader.py

@@ -0,0 +1,565 @@
+# SPDX-License-Identifier: Apache-2.0
+"""IPC Model Loader — loads model weights from a Weight Cache Daemon via CUDA IPC.
+
+Zero-copy mode: param.data points directly to IPC-mapped GPU memory. Only 1x GPU
+memory needed — engine and daemon share the same physical GPU memory via CUDA IPC.
+Engine depends on daemon staying alive.
+"""
+
+import logging
+import os
+import signal
+import stat
+import threading
+import time
+from typing import Optional
+
+import torch
+import torch.nn as nn
+
+from sglang.srt.configs.load_config import LoadConfig
+from sglang.srt.model_loader.loader import (
+    BaseModelLoader,
+    _initialize_model,
+)
+from sglang.srt.utils import MultiprocessingSerializer
+
+from .protocol import (
+    CacheConfig,
+    check_ipc_quant_support,
+    compute_env_stamp,
+    get_quant_method_name,
+    hash_quant_config,
+    recv_msg,
+    send_msg,
+)
+
+logger = logging.getLogger(__name__)
+
+# How often the client polls the serving daemon's PID for liveness.
+_DAEMON_LIVENESS_POLL_INTERVAL = 5.0
+
+
+class IpcModelLoader(BaseModelLoader):
+    """Load model weights from a Weight Cache Daemon via CUDA IPC handles.
+
+    In daemon mode (weight_cache_mode="daemon"), the engine and daemon share
+    the same GPU. Falling back to disk loading would cause OOM because both
+    processes would hold weights on the same GPU. Therefore, daemon mode
+    raises an error if the daemon is unavailable instead of falling back.
+
+    In client mode, disk fallback is allowed ONLY when the daemon is genuinely
+    absent (its Unix socket file does not exist). Every other failure is a hard
+    error rather than a silent fallback, so a broken IPC path never masquerades
+    as a healthy (but slow, disk-loaded) server:
+
+    - socket file missing            -> fall back to disk load
+    - connection refused             -> raise (daemon crashed after binding)
+    - CacheConfig mismatch           -> raise (do NOT disk-load on a shared GPU
+                                        holding a different config's weights;
+                                        also surfaces fingerprint drift bugs)
+    - any protocol / transfer error  -> raise
+
+    See _fetch_from_cache for the authoritative fallback-vs-raise contract.
+    """
+
+    def __init__(
+        self,
+        load_config: LoadConfig,
+        socket_path: str,
+        fallback_loader_cls=None,
+        weight_cache_mode: str = "client",
+        fallback_load_format: str = "auto",
+    ):
+        super().__init__(load_config)
+        self.socket_path = socket_path
+        self.weight_cache_mode = weight_cache_mode
+        self._fallback_loader_cls = fallback_loader_cls
+        self._fallback_load_format = fallback_load_format
+
+    def load_model(
+        self,
+        *,
+        model_config,
+        device_config,
+    ) -> nn.Module:
+        """Load model weights from the weight cache daemon.
+
+        In daemon mode, raises RuntimeError if the daemon is unavailable
+        (fallback to disk loading would cause OOM on shared GPUs).
+        In client mode, falls back to DefaultModelLoader.
+        """
+        tic = time.perf_counter()
+
+        # Hard-gate unsupported quant methods before touching the daemon, so an
+        # unsupported model fails explicitly instead of silently disk-loading
+        # (client mode) or serving wrong-numerics IPC weights. Checked here so
+        # it applies regardless of whether the daemon is reachable.
+        quant_method, engine_quant_config = self._resolve_engine_quant(model_config)
+        check_ipc_quant_support(quant_method, engine_quant_config, where="client")
+
+        # Try to fetch state from daemon
+        cache_data = self._fetch_from_cache(model_config)
+
+        if cache_data is None:
+            if self.weight_cache_mode == "daemon":
+                raise RuntimeError(
+                    f"[IpcModelLoader] Weight cache daemon not available at "
+                    f"{self.socket_path}. In daemon mode, fallback to disk "
+                    f"loading is disabled because the daemon process already "
+                    f"holds weights on the same GPU — loading from disk would "
+                    f"cause OOM. Please ensure the weight cache daemon is "
+                    f"running and the config matches."
