file_path
stringlengths
3
280
file_language
stringclasses
66 values
content
stringlengths
1
1.04M
repo_name
stringlengths
5
92
repo_stars
int64
0
154k
repo_description
stringlengths
0
402
repo_primary_language
stringclasses
108 values
developer_username
stringlengths
1
25
developer_name
stringlengths
0
30
developer_company
stringlengths
0
82
python/ray/tune/suggest/nevergrad.py
Python
import logging import pickle try: import nevergrad as ng except ImportError: ng = None from ray.tune.suggest.suggestion import SuggestionAlgorithm logger = logging.getLogger(__name__) class NevergradSearch(SuggestionAlgorithm): """A wrapper around Nevergrad to provide trial suggestions. Requires Nevergrad to be installed. Nevergrad is an open source tool from Facebook for derivative free optimization of parameters and/or hyperparameters. It features a wide range of optimizers in a standard ask and tell interface. More information can be found at https://github.com/facebookresearch/nevergrad. Parameters: optimizer (nevergrad.optimization.Optimizer): Optimizer provided from Nevergrad. parameter_names (list): List of parameter names. Should match the dimension of the optimizer output. Alternatively, set to None if the optimizer is already instrumented with kwargs (see nevergrad v0.2.0+). max_concurrent (int): Number of maximum concurrent trials. Defaults to 10. metric (str): The training result objective value attribute. mode (str): One of {min, max}. Determines whether objective is minimizing or maximizing the metric attribute. use_early_stopped_trials (bool): Whether to use early terminated trial results in the optimization process. Example: >>> from nevergrad.optimization import optimizerlib >>> instrumentation = 1 >>> optimizer = optimizerlib.OnePlusOne(instrumentation, budget=100) >>> algo = NevergradSearch(optimizer, ["lr"], max_concurrent=4, >>> metric="mean_loss", mode="min") Note: In nevergrad v0.2.0+, optimizers can be instrumented. For instance, the following will specifies searching for "lr" from 1 to 2. >>> from nevergrad.optimization import optimizerlib >>> from nevergrad import instrumentation as inst >>> lr = inst.var.Array(1).bounded(1, 2).asfloat() >>> instrumentation = inst.Instrumentation(lr=lr) >>> optimizer = optimizerlib.OnePlusOne(instrumentation, budget=100) >>> algo = NevergradSearch(optimizer, None, max_concurrent=4, >>> metric="mean_loss", mode="min") """ def __init__(self, optimizer, parameter_names, max_concurrent=10, reward_attr=None, metric="episode_reward_mean", mode="max", **kwargs): assert ng is not None, "Nevergrad must be installed!" assert type(max_concurrent) is int and max_concurrent > 0 assert mode in ["min", "max"], "`mode` must be 'min' or 'max'!" if reward_attr is not None: mode = "max" metric = reward_attr logger.warning( "`reward_attr` is deprecated and will be removed in a future " "version of Tune. " "Setting `metric={}` and `mode=max`.".format(reward_attr)) self._max_concurrent = max_concurrent self._parameters = parameter_names self._metric = metric # nevergrad.tell internally minimizes, so "max" => -1 if mode == "max": self._metric_op = -1. elif mode == "min": self._metric_op = 1. self._nevergrad_opt = optimizer self._live_trial_mapping = {} super(NevergradSearch, self).__init__(**kwargs) # validate parameters if hasattr(optimizer, "instrumentation"): # added in v0.2.0 if optimizer.instrumentation.kwargs: if optimizer.instrumentation.args: raise ValueError( "Instrumented optimizers should use kwargs only") if parameter_names is not None: raise ValueError("Instrumented optimizers should provide " "None as parameter_names") else: if parameter_names is None: raise ValueError("Non-instrumented optimizers should have " "a list of parameter_names") if len(optimizer.instrumentation.args) != 1: raise ValueError( "Instrumented optimizers should use kwargs only") if parameter_names is not None and optimizer.dimension != len( parameter_names): raise ValueError("len(parameters_names) must match optimizer " "dimension for non-instrumented optimizers") def _suggest(self, trial_id): if self._num_live_trials() >= self._max_concurrent: return None suggested_config = self._nevergrad_opt.ask() self._live_trial_mapping[trial_id] = suggested_config # in v0.2.0+, output of ask() is a Candidate, # with fields args and kwargs if hasattr(self._nevergrad_opt, "instrumentation"): if not suggested_config.kwargs: return dict(zip(self._parameters, suggested_config.args[0])) else: return suggested_config.kwargs # legacy: output of ask() is a np.ndarray return dict(zip(self._parameters, suggested_config)) def on_trial_result(self, trial_id, result): pass def on_trial_complete(self, trial_id, result=None, error=False, early_terminated=False): """Notification for the completion of trial. The result is internally negated when interacting with Nevergrad so that Nevergrad Optimizers can "maximize" this value, as it minimizes on default. """ if result: self._process_result(trial_id, result, early_terminated) self._live_trial_mapping.pop(trial_id) def _process_result(self, trial_id, result, early_terminated=False): if early_terminated and self._use_early_stopped is False: return ng_trial_info = self._live_trial_mapping[trial_id] self._nevergrad_opt.tell(ng_trial_info, self._metric_op * result[self._metric]) def _num_live_trials(self): return len(self._live_trial_mapping) def save(self, checkpoint_dir): trials_object = (self._nevergrad_opt, self._parameters) with open(checkpoint_dir, "wb") as outputFile: pickle.dump(trials_object, outputFile) def restore(self, checkpoint_dir): with open(checkpoint_dir, "rb") as inputFile: trials_object = pickle.load(inputFile) self._nevergrad_opt = trials_object[0] self._parameters = trials_object[1]
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/suggest/search.py
Python
class SearchAlgorithm: """Interface of an event handler API for hyperparameter search. Unlike TrialSchedulers, SearchAlgorithms will not have the ability to modify the execution (i.e., stop and pause trials). Trials added manually (i.e., via the Client API) will also notify this class upon new events, so custom search algorithms should maintain a list of trials ID generated from this class. See also: `ray.tune.suggest.BasicVariantGenerator`. """ def add_configurations(self, experiments): """Tracks given experiment specifications. Arguments: experiments (Experiment | list | dict): Experiments to run. """ raise NotImplementedError def next_trials(self): """Provides Trial objects to be queued into the TrialRunner. Returns: trials (list): Returns a list of trials. """ raise NotImplementedError def on_trial_result(self, trial_id, result): """Called on each intermediate result returned by a trial. This will only be called when the trial is in the RUNNING state. Arguments: trial_id: Identifier for the trial. """ pass def on_trial_complete(self, trial_id, result=None, error=False, early_terminated=False): """Notification for the completion of trial. Arguments: trial_id: Identifier for the trial. result (dict): Defaults to None. A dict will be provided with this notification when the trial is in the RUNNING state AND either completes naturally or by manual termination. error (bool): Defaults to False. True if the trial is in the RUNNING state and errors. early_terminated (bool): Defaults to False. True if the trial is stopped while in PAUSED or PENDING state. """ pass def is_finished(self): """Returns True if no trials left to be queued into TrialRunner. Can return True before all trials have finished executing. """ raise NotImplementedError
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/suggest/sigopt.py
Python
import copy import os import logging import pickle try: import sigopt as sgo except ImportError: sgo = None from ray.tune.suggest.suggestion import SuggestionAlgorithm logger = logging.getLogger(__name__) class SigOptSearch(SuggestionAlgorithm): """A wrapper around SigOpt to provide trial suggestions. Requires SigOpt to be installed. Requires user to store their SigOpt API key locally as an environment variable at `SIGOPT_KEY`. Parameters: space (list of dict): SigOpt configuration. Parameters will be sampled from this configuration and will be used to override parameters generated in the variant generation process. name (str): Name of experiment. Required by SigOpt. max_concurrent (int): Number of maximum concurrent trials supported based on the user's SigOpt plan. Defaults to 1. metric (str): The training result objective value attribute. mode (str): One of {min, max}. Determines whether objective is minimizing or maximizing the metric attribute. Example: >>> space = [ >>> { >>> 'name': 'width', >>> 'type': 'int', >>> 'bounds': { >>> 'min': 0, >>> 'max': 20 >>> }, >>> }, >>> { >>> 'name': 'height', >>> 'type': 'int', >>> 'bounds': { >>> 'min': -100, >>> 'max': 100 >>> }, >>> }, >>> ] >>> algo = SigOptSearch( >>> space, name="SigOpt Example Experiment", >>> max_concurrent=1, metric="mean_loss", mode="min") """ def __init__(self, space, name="Default Tune Experiment", max_concurrent=1, reward_attr=None, metric="episode_reward_mean", mode="max", **kwargs): assert sgo is not None, "SigOpt must be installed!" assert type(max_concurrent) is int and max_concurrent > 0 assert "SIGOPT_KEY" in os.environ, \ "SigOpt API key must be stored as environ variable at SIGOPT_KEY" assert mode in ["min", "max"], "`mode` must be 'min' or 'max'!" if reward_attr is not None: mode = "max" metric = reward_attr logger.warning( "`reward_attr` is deprecated and will be removed in a future " "version of Tune. " "Setting `metric={}` and `mode=max`.".format(reward_attr)) if "use_early_stopped_trials" in kwargs: logger.warning( "`use_early_stopped_trials` is not used in SigOptSearch.") self._max_concurrent = max_concurrent self._metric = metric if mode == "max": self._metric_op = 1. elif mode == "min": self._metric_op = -1. self._live_trial_mapping = {} # Create a connection with SigOpt API, requires API key self.conn = sgo.Connection(client_token=os.environ["SIGOPT_KEY"]) self.experiment = self.conn.experiments().create( name=name, parameters=space, parallel_bandwidth=self._max_concurrent, ) super(SigOptSearch, self).__init__(**kwargs) def _suggest(self, trial_id): if self._num_live_trials() >= self._max_concurrent: return None # Get new suggestion from SigOpt suggestion = self.conn.experiments( self.experiment.id).suggestions().create() self._live_trial_mapping[trial_id] = suggestion return copy.deepcopy(suggestion.assignments) def on_trial_result(self, trial_id, result): pass def on_trial_complete(self, trial_id, result=None, error=False, early_terminated=False): """Notification for the completion of trial. If a trial fails, it will be reported as a failed Observation, telling the optimizer that the Suggestion led to a metric failure, which updates the feasible region and improves parameter recommendation. Creates SigOpt Observation object for trial. """ if result: self.conn.experiments(self.experiment.id).observations().create( suggestion=self._live_trial_mapping[trial_id].id, value=self._metric_op * result[self._metric], ) # Update the experiment object self.experiment = self.conn.experiments(self.experiment.id).fetch() elif error or early_terminated: # Reports a failed Observation self.conn.experiments(self.experiment.id).observations().create( failed=True, suggestion=self._live_trial_mapping[trial_id].id) del self._live_trial_mapping[trial_id] def _num_live_trials(self): return len(self._live_trial_mapping) def save(self, checkpoint_dir): trials_object = (self.conn, self.experiment) with open(checkpoint_dir, "wb") as outputFile: pickle.dump(trials_object, outputFile) def restore(self, checkpoint_dir): with open(checkpoint_dir, "rb") as inputFile: trials_object = pickle.load(inputFile) self.conn = trials_object[0] self.experiment = trials_object[1]
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/suggest/skopt.py
Python
import logging import pickle try: import skopt as sko except ImportError: sko = None from ray.tune.suggest.suggestion import SuggestionAlgorithm logger = logging.getLogger(__name__) def _validate_warmstart(parameter_names, points_to_evaluate, evaluated_rewards): if points_to_evaluate: if not isinstance(points_to_evaluate, list): raise TypeError( "points_to_evaluate expected to be a list, got {}.".format( type(points_to_evaluate))) for point in points_to_evaluate: if not isinstance(point, list): raise TypeError( "points_to_evaluate expected to include list, got {}.". format(point)) if not len(point) == len(parameter_names): raise ValueError("Dim of point {}".format(point) + " and parameter_names {}".format( parameter_names) + " do not match.") if points_to_evaluate and evaluated_rewards: if not isinstance(evaluated_rewards, list): raise TypeError( "evaluated_rewards expected to be a list, got {}.".format( type(evaluated_rewards))) if not len(evaluated_rewards) == len(points_to_evaluate): raise ValueError( "Dim of evaluated_rewards {}".format(evaluated_rewards) + " and points_to_evaluate {}".format(points_to_evaluate) + " do not match.") class SkOptSearch(SuggestionAlgorithm): """A wrapper around skopt to provide trial suggestions. Requires skopt to be installed. Parameters: optimizer (skopt.optimizer.Optimizer): Optimizer provided from skopt. parameter_names (list): List of parameter names. Should match the dimension of the optimizer output. max_concurrent (int): Number of maximum concurrent trials. Defaults to 10. metric (str): The training result objective value attribute. mode (str): One of {min, max}. Determines whether objective is minimizing or maximizing the metric attribute. points_to_evaluate (list of lists): A list of points you'd like to run first before sampling from the optimiser, e.g. these could be parameter configurations you already know work well to help the optimiser select good values. Each point is a list of the parameters using the order definition given by parameter_names. evaluated_rewards (list): If you have previously evaluated the parameters passed in as points_to_evaluate you can avoid re-running those trials by passing in the reward attributes as a list so the optimiser can be told the results without needing to re-compute the trial. Must be the same length as points_to_evaluate. (See tune/examples/skopt_example.py) use_early_stopped_trials (bool): Whether to use early terminated trial results in the optimization process. Example: >>> from skopt import Optimizer >>> optimizer = Optimizer([(0,20),(-100,100)]) >>> current_best_params = [[10, 0], [15, -20]] >>> algo = SkOptSearch(optimizer, >>> ["width", "height"], >>> max_concurrent=4, >>> metric="mean_loss", >>> mode="min", >>> points_to_evaluate=current_best_params) """ def __init__(self, optimizer, parameter_names, max_concurrent=10, reward_attr=None, metric="episode_reward_mean", mode="max", points_to_evaluate=None, evaluated_rewards=None, **kwargs): assert sko is not None, """skopt must be installed! You can install Skopt with the command: `pip install scikit-optimize`.""" assert type(max_concurrent) is int and max_concurrent > 0 _validate_warmstart(parameter_names, points_to_evaluate, evaluated_rewards) assert mode in ["min", "max"], "`mode` must be 'min' or 'max'!" if reward_attr is not None: mode = "max" metric = reward_attr logger.warning( "`reward_attr` is deprecated and will be removed in a future " "version of Tune. " "Setting `metric={}` and `mode=max`.".format(reward_attr)) self._initial_points = [] if points_to_evaluate and evaluated_rewards: optimizer.tell(points_to_evaluate, evaluated_rewards) elif points_to_evaluate: self._initial_points = points_to_evaluate self._max_concurrent = max_concurrent self._parameters = parameter_names self._metric = metric # Skopt internally minimizes, so "max" => -1 if mode == "max": self._metric_op = -1. elif mode == "min": self._metric_op = 1. self._skopt_opt = optimizer self._live_trial_mapping = {} super(SkOptSearch, self).__init__(**kwargs) def _suggest(self, trial_id): if self._num_live_trials() >= self._max_concurrent: return None if self._initial_points: suggested_config = self._initial_points[0] del self._initial_points[0] else: suggested_config = self._skopt_opt.ask() self._live_trial_mapping[trial_id] = suggested_config return dict(zip(self._parameters, suggested_config)) def on_trial_result(self, trial_id, result): pass def on_trial_complete(self, trial_id, result=None, error=False, early_terminated=False): """Notification for the completion of trial. The result is internally negated when interacting with Skopt so that Skopt Optimizers can "maximize" this value, as it minimizes on default. """ if result: self._process_result(trial_id, result, early_terminated) self._live_trial_mapping.pop(trial_id) def _process_result(self, trial_id, result, early_terminated=False): if early_terminated and self._use_early_stopped is False: return skopt_trial_info = self._live_trial_mapping[trial_id] self._skopt_opt.tell(skopt_trial_info, self._metric_op * result[self._metric]) def _num_live_trials(self): return len(self._live_trial_mapping) def save(self, checkpoint_dir): trials_object = (self._initial_points, self._skopt_opt) with open(checkpoint_dir, "wb") as outputFile: pickle.dump(trials_object, outputFile) def restore(self, checkpoint_dir): with open(checkpoint_dir, "rb") as inputFile: trials_object = pickle.load(inputFile) self._initial_points = trials_object[0] self._skopt_opt = trials_object[1]
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/suggest/suggestion.py
Python
import itertools import copy from ray.tune.error import TuneError from ray.tune.experiment import convert_to_experiment_list from ray.tune.config_parser import make_parser, create_trial_from_spec from ray.tune.suggest.search import SearchAlgorithm from ray.tune.suggest.variant_generator import format_vars, resolve_nested_dict from ray.tune.trial import Trial from ray.tune.utils import merge_dicts, flatten_dict class SuggestionAlgorithm(SearchAlgorithm): """Abstract class for suggestion-based algorithms. Custom search algorithms can extend this class easily by overriding the `_suggest` method provide generated parameters for the trials. To track suggestions and their corresponding evaluations, the method `_suggest` will be passed a trial_id, which will be used in subsequent notifications. Example: >>> suggester = SuggestionAlgorithm() >>> suggester.add_configurations({ ... }) >>> new_parameters = suggester._suggest() >>> suggester.on_trial_complete(trial_id, result) >>> better_parameters = suggester._suggest() """ def __init__(self, use_early_stopped_trials=True): """Constructs a generator given experiment specifications. """ self._parser = make_parser() self._trial_generator = [] self._counter = 0 self._finished = False self._use_early_stopped = use_early_stopped_trials def add_configurations(self, experiments): """Chains generator given experiment specifications. Arguments: experiments (Experiment | list | dict): Experiments to run. """ experiment_list = convert_to_experiment_list(experiments) for experiment in experiment_list: self._trial_generator = itertools.chain( self._trial_generator, self._generate_trials(experiment.spec, experiment.name)) def next_trials(self): """Provides a batch of Trial objects to be queued into the TrialRunner. A batch ends when self._trial_generator returns None. Returns: trials (list): Returns a list of trials. """ trials = [] for trial in self._trial_generator: if trial is None: return trials trials += [trial] self._finished = True return trials def _generate_trials(self, experiment_spec, output_path=""): """Generates trials with configurations from `_suggest`. Creates a trial_id that is passed into `_suggest`. Yields: Trial objects constructed according to `spec` """ if "run" not in experiment_spec: raise TuneError("Must specify `run` in {}".format(experiment_spec)) for _ in range(experiment_spec.get("num_samples", 1)): trial_id = Trial.generate_id() while True: suggested_config = self._suggest(trial_id) if suggested_config is None: yield None else: break spec = copy.deepcopy(experiment_spec) spec["config"] = merge_dicts(spec["config"], copy.deepcopy(suggested_config)) flattened_config = resolve_nested_dict(spec["config"]) self._counter += 1 tag = "{0}_{1}".format( str(self._counter), format_vars(flattened_config)) yield create_trial_from_spec( spec, output_path, self._parser, evaluated_params=flatten_dict(suggested_config), experiment_tag=tag, trial_id=trial_id) def is_finished(self): return self._finished def _suggest(self, trial_id): """Queries the algorithm to retrieve the next set of parameters. Arguments: trial_id: Trial ID used for subsequent notifications. Returns: dict|None: Configuration for a trial, if possible. Else, returns None, which will temporarily stop the TrialRunner from querying. Example: >>> suggester = SuggestionAlgorithm(max_concurrent=1) >>> suggester.add_configurations({ ... }) >>> parameters_1 = suggester._suggest() >>> parameters_2 = suggester._suggest() >>> parameters_2 is None >>> suggester.on_trial_complete(trial_id, result) >>> parameters_2 = suggester._suggest() >>> parameters_2 is not None """ raise NotImplementedError def save(self, checkpoint_dir): raise NotImplementedError def restore(self, checkpoint_dir): raise NotImplementedError class _MockSuggestionAlgorithm(SuggestionAlgorithm): def __init__(self, max_concurrent=2, **kwargs): self._max_concurrent = max_concurrent self.live_trials = {} self.counter = {"result": 0, "complete": 0} self.final_results = [] self.stall = False self.results = [] super(_MockSuggestionAlgorithm, self).__init__(**kwargs) def _suggest(self, trial_id): if len(self.live_trials) < self._max_concurrent and not self.stall: self.live_trials[trial_id] = 1 return {"test_variable": 2} return None def on_trial_result(self, trial_id, result): self.counter["result"] += 1 self.results += [result] def on_trial_complete(self, trial_id, result=None, error=False, early_terminated=False): self.counter["complete"] += 1 if result: self._process_result(result, early_terminated) del self.live_trials[trial_id] def _process_result(self, result, early_terminated): if early_terminated and self._use_early_stopped: self.final_results += [result]
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/suggest/variant_generator.py
Python
import copy import logging import numpy import random import types from ray.tune import TuneError from ray.tune.sample import sample_from logger = logging.getLogger(__name__) def generate_variants(unresolved_spec): """Generates variants from a spec (dict) with unresolved values. There are two types of unresolved values: Grid search: These define a grid search over values. For example, the following grid search values in a spec will produce six distinct variants in combination: "activation": grid_search(["relu", "tanh"]) "learning_rate": grid_search([1e-3, 1e-4, 1e-5]) Lambda functions: These are evaluated to produce a concrete value, and can express dependencies or conditional distributions between values. They can also be used to express random search (e.g., by calling into the `random` or `np` module). "cpu": lambda spec: spec.config.num_workers "batch_size": lambda spec: random.uniform(1, 1000) Finally, to support defining specs in plain JSON / YAML, grid search and lambda functions can also be defined alternatively as follows: "activation": {"grid_search": ["relu", "tanh"]} "cpu": {"eval": "spec.config.num_workers"} Use `format_vars` to format the returned dict of hyperparameters. Yields: (Dict of resolved variables, Spec object) """ for resolved_vars, spec in _generate_variants(unresolved_spec): assert not _unresolved_values(spec) yield resolved_vars, spec def grid_search(values): """Convenience method for specifying grid search over a value. Arguments: values: An iterable whose parameters will be gridded. """ return {"grid_search": values} _STANDARD_IMPORTS = { "random": random, "np": numpy, } _MAX_RESOLUTION_PASSES = 20 def resolve_nested_dict(nested_dict): """Flattens a nested dict by joining keys into tuple of paths. Can then be passed into `format_vars`. """ res = {} for k, v in nested_dict.items(): if isinstance(v, dict): for k_, v_ in resolve_nested_dict(v).items(): res[(k, ) + k_] = v_ else: res[(k, )] = v return res def format_vars(resolved_vars): """Formats the resolved variable dict into a single string.""" out = [] for path, value in sorted(resolved_vars.items()): if path[0] in ["run", "env", "resources_per_trial"]: continue # TrialRunner already has these in the experiment_tag pieces = [] last_string = True for k in path[::-1]: if isinstance(k, int): pieces.append(str(k)) elif last_string: last_string = False pieces.append(k) pieces.reverse() out.append(_clean_value("_".join(pieces)) + "=" + _clean_value(value)) return ",".join(out) def flatten_resolved_vars(resolved_vars): """Formats the resolved variable dict into a mapping of (str -> value).""" flattened_resolved_vars_dict = {} for pieces, value in resolved_vars.items(): if pieces[0] == "config": pieces = pieces[1:] pieces = [str(piece) for piece in pieces] flattened_resolved_vars_dict["/".join(pieces)] = value return flattened_resolved_vars_dict def _clean_value(value): if isinstance(value, float): return "{:.5}".format(value) else: return str(value).replace("/", "_") def _generate_variants(spec): spec = copy.deepcopy(spec) unresolved = _unresolved_values(spec) if not unresolved: yield {}, spec return grid_vars = [] lambda_vars = [] for path, value in unresolved.items(): if isinstance(value, types.FunctionType): lambda_vars.append((path, value)) else: grid_vars.append((path, value)) grid_vars.sort() grid_search = _grid_search_generator(spec, grid_vars) for resolved_spec in grid_search: resolved_vars = _resolve_lambda_vars(resolved_spec, lambda_vars) for resolved, spec in _generate_variants(resolved_spec): for path, value in grid_vars: resolved_vars[path] = _get_value(spec, path) for k, v in resolved.items(): if (k in resolved_vars and v != resolved_vars[k] and _is_resolved(resolved_vars[k])): raise ValueError( "The variable `{}` could not be unambiguously " "resolved to a single value. Consider simplifying " "your configuration.".format(k)) resolved_vars[k] = v yield resolved_vars, spec def _assign_value(spec, path, value): for k in path[:-1]: spec = spec[k] spec[path[-1]] = value def _get_value(spec, path): for k in path: spec = spec[k] return spec def _resolve_lambda_vars(spec, lambda_vars): resolved = {} error = True num_passes = 0 while error and num_passes < _MAX_RESOLUTION_PASSES: num_passes += 1 error = False for path, fn in lambda_vars: try: value = fn(_UnresolvedAccessGuard(spec)) except RecursiveDependencyError as e: error = e except Exception: raise ValueError( "Failed to evaluate expression: {}: {}".format(path, fn)) else: _assign_value(spec, path, value) resolved[path] = value if error: raise error return resolved def _grid_search_generator(unresolved_spec, grid_vars): value_indices = [0] * len(grid_vars) def increment(i): value_indices[i] += 1 if value_indices[i] >= len(grid_vars[i][1]): value_indices[i] = 0 if i + 1 < len(value_indices): return increment(i + 1) else: return True return False if not grid_vars: yield unresolved_spec return while value_indices[-1] < len(grid_vars[-1][1]): spec = copy.deepcopy(unresolved_spec) for i, (path, values) in enumerate(grid_vars): _assign_value(spec, path, values[value_indices[i]]) yield spec if grid_vars: done = increment(0) if done: break def _is_resolved(v): resolved, _ = _try_resolve(v) return resolved def _try_resolve(v): if isinstance(v, sample_from): # Function to sample from return False, v.func elif isinstance(v, dict) and len(v) == 1 and "eval" in v: # Lambda function in eval syntax return False, lambda spec: eval( v["eval"], _STANDARD_IMPORTS, {"spec": spec}) elif isinstance(v, dict) and len(v) == 1 and "grid_search" in v: # Grid search values grid_values = v["grid_search"] if not isinstance(grid_values, list): raise TuneError( "Grid search expected list of values, got: {}".format( grid_values)) return False, grid_values return True, v def _unresolved_values(spec): found = {} for k, v in spec.items(): resolved, v = _try_resolve(v) if not resolved: found[(k, )] = v elif isinstance(v, dict): # Recurse into a dict for (path, value) in _unresolved_values(v).items(): found[(k, ) + path] = value elif isinstance(v, list): # Recurse into a list for i, elem in enumerate(v): for (path, value) in _unresolved_values({i: elem}).items(): found[(k, ) + path] = value return found class _UnresolvedAccessGuard(dict): def __init__(self, *args, **kwds): super(_UnresolvedAccessGuard, self).__init__(*args, **kwds) self.__dict__ = self def __getattribute__(self, item): value = dict.__getattribute__(self, item) if not _is_resolved(value): raise RecursiveDependencyError( "`{}` recursively depends on {}".format(item, value)) elif isinstance(value, dict): return _UnresolvedAccessGuard(value) else: return value class RecursiveDependencyError(Exception): def __init__(self, msg): Exception.__init__(self, msg)
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/sync_client.py
Python
import distutils import distutils.spawn import logging import subprocess import tempfile import types from shlex import quote from ray.tune.error import TuneError logger = logging.getLogger(__name__) S3_PREFIX = "s3://" GS_PREFIX = "gs://" ALLOWED_REMOTE_PREFIXES = (S3_PREFIX, GS_PREFIX) noop_template = ": {target}" # noop in bash def noop(*args): return def get_sync_client(sync_function, delete_function=None): """Returns a sync client. Args: sync_function (Optional[str|function]): Sync function. delete_function (Optional[str|function]): Delete function. Must be the same type as sync_function if it is provided. Raises: ValueError if sync_function or delete_function are malformed. """ if sync_function is None: return None if delete_function and type(sync_function) != type(delete_function): raise ValueError("Sync and delete functions must be of same type.") if isinstance(sync_function, types.FunctionType): delete_function = delete_function or noop client_cls = FunctionBasedClient elif isinstance(sync_function, str): delete_function = delete_function or noop_template client_cls = CommandBasedClient else: raise ValueError("Sync function {} must be string or function".format( sync_function)) return client_cls(sync_function, sync_function, delete_function) def get_cloud_sync_client(remote_path): """Returns a CommandBasedClient that can sync to/from remote storage. Args: remote_path (str): Path to remote storage (S3 or GS). Raises: ValueError if malformed remote_dir. """ if remote_path.startswith(S3_PREFIX): if not distutils.spawn.find_executable("aws"): raise ValueError( "Upload uri starting with '{}' requires awscli tool" " to be installed".format(S3_PREFIX)) template = "aws s3 sync {source} {target} --only-show-errors" delete_template = "aws s3 rm {target} --recursive --only-show-errors" elif remote_path.startswith(GS_PREFIX): if not distutils.spawn.find_executable("gsutil"): raise ValueError( "Upload uri starting with '{}' requires gsutil tool" " to be installed".format(GS_PREFIX)) template = "gsutil rsync -r {source} {target}" delete_template = "gsutil rm -r {target}" else: raise ValueError("Upload uri must start with one of: {}" "".format(ALLOWED_REMOTE_PREFIXES)) return CommandBasedClient(template, template, delete_template) class SyncClient: """Client interface for interacting with remote storage options.""" def sync_up(self, source, target): """Syncs up from source to target. Args: source (str): Source path. target (str): Target path. Returns: True if sync initiation successful, False otherwise. """ raise NotImplementedError def sync_down(self, source, target): """Syncs down from source to target. Args: source (str): Source path. target (str): Target path. Returns: True if sync initiation successful, False otherwise. """ raise NotImplementedError def delete(self, target): """Deletes target. Args: target (str): Target path. Returns: True if delete initiation successful, False otherwise. """ raise NotImplementedError def wait(self): """Waits for current sync to complete, if asynchronously started.""" pass def reset(self): """Resets state.""" pass class FunctionBasedClient(SyncClient): def __init__(self, sync_up_func, sync_down_func, delete_func=None): self.sync_up_func = sync_up_func self.sync_down_func = sync_down_func self.delete_func = delete_func or noop def sync_up(self, source, target): self.sync_up_func(source, target) return True def sync_down(self, source, target): self.sync_down_func(source, target) return True def delete(self, target): self.delete_func(target) return True NOOP = FunctionBasedClient(noop, noop) class CommandBasedClient(SyncClient): def __init__(self, sync_up_template, sync_down_template, delete_template=noop_template): """Syncs between two directories with the given command. Arguments: sync_up_template (str): A runnable string template; needs to include replacement fields '{source}' and '{target}'. sync_down_template (str): A runnable string template; needs to include replacement fields '{source}' and '{target}'. delete_template (Optional[str]): A runnable string template; needs to include replacement field '{target}'. Noop by default. """ self._validate_sync_string(sync_up_template) self._validate_sync_string(sync_down_template) self.sync_up_template = sync_up_template self.sync_down_template = sync_down_template self.delete_template = delete_template self.logfile = None self.cmd_process = None def set_logdir(self, logdir): """Sets the directory to log sync execution output in. Args: logdir (str): Log directory. """ self.logfile = tempfile.NamedTemporaryFile( prefix="log_sync_out", dir=logdir, suffix=".log", delete=False) def sync_up(self, source, target): return self._execute(self.sync_up_template, source, target) def sync_down(self, source, target): return self._execute(self.sync_down_template, source, target) def delete(self, target): if self.is_running: logger.warning("Last sync client cmd still in progress, skipping.") return False final_cmd = self.delete_template.format(target=quote(target)) logger.debug("Running delete: {}".format(final_cmd)) self.cmd_process = subprocess.Popen( final_cmd, shell=True, stderr=subprocess.PIPE, stdout=self.logfile) return True def wait(self): if self.cmd_process: _, error_msg = self.cmd_process.communicate() error_msg = error_msg.decode("ascii") code = self.cmd_process.returncode args = self.cmd_process.args self.cmd_process = None if code != 0: raise TuneError("Sync error. Ran command: {}\n" "Error message ({}): {}".format( args, code, error_msg)) def reset(self): if self.is_running: logger.warning("Sync process still running but resetting anyways.") self.cmd_process = None @property def is_running(self): """Returns whether a sync or delete process is running.""" if self.cmd_process: self.cmd_process.poll() return self.cmd_process.returncode is None return False def _execute(self, sync_template, source, target): """Executes sync_template on source and target.""" if self.is_running: logger.warning("Last sync client cmd still in progress, skipping.") return False final_cmd = sync_template.format( source=quote(source), target=quote(target)) logger.debug("Running sync: {}".format(final_cmd)) self.cmd_process = subprocess.Popen( final_cmd, shell=True, stderr=subprocess.PIPE, stdout=self.logfile) return True @staticmethod def _validate_sync_string(sync_string): if not isinstance(sync_string, str): raise ValueError("{} is not a string.".format(sync_string)) if "{source}" not in sync_string: raise ValueError("Sync template missing '{source}'.") if "{target}" not in sync_string: raise ValueError("Sync template missing '{target}'.")
