tvm.rpc

Lightweight TVM RPC module.

RPC enables connect to a remote server, upload and launch functions. This is useful to for cross-compile and remote testing, The compiler stack runs on local server, while we use RPC server to run on remote runtime which don’t have a compiler available.

The test program compiles the program on local server, upload and run remote RPC server, get the result back to verify correctness.

class tvm.rpc.LocalSession

RPCSession interface backed by local environment.

This class can be used to implement functions that need to be ran both locally and remotely.

class tvm.rpc.PopenSession(binary)

RPCSession interface backed by popen.

Parameters:

binary (List[Union[str, bytes]]) – The binary to be executed.

class tvm.rpc.RPCSession(sess)

RPC Client session module

Do not directly create the object, call connect

cl(dev_id=0)

Construct OpenCL device.

cpu(dev_id=0)

Construct CPU device.

cuda(dev_id=0)

Construct CUDA GPU device.

device(dev_type, dev_id=0)

Construct a remote device.

Parameters:
  • dev_type (int or str) –

  • dev_id (int, optional) –

Returns:

dev – The corresponding encoded remote device.

Return type:

Device

download(path)

Download file from remote temp folder.

Parameters:

path (str) – The relative location to remote temp folder.

Returns:

blob – The result blob from the file.

Return type:

bytearray

download_linked_module(path)

Link a module in the remote and download it.

Parameters:

path (str) – The relative location to remote temp folder.

Returns:

blob – The result blob from the file.

Return type:

bytearray

Note

This function can be helpful when a linker is not available on the local client.

Examples

mod = build_module_with_cross_compilation()
# export the module as tar because a local linker is not available
mod.export_library("lib.tar")
remote.upload("lib.tar")
# invoke the linker on the remote to link the module as a library
# note that the library can only run on the same env as the remote
with open("lib.so", "wb") as file:
    file.write(remote.download_linked_module("lib.tar"))
ext_dev(dev_id=0)

Construct extension device.

get_function(name)

Get function from the session.

Parameters:

name (str) – The name of the function

Returns:

f – The result function.

Return type:

Function

hexagon(dev_id=0)

Construct Hexagon device.

listdir(path)

ls files from remote temp folder.

Parameters:

path (str) – The relative location to remote temp folder.

Returns:

dirs – The files in the given directory with split token ‘,’.

Return type:

str

load_module(path)

Load a remote module, the file need to be uploaded first.

Parameters:

path (str) – The relative location to remote temp folder.

Returns:

m – The remote module containing remote function.

Return type:

Module

metal(dev_id=0)

Construct Metal device.

remove(path)

Remove file from remote temp folder.

Parameters:

path (str) – The relative location to remote temp folder.

rocm(dev_id=0)

Construct ROCm device.

system_lib()

Get system-wide library module.

Returns:

module – The system-wide library module.

Return type:

runtime.Module

upload(data, target=None)

Upload file to remote runtime temp folder

Parameters:
  • data (str or bytearray) – The file name or binary in local to upload.

  • target (str, optional) – The path in remote

vulkan(dev_id=0)

Construct Vulkan device.

webgpu(dev_id=0)

Construct WebGPU device.

class tvm.rpc.Server(host='0.0.0.0', port=9091, port_end=9199, is_proxy=False, tracker_addr=None, key='', load_library=None, custom_addr=None, silent=False, no_fork=False, server_init_callback=None, reuse_addr=True, timeout=None)

Start RPC server on a separate process.

This is a simple python implementation based on multi-processing. It is also possible to implement a similar C based server with TVM runtime which does not depend on the python.

Parameters:
  • host (str) – The host url of the server.

  • port (int) – The port to be bind to

  • port_end (int, optional) – The end port to search

  • is_proxy (bool, optional) – Whether the address specified is a proxy. If this is true, the host and port actually corresponds to the address of the proxy server.

  • tracker_addr (Tuple (str, int) , optional) – The address of RPC Tracker in tuple(host, ip) format. If is not None, the server will register itself to the tracker.

  • key (str, optional) – The key used to identify the device type in tracker.

  • load_library (str, optional) – List of additional libraries to be loaded during execution.