+                )
+            logger.warning(
+                "[IpcModelLoader] Weight cache not available or config mismatch, "
+                "falling back to disk load"
+            )
+            return self._fallback_load(model_config, device_config)
+
+        entries = cache_data["entries"]
+        logger.info(
+            f"[IpcModelLoader] Fetched {len(entries)} IPC handles from daemon "
+            f"in {time.perf_counter() - tic:.2f}s"
+        )
+
+        from sglang.srt.model_loader.loader import (
+            _get_quantization_config,
+        )
+
+        quant_config = _get_quantization_config(model_config, self.load_config)
+
+        model = self._load_zero_copy_mode(
+            model_config,
+            device_config,
+            entries,
+            quant_config,
+        )
+
+        # Skip _post_load_weights: the daemon already ran
+        # process_weights_after_loading on the weights before exporting
+        # IPC handles. Running it again would double-process (e.g.,
+        # re-quantize already-quantized weights), corrupting tensor data.
+
+        # Rebuild stale tensor views. Some modules store tensor views as
+        # plain attributes (not parameters/buffers) during __init__. When
+        # the model is initialized on meta device and then weights are
+        # replaced via IPC mapping, these views still point to the old
+        # meta storage. We must recreate them from the now-valid tensors.
+        self._rebuild_stale_views(model)
+
+        # The model now points into the daemon's GPU memory via CUDA IPC. If the
+        # daemon dies, those pointers dangle, so watch it and fail loud.
+        self._start_daemon_liveness_watchdog(cache_data.get("pid"))
+
+        logger.info(
+            f"[IpcModelLoader] Loaded model via IPC (mode={self.weight_cache_mode}), "
+            f"total={time.perf_counter() - tic:.2f}s"
+        )
+
+        return model.eval()
+
+    def _start_daemon_liveness_watchdog(self, daemon_pid: Optional[int]) -> None:
+        """Fail loud if the serving daemon dies while we hold its weights.
+
+        In both client and (engine-spawned) daemon mode, the model's param.data
+        points into the daemon's GPU memory via CUDA IPC, and CUDA graphs may
+        capture those addresses. If the daemon exits, the pointers dangle:
+        forward passes would read freed GPU memory -> illegal-address crashes or
+        silent garbage. There is no safe in-place recovery, so a background
+        thread polls the daemon PID and, on death, SIGKILLs this process with a
+        clear message instead of letting it serve corrupt results.
+        """
+        if not daemon_pid or daemon_pid <= 0:
+            logger.warning(
+                "[IpcModelLoader] Daemon did not report a PID; skipping the "
+                "daemon-liveness watchdog. A daemon crash will not be detected."
+            )
+            return
+
+        def _daemon_alive(pid: int) -> bool:
+            try:
+                os.kill(pid, 0)
+            except ProcessLookupError:
+                return False
+            except PermissionError:
+                return True  # exists but owned by another user
+            return True
+
+        def _watch() -> None:
+            while True:
+                time.sleep(_DAEMON_LIVENESS_POLL_INTERVAL)
+                if not _daemon_alive(daemon_pid):
+                    logger.critical(
+                        f"[IpcModelLoader] Weight cache daemon (pid={daemon_pid}) "
+                        f"died while this engine holds its weights via CUDA IPC. "
+                        f"The mapped weight pointers are now dangling; continuing "
+                        f"would read freed GPU memory. Terminating this process."
+                    )
+                    os.kill(os.getpid(), signal.SIGKILL)
+                    return
+
+        threading.Thread(
+            target=_watch, name="weight-cache-daemon-watchdog", daemon=True
+        ).start()
+        logger.info(
+            f"[IpcModelLoader] Started daemon-liveness watchdog for pid={daemon_pid}"
+        )
+
+    def _resolve_engine_quant(self, model_config):
+        """Return (quant_method, quant_config) matching the daemon's fingerprint.
+
+        Shared by the IPC allowlist gate and the CacheConfig fingerprint so the
+        two can never drift apart. ModelConfig always exposes
+        hf_config/quantization directly; quantization_config is the only
+        genuinely-optional attribute.
+        """
+        quant_config = getattr(model_config.hf_config, "quantization_config", None)
+        quant_method = get_quant_method_name(model_config.quantization)
+        if not quant_method and quant_config is not None:
+            quant_method = get_quant_method_name(quant_config)
+        return quant_method, quant_config
+
+    @staticmethod
+    def _rebuild_stale_views(model):
+        """Rebuild tensor views that went stale after IPC weight replacement.
+
+        RadixLinearAttention.conv_weights is a view of conv1d.weight created
+        during __init__. After IPC mapping replaces conv1d.weight with a new
+        tensor, the old view still points to meta-device storage. Recreate
+        it from the now-valid parameter.