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/syncer.py
Python
import distutils import logging import os import time from shlex import quote from ray import services from ray.tune.cluster_info import get_ssh_key, get_ssh_user from ray.tune.sync_client import (CommandBasedClient, get_sync_client, get_cloud_sync_client, NOOP) logger = logging.getLogger(__name__) SYNC_PERIOD = 300 _log_sync_warned = False _syncers = {} def wait_for_sync(): for syncer in _syncers.values(): syncer.wait() def log_sync_template(options=""): """Template enabling syncs between driver and worker when possible. Requires ray cluster to be started with the autoscaler. Also requires rsync to be installed. Args: options (str): Additional rsync options. Returns: Sync template with source and target parameters. None if rsync unavailable. """ if not distutils.spawn.find_executable("rsync"): logger.error("Log sync requires rsync to be installed.") return None global _log_sync_warned ssh_key = get_ssh_key() if ssh_key is None: if not _log_sync_warned: logger.debug("Log sync requires cluster to be setup with " "`ray up`.") _log_sync_warned = True return None rsh = "ssh -i {ssh_key} -o ConnectTimeout=120s -o StrictHostKeyChecking=no" rsh = rsh.format(ssh_key=quote(ssh_key)) template = "rsync {options} -savz -e {rsh} {{source}} {{target}}" return template.format(options=options, rsh=quote(rsh)) class Syncer: def __init__(self, local_dir, remote_dir, sync_client=NOOP): """Syncs between two directories with the sync_function. Arguments: local_dir (str): Directory to sync. Uniquely identifies the syncer. remote_dir (str): Remote directory to sync with. sync_client (SyncClient): Client for syncing between local_dir and remote_dir. Defaults to a Noop. """ self._local_dir = (os.path.join(local_dir, "") if local_dir else local_dir) self._remote_dir = remote_dir self.last_sync_up_time = float("-inf") self.last_sync_down_time = float("-inf") self.sync_client = sync_client def sync_up_if_needed(self): if time.time() - self.last_sync_up_time > SYNC_PERIOD: self.sync_up() def sync_down_if_needed(self): if time.time() - self.last_sync_down_time > SYNC_PERIOD: self.sync_down() def sync_up(self): """Attempts to start the sync-up to the remote path. Returns: Whether the sync (if feasible) was successfully started. """ result = False if self.validate_hosts(self._local_dir, self._remote_path): try: result = self.sync_client.sync_up(self._local_dir, self._remote_path) self.last_sync_up_time = time.time() except Exception: logger.exception("Sync execution failed.") return result def sync_down(self): """Attempts to start the sync-down from the remote path. Returns: Whether the sync (if feasible) was successfully started. """ result = False if self.validate_hosts(self._local_dir, self._remote_path): try: result = self.sync_client.sync_down(self._remote_path, self._local_dir) self.last_sync_down_time = time.time() except Exception: logger.exception("Sync execution failed.") return result def validate_hosts(self, source, target): if not (source and target): logger.debug("Source or target is empty, skipping log sync for " "{}".format(self._local_dir)) return False return True def wait(self): """Waits for the sync client to complete the current sync.""" self.sync_client.wait() def reset(self): self.last_sync_up_time = float("-inf") self.last_sync_down_time = float("-inf") self.sync_client.reset() @property def _remote_path(self): return self._remote_dir class NodeSyncer(Syncer): """Syncer for syncing files to/from a remote dir to a local dir.""" def __init__(self, local_dir, remote_dir, sync_client): self.local_ip = services.get_node_ip_address() self.worker_ip = None super(NodeSyncer, self).__init__(local_dir, remote_dir, sync_client) def set_worker_ip(self, worker_ip): """Sets the worker IP to sync logs from.""" self.worker_ip = worker_ip def has_remote_target(self): """Returns whether the Syncer has a remote target.""" if not self.worker_ip: logger.debug("Worker IP unknown, skipping sync for %s", self._local_dir) return False if self.worker_ip == self.local_ip: logger.debug("Worker IP is local IP, skipping sync for %s", self._local_dir) return False return True def sync_up_if_needed(self): if not self.has_remote_target(): return True return super(NodeSyncer, self).sync_up_if_needed() def sync_down_if_needed(self): if not self.has_remote_target(): return True return super(NodeSyncer, self).sync_down_if_needed() def sync_up_to_new_location(self, worker_ip): if worker_ip != self.worker_ip: logger.debug("Setting new worker IP to %s", worker_ip) self.set_worker_ip(worker_ip) self.reset() if not self.sync_up(): logger.warning( "Sync up to new location skipped. This should not occur.") else: logger.warning("Sync attempted to same IP %s.", worker_ip) def sync_up(self): if not self.has_remote_target(): return True return super(NodeSyncer, self).sync_up() def sync_down(self): if not self.has_remote_target(): return True logger.debug("Syncing from %s to %s", self._remote_path, self._local_dir) return super(NodeSyncer, self).sync_down() @property def _remote_path(self): ssh_user = get_ssh_user() global _log_sync_warned if not self.has_remote_target(): return None if ssh_user is None: if not _log_sync_warned: logger.error("Syncer requires cluster to be setup with " "`ray up`.") _log_sync_warned = True return None return "{}@{}:{}/".format(ssh_user, self.worker_ip, self._remote_dir) def get_cloud_syncer(local_dir, remote_dir=None, sync_function=None): """Returns a Syncer. This syncer is in charge of syncing the local_dir with upload_dir. Args: local_dir (str): Source directory for syncing. remote_dir (str): Target directory for syncing. If not provided, a no-op Syncer is returned. sync_function (func | str): Function for syncing the local_dir to remote_dir. If string, then it must be a string template for syncer to run. If not provided, it defaults to standard S3 or gsutil sync commands. Raises: ValueError if malformed remote_dir. """ key = (local_dir, remote_dir) if key in _syncers: return _syncers[key] if not remote_dir: _syncers[key] = Syncer(local_dir, remote_dir, NOOP) return _syncers[key] client = get_sync_client(sync_function) if client: _syncers[key] = Syncer(local_dir, remote_dir, client) return _syncers[key] sync_client = get_cloud_sync_client(remote_dir) _syncers[key] = Syncer(local_dir, remote_dir, sync_client) return _syncers[key] def get_node_syncer(local_dir, remote_dir=None, sync_function=None): """Returns a NodeSyncer. Args: local_dir (str): Source directory for syncing. remote_dir (str): Target directory for syncing. If not provided, a noop Syncer is returned. sync_function (func|str|bool): Function for syncing the local_dir to remote_dir. If string, then it must be a string template for syncer to run. If True or not provided, it defaults rsync. If False, a noop Syncer is returned. """ key = (local_dir, remote_dir) if key in _syncers: return _syncers[key] elif not remote_dir or sync_function is False: sync_client = NOOP elif sync_function and sync_function is not True: sync_client = get_sync_client(sync_function) else: sync = log_sync_template() if sync: sync_client = CommandBasedClient(sync, sync) sync_client.set_logdir(local_dir) else: sync_client = NOOP _syncers[key] = NodeSyncer(local_dir, remote_dir, sync_client) return _syncers[key]
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/example.py
Python
# flake8: noqa # This is an example quickstart for Tune. # To connect to a cluster, uncomment below: # import ray # import argparse # parser = argparse.ArgumentParser() # parser.add_argument("--address") # args = parser.parse_args() # ray.init(address=args.address) # __quick_start_begin__ import torch.optim as optim from ray import tune from ray.tune.examples.mnist_pytorch import get_data_loaders, ConvNet, train, test def train_mnist(config): train_loader, test_loader = get_data_loaders() model = ConvNet() optimizer = optim.SGD(model.parameters(), lr=config["lr"]) for i in range(10): train(model, optimizer, train_loader) acc = test(model, test_loader) tune.track.log(mean_accuracy=acc) analysis = tune.run( train_mnist, config={"lr": tune.grid_search([0.001, 0.01, 0.1])}) print("Best config: ", analysis.get_best_config(metric="mean_accuracy")) # Get a dataframe for analyzing trial results. df = analysis.dataframe() # __quick_start_end__
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/test_actor_reuse.py
Python
import unittest import ray from ray.tune import Trainable, run_experiments from ray.tune.error import TuneError from ray.tune.schedulers.trial_scheduler import FIFOScheduler, TrialScheduler class FrequentPausesScheduler(FIFOScheduler): def on_trial_result(self, trial_runner, trial, result): return TrialScheduler.PAUSE def create_resettable_class(): class MyResettableClass(Trainable): def _setup(self, config): self.config = config self.num_resets = 0 self.iter = 0 def _train(self): self.iter += 1 return {"num_resets": self.num_resets, "done": self.iter > 1} def _save(self, chkpt_dir): return {"iter": self.iter} def _restore(self, item): self.iter = item["iter"] def reset_config(self, new_config): if "fake_reset_not_supported" in self.config: return False self.num_resets += 1 return True return MyResettableClass class ActorReuseTest(unittest.TestCase): def setUp(self): ray.init(num_cpus=1, num_gpus=0) def tearDown(self): ray.shutdown() def testTrialReuseDisabled(self): trials = run_experiments( { "foo": { "run": create_resettable_class(), "num_samples": 4, "config": {}, } }, reuse_actors=False, scheduler=FrequentPausesScheduler()) self.assertEqual([t.last_result["num_resets"] for t in trials], [0, 0, 0, 0]) def testTrialReuseEnabled(self): trials = run_experiments( { "foo": { "run": create_resettable_class(), "num_samples": 4, "config": {}, } }, reuse_actors=True, scheduler=FrequentPausesScheduler()) self.assertEqual([t.last_result["num_resets"] for t in trials], [1, 2, 3, 4]) def testTrialReuseEnabledError(self): def run(): run_experiments( { "foo": { "run": create_resettable_class(), "max_failures": 1, "num_samples": 4, "config": { "fake_reset_not_supported": True }, } }, reuse_actors=True, scheduler=FrequentPausesScheduler()) self.assertRaises(TuneError, lambda: run()) if __name__ == "__main__": import pytest import sys sys.exit(pytest.main(["-v", __file__]))
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/test_api.py
Python
import shutil import copy import os import time import unittest from unittest.mock import patch import ray from ray.rllib import _register_all from ray import tune from ray.tune import DurableTrainable, Trainable, TuneError from ray.tune import register_env, register_trainable, run_experiments from ray.tune.schedulers import TrialScheduler, FIFOScheduler from ray.tune.trial import Trial from ray.tune.result import (TIMESTEPS_TOTAL, DONE, HOSTNAME, NODE_IP, PID, EPISODES_TOTAL, TRAINING_ITERATION, TIMESTEPS_THIS_ITER, TIME_THIS_ITER_S, TIME_TOTAL_S, TRIAL_ID, EXPERIMENT_TAG) from ray.tune.logger import Logger from ray.tune.experiment import Experiment from ray.tune.resources import Resources from ray.tune.suggest import grid_search from ray.tune.suggest.suggestion import _MockSuggestionAlgorithm from ray.tune.utils import (flatten_dict, get_pinned_object, pin_in_object_store) from ray.tune.utils.mock import mock_storage_client, MOCK_REMOTE_DIR class TrainableFunctionApiTest(unittest.TestCase): def setUp(self): ray.init(num_cpus=4, num_gpus=0, object_store_memory=150 * 1024 * 1024) def tearDown(self): ray.shutdown() _register_all() # re-register the evicted objects def checkAndReturnConsistentLogs(self, results, sleep_per_iter=None): """Checks logging is the same between APIs. Ignore "DONE" for logging but checks that the scheduler is notified properly with the last result. """ class_results = copy.deepcopy(results) function_results = copy.deepcopy(results) class_output = [] function_output = [] scheduler_notif = [] class MockScheduler(FIFOScheduler): def on_trial_complete(self, runner, trial, result): scheduler_notif.append(result) class ClassAPILogger(Logger): def on_result(self, result): class_output.append(result) class FunctionAPILogger(Logger): def on_result(self, result): function_output.append(result) class _WrappedTrainable(Trainable): def _setup(self, config): del config self._result_iter = copy.deepcopy(class_results) def _train(self): if sleep_per_iter: time.sleep(sleep_per_iter) res = self._result_iter.pop(0) # This should not fail if not self._result_iter: # Mark "Done" for last result res[DONE] = True return res def _function_trainable(config, reporter): for result in function_results: if sleep_per_iter: time.sleep(sleep_per_iter) reporter(**result) class_trainable_name = "class_trainable" register_trainable(class_trainable_name, _WrappedTrainable) trials = run_experiments( { "function_api": { "run": _function_trainable, "loggers": [FunctionAPILogger], }, "class_api": { "run": class_trainable_name, "loggers": [ClassAPILogger], }, }, raise_on_failed_trial=False, scheduler=MockScheduler()) # Ignore these fields NO_COMPARE_FIELDS = { HOSTNAME, NODE_IP, TRIAL_ID, EXPERIMENT_TAG, PID, TIME_THIS_ITER_S, TIME_TOTAL_S, DONE, # This is ignored because FunctionAPI has different handling "timestamp", "time_since_restore", "experiment_id", "date", } self.assertEqual(len(class_output), len(results)) self.assertEqual(len(function_output), len(results)) def as_comparable_result(result): return { k: v for k, v in result.items() if k not in NO_COMPARE_FIELDS } function_comparable = [ as_comparable_result(result) for result in function_output ] class_comparable = [ as_comparable_result(result) for result in class_output ] self.assertEqual(function_comparable, class_comparable) self.assertEqual(sum(t.get(DONE) for t in scheduler_notif), 2) self.assertEqual( as_comparable_result(scheduler_notif[0]), as_comparable_result(scheduler_notif[1])) # Make sure the last result is the same. self.assertEqual( as_comparable_result(trials[0].last_result), as_comparable_result(trials[1].last_result)) return function_output, trials def testPinObject(self): X = pin_in_object_store("hello") @ray.remote def f(): return get_pinned_object(X) self.assertEqual(ray.get(f.remote()), "hello") def testFetchPinned(self): X = pin_in_object_store("hello") def train(config, reporter): get_pinned_object(X) reporter(timesteps_total=100, done=True) register_trainable("f1", train) [trial] = run_experiments({ "foo": { "run": "f1", } }) self.assertEqual(trial.status, Trial.TERMINATED) self.assertEqual(trial.last_result[TIMESTEPS_TOTAL], 100) def testRegisterEnv(self): register_env("foo", lambda: None) self.assertRaises(TypeError, lambda: register_env("foo", 2)) def testRegisterEnvOverwrite(self): def train(config, reporter): reporter(timesteps_total=100, done=True) def train2(config, reporter): reporter(timesteps_total=200, done=True) register_trainable("f1", train) register_trainable("f1", train2) [trial] = run_experiments({ "foo": { "run": "f1", } }) self.assertEqual(trial.status, Trial.TERMINATED) self.assertEqual(trial.last_result[TIMESTEPS_TOTAL], 200) def testRegisterTrainable(self): def train(config, reporter): pass class A: pass class B(Trainable): pass register_trainable("foo", train) Experiment("test", train) register_trainable("foo", B) Experiment("test", B) self.assertRaises(TypeError, lambda: register_trainable("foo", B())) self.assertRaises(TuneError, lambda: Experiment("foo", B())) self.assertRaises(TypeError, lambda: register_trainable("foo", A)) self.assertRaises(TypeError, lambda: Experiment("foo", A)) def testTrainableCallable(self): def dummy_fn(config, reporter, steps): reporter(timesteps_total=steps, done=True) from functools import partial steps = 500 register_trainable("test", partial(dummy_fn, steps=steps)) [trial] = run_experiments({ "foo": { "run": "test", } }) self.assertEqual(trial.status, Trial.TERMINATED) self.assertEqual(trial.last_result[TIMESTEPS_TOTAL], steps) [trial] = tune.run(partial(dummy_fn, steps=steps)).trials self.assertEqual(trial.status, Trial.TERMINATED) self.assertEqual(trial.last_result[TIMESTEPS_TOTAL], steps) def testBuiltInTrainableResources(self): class B(Trainable): @classmethod def default_resource_request(cls, config): return Resources(cpu=config["cpu"], gpu=config["gpu"]) def _train(self): return {"timesteps_this_iter": 1, "done": True} register_trainable("B", B) def f(cpus, gpus, queue_trials): return run_experiments( { "foo": { "run": "B", "config": { "cpu": cpus, "gpu": gpus, }, } }, queue_trials=queue_trials)[0] # Should all succeed self.assertEqual(f(0, 0, False).status, Trial.TERMINATED) self.assertEqual(f(1, 0, True).status, Trial.TERMINATED) self.assertEqual(f(1, 0, True).status, Trial.TERMINATED) # Too large resource request self.assertRaises(TuneError, lambda: f(100, 100, False)) self.assertRaises(TuneError, lambda: f(0, 100, False)) self.assertRaises(TuneError, lambda: f(100, 0, False)) # TODO(ekl) how can we test this is queued (hangs)? # f(100, 0, True) def testRewriteEnv(self): def train(config, reporter): reporter(timesteps_total=1) register_trainable("f1", train) [trial] = run_experiments({ "foo": { "run": "f1", "env": "CartPole-v0", } }) self.assertEqual(trial.config["env"], "CartPole-v0") def testConfigPurity(self): def train(config, reporter): assert config == {"a": "b"}, config reporter(timesteps_total=1) register_trainable("f1", train) run_experiments({ "foo": { "run": "f1", "config": { "a": "b" }, } }) def testLogdir(self): def train(config, reporter): assert "/tmp/logdir/foo" in os.getcwd(), os.getcwd() reporter(timesteps_total=1) register_trainable("f1", train) run_experiments({ "foo": { "run": "f1", "local_dir": "/tmp/logdir", "config": { "a": "b" }, } }) def testLogdirStartingWithTilde(self): local_dir = "~/ray_results/local_dir" def train(config, reporter): cwd = os.getcwd() assert cwd.startswith(os.path.expanduser(local_dir)), cwd assert not cwd.startswith("~"), cwd reporter(timesteps_total=1) register_trainable("f1", train) run_experiments({ "foo": { "run": "f1", "local_dir": local_dir, "config": { "a": "b" }, } }) def testLongFilename(self): def train(config, reporter): assert "/tmp/logdir/foo" in os.getcwd(), os.getcwd() reporter(timesteps_total=1) register_trainable("f1", train) run_experiments({ "foo": { "run": "f1", "local_dir": "/tmp/logdir", "config": { "a" * 50: tune.sample_from(lambda spec: 5.0 / 7), "b" * 50: tune.sample_from(lambda spec: "long" * 40), }, } }) def testBadParams(self): def f(): run_experiments({"foo": {}}) self.assertRaises(TuneError, f) def testBadParams2(self): def f(): run_experiments({ "foo": { "run": "asdf", "bah": "this param is not allowed", } }) self.assertRaises(TuneError, f) def testBadParams3(self): def f(): run_experiments({ "foo": { "run": grid_search("invalid grid search"), } }) self.assertRaises(TuneError, f) def testBadParams4(self): def f(): run_experiments({ "foo": { "run": "asdf", } }) self.assertRaises(TuneError, f) def testBadParams5(self): def f(): run_experiments({"foo": {"run": "PPO", "stop": {"asdf": 1}}}) self.assertRaises(TuneError, f) def testBadParams6(self): def f(): run_experiments({ "foo": { "run": "PPO", "resources_per_trial": { "asdf": 1 } } }) self.assertRaises(TuneError, f) def testBadStoppingReturn(self): def train(config, reporter): reporter() register_trainable("f1", train) def f(): run_experiments({ "foo": { "run": "f1", "stop": { "time": 10 }, } }) self.assertRaises(TuneError, f) def testNestedStoppingReturn(self): def train(config, reporter): for i in range(10): reporter(test={"test1": {"test2": i}}) with self.assertRaises(TuneError): [trial] = tune.run( train, stop={ "test": { "test1": { "test2": 6 } } }).trials [trial] = tune.run(train, stop={"test/test1/test2": 6}).trials self.assertEqual(trial.last_result["training_iteration"], 7) def testStoppingFunction(self): def train(config, reporter): for i in range(10): reporter(test=i) def stop(trial_id, result): return result["test"] > 6 [trial] = tune.run(train, stop=stop).trials self.assertEqual(trial.last_result["training_iteration"], 8) def testStoppingMemberFunction(self): def train(config, reporter): for i in range(10): reporter(test=i) class Stopper: def stop(self, trial_id, result): return result["test"] > 6 [trial] = tune.run(train, stop=Stopper().stop).trials self.assertEqual(trial.last_result["training_iteration"], 8) def testBadStoppingFunction(self): def train(config, reporter): for i in range(10): reporter(test=i) class Stopper: def stop(self, result): return result["test"] > 6 def stop(result): return result["test"] > 6 with self.assertRaises(ValueError): tune.run(train, stop=Stopper().stop) with self.assertRaises(ValueError): tune.run(train, stop=stop) def testEarlyReturn(self): def train(config, reporter): reporter(timesteps_total=100, done=True) time.sleep(99999) register_trainable("f1", train) [trial] = run_experiments({ "foo": { "run": "f1", } }) self.assertEqual(trial.status, Trial.TERMINATED) self.assertEqual(trial.last_result[TIMESTEPS_TOTAL], 100) def testReporterNoUsage(self): def run_task(config, reporter): print("hello") experiment = Experiment(run=run_task, name="ray_crash_repro") [trial] = ray.tune.run(experiment).trials print(trial.last_result) self.assertEqual(trial.last_result[DONE], True) def testErrorReturn(self): def train(config, reporter): raise Exception("uh oh") register_trainable("f1", train) def f(): run_experiments({ "foo": { "run": "f1", } }) self.assertRaises(TuneError, f) def testSuccess(self): def train(config, reporter): for i in range(100): reporter(timesteps_total=i) register_trainable("f1", train) [trial] = run_experiments({ "foo": { "run": "f1", } }) self.assertEqual(trial.status, Trial.TERMINATED) self.assertEqual(trial.last_result[TIMESTEPS_TOTAL], 99) def testNoRaiseFlag(self): def train(config, reporter): raise Exception() register_trainable("f1", train) [trial] = run_experiments( { "foo": { "run": "f1", } }, raise_on_failed_trial=False) self.assertEqual(trial.status, Trial.ERROR) def testReportInfinity(self): def train(config, reporter): for i in range(100): reporter(mean_accuracy=float("inf")) register_trainable("f1", train) [trial] = run_experiments({ "foo": { "run": "f1", } }) self.assertEqual(trial.status, Trial.TERMINATED) self.assertEqual(trial.last_result["mean_accuracy"], float("inf")) def testNestedResults(self): def create_result(i): return {"test": {"1": {"2": {"3": i, "4": False}}}} flattened_keys = list(flatten_dict(create_result(0))) class _MockScheduler(FIFOScheduler): results = [] def on_trial_result(self, trial_runner, trial, result): self.results += [result] return TrialScheduler.CONTINUE def on_trial_complete(self, trial_runner, trial, result): self.complete_result = result def train(config, reporter): for i in range(100): reporter(**create_result(i)) algo = _MockSuggestionAlgorithm() scheduler = _MockScheduler() [trial] = tune.run( train, scheduler=scheduler, search_alg=algo, stop={ "test/1/2/3": 20 }).trials self.assertEqual(trial.status, Trial.TERMINATED) self.assertEqual(trial.last_result["test"]["1"]["2"]["3"], 20) self.assertEqual(trial.last_result["test"]["1"]["2"]["4"], False) self.assertEqual(trial.last_result[TRAINING_ITERATION], 21) self.assertEqual(len(scheduler.results), 20) self.assertTrue( all( set(result) >= set(flattened_keys) for result in scheduler.results)) self.assertTrue(set(scheduler.complete_result) >= set(flattened_keys)) self.assertEqual(len(algo.results), 20) self.assertTrue( all(set(result) >= set(flattened_keys) for result in algo.results)) with self.assertRaises(TuneError): [trial] = tune.run(train, stop={"1/2/3": 20}) with self.assertRaises(TuneError): [trial] = tune.run(train, stop={"test": 1}).trials def testReportTimeStep(self): # Test that no timestep count are logged if never the Trainable never # returns any. results1 = [dict(mean_accuracy=5, done=i == 99) for i in range(100)] logs1, _ = self.checkAndReturnConsistentLogs(results1) self.assertTrue(all(log[TIMESTEPS_TOTAL] is None for log in logs1)) # Test that no timesteps_this_iter are logged if only timesteps_total # are returned. results2 = [dict(timesteps_total=5, done=i == 9) for i in range(10)] logs2, _ = self.checkAndReturnConsistentLogs(results2) # Re-run the same trials but with added delay. This is to catch some # inconsistent timestep counting that was present in the multi-threaded # FunctionRunner. This part of the test can be removed once the # multi-threaded FunctionRunner is removed from ray/tune. # TODO: remove once the multi-threaded function runner is gone. logs2, _ = self.checkAndReturnConsistentLogs(results2, 0.5) # check all timesteps_total report the same value self.assertTrue(all(log[TIMESTEPS_TOTAL] == 5 for log in logs2)) # check that none of the logs report timesteps_this_iter self.assertFalse( any(hasattr(log, TIMESTEPS_THIS_ITER) for log in logs2)) # Test that timesteps_total and episodes_total are reported when # timesteps_this_iter and episodes_this_iter despite only return zeros. results3 = [ dict(timesteps_this_iter=0, episodes_this_iter=0) for i in range(10) ] logs3, _ = self.checkAndReturnConsistentLogs(results3) self.assertTrue(all(log[TIMESTEPS_TOTAL] == 0 for log in logs3)) self.assertTrue(all(log[EPISODES_TOTAL] == 0 for log in logs3)) # Test that timesteps_total and episodes_total are properly counted # when timesteps_this_iter and episodes_this_iter report non-zero # values. results4 = [ dict(timesteps_this_iter=3, episodes_this_iter=i) for i in range(10) ] logs4, _ = self.checkAndReturnConsistentLogs(results4) # The last reported result should not be double-logged. self.assertEqual(logs4[-1][TIMESTEPS_TOTAL], 30) self.assertNotEqual(logs4[-2][TIMESTEPS_TOTAL], logs4[-1][TIMESTEPS_TOTAL]) self.assertEqual(logs4[-1][EPISODES_TOTAL], 45) self.assertNotEqual(logs4[-2][EPISODES_TOTAL], logs4[-1][EPISODES_TOTAL]) def testAllValuesReceived(self): results1 = [ dict(timesteps_total=(i + 1), my_score=i**2, done=i == 4) for i in range(5) ] logs1, _ = self.checkAndReturnConsistentLogs(results1) # check if the correct number of results were reported self.assertEqual(len(logs1), len(results1)) def check_no_missing(reported_result, result): common_results = [reported_result[k] == result[k] for k in result] return all(common_results) # check that no result was dropped or modified complete_results = [ check_no_missing(log, result) for log, result in zip(logs1, results1) ] self.assertTrue(all(complete_results)) # check if done was logged exactly once self.assertEqual(len([r for r in logs1 if r.get("done")]), 1) def testNoDoneReceived(self): # repeat same test but without explicitly reporting done=True results1 = [ dict(timesteps_total=(i + 1), my_score=i**2) for i in range(5) ] logs1, trials = self.checkAndReturnConsistentLogs(results1) # check if the correct number of results were reported. self.assertEqual(len(logs1), len(results1)) def check_no_missing(reported_result, result): common_results = [reported_result[k] == result[k] for k in result] return all(common_results) # check that no result was dropped or modified complete_results1 = [ check_no_missing(log, result) for log, result in zip(logs1, results1) ] self.assertTrue(all(complete_results1)) def testDurableTrainable(self): class TestTrain(DurableTrainable): def _setup(self, config): self.state = {"hi": 1, "iter": 0} def _train(self): self.state["iter"] += 1 return {"timesteps_this_iter": 1, "done": True} def _save(self, path): return self.state def _restore(self, state): self.state = state sync_client = mock_storage_client() mock_get_client = "ray.tune.durable_trainable.get_cloud_sync_client" with patch(mock_get_client) as mock_get_cloud_sync_client: mock_get_cloud_sync_client.return_value = sync_client test_trainable = TestTrain(remote_checkpoint_dir=MOCK_REMOTE_DIR) checkpoint_path = test_trainable.save() test_trainable.train() test_trainable.state["hi"] = 2 test_trainable.restore(checkpoint_path) self.assertEqual(test_trainable.state["hi"], 1) self.addCleanup(shutil.rmtree, MOCK_REMOTE_DIR) def testCheckpointDict(self): class TestTrain(Trainable): def _setup(self, config): self.state = {"hi": 1} def _train(self): return {"timesteps_this_iter": 1, "done": True} def _save(self, path): return self.state def _restore(self, state): self.state = state test_trainable = TestTrain() result = test_trainable.save() test_trainable.state["hi"] = 2 test_trainable.restore(result) self.assertEqual(test_trainable.state["hi"], 1) trials = run_experiments({ "foo": { "run": TestTrain, "checkpoint_at_end": True } }) for trial in trials: self.assertEqual(trial.status, Trial.TERMINATED) self.assertTrue(trial.has_checkpoint()) def testMultipleCheckpoints(self): class TestTrain(Trainable): def _setup(self, config): self.state = {"hi": 1, "iter": 0} def _train(self): self.state["iter"] += 1 return {"timesteps_this_iter": 1, "done": True} def _save(self, path): return self.state def _restore(self, state): self.state = state test_trainable = TestTrain() checkpoint_1 = test_trainable.save() test_trainable.train() checkpoint_2 = test_trainable.save() self.assertNotEqual(checkpoint_1, checkpoint_2) test_trainable.restore(checkpoint_2) self.assertEqual(test_trainable.state["iter"], 1) test_trainable.restore(checkpoint_1) self.assertEqual(test_trainable.state["iter"], 0) trials = run_experiments({ "foo": { "run": TestTrain, "checkpoint_at_end": True } }) for trial in trials: self.assertEqual(trial.status, Trial.TERMINATED) self.assertTrue(trial.has_checkpoint()) def testIterationCounter(self): def train(config, reporter): for i in range(100): reporter(itr=i, timesteps_this_iter=1) register_trainable("exp", train) config = { "my_exp": { "run": "exp", "config": { "iterations": 100, }, "stop": { "timesteps_total": 100 }, } } [trial] = run_experiments(config) self.assertEqual(trial.status, Trial.TERMINATED) self.assertEqual(trial.last_result[TRAINING_ITERATION], 100) self.assertEqual(trial.last_result["itr"], 99) if __name__ == "__main__": import pytest import sys sys.exit(pytest.main(["-v", __file__]))
