Support learning rate schedulers in ExpertBackend (#196)

* Add empty __init__ to hivemind_cli for correct package discovery

* Support learning rate schedulers in ExpertBackend

* Save/load full expert state

* Don't pass compression to make_empty

* spawn -> fork

* Remove load_expert_states

* Make TaskPoolBase an abstract class

* Output warning if some of the keys in state_dict are missing

Co-authored-by: justheuristic <[email protected]>

.circleci/config.yml

@@ -8,11 +8,11 @@ jobs:
       - checkout
       - restore_cache:
           keys:
-            - v1-{{ checksum "requirements.txt" }}-{{ checksum "requirements-dev.txt" }}
+            - py37-v1-{{ checksum "requirements.txt" }}-{{ checksum "requirements-dev.txt" }}
       - run: pip install -r requirements.txt
       - run: pip install -r requirements-dev.txt
       - save_cache:
-          key: v1-{{ checksum "requirements.txt" }}-{{ checksum "requirements-dev.txt" }}
+          key: py37-v1-{{ checksum "requirements.txt" }}-{{ checksum "requirements-dev.txt" }}
           paths:
             - '~/.cache/pip'
       - run:
@@ -28,11 +28,11 @@ jobs:
       - checkout
       - restore_cache:
           keys:
-            - v1-{{ checksum "requirements.txt" }}-{{ checksum "requirements-dev.txt" }}
+            - py38-v1-{{ checksum "requirements.txt" }}-{{ checksum "requirements-dev.txt" }}
       - run: pip install -r requirements.txt
       - run: pip install -r requirements-dev.txt
       - save_cache:
-          key: v1-{{ checksum "requirements.txt" }}-{{ checksum "requirements-dev.txt" }}
+          key: py38-v1-{{ checksum "requirements.txt" }}-{{ checksum "requirements-dev.txt" }}
           paths:
             - '~/.cache/pip'
       - run:
@@ -48,11 +48,11 @@ jobs:
       - checkout
       - restore_cache:
           keys:
-            - v1-{{ checksum "requirements.txt" }}-{{ checksum "requirements-dev.txt" }}
+            - py39-v1-{{ checksum "requirements.txt" }}-{{ checksum "requirements-dev.txt" }}
       - run: pip install -r requirements.txt
       - run: pip install -r requirements-dev.txt
       - save_cache:
-          key: v1-{{ checksum "requirements.txt" }}-{{ checksum "requirements-dev.txt" }}
+          key: py39-v1-{{ checksum "requirements.txt" }}-{{ checksum "requirements-dev.txt" }}
           paths:
             - '~/.cache/pip'
       - run:

hivemind/hivemind_cli/run_server.py

@@ -8,6 +8,7 @@
 from hivemind.server import Server
 from hivemind.utils.threading import increase_file_limit
 from hivemind.utils.logging import get_logger
+from hivemind.server.layers import schedule_name_to_scheduler
 
 logger = get_logger(__name__)
 
@@ -28,13 +29,20 @@ def main():
     parser.add_argument('--expert_cls', type=str, default='ffn', required=False,
                         help="expert type from test_utils.layers, e.g. 'ffn', 'transformer', 'det_dropout' or 'nop'.")
     parser.add_argument('--hidden_dim', type=int, default=1024, required=False, help='main dimension for expert_cls')
+
     parser.add_argument('--num_handlers', type=int, default=None, required=False,
                         help='server will use this many processes to handle incoming requests')
     parser.add_argument('--max_batch_size', type=int, default=16384, required=False,
                         help='The total number of examples in the same batch will not exceed this value')
     parser.add_argument('--device', type=str, default=None, required=False,
                         help='all experts will use this device in torch notation; default: cuda if available else cpu')
+
     parser.add_argument('--optimizer', type=str, default='adam', required=False, help='adam, sgd or none')
+    parser.add_argument('--scheduler', type=str, choices=schedule_name_to_scheduler.keys(), default='none',
+                        help='LR scheduler type to use')
+    parser.add_argument('--num-warmup-steps', type=int, required=False, help='the number of warmup steps for LR schedule')
+    parser.add_argument('--num-training-steps', type=int, required=False, help='the total number of steps for LR schedule')
+
     parser.add_argument('--no_dht', action='store_true', help='if specified, the server will not be attached to a dht')
     parser.add_argument('--initial_peers', type=str, nargs='*', required=False, default=[],
                         help='one or more peers that can welcome you to the dht, e.g. 1.2.3.4:1337 192.132.231.4:4321')
@@ -45,7 +53,6 @@ def main():
     parser.add_argument('--compression', type=str, default='NONE', required=False, help='Tensor compression '
                         'parameter for grpc. Can be NONE, MEANSTD or FLOAT16')
     parser.add_argument('--checkpoint_dir', type=Path, required=False, help='Directory to store expert checkpoints')
-    parser.add_argument('--load_experts', action='store_true', help='Load experts from the checkpoint directory')
 