+        """
+        try:
+            from sglang.srt.layers.radix_linear_attention import RadixLinearAttention
+        except ImportError:
+            return
+
+        count = 0
+        for _, module in model.named_modules():
+            conv1d = getattr(module, "conv1d", None)
+            attn = getattr(module, "attn", None)
+            if conv1d is not None and isinstance(attn, RadixLinearAttention):
+                if hasattr(conv1d, "weight") and conv1d.weight is not None:
+                    attn.conv_weights = conv1d.weight.view(
+                        conv1d.weight.size(0), conv1d.weight.size(2)
+                    )
+                    if hasattr(conv1d, "bias") and conv1d.bias is not None:
+                        attn.bias = conv1d.bias
+                    count += 1
+
+        if count > 0:
+            logger.info(f"[IpcModelLoader] Rebuilt {count} stale conv_weights views")
+
+    @staticmethod
+    def _set_module_tensor(model, name, tensor, is_param=True):
+        """Replace or register a parameter/buffer in the model by its full dotted name.
+
+        This is necessary because setting param.data on a meta-device tensor
+        raises a type mismatch error (meta and CUDA tensors have incompatible
+        dispatch keys). Instead, we walk the module tree and use setattr to
+        replace the entire parameter/buffer object.
+
+        If the attribute already exists as a parameter/buffer, it is replaced.
+        If it doesn't exist (e.g. post-quantization params like weight_scale),
+        it is registered as a new parameter or buffer.
+        """
+        parts = name.split(".")
+        obj = model
+        for part in parts[:-1]:
+            obj = getattr(obj, part)
+        leaf_name = parts[-1]
+        if is_param:
+            # requires_grad=False: the IPC memory is shared/read-only and SGLang
+            # is inference-only, so autograd must never write into it.
+            new_param = nn.Parameter(tensor, requires_grad=False)
+            setattr(obj, leaf_name, new_param)
+        else:
+            # register_buffer raises KeyError if the name already exists as a
+            # parameter or plain attribute (not a buffer). This happens when
+            # process_weights_after_loading converts a parameter to a buffer
+            # (e.g. Mamba's A_log). Remove the old attribute first.
+            if leaf_name in obj._parameters:
+                del obj._parameters[leaf_name]
+            elif hasattr(obj, leaf_name) and leaf_name not in obj._buffers:
+                delattr(obj, leaf_name)
+            obj.register_buffer(leaf_name, tensor)
+
+    def _load_zero_copy_mode(
+        self,
+        model_config,
+        device_config,
+        entries,
+        quant_config,
+    ) -> nn.Module:
+        """Zero-copy load: map IPC tensors directly as param.data.
+
+        The model is initialized on the meta device (no memory allocation),
+        then each parameter's data is replaced with the IPC-mapped GPU tensor.
+        The engine and daemon share the same physical GPU memory via CUDA IPC.
+        """
+        from sglang.srt.model_loader.utils import set_default_torch_dtype
+
+        # Initialize model on meta device to avoid any GPU/CPU memory allocation.
+        # This creates the model structure with the correct parameter shapes/dtypes
+        # but without allocating actual storage.
+        with set_default_torch_dtype(model_config.dtype):
+            with torch.device("meta"):
+                model = _initialize_model(
+                    model_config,
+                    self.load_config,
+                    quant_config,
+                )
+
+        # Build lookup dicts of existing parameter/buffer names in the
+        # meta-device model. Post-quantization parameters (e.g. weight_scale
+        # from FP8) are created by process_weights_after_loading, which the
+        # daemon already ran. These params exist in the daemon's entries but
+        # NOT in the meta-device model — we must register them as new attrs.
+        # Use dicts (not sets) so we can do O(1) shape/dtype validation
+        # without re-traversing the model tree on every lookup.
+        # remove_duplicate=False mirrors the daemon's export (which keys tied
+        # weights under every name) so a tied parameter is recognized under all
+        # of its names here too.
+        existing_params = {
+            name: param
+            for name, param in model.named_parameters(remove_duplicate=False)
+        }
+        existing_buffers = {name: buf for name, buf in model.named_buffers()}
+        existing_names = set(existing_params) | set(existing_buffers)
+
+        imported_refs = []
+        imported_count = 0
+        mismatched = []
+        new_params_count = 0
+        map_tic = time.perf_counter()
+
+        # Iterate over ALL daemon entries (not just model params/buffers).
+        # This ensures post-quantization parameters (weight_scale, etc.)
+        # that were created by process_weights_after_loading are also mapped.