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/test_automl_searcher.py
Python
import random import unittest from ray.tune import register_trainable from ray.tune.automl import SearchSpace, DiscreteSpace, GridSearch class AutoMLSearcherTest(unittest.TestCase): def setUp(self): def dummy_train(config, reporter): reporter(timesteps_total=100, done=True) register_trainable("f1", dummy_train) def testExpandSearchSpace(self): exp = {"test-exp": {"run": "f1", "config": {"a": {"d": "dummy"}}}} space = SearchSpace([ DiscreteSpace("a.b.c", [1, 2]), DiscreteSpace("a.d", ["a", "b"]), ]) searcher = GridSearch(space, "reward") searcher.add_configurations(exp) trials = searcher.next_trials() self.assertEqual(len(trials), 4) self.assertTrue(trials[0].config["a"]["b"]["c"] in [1, 2]) self.assertTrue(trials[1].config["a"]["d"] in ["a", "b"]) def testSearchRound(self): exp = {"test-exp": {"run": "f1", "config": {"a": {"d": "dummy"}}}} space = SearchSpace([ DiscreteSpace("a.b.c", [1, 2]), DiscreteSpace("a.d", ["a", "b"]), ]) searcher = GridSearch(space, "reward") searcher.add_configurations(exp) trials = searcher.next_trials() self.assertEqual(len(searcher.next_trials()), 0) for trial in trials[1:]: searcher.on_trial_complete(trial.trial_id) searcher.on_trial_complete(trials[0].trial_id, error=True) self.assertTrue(searcher.is_finished()) def testBestTrial(self): exp = {"test-exp": {"run": "f1", "config": {"a": {"d": "dummy"}}}} space = SearchSpace([ DiscreteSpace("a.b.c", [1, 2]), DiscreteSpace("a.d", ["a", "b"]), ]) searcher = GridSearch(space, "reward") searcher.add_configurations(exp) trials = searcher.next_trials() self.assertEqual(len(searcher.next_trials()), 0) for i, trial in enumerate(trials): rewards = list(range(i, i + 10)) random.shuffle(rewards) for reward in rewards: searcher.on_trial_result(trial.trial_id, {"reward": reward}) best_trial = searcher.get_best_trial() self.assertEqual(best_trial, trials[-1]) self.assertEqual(best_trial.best_result["reward"], 3 + 10 - 1) if __name__ == "__main__": import pytest import sys sys.exit(pytest.main(["-v", __file__]))
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/test_checkpoint_manager.py
Python
# coding: utf-8 import random import sys import unittest from unittest.mock import patch from ray.tune.checkpoint_manager import Checkpoint, CheckpointManager, logger class CheckpointManagerTest(unittest.TestCase): @staticmethod def mock_result(i): return {"i": i} def checkpoint_manager(self, keep_checkpoints_num): return CheckpointManager( keep_checkpoints_num, "i", delete_fn=lambda c: None) def testOnCheckpointOrdered(self): """ Tests increasing priorities. Also tests that that the worst checkpoints are deleted when necessary. """ keep_checkpoints_num = 2 checkpoint_manager = self.checkpoint_manager(keep_checkpoints_num) checkpoints = [ Checkpoint(Checkpoint.PERSISTENT, {i}, self.mock_result(i)) for i in range(3) ] with patch.object(checkpoint_manager, "delete") as \ delete_mock: for j in range(3): checkpoint_manager.on_checkpoint(checkpoints[j]) expected_deletes = 0 if j != 2 else 1 self.assertEqual(delete_mock.call_count, expected_deletes, j) self.assertEqual(checkpoint_manager.newest_checkpoint, checkpoints[j]) best_checkpoints = checkpoint_manager.best_checkpoints() self.assertEqual(len(best_checkpoints), keep_checkpoints_num) self.assertIn(checkpoints[1], best_checkpoints) self.assertIn(checkpoints[2], best_checkpoints) def testOnCheckpointUnordered(self): """ Tests priorities that aren't inserted in ascending order. Also tests that the worst checkpoints are deleted when necessary. """ keep_checkpoints_num = 2 checkpoint_manager = self.checkpoint_manager(keep_checkpoints_num) checkpoints = [ Checkpoint(Checkpoint.PERSISTENT, {i}, self.mock_result(i)) for i in range(3, -1, -1) ] with patch.object(checkpoint_manager, "delete") as delete_mock: for j in range(0, len(checkpoints)): checkpoint_manager.on_checkpoint(checkpoints[j]) expected_deletes = 0 if j != 3 else 1 self.assertEqual(delete_mock.call_count, expected_deletes) self.assertEqual(checkpoint_manager.newest_checkpoint, checkpoints[j]) best_checkpoints = checkpoint_manager.best_checkpoints() self.assertEqual(len(best_checkpoints), keep_checkpoints_num) self.assertIn(checkpoints[0], best_checkpoints) self.assertIn(checkpoints[1], best_checkpoints) def testBestCheckpoints(self): """ Tests that the best checkpoints are tracked and ordered correctly. """ keep_checkpoints_num = 4 checkpoint_manager = self.checkpoint_manager(keep_checkpoints_num) checkpoints = [ Checkpoint(Checkpoint.MEMORY, i, self.mock_result(i)) for i in range(16) ] random.shuffle(checkpoints) for checkpoint in checkpoints: checkpoint_manager.on_checkpoint(checkpoint) best_checkpoints = checkpoint_manager.best_checkpoints() self.assertEqual(len(best_checkpoints), keep_checkpoints_num) for i in range(len(best_checkpoints)): self.assertEqual(best_checkpoints[i].value, i + 12) def testOnCheckpointUnavailableAttribute(self): """ Tests that an error is logged when the associated result of the checkpoint has no checkpoint score attribute. """ keep_checkpoints_num = 1 checkpoint_manager = self.checkpoint_manager(keep_checkpoints_num) no_attr_checkpoint = Checkpoint(Checkpoint.MEMORY, 0, {}) with patch.object(logger, "error") as log_error_mock: checkpoint_manager.on_checkpoint(no_attr_checkpoint) log_error_mock.assert_called_once() # The newest checkpoint should still be set despite this error. assert checkpoint_manager.newest_checkpoint == no_attr_checkpoint if __name__ == "__main__": import pytest sys.exit(pytest.main(["-v", __file__]))
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/test_cluster.py
Python
import inspect import json import time import os import pytest import shutil import sys from unittest.mock import MagicMock, patch import ray from ray import tune from ray.rllib import _register_all from ray.cluster_utils import Cluster from ray.test_utils import run_string_as_driver_nonblocking from ray.tune import register_trainable from ray.tune.experiment import Experiment from ray.tune.error import TuneError from ray.tune.ray_trial_executor import RayTrialExecutor from ray.tune.resources import Resources from ray.tune.suggest import BasicVariantGenerator from ray.tune.syncer import Syncer from ray.tune.trainable import TrainableUtil from ray.tune.trial import Trial from ray.tune.trial_runner import TrialRunner from ray.tune.utils.mock import (MockDurableTrainer, MockRemoteTrainer, MockNodeSyncer, mock_storage_client, MOCK_REMOTE_DIR) def _start_new_cluster(): cluster = Cluster( initialize_head=True, connect=True, head_node_args={ "num_cpus": 1, "_internal_config": json.dumps({ "num_heartbeats_timeout": 10 }) }) # Pytest doesn't play nicely with imports register_trainable("__fake_remote", MockRemoteTrainer) register_trainable("__fake_durable", MockDurableTrainer) _register_all() return cluster @pytest.fixture def start_connected_cluster(): # Start the Ray processes. cluster = _start_new_cluster() yield cluster # The code after the yield will run as teardown code. ray.shutdown() cluster.shutdown() @pytest.fixture def start_connected_emptyhead_cluster(): """Starts head with no resources.""" cluster = Cluster( initialize_head=True, connect=True, head_node_args={ "num_cpus": 0, "_internal_config": json.dumps({ "num_heartbeats_timeout": 10 }) }) # Pytest doesn't play nicely with imports _register_all() register_trainable("__fake_remote", MockRemoteTrainer) register_trainable("__fake_durable", MockDurableTrainer) yield cluster # The code after the yield will run as teardown code. ray.shutdown() cluster.shutdown() def test_counting_resources(start_connected_cluster): """Tests that Tune accounting is consistent with actual cluster.""" cluster = start_connected_cluster nodes = [] assert ray.cluster_resources()["CPU"] == 1 runner = TrialRunner(BasicVariantGenerator()) kwargs = {"stopping_criterion": {"training_iteration": 10}} trials = [Trial("__fake", **kwargs), Trial("__fake", **kwargs)] for t in trials: runner.add_trial(t) runner.step() # run 1 nodes += [cluster.add_node(num_cpus=1)] cluster.wait_for_nodes() assert ray.cluster_resources()["CPU"] == 2 cluster.remove_node(nodes.pop()) cluster.wait_for_nodes() assert ray.cluster_resources()["CPU"] == 1 runner.step() # run 2 assert sum(t.status == Trial.RUNNING for t in runner.get_trials()) == 1 for i in range(5): nodes += [cluster.add_node(num_cpus=1)] cluster.wait_for_nodes() assert ray.cluster_resources()["CPU"] == 6 runner.step() # 1 result assert sum(t.status == Trial.RUNNING for t in runner.get_trials()) == 2 def test_trial_processed_after_node_failure(start_connected_emptyhead_cluster): """Tests that Tune processes a trial as failed if its node died.""" cluster = start_connected_emptyhead_cluster node = cluster.add_node(num_cpus=1) cluster.wait_for_nodes() runner = TrialRunner(BasicVariantGenerator()) mock_process_failure = MagicMock(side_effect=runner._process_trial_failure) runner._process_trial_failure = mock_process_failure runner.add_trial(Trial("__fake")) runner.step() runner.step() assert not mock_process_failure.called cluster.remove_node(node) runner.step() if not mock_process_failure.called: runner.step() assert mock_process_failure.called def test_remove_node_before_result(start_connected_emptyhead_cluster): """Tune continues when node is removed before trial returns.""" cluster = start_connected_emptyhead_cluster node = cluster.add_node(num_cpus=1) cluster.wait_for_nodes() runner = TrialRunner(BasicVariantGenerator()) kwargs = { "stopping_criterion": { "training_iteration": 3 }, "checkpoint_freq": 2, "max_failures": 2 } trial = Trial("__fake", **kwargs) runner.add_trial(trial) runner.step() # run 1 assert trial.status == Trial.RUNNING cluster.remove_node(node) cluster.add_node(num_cpus=1) cluster.wait_for_nodes() assert ray.cluster_resources()["CPU"] == 1 for i in range(3): runner.step() assert trial.status == Trial.TERMINATED with pytest.raises(TuneError): runner.step() def test_queue_trials(start_connected_emptyhead_cluster): """Tests explicit oversubscription for autoscaling. Tune oversubscribes a trial when `queue_trials=True`, but does not block other trials from running. """ cluster = start_connected_emptyhead_cluster runner = TrialRunner() def create_trial(cpu, gpu=0): kwargs = { "resources": Resources(cpu=cpu, gpu=gpu), "stopping_criterion": { "training_iteration": 3 } } return Trial("__fake", **kwargs) runner.add_trial(create_trial(cpu=1)) with pytest.raises(TuneError): runner.step() # run 1 del runner executor = RayTrialExecutor(queue_trials=True) runner = TrialRunner(trial_executor=executor) cluster.add_node(num_cpus=2) cluster.wait_for_nodes() cpu_only = create_trial(cpu=1) runner.add_trial(cpu_only) runner.step() # add cpu_only trial gpu_trial = create_trial(cpu=1, gpu=1) runner.add_trial(gpu_trial) runner.step() # queue gpu_trial # This tests that the cpu_only trial should bypass the queued trial. for i in range(3): runner.step() assert cpu_only.status == Trial.TERMINATED assert gpu_trial.status == Trial.RUNNING # Scale up cluster.add_node(num_cpus=1, num_gpus=1) cluster.wait_for_nodes() for i in range(3): runner.step() assert gpu_trial.status == Trial.TERMINATED @pytest.mark.parametrize("trainable_id", ["__fake", "__fake_durable"]) def test_trial_migration(start_connected_emptyhead_cluster, trainable_id): """Removing a node while cluster has space should migrate trial. The trial state should also be consistent with the checkpoint. """ cluster = start_connected_emptyhead_cluster node = cluster.add_node(num_cpus=1) cluster.wait_for_nodes() runner = TrialRunner(BasicVariantGenerator()) kwargs = { "stopping_criterion": { "training_iteration": 4 }, "checkpoint_freq": 2, "max_failures": 2, "remote_checkpoint_dir": MOCK_REMOTE_DIR, "sync_to_driver_fn": trainable_id == "__fake", } # Test recovery of trial that hasn't been checkpointed t = Trial(trainable_id, **kwargs) runner.add_trial(t) runner.step() # start runner.step() # 1 result assert t.last_result node2 = cluster.add_node(num_cpus=1) cluster.remove_node(node) cluster.wait_for_nodes() runner.step() # Recovery step # TODO(rliaw): This assertion is not critical but will not pass # because checkpoint handling is messy and should be refactored # rather than hotfixed. # assert t.last_result is None, "Trial result not restored correctly." for i in range(4): runner.step() assert t.status == Trial.TERMINATED # Test recovery of trial that has been checkpointed t2 = Trial(trainable_id, **kwargs) runner.add_trial(t2) runner.step() # start runner.step() # 1 result runner.step() # 2 result and checkpoint assert t2.has_checkpoint() node3 = cluster.add_node(num_cpus=1) cluster.remove_node(node2) cluster.wait_for_nodes() runner.step() # 3 result + start and fail 4 result runner.step() # Recovery step runner.step() # Process recovery runner.step() # result if t2.status != Trial.TERMINATED: runner.step() assert t2.status == Trial.TERMINATED, runner.debug_string() # Test recovery of trial that won't be checkpointed kwargs = { "stopping_criterion": { "training_iteration": 3 }, "remote_checkpoint_dir": MOCK_REMOTE_DIR, "sync_to_driver_fn": trainable_id == "__fake", } t3 = Trial(trainable_id, **kwargs) runner.add_trial(t3) runner.step() # start runner.step() # 1 result cluster.add_node(num_cpus=1) cluster.remove_node(node3) cluster.wait_for_nodes() runner.step() # Error handling step if t3.status != Trial.ERROR: runner.step() assert t3.status == Trial.ERROR, runner.debug_string() with pytest.raises(TuneError): runner.step() @pytest.mark.parametrize("trainable_id", ["__fake", "__fake_durable"]) def test_trial_requeue(start_connected_emptyhead_cluster, trainable_id): """Removing a node in full cluster causes Trial to be requeued.""" cluster = start_connected_emptyhead_cluster node = cluster.add_node(num_cpus=1) cluster.wait_for_nodes() runner = TrialRunner(BasicVariantGenerator()) kwargs = { "stopping_criterion": { "training_iteration": 5 }, "checkpoint_freq": 1, "max_failures": 1, "remote_checkpoint_dir": MOCK_REMOTE_DIR, "sync_to_driver_fn": trainable_id == "__fake", } trials = [Trial(trainable_id, **kwargs), Trial(trainable_id, **kwargs)] for t in trials: runner.add_trial(t) runner.step() # start runner.step() # 1 result cluster.remove_node(node) cluster.wait_for_nodes() runner.step() assert all(t.status == Trial.PENDING for t in trials) with pytest.raises(TuneError): runner.step() @pytest.mark.parametrize("trainable_id", ["__fake_remote", "__fake_durable"]) def test_migration_checkpoint_removal(start_connected_emptyhead_cluster, trainable_id): """Test checks that trial restarts if checkpoint is lost w/ node fail.""" cluster = start_connected_emptyhead_cluster node = cluster.add_node(num_cpus=1) cluster.wait_for_nodes() runner = TrialRunner(BasicVariantGenerator()) kwargs = { "stopping_criterion": { "training_iteration": 4 }, "checkpoint_freq": 2, "max_failures": 2, "remote_checkpoint_dir": MOCK_REMOTE_DIR, "sync_to_driver_fn": trainable_id == "__fake_remote", } # The following patches only affect __fake_remote. find_checkpoint_dir = TrainableUtil.find_checkpoint_dir with patch("ray.tune.logger.get_node_syncer") as mock_get_node_syncer: trainable_util = "ray.tune.ray_trial_executor.TrainableUtil" with patch(trainable_util + ".find_checkpoint_dir") as mock_find_dir: def mock_get_syncer_fn(local_dir, remote_dir, sync_function): client = mock_storage_client() return MockNodeSyncer(local_dir, remote_dir, client) mock_get_node_syncer.side_effect = mock_get_syncer_fn def mock_find_dir_fn(checkpoint_path): """Converts back to local path first.""" checkpoint_path = checkpoint_path[len(MOCK_REMOTE_DIR):] checkpoint_path = os.path.join("/", checkpoint_path) return find_checkpoint_dir(checkpoint_path) # __fake_remote trainables save to a separate "remote" directory. # TrainableUtil will not check this path unless we mock it. mock_find_dir.side_effect = mock_find_dir_fn # Test recovery of trial that has been checkpointed t1 = Trial(trainable_id, **kwargs) runner.add_trial(t1) runner.step() # start runner.step() # 1 result runner.step() # 2 result and checkpoint assert t1.has_checkpoint() cluster.add_node(num_cpus=1) cluster.remove_node(node) cluster.wait_for_nodes() shutil.rmtree(os.path.dirname(t1.checkpoint.value)) runner.step() # collect result 3, kick off + fail result 4 runner.step() # Recovery step runner.step() # Process Recovery + step 4 for i in range(3): if t1.status != Trial.TERMINATED: runner.step() assert t1.status == Trial.TERMINATED, runner.debug_string() @pytest.mark.parametrize("trainable_id", ["__fake", "__fake_durable"]) def test_cluster_down_simple(start_connected_cluster, tmpdir, trainable_id): """Tests that TrialRunner save/restore works on cluster shutdown.""" cluster = start_connected_cluster cluster.add_node(num_cpus=1) cluster.wait_for_nodes() dirpath = str(tmpdir) runner = TrialRunner(local_checkpoint_dir=dirpath, checkpoint_period=0) kwargs = { "stopping_criterion": { "training_iteration": 2 }, "checkpoint_freq": 1, "max_failures": 1, "remote_checkpoint_dir": MOCK_REMOTE_DIR, "sync_to_driver_fn": trainable_id == "__fake", } trials = [Trial(trainable_id, **kwargs), Trial(trainable_id, **kwargs)] for t in trials: runner.add_trial(t) runner.step() # start runner.step() # start2 runner.step() # step assert all(t.status == Trial.RUNNING for t in runner.get_trials()) runner.checkpoint() ray.shutdown() cluster.shutdown() cluster = _start_new_cluster() runner = TrialRunner(resume="LOCAL", local_checkpoint_dir=dirpath) runner.step() # start runner.step() # process restore runner.step() # start2 for i in range(3): runner.step() with pytest.raises(TuneError): runner.step() assert all(t.status == Trial.TERMINATED for t in runner.get_trials()) ray.shutdown() cluster.shutdown() @pytest.mark.parametrize("trainable_id", ["__fake", "__fake_durable"]) def test_cluster_down_full(start_connected_cluster, tmpdir, trainable_id): """Tests that run_experiment restoring works on cluster shutdown.""" cluster = start_connected_cluster dirpath = str(tmpdir) use_default_sync = trainable_id == "__fake" from ray.tune.result import DEFAULT_RESULTS_DIR local_dir = DEFAULT_RESULTS_DIR upload_dir = None if use_default_sync else MOCK_REMOTE_DIR base_dict = dict( run=trainable_id, stop=dict(training_iteration=3), local_dir=local_dir, upload_dir=upload_dir, sync_to_driver=use_default_sync, ) exp1_args = base_dict exp2_args = dict(base_dict.items(), local_dir=dirpath, checkpoint_freq=1) exp3_args = dict(base_dict.items(), config=dict(mock_error=True)) exp4_args = dict( base_dict.items(), config=dict(mock_error=True), checkpoint_freq=1) all_experiments = { "exp1": exp1_args, "exp2": exp2_args, "exp3": exp3_args, "exp4": exp4_args } mock_get_client = "ray.tune.trial_runner.get_cloud_syncer" with patch(mock_get_client) as mock_get_cloud_syncer: mock_syncer = Syncer(local_dir, upload_dir, mock_storage_client()) mock_get_cloud_syncer.return_value = mock_syncer tune.run_experiments(all_experiments, raise_on_failed_trial=False) ray.shutdown() cluster.shutdown() cluster = _start_new_cluster() trials = tune.run_experiments( all_experiments, resume=True, raise_on_failed_trial=False) assert len(trials) == 4 assert all(t.status in [Trial.TERMINATED, Trial.ERROR] for t in trials) ray.shutdown() cluster.shutdown() @pytest.mark.skip(reason="Not very consistent.") def test_cluster_rllib_restore(start_connected_cluster, tmpdir): cluster = start_connected_cluster dirpath = str(tmpdir) script = """ import time import ray from ray import tune ray.init(address="{address}") tune.run( "PG", name="experiment", config=dict(env="CartPole-v1"), stop=dict(training_iteration=10), local_dir="{checkpoint_dir}", checkpoint_freq=1, max_failures=1, dict(experiment=kwargs), raise_on_failed_trial=False) """.format( address=cluster.address, checkpoint_dir=dirpath) run_string_as_driver_nonblocking(script) # Wait until the right checkpoint is saved. # The trainable returns every 0.5 seconds, so this should not miss # the checkpoint. local_checkpoint_dir = os.path.join(dirpath, "experiment") for i in range(100): if TrialRunner.checkpoint_exists(local_checkpoint_dir): # Inspect the internal trialrunner runner = TrialRunner( resume="LOCAL", local_checkpoint_dir=local_checkpoint_dir) trials = runner.get_trials() last_res = trials[0].last_result if last_res and last_res.get("training_iteration"): break time.sleep(0.3) if not TrialRunner.checkpoint_exists(local_checkpoint_dir): raise RuntimeError("Checkpoint file didn't appear.") ray.shutdown() cluster.shutdown() cluster = _start_new_cluster() cluster.wait_for_nodes() # Restore properly from checkpoint trials2 = tune.run_experiments( { "experiment": { "run": "PG", "checkpoint_freq": 1, "local_dir": dirpath } }, resume=True) assert all(t.status == Trial.TERMINATED for t in trials2) ray.shutdown() cluster.shutdown() def test_cluster_interrupt(start_connected_cluster, tmpdir): """Tests run_experiment on cluster shutdown with actual interrupt. This is an end-to-end test. """ cluster = start_connected_cluster dirpath = str(tmpdir) # Needs to be in scope for pytest class _Mock(tune.Trainable): """Finishes on the 4th iteration.""" def _setup(self, config): self.state = {"hi": 0} def _train(self): self.state["hi"] += 1 time.sleep(0.5) return {"done": self.state["hi"] >= 4} def _save(self, path): return self.state def _restore(self, state): self.state = state # Removes indent from class. reformatted = "\n".join(line[4:] if len(line) else line for line in inspect.getsource(_Mock).split("\n")) script = """ import time import ray from ray import tune ray.init(address="{address}") {fail_class_code} tune.run( {fail_class}, name="experiment", stop=dict(training_iteration=5), local_dir="{checkpoint_dir}", checkpoint_freq=1, global_checkpoint_period=0, max_failures=1, raise_on_failed_trial=False) """.format( address=cluster.address, checkpoint_dir=dirpath, fail_class_code=reformatted, fail_class=_Mock.__name__) run_string_as_driver_nonblocking(script) # Wait until the right checkpoint is saved. # The trainable returns every 0.5 seconds, so this should not miss # the checkpoint. local_checkpoint_dir = os.path.join(dirpath, "experiment") for i in range(50): if TrialRunner.checkpoint_exists(local_checkpoint_dir): # Inspect the internal trialrunner runner = TrialRunner( resume="LOCAL", local_checkpoint_dir=local_checkpoint_dir) trials = runner.get_trials() last_res = trials[0].last_result if last_res and last_res.get("training_iteration") == 3: break time.sleep(0.2) if not TrialRunner.checkpoint_exists(local_checkpoint_dir): raise RuntimeError("Checkpoint file didn't appear.") ray.shutdown() cluster.shutdown() cluster = _start_new_cluster() Experiment.register_if_needed(_Mock) # Inspect the internal trialrunner runner = TrialRunner( resume="LOCAL", local_checkpoint_dir=local_checkpoint_dir) trials = runner.get_trials() assert trials[0].last_result["training_iteration"] == 3 assert trials[0].status == Trial.PENDING # Restore properly from checkpoint trials2 = tune.run_experiments( { "experiment": { "run": _Mock, "local_dir": dirpath, "checkpoint_freq": 1 } }, resume=True, raise_on_failed_trial=False) assert all(t.status == Trial.TERMINATED for t in trials2) assert {t.trial_id for t in trials2} == {t.trial_id for t in trials} ray.shutdown() cluster.shutdown() if __name__ == "__main__": import pytest sys.exit(pytest.main(["-v", __file__]))
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/test_commands.py
Python
import click import os import pytest import subprocess import sys import time try: from cStringIO import StringIO except ImportError: from io import StringIO import ray from ray import tune from ray.rllib import _register_all from ray.tune import commands from ray.tune.result import CONFIG_PREFIX class Capturing: def __enter__(self): self._stdout = sys.stdout sys.stdout = self._stringio = StringIO() self.captured = [] return self def __exit__(self, *args): self.captured.extend(self._stringio.getvalue().splitlines()) del self._stringio # free up some memory sys.stdout = self._stdout @pytest.fixture def start_ray(): ray.init(log_to_driver=False, local_mode=True) _register_all() yield ray.shutdown() def test_time(start_ray, tmpdir): experiment_name = "test_time" experiment_path = os.path.join(str(tmpdir), experiment_name) num_samples = 2 tune.run_experiments({ experiment_name: { "run": "__fake", "stop": { "training_iteration": 1 }, "num_samples": num_samples, "local_dir": str(tmpdir) } }) times = [] for i in range(5): start = time.time() subprocess.check_call(["tune", "ls", experiment_path]) times += [time.time() - start] assert sum(times) / len(times) < 2.0, "CLI is taking too long!" def test_ls(start_ray, tmpdir): """This test captures output of list_trials.""" experiment_name = "test_ls" experiment_path = os.path.join(str(tmpdir), experiment_name) num_samples = 3 tune.run( "__fake", name=experiment_name, stop={"training_iteration": 1}, num_samples=num_samples, local_dir=str(tmpdir)) columns = ["episode_reward_mean", "training_iteration", "trial_id"] limit = 2 with Capturing() as output: commands.list_trials(experiment_path, info_keys=columns, limit=limit) lines = output.captured assert all(col in lines[1] for col in columns) assert lines[1].count("|") == len(columns) + 1 assert len(lines) == 3 + limit + 1 with Capturing() as output: commands.list_trials( experiment_path, sort=["trial_id"], info_keys=("trial_id", "training_iteration"), filter_op="training_iteration == 1") lines = output.captured assert len(lines) == 3 + num_samples + 1 with pytest.raises(click.ClickException): commands.list_trials( experiment_path, sort=["trial_id"], info_keys=("training_iteration", )) with pytest.raises(click.ClickException): commands.list_trials(experiment_path, info_keys=("asdf", )) def test_ls_with_cfg(start_ray, tmpdir): experiment_name = "test_ls_with_cfg" experiment_path = os.path.join(str(tmpdir), experiment_name) tune.run( "__fake", name=experiment_name, stop={"training_iteration": 1}, config={"test_variable": tune.grid_search(list(range(5)))}, local_dir=str(tmpdir)) columns = [CONFIG_PREFIX + "test_variable", "trial_id"] limit = 4 with Capturing() as output: commands.list_trials(experiment_path, info_keys=columns, limit=limit) lines = output.captured assert all(col in lines[1] for col in columns) assert lines[1].count("|") == len(columns) + 1 assert len(lines) == 3 + limit + 1 def test_lsx(start_ray, tmpdir): """This test captures output of list_experiments.""" project_path = str(tmpdir) num_experiments = 3 for i in range(num_experiments): experiment_name = "test_lsx{}".format(i) tune.run( "__fake", name=experiment_name, stop={"training_iteration": 1}, num_samples=1, local_dir=project_path) limit = 2 with Capturing() as output: commands.list_experiments( project_path, info_keys=("total_trials", ), limit=limit) lines = output.captured assert "total_trials" in lines[1] assert lines[1].count("|") == 2 assert len(lines) == 3 + limit + 1 with Capturing() as output: commands.list_experiments( project_path, sort=["total_trials"], info_keys=("total_trials", ), filter_op="total_trials == 1") lines = output.captured assert sum("1" in line for line in lines) >= num_experiments assert len(lines) == 3 + num_experiments + 1 if __name__ == "__main__": # Make click happy in bazel. os.environ["LC_ALL"] = "en_US.UTF-8" os.environ["LANG"] = "en_US.UTF-8" sys.exit(pytest.main([__file__]))
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/test_dependency.py
Python
#!/usr/bin/env python import sys import ray from ray.tune import register_trainable, run_experiments def f(config, reporter): reporter(timesteps_total=1) if __name__ == "__main__": ray.init() register_trainable("my_class", f) run_experiments({ "test": { "run": "my_class", "stop": { "training_iteration": 1 } } }) assert "ray.rllib" not in sys.modules, "RLlib should not be imported"
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/test_experiment.py
Python
import unittest import ray from ray.rllib import _register_all from ray.tune import register_trainable from ray.tune.experiment import Experiment, convert_to_experiment_list from ray.tune.error import TuneError class ExperimentTest(unittest.TestCase): def tearDown(self): ray.shutdown() _register_all() # re-register the evicted objects def setUp(self): def train(config, reporter): for i in range(100): reporter(timesteps_total=i) register_trainable("f1", train) def testConvertExperimentFromExperiment(self): exp1 = Experiment(**{ "name": "foo", "run": "f1", "config": { "script_min_iter_time_s": 0 } }) result = convert_to_experiment_list(exp1) self.assertEqual(len(result), 1) self.assertEqual(type(result), list) def testConvertExperimentNone(self): result = convert_to_experiment_list(None) self.assertEqual(len(result), 0) self.assertEqual(type(result), list) def testConvertExperimentList(self): exp1 = Experiment(**{ "name": "foo", "run": "f1", "config": { "script_min_iter_time_s": 0 } }) result = convert_to_experiment_list([exp1, exp1]) self.assertEqual(len(result), 2) self.assertEqual(type(result), list) def testConvertExperimentJSON(self): experiment = { "name": { "run": "f1", "config": { "script_min_iter_time_s": 0 } }, "named": { "run": "f1", "config": { "script_min_iter_time_s": 0 } } } result = convert_to_experiment_list(experiment) self.assertEqual(len(result), 2) self.assertEqual(type(result), list) def testConvertExperimentIncorrect(self): self.assertRaises(TuneError, lambda: convert_to_experiment_list("hi")) if __name__ == "__main__": import pytest import sys sys.exit(pytest.main(["-v", __file__]))
