RPC is broken due to change of interface in llama.cpp main repository (rpc : early register backend devices #11262)

[j-lag]

RPC does not work in llama-cpp-python since this commit in llama.cpp

Revision: 667d72846c06b2cf4f7c8a4265e210991a49706b
Author: Radoslav Gerganov <rgerganov@gmail.com>
Date: 17/01/2025 09:57:09
Message:
rpc : early register backend devices (#11262)
Early register RPC devices and do not propagate RPC specifics in the
llama model structures.

This commit change the way rpc_servers are passed to llama.cpp, (in llama.cpp\common\arg.cpp) it is not a parameter anymore (params.rpc_servers = value;) but need to call add_rpc_devices(value);

To reproduce:

from llama_cpp import Llama
llm = Llama(
    model_path=MODEL,
	rpc_servers=RPC_NODES,
	n_ctx = 4096,
	flash_attn = False,
	n_gpu_layers = 99,
    verbose=True
)

--> should init with RPC offloading

[Hofi]

here is a quick fix i created with chatgpt to get rpc working:

diff --git a/llama_cpp/llama.py b/llama_cpp/llama.py
index 7e9a6af..5274e98 100644
--- a/llama_cpp/llama.py
+++ b/llama_cpp/llama.py
@@ -51,6 +51,45 @@ import llama_cpp._internals as internals
 from ._logger import set_verbose
 from ._utils import suppress_stdout_stderr
 
+_rpc_devices_registered = False
+from llama_cpp.llama_cpp import _lib  # now _lib holds the shared library handle
+
+def add_rpc_devices(servers: str) -> None:
+    """
+    Register RPC devices with the llama backend using the provided comma‐separated string.
+    """
+    global _rpc_devices_registered
+    if _rpc_devices_registered:
+        return  # Already registered, so do nothing
+    lib = _lib  # use the imported shared library handle
+    # Bind ggml_backend_reg_by_name
+    lib.ggml_backend_reg_by_name.argtypes = [ctypes.c_char_p]
+    lib.ggml_backend_reg_by_name.restype = ctypes.c_void_p
+    rpc_reg = lib.ggml_backend_reg_by_name(b"RPC")
+    if not rpc_reg:
+        raise ValueError("failed to find RPC backend")
+    # Bind ggml_backend_reg_get_proc_address
+    lib.ggml_backend_reg_get_proc_address.argtypes = [ctypes.c_void_p, ctypes.c_char_p]
+    lib.ggml_backend_reg_get_proc_address.restype = ctypes.c_void_p
+    rpc_add_fn_ptr = lib.ggml_backend_reg_get_proc_address(rpc_reg, b"ggml_backend_rpc_add_device")
+    if not rpc_add_fn_ptr:
+        raise ValueError("failed to find RPC device add function")
+    # Create a callable from the function pointer: returns a void pointer given a char*
+    PROTOTYPE = ctypes.CFUNCTYPE(ctypes.c_void_p, ctypes.c_char_p)
+    rpc_add_fn = PROTOTYPE(rpc_add_fn_ptr)
+    # Bind ggml_backend_device_register
+    lib.ggml_backend_device_register.argtypes = [ctypes.c_void_p]
+    lib.ggml_backend_device_register.restype = None
+    # For each server in the comma-separated list, register the device.
+    for server in servers.split(','):
+        server = server.strip().encode("utf-8")
+        dev = rpc_add_fn(server)
+        if dev:
+            lib.ggml_backend_device_register(dev)
+        else:
+            raise ValueError(f"failed to register RPC device for server: {server.decode('utf-8')}")
+    _rpc_devices_registered = True
+
 
 class Llama:
     """High-level Python wrapper for a llama.cpp model."""
@@ -227,7 +266,7 @@ class Llama:
         self.model_params.split_mode = split_mode
         self.model_params.main_gpu = main_gpu
         if rpc_servers is not None:
-            self.model_params.rpc_servers = rpc_servers.encode("utf-8")
+            add_rpc_devices(rpc_servers)
             self._rpc_servers = rpc_servers
         else:
             self._rpc_servers = None

[LeaveNhA]

What are we waiting for to fix it? Is there any road-blocker?

[szwedek]

Does anyone can merge it? I think this is an important feature that doesn't work.