+        for name, entry in entries.items():
+            imported_tensor = MultiprocessingSerializer.deserialize(entry["handle"])
+            is_param = entry.get("is_param", True)
+
+            if name in existing_names:
+                # Existing parameter/buffer — validate shape/dtype
+                if name in existing_params:
+                    ref_param = existing_params[name]
+                else:
+                    ref_param = existing_buffers[name]
+                if (
+                    imported_tensor.shape != ref_param.shape
+                    or imported_tensor.dtype != ref_param.dtype
+                ):
+                    mismatched.append(
+                        f"  {name}: IPC={imported_tensor.shape}/{imported_tensor.dtype} "
+                        f"vs model={ref_param.shape}/{ref_param.dtype}"
+                    )
+                    del imported_tensor
+                    continue
+
+            # Replace or register the tensor in the model
+            self._set_module_tensor(model, name, imported_tensor, is_param=is_param)
+            imported_refs.append(imported_tensor)
+            imported_count += 1
+
+            if name not in existing_names:
+                new_params_count += 1
+
+        if mismatched:
+            raise RuntimeError(
+                f"[IpcModelLoader] {len(mismatched)} tensor(s) have shape/dtype "
+                f"mismatch between the IPC daemon and the meta-initialized model. "
+                f"The quantization method passed the IPC allowlist gate "
+                f"(check_ipc_quant_support), so this is NOT an unsupported-quant "
+                f"case — it indicates the daemon's weight fingerprint is "
+                f"incomplete or the daemon/client configs drifted (a bug to fix), "
+                f"not merely uninitialized weights:\n" + "\n".join(mismatched)
+            )
+
+        # After mapping every daemon entry, any tensor still on the meta device
+        # is one the daemon did NOT provide. Filling it with torch.empty() would
+        # hand the model uninitialized GPU memory — silently producing wrong
+        # output, the worst failure mode for a load path. Hard-error and list the
+        # offenders instead.
+        #
+        # The daemon exports the full state_dict AND non-persistent buffers
+        # (e.g. rotary embedding cos_sin_cache), so a correct setup leaves nothing
+        # on meta here. A non-empty list means the daemon's export is incomplete,
+        # or the model has a genuinely-recomputable buffer that must be recomputed
+        # explicitly (not filled with garbage) — add that handling here if needed.
+        still_on_meta_params = [
+            name
+            for name, param in model.named_parameters()
+            if param.device.type == "meta"
+        ]
+        still_on_meta_buffers = [
+            name for name, buf in model.named_buffers() if buf.device.type == "meta"
+        ]
+
+        if still_on_meta_params or still_on_meta_buffers:
+            raise RuntimeError(
+                f"[IpcModelLoader] After IPC mapping, "
+                f"{len(still_on_meta_params)} parameter(s) and "
+                f"{len(still_on_meta_buffers)} buffer(s) remain on the meta device "
+                f"— the daemon did not export them. Refusing to fill them with "
+                f"uninitialized memory, which would silently produce wrong output. "
+                f"This means the daemon's export is incomplete, or a recomputable "
+                f"buffer needs explicit recompute logic here.\n"
+                f"  params: {still_on_meta_params[:10]}"
+                f"{'...' if len(still_on_meta_params) > 10 else ''}\n"
+                f"  buffers: {still_on_meta_buffers[:10]}"
+                f"{'...' if len(still_on_meta_buffers) > 10 else ''}"
+            )
+
+        map_elapsed = time.perf_counter() - map_tic
+
+        # Stash IPC refs on the model to prevent GC (which would unmap the memory)
+        if imported_refs:
+            model._ipc_imported_tensors = imported_refs
+
+        logger.info(
+            f"[IpcModelLoader] Zero-copy: mapped {imported_count} tensors "
+            f"({new_params_count} new post-quant), time={map_elapsed:.3f}s"
+        )
+
+        return model
+
+    def _fetch_from_cache(self, model_config) -> Optional[dict]:
+        """Connect to daemon, validate config, fetch IPC handles.
+
+        Returns the daemon response dict on success, None if the daemon is
+        genuinely absent (socket file doesn't exist). Raises on all other
+        failures so they are never silently swallowed as a disk-load fallback.
+        """
+        import socket as socket_mod
+
+        # Only connect to a real socket node owned by us: reject a symlink, a
+        # plain file, or another user's socket planted at this /tmp path. An
+        # absent socket means no daemon -> fall back to disk (return None).
+        try:
+            st = os.lstat(self.socket_path)
+        except FileNotFoundError:
+            logger.info(
+                f"[IpcModelLoader] Daemon socket not found at {self.socket_path}."
+            )
+            return None
+        if not stat.S_ISSOCK(st.st_mode) or st.st_uid != os.getuid():
+            raise RuntimeError(
+                f"[IpcModelLoader] Refusing to connect: {self.socket_path} is not "
+                f"a socket owned by this user."
+            )
+
+        sock = socket_mod.socket(socket_mod.AF_UNIX, socket_mod.SOCK_STREAM)
+        try:
+            sock.settimeout(30)
+            sock.connect(self.socket_path)
+        except FileNotFoundError:
+            # Raced: socket removed between lstat and connect -> treat as absent.