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/test_experiment_analysis.py
Python
import unittest import shutil import tempfile import random import os import pandas as pd import ray from ray.tune import run, sample_from from ray.tune.examples.async_hyperband_example import MyTrainableClass class ExperimentAnalysisSuite(unittest.TestCase): def setUp(self): ray.init(local_mode=False) self.test_dir = tempfile.mkdtemp() self.test_name = "analysis_exp" self.num_samples = 10 self.metric = "episode_reward_mean" self.test_path = os.path.join(self.test_dir, self.test_name) self.run_test_exp() def tearDown(self): shutil.rmtree(self.test_dir, ignore_errors=True) ray.shutdown() def run_test_exp(self): self.ea = run( MyTrainableClass, name=self.test_name, local_dir=self.test_dir, stop={"training_iteration": 1}, checkpoint_freq=1, num_samples=self.num_samples, config={ "width": sample_from( lambda spec: 10 + int(90 * random.random())), "height": sample_from(lambda spec: int(100 * random.random())), }) def testDataframe(self): df = self.ea.dataframe() self.assertTrue(isinstance(df, pd.DataFrame)) self.assertEquals(df.shape[0], self.num_samples) def testStats(self): assert self.ea.stats() assert self.ea.runner_data() def testTrialDataframe(self): checkpoints = self.ea._checkpoints idx = random.randint(0, len(checkpoints) - 1) trial_df = self.ea.trial_dataframes[checkpoints[idx]["logdir"]] self.assertTrue(isinstance(trial_df, pd.DataFrame)) self.assertEqual(trial_df.shape[0], 1) def testBestConfig(self): best_config = self.ea.get_best_config(self.metric) self.assertTrue(isinstance(best_config, dict)) self.assertTrue("width" in best_config) self.assertTrue("height" in best_config) def testBestLogdir(self): logdir = self.ea.get_best_logdir(self.metric) self.assertTrue(logdir.startswith(self.test_path)) logdir2 = self.ea.get_best_logdir(self.metric, mode="min") self.assertTrue(logdir2.startswith(self.test_path)) self.assertNotEquals(logdir, logdir2) def testGetTrialCheckpointsPathsByTrial(self): best_trial = self.ea.get_best_trial(self.metric) checkpoints_metrics = self.ea.get_trial_checkpoints_paths(best_trial) logdir = self.ea.get_best_logdir(self.metric) expected_path = os.path.join(logdir, "checkpoint_1", "checkpoint") assert checkpoints_metrics[0][0] == expected_path assert checkpoints_metrics[0][1] == 1 def testGetTrialCheckpointsPathsByPath(self): logdir = self.ea.get_best_logdir(self.metric) checkpoints_metrics = self.ea.get_trial_checkpoints_paths(logdir) expected_path = os.path.join(logdir, "checkpoint_1/", "checkpoint") assert checkpoints_metrics[0][0] == expected_path assert checkpoints_metrics[0][1] == 1 def testGetTrialCheckpointsPathsWithMetricByTrial(self): best_trial = self.ea.get_best_trial(self.metric) paths = self.ea.get_trial_checkpoints_paths(best_trial, self.metric) logdir = self.ea.get_best_logdir(self.metric) expected_path = os.path.join(logdir, "checkpoint_1", "checkpoint") assert paths[0][0] == expected_path assert paths[0][1] == best_trial.metric_analysis[self.metric]["last"] def testGetTrialCheckpointsPathsWithMetricByPath(self): best_trial = self.ea.get_best_trial(self.metric) logdir = self.ea.get_best_logdir(self.metric) paths = self.ea.get_trial_checkpoints_paths(best_trial, self.metric) expected_path = os.path.join(logdir, "checkpoint_1", "checkpoint") assert paths[0][0] == expected_path assert paths[0][1] == best_trial.metric_analysis[self.metric]["last"] def testAllDataframes(self): dataframes = self.ea.trial_dataframes self.assertTrue(len(dataframes) == self.num_samples) self.assertTrue(isinstance(dataframes, dict)) for df in dataframes.values(): self.assertEqual(df.training_iteration.max(), 1) def testIgnoreOtherExperiment(self): analysis = run( MyTrainableClass, name="test_example", local_dir=self.test_dir, return_trials=False, stop={"training_iteration": 1}, num_samples=1, config={ "width": sample_from( lambda spec: 10 + int(90 * random.random())), "height": sample_from(lambda spec: int(100 * random.random())), }) df = analysis.dataframe() self.assertEquals(df.shape[0], 1) if __name__ == "__main__": import pytest import sys sys.exit(pytest.main(["-v", __file__]))
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/test_experiment_analysis_mem.py
Python
import unittest import shutil import tempfile import random import pandas as pd import ray from ray.tune import run, Trainable, sample_from, Analysis, grid_search from ray.tune.examples.async_hyperband_example import MyTrainableClass class ExperimentAnalysisInMemorySuite(unittest.TestCase): def setUp(self): class MockTrainable(Trainable): def _setup(self, config): self.id = config["id"] self.idx = 0 self.scores_dict = { 0: [5, 0], 1: [4, 1], 2: [2, 8], 3: [9, 6], 4: [7, 3] } def _train(self): val = self.scores_dict[self.id][self.idx] self.idx += 1 return {"score": val} def _save(self, checkpoint_dir): pass def _restore(self, checkpoint_path): pass self.MockTrainable = MockTrainable ray.init(local_mode=False, num_cpus=1) def tearDown(self): shutil.rmtree(self.test_dir, ignore_errors=True) ray.shutdown() def testCompareTrials(self): self.test_dir = tempfile.mkdtemp() scores_all = [5, 4, 2, 9, 7, 0, 1, 8, 6, 3] scores_last = scores_all[5:] ea = run( self.MockTrainable, name="analysis_exp", local_dir=self.test_dir, stop={"training_iteration": 2}, num_samples=1, config={"id": grid_search(list(range(5)))}) max_all = ea.get_best_trial("score", "max").metric_analysis["score"]["max"] min_all = ea.get_best_trial("score", "min").metric_analysis["score"]["min"] max_last = ea.get_best_trial("score", "max", "last").metric_analysis["score"]["last"] self.assertEqual(max_all, max(scores_all)) self.assertEqual(min_all, min(scores_all)) self.assertEqual(max_last, max(scores_last)) self.assertNotEqual(max_last, max(scores_all)) class AnalysisSuite(unittest.TestCase): def setUp(self): ray.init(local_mode=True) self.test_dir = tempfile.mkdtemp() self.num_samples = 10 self.metric = "episode_reward_mean" self.run_test_exp(test_name="analysis_exp1") self.run_test_exp(test_name="analysis_exp2") def run_test_exp(self, test_name=None): run(MyTrainableClass, name=test_name, local_dir=self.test_dir, return_trials=False, stop={"training_iteration": 1}, num_samples=self.num_samples, config={ "width": sample_from( lambda spec: 10 + int(90 * random.random())), "height": sample_from(lambda spec: int(100 * random.random())), }) def tearDown(self): shutil.rmtree(self.test_dir, ignore_errors=True) ray.shutdown() def testDataframe(self): analysis = Analysis(self.test_dir) df = analysis.dataframe() self.assertTrue(isinstance(df, pd.DataFrame)) self.assertEquals(df.shape[0], self.num_samples * 2) def testBestLogdir(self): analysis = Analysis(self.test_dir) logdir = analysis.get_best_logdir(self.metric) self.assertTrue(logdir.startswith(self.test_dir)) logdir2 = analysis.get_best_logdir(self.metric, mode="min") self.assertTrue(logdir2.startswith(self.test_dir)) self.assertNotEquals(logdir, logdir2) def testBestConfigIsLogdir(self): analysis = Analysis(self.test_dir) for metric, mode in [(self.metric, "min"), (self.metric, "max")]: logdir = analysis.get_best_logdir(metric, mode=mode) best_config = analysis.get_best_config(metric, mode=mode) self.assertEquals(analysis.get_all_configs()[logdir], best_config) if __name__ == "__main__": import pytest import sys sys.exit(pytest.main(["-v", __file__]))
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/test_logger.py
Python
from collections import namedtuple import unittest import tempfile import shutil from ray.tune.logger import tf2_compat_logger, JsonLogger, CSVLogger, TBXLogger Trial = namedtuple("MockTrial", ["evaluated_params", "trial_id"]) def result(t, rew): return dict( time_total_s=t, episode_reward_mean=rew, mean_accuracy=rew * 2, training_iteration=int(t)) class LoggerSuite(unittest.TestCase): """Test built-in loggers.""" def setUp(self): self.test_dir = tempfile.mkdtemp() def tearDown(self): shutil.rmtree(self.test_dir, ignore_errors=True) def testTensorBoardLogger(self): config = {"a": 2, "b": 5} t = Trial(evaluated_params=config, trial_id=5342) logger = tf2_compat_logger( config=config, logdir=self.test_dir, trial=t) logger.on_result(result(2, 4)) logger.on_result(result(2, 4)) logger.close() def testCSV(self): config = {"a": 2, "b": 5} t = Trial(evaluated_params=config, trial_id="csv") logger = CSVLogger(config=config, logdir=self.test_dir, trial=t) logger.on_result(result(2, 4)) logger.on_result(result(2, 4)) logger.close() def testJSON(self): config = {"a": 2, "b": 5} t = Trial(evaluated_params=config, trial_id="json") logger = JsonLogger(config=config, logdir=self.test_dir, trial=t) logger.on_result(result(2, 4)) logger.on_result(result(2, 4)) logger.close() def testTBX(self): config = {"a": 2, "b": 5} t = Trial(evaluated_params=config, trial_id="tbx") logger = TBXLogger(config=config, logdir=self.test_dir, trial=t) logger.on_result(result(2, 4)) logger.on_result(result(2, 4)) logger.close() if __name__ == "__main__": import pytest import sys sys.exit(pytest.main(["-v", __file__]))
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/test_progress_reporter.py
Python
import collections import time import unittest from unittest.mock import MagicMock from ray.tune.trial import Trial from ray.tune.progress_reporter import _fair_filter_trials class ProgressReporterTest(unittest.TestCase): def mock_trial(self, status, start_time): mock = MagicMock() mock.status = status mock.start_time = start_time return mock def testFairFilterTrials(self): """Tests that trials are represented fairly.""" trials_by_state = collections.defaultdict(list) # States for which trials are under and overrepresented states_under = (Trial.PAUSED, Trial.ERROR) states_over = (Trial.PENDING, Trial.RUNNING, Trial.TERMINATED) max_trials = 13 num_trials_under = 2 # num of trials for each underrepresented state num_trials_over = 10 # num of trials for each overrepresented state for state in states_under: for _ in range(num_trials_under): trials_by_state[state].append( self.mock_trial(state, time.time())) for state in states_over: for _ in range(num_trials_over): trials_by_state[state].append( self.mock_trial(state, time.time())) filtered_trials_by_state = _fair_filter_trials( trials_by_state, max_trials=max_trials) for state in trials_by_state: if state in states_under: expected_num_trials = num_trials_under else: expected_num_trials = (max_trials - num_trials_under * len(states_under)) / len(states_over) state_trials = filtered_trials_by_state[state] self.assertEqual(len(state_trials), expected_num_trials) # Make sure trials are sorted newest-first within state. for i in range(len(state_trials) - 1): self.assertGreaterEqual(state_trials[i].start_time, state_trials[i + 1].start_time)
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/test_ray_trial_executor.py
Python
# coding: utf-8 import json import unittest import ray from ray.rllib import _register_all from ray.tune import Trainable from ray.tune.ray_trial_executor import RayTrialExecutor from ray.tune.registry import _global_registry, TRAINABLE_CLASS from ray.tune.suggest import BasicVariantGenerator from ray.tune.trial import Trial, Checkpoint from ray.tune.resources import Resources from ray.cluster_utils import Cluster class RayTrialExecutorTest(unittest.TestCase): def setUp(self): self.trial_executor = RayTrialExecutor(queue_trials=False) ray.init() _register_all() # Needed for flaky tests def tearDown(self): ray.shutdown() _register_all() # re-register the evicted objects def testStartStop(self): trial = Trial("__fake") self.trial_executor.start_trial(trial) running = self.trial_executor.get_running_trials() self.assertEqual(1, len(running)) self.trial_executor.stop_trial(trial) def testSaveRestore(self): trial = Trial("__fake") self.trial_executor.start_trial(trial) self.assertEqual(Trial.RUNNING, trial.status) self.trial_executor.save(trial, Checkpoint.PERSISTENT) self.trial_executor.restore(trial) self.trial_executor.stop_trial(trial) self.assertEqual(Trial.TERMINATED, trial.status) def testPauseResume(self): """Tests that pausing works for trials in flight.""" trial = Trial("__fake") self.trial_executor.start_trial(trial) self.assertEqual(Trial.RUNNING, trial.status) self.trial_executor.pause_trial(trial) self.assertEqual(Trial.PAUSED, trial.status) self.trial_executor.start_trial(trial) self.assertEqual(Trial.RUNNING, trial.status) self.trial_executor.stop_trial(trial) self.assertEqual(Trial.TERMINATED, trial.status) def testStartFailure(self): _global_registry.register(TRAINABLE_CLASS, "asdf", None) trial = Trial("asdf", resources=Resources(1, 0)) self.trial_executor.start_trial(trial) self.assertEqual(Trial.ERROR, trial.status) def testPauseResume2(self): """Tests that pausing works for trials being processed.""" trial = Trial("__fake") self.trial_executor.start_trial(trial) self.assertEqual(Trial.RUNNING, trial.status) self.trial_executor.fetch_result(trial) self.trial_executor.pause_trial(trial) self.assertEqual(Trial.PAUSED, trial.status) self.trial_executor.start_trial(trial) self.assertEqual(Trial.RUNNING, trial.status) self.trial_executor.stop_trial(trial) self.assertEqual(Trial.TERMINATED, trial.status) def testNoResetTrial(self): """Tests that reset handles NotImplemented properly.""" trial = Trial("__fake") self.trial_executor.start_trial(trial) exists = self.trial_executor.reset_trial(trial, {}, "modified_mock") self.assertEqual(exists, False) self.assertEqual(Trial.RUNNING, trial.status) def testResetTrial(self): """Tests that reset works as expected.""" class B(Trainable): def _train(self): return dict(timesteps_this_iter=1, done=True) def reset_config(self, config): self.config = config return True trials = self.generate_trials({ "run": B, "config": { "foo": 0 }, }, "grid_search") trial = trials[0] self.trial_executor.start_trial(trial) exists = self.trial_executor.reset_trial(trial, {"hi": 1}, "modified_mock") self.assertEqual(exists, True) self.assertEqual(trial.config.get("hi"), 1) self.assertEqual(trial.experiment_tag, "modified_mock") self.assertEqual(Trial.RUNNING, trial.status) def generate_trials(self, spec, name): suggester = BasicVariantGenerator() suggester.add_configurations({name: spec}) return suggester.next_trials() class RayExecutorQueueTest(unittest.TestCase): def setUp(self): self.trial_executor = RayTrialExecutor( queue_trials=True, refresh_period=0) self.cluster = Cluster( initialize_head=True, connect=True, head_node_args={ "num_cpus": 1, "_internal_config": json.dumps({ "num_heartbeats_timeout": 10 }) }) # Pytest doesn't play nicely with imports _register_all() def tearDown(self): ray.shutdown() self.cluster.shutdown() _register_all() # re-register the evicted objects def testQueueTrial(self): """Tests that reset handles NotImplemented properly.""" def create_trial(cpu, gpu=0): return Trial("__fake", resources=Resources(cpu=cpu, gpu=gpu)) cpu_only = create_trial(1, 0) self.assertTrue(self.trial_executor.has_resources(cpu_only.resources)) self.trial_executor.start_trial(cpu_only) gpu_only = create_trial(0, 1) self.assertTrue(self.trial_executor.has_resources(gpu_only.resources)) def testHeadBlocking(self): def create_trial(cpu, gpu=0): return Trial("__fake", resources=Resources(cpu=cpu, gpu=gpu)) gpu_trial = create_trial(1, 1) self.assertTrue(self.trial_executor.has_resources(gpu_trial.resources)) self.trial_executor.start_trial(gpu_trial) # TODO(rliaw): This behavior is probably undesirable, but right now # trials with different resource requirements is not often used. cpu_only_trial = create_trial(1, 0) self.assertFalse( self.trial_executor.has_resources(cpu_only_trial.resources)) self.cluster.add_node(num_cpus=1, num_gpus=1) self.cluster.wait_for_nodes() self.assertTrue( self.trial_executor.has_resources(cpu_only_trial.resources)) self.trial_executor.start_trial(cpu_only_trial) cpu_only_trial2 = create_trial(1, 0) self.assertTrue( self.trial_executor.has_resources(cpu_only_trial2.resources)) self.trial_executor.start_trial(cpu_only_trial2) cpu_only_trial3 = create_trial(1, 0) self.assertFalse( self.trial_executor.has_resources(cpu_only_trial3.resources)) class LocalModeExecutorTest(RayTrialExecutorTest): def setUp(self): self.trial_executor = RayTrialExecutor(queue_trials=False) ray.init(local_mode=True) def tearDown(self): ray.shutdown() _register_all() # re-register the evicted objects if __name__ == "__main__": import pytest import sys sys.exit(pytest.main(["-v", __file__]))
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/test_run_experiment.py
Python
import os import unittest import ray from ray.rllib import _register_all from ray.tune.result import TIMESTEPS_TOTAL from ray.tune import Trainable, TuneError from ray.tune import register_trainable, run_experiments from ray.tune.logger import Logger from ray.tune.experiment import Experiment from ray.tune.trial import Trial, ExportFormat class RunExperimentTest(unittest.TestCase): def tearDown(self): ray.shutdown() _register_all() # re-register the evicted objects def testDict(self): def train(config, reporter): for i in range(100): reporter(timesteps_total=i) register_trainable("f1", train) trials = run_experiments({ "foo": { "run": "f1", }, "bar": { "run": "f1", } }) for trial in trials: self.assertEqual(trial.status, Trial.TERMINATED) self.assertEqual(trial.last_result[TIMESTEPS_TOTAL], 99) def testExperiment(self): def train(config, reporter): for i in range(100): reporter(timesteps_total=i) register_trainable("f1", train) exp1 = Experiment(**{ "name": "foo", "run": "f1", }) [trial] = run_experiments(exp1) self.assertEqual(trial.status, Trial.TERMINATED) self.assertEqual(trial.last_result[TIMESTEPS_TOTAL], 99) def testExperimentList(self): def train(config, reporter): for i in range(100): reporter(timesteps_total=i) register_trainable("f1", train) exp1 = Experiment(**{ "name": "foo", "run": "f1", }) exp2 = Experiment(**{ "name": "bar", "run": "f1", }) trials = run_experiments([exp1, exp2]) for trial in trials: self.assertEqual(trial.status, Trial.TERMINATED) self.assertEqual(trial.last_result[TIMESTEPS_TOTAL], 99) def testAutoregisterTrainable(self): def train(config, reporter): for i in range(100): reporter(timesteps_total=i) class B(Trainable): def _train(self): return {"timesteps_this_iter": 1, "done": True} register_trainable("f1", train) trials = run_experiments({ "foo": { "run": train, }, "bar": { "run": B } }) for trial in trials: self.assertEqual(trial.status, Trial.TERMINATED) def testCheckpointAtEnd(self): class train(Trainable): def _train(self): return {"timesteps_this_iter": 1, "done": True} def _save(self, path): checkpoint = path + "/checkpoint" with open(checkpoint, "w") as f: f.write("OK") return checkpoint trials = run_experiments({ "foo": { "run": train, "checkpoint_at_end": True } }) for trial in trials: self.assertEqual(trial.status, Trial.TERMINATED) self.assertTrue(trial.has_checkpoint()) def testExportFormats(self): class train(Trainable): def _train(self): return {"timesteps_this_iter": 1, "done": True} def _export_model(self, export_formats, export_dir): path = export_dir + "/exported" with open(path, "w") as f: f.write("OK") return {export_formats[0]: path} trials = run_experiments({ "foo": { "run": train, "export_formats": ["format"] } }) for trial in trials: self.assertEqual(trial.status, Trial.TERMINATED) self.assertTrue( os.path.exists(os.path.join(trial.logdir, "exported"))) def testInvalidExportFormats(self): class train(Trainable): def _train(self): return {"timesteps_this_iter": 1, "done": True} def _export_model(self, export_formats, export_dir): ExportFormat.validate(export_formats) return {} def fail_trial(): run_experiments({ "foo": { "run": train, "export_formats": ["format"] } }) self.assertRaises(TuneError, fail_trial) def testCustomResources(self): ray.shutdown() ray.init(resources={"hi": 3}) class train(Trainable): def _train(self): return {"timesteps_this_iter": 1, "done": True} trials = run_experiments({ "foo": { "run": train, "resources_per_trial": { "cpu": 1, "custom_resources": { "hi": 2 } } } }) for trial in trials: self.assertEqual(trial.status, Trial.TERMINATED) def testCustomLogger(self): class CustomLogger(Logger): def on_result(self, result): with open(os.path.join(self.logdir, "test.log"), "w") as f: f.write("hi") [trial] = run_experiments({ "foo": { "run": "__fake", "stop": { "training_iteration": 1 }, "loggers": [CustomLogger] } }) self.assertTrue(os.path.exists(os.path.join(trial.logdir, "test.log"))) self.assertFalse( os.path.exists(os.path.join(trial.logdir, "params.json"))) [trial] = run_experiments({ "foo": { "run": "__fake", "stop": { "training_iteration": 1 } } }) self.assertTrue( os.path.exists(os.path.join(trial.logdir, "params.json"))) [trial] = run_experiments({ "foo": { "run": "__fake", "stop": { "training_iteration": 1 }, "loggers": [] } }) self.assertFalse( os.path.exists(os.path.join(trial.logdir, "params.json"))) def testCustomTrialString(self): [trial] = run_experiments({ "foo": { "run": "__fake", "stop": { "training_iteration": 1 }, "trial_name_creator": lambda t: "{}_{}_321".format(t.trainable_name, t.trial_id) } }) self.assertEquals( str(trial), "{}_{}_321".format(trial.trainable_name, trial.trial_id)) if __name__ == "__main__": import pytest import sys sys.exit(pytest.main(["-v", __file__]))
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/test_sync.py
Python
import glob import os import shutil import sys import tempfile import unittest from unittest.mock import patch import ray from ray.rllib import _register_all from ray import tune from ray.tune import TuneError from ray.tune.syncer import CommandBasedClient class TestSyncFunctionality(unittest.TestCase): def setUp(self): ray.init() def tearDown(self): ray.shutdown() _register_all() # re-register the evicted objects @patch("ray.tune.sync_client.S3_PREFIX", "test") def testNoUploadDir(self): """No Upload Dir is given.""" with self.assertRaises(AssertionError): [trial] = tune.run( "__fake", name="foo", max_failures=0, **{ "stop": { "training_iteration": 1 }, "sync_to_cloud": "echo {source} {target}" }).trials @patch("ray.tune.sync_client.S3_PREFIX", "test") def testCloudProperString(self): with self.assertRaises(ValueError): [trial] = tune.run( "__fake", name="foo", max_failures=0, **{ "stop": { "training_iteration": 1 }, "upload_dir": "test", "sync_to_cloud": "ls {target}" }).trials with self.assertRaises(ValueError): [trial] = tune.run( "__fake", name="foo", max_failures=0, **{ "stop": { "training_iteration": 1 }, "upload_dir": "test", "sync_to_cloud": "ls {source}" }).trials tmpdir = tempfile.mkdtemp() logfile = os.path.join(tmpdir, "test.log") [trial] = tune.run( "__fake", name="foo", max_failures=0, **{ "stop": { "training_iteration": 1 }, "upload_dir": "test", "sync_to_cloud": "echo {source} {target} > " + logfile }).trials with open(logfile) as f: lines = f.read() self.assertTrue("test" in lines) shutil.rmtree(tmpdir) def testClusterProperString(self): """Tests that invalid commands throw..""" with self.assertRaises(TuneError): # This raises TuneError because logger is init in safe zone. [trial] = tune.run( "__fake", name="foo", max_failures=0, **{ "stop": { "training_iteration": 1 }, "sync_to_driver": "ls {target}" }).trials with self.assertRaises(TuneError): # This raises TuneError because logger is init in safe zone. [trial] = tune.run( "__fake", name="foo", max_failures=0, **{ "stop": { "training_iteration": 1 }, "sync_to_driver": "ls {source}" }).trials with patch.object(CommandBasedClient, "_execute") as mock_fn: with patch("ray.services.get_node_ip_address") as mock_sync: mock_sync.return_value = "0.0.0.0" [trial] = tune.run( "__fake", name="foo", max_failures=0, **{ "stop": { "training_iteration": 1 }, "sync_to_driver": "echo {source} {target}" }).trials self.assertGreater(mock_fn.call_count, 0) def testCloudFunctions(self): tmpdir = tempfile.mkdtemp() tmpdir2 = tempfile.mkdtemp() os.mkdir(os.path.join(tmpdir2, "foo")) def sync_func(local, remote): for filename in glob.glob(os.path.join(local, "*.json")): shutil.copy(filename, remote) [trial] = tune.run( "__fake", name="foo", max_failures=0, local_dir=tmpdir, stop={ "training_iteration": 1 }, upload_dir=tmpdir2, sync_to_cloud=sync_func).trials test_file_path = glob.glob(os.path.join(tmpdir2, "foo", "*.json")) self.assertTrue(test_file_path) shutil.rmtree(tmpdir) shutil.rmtree(tmpdir2) def testClusterSyncFunction(self): def sync_func_driver(source, target): assert ":" in source, "Source {} not a remote path.".format(source) assert ":" not in target, "Target is supposed to be local." with open(os.path.join(target, "test.log2"), "w") as f: print("writing to", f.name) f.write(source) [trial] = tune.run( "__fake", name="foo", max_failures=0, stop={ "training_iteration": 1 }, sync_to_driver=sync_func_driver).trials test_file_path = os.path.join(trial.logdir, "test.log2") self.assertFalse(os.path.exists(test_file_path)) with patch("ray.services.get_node_ip_address") as mock_sync: mock_sync.return_value = "0.0.0.0" [trial] = tune.run( "__fake", name="foo", max_failures=0, stop={ "training_iteration": 1 }, sync_to_driver=sync_func_driver).trials test_file_path = os.path.join(trial.logdir, "test.log2") self.assertTrue(os.path.exists(test_file_path)) os.remove(test_file_path) def testNoSync(self): """Sync should not run on a single node.""" def sync_func(source, target): pass with patch.object(CommandBasedClient, "_execute") as mock_sync: [trial] = tune.run( "__fake", name="foo", max_failures=0, **{ "stop": { "training_iteration": 1 }, "sync_to_driver": sync_func }).trials self.assertEqual(mock_sync.call_count, 0) if __name__ == "__main__": import pytest sys.exit(pytest.main(["-v", __file__]))
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/test_track.py
Python
import os import pandas as pd import unittest import ray from ray import tune from ray.tune import track from ray.tune.result import EXPR_PARAM_FILE, EXPR_RESULT_FILE def _check_json_val(fname, key, val): with open(fname, "r") as f: df = pd.read_json(f, typ="frame", lines=True) return key in df.columns and (df[key].tail(n=1) == val).all() class TrackApiTest(unittest.TestCase): def tearDown(self): track.shutdown() ray.shutdown() def testSessionInitShutdown(self): self.assertTrue(track._session is None) # Checks that the singleton _session is created/destroyed # by track.init() and track.shutdown() for _ in range(2): # do it twice to see that we can reopen the session track.init(trial_name="test_init") self.assertTrue(track._session is not None) track.shutdown() self.assertTrue(track._session is None) def testLogCreation(self): """Checks that track.init() starts logger and creates log files.""" track.init(trial_name="test_init") session = track.get_session() self.assertTrue(session is not None) self.assertTrue(os.path.isdir(session.logdir)) params_path = os.path.join(session.logdir, EXPR_PARAM_FILE) result_path = os.path.join(session.logdir, EXPR_RESULT_FILE) self.assertTrue(os.path.exists(params_path)) self.assertTrue(os.path.exists(result_path)) self.assertTrue(session.logdir == track.trial_dir()) def testMetric(self): track.init(trial_name="test_log") session = track.get_session() for i in range(5): track.log(test=i) result_path = os.path.join(session.logdir, EXPR_RESULT_FILE) self.assertTrue(_check_json_val(result_path, "test", i)) def testRayOutput(self): """Checks that local and remote log format are the same.""" ray.init() def testme(config): for i in range(config["iters"]): track.log(iteration=i, hi="test") trials = tune.run(testme, config={"iters": 5}).trials trial_res = trials[0].last_result self.assertTrue(trial_res["hi"], "test") self.assertTrue(trial_res["training_iteration"], 5) def testLocalMetrics(self): """Checks that metric state is updated correctly.""" track.init(trial_name="test_logs") session = track.get_session() self.assertEqual(set(session.trial_config.keys()), {"trial_id"}) result_path = os.path.join(session.logdir, EXPR_RESULT_FILE) track.log(test=1) self.assertTrue(_check_json_val(result_path, "test", 1)) track.log(iteration=1, test=2) self.assertTrue(_check_json_val(result_path, "test", 2)) if __name__ == "__main__": import pytest import sys sys.exit(pytest.main(["-v", __file__]))
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/test_trainable_util.py
Python
import os import pickle import shutil import unittest from ray.tune.trainable import TrainableUtil class TrainableUtilTest(unittest.TestCase): def setUp(self): self.checkpoint_dir = "/tmp/tune/MyTrainable123" TrainableUtil.make_checkpoint_dir(self.checkpoint_dir) def tearDown(self): self.addCleanup(shutil.rmtree, self.checkpoint_dir) def testFindCheckpointDir(self): checkpoint_path = os.path.join(self.checkpoint_dir, "my/nested/chkpt") os.makedirs(checkpoint_path) found_dir = TrainableUtil.find_checkpoint_dir(checkpoint_path) self.assertEquals(self.checkpoint_dir, found_dir) with self.assertRaises(FileNotFoundError): parent = os.path.dirname(found_dir) TrainableUtil.find_checkpoint_dir(parent) def testPickleCheckpoint(self): for i in range(5): path = os.path.join(self.checkpoint_dir, str(i)) with open(path, "w") as f: f.write(str(i)) checkpoint_path = os.path.join(self.checkpoint_dir, "0") data_dict = TrainableUtil.pickle_checkpoint(checkpoint_path) loaded = pickle.loads(data_dict) checkpoint_name = os.path.basename(checkpoint_path) self.assertEqual(loaded["checkpoint_name"], checkpoint_name) for i in range(5): path = os.path.join(self.checkpoint_dir, str(i)) self.assertEquals(loaded["data"][str(i)], open(path, "rb").read())