     # fmt:on
     args = vars(parser.parse_args())

hivemind/server/__init__.py

@@ -2,23 +2,22 @@
 
 import multiprocessing as mp
 import multiprocessing.synchronize
-import random
 import threading
 from contextlib import contextmanager
 from functools import partial
-from typing import Dict, Optional, Tuple, List
+from typing import Dict, Optional, Tuple
 from pathlib import Path
 
 import torch
 
 import hivemind
 from hivemind.dht import DHT
-from hivemind.server.expert_uid import UID_DELIMITER
-from hivemind.server.checkpoints import CheckpointSaver, load_weights, dir_is_correct
+from hivemind.server.expert_uid import UID_DELIMITER, generate_uids_from_pattern
+from hivemind.server.checkpoints import CheckpointSaver, load_experts, is_directory
 from hivemind.server.connection_handler import ConnectionHandler
 from hivemind.server.dht_handler import DHTHandlerThread, declare_experts, get_experts
 from hivemind.server.expert_backend import ExpertBackend
-from hivemind.server.layers import name_to_block, name_to_input
+from hivemind.server.layers import name_to_block, name_to_input, schedule_name_to_scheduler
 from hivemind.server.runtime import Runtime
 from hivemind.server.task_pool import Task, TaskPool, TaskPoolBase
 from hivemind.utils import Endpoint, get_port, replace_port, find_open_port, get_logger
@@ -68,11 +67,12 @@ def __init__(
         if start:
             self.run_in_background(await_ready=True)
 
-    @staticmethod
-    def create(listen_on='0.0.0.0:*', num_experts: int = None, expert_uids: str = None, expert_pattern: str = None,
-               expert_cls='ffn', hidden_dim=1024, optim_cls=torch.optim.Adam, num_handlers=None, max_batch_size=4096,
-               device=None, no_dht=False, initial_peers=(), dht_port=None, checkpoint_dir: Optional[Path] = None,
-               load_experts=False, compression=CompressionType.NONE, *, start: bool, **kwargs) -> Server:
+    @classmethod
+    def create(cls, listen_on='0.0.0.0:*', num_experts: int = None, expert_uids: str = None, expert_pattern: str = None,
+               expert_cls='ffn', hidden_dim=1024, optim_cls=torch.optim.Adam, scheduler: str = 'none',
+               num_warmup_steps=None, num_training_steps=None, num_handlers=None, max_batch_size=4096, device=None,
+               no_dht=False, initial_peers=(), dht_port=None, checkpoint_dir: Optional[Path] = None,
+               compression=CompressionType.NONE, *, start: bool, **kwargs) -> Server:
         """
         Instantiate a server with several identical experts. See argparse comments below for details
         :param listen_on: network interface with address and (optional) port, e.g. "127.0.0.1:1337" or "[::]:80"
@@ -85,16 +85,20 @@ def create(listen_on='0.0.0.0:*', num_experts: int = None, expert_uids: str = No
         :param num_handlers: server will use this many parallel processes to handle incoming requests
         :param max_batch_size: total num examples in the same batch will not exceed this value
         :param device: all experts will use this device in torch notation; default: cuda if available else cpu
+
         :param optim_cls: uses this optimizer to train all experts
+        :param scheduler: if not `none`, the name of the expert LR scheduler
+        :param num_warmup_steps: the number of warmup steps for LR schedule
+        :param num_training_steps: the total number of steps for LR schedule
+
         :param no_dht: if specified, the server will not be attached to a dht
         :param initial_peers: a list of peers that will introduce this node to the dht,\
            e.g. ('123.11.22.33:1337', '[fe80::abe2:db1c:be7d:5a85]:4567'), default = no peers
 