+            sock.close()
+            return None
+        except ConnectionRefusedError:
+            sock.close()
+            raise RuntimeError(
+                f"[IpcModelLoader] Daemon socket exists at {self.socket_path} but "
+                f"refused the connection. The daemon may have crashed after "
+                f"creating the socket. Check daemon logs."
+            )
+        except Exception as e:
+            sock.close()
+            raise RuntimeError(
+                f"[IpcModelLoader] Failed to connect to daemon at "
+                f"{self.socket_path}: {e}"
+            ) from e
+
+        try:
+            # Build engine's config fingerprint
+            from sglang.srt.runtime_context import get_parallel
+
+            ps = get_parallel()
+            tp_size = ps.tp_size
+            tp_rank = ps.tp_rank
+
+            pp_size = ps.pp_size
+            pp_rank = ps.pp_rank
+
+            ep_size = ps.moe_ep_size
+
+            from sglang.srt.runtime_context import get_server_args
+
+            dp_size = get_server_args().dp_size
+
+            quant_method, quant_config = self._resolve_engine_quant(model_config)
+
+            engine_config = CacheConfig(
+                model_path=model_config.model_path,
+                model_arch=(
+                    model_config.hf_config.architectures[0]
+                    if model_config.hf_config.architectures
+                    else ""
+                ),
+                tp_size=tp_size,
+                tp_rank=tp_rank,
+                pp_size=pp_size,
+                pp_rank=pp_rank,
+                dp_size=dp_size,
+                ep_size=ep_size,
+                quant_method=quant_method,
+                quant_config_hash=hash_quant_config(quant_config),
+                dtype=str(model_config.dtype),
+                revision=model_config.revision or "",
+                **compute_env_stamp(),
+            )
+
+            logger.info(
+                f"[IpcModelLoader] Requesting weights from daemon at "
+                f"{self.socket_path} with config: "
+                f"model={engine_config.model_path}, "
+                f"arch={engine_config.model_arch}, "
+                f"tp={engine_config.tp_size}/{engine_config.tp_rank}, "
+                f"quant={engine_config.quant_method}, "
+                f"dtype={engine_config.dtype}"
+            )
+
+            send_msg(sock, {"type": "fetch_state", "config": engine_config.to_dict()})
+            result = recv_msg(sock)
+
+            if result.get("status") != "ok":
+                daemon_config = result.get("daemon_config", {})
+                raise RuntimeError(
+                    f"[IpcModelLoader] Daemon config mismatch!\n"
+                    f"  Engine config: {engine_config.to_dict()}\n"
+                    f"  Daemon config: {daemon_config}"
+                )
+
+            return result
+
+        except RuntimeError:
+            raise
+        except Exception as e:
+            raise RuntimeError(
+                f"[IpcModelLoader] Error communicating with daemon at "
+                f"{self.socket_path}: {e}"
+            ) from e
+        finally:
+            sock.close()
+
+    def _fallback_load(self, model_config, device_config) -> nn.Module:
+        """Fall back to DefaultModelLoader for disk-based loading."""
+        from sglang.srt.configs.load_config import LoadConfig
+        from sglang.srt.model_loader.loader import DefaultModelLoader
+
+        fallback_config = LoadConfig(
+            load_format=self._fallback_load_format,
+            download_dir=self.load_config.download_dir,
+            model_loader_extra_config=self.load_config.model_loader_extra_config,
+            tp_rank=self.load_config.tp_rank,
+        )
+        loader_cls = self._fallback_loader_cls or DefaultModelLoader
+        fallback = loader_cls(fallback_config)
+        return fallback.load_model(
+            model_config=model_config, device_config=device_config
+        )
+
+    def download_model(self, model_config) -> None:
+        """No-op: daemon handles its own model downloading."""
+        pass

python/sglang/srt/weight_cache/protocol.py

@@ -0,0 +1,380 @@
+# SPDX-License-Identifier: Apache-2.0
+"""Protocol definitions for the weight cache daemon.
+
+Defines CacheConfig for validation and socket message protocol helpers.
+"""
+
+import hashlib
+import json
+import logging
+import os
+import pickle
+import signal
+import struct
+from typing import Any, Dict, Optional
+
+import msgspec
+
+from sglang.srt.utils.common import safe_pickle_loads
+
+logger = logging.getLogger(__name__)
+
+# Socket path template for weight cache daemons (keyed by global rank
+# = tp_size * pp_rank + tp_rank, so multi-node / multi-PP don't collide)
+WEIGHT_CACHE_SOCKET_TEMPLATE = "/tmp/sglang_weight_cache_rank{global_rank}.sock"
+
+# Ready file template — daemon writes this after loading completes
+WEIGHT_CACHE_READY_TEMPLATE = "/tmp/sglang_weight_cache_rank{global_rank}.ready"
+
+
+class CacheConfig(msgspec.Struct):
+    """Fingerprint of the cached weights. Used to validate compatibility
+    between a daemon's cached state and a requesting engine process.