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/test_trial_runner.py
Python
import sys import unittest import ray from ray.rllib import _register_all from ray import tune from ray.tune import TuneError, register_trainable from ray.tune.ray_trial_executor import RayTrialExecutor from ray.tune.schedulers import TrialScheduler, FIFOScheduler from ray.tune.trial import Trial from ray.tune.trial_runner import TrialRunner from ray.tune.resources import Resources from ray.tune.suggest import BasicVariantGenerator class TrialRunnerTest(unittest.TestCase): def setUp(self): _register_all() # re-register the evicted objects def tearDown(self): ray.shutdown() def testTrialStatus(self): ray.init() trial = Trial("__fake") trial_executor = RayTrialExecutor() self.assertEqual(trial.status, Trial.PENDING) trial_executor.start_trial(trial) self.assertEqual(trial.status, Trial.RUNNING) trial_executor.stop_trial(trial) self.assertEqual(trial.status, Trial.TERMINATED) trial_executor.stop_trial(trial, error=True) self.assertEqual(trial.status, Trial.ERROR) def testExperimentTagTruncation(self): ray.init() def train(config, reporter): reporter(timesteps_total=1) trial_executor = RayTrialExecutor() register_trainable("f1", train) experiments = { "foo": { "run": "f1", "config": { "a" * 50: tune.sample_from(lambda spec: 5.0 / 7), "b" * 50: tune.sample_from(lambda spec: "long" * 40) }, } } for name, spec in experiments.items(): trial_generator = BasicVariantGenerator() trial_generator.add_configurations({name: spec}) for trial in trial_generator.next_trials(): trial_executor.start_trial(trial) self.assertLessEqual(len(trial.logdir), 200) trial_executor.stop_trial(trial) def testExtraResources(self): ray.init(num_cpus=4, num_gpus=2) runner = TrialRunner() kwargs = { "stopping_criterion": { "training_iteration": 1 }, "resources": Resources(cpu=1, gpu=0, extra_cpu=3, extra_gpu=1), } trials = [Trial("__fake", **kwargs), Trial("__fake", **kwargs)] for t in trials: runner.add_trial(t) runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) self.assertEqual(trials[1].status, Trial.PENDING) runner.step() self.assertEqual(trials[0].status, Trial.TERMINATED) self.assertEqual(trials[1].status, Trial.PENDING) def testCustomResources(self): ray.init(num_cpus=4, num_gpus=2, resources={"a": 2}) runner = TrialRunner() kwargs = { "stopping_criterion": { "training_iteration": 1 }, "resources": Resources(cpu=1, gpu=0, custom_resources={"a": 2}), } trials = [Trial("__fake", **kwargs), Trial("__fake", **kwargs)] for t in trials: runner.add_trial(t) runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) self.assertEqual(trials[1].status, Trial.PENDING) runner.step() self.assertEqual(trials[0].status, Trial.TERMINATED) self.assertEqual(trials[1].status, Trial.PENDING) def testExtraCustomResources(self): ray.init(num_cpus=4, num_gpus=2, resources={"a": 2}) runner = TrialRunner() kwargs = { "stopping_criterion": { "training_iteration": 1 }, "resources": Resources( cpu=1, gpu=0, extra_custom_resources={"a": 2}), } trials = [Trial("__fake", **kwargs), Trial("__fake", **kwargs)] for t in trials: runner.add_trial(t) runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) self.assertEqual(trials[1].status, Trial.PENDING) runner.step() self.assertTrue(sum(t.status == Trial.RUNNING for t in trials) < 2) self.assertEqual(trials[0].status, Trial.TERMINATED) self.assertEqual(trials[1].status, Trial.PENDING) def testCustomResources2(self): ray.init(num_cpus=4, num_gpus=2, resources={"a": 2}) runner = TrialRunner() resource1 = Resources(cpu=1, gpu=0, extra_custom_resources={"a": 2}) self.assertTrue(runner.has_resources(resource1)) resource2 = Resources(cpu=1, gpu=0, custom_resources={"a": 2}) self.assertTrue(runner.has_resources(resource2)) resource3 = Resources(cpu=1, gpu=0, custom_resources={"a": 3}) self.assertFalse(runner.has_resources(resource3)) resource4 = Resources(cpu=1, gpu=0, extra_custom_resources={"a": 3}) self.assertFalse(runner.has_resources(resource4)) def testFractionalGpus(self): ray.init(num_cpus=4, num_gpus=1) runner = TrialRunner() kwargs = { "resources": Resources(cpu=1, gpu=0.5), } trials = [ Trial("__fake", **kwargs), Trial("__fake", **kwargs), Trial("__fake", **kwargs), Trial("__fake", **kwargs) ] for t in trials: runner.add_trial(t) for _ in range(10): runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) self.assertEqual(trials[1].status, Trial.RUNNING) self.assertEqual(trials[2].status, Trial.PENDING) self.assertEqual(trials[3].status, Trial.PENDING) def testResourceNumericalError(self): resource = Resources(cpu=0.99, gpu=0.99, custom_resources={"a": 0.99}) small_resource = Resources( cpu=0.33, gpu=0.33, custom_resources={"a": 0.33}) for i in range(3): resource = Resources.subtract(resource, small_resource) self.assertTrue(resource.is_nonnegative()) def testResourceScheduler(self): ray.init(num_cpus=4, num_gpus=1) runner = TrialRunner() kwargs = { "stopping_criterion": { "training_iteration": 1 }, "resources": Resources(cpu=1, gpu=1), } trials = [Trial("__fake", **kwargs), Trial("__fake", **kwargs)] for t in trials: runner.add_trial(t) runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) self.assertEqual(trials[1].status, Trial.PENDING) runner.step() self.assertEqual(trials[0].status, Trial.TERMINATED) self.assertEqual(trials[1].status, Trial.PENDING) runner.step() self.assertEqual(trials[0].status, Trial.TERMINATED) self.assertEqual(trials[1].status, Trial.RUNNING) runner.step() self.assertEqual(trials[0].status, Trial.TERMINATED) self.assertEqual(trials[1].status, Trial.TERMINATED) def testMultiStepRun(self): ray.init(num_cpus=4, num_gpus=2) runner = TrialRunner() kwargs = { "stopping_criterion": { "training_iteration": 5 }, "resources": Resources(cpu=1, gpu=1), } trials = [Trial("__fake", **kwargs), Trial("__fake", **kwargs)] for t in trials: runner.add_trial(t) runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) self.assertEqual(trials[1].status, Trial.PENDING) runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) self.assertEqual(trials[1].status, Trial.RUNNING) runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) self.assertEqual(trials[1].status, Trial.RUNNING) runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) self.assertEqual(trials[1].status, Trial.RUNNING) def testMultiStepRun2(self): """Checks that runner.step throws when overstepping.""" ray.init(num_cpus=1) runner = TrialRunner() kwargs = { "stopping_criterion": { "training_iteration": 2 }, "resources": Resources(cpu=1, gpu=0), } trials = [Trial("__fake", **kwargs)] for t in trials: runner.add_trial(t) runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) runner.step() self.assertEqual(trials[0].status, Trial.TERMINATED) self.assertRaises(TuneError, runner.step) def testChangeResources(self): """Checks that resource requirements can be changed on fly.""" ray.init(num_cpus=2) class ChangingScheduler(FIFOScheduler): def on_trial_result(self, trial_runner, trial, result): if result["training_iteration"] == 1: executor = trial_runner.trial_executor executor.stop_trial(trial, stop_logger=False) trial.update_resources(2, 0) executor.start_trial(trial) return TrialScheduler.CONTINUE runner = TrialRunner(scheduler=ChangingScheduler()) kwargs = { "stopping_criterion": { "training_iteration": 2 }, "resources": Resources(cpu=1, gpu=0), } trials = [Trial("__fake", **kwargs)] for t in trials: runner.add_trial(t) runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) self.assertEqual(runner.trial_executor._committed_resources.cpu, 1) self.assertRaises(ValueError, lambda: trials[0].update_resources(2, 0)) runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) self.assertEqual(runner.trial_executor._committed_resources.cpu, 2) if __name__ == "__main__": import pytest sys.exit(pytest.main(["-v", __file__]))
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/test_trial_runner_2.py
Python
import os import sys import unittest from unittest.mock import patch import ray from ray.rllib import _register_all from ray.tune import TuneError from ray.tune.schedulers import FIFOScheduler from ray.tune.result import DONE from ray.tune.registry import _global_registry, TRAINABLE_CLASS from ray.tune.trial import Trial from ray.tune.trial_runner import TrialRunner from ray.tune.resources import Resources from ray.tune.suggest import BasicVariantGenerator def create_mock_components(): class _MockScheduler(FIFOScheduler): errored_trials = [] def on_trial_error(self, trial_runner, trial): self.errored_trials += [trial] class _MockSearchAlg(BasicVariantGenerator): errored_trials = [] def on_trial_complete(self, trial_id, error=False, **kwargs): if error: self.errored_trials += [trial_id] searchalg = _MockSearchAlg() scheduler = _MockScheduler() return searchalg, scheduler class TrialRunnerTest2(unittest.TestCase): def tearDown(self): ray.shutdown() _register_all() # re-register the evicted objects def testErrorHandling(self): ray.init(num_cpus=4, num_gpus=2) runner = TrialRunner() kwargs = { "stopping_criterion": { "training_iteration": 1 }, "resources": Resources(cpu=1, gpu=1), } _global_registry.register(TRAINABLE_CLASS, "asdf", None) trials = [Trial("asdf", **kwargs), Trial("__fake", **kwargs)] for t in trials: runner.add_trial(t) runner.step() self.assertEqual(trials[0].status, Trial.ERROR) self.assertEqual(trials[1].status, Trial.PENDING) runner.step() self.assertEqual(trials[0].status, Trial.ERROR) self.assertEqual(trials[1].status, Trial.RUNNING) def testThrowOnOverstep(self): ray.init(num_cpus=1, num_gpus=1) runner = TrialRunner() runner.step() self.assertRaises(TuneError, runner.step) def testFailureRecoveryDisabled(self): ray.init(num_cpus=1, num_gpus=1) searchalg, scheduler = create_mock_components() runner = TrialRunner(searchalg, scheduler=scheduler) kwargs = { "resources": Resources(cpu=1, gpu=1), "checkpoint_freq": 1, "max_failures": 0, "config": { "mock_error": True, }, } runner.add_trial(Trial("__fake", **kwargs)) trials = runner.get_trials() runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) runner.step() self.assertEqual(trials[0].status, Trial.ERROR) self.assertEqual(trials[0].num_failures, 1) self.assertEqual(len(searchalg.errored_trials), 1) self.assertEqual(len(scheduler.errored_trials), 1) def testFailureRecoveryEnabled(self): ray.init(num_cpus=1, num_gpus=1) searchalg, scheduler = create_mock_components() runner = TrialRunner(searchalg, scheduler=scheduler) kwargs = { "resources": Resources(cpu=1, gpu=1), "checkpoint_freq": 1, "max_failures": 1, "config": { "mock_error": True, }, } runner.add_trial(Trial("__fake", **kwargs)) trials = runner.get_trials() runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) self.assertEqual(trials[0].num_failures, 1) runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) self.assertEqual(len(searchalg.errored_trials), 0) self.assertEqual(len(scheduler.errored_trials), 0) def testFailureRecoveryNodeRemoval(self): ray.init(num_cpus=1, num_gpus=1) searchalg, scheduler = create_mock_components() runner = TrialRunner(searchalg, scheduler=scheduler) kwargs = { "resources": Resources(cpu=1, gpu=1), "checkpoint_freq": 1, "max_failures": 1, "config": { "mock_error": True, }, } runner.add_trial(Trial("__fake", **kwargs)) trials = runner.get_trials() with patch("ray.cluster_resources") as resource_mock: resource_mock.return_value = {"CPU": 1, "GPU": 1} runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) # Mimic a node failure resource_mock.return_value = {"CPU": 0, "GPU": 0} runner.step() self.assertEqual(trials[0].status, Trial.PENDING) self.assertEqual(trials[0].num_failures, 1) self.assertEqual(len(searchalg.errored_trials), 0) self.assertEqual(len(scheduler.errored_trials), 1) def testFailureRecoveryMaxFailures(self): ray.init(num_cpus=1, num_gpus=1) runner = TrialRunner() kwargs = { "resources": Resources(cpu=1, gpu=1), "checkpoint_freq": 1, "max_failures": 2, "config": { "mock_error": True, "persistent_error": True, }, } runner.add_trial(Trial("__fake", **kwargs)) trials = runner.get_trials() runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) self.assertEqual(trials[0].num_failures, 1) runner.step() # Restore step runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) self.assertEqual(trials[0].num_failures, 2) runner.step() # Restore step runner.step() self.assertEqual(trials[0].status, Trial.ERROR) self.assertEqual(trials[0].num_failures, 3) def testCheckpointing(self): ray.init(num_cpus=1, num_gpus=1) runner = TrialRunner() kwargs = { "stopping_criterion": { "training_iteration": 1 }, "resources": Resources(cpu=1, gpu=1), } runner.add_trial(Trial("__fake", **kwargs)) trials = runner.get_trials() runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) self.assertEqual(ray.get(trials[0].runner.set_info.remote(1)), 1) path = runner.trial_executor.save(trials[0]) kwargs["restore_path"] = path runner.add_trial(Trial("__fake", **kwargs)) trials = runner.get_trials() runner.step() self.assertEqual(trials[0].status, Trial.TERMINATED) self.assertEqual(trials[1].status, Trial.PENDING) runner.step() self.assertEqual(trials[0].status, Trial.TERMINATED) self.assertEqual(trials[1].status, Trial.RUNNING) self.assertEqual(ray.get(trials[1].runner.get_info.remote()), 1) self.addCleanup(os.remove, path) def testRestoreMetricsAfterCheckpointing(self): ray.init(num_cpus=1, num_gpus=1) runner = TrialRunner() kwargs = { "resources": Resources(cpu=1, gpu=1), } runner.add_trial(Trial("__fake", **kwargs)) trials = runner.get_trials() runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) self.assertEqual(ray.get(trials[0].runner.set_info.remote(1)), 1) path = runner.trial_executor.save(trials[0]) runner.trial_executor.stop_trial(trials[0]) kwargs["restore_path"] = path runner.add_trial(Trial("__fake", **kwargs)) trials = runner.get_trials() runner.step() self.assertEqual(trials[0].status, Trial.TERMINATED) self.assertEqual(trials[1].status, Trial.RUNNING) runner.step() # Restore step runner.step() self.assertEqual(trials[1].last_result["timesteps_since_restore"], 10) self.assertEqual(trials[1].last_result["iterations_since_restore"], 1) self.assertGreater(trials[1].last_result["time_since_restore"], 0) runner.step() self.assertEqual(trials[1].last_result["timesteps_since_restore"], 20) self.assertEqual(trials[1].last_result["iterations_since_restore"], 2) self.assertGreater(trials[1].last_result["time_since_restore"], 0) self.addCleanup(os.remove, path) def testCheckpointingAtEnd(self): ray.init(num_cpus=1, num_gpus=1) runner = TrialRunner() kwargs = { "stopping_criterion": { "training_iteration": 2 }, "checkpoint_at_end": True, "resources": Resources(cpu=1, gpu=1), } runner.add_trial(Trial("__fake", **kwargs)) trials = runner.get_trials() runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) runner.step() runner.step() self.assertEqual(trials[0].last_result[DONE], True) self.assertEqual(trials[0].has_checkpoint(), True) def testResultDone(self): """Tests that last_result is marked `done` after trial is complete.""" ray.init(num_cpus=1, num_gpus=1) runner = TrialRunner() kwargs = { "stopping_criterion": { "training_iteration": 2 }, "resources": Resources(cpu=1, gpu=1), } runner.add_trial(Trial("__fake", **kwargs)) trials = runner.get_trials() runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) runner.step() self.assertNotEqual(trials[0].last_result[DONE], True) runner.step() self.assertEqual(trials[0].last_result[DONE], True) def testPauseThenResume(self): ray.init(num_cpus=1, num_gpus=1) runner = TrialRunner() kwargs = { "stopping_criterion": { "training_iteration": 2 }, "resources": Resources(cpu=1, gpu=1), } runner.add_trial(Trial("__fake", **kwargs)) trials = runner.get_trials() runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) self.assertEqual(ray.get(trials[0].runner.get_info.remote()), None) self.assertEqual(ray.get(trials[0].runner.set_info.remote(1)), 1) runner.trial_executor.pause_trial(trials[0]) self.assertEqual(trials[0].status, Trial.PAUSED) runner.trial_executor.resume_trial(trials[0]) self.assertEqual(trials[0].status, Trial.RUNNING) runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) self.assertEqual(ray.get(trials[0].runner.get_info.remote()), 1) runner.step() self.assertEqual(trials[0].status, Trial.TERMINATED) if __name__ == "__main__": import pytest sys.exit(pytest.main(["-v", __file__]))
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/test_trial_runner_3.py
Python
import os import shutil import sys import tempfile import unittest import ray from ray.rllib import _register_all from ray.tune import TuneError from ray.tune.schedulers import TrialScheduler, FIFOScheduler from ray.tune.experiment import Experiment from ray.tune.trial import Trial from ray.tune.trial_runner import TrialRunner from ray.tune.resources import Resources, json_to_resources, resources_to_json from ray.tune.suggest.suggestion import (_MockSuggestionAlgorithm, SuggestionAlgorithm) class TrialRunnerTest3(unittest.TestCase): def tearDown(self): ray.shutdown() _register_all() # re-register the evicted objects def testStepHook(self): ray.init(num_cpus=4, num_gpus=2) runner = TrialRunner() def on_step_begin(self, trialrunner): self._update_avail_resources() cnt = self.pre_step if hasattr(self, "pre_step") else 0 setattr(self, "pre_step", cnt + 1) def on_step_end(self, trialrunner): cnt = self.pre_step if hasattr(self, "post_step") else 0 setattr(self, "post_step", 1 + cnt) import types runner.trial_executor.on_step_begin = types.MethodType( on_step_begin, runner.trial_executor) runner.trial_executor.on_step_end = types.MethodType( on_step_end, runner.trial_executor) kwargs = { "stopping_criterion": { "training_iteration": 5 }, "resources": Resources(cpu=1, gpu=1), } runner.add_trial(Trial("__fake", **kwargs)) runner.step() self.assertEqual(runner.trial_executor.pre_step, 1) self.assertEqual(runner.trial_executor.post_step, 1) def testStopTrial(self): ray.init(num_cpus=4, num_gpus=2) runner = TrialRunner() kwargs = { "stopping_criterion": { "training_iteration": 5 }, "resources": Resources(cpu=1, gpu=1), } trials = [ Trial("__fake", **kwargs), Trial("__fake", **kwargs), Trial("__fake", **kwargs), Trial("__fake", **kwargs) ] for t in trials: runner.add_trial(t) runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) self.assertEqual(trials[1].status, Trial.PENDING) # Stop trial while running runner.stop_trial(trials[0]) self.assertEqual(trials[0].status, Trial.TERMINATED) self.assertEqual(trials[1].status, Trial.PENDING) runner.step() self.assertEqual(trials[0].status, Trial.TERMINATED) self.assertEqual(trials[1].status, Trial.RUNNING) self.assertEqual(trials[-1].status, Trial.PENDING) # Stop trial while pending runner.stop_trial(trials[-1]) self.assertEqual(trials[0].status, Trial.TERMINATED) self.assertEqual(trials[1].status, Trial.RUNNING) self.assertEqual(trials[-1].status, Trial.TERMINATED) runner.step() self.assertEqual(trials[0].status, Trial.TERMINATED) self.assertEqual(trials[1].status, Trial.RUNNING) self.assertEqual(trials[2].status, Trial.RUNNING) self.assertEqual(trials[-1].status, Trial.TERMINATED) def testSearchAlgNotification(self): """Checks notification of trial to the Search Algorithm.""" ray.init(num_cpus=4, num_gpus=2) experiment_spec = {"run": "__fake", "stop": {"training_iteration": 2}} experiments = [Experiment.from_json("test", experiment_spec)] searcher = _MockSuggestionAlgorithm(max_concurrent=10) searcher.add_configurations(experiments) runner = TrialRunner(search_alg=searcher) runner.step() trials = runner.get_trials() self.assertEqual(trials[0].status, Trial.RUNNING) runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) runner.step() self.assertEqual(trials[0].status, Trial.TERMINATED) self.assertEqual(searcher.counter["result"], 1) self.assertEqual(searcher.counter["complete"], 1) def testSearchAlgFinished(self): """Checks that SearchAlg is Finished before all trials are done.""" ray.init(num_cpus=4, num_gpus=2) experiment_spec = {"run": "__fake", "stop": {"training_iteration": 1}} experiments = [Experiment.from_json("test", experiment_spec)] searcher = _MockSuggestionAlgorithm(max_concurrent=10) searcher.add_configurations(experiments) runner = TrialRunner(search_alg=searcher) runner.step() trials = runner.get_trials() self.assertEqual(trials[0].status, Trial.RUNNING) self.assertTrue(searcher.is_finished()) self.assertFalse(runner.is_finished()) runner.step() self.assertEqual(trials[0].status, Trial.TERMINATED) self.assertEqual(len(searcher.live_trials), 0) self.assertTrue(searcher.is_finished()) self.assertTrue(runner.is_finished()) def testSearchAlgSchedulerInteraction(self): """Checks that TrialScheduler killing trial will notify SearchAlg.""" class _MockScheduler(FIFOScheduler): def on_trial_result(self, *args, **kwargs): return TrialScheduler.STOP ray.init(num_cpus=4, num_gpus=2) experiment_spec = {"run": "__fake", "stop": {"training_iteration": 2}} experiments = [Experiment.from_json("test", experiment_spec)] searcher = _MockSuggestionAlgorithm(max_concurrent=10) searcher.add_configurations(experiments) runner = TrialRunner(search_alg=searcher, scheduler=_MockScheduler()) runner.step() trials = runner.get_trials() self.assertEqual(trials[0].status, Trial.RUNNING) self.assertTrue(searcher.is_finished()) self.assertFalse(runner.is_finished()) runner.step() self.assertEqual(trials[0].status, Trial.TERMINATED) self.assertEqual(len(searcher.live_trials), 0) self.assertTrue(searcher.is_finished()) self.assertTrue(runner.is_finished()) def testSearchAlgSchedulerEarlyStop(self): """Early termination notif to Searcher can be turned off.""" class _MockScheduler(FIFOScheduler): def on_trial_result(self, *args, **kwargs): return TrialScheduler.STOP ray.init(num_cpus=4, num_gpus=2) experiment_spec = {"run": "__fake", "stop": {"training_iteration": 2}} experiments = [Experiment.from_json("test", experiment_spec)] searcher = _MockSuggestionAlgorithm(use_early_stopped_trials=True) searcher.add_configurations(experiments) runner = TrialRunner(search_alg=searcher, scheduler=_MockScheduler()) runner.step() runner.step() self.assertEqual(len(searcher.final_results), 1) searcher = _MockSuggestionAlgorithm(use_early_stopped_trials=False) searcher.add_configurations(experiments) runner = TrialRunner(search_alg=searcher, scheduler=_MockScheduler()) runner.step() runner.step() self.assertEqual(len(searcher.final_results), 0) def testSearchAlgStalled(self): """Checks that runner and searcher state is maintained when stalled.""" ray.init(num_cpus=4, num_gpus=2) experiment_spec = { "run": "__fake", "num_samples": 3, "stop": { "training_iteration": 1 } } experiments = [Experiment.from_json("test", experiment_spec)] searcher = _MockSuggestionAlgorithm(max_concurrent=1) searcher.add_configurations(experiments) runner = TrialRunner(search_alg=searcher) runner.step() trials = runner.get_trials() self.assertEqual(trials[0].status, Trial.RUNNING) runner.step() self.assertEqual(trials[0].status, Trial.TERMINATED) trials = runner.get_trials() runner.step() self.assertEqual(trials[1].status, Trial.RUNNING) self.assertEqual(len(searcher.live_trials), 1) searcher.stall = True runner.step() self.assertEqual(trials[1].status, Trial.TERMINATED) self.assertEqual(len(searcher.live_trials), 0) self.assertTrue(all(trial.is_finished() for trial in trials)) self.assertFalse(searcher.is_finished()) self.assertFalse(runner.is_finished()) searcher.stall = False runner.step() trials = runner.get_trials() self.assertEqual(trials[2].status, Trial.RUNNING) self.assertEqual(len(searcher.live_trials), 1) runner.step() self.assertEqual(trials[2].status, Trial.TERMINATED) self.assertEqual(len(searcher.live_trials), 0) self.assertTrue(searcher.is_finished()) self.assertTrue(runner.is_finished()) def testSearchAlgFinishes(self): """Empty SearchAlg changing state in `next_trials` does not crash.""" class FinishFastAlg(SuggestionAlgorithm): _index = 0 def next_trials(self): trials = [] self._index += 1 for trial in self._trial_generator: trials += [trial] break if self._index > 4: self._finished = True return trials def _suggest(self, trial_id): return {} ray.init(num_cpus=2) experiment_spec = { "run": "__fake", "num_samples": 2, "stop": { "training_iteration": 1 } } searcher = FinishFastAlg() experiments = [Experiment.from_json("test", experiment_spec)] searcher.add_configurations(experiments) runner = TrialRunner(search_alg=searcher) self.assertFalse(runner.is_finished()) runner.step() # This launches a new run runner.step() # This launches a 2nd run self.assertFalse(searcher.is_finished()) self.assertFalse(runner.is_finished()) runner.step() # This kills the first run self.assertFalse(searcher.is_finished()) self.assertFalse(runner.is_finished()) runner.step() # This kills the 2nd run self.assertFalse(searcher.is_finished()) self.assertFalse(runner.is_finished()) runner.step() # this converts self._finished to True self.assertTrue(searcher.is_finished()) self.assertRaises(TuneError, runner.step) def testTrialSaveRestore(self): """Creates different trials to test runner.checkpoint/restore.""" ray.init(num_cpus=3) tmpdir = tempfile.mkdtemp() runner = TrialRunner(local_checkpoint_dir=tmpdir, checkpoint_period=0) trials = [ Trial( "__fake", trial_id="trial_terminate", stopping_criterion={"training_iteration": 1}, checkpoint_freq=1) ] runner.add_trial(trials[0]) runner.step() # start runner.step() self.assertEquals(trials[0].status, Trial.TERMINATED) trials += [ Trial( "__fake", trial_id="trial_fail", stopping_criterion={"training_iteration": 3}, checkpoint_freq=1, config={"mock_error": True}) ] runner.add_trial(trials[1]) runner.step() runner.step() runner.step() self.assertEquals(trials[1].status, Trial.ERROR) trials += [ Trial( "__fake", trial_id="trial_succ", stopping_criterion={"training_iteration": 2}, checkpoint_freq=1) ] runner.add_trial(trials[2]) runner.step() self.assertEquals(len(runner.trial_executor.get_checkpoints()), 3) self.assertEquals(trials[2].status, Trial.RUNNING) runner2 = TrialRunner(resume="LOCAL", local_checkpoint_dir=tmpdir) for tid in ["trial_terminate", "trial_fail"]: original_trial = runner.get_trial(tid) restored_trial = runner2.get_trial(tid) self.assertEqual(original_trial.status, restored_trial.status) restored_trial = runner2.get_trial("trial_succ") self.assertEqual(Trial.PENDING, restored_trial.status) runner2.step() runner2.step() runner2.step() self.assertRaises(TuneError, runner2.step) shutil.rmtree(tmpdir) def testTrialNoSave(self): """Check that non-checkpointing trials are not saved.""" ray.init(num_cpus=3) tmpdir = tempfile.mkdtemp() runner = TrialRunner(local_checkpoint_dir=tmpdir, checkpoint_period=0) runner.add_trial( Trial( "__fake", trial_id="non_checkpoint", stopping_criterion={"training_iteration": 2})) while not all(t.status == Trial.TERMINATED for t in runner.get_trials()): runner.step() runner.add_trial( Trial( "__fake", trial_id="checkpoint", checkpoint_at_end=True, stopping_criterion={"training_iteration": 2})) while not all(t.status == Trial.TERMINATED for t in runner.get_trials()): runner.step() runner.add_trial( Trial( "__fake", trial_id="pending", stopping_criterion={"training_iteration": 2})) runner.step() runner.step() runner2 = TrialRunner(resume="LOCAL", local_checkpoint_dir=tmpdir) new_trials = runner2.get_trials() self.assertEquals(len(new_trials), 3) self.assertTrue( runner2.get_trial("non_checkpoint").status == Trial.TERMINATED) self.assertTrue( runner2.get_trial("checkpoint").status == Trial.TERMINATED) self.assertTrue(runner2.get_trial("pending").status == Trial.PENDING) self.assertTrue(not runner2.get_trial("pending").last_result) runner2.step() shutil.rmtree(tmpdir) def testCheckpointWithFunction(self): ray.init() trial = Trial( "__fake", config={"callbacks": { "on_episode_start": lambda i: i, }}, checkpoint_freq=1) tmpdir = tempfile.mkdtemp() runner = TrialRunner(local_checkpoint_dir=tmpdir, checkpoint_period=0) runner.add_trial(trial) for i in range(5): runner.step() # force checkpoint runner.checkpoint() runner2 = TrialRunner(resume="LOCAL", local_checkpoint_dir=tmpdir) new_trial = runner2.get_trials()[0] self.assertTrue("callbacks" in new_trial.config) self.assertTrue("on_episode_start" in new_trial.config["callbacks"]) shutil.rmtree(tmpdir) def testCheckpointOverwrite(self): def count_checkpoints(cdir): return sum((fname.startswith("experiment_state") and fname.endswith(".json")) for fname in os.listdir(cdir)) ray.init() trial = Trial("__fake", checkpoint_freq=1) tmpdir = tempfile.mkdtemp() runner = TrialRunner(local_checkpoint_dir=tmpdir, checkpoint_period=0) runner.add_trial(trial) for i in range(5): runner.step() # force checkpoint runner.checkpoint() self.assertEquals(count_checkpoints(tmpdir), 1) runner2 = TrialRunner(resume="LOCAL", local_checkpoint_dir=tmpdir) for i in range(5): runner2.step() self.assertEquals(count_checkpoints(tmpdir), 2) runner2.checkpoint() self.assertEquals(count_checkpoints(tmpdir), 2) shutil.rmtree(tmpdir) def testUserCheckpoint(self): ray.init(num_cpus=3) tmpdir = tempfile.mkdtemp() runner = TrialRunner(local_checkpoint_dir=tmpdir, checkpoint_period=0) runner.add_trial(Trial("__fake", config={"user_checkpoint_freq": 2})) trials = runner.get_trials() runner.step() self.assertEqual(trials[0].status, Trial.RUNNING) self.assertEqual(ray.get(trials[0].runner.set_info.remote(1)), 1) runner.step() # 0 self.assertFalse(trials[0].has_checkpoint()) runner.step() # 1 self.assertFalse(trials[0].has_checkpoint()) runner.step() # 2 self.assertTrue(trials[0].has_checkpoint()) runner2 = TrialRunner(resume="LOCAL", local_checkpoint_dir=tmpdir) runner2.step() trials2 = runner2.get_trials() self.assertEqual(ray.get(trials2[0].runner.get_info.remote()), 1) shutil.rmtree(tmpdir) class SearchAlgorithmTest(unittest.TestCase): def testNestedSuggestion(self): class TestSuggestion(SuggestionAlgorithm): def _suggest(self, trial_id): return {"a": {"b": {"c": {"d": 4, "e": 5}}}} alg = TestSuggestion() alg.add_configurations({"test": {"run": "__fake"}}) trial = alg.next_trials()[0] self.assertTrue("e=5" in trial.experiment_tag) self.assertTrue("d=4" in trial.experiment_tag) class ResourcesTest(unittest.TestCase): def testSubtraction(self): resource_1 = Resources( 1, 0, 0, 1, custom_resources={ "a": 1, "b": 2 }, extra_custom_resources={ "a": 1, "b": 1 }) resource_2 = Resources( 1, 0, 0, 1, custom_resources={ "a": 1, "b": 2 }, extra_custom_resources={ "a": 1, "b": 1 }) new_res = Resources.subtract(resource_1, resource_2) self.assertTrue(new_res.cpu == 0) self.assertTrue(new_res.gpu == 0) self.assertTrue(new_res.extra_cpu == 0) self.assertTrue(new_res.extra_gpu == 0) self.assertTrue(all(k == 0 for k in new_res.custom_resources.values())) self.assertTrue( all(k == 0 for k in new_res.extra_custom_resources.values())) def testDifferentResources(self): resource_1 = Resources(1, 0, 0, 1, custom_resources={"a": 1, "b": 2}) resource_2 = Resources(1, 0, 0, 1, custom_resources={"a": 1, "c": 2}) new_res = Resources.subtract(resource_1, resource_2) assert "c" in new_res.custom_resources assert "b" in new_res.custom_resources self.assertTrue(new_res.cpu == 0) self.assertTrue(new_res.gpu == 0) self.assertTrue(new_res.extra_cpu == 0) self.assertTrue(new_res.extra_gpu == 0) self.assertTrue(new_res.get("a") == 0) def testSerialization(self): original = Resources(1, 0, 0, 1, custom_resources={"a": 1, "b": 2}) jsoned = resources_to_json(original) new_resource = json_to_resources(jsoned) self.assertEquals(original, new_resource) if __name__ == "__main__": import pytest sys.exit(pytest.main(["-v", __file__]))