  • custom_addr (str, optional) – Custom IP Address to Report to RPC Tracker

  • silent (bool, optional) – Whether run this server in silent mode.

  • no_fork (bool, optional) – Whether forbid fork in multiprocessing.

  • server_init_callback (Callable, optional) – Additional initialization function when starting the server.

  • reuse_addr (bool, optional) – Allows the kernel to reuse a local socket in TIME_WAIT state.

  • timeout (float, optional) – set a timeout for all operations on the socket

Note

The RPC server only sees functions in the tvm namespace. To bring additional custom functions to the server env, you can use server_init_callback.

def server_init_callback():
    import tvm
    # must import mypackage here
    import mypackage

    tvm.register_func("function", mypackage.func)

server = rpc.Server(host, server_init_callback=server_init_callback)
terminate()

Terminate the server process

class tvm.rpc.TrackerSession(addr)

Tracker client session.

Parameters:

addr (tuple) – The address tuple

close()

Close the tracker connection.

request(key, priority=1, session_timeout=0, max_retry=5, session_constructor_args=None)

Request a new connection from the tracker.

Parameters:
  • key (str) – The type key of the device.

  • priority (int, optional) – The priority of the request.

  • session_timeout (float, optional) – The duration of the session, allows server to kill the connection when duration is longer than this value. When duration is zero, it means the request must always be kept alive.

  • max_retry (int, optional) – Maximum number of times to retry before give up.

  • session_constructor_args (list, optional) – List of additional arguments to passed as the remote session constructor. The first element of the list is always a string specifying the name of the session constructor, the following args are the positional args to that function.

request_and_run(key, func, priority=1, session_timeout=0, max_retry=2)

Request a resource from tracker and run the func.

This function safe-guard rare server node dropout during execution. In such case, a new resource will be requested and func will be ran again.

Parameters:
  • key (str) – The type key of the device.

  • func (function of session -> value) – A stateless function

  • priority (int, optional) – The priority of the request.

  • session_timeout (float, optional) – The duration of the session, allows server to kill the connection when duration is longer than this value. When duration is zero, it means the request must always be kept alive.

  • max_retry (int, optional) – Maximum number of times to retry the function before give up.

summary()

Get the summary dict of the tracker.

text_summary()

Get a text summary of the tracker.

tvm.rpc.connect(url, port, key='', session_timeout=0, session_constructor_args=None, enable_logging=False)

Connect to RPC Server

Parameters:
  • url (str) – The url of the host

  • port (int) – The port to connect to

  • key (str, optional) – Additional key to match server

  • session_timeout (float, optional) – The duration of the session in seconds, allows server to kill the connection when duration is longer than this value. When duration is zero, it means the request must always be kept alive.

  • session_constructor_args (List) – List of additional arguments to passed as the remote session constructor. The first element of the list is always a string specifying the name of the session constructor, the following args are the positional args to that function.

  • enable_logging (boolean) – flag to enable/disable logging. Logging is disabled by default.

Returns:

sess – The connected session.

Return type:

RPCSession

Examples

Normal usage .. code-block:: python

client = rpc.connect(server_url, server_port, server_key)

Session_constructor can be used to customize the session in the remote The following code connects to a remote internal server via a proxy by constructing another RPCClientSession on the proxy machine and use that as the serving session of the proxy endpoint.

client_via_proxy = rpc.connect(
    proxy_server_url, proxy_server_port, proxy_server_key, enable_logging
    session_constructor_args=[
        "rpc.Connect", internal_url, internal_port, internal_key, internal_logging])
tvm.rpc.connect_tracker(url, port)

Connect to a RPC tracker

Parameters:
  • url (str) – The url of the host

  • port (int) – The port to connect to

Returns:

sess – The connected tracker session.

Return type:

TrackerSession

tvm.rpc.with_minrpc(compile_func, server='posix_popen_server', runtime='libtvm')

Attach the compiler function with minrpc related options.

Parameters:
  • compile_func (Union[str, Callable[[str, str, Optional[str]], None]]) – The compilation function to decorate.

  • server (str) – The server type.

  • runtime (str) – The runtime library.

Returns:

fcompile – The return compilation.

Return type:

function