+
+    Any mismatch triggers a fallback to disk loading.
+    """
+
+    model_path: str
+    model_arch: str
+    tp_size: int
+    tp_rank: int
+    pp_size: int
+    pp_rank: int
+    dp_size: int
+    ep_size: int
+    quant_method: str  # e.g. "fp8", "gptq_marlin", "" for unquantized
+    quant_config_hash: str  # SHA-256 hash of quantization config
+    dtype: str  # e.g. "torch.float16"
+    revision: str  # model revision the weights were loaded from ("" if unset)
+    # Environment stamp: a daemon and a client that ran different post-processing
+    # branches (different GPU compute capability or torch/kernel version) can
+    # produce incompatible weights that would map cleanly yet serve garbage.
+    # Comparing these turns that into a clean mismatch. See compute_env_stamp().
+    device_capability: str  # local compute capability, e.g. "8.0" ("" if N/A)
+    torch_version: str  # torch.__version__ of the process that built the weights
+
+    def matches(self, other: "CacheConfig") -> bool:
+        """Check if two configs are compatible for weight sharing."""
+        return self == other
+
+    def to_dict(self) -> Dict[str, Any]:
+        return {f: getattr(self, f) for f in self.__struct_fields__}
+
+    @classmethod
+    def from_dict(cls, d: Dict[str, Any]) -> "CacheConfig":
+        return cls(**d)
+
+
+def hash_quant_config(quant_config: Any) -> str:
+    """Compute a stable hash of the quantization config.
+
+    Avoids str()/repr() on arbitrary objects because those embed memory
+    addresses (e.g. "at 0x7f..."), producing different hashes across
+    processes and causing permanent config mismatch.
+    """
+    if quant_config is None:
+        return ""
+    try:
+        if hasattr(quant_config, "to_dict"):
+            config_str = json.dumps(quant_config.to_dict(), sort_keys=True)
+        elif isinstance(quant_config, dict):
+            config_str = json.dumps(quant_config, sort_keys=True)
+        elif hasattr(quant_config, "__dict__"):
+            config_str = (
+                type(quant_config).__name__
+                + ":"
+                + json.dumps(
+                    {
+                        k: v
+                        for k, v in sorted(quant_config.__dict__.items())
+                        if not k.startswith("_")
+                        and isinstance(
+                            v, (str, int, float, bool, type(None), list, dict)
+                        )
+                    },
+                    sort_keys=True,
+                )
+            )
+        else:
+            config_str = type(quant_config).__name__
+        return hashlib.sha256(config_str.encode()).hexdigest()
+    except Exception:
+        config_str = type(quant_config).__name__
+        return hashlib.sha256(config_str.encode()).hexdigest()
+
+
+def get_quant_method_name(quant_config: Any) -> str:
+    """Extract the quantization method name from config."""
+    if quant_config is None:
+        return ""
+    if isinstance(quant_config, str):
+        return quant_config
+    if hasattr(quant_config, "get_name"):
+        return quant_config.get_name()
+    if hasattr(quant_config, "name"):
+        return quant_config.name
+    return type(quant_config).__name__
+
+
+# ---------------------------------------------------------------------------
+# IPC quantization-method allowlist
+# ---------------------------------------------------------------------------
+#
+# CUDA IPC zero-copy sharing exports ONLY raw tensor data, so it is correct only
+# when process_weights_after_loading's entire effect is captured by that data.
+# Methods that stamp Python-side metadata (e.g. block-FP8's format_ue8m0) or
+# repack/transpose weights into shapes the meta-init client can't reproduce
+# (per-tensor FP8, Marlin, AWQ/GPTQ) would serve silently-wrong numerics. Only
+# methods verified to round-trip through pure tensor export are allowed; every
+# other method hard-errors. Extend the registry below only after verifying a
+# method end-to-end.
+
+
+class UnsupportedQuantForIPCError(RuntimeError):
+    """Raised when a quantization method is not on the verified allowlist for
+    CUDA IPC zero-copy weight sharing."""
+
+
+def _get_quant_field(quant_config: Any, key: str) -> Any:
+    """Read a field from a quant config that may be a dict or an object."""