         :param dht_port:  DHT node will listen on this port, default = find open port
            You can then use this node as initial peer for subsequent servers.
 
-        :param checkpoint_dir: directory to save expert checkpoints
-        :param load_experts: whether to load expert checkpoints from checkpoint_dir
+        :param checkpoint_dir: directory to save and load expert checkpoints
 
         :param compression: if specified, use this compression to pack all inputs, outputs and gradients by all experts
             hosted on this server. For a more fine-grained compression, start server in python and specify compression
@@ -113,23 +117,29 @@ def create(listen_on='0.0.0.0:*', num_experts: int = None, expert_uids: str = No
             dht = hivemind.DHT(initial_peers=initial_peers, start=True, listen_on=dht_endpoint)
             logger.info(f"Running DHT node on port {dht.port}, initial peers = {initial_peers}")
 
-        if load_experts:
-            assert dir_is_correct(checkpoint_dir)
-            assert expert_uids is None, "Can't both load saved experts and create new ones from given UIDs"
-            expert_uids = [child.name for child in checkpoint_dir.iterdir() if (child / 'checkpoint_last.pt').exists()]
-            if expert_uids:
-                logger.info(f"Located checkpoints for experts {expert_uids}, ignoring UID generation options")
-            else:
-                logger.info(f"No expert checkpoints found in {checkpoint_dir}, generating...")
-
-        assert (expert_pattern is None and num_experts is None) or (expert_uids is None) or (num_experts == 0), \
-            "Please provide either expert_uids *or* num_experts and expert_pattern, but not both"
+        assert ((expert_pattern is None and num_experts is None and expert_uids is not None) or
+                (num_experts is not None and expert_uids is None)), \
+            "Please provide either expert_uids *or* num_experts (possibly with expert_pattern), but not both"
 
-        # get expert uids if not loaded previously
         if expert_uids is None:
-            assert num_experts is not None, "Please specify either expert_uids or num_experts [and expert_pattern]"
-            logger.info(f"Generating expert uids from pattern {expert_pattern}")
-            expert_uids = generate_uids_from_pattern(num_experts, expert_pattern, dht=dht)
+            if checkpoint_dir is not None:
+                assert is_directory(checkpoint_dir)
+                expert_uids = [child.name for child in checkpoint_dir.iterdir() if
+                               (child / 'checkpoint_last.pt').exists()]
+                total_experts_in_checkpoint = len(expert_uids)
+                logger.info(f"Located {total_experts_in_checkpoint} checkpoints for experts {expert_uids}")
+
+                if total_experts_in_checkpoint > num_experts:
+                    raise ValueError(
+                        f"Found {total_experts_in_checkpoint} checkpoints, but num_experts is set to {num_experts}, "
+                        f"which is smaller. Either increase num_experts or remove unneeded checkpoints.")
+            else:
+                expert_uids = []
+
+            uids_to_generate = num_experts - len(expert_uids)
+            if uids_to_generate > 0:
+                logger.info(f"Generating {uids_to_generate} expert uids from pattern {expert_pattern}")
+                expert_uids.extend(generate_uids_from_pattern(uids_to_generate, expert_pattern, dht))
 
         num_experts = len(expert_uids)
         num_handlers = num_handlers if num_handlers is not None else num_experts * 8
@@ -142,6 +152,8 @@ def create(listen_on='0.0.0.0:*', num_experts: int = None, expert_uids: str = No
         else:
             args_schema = (hivemind.BatchTensorDescriptor.from_tensor(sample_input, compression),)
 