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/test_trial_scheduler.py
Python
import os import json import random import unittest import numpy as np import sys import tempfile import shutil from unittest.mock import MagicMock import ray from ray.tune.result import TRAINING_ITERATION from ray.tune.schedulers import (HyperBandScheduler, AsyncHyperBandScheduler, PopulationBasedTraining, MedianStoppingRule, TrialScheduler, HyperBandForBOHB) from ray.tune.schedulers.pbt import explore from ray.tune.trial import Trial, Checkpoint from ray.tune.trial_executor import TrialExecutor from ray.tune.resources import Resources from ray.rllib import _register_all _register_all() def result(t, rew): return dict( time_total_s=t, episode_reward_mean=rew, training_iteration=int(t)) def mock_trial_runner(trials=None): trial_runner = MagicMock() trial_runner.get_trials.return_value = trials or [] return trial_runner class EarlyStoppingSuite(unittest.TestCase): def setUp(self): ray.init() def tearDown(self): ray.shutdown() _register_all() # re-register the evicted objects def basicSetup(self, rule): t1 = Trial("PPO") # mean is 450, max 900, t_max=10 t2 = Trial("PPO") # mean is 450, max 450, t_max=5 runner = mock_trial_runner() for i in range(10): r1 = result(i, i * 100) print("basicSetup:", i) self.assertEqual( rule.on_trial_result(runner, t1, r1), TrialScheduler.CONTINUE) for i in range(5): r2 = result(i, 450) self.assertEqual( rule.on_trial_result(runner, t2, r2), TrialScheduler.CONTINUE) return t1, t2 def testMedianStoppingConstantPerf(self): rule = MedianStoppingRule(grace_period=0, min_samples_required=1) t1, t2 = self.basicSetup(rule) runner = mock_trial_runner() rule.on_trial_complete(runner, t1, result(10, 1000)) self.assertEqual( rule.on_trial_result(runner, t2, result(5, 450)), TrialScheduler.CONTINUE) self.assertEqual( rule.on_trial_result(runner, t2, result(6, 0)), TrialScheduler.CONTINUE) self.assertEqual( rule.on_trial_result(runner, t2, result(10, 450)), TrialScheduler.STOP) def testMedianStoppingOnCompleteOnly(self): rule = MedianStoppingRule(grace_period=0, min_samples_required=1) t1, t2 = self.basicSetup(rule) runner = mock_trial_runner() self.assertEqual( rule.on_trial_result(runner, t2, result(100, 0)), TrialScheduler.CONTINUE) rule.on_trial_complete(runner, t1, result(101, 1000)) self.assertEqual( rule.on_trial_result(runner, t2, result(101, 0)), TrialScheduler.STOP) def testMedianStoppingGracePeriod(self): rule = MedianStoppingRule(grace_period=2.5, min_samples_required=1) t1, t2 = self.basicSetup(rule) runner = mock_trial_runner() rule.on_trial_complete(runner, t1, result(10, 1000)) rule.on_trial_complete(runner, t2, result(10, 1000)) t3 = Trial("PPO") self.assertEqual( rule.on_trial_result(runner, t3, result(1, 10)), TrialScheduler.CONTINUE) self.assertEqual( rule.on_trial_result(runner, t3, result(2, 10)), TrialScheduler.CONTINUE) self.assertEqual( rule.on_trial_result(runner, t3, result(3, 10)), TrialScheduler.STOP) def testMedianStoppingMinSamples(self): rule = MedianStoppingRule(grace_period=0, min_samples_required=2) t1, t2 = self.basicSetup(rule) runner = mock_trial_runner() rule.on_trial_complete(runner, t1, result(10, 1000)) t3 = Trial("PPO") # Insufficient samples to evaluate t3 self.assertEqual( rule.on_trial_result(runner, t3, result(5, 10)), TrialScheduler.CONTINUE) rule.on_trial_complete(runner, t2, result(5, 1000)) # Sufficient samples to evaluate t3 self.assertEqual( rule.on_trial_result(runner, t3, result(5, 10)), TrialScheduler.STOP) def testMedianStoppingUsesMedian(self): rule = MedianStoppingRule(grace_period=0, min_samples_required=1) t1, t2 = self.basicSetup(rule) runner = mock_trial_runner() rule.on_trial_complete(runner, t1, result(10, 1000)) rule.on_trial_complete(runner, t2, result(10, 1000)) t3 = Trial("PPO") self.assertEqual( rule.on_trial_result(runner, t3, result(1, 260)), TrialScheduler.CONTINUE) self.assertEqual( rule.on_trial_result(runner, t3, result(2, 260)), TrialScheduler.STOP) def testMedianStoppingSoftStop(self): rule = MedianStoppingRule( grace_period=0, min_samples_required=1, hard_stop=False) t1, t2 = self.basicSetup(rule) runner = mock_trial_runner() rule.on_trial_complete(runner, t1, result(10, 1000)) rule.on_trial_complete(runner, t2, result(10, 1000)) t3 = Trial("PPO") self.assertEqual( rule.on_trial_result(runner, t3, result(1, 260)), TrialScheduler.CONTINUE) self.assertEqual( rule.on_trial_result(runner, t3, result(2, 260)), TrialScheduler.PAUSE) def _test_metrics(self, result_func, metric, mode): rule = MedianStoppingRule( grace_period=0, min_samples_required=1, time_attr="training_iteration", metric=metric, mode=mode) t1 = Trial("PPO") # mean is 450, max 900, t_max=10 t2 = Trial("PPO") # mean is 450, max 450, t_max=5 runner = mock_trial_runner() for i in range(10): self.assertEqual( rule.on_trial_result(runner, t1, result_func(i, i * 100)), TrialScheduler.CONTINUE) for i in range(5): self.assertEqual( rule.on_trial_result(runner, t2, result_func(i, 450)), TrialScheduler.CONTINUE) rule.on_trial_complete(runner, t1, result_func(10, 1000)) self.assertEqual( rule.on_trial_result(runner, t2, result_func(5, 450)), TrialScheduler.CONTINUE) self.assertEqual( rule.on_trial_result(runner, t2, result_func(6, 0)), TrialScheduler.CONTINUE) def testAlternateMetrics(self): def result2(t, rew): return dict(training_iteration=t, neg_mean_loss=rew) self._test_metrics(result2, "neg_mean_loss", "max") def testAlternateMetricsMin(self): def result2(t, rew): return dict(training_iteration=t, mean_loss=-rew) self._test_metrics(result2, "mean_loss", "min") class _MockTrialExecutor(TrialExecutor): def start_trial(self, trial, checkpoint_obj=None): trial.logger_running = True trial.restored_checkpoint = checkpoint_obj.value trial.status = Trial.RUNNING def stop_trial(self, trial, error=False, error_msg=None, stop_logger=True): trial.status = Trial.ERROR if error else Trial.TERMINATED if stop_logger: trial.logger_running = False def restore(self, trial, checkpoint=None): pass def save(self, trial, type=Checkpoint.PERSISTENT, result=None): return trial.trainable_name def reset_trial(self, trial, new_config, new_experiment_tag): return False class _MockTrialRunner(): def __init__(self, scheduler): self._scheduler_alg = scheduler self.trials = [] self.trial_executor = _MockTrialExecutor() def process_action(self, trial, action): if action == TrialScheduler.CONTINUE: pass elif action == TrialScheduler.PAUSE: self._pause_trial(trial) elif action == TrialScheduler.STOP: self.trial_executor.stop_trial(trial) def stop_trial(self, trial): if trial.status in [Trial.ERROR, Trial.TERMINATED]: return elif trial.status in [Trial.PENDING, Trial.PAUSED]: self._scheduler_alg.on_trial_remove(self, trial) else: self._scheduler_alg.on_trial_complete(self, trial, result(100, 10)) def add_trial(self, trial): self.trials.append(trial) self._scheduler_alg.on_trial_add(self, trial) def get_trials(self): return self.trials def has_resources(self, resources): return True def _pause_trial(self, trial): trial.status = Trial.PAUSED def _launch_trial(self, trial): trial.status = Trial.RUNNING class HyperbandSuite(unittest.TestCase): def setUp(self): ray.init(object_store_memory=int(1e8)) def tearDown(self): ray.shutdown() _register_all() # re-register the evicted objects def schedulerSetup(self, num_trials, max_t=81): """Setup a scheduler and Runner with max Iter = 9. Bracketing is placed as follows: (5, 81); (8, 27) -> (3, 54); (15, 9) -> (5, 27) -> (2, 45); (34, 3) -> (12, 9) -> (4, 27) -> (2, 42); (81, 1) -> (27, 3) -> (9, 9) -> (3, 27) -> (1, 41);""" sched = HyperBandScheduler(max_t=max_t) for i in range(num_trials): t = Trial("__fake") sched.on_trial_add(None, t) runner = _MockTrialRunner(sched) return sched, runner def default_statistics(self): """Default statistics for HyperBand.""" sched = HyperBandScheduler() res = { str(s): { "n": sched._get_n0(s), "r": sched._get_r0(s) } for s in range(sched._s_max_1) } res["max_trials"] = sum(v["n"] for v in res.values()) res["brack_count"] = sched._s_max_1 res["s_max"] = sched._s_max_1 - 1 return res def downscale(self, n, sched): return int(np.ceil(n / sched._eta)) def basicSetup(self): """Setup and verify full band.""" stats = self.default_statistics() sched, _ = self.schedulerSetup(stats["max_trials"]) self.assertEqual(len(sched._hyperbands), 1) self.assertEqual(sched._cur_band_filled(), True) filled_band = sched._hyperbands[0] for bracket in filled_band: self.assertEqual(bracket.filled(), True) return sched def advancedSetup(self): sched = self.basicSetup() for i in range(4): t = Trial("__fake") sched.on_trial_add(None, t) self.assertEqual(sched._cur_band_filled(), False) unfilled_band = sched._hyperbands[-1] self.assertEqual(len(unfilled_band), 2) bracket = unfilled_band[-1] self.assertEqual(bracket.filled(), False) self.assertEqual(len(bracket.current_trials()), 7) return sched def testConfigSameEta(self): sched = HyperBandScheduler() i = 0 while not sched._cur_band_filled(): t = Trial("__fake") sched.on_trial_add(None, t) i += 1 self.assertEqual(len(sched._hyperbands[0]), 5) self.assertEqual(sched._hyperbands[0][0]._n, 5) self.assertEqual(sched._hyperbands[0][0]._r, 81) self.assertEqual(sched._hyperbands[0][-1]._n, 81) self.assertEqual(sched._hyperbands[0][-1]._r, 1) reduction_factor = 10 sched = HyperBandScheduler( max_t=1000, reduction_factor=reduction_factor) i = 0 while not sched._cur_band_filled(): t = Trial("__fake") sched.on_trial_add(None, t) i += 1 self.assertEqual(len(sched._hyperbands[0]), 4) self.assertEqual(sched._hyperbands[0][0]._n, 4) self.assertEqual(sched._hyperbands[0][0]._r, 1000) self.assertEqual(sched._hyperbands[0][-1]._n, 1000) self.assertEqual(sched._hyperbands[0][-1]._r, 1) def testConfigSameEtaSmall(self): sched = HyperBandScheduler(max_t=1) i = 0 while len(sched._hyperbands) < 2: t = Trial("__fake") sched.on_trial_add(None, t) i += 1 self.assertEqual(len(sched._hyperbands[0]), 1) def testSuccessiveHalving(self): """Setup full band, then iterate through last bracket (n=81) to make sure successive halving is correct.""" stats = self.default_statistics() sched, mock_runner = self.schedulerSetup(stats["max_trials"]) big_bracket = sched._state["bracket"] cur_units = stats[str(stats["s_max"])]["r"] # The last bracket will downscale 4 times for x in range(stats["brack_count"] - 1): trials = big_bracket.current_trials() current_length = len(trials) for trl in trials: mock_runner._launch_trial(trl) # Provides results from 0 to 8 in order, keeping last one running for i, trl in enumerate(trials): action = sched.on_trial_result(mock_runner, trl, result(cur_units, i)) if i < current_length - 1: self.assertEqual(action, TrialScheduler.PAUSE) mock_runner.process_action(trl, action) self.assertEqual(action, TrialScheduler.CONTINUE) new_length = len(big_bracket.current_trials()) self.assertEqual(new_length, self.downscale(current_length, sched)) cur_units = int(cur_units * sched._eta) self.assertEqual(len(big_bracket.current_trials()), 1) def testHalvingStop(self): stats = self.default_statistics() num_trials = stats[str(0)]["n"] + stats[str(1)]["n"] sched, mock_runner = self.schedulerSetup(num_trials) big_bracket = sched._state["bracket"] for trl in big_bracket.current_trials(): mock_runner._launch_trial(trl) # # Provides result in reverse order, killing the last one cur_units = stats[str(1)]["r"] for i, trl in reversed(list(enumerate(big_bracket.current_trials()))): action = sched.on_trial_result(mock_runner, trl, result(cur_units, i)) mock_runner.process_action(trl, action) self.assertEqual(action, TrialScheduler.STOP) def testStopsLastOne(self): stats = self.default_statistics() num_trials = stats[str(0)]["n"] # setup one bracket sched, mock_runner = self.schedulerSetup(num_trials) big_bracket = sched._state["bracket"] for trl in big_bracket.current_trials(): mock_runner._launch_trial(trl) # # Provides result in reverse order, killing the last one cur_units = stats[str(0)]["r"] for i, trl in enumerate(big_bracket.current_trials()): action = sched.on_trial_result(mock_runner, trl, result(cur_units, i)) mock_runner.process_action(trl, action) self.assertEqual(action, TrialScheduler.STOP) def testTrialErrored(self): """If a trial errored, make sure successive halving still happens""" stats = self.default_statistics() trial_count = stats[str(0)]["n"] + 3 sched, mock_runner = self.schedulerSetup(trial_count) t1, t2, t3 = sched._state["bracket"].current_trials() for t in [t1, t2, t3]: mock_runner._launch_trial(t) sched.on_trial_error(mock_runner, t3) self.assertEqual( TrialScheduler.PAUSE, sched.on_trial_result(mock_runner, t1, result(stats[str(1)]["r"], 10))) self.assertEqual( TrialScheduler.CONTINUE, sched.on_trial_result(mock_runner, t2, result(stats[str(1)]["r"], 10))) def testTrialErrored2(self): """Check successive halving happened even when last trial failed""" stats = self.default_statistics() trial_count = stats[str(0)]["n"] + stats[str(1)]["n"] sched, mock_runner = self.schedulerSetup(trial_count) trials = sched._state["bracket"].current_trials() for t in trials[:-1]: mock_runner._launch_trial(t) sched.on_trial_result(mock_runner, t, result( stats[str(1)]["r"], 10)) mock_runner._launch_trial(trials[-1]) sched.on_trial_error(mock_runner, trials[-1]) self.assertEqual( len(sched._state["bracket"].current_trials()), self.downscale(stats[str(1)]["n"], sched)) def testTrialEndedEarly(self): """Check successive halving happened even when one trial failed""" stats = self.default_statistics() trial_count = stats[str(0)]["n"] + 3 sched, mock_runner = self.schedulerSetup(trial_count) t1, t2, t3 = sched._state["bracket"].current_trials() for t in [t1, t2, t3]: mock_runner._launch_trial(t) sched.on_trial_complete(mock_runner, t3, result(1, 12)) self.assertEqual( TrialScheduler.PAUSE, sched.on_trial_result(mock_runner, t1, result(stats[str(1)]["r"], 10))) self.assertEqual( TrialScheduler.CONTINUE, sched.on_trial_result(mock_runner, t2, result(stats[str(1)]["r"], 10))) def testTrialEndedEarly2(self): """Check successive halving happened even when last trial failed""" stats = self.default_statistics() trial_count = stats[str(0)]["n"] + stats[str(1)]["n"] sched, mock_runner = self.schedulerSetup(trial_count) trials = sched._state["bracket"].current_trials() for t in trials[:-1]: mock_runner._launch_trial(t) sched.on_trial_result(mock_runner, t, result( stats[str(1)]["r"], 10)) mock_runner._launch_trial(trials[-1]) sched.on_trial_complete(mock_runner, trials[-1], result(100, 12)) self.assertEqual( len(sched._state["bracket"].current_trials()), self.downscale(stats[str(1)]["n"], sched)) def testAddAfterHalving(self): stats = self.default_statistics() trial_count = stats[str(0)]["n"] + 1 sched, mock_runner = self.schedulerSetup(trial_count) bracket_trials = sched._state["bracket"].current_trials() init_units = stats[str(1)]["r"] for t in bracket_trials: mock_runner._launch_trial(t) for i, t in enumerate(bracket_trials): action = sched.on_trial_result(mock_runner, t, result( init_units, i)) self.assertEqual(action, TrialScheduler.CONTINUE) t = Trial("__fake") sched.on_trial_add(None, t) mock_runner._launch_trial(t) self.assertEqual(len(sched._state["bracket"].current_trials()), 2) # Make sure that newly added trial gets fair computation (not just 1) self.assertEqual( TrialScheduler.CONTINUE, sched.on_trial_result(mock_runner, t, result(init_units, 12))) new_units = init_units + int(init_units * sched._eta) self.assertEqual( TrialScheduler.PAUSE, sched.on_trial_result(mock_runner, t, result(new_units, 12))) def _test_metrics(self, result_func, metric, mode): sched = HyperBandScheduler( time_attr="time_total_s", metric=metric, mode=mode) stats = self.default_statistics() for i in range(stats["max_trials"]): t = Trial("__fake") sched.on_trial_add(None, t) runner = _MockTrialRunner(sched) big_bracket = sched._hyperbands[0][-1] for trl in big_bracket.current_trials(): runner._launch_trial(trl) current_length = len(big_bracket.current_trials()) # Provides results from 0 to 8 in order, keeping the last one running for i, trl in enumerate(big_bracket.current_trials()): action = sched.on_trial_result(runner, trl, result_func(1, i)) runner.process_action(trl, action) new_length = len(big_bracket.current_trials()) self.assertEqual(action, TrialScheduler.CONTINUE) self.assertEqual(new_length, self.downscale(current_length, sched)) def testAlternateMetrics(self): """Checking that alternate metrics will pass.""" def result2(t, rew): return dict(time_total_s=t, neg_mean_loss=rew) self._test_metrics(result2, "neg_mean_loss", "max") def testAlternateMetricsMin(self): """Checking that alternate metrics will pass.""" def result2(t, rew): return dict(time_total_s=t, mean_loss=-rew) self._test_metrics(result2, "mean_loss", "min") def testJumpingTime(self): sched, mock_runner = self.schedulerSetup(81) big_bracket = sched._hyperbands[0][-1] for trl in big_bracket.current_trials(): mock_runner._launch_trial(trl) # Provides results from 0 to 8 in order, keeping the last one running main_trials = big_bracket.current_trials()[:-1] jump = big_bracket.current_trials()[-1] for i, trl in enumerate(main_trials): action = sched.on_trial_result(mock_runner, trl, result(1, i)) mock_runner.process_action(trl, action) action = sched.on_trial_result(mock_runner, jump, result(4, i)) self.assertEqual(action, TrialScheduler.PAUSE) current_length = len(big_bracket.current_trials()) self.assertLess(current_length, 27) def testRemove(self): """Test with 4: start 1, remove 1 pending, add 2, remove 1 pending.""" sched, runner = self.schedulerSetup(4) trials = sorted(sched._trial_info, key=lambda t: t.trial_id) runner._launch_trial(trials[0]) sched.on_trial_result(runner, trials[0], result(1, 5)) self.assertEqual(trials[0].status, Trial.RUNNING) self.assertEqual(trials[1].status, Trial.PENDING) bracket, _ = sched._trial_info[trials[1]] self.assertTrue(trials[1] in bracket._live_trials) sched.on_trial_remove(runner, trials[1]) self.assertFalse(trials[1] in bracket._live_trials) for i in range(2): trial = Trial("__fake") sched.on_trial_add(None, trial) bracket, _ = sched._trial_info[trial] self.assertTrue(trial in bracket._live_trials) sched.on_trial_remove(runner, trial) # where trial is not running self.assertFalse(trial in bracket._live_trials) def testFilterNoneBracket(self): sched, runner = self.schedulerSetup(100, 20) # "sched" should contains None brackets non_brackets = [ b for hyperband in sched._hyperbands for b in hyperband if b is None ] self.assertTrue(non_brackets) # Make sure "choose_trial_to_run" still works trial = sched.choose_trial_to_run(runner) self.assertIsNotNone(trial) class BOHBSuite(unittest.TestCase): def setUp(self): ray.init(object_store_memory=int(1e8)) def tearDown(self): ray.shutdown() _register_all() # re-register the evicted objects def testLargestBracketFirst(self): sched = HyperBandForBOHB(max_t=3, reduction_factor=3) runner = _MockTrialRunner(sched) for i in range(3): t = Trial("__fake") sched.on_trial_add(runner, t) runner._launch_trial(t) self.assertEqual(sched.state()["num_brackets"], 1) sched.on_trial_add(runner, Trial("__fake")) self.assertEqual(sched.state()["num_brackets"], 2) def testCheckTrialInfoUpdate(self): def result(score, ts): return {"episode_reward_mean": score, TRAINING_ITERATION: ts} sched = HyperBandForBOHB(max_t=3, reduction_factor=3) runner = _MockTrialRunner(sched) runner._search_alg = MagicMock() trials = [Trial("__fake") for i in range(3)] for t in trials: runner.add_trial(t) runner._launch_trial(t) for trial, trial_result in zip(trials, [result(1, 1), result(2, 1)]): decision = sched.on_trial_result(runner, trial, trial_result) self.assertEqual(decision, TrialScheduler.PAUSE) runner._pause_trial(trial) spy_result = result(0, 1) decision = sched.on_trial_result(runner, trials[-1], spy_result) self.assertEqual(decision, TrialScheduler.STOP) sched.choose_trial_to_run(runner) self.assertEqual(runner._search_alg.on_pause.call_count, 2) self.assertEqual(runner._search_alg.on_unpause.call_count, 1) self.assertTrue("hyperband_info" in spy_result) self.assertEquals(spy_result["hyperband_info"]["budget"], 1) def testCheckTrialInfoUpdateMin(self): def result(score, ts): return {"episode_reward_mean": score, TRAINING_ITERATION: ts} sched = HyperBandForBOHB(max_t=3, reduction_factor=3, mode="min") runner = _MockTrialRunner(sched) runner._search_alg = MagicMock() trials = [Trial("__fake") for i in range(3)] for t in trials: runner.add_trial(t) runner._launch_trial(t) for trial, trial_result in zip(trials, [result(1, 1), result(2, 1)]): decision = sched.on_trial_result(runner, trial, trial_result) self.assertEqual(decision, TrialScheduler.PAUSE) runner._pause_trial(trial) spy_result = result(0, 1) decision = sched.on_trial_result(runner, trials[-1], spy_result) self.assertEqual(decision, TrialScheduler.CONTINUE) sched.choose_trial_to_run(runner) self.assertEqual(runner._search_alg.on_pause.call_count, 2) self.assertTrue("hyperband_info" in spy_result) self.assertEquals(spy_result["hyperband_info"]["budget"], 1) class _MockTrial(Trial): def __init__(self, i, config): self.trainable_name = "trial_{}".format(i) self.trial_id = Trial.generate_id() self.config = config self.experiment_tag = "{}tag".format(i) self.trial_name_creator = None self.logger_running = False self.restored_checkpoint = None self.resources = Resources(1, 0) self.custom_trial_name = None class PopulationBasedTestingSuite(unittest.TestCase): def setUp(self): ray.init() def tearDown(self): ray.shutdown() _register_all() # re-register the evicted objects def basicSetup(self, resample_prob=0.0, explore=None, perturbation_interval=10, log_config=False, step_once=True): pbt = PopulationBasedTraining( time_attr="training_iteration", perturbation_interval=perturbation_interval, resample_probability=resample_prob, quantile_fraction=0.25, hyperparam_mutations={ "id_factor": [100], "float_factor": lambda: 100.0, "int_factor": lambda: 10, }, custom_explore_fn=explore, log_config=log_config) runner = _MockTrialRunner(pbt) for i in range(5): trial = _MockTrial( i, { "id_factor": i, "float_factor": 2.0, "const_factor": 3, "int_factor": 10 }) runner.add_trial(trial) trial.status = Trial.RUNNING if step_once: self.assertEqual( pbt.on_trial_result(runner, trial, result(10, 50 * i)), TrialScheduler.CONTINUE) pbt.reset_stats() return pbt, runner def testCheckpointsMostPromisingTrials(self): pbt, runner = self.basicSetup() trials = runner.get_trials() # no checkpoint: haven't hit next perturbation interval yet self.assertEqual(pbt.last_scores(trials), [0, 50, 100, 150, 200]) self.assertEqual( pbt.on_trial_result(runner, trials[0], result(15, 200)), TrialScheduler.CONTINUE) self.assertEqual(pbt.last_scores(trials), [0, 50, 100, 150, 200]) self.assertEqual(pbt._num_checkpoints, 0) # checkpoint: both past interval and upper quantile self.assertEqual( pbt.on_trial_result(runner, trials[0], result(20, 200)), TrialScheduler.CONTINUE) self.assertEqual(pbt.last_scores(trials), [200, 50, 100, 150, 200]) self.assertEqual(pbt._num_checkpoints, 1) self.assertEqual( pbt.on_trial_result(runner, trials[1], result(30, 201)), TrialScheduler.CONTINUE) self.assertEqual(pbt.last_scores(trials), [200, 201, 100, 150, 200]) self.assertEqual(pbt._num_checkpoints, 2) # not upper quantile any more self.assertEqual( pbt.on_trial_result(runner, trials[4], result(30, 199)), TrialScheduler.CONTINUE) self.assertEqual(pbt._num_checkpoints, 2) self.assertEqual(pbt._num_perturbations, 0) def testPerturbsLowPerformingTrials(self): pbt, runner = self.basicSetup() trials = runner.get_trials() # no perturbation: haven't hit next perturbation interval self.assertEqual( pbt.on_trial_result(runner, trials[0], result(15, -100)), TrialScheduler.CONTINUE) self.assertEqual(pbt.last_scores(trials), [0, 50, 100, 150, 200]) self.assertTrue("@perturbed" not in trials[0].experiment_tag) self.assertEqual(pbt._num_perturbations, 0) # perturb since it's lower quantile self.assertEqual( pbt.on_trial_result(runner, trials[0], result(20, -100)), TrialScheduler.CONTINUE) self.assertEqual(pbt.last_scores(trials), [-100, 50, 100, 150, 200]) self.assertTrue("@perturbed" in trials[0].experiment_tag) self.assertIn(trials[0].restored_checkpoint, ["trial_3", "trial_4"]) self.assertEqual(pbt._num_perturbations, 1) # also perturbed self.assertEqual( pbt.on_trial_result(runner, trials[2], result(20, 40)), TrialScheduler.CONTINUE) self.assertEqual(pbt.last_scores(trials), [-100, 50, 40, 150, 200]) self.assertEqual(pbt._num_perturbations, 2) self.assertIn(trials[0].restored_checkpoint, ["trial_3", "trial_4"]) self.assertTrue("@perturbed" in trials[2].experiment_tag) def testPerturbWithoutResample(self): pbt, runner = self.basicSetup(resample_prob=0.0) trials = runner.get_trials() self.assertEqual( pbt.on_trial_result(runner, trials[0], result(20, -100)), TrialScheduler.CONTINUE) self.assertIn(trials[0].restored_checkpoint, ["trial_3", "trial_4"]) self.assertIn(trials[0].config["id_factor"], [100]) self.assertIn(trials[0].config["float_factor"], [2.4, 1.6]) self.assertEqual(type(trials[0].config["float_factor"]), float) self.assertIn(trials[0].config["int_factor"], [8, 12]) self.assertEqual(type(trials[0].config["int_factor"]), int) self.assertEqual(trials[0].config["const_factor"], 3) def testPerturbWithResample(self): pbt, runner = self.basicSetup(resample_prob=1.0) trials = runner.get_trials() self.assertEqual( pbt.on_trial_result(runner, trials[0], result(20, -100)), TrialScheduler.CONTINUE) self.assertIn(trials[0].restored_checkpoint, ["trial_3", "trial_4"]) self.assertEqual(trials[0].config["id_factor"], 100) self.assertEqual(trials[0].config["float_factor"], 100.0) self.assertEqual(type(trials[0].config["float_factor"]), float) self.assertEqual(trials[0].config["int_factor"], 10) self.assertEqual(type(trials[0].config["int_factor"]), int) self.assertEqual(trials[0].config["const_factor"], 3) def testPerturbationValues(self): def assertProduces(fn, values): random.seed(0) seen = set() for _ in range(100): seen.add(fn()["v"]) self.assertEqual(seen, values) # Categorical case assertProduces( lambda: explore({"v": 4}, {"v": [3, 4, 8, 10]}, 0.0, lambda x: x), {3, 8}) assertProduces( lambda: explore({"v": 3}, {"v": [3, 4, 8, 10]}, 0.0, lambda x: x), {3, 4}) assertProduces( lambda: explore({"v": 10}, {"v": [3, 4, 8, 10]}, 0.0, lambda x: x), {8, 10}) assertProduces( lambda: explore({"v": 7}, {"v": [3, 4, 8, 10]}, 0.0, lambda x: x), {3, 4, 8, 10}) assertProduces( lambda: explore({"v": 4}, {"v": [3, 4, 8, 10]}, 1.0, lambda x: x), {3, 4, 8, 10}) # Continuous case assertProduces( lambda: explore({"v": 100}, { "v": lambda: random.choice([10, 100]) }, 0.0, lambda x: x), {80, 120}) assertProduces( lambda: explore({"v": 100.0}, { "v": lambda: random.choice([10, 100]) }, 0.0, lambda x: x), {80.0, 120.0}) assertProduces( lambda: explore({"v": 100.0}, { "v": lambda: random.choice([10, 100]) }, 1.0, lambda x: x), {10.0, 100.0}) def deep_add(seen, new_values): for k, new_value in new_values.items(): if isinstance(new_value, dict): if k not in seen: seen[k] = {} seen[k].update(deep_add(seen[k], new_value)) else: if k not in seen: seen[k] = set() seen[k].add(new_value) return seen def assertNestedProduces(fn, values): random.seed(0) seen = {} for _ in range(100): new_config = fn() seen = deep_add(seen, new_config) self.assertEqual(seen, values) # Nested mutation and spec assertNestedProduces( lambda: explore({ "a": { "b": 4 }, "1": { "2": { "3": 100 } }, }, { "a": { "b": [3, 4, 8, 10] }, "1": { "2": { "3": lambda: random.choice([10, 100]) } }, }, 0.0, lambda x: x), { "a": { "b": {3, 8} }, "1": { "2": { "3": {80, 120} } }, }) custom_explore_fn = MagicMock(side_effect=lambda x: x) # Nested mutation and spec assertNestedProduces( lambda: explore({ "a": { "b": 4 }, "1": { "2": { "3": 100 } }, }, { "a": { "b": [3, 4, 8, 10] }, "1": { "2": { "3": lambda: random.choice([10, 100]) } }, }, 0.0, custom_explore_fn), { "a": { "b": {3, 8} }, "1": { "2": { "3": {80, 120} } }, }) # Expect call count to be 100 because we call explore 100 times self.assertEqual(custom_explore_fn.call_count, 100) def testYieldsTimeToOtherTrials(self): pbt, runner = self.basicSetup() trials = runner.get_trials() trials[0].status = Trial.PENDING # simulate not enough resources self.assertEqual( pbt.on_trial_result(runner, trials[1], result(20, 1000)), TrialScheduler.PAUSE) self.assertEqual(pbt.last_scores(trials), [0, 1000, 100, 150, 200]) self.assertEqual(pbt.choose_trial_to_run(runner), trials[0]) def testSchedulesMostBehindTrialToRun(self): pbt, runner = self.basicSetup() trials = runner.get_trials() pbt.on_trial_result(runner, trials[0], result(800, 1000)) pbt.on_trial_result(runner, trials[1], result(700, 1001)) pbt.on_trial_result(runner, trials[2], result(600, 1002)) pbt.on_trial_result(runner, trials[3], result(500, 1003)) pbt.on_trial_result(runner, trials[4], result(700, 1004)) self.assertEqual(pbt.choose_trial_to_run(runner), None) for i in range(5): trials[i].status = Trial.PENDING self.assertEqual(pbt.choose_trial_to_run(runner), trials[3]) def testPerturbationResetsLastPerturbTime(self): pbt, runner = self.basicSetup() trials = runner.get_trials() pbt.on_trial_result(runner, trials[0], result(10000, 1005)) pbt.on_trial_result(runner, trials[1], result(10000, 1004)) pbt.on_trial_result(runner, trials[2], result(600, 1003)) self.assertEqual(pbt._num_perturbations, 0) pbt.on_trial_result(runner, trials[3], result(500, 1002)) self.assertEqual(pbt._num_perturbations, 1) pbt.on_trial_result(runner, trials[3], result(600, 100)) self.assertEqual(pbt._num_perturbations, 1) pbt.on_trial_result(runner, trials[3], result(11000, 100)) self.assertEqual(pbt._num_perturbations, 2) def testLogConfig(self): def check_policy(policy): self.assertIsInstance(policy[2], int) self.assertIsInstance(policy[3], int) self.assertIn(policy[0], ["0tag", "2tag", "3tag", "4tag"]) self.assertIn(policy[1], ["0tag", "2tag", "3tag", "4tag"]) self.assertIn(policy[2], [0, 2, 3, 4]) self.assertIn(policy[3], [0, 2, 3, 4]) for i in [4, 5]: self.assertIsInstance(policy[i], dict) for key in [ "const_factor", "int_factor", "float_factor", "id_factor" ]: self.assertIn(key, policy[i]) self.assertIsInstance(policy[i]["float_factor"], float) self.assertIsInstance(policy[i]["int_factor"], int) self.assertIn(policy[i]["const_factor"], [3]) self.assertIn(policy[i]["int_factor"], [8, 10, 12]) self.assertIn(policy[i]["float_factor"], [2.4, 2, 1.6]) self.assertIn(policy[i]["id_factor"], [3, 4, 100]) pbt, runner = self.basicSetup(log_config=True) trials = runner.get_trials() tmpdir = tempfile.mkdtemp() for i, trial in enumerate(trials): trial.local_dir = tmpdir trial.last_result = {TRAINING_ITERATION: i} pbt.on_trial_result(runner, trials[0], result(15, -100)) pbt.on_trial_result(runner, trials[0], result(20, -100)) pbt.on_trial_result(runner, trials[2], result(20, 40)) log_files = ["pbt_global.txt", "pbt_policy_0.txt", "pbt_policy_2.txt"] for log_file in log_files: self.assertTrue(os.path.exists(os.path.join(tmpdir, log_file))) raw_policy = open(os.path.join(tmpdir, log_file), "r").readlines() for line in raw_policy: check_policy(json.loads(line)) shutil.rmtree(tmpdir) def testPostprocessingHook(self): def explore(new_config): new_config["id_factor"] = 42 new_config["float_factor"] = 43 return new_config pbt, runner = self.basicSetup(resample_prob=0.0, explore=explore) trials = runner.get_trials() self.assertEqual( pbt.on_trial_result(runner, trials[0], result(20, -100)), TrialScheduler.CONTINUE) self.assertEqual(trials[0].config["id_factor"], 42) self.assertEqual(trials[0].config["float_factor"], 43) def testFastPerturb(self): pbt, runner = self.basicSetup( perturbation_interval=1, step_once=False, log_config=True) trials = runner.get_trials() tmpdir = tempfile.mkdtemp() for i, trial in enumerate(trials): trial.local_dir = tmpdir trial.last_result = {} pbt.on_trial_result(runner, trials[0], result(1, 10)) self.assertEqual( pbt.on_trial_result(runner, trials[2], result(1, 200)), TrialScheduler.CONTINUE) self.assertEqual(pbt._num_checkpoints, 1) pbt._exploit(runner.trial_executor, trials[1], trials[2]) shutil.rmtree(tmpdir) class AsyncHyperBandSuite(unittest.TestCase): def setUp(self): ray.init() def tearDown(self): ray.shutdown() _register_all() # re-register the evicted objects def basicSetup(self, scheduler): t1 = Trial("PPO") # mean is 450, max 900, t_max=10 t2 = Trial("PPO") # mean is 450, max 450, t_max=5 scheduler.on_trial_add(None, t1) scheduler.on_trial_add(None, t2) for i in range(10): self.assertEqual( scheduler.on_trial_result(None, t1, result(i, i * 100)), TrialScheduler.CONTINUE) for i in range(5): self.assertEqual( scheduler.on_trial_result(None, t2, result(i, 450)), TrialScheduler.CONTINUE) return t1, t2 def testAsyncHBOnComplete(self): scheduler = AsyncHyperBandScheduler(max_t=10, brackets=1) t1, t2 = self.basicSetup(scheduler) t3 = Trial("PPO") scheduler.on_trial_add(None, t3) scheduler.on_trial_complete(None, t3, result(10, 1000)) self.assertEqual( scheduler.on_trial_result(None, t2, result(101, 0)), TrialScheduler.STOP) def testAsyncHBGracePeriod(self): scheduler = AsyncHyperBandScheduler( grace_period=2.5, reduction_factor=3, brackets=1) t1, t2 = self.basicSetup(scheduler) scheduler.on_trial_complete(None, t1, result(10, 1000)) scheduler.on_trial_complete(None, t2, result(10, 1000)) t3 = Trial("PPO") scheduler.on_trial_add(None, t3) self.assertEqual( scheduler.on_trial_result(None, t3, result(1, 10)), TrialScheduler.CONTINUE) self.assertEqual( scheduler.on_trial_result(None, t3, result(2, 10)), TrialScheduler.CONTINUE) self.assertEqual( scheduler.on_trial_result(None, t3, result(3, 10)), TrialScheduler.STOP) def testAsyncHBAllCompletes(self): scheduler = AsyncHyperBandScheduler(max_t=10, brackets=10) trials = [Trial("PPO") for i in range(10)] for t in trials: scheduler.on_trial_add(None, t) for t in trials: self.assertEqual( scheduler.on_trial_result(None, t, result(10, -2)), TrialScheduler.STOP) def testAsyncHBUsesPercentile(self): scheduler = AsyncHyperBandScheduler( grace_period=1, max_t=10, reduction_factor=2, brackets=1) t1, t2 = self.basicSetup(scheduler) scheduler.on_trial_complete(None, t1, result(10, 1000)) scheduler.on_trial_complete(None, t2, result(10, 1000)) t3 = Trial("PPO") scheduler.on_trial_add(None, t3) self.assertEqual( scheduler.on_trial_result(None, t3, result(1, 260)), TrialScheduler.STOP) self.assertEqual( scheduler.on_trial_result(None, t3, result(2, 260)), TrialScheduler.STOP) def _test_metrics(self, result_func, metric, mode): scheduler = AsyncHyperBandScheduler( grace_period=1, time_attr="training_iteration", metric=metric, mode=mode, brackets=1) t1 = Trial("PPO") # mean is 450, max 900, t_max=10 t2 = Trial("PPO") # mean is 450, max 450, t_max=5 scheduler.on_trial_add(None, t1) scheduler.on_trial_add(None, t2) for i in range(10): self.assertEqual( scheduler.on_trial_result(None, t1, result_func(i, i * 100)), TrialScheduler.CONTINUE) for i in range(5): self.assertEqual( scheduler.on_trial_result(None, t2, result_func(i, 450)), TrialScheduler.CONTINUE) scheduler.on_trial_complete(None, t1, result_func(10, 1000)) self.assertEqual( scheduler.on_trial_result(None, t2, result_func(5, 450)), TrialScheduler.CONTINUE) self.assertEqual( scheduler.on_trial_result(None, t2, result_func(6, 0)), TrialScheduler.CONTINUE) def testAlternateMetrics(self): def result2(t, rew): return dict(training_iteration=t, neg_mean_loss=rew) self._test_metrics(result2, "neg_mean_loss", "max") def testAlternateMetricsMin(self): def result2(t, rew): return dict(training_iteration=t, mean_loss=-rew) self._test_metrics(result2, "mean_loss", "min") if __name__ == "__main__": import pytest sys.exit(pytest.main(["-v", __file__]))
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/test_tune_restore.py
Python
# coding: utf-8 import os import shutil import tempfile import unittest import skopt import numpy as np from hyperopt import hp from nevergrad.optimization import optimizerlib import ray from ray import tune from ray.test_utils import recursive_fnmatch from ray.rllib import _register_all from ray.tune.suggest.hyperopt import HyperOptSearch from ray.tune.suggest.bayesopt import BayesOptSearch from ray.tune.suggest.skopt import SkOptSearch from ray.tune.suggest.nevergrad import NevergradSearch from ray.tune.suggest.sigopt import SigOptSearch from ray.tune.utils import validate_save_restore class TuneRestoreTest(unittest.TestCase): def setUp(self): ray.init(num_cpus=1, num_gpus=0, local_mode=True) tmpdir = tempfile.mkdtemp() test_name = "TuneRestoreTest" tune.run( "PG", name=test_name, stop={"training_iteration": 1}, checkpoint_freq=1, local_dir=tmpdir, config={ "env": "CartPole-v0", }, ) logdir = os.path.expanduser(os.path.join(tmpdir, test_name)) self.logdir = logdir self.checkpoint_path = recursive_fnmatch(logdir, "checkpoint-1")[0] def tearDown(self): shutil.rmtree(self.logdir) ray.shutdown() _register_all() def testTuneRestore(self): self.assertTrue(os.path.isfile(self.checkpoint_path)) tune.run( "PG", name="TuneRestoreTest", stop={"training_iteration": 2}, # train one more iteration. checkpoint_freq=1, restore=self.checkpoint_path, # Restore the checkpoint config={ "env": "CartPole-v0", }, ) class TuneExampleTest(unittest.TestCase): def setUp(self): ray.init() def tearDown(self): ray.shutdown() _register_all() def testPBTKeras(self): from ray.tune.examples.pbt_tune_cifar10_with_keras import Cifar10Model from tensorflow.python.keras.datasets import cifar10 cifar10.load_data() validate_save_restore(Cifar10Model) validate_save_restore(Cifar10Model, use_object_store=True) def testPyTorchMNIST(self): from ray.tune.examples.mnist_pytorch_trainable import TrainMNIST from torchvision import datasets datasets.MNIST("~/data", train=True, download=True) validate_save_restore(TrainMNIST) validate_save_restore(TrainMNIST, use_object_store=True) def testLogging(self): from ray.tune.examples.logging_example import MyTrainableClass validate_save_restore(MyTrainableClass) validate_save_restore(MyTrainableClass, use_object_store=True) def testHyperbandExample(self): from ray.tune.examples.hyperband_example import MyTrainableClass validate_save_restore(MyTrainableClass) validate_save_restore(MyTrainableClass, use_object_store=True) def testAsyncHyperbandExample(self): from ray.tune.examples.async_hyperband_example import MyTrainableClass validate_save_restore(MyTrainableClass) validate_save_restore(MyTrainableClass, use_object_store=True) class AutoInitTest(unittest.TestCase): def testTuneRestore(self): self.assertFalse(ray.is_initialized()) tune.run( "__fake", name="TestAutoInit", stop={"training_iteration": 1}, ray_auto_init=True) self.assertTrue(ray.is_initialized()) def tearDown(self): ray.shutdown() _register_all() class AbstractWarmStartTest: def setUp(self): ray.init(local_mode=True) self.tmpdir = tempfile.mkdtemp() def tearDown(self): shutil.rmtree(self.tmpdir) ray.shutdown() _register_all() def set_basic_conf(self): raise NotImplementedError() def run_exp_1(self): np.random.seed(162) search_alg, cost = self.set_basic_conf() results_exp_1 = tune.run(cost, num_samples=15, search_alg=search_alg) self.log_dir = os.path.join(self.tmpdir, "warmStartTest.pkl") search_alg.save(self.log_dir) return results_exp_1 def run_exp_2(self): search_alg2, cost = self.set_basic_conf() search_alg2.restore(self.log_dir) return tune.run(cost, num_samples=15, search_alg=search_alg2) def run_exp_3(self): np.random.seed(162) search_alg3, cost = self.set_basic_conf() return tune.run(cost, num_samples=30, search_alg=search_alg3) def testWarmStart(self): results_exp_1 = self.run_exp_1() results_exp_2 = self.run_exp_2() results_exp_3 = self.run_exp_3() trials_1_config = [trial.config for trial in results_exp_1.trials] trials_2_config = [trial.config for trial in results_exp_2.trials] trials_3_config = [trial.config for trial in results_exp_3.trials] self.assertEqual(trials_1_config + trials_2_config, trials_3_config) class HyperoptWarmStartTest(AbstractWarmStartTest, unittest.TestCase): def set_basic_conf(self): space = { "x": hp.uniform("x", 0, 10), "y": hp.uniform("y", -10, 10), "z": hp.uniform("z", -10, 0) } def cost(space, reporter): loss = space["x"]**2 + space["y"]**2 + space["z"]**2 reporter(loss=loss) search_alg = HyperOptSearch( space, max_concurrent=1, metric="loss", mode="min", random_state_seed=5) return search_alg, cost class BayesoptWarmStartTest(AbstractWarmStartTest, unittest.TestCase): def set_basic_conf(self): space = {"width": (0, 20), "height": (-100, 100)} def cost(space, reporter): reporter(loss=(space["height"] - 14)**2 - abs(space["width"] - 3)) search_alg = BayesOptSearch( space, max_concurrent=1, metric="loss", mode="min", utility_kwargs={ "kind": "ucb", "kappa": 2.5, "xi": 0.0 }) return search_alg, cost class SkoptWarmStartTest(AbstractWarmStartTest, unittest.TestCase): def set_basic_conf(self): optimizer = skopt.Optimizer([(0, 20), (-100, 100)]) previously_run_params = [[10, 0], [15, -20]] known_rewards = [-189, -1144] def cost(space, reporter): reporter(loss=(space["height"]**2 + space["width"]**2)) search_alg = SkOptSearch( optimizer, ["width", "height"], max_concurrent=1, metric="loss", mode="min", points_to_evaluate=previously_run_params, evaluated_rewards=known_rewards) return search_alg, cost class NevergradWarmStartTest(AbstractWarmStartTest, unittest.TestCase): def set_basic_conf(self): instrumentation = 2 parameter_names = ["height", "width"] optimizer = optimizerlib.OnePlusOne(instrumentation) def cost(space, reporter): reporter( mean_loss=(space["height"] - 14)**2 - abs(space["width"] - 3)) search_alg = NevergradSearch( optimizer, parameter_names, max_concurrent=1, metric="mean_loss", mode="min") return search_alg, cost class SigOptWarmStartTest(AbstractWarmStartTest, unittest.TestCase): def set_basic_conf(self): space = [ { "name": "width", "type": "int", "bounds": { "min": 0, "max": 20 }, }, { "name": "height", "type": "int", "bounds": { "min": -100, "max": 100 }, }, ] def cost(space, reporter): reporter( mean_loss=(space["height"] - 14)**2 - abs(space["width"] - 3)) search_alg = SigOptSearch( space, name="SigOpt Example Experiment", max_concurrent=1, metric="mean_loss", mode="min") return search_alg, cost def testWarmStart(self): if ("SIGOPT_KEY" not in os.environ): return super().testWarmStart() if __name__ == "__main__": import pytest import sys sys.exit(pytest.main(["-v", __file__]))
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/test_tune_save_restore.py
Python
# coding: utf-8 import os import pickle import shutil import tempfile import unittest import ray from ray import tune from ray.rllib import _register_all from ray.tune import Trainable class SerialTuneRelativeLocalDirTest(unittest.TestCase): local_mode = True prefix = "Serial" class MockTrainable(Trainable): _name = "MockTrainable" def _setup(self, config): self.state = {"hi": 1} def _train(self): return {"timesteps_this_iter": 1, "done": True} def _save(self, checkpoint_dir): checkpoint_path = os.path.join( checkpoint_dir, "checkpoint-{}".format(self._iteration)) with open(checkpoint_path, "wb") as f: pickle.dump(self.state, f) return checkpoint_path def _restore(self, checkpoint_path): with open(checkpoint_path, "rb") as f: extra_data = pickle.load(f) self.state.update(extra_data) def setUp(self): ray.init(num_cpus=1, num_gpus=0, local_mode=self.local_mode) def tearDown(self): shutil.rmtree(self.absolute_local_dir, ignore_errors=True) self.absolute_local_dir = None ray.shutdown() # Without this line, test_tune_server.testAddTrial would fail. _register_all() def _get_trial_dir(self, absoulte_exp_dir): trial_dirname = next( (child_dir for child_dir in os.listdir(absoulte_exp_dir) if (os.path.isdir(os.path.join(absoulte_exp_dir, child_dir)) and child_dir.startswith(self.MockTrainable._name)))) trial_absolute_dir = os.path.join(absoulte_exp_dir, trial_dirname) return trial_dirname, trial_absolute_dir def _train(self, exp_name, local_dir, absolute_local_dir): trial, = tune.run( self.MockTrainable, name=exp_name, stop={ "training_iteration": 1 }, checkpoint_freq=1, local_dir=local_dir, config={ "env": "CartPole-v0", "log_level": "DEBUG" }).trials exp_dir = os.path.join(absolute_local_dir, exp_name) _, abs_trial_dir = self._get_trial_dir(exp_dir) self.assertIsNone(trial.error_file) self.assertEqual(trial.local_dir, exp_dir) self.assertEqual(trial.logdir, abs_trial_dir) self.assertTrue(os.path.isdir(absolute_local_dir), absolute_local_dir) self.assertTrue(os.path.isdir(exp_dir)) self.assertTrue(os.path.isdir(abs_trial_dir)) self.assertTrue( os.path.isfile( os.path.join(abs_trial_dir, "checkpoint_1/checkpoint-1"))) def _restore(self, exp_name, local_dir, absolute_local_dir): trial_name, abs_trial_dir = self._get_trial_dir( os.path.join(absolute_local_dir, exp_name)) checkpoint_path = os.path.join( local_dir, exp_name, trial_name, "checkpoint_1/checkpoint-1") # Relative checkpoint path # The file tune would find. The absolute checkpoint path. tune_find_file = os.path.abspath(os.path.expanduser(checkpoint_path)) self.assertTrue( os.path.isfile(tune_find_file), "{} is not exist!".format(tune_find_file)) trial, = tune.run( self.MockTrainable, name=exp_name, stop={ "training_iteration": 2 }, # train one more iteration. restore=checkpoint_path, # Restore the checkpoint config={ "env": "CartPole-v0", "log_level": "DEBUG" }).trials self.assertIsNone(trial.error_file) def testDottedRelativePath(self): local_dir = "./test_dotted_relative_local_dir" exp_name = self.prefix + "DottedRelativeLocalDir" absolute_local_dir = os.path.abspath(local_dir) self.absolute_local_dir = absolute_local_dir self.assertFalse(os.path.exists(absolute_local_dir)) self._train(exp_name, local_dir, absolute_local_dir) self._restore(exp_name, local_dir, absolute_local_dir) def testRelativePath(self): local_dir = "test_relative_local_dir" exp_name = self.prefix + "RelativePath" absolute_local_dir = os.path.abspath(local_dir) self.absolute_local_dir = absolute_local_dir self.assertFalse(os.path.exists(absolute_local_dir)) self._train(exp_name, local_dir, absolute_local_dir) self._restore(exp_name, local_dir, absolute_local_dir) def testTildeAbsolutePath(self): local_dir = "~/test_tilde_absolute_local_dir" exp_name = self.prefix + "TildeAbsolutePath" absolute_local_dir = os.path.abspath(os.path.expanduser(local_dir)) self.absolute_local_dir = absolute_local_dir self.assertFalse(os.path.exists(absolute_local_dir)) self._train(exp_name, local_dir, absolute_local_dir) self._restore(exp_name, local_dir, absolute_local_dir) def testTempfile(self): local_dir = tempfile.mkdtemp() exp_name = self.prefix + "Tempfile" self.absolute_local_dir = local_dir self._train(exp_name, local_dir, local_dir) self._restore(exp_name, local_dir, local_dir) if __name__ == "__main__": import pytest import sys sys.exit(pytest.main(["-v", __file__]))
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/test_tune_server.py
Python
import unittest import socket import subprocess import json import ray from ray.rllib import _register_all from ray.tune.trial import Trial, Resources from ray.tune.web_server import TuneClient from ray.tune.trial_runner import TrialRunner def get_valid_port(): port = 4321 while True: try: print("Trying port", port) port_test_socket = socket.socket() port_test_socket.bind(("127.0.0.1", port)) port_test_socket.close() break except socket.error: port += 1 return port class TuneServerSuite(unittest.TestCase): def basicSetup(self): ray.init(num_cpus=4, num_gpus=1) port = get_valid_port() self.runner = TrialRunner(launch_web_server=True, server_port=port) runner = self.runner kwargs = { "stopping_criterion": { "training_iteration": 3 }, "resources": Resources(cpu=1, gpu=1), } trials = [Trial("__fake", **kwargs), Trial("__fake", **kwargs)] for t in trials: runner.add_trial(t) client = TuneClient("localhost", port) return runner, client def tearDown(self): print("Tearing down....") try: self.runner._server.shutdown() self.runner = None except Exception as e: print(e) ray.shutdown() _register_all() def testAddTrial(self): runner, client = self.basicSetup() for i in range(3): runner.step() spec = { "run": "__fake", "stop": { "training_iteration": 3 }, "resources_per_trial": { "cpu": 1, "gpu": 1 }, } client.add_trial("test", spec) runner.step() all_trials = client.get_all_trials()["trials"] runner.step() self.assertEqual(len(all_trials), 3) def testGetTrials(self): runner, client = self.basicSetup() for i in range(3): runner.step() all_trials = client.get_all_trials()["trials"] self.assertEqual(len(all_trials), 2) tid = all_trials[0]["id"] client.get_trial(tid) runner.step() self.assertEqual(len(all_trials), 2) def testGetTrialsWithFunction(self): runner, client = self.basicSetup() test_trial = Trial( "__fake", trial_id="function_trial", stopping_criterion={"training_iteration": 3}, config={"callbacks": { "on_episode_start": lambda x: None }}) runner.add_trial(test_trial) for i in range(3): runner.step() all_trials = client.get_all_trials()["trials"] self.assertEqual(len(all_trials), 3) client.get_trial("function_trial") runner.step() self.assertEqual(len(all_trials), 3) def testStopTrial(self): """Check if Stop Trial works.""" runner, client = self.basicSetup() for i in range(2): runner.step() all_trials = client.get_all_trials()["trials"] self.assertEqual( len([t for t in all_trials if t["status"] == Trial.RUNNING]), 1) tid = [t for t in all_trials if t["status"] == Trial.RUNNING][0]["id"] client.stop_trial(tid) runner.step() all_trials = client.get_all_trials()["trials"] self.assertEqual( len([t for t in all_trials if t["status"] == Trial.RUNNING]), 0) def testCurlCommand(self): """Check if Stop Trial works.""" runner, client = self.basicSetup() for i in range(2): runner.step() stdout = subprocess.check_output( "curl \"http://{}:{}/trials\"".format(client.server_address, client.server_port), shell=True) self.assertNotEqual(stdout, None) curl_trials = json.loads(stdout.decode())["trials"] client_trials = client.get_all_trials()["trials"] for curl_trial, client_trial in zip(curl_trials, client_trials): self.assertEqual(curl_trial.keys(), client_trial.keys()) self.assertEqual(curl_trial["id"], client_trial["id"]) self.assertEqual(curl_trial["trainable_name"], client_trial["trainable_name"]) self.assertEqual(curl_trial["status"], client_trial["status"]) if __name__ == "__main__": import pytest import sys sys.exit(pytest.main(["-v", __file__]))
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/test_var.py
Python
import os import numpy as np import unittest import ray from ray.rllib import _register_all from ray import tune from ray.tune.result import DEFAULT_RESULTS_DIR from ray.tune.experiment import Experiment from ray.tune.suggest import grid_search, BasicVariantGenerator from ray.tune.suggest.suggestion import _MockSuggestionAlgorithm from ray.tune.suggest.variant_generator import (RecursiveDependencyError, resolve_nested_dict) class VariantGeneratorTest(unittest.TestCase): def setUp(self): ray.init() def tearDown(self): ray.shutdown() _register_all() # re-register the evicted objects def generate_trials(self, spec, name): suggester = BasicVariantGenerator() suggester.add_configurations({name: spec}) return suggester.next_trials() def testParseToTrials(self): trials = self.generate_trials({ "run": "PPO", "num_samples": 2, "max_failures": 5, "config": { "env": "Pong-v0", "foo": "bar" }, }, "tune-pong") trials = list(trials) self.assertEqual(len(trials), 2) self.assertTrue("PPO_Pong-v0" in str(trials[0])) self.assertEqual(trials[0].config, {"foo": "bar", "env": "Pong-v0"}) self.assertEqual(trials[0].trainable_name, "PPO") self.assertEqual(trials[0].experiment_tag, "0") self.assertEqual(trials[0].max_failures, 5) self.assertEqual(trials[0].evaluated_params, {}) self.assertEqual(trials[0].local_dir, os.path.join(DEFAULT_RESULTS_DIR, "tune-pong")) self.assertEqual(trials[1].experiment_tag, "1") def testEval(self): trials = self.generate_trials({ "run": "PPO", "config": { "foo": { "eval": "2 + 2" }, }, }, "eval") trials = list(trials) self.assertEqual(len(trials), 1) self.assertEqual(trials[0].config, {"foo": 4}) self.assertEqual(trials[0].evaluated_params, {"foo": 4}) self.assertEqual(trials[0].experiment_tag, "0_foo=4") def testGridSearch(self): trials = self.generate_trials({ "run": "PPO", "config": { "bar": { "grid_search": [True, False] }, "foo": { "grid_search": [1, 2, 3] }, "baz": "asd", }, }, "grid_search") trials = list(trials) self.assertEqual(len(trials), 6) self.assertEqual(trials[0].config, { "bar": True, "foo": 1, "baz": "asd", }) self.assertEqual(trials[0].evaluated_params, { "bar": True, "foo": 1, }) self.assertEqual(trials[0].experiment_tag, "0_bar=True,foo=1") self.assertEqual(trials[1].config, { "bar": False, "foo": 1, "baz": "asd", }) self.assertEqual(trials[1].evaluated_params, { "bar": False, "foo": 1, }) self.assertEqual(trials[1].experiment_tag, "1_bar=False,foo=1") self.assertEqual(trials[2].config, { "bar": True, "foo": 2, "baz": "asd", }) self.assertEqual(trials[2].evaluated_params, { "bar": True, "foo": 2, }) self.assertEqual(trials[3].config, { "bar": False, "foo": 2, "baz": "asd", }) self.assertEqual(trials[3].evaluated_params, { "bar": False, "foo": 2, }) self.assertEqual(trials[4].config, { "bar": True, "foo": 3, "baz": "asd", }) self.assertEqual(trials[4].evaluated_params, { "bar": True, "foo": 3, }) self.assertEqual(trials[5].config, { "bar": False, "foo": 3, "baz": "asd", }) self.assertEqual(trials[5].evaluated_params, { "bar": False, "foo": 3, }) def testGridSearchAndEval(self): trials = self.generate_trials({ "run": "PPO", "config": { "qux": tune.sample_from(lambda spec: 2 + 2), "bar": grid_search([True, False]), "foo": grid_search([1, 2, 3]), "baz": "asd", }, }, "grid_eval") trials = list(trials) self.assertEqual(len(trials), 6) self.assertEqual(trials[0].config, { "bar": True, "foo": 1, "qux": 4, "baz": "asd", }) self.assertEqual(trials[0].evaluated_params, { "bar": True, "foo": 1, "qux": 4, }) self.assertEqual(trials[0].experiment_tag, "0_bar=True,foo=1,qux=4") def testConditionResolution(self): trials = self.generate_trials({ "run": "PPO", "config": { "x": 1, "y": tune.sample_from(lambda spec: spec.config.x + 1), "z": tune.sample_from(lambda spec: spec.config.y + 1), }, }, "condition_resolution") trials = list(trials) self.assertEqual(len(trials), 1) self.assertEqual(trials[0].config, {"x": 1, "y": 2, "z": 3}) self.assertEqual(trials[0].evaluated_params, {"y": 2, "z": 3}) self.assertEqual(trials[0].experiment_tag, "0_y=2,z=3") def testDependentLambda(self): trials = self.generate_trials({ "run": "PPO", "config": { "x": grid_search([1, 2]), "y": tune.sample_from(lambda spec: spec.config.x * 100), }, }, "dependent_lambda") trials = list(trials) self.assertEqual(len(trials), 2) self.assertEqual(trials[0].config, {"x": 1, "y": 100}) self.assertEqual(trials[1].config, {"x": 2, "y": 200}) def testDependentGridSearch(self): trials = self.generate_trials({ "run": "PPO", "config": { "x": grid_search([ tune.sample_from(lambda spec: spec.config.y * 100), tune.sample_from(lambda spec: spec.config.y * 200) ]), "y": tune.sample_from(lambda spec: 1), }, }, "dependent_grid_search") trials = list(trials) self.assertEqual(len(trials), 2) self.assertEqual(trials[0].config, {"x": 100, "y": 1}) self.assertEqual(trials[1].config, {"x": 200, "y": 1}) def testNestedValues(self): trials = self.generate_trials({ "run": "PPO", "config": { "x": { "y": { "z": tune.sample_from(lambda spec: 1) } }, "y": tune.sample_from(lambda spec: 12), "z": tune.sample_from(lambda spec: spec.config.x.y.z * 100), }, }, "nested_values") trials = list(trials) self.assertEqual(len(trials), 1) self.assertEqual(trials[0].config, { "x": { "y": { "z": 1 } }, "y": 12, "z": 100 }) self.assertEqual(trials[0].evaluated_params, { "x/y/z": 1, "y": 12, "z": 100 }) def testLogUniform(self): sampler = tune.loguniform(1e-10, 1e-1).func results = [sampler(None) for i in range(1000)] assert abs(np.log(min(results)) / np.log(10) - -10) < 0.1 assert abs(np.log(max(results)) / np.log(10) - -1) < 0.1 sampler_e = tune.loguniform(np.e**-4, np.e, base=np.e).func results_e = [sampler_e(None) for i in range(1000)] assert abs(np.log(min(results_e)) - -4) < 0.1 assert abs(np.log(max(results_e)) - 1) < 0.1 def test_resolve_dict(self): config = { "a": { "b": 1, "c": 2, }, "b": { "a": 3 } } resolved = resolve_nested_dict(config) for k, v in [(("a", "b"), 1), (("a", "c"), 2), (("b", "a"), 3)]: self.assertEqual(resolved.get(k), v) def testRecursiveDep(self): try: list( self.generate_trials({ "run": "PPO", "config": { "foo": tune.sample_from(lambda spec: spec.config.foo), }, }, "recursive_dep")) except RecursiveDependencyError as e: assert "`foo` recursively depends on" in str(e), e else: assert False def testMaxConcurrentSuggestions(self): """Checks that next_trials() supports throttling.""" experiment_spec = { "run": "PPO", "num_samples": 6, } experiments = [Experiment.from_json("test", experiment_spec)] searcher = _MockSuggestionAlgorithm(max_concurrent=4) searcher.add_configurations(experiments) trials = searcher.next_trials() self.assertEqual(len(trials), 4) self.assertEqual(searcher.next_trials(), []) finished_trial = trials.pop() searcher.on_trial_complete(finished_trial.trial_id) self.assertEqual(len(searcher.next_trials()), 1) finished_trial = trials.pop() searcher.on_trial_complete(finished_trial.trial_id) finished_trial = trials.pop() searcher.on_trial_complete(finished_trial.trial_id) finished_trial = trials.pop() searcher.on_trial_complete(finished_trial.trial_id) self.assertEqual(len(searcher.next_trials()), 1) self.assertEqual(len(searcher.next_trials()), 0) if __name__ == "__main__": import pytest import sys sys.exit(pytest.main(["-v", __file__]))