+    if quant_config is None:
+        return None
+    if isinstance(quant_config, dict):
+        return quant_config.get(key)
+    return getattr(quant_config, key, None)
+
+
+def _fp8_round_trips_via_ipc(quant_config: Any) -> bool:
+    """Only block-wise FP8 is verified.
+
+    Block-wise FP8 (weight_block_size set) preserves weight shape and the only
+    post-load metadata it stamps is accounted for. Per-tensor FP8 transposes
+    `layer.weight` during post-processing, a shape change the meta-init client
+    cannot reproduce, so it is not supported.
+    """
+    return _get_quant_field(quant_config, "weight_block_size") is not None
+
+
+# quant_method name -> predicate(quant_config) -> bool (True == verified safe).
+# A method absent from this registry is unsupported and hard-errors.
+IPC_QUANT_ALLOWLIST = {
+    "": lambda _quant_config: True,  # unquantized
+    "fp8": _fp8_round_trips_via_ipc,  # only block-wise FP8 verified
+}
+
+
+def is_ipc_quant_supported(quant_method: str, quant_config: Any) -> bool:
+    """Return True if `quant_method` is verified safe for IPC zero-copy sharing."""
+    predicate = IPC_QUANT_ALLOWLIST.get(quant_method)
+    if predicate is None:
+        return False
+    return bool(predicate(quant_config))
+
+
+def check_ipc_quant_support(
+    quant_method: str, quant_config: Any, *, where: str
+) -> None:
+    """Hard-error unless `quant_method` is verified safe for IPC zero-copy sharing.
+
+    `where` is a short tag (e.g. "daemon"/"client") used only in the error
+    message. Raises UnsupportedQuantForIPCError with an actionable message.
+    """
+    if is_ipc_quant_supported(quant_method, quant_config):
+        return
+    verified = ", ".join(
+        (repr(m) if m else "'' (unquantized)") for m in IPC_QUANT_ALLOWLIST
+    )
+    raise UnsupportedQuantForIPCError(
+        f"[weight_cache:{where}] quantization method {quant_method!r} is not "
+        f"verified for CUDA IPC zero-copy weight sharing. Its "
+        f"process_weights_after_loading may stamp Python-side metadata "
+        f"(e.g. format_ue8m0) or repack/transpose weights into shapes the "
+        f"meta-initialized client cannot reproduce, which would silently serve "
+        f"wrong-numerics weights. Verified methods: {verified}. Note: FP8 is "
+        f"only verified for block-wise configs (weight_block_size set), not "
+        f"per-tensor FP8. Disable the weight cache (--weight-cache-mode off) "
+        f"for this model."
+    )
+
+
+# ---------------------------------------------------------------------------
+# Socket protocol helpers
+# ---------------------------------------------------------------------------
+
+
+MAX_MSG_SIZE = 256 * 1024 * 1024  # 256 MiB
+
+
+def send_msg(sock, obj: Any) -> None:
+    """Send a length-prefixed pickled message over a socket."""
+    data = pickle.dumps(obj, protocol=pickle.HIGHEST_PROTOCOL)
+    header = struct.pack("!I", len(data))
+    sock.sendall(header + data)
+
+
+def recv_msg(sock) -> Any:
+    """Receive a length-prefixed pickled message from a socket."""
+    header = _recv_exact(sock, 4)
+    if header is None:
+        raise ConnectionError("Connection closed while reading message header")
+    length = struct.unpack("!I", header)[0]
+    if length > MAX_MSG_SIZE:
+        raise ValueError(f"Message size {length} exceeds {MAX_MSG_SIZE} byte cap")
+    data = _recv_exact(sock, length)
+    if data is None:
+        raise ConnectionError("Connection closed while reading message body")
+    return safe_pickle_loads(data)
+
+
+def _recv_exact(sock, n: int) -> Optional[bytes]:
+    """Receive exactly n bytes from a socket."""
+    buf = bytearray()
+    while len(buf) < n:
+        chunk = sock.recv(n - len(buf))
+        if not chunk:
+            return None
+        buf.extend(chunk)
+    return bytes(buf)
+
+
+def compute_env_stamp() -> Dict[str, str]:
+    """Local environment fingerprint for the IPC weight cache.
+
+    Returns the device compute capability and torch version of the current
+    process. A daemon and a connecting client that differ on either may have run
+    different post-processing / kernel-selection branches, producing weights that
+    map cleanly over IPC yet serve garbage; stamping these into CacheConfig turns
+    that into a clean mismatch. Imported lazily so protocol.py stays cheap to
+    import and usable on CPU-only hosts (both fields degrade to "").