+        scheduler = schedule_name_to_scheduler[scheduler]
+
         # initialize experts
         experts = {}
         for expert_uid in expert_uids:
@@ -150,15 +162,17 @@ def create(listen_on='0.0.0.0:*', num_experts: int = None, expert_uids: str = No
                                                          args_schema=args_schema,
                                                          outputs_schema=hivemind.BatchTensorDescriptor(
                                                              hidden_dim, compression=compression),
-                                                         opt=optim_cls(expert.parameters()),
+                                                         optimizer=optim_cls(expert.parameters()),
+                                                         scheduler=scheduler,
+                                                         num_warmup_steps=num_warmup_steps,
+                                                         num_training_steps=num_training_steps,
                                                          max_batch_size=max_batch_size)
 
-        if load_experts:
-            load_weights(experts, checkpoint_dir)
+        if checkpoint_dir is not None:
+            load_experts(experts, checkpoint_dir)
 
-        server = Server(dht, experts, listen_on=listen_on, num_connection_handlers=num_handlers, device=device,
-                        start=start)
-        return server
+        return cls(dht, experts, listen_on=listen_on, num_connection_handlers=num_handlers, device=device,
+                   checkpoint_dir=checkpoint_dir, start=start)
 
     def run(self):
         """
@@ -241,7 +255,7 @@ def shutdown(self):
 def background_server(*args, shutdown_timeout=5, **kwargs) -> Tuple[hivemind.Endpoint, hivemind.Endpoint]:
     """ A context manager that creates server in a background thread, awaits .ready on entry and shutdowns on exit """
     pipe, runners_pipe = mp.Pipe(duplex=True)
-    runner = mp.get_context("spawn").Process(target=_server_runner, args=(runners_pipe, *args), kwargs=kwargs)
+    runner = mp.Process(target=_server_runner, args=(runners_pipe, *args), kwargs=kwargs)
 
     try:
         runner.start()
@@ -269,63 +283,3 @@ def _server_runner(pipe, *args, **kwargs):
         server.shutdown()
         server.join()
         logger.info("Server shut down.")
-
-
-def generate_uids_from_pattern(num_experts: int, expert_pattern: Optional[str], dht: Optional[DHT] = None,
-                               attempts_per_expert=10) -> List[str]:
-    """
-    Sample experts from a given pattern, remove duplicates.
-    :param num_experts: sample this many unique expert uids
-    :param expert_pattern: a string pattern or a list of expert uids,  example: myprefix.[0:32].[0:256]\
-     means "sample random experts between myprefix.0.0 and myprefix.255.255;
-    :param dht: if specified, uses this DHT to check that expert uids are not yet occupied by other peers
-    :param attempts_per_expert: give up if unable to generate a new expert uid after this many attempts per uid
-    :note: this method is not strictly process-safe. If several servers run it concurrently, they have
-     a small chance of sampling duplicate expert uids.
-    """
-    remaining_attempts = attempts_per_expert * num_experts
-    found_uids, attempted_uids = list(), set()
-
-    def _generate_uid():
-        if expert_pattern is None:
-            return f"expert{UID_DELIMITER}{attempts_per_expert * num_experts - remaining_attempts}"
-
-        uid = []
-        for block in expert_pattern.split(UID_DELIMITER):
-            try:
-                if '[' not in block and ']' not in block:
-                    uid.append(block)
-                elif block.startswith('[') and block.endswith(']') and ':' in block:
-                    slice_start, slice_end = map(int, block[1:-1].split(':'))
-                    uid.append(str(random.randint(slice_start, slice_end - 1)))
-                else:
-                    raise ValueError("Block must be either fixed or a range [from:to]")
-            except KeyboardInterrupt as e:
-                raise e
-            except Exception as e:
-                raise ValueError(f"Expert pattern {expert_pattern} has invalid block {block}, {e}")
-        return UID_DELIMITER.join(uid)
-
-    while remaining_attempts > 0 and len(found_uids) < num_experts:
-
-        # 1. sample new expert uids at random
-        new_uids = []
-        while len(new_uids) + len(found_uids) < num_experts and remaining_attempts > 0:
-            new_uid = _generate_uid()
-            remaining_attempts -= 1
-            if new_uid not in attempted_uids:
-                attempted_uids.add(new_uid)
-                new_uids.append(new_uid)
-
-        # 2. look into DHT (if given) and remove duplicates
-        if dht:
-            existing_expert_uids = {found_expert.uid for found_expert in dht.get_experts(new_uids)
-                                    if found_expert is not None}
-            new_uids = [new_uid for new_uid in new_uids if new_uid not in existing_expert_uids]
-
-        found_uids += new_uids
-
-    if len(found_uids) != num_experts:
-        logger.warning(f"Found only {len(found_uids)} out of {num_experts} free expert uids after "
-                       f"{attempts_per_expert * num_experts} attempts")
-    return found_uids

hivemind/server/checkpoints.py

@@ -1,17 +1,20 @@
+import os
 import threading
 from datetime import datetime
 from pathlib import Path
 from shutil import copy2
 from tempfile import TemporaryDirectory
 from typing import Dict
-import os
 
 import torch
 
 from hivemind.server.expert_backend import ExpertBackend
+from hivemind.utils.logging import get_logger
 
+logger = get_logger(__name__)
 
-def dir_is_correct(directory: Path):
+
+def is_directory(directory: Path):
     assert directory is not None
     assert directory.exists()
     assert directory.is_dir()
@@ -33,7 +36,7 @@ def copy_tree(src: str, dst: str):
 class CheckpointSaver(threading.Thread):
     def __init__(self, expert_backends: Dict[str, ExpertBackend], checkpoint_dir: Path, update_period: int):
         super().__init__()
-        assert dir_is_correct(checkpoint_dir)
+        assert is_directory(checkpoint_dir)
         self.expert_backends = expert_backends
         self.update_period = update_period
         self.checkpoint_dir = checkpoint_dir
@@ -48,21 +51,25 @@ def run(self) -> None:
 
 
 def store_experts(experts: Dict[str, ExpertBackend], checkpoint_dir: Path):
-    assert dir_is_correct(checkpoint_dir)
+    logger.debug(f'Storing experts at {checkpoint_dir.absolute()}')
+    assert is_directory(checkpoint_dir)
     timestamp = datetime.now().isoformat(sep='_')
     with TemporaryDirectory() as tmpdirname:
         for expert_name, expert_backend in experts.items():
             expert_dir = Path(tmpdirname) / expert_name
             expert_dir.mkdir()
             checkpoint_name = expert_dir / f'checkpoint_{timestamp}.pt'
-            torch.save(expert_backend.state_dict(), checkpoint_name)
+            torch.save(expert_backend.get_full_state(), checkpoint_name)
             os.symlink(checkpoint_name, expert_dir / 'checkpoint_last.pt')
         copy_tree(tmpdirname, str(checkpoint_dir))
 
 
-def load_weights(experts: Dict[str, ExpertBackend], checkpoint_dir: Path):
-    assert dir_is_correct(checkpoint_dir)
+def load_experts(experts: Dict[str, ExpertBackend], checkpoint_dir: Path):
+    assert is_directory(checkpoint_dir)
     for expert_name, expert in experts.items():
         checkpoints_folder = checkpoint_dir / expert_name
         latest_checkpoint = checkpoints_folder / 'checkpoint_last.pt'
-        expert.load_state_dict(torch.load(latest_checkpoint))
+        if latest_checkpoint.exists():
+            expert.load_full_state(torch.load(latest_checkpoint))
+        else:
+            logger.warning(f'Failed to load checkpoint for expert {expert_name}')

hivemind/server/connection_handler.py

@@ -16,7 +16,7 @@
 logger = get_logger(__name__)
 
 
-class ConnectionHandler(mp.Process):
+class ConnectionHandler(mp.context.ForkProcess):
     """
     A process that accepts incoming requests to experts and submits them into the corresponding TaskPool.