+    """
+    device_capability = ""
+    torch_version = ""
+    try:
+        import torch
+
+        torch_version = str(torch.__version__)
+    except Exception:
+        pass
+    try:
+        from sglang.srt.platforms import current_platform
+
+        cap = current_platform.get_device_capability()
+        if cap is not None:
+            device_capability = f"{cap.major}.{cap.minor}"
+    except Exception:
+        pass
+    return {"device_capability": device_capability, "torch_version": torch_version}
+
+
+def compute_global_rank(tp_size: int, pp_rank: int, tp_rank: int) -> int:
+    """Single source of truth for the daemon rank formula.
+
+    global_rank = tp_size * pp_rank + tp_rank, so each daemon gets a unique
+    socket/ready path even across PP stages and nodes. Every call site (engine,
+    loader, model_runner, daemon) must go through this so the copies can't drift.
+    """
+    return tp_size * pp_rank + tp_rank
+
+
+def compute_local_gpu_id(
+    pp_rank: int,
+    tp_rank: int,
+    pp_size_per_node: int,
+    tp_size_per_node: int,
+    base_gpu_id: int = 0,
+    gpu_id_step: int = 1,
+) -> int:
+    """Single source of truth for the local GPU id a daemon rank runs on.
+
+    Mirrors the engine's device assignment so a daemon and the engine rank it
+    serves always land on the same physical GPU (a prerequisite for CUDA IPC).
+    ``base_gpu_id``/``gpu_id_step`` default to the identity mapping used by the
+    standalone launcher; the engine passes its real ``--base-gpu-id`` /
+    ``--gpu-id-step`` so every call site computes the id the same way instead of
+    keeping three drifting copies of the formula.
+    """
+    return (
+        base_gpu_id
+        + (pp_rank % pp_size_per_node) * tp_size_per_node
+        + (tp_rank % tp_size_per_node) * gpu_id_step
+    )
+
+
+def get_socket_path(global_rank: int) -> str:
+    """Get the Unix socket path for a weight cache daemon.
+
+    global_rank = tp_size * pp_rank + tp_rank
+    """
+    return WEIGHT_CACHE_SOCKET_TEMPLATE.format(global_rank=global_rank)
+
+
+def get_ready_path(global_rank: int) -> str:
+    """Get the ready-file path for a weight cache daemon.
+
+    global_rank = tp_size * pp_rank + tp_rank
+    """
+    return WEIGHT_CACHE_READY_TEMPLATE.format(global_rank=global_rank)
+
+
+def _read_ready_pid(ready_path: str) -> Optional[int]:
+    """Read the daemon PID from a .ready file. Returns None if unreadable."""
+    try:
+        with open(ready_path) as f:
+            for line in f:
+                if line.startswith("pid="):
+                    return int(line.strip().split("=", 1)[1])
+    except (OSError, ValueError):
+        pass
+    return None
+
+
+def _is_pid_alive(pid: int) -> bool:
+    """Check whether a process is still running."""
+    try:
+        os.kill(pid, 0)
+        return True
+    except ProcessLookupError:
+        return False
+    except PermissionError:
+        return True
+
+
+def cleanup_stale_daemon_files(global_rank: int, *, force: bool = False) -> None:
+    """Validate and clean up .ready/.sock files for a daemon rank.
+
+    If the .ready file exists and the recorded PID is still alive, the daemon
+    is still running — raise RuntimeError so the caller doesn't clobber it,
+    unless ``force`` is set, in which case the running daemon is killed and its
+    files are taken over (stale-takeover path for a wedged/orphaned daemon).
+    If the PID is dead (or unreadable), the files are stale leftovers from a
+    crashed/killed daemon and are safe to remove.
+    """
+    ready_path = get_ready_path(global_rank)
+    socket_path = get_socket_path(global_rank)
+
+    if not os.path.exists(ready_path) and not os.path.exists(socket_path):
+        return
+
+    pid = _read_ready_pid(ready_path) if os.path.exists(ready_path) else None
+
+    if pid is not None and _is_pid_alive(pid):
+        if not force:
+            raise RuntimeError(
+                f"Weight cache daemon for rank {global_rank} is already running "
+                f"(pid={pid}, ready={ready_path}). Stop the existing daemon before "
+                f"launching a new one, or pass force=True (--force) to kill it and "
+                f"take over."
+            )
+        logger.warning(
+            f"[weight_cache] force takeover: killing existing daemon pid={pid} "
+            f"for rank {global_rank} and reclaiming its socket/ready files."
+        )
+        try:
+            os.kill(pid, signal.SIGKILL)
+        except ProcessLookupError:
+            pass
+
+    for path in (ready_path, socket_path):
+        if os.path.exists(path):
+            os.unlink(path)
+            logger.info(f"Removed stale daemon file: {path}")