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tests/tutorial.py
Python
# flake8: noqa # Original Code: https://github.com/pytorch/examples/blob/master/mnist/main.py # yapf: disable # __tutorial_imports_begin__ import numpy as np import torch import torch.optim as optim from torchvision import datasets from ray import tune from ray.tune import track from ray.tune.schedulers import ASHAScheduler from ray.tune.examples.mnist_pytorch import get_data_loaders, ConvNet, train, test # __tutorial_imports_end__ # yapf: enable # yapf: disable # __train_func_begin__ def train_mnist(config): model = ConvNet() train_loader, test_loader = get_data_loaders() optimizer = optim.SGD( model.parameters(), lr=config["lr"], momentum=config["momentum"]) for i in range(10): train(model, optimizer, train_loader) acc = test(model, test_loader) track.log(mean_accuracy=acc) if i % 5 == 0: # This saves the model to the trial directory torch.save(model, "./model.pth") # __train_func_end__ # yapf: enable # __eval_func_begin__ search_space = { "lr": tune.sample_from(lambda spec: 10**(-10 * np.random.rand())), "momentum": tune.uniform(0.1, 0.9) } # Uncomment this to enable distributed execution # `ray.init(address=...)` analysis = tune.run(train_mnist, config=search_space) # __eval_func_end__ #__plot_begin__ dfs = analysis.trial_dataframes [d.mean_accuracy.plot() for d in dfs.values()] #__plot_end__ # __run_scheduler_begin__ analysis = tune.run( train_mnist, num_samples=30, scheduler=ASHAScheduler(metric="mean_accuracy", mode="max"), config=search_space) # Obtain a trial dataframe from all run trials of this `tune.run` call. dfs = analysis.trial_dataframes # __run_scheduler_end__ # yapf: disable # __plot_scheduler_begin__ # Plot by epoch ax = None # This plots everything on the same plot for d in dfs.values(): ax = d.mean_accuracy.plot(ax=ax, legend=False) # __plot_scheduler_end__ # yapf: enable # __run_searchalg_begin__ from hyperopt import hp from ray.tune.suggest.hyperopt import HyperOptSearch space = { "lr": hp.loguniform("lr", 1e-10, 0.1), "momentum": hp.uniform("momentum", 0.1, 0.9), } hyperopt_search = HyperOptSearch( space, max_concurrent=2, reward_attr="mean_accuracy") analysis = tune.run(train_mnist, num_samples=10, search_alg=hyperopt_search) # __run_searchalg_end__ # __run_analysis_begin__ import os df = analysis.dataframe() logdir = analysis.get_best_logdir("mean_accuracy", mode="max") model = torch.load(os.path.join(logdir, "model.pth")) # __run_analysis_end__ from ray.tune.examples.mnist_pytorch_trainable import TrainMNIST # __trainable_run_begin__ search_space = { "lr": tune.sample_from(lambda spec: 10**(-10 * np.random.rand())), "momentum": tune.uniform(0.1, 0.9) } analysis = tune.run( TrainMNIST, config=search_space, stop={"training_iteration": 10}) # __trainable_run_end__
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/track/__init__.py
Python
import logging from ray.tune.track.session import TrackSession logger = logging.getLogger(__name__) _session = None def get_session(): global _session if not _session: raise ValueError("Session not detected. Try `track.init()`?") return _session def init(ignore_reinit_error=True, **session_kwargs): """Initializes the global trial context for this process. This creates a TrackSession object and the corresponding hooks for logging. Examples: >>> from ray.tune import track >>> track.init() """ global _session if _session: # TODO(ng): would be nice to stack crawl at creation time to report # where that initial trial was created, and that creation line # info is helpful to keep around anyway. reinit_msg = "A session already exists in the current context." if ignore_reinit_error: if not _session.is_tune_session: logger.warning(reinit_msg) return else: raise ValueError(reinit_msg) _session = TrackSession(**session_kwargs) def shutdown(): """Cleans up the trial and removes it from the global context.""" global _session if _session: _session.close() _session = None def log(**kwargs): """Applies TrackSession.log to the trial in the current context.""" _session = get_session() return _session.log(**kwargs) def trial_dir(): """Returns the directory where trial results are saved. This includes json data containing the session's parameters and metrics. """ _session = get_session() return _session.logdir __all__ = ["TrackSession", "session", "log", "trial_dir", "init", "shutdown"]
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/track/session.py
Python
import os from datetime import datetime from ray.tune.trial import Trial from ray.tune.result import DEFAULT_RESULTS_DIR, TRAINING_ITERATION from ray.tune.logger import UnifiedLogger, Logger class _ReporterHook(Logger): def __init__(self, tune_reporter): self.tune_reporter = tune_reporter def on_result(self, metrics): return self.tune_reporter(**metrics) class TrackSession: """Manages results for a single session. Represents a single Trial in an experiment. Attributes: trial_name (str): Custom trial name. experiment_dir (str): Directory where results for all trials are stored. Each session is stored into a unique directory inside experiment_dir. upload_dir (str): Directory to sync results to. trial_config (dict): Parameters that will be logged to disk. _tune_reporter (StatusReporter): For rerouting when using Tune. Will not instantiate logging if not None. """ def __init__(self, trial_name="", experiment_dir=None, upload_dir=None, trial_config=None, _tune_reporter=None): self._experiment_dir = None self._logdir = None self._upload_dir = None self.trial_config = None self._iteration = -1 self.is_tune_session = bool(_tune_reporter) self.trial_id = Trial.generate_id() if trial_name: self.trial_id = trial_name + "_" + self.trial_id if self.is_tune_session: self._logger = _ReporterHook(_tune_reporter) self._logdir = _tune_reporter.logdir else: self._initialize_logging(trial_name, experiment_dir, upload_dir, trial_config) def _initialize_logging(self, trial_name="", experiment_dir=None, upload_dir=None, trial_config=None): if upload_dir: raise NotImplementedError("Upload Dir is not yet implemented.") # TODO(rliaw): In other parts of the code, this is `local_dir`. if experiment_dir is None: experiment_dir = os.path.join(DEFAULT_RESULTS_DIR, "default") self._experiment_dir = os.path.expanduser(experiment_dir) # TODO(rliaw): Refactor `logdir` to `trial_dir`. self._logdir = Trial.create_logdir(trial_name, self._experiment_dir) self._upload_dir = upload_dir self.trial_config = trial_config or {} # misc metadata to save as well self.trial_config["trial_id"] = self.trial_id self._logger = UnifiedLogger(self.trial_config, self._logdir) def log(self, **metrics): """Logs all named arguments specified in `metrics`. This will log trial metrics locally, and they will be synchronized with the driver periodically through ray. Arguments: metrics: named arguments with corresponding values to log. """ self._iteration += 1 # TODO: Implement a batching mechanism for multiple calls to `log` # within the same iteration. metrics_dict = metrics.copy() metrics_dict.update({"trial_id": self.trial_id}) # TODO: Move Trainable autopopulation to a util function metrics_dict.setdefault(TRAINING_ITERATION, self._iteration) self._logger.on_result(metrics_dict) def close(self): self.trial_config["trial_completed"] = True self.trial_config["end_time"] = datetime.now().isoformat() # TODO(rliaw): Have Tune support updated configs self._logger.update_config(self.trial_config) self._logger.flush() self._logger.close() @property def logdir(self): """Trial logdir (subdir of given experiment directory)""" return self._logdir
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/trainable.py
Python
from datetime import datetime import copy import io import logging import glob import os import pickle import pandas as pd from six import string_types import shutil import tempfile import time import uuid import ray from ray.tune.logger import UnifiedLogger from ray.tune.result import (DEFAULT_RESULTS_DIR, TIME_THIS_ITER_S, TIMESTEPS_THIS_ITER, DONE, TIMESTEPS_TOTAL, EPISODES_THIS_ITER, EPISODES_TOTAL, TRAINING_ITERATION, RESULT_DUPLICATE) from ray.tune.utils import UtilMonitor logger = logging.getLogger(__name__) SETUP_TIME_THRESHOLD = 10 class TrainableUtil: @staticmethod def pickle_checkpoint(checkpoint_path): """Pickles checkpoint data.""" checkpoint_dir = TrainableUtil.find_checkpoint_dir(checkpoint_path) data = {} for basedir, _, file_names in os.walk(checkpoint_dir): for file_name in file_names: path = os.path.join(basedir, file_name) with open(path, "rb") as f: data[os.path.relpath(path, checkpoint_dir)] = f.read() # Use normpath so that a directory path isn't mapped to empty string. name = os.path.basename(os.path.normpath(checkpoint_path)) name += os.path.sep if os.path.isdir(checkpoint_path) else "" data_dict = pickle.dumps({ "checkpoint_name": name, "data": data, }) return data_dict @staticmethod def find_checkpoint_dir(checkpoint_path): """Returns the directory containing the checkpoint path. Raises: FileNotFoundError if the directory is not found. """ if not os.path.exists(checkpoint_path): raise FileNotFoundError("Path does not exist", checkpoint_path) if os.path.isdir(checkpoint_path): checkpoint_dir = checkpoint_path else: checkpoint_dir = os.path.dirname(checkpoint_path) while checkpoint_dir != os.path.dirname(checkpoint_dir): if os.path.exists(os.path.join(checkpoint_dir, ".is_checkpoint")): break checkpoint_dir = os.path.dirname(checkpoint_dir) else: raise FileNotFoundError("Checkpoint directory not found for {}" .format(checkpoint_path)) return checkpoint_dir @staticmethod def make_checkpoint_dir(checkpoint_dir): """Creates a checkpoint directory at the provided path.""" os.makedirs(checkpoint_dir, exist_ok=True) # Drop marker in directory to identify it as a checkpoint dir. open(os.path.join(checkpoint_dir, ".is_checkpoint"), "a").close() @staticmethod def get_checkpoints_paths(logdir): """ Finds the checkpoints within a specific folder. Returns a pandas DataFrame of training iterations and checkpoint paths within a specific folder. Raises: FileNotFoundError if the directory is not found. """ marker_paths = glob.glob( os.path.join(logdir, "checkpoint_*/.is_checkpoint")) iter_chkpt_pairs = [] for marker_path in marker_paths: chkpt_dir = os.path.dirname(marker_path) metadata_file = glob.glob( os.path.join(chkpt_dir, "*.tune_metadata")) if len(metadata_file) != 1: raise ValueError( "{} has zero or more than one tune_metadata.".format( chkpt_dir)) chkpt_path = metadata_file[0][:-len(".tune_metadata")] chkpt_iter = int(chkpt_dir[chkpt_dir.rfind("_") + 1:]) iter_chkpt_pairs.append([chkpt_iter, chkpt_path]) chkpt_df = pd.DataFrame( iter_chkpt_pairs, columns=["training_iteration", "chkpt_path"]) return chkpt_df class Trainable: """Abstract class for trainable models, functions, etc. A call to ``train()`` on a trainable will execute one logical iteration of training. As a rule of thumb, the execution time of one train call should be large enough to avoid overheads (i.e. more than a few seconds), but short enough to report progress periodically (i.e. at most a few minutes). Calling ``save()`` should save the training state of a trainable to disk, and ``restore(path)`` should restore a trainable to the given state. Generally you only need to implement ``_setup``, ``_train``, ``_save``, and ``_restore`` when subclassing Trainable. Other implementation methods that may be helpful to override are ``_log_result``, ``reset_config``, ``_stop``, and ``_export_model``. When using Tune, Tune will convert this class into a Ray actor, which runs on a separate process. Tune will also change the current working directory of this process to `self.logdir`. """ def __init__(self, config=None, logger_creator=None): """Initialize an Trainable. Sets up logging and points ``self.logdir`` to a directory in which training outputs should be placed. Subclasses should prefer defining ``_setup()`` instead of overriding ``__init__()`` directly. Args: config (dict): Trainable-specific configuration data. By default will be saved as ``self.config``. logger_creator (func): Function that creates a ray.tune.Logger object. If unspecified, a default logger is created. """ self._experiment_id = uuid.uuid4().hex self.config = config or {} if logger_creator: self._result_logger = logger_creator(self.config) self._logdir = self._result_logger.logdir else: logdir_prefix = datetime.today().strftime("%Y-%m-%d_%H-%M-%S") ray.utils.try_to_create_directory(DEFAULT_RESULTS_DIR) self._logdir = tempfile.mkdtemp( prefix=logdir_prefix, dir=DEFAULT_RESULTS_DIR) self._result_logger = UnifiedLogger( self.config, self._logdir, loggers=None) self._iteration = 0 self._time_total = 0.0 self._timesteps_total = None self._episodes_total = None self._time_since_restore = 0.0 self._timesteps_since_restore = 0 self._iterations_since_restore = 0 self._restored = False start_time = time.time() self._setup(copy.deepcopy(self.config)) setup_time = time.time() - start_time if setup_time > SETUP_TIME_THRESHOLD: logger.info("_setup took {:.3f} seconds. If your trainable is " "slow to initialize, consider setting " "reuse_actors=True to reduce actor creation " "overheads.".format(setup_time)) self._local_ip = ray.services.get_node_ip_address() log_sys_usage = self.config.get("log_sys_usage", False) self._monitor = UtilMonitor(start=log_sys_usage) @classmethod def default_resource_request(cls, config): """Returns the resource requirement for the given configuration. This can be overriden by sub-classes to set the correct trial resource allocation, so the user does not need to. Example: >>> def default_resource_request(cls, config): >>> return Resources( >>> cpu=0, >>> gpu=0, >>> extra_cpu=config["workers"], >>> extra_gpu=int(config["use_gpu"]) * config["workers"]) """ return None @classmethod def resource_help(cls, config): """Returns a help string for configuring this trainable's resources. Args: config (dict): The Trainer's config dict. """ return "" def current_ip(self): logger.warning("Getting current IP.") self._local_ip = ray.services.get_node_ip_address() return self._local_ip def train(self): """Runs one logical iteration of training. Subclasses should override ``_train()`` instead to return results. This class automatically fills the following fields in the result: `done` (bool): training is terminated. Filled only if not provided. `time_this_iter_s` (float): Time in seconds this iteration took to run. This may be overriden in order to override the system-computed time difference. `time_total_s` (float): Accumulated time in seconds for this entire experiment. `experiment_id` (str): Unique string identifier for this experiment. This id is preserved across checkpoint / restore calls. `training_iteration` (int): The index of this training iteration, e.g. call to train(). This is incremented after `_train()` is called. `pid` (str): The pid of the training process. `date` (str): A formatted date of when the result was processed. `timestamp` (str): A UNIX timestamp of when the result was processed. `hostname` (str): Hostname of the machine hosting the training process. `node_ip` (str): Node ip of the machine hosting the training process. Returns: A dict that describes training progress. """ start = time.time() result = self._train() assert isinstance(result, dict), "_train() needs to return a dict." # We do not modify internal state nor update this result if duplicate. if RESULT_DUPLICATE in result: return result result = result.copy() self._iteration += 1 self._iterations_since_restore += 1 if result.get(TIME_THIS_ITER_S) is not None: time_this_iter = result[TIME_THIS_ITER_S] else: time_this_iter = time.time() - start self._time_total += time_this_iter self._time_since_restore += time_this_iter result.setdefault(DONE, False) # self._timesteps_total should only be tracked if increments provided if result.get(TIMESTEPS_THIS_ITER) is not None: if self._timesteps_total is None: self._timesteps_total = 0 self._timesteps_total += result[TIMESTEPS_THIS_ITER] self._timesteps_since_restore += result[TIMESTEPS_THIS_ITER] # self._episodes_total should only be tracked if increments provided if result.get(EPISODES_THIS_ITER) is not None: if self._episodes_total is None: self._episodes_total = 0 self._episodes_total += result[EPISODES_THIS_ITER] # self._timesteps_total should not override user-provided total result.setdefault(TIMESTEPS_TOTAL, self._timesteps_total) result.setdefault(EPISODES_TOTAL, self._episodes_total) result.setdefault(TRAINING_ITERATION, self._iteration) # Provides auto-filled neg_mean_loss for avoiding regressions if result.get("mean_loss"): result.setdefault("neg_mean_loss", -result["mean_loss"]) now = datetime.today() result.update( experiment_id=self._experiment_id, date=now.strftime("%Y-%m-%d_%H-%M-%S"), timestamp=int(time.mktime(now.timetuple())), time_this_iter_s=time_this_iter, time_total_s=self._time_total, pid=os.getpid(), hostname=os.uname()[1], node_ip=self._local_ip, config=self.config, time_since_restore=self._time_since_restore, timesteps_since_restore=self._timesteps_since_restore, iterations_since_restore=self._iterations_since_restore) monitor_data = self._monitor.get_data() if monitor_data: result.update(monitor_data) self._log_result(result) return result def save(self, checkpoint_dir=None): """Saves the current model state to a checkpoint. Subclasses should override ``_save()`` instead to save state. This method dumps additional metadata alongside the saved path. Args: checkpoint_dir (str): Optional dir to place the checkpoint. Returns: Checkpoint path or prefix that may be passed to restore(). """ checkpoint_dir = os.path.join(checkpoint_dir or self.logdir, "checkpoint_{}".format(self._iteration)) TrainableUtil.make_checkpoint_dir(checkpoint_dir) checkpoint = self._save(checkpoint_dir) saved_as_dict = False if isinstance(checkpoint, string_types): if not checkpoint.startswith(checkpoint_dir): raise ValueError( "The returned checkpoint path must be within the " "given checkpoint dir {}: {}".format( checkpoint_dir, checkpoint)) checkpoint_path = checkpoint if os.path.isdir(checkpoint_path): # Add trailing slash to prevent tune metadata from # being written outside the directory. checkpoint_path = os.path.join(checkpoint_path, "") elif isinstance(checkpoint, dict): saved_as_dict = True checkpoint_path = os.path.join(checkpoint_dir, "checkpoint") with open(checkpoint_path, "wb") as f: pickle.dump(checkpoint, f) else: raise ValueError("Returned unexpected type {}. " "Expected str or dict.".format(type(checkpoint))) with open(checkpoint_path + ".tune_metadata", "wb") as f: pickle.dump({ "experiment_id": self._experiment_id, "iteration": self._iteration, "timesteps_total": self._timesteps_total, "time_total": self._time_total, "episodes_total": self._episodes_total, "saved_as_dict": saved_as_dict, "ray_version": ray.__version__, }, f) return checkpoint_path def save_to_object(self): """Saves the current model state to a Python object. It also saves to disk but does not return the checkpoint path. Returns: Object holding checkpoint data. """ tmpdir = tempfile.mkdtemp("save_to_object", dir=self.logdir) checkpoint_path = self.save(tmpdir) # Save all files in subtree. data_dict = TrainableUtil.pickle_checkpoint(checkpoint_path) out = io.BytesIO() if len(data_dict) > 10e6: # getting pretty large logger.info("Checkpoint size is {} bytes".format(len(data_dict))) out.write(data_dict) shutil.rmtree(tmpdir) return out.getvalue() def restore(self, checkpoint_path): """Restores training state from a given model checkpoint. These checkpoints are returned from calls to save(). Subclasses should override ``_restore()`` instead to restore state. This method restores additional metadata saved with the checkpoint. """ with open(checkpoint_path + ".tune_metadata", "rb") as f: metadata = pickle.load(f) self._experiment_id = metadata["experiment_id"] self._iteration = metadata["iteration"] self._timesteps_total = metadata["timesteps_total"] self._time_total = metadata["time_total"] self._episodes_total = metadata["episodes_total"] saved_as_dict = metadata["saved_as_dict"] if saved_as_dict: with open(checkpoint_path, "rb") as loaded_state: checkpoint_dict = pickle.load(loaded_state) checkpoint_dict.update(tune_checkpoint_path=checkpoint_path) self._restore(checkpoint_dict) else: self._restore(checkpoint_path) self._time_since_restore = 0.0 self._timesteps_since_restore = 0 self._iterations_since_restore = 0 self._restored = True logger.info("Restored on %s from checkpoint: %s", self.current_ip(), checkpoint_path) state = { "_iteration": self._iteration, "_timesteps_total": self._timesteps_total, "_time_total": self._time_total, "_episodes_total": self._episodes_total, } logger.info("Current state after restoring: %s", state) def restore_from_object(self, obj): """Restores training state from a checkpoint object. These checkpoints are returned from calls to save_to_object(). """ info = pickle.loads(obj) data = info["data"] tmpdir = tempfile.mkdtemp("restore_from_object", dir=self.logdir) checkpoint_path = os.path.join(tmpdir, info["checkpoint_name"]) for relpath_name, file_contents in data.items(): path = os.path.join(tmpdir, relpath_name) # This may be a subdirectory, hence not just using tmpdir os.makedirs(os.path.dirname(path), exist_ok=True) with open(path, "wb") as f: f.write(file_contents) self.restore(checkpoint_path) shutil.rmtree(tmpdir) def delete_checkpoint(self, checkpoint_path): """Deletes local copy of checkpoint. Args: checkpoint_path (str): Path to checkpoint. """ try: checkpoint_dir = TrainableUtil.find_checkpoint_dir(checkpoint_path) except FileNotFoundError: # The checkpoint won't exist locally if the # trial was rescheduled to another worker. logger.debug("Checkpoint not found during garbage collection.") return if os.path.exists(checkpoint_dir): shutil.rmtree(checkpoint_dir) def export_model(self, export_formats, export_dir=None): """Exports model based on export_formats. Subclasses should override _export_model() to actually export model to local directory. Args: export_formats (list): List of formats that should be exported. export_dir (str): Optional dir to place the exported model. Defaults to self.logdir. Returns: A dict that maps ExportFormats to successfully exported models. """ export_dir = export_dir or self.logdir return self._export_model(export_formats, export_dir) def reset_config(self, new_config): """Resets configuration without restarting the trial. This method is optional, but can be implemented to speed up algorithms such as PBT, and to allow performance optimizations such as running experiments with reuse_actors=True. Note that self.config need to be updated to reflect the latest parameter information in Ray logs. Args: new_config (dir): Updated hyperparameter configuration for the trainable. Returns: True if reset was successful else False. """ return False def stop(self): """Releases all resources used by this trainable.""" self._result_logger.flush() self._result_logger.close() self._stop() @property def logdir(self): """Directory of the results and checkpoints for this Trainable. Tune will automatically sync this folder with the driver if execution is distributed. Note that the current working directory will also be changed to this. """ return os.path.join(self._logdir, "") @property def iteration(self): """Current training iteration. This value is automatically incremented every time `train()` is called and is automatically inserted into the training result dict. """ return self._iteration def get_config(self): """Returns configuration passed in by Tune.""" return self.config def _train(self): """Subclasses should override this to implement train(). The return value will be automatically passed to the loggers. Users can also return `tune.result.DONE` or `tune.result.SHOULD_CHECKPOINT` as a key to manually trigger termination or checkpointing of this trial. Note that manual checkpointing only works when subclassing Trainables. Returns: A dict that describes training progress. """ raise NotImplementedError def _save(self, tmp_checkpoint_dir): """Subclasses should override this to implement ``save()``. Warning: Do not rely on absolute paths in the implementation of ``_save`` and ``_restore``. Use ``validate_save_restore`` to catch ``_save``/``_restore`` errors before execution. >>> from ray.tune.utils import validate_save_restore >>> validate_save_restore(MyTrainableClass) >>> validate_save_restore(MyTrainableClass, use_object_store=True) Args: tmp_checkpoint_dir (str): The directory where the checkpoint file must be stored. In a Tune run, if the trial is paused, the provided path may be temporary and moved. Returns: A dict or string. If string, the return value is expected to be prefixed by `tmp_checkpoint_dir`. If dict, the return value will be automatically serialized by Tune and passed to `_restore()`. Examples: >>> print(trainable1._save("/tmp/checkpoint_1")) "/tmp/checkpoint_1/my_checkpoint_file" >>> print(trainable2._save("/tmp/checkpoint_2")) {"some": "data"} >>> trainable._save("/tmp/bad_example") "/tmp/NEW_CHECKPOINT_PATH/my_checkpoint_file" # This will error. """ raise NotImplementedError def _restore(self, checkpoint): """Subclasses should override this to implement restore(). Warning: In this method, do not rely on absolute paths. The absolute path of the checkpoint_dir used in ``_save`` may be changed. If ``_save`` returned a prefixed string, the prefix of the checkpoint string returned by ``_save`` may be changed. This is because trial pausing depends on temporary directories. The directory structure under the checkpoint_dir provided to ``_save`` is preserved. See the example below. .. code-block:: python class Example(Trainable): def _save(self, checkpoint_path): print(checkpoint_path) return os.path.join(checkpoint_path, "my/check/point") def _restore(self, checkpoint): print(checkpoint) >>> trainer = Example() >>> obj = trainer.save_to_object() # This is used when PAUSED. <logdir>/tmpc8k_c_6hsave_to_object/checkpoint_0/my/check/point >>> trainer.restore_from_object(obj) # Note the different prefix. <logdir>/tmpb87b5axfrestore_from_object/checkpoint_0/my/check/point Args: checkpoint (str|dict): If dict, the return value is as returned by `_save`. If a string, then it is a checkpoint path that may have a different prefix than that returned by `_save`. The directory structure underneath the `checkpoint_dir` `_save` is preserved. """ raise NotImplementedError def _setup(self, config): """Subclasses should override this for custom initialization. Args: config (dict): Hyperparameters and other configs given. Copy of `self.config`. """ pass def _log_result(self, result): """Subclasses can optionally override this to customize logging. Args: result (dict): Training result returned by _train(). """ self._result_logger.on_result(result) def _stop(self): """Subclasses should override this for any cleanup on stop. If any Ray actors are launched in the Trainable (i.e., with a RLlib trainer), be sure to kill the Ray actor process here. You can kill a Ray actor by calling `actor.__ray_terminate__.remote()` on the actor. """ pass def _export_model(self, export_formats, export_dir): """Subclasses should override this to export model. Args: export_formats (list): List of formats that should be exported. export_dir (str): Directory to place exported models. Return: A dict that maps ExportFormats to successfully exported models. """ return {}
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/trial.py
Python
import ray.cloudpickle as cloudpickle import copy from datetime import datetime import logging import shutil import uuid import time import tempfile import os from numbers import Number from ray.tune import TuneError from ray.tune.checkpoint_manager import Checkpoint, CheckpointManager from ray.tune.durable_trainable import DurableTrainable from ray.tune.logger import pretty_print, UnifiedLogger # NOTE(rkn): We import ray.tune.registry here instead of importing the names we # need because there are cyclic imports that may cause specific names to not # have been defined yet. See https://github.com/ray-project/ray/issues/1716. from ray.tune.registry import get_trainable_cls, validate_trainable from ray.tune.result import DEFAULT_RESULTS_DIR, DONE, TRAINING_ITERATION from ray.tune.resources import Resources, json_to_resources, resources_to_json from ray.tune.trainable import TrainableUtil from ray.tune.utils import flatten_dict from ray.utils import binary_to_hex, hex_to_binary DEBUG_PRINT_INTERVAL = 5 MAX_LEN_IDENTIFIER = int(os.environ.get("MAX_LEN_IDENTIFIER", 130)) logger = logging.getLogger(__name__) def date_str(): return datetime.today().strftime("%Y-%m-%d_%H-%M-%S") class Location: """Describes the location at which Trial is placed to run.""" def __init__(self, hostname=None, pid=None): self.hostname = hostname self.pid = pid def __str__(self): if not self.pid: return "" elif self.hostname == os.uname()[1]: return "pid={}".format(self.pid) else: return "{}:{}".format(self.hostname, self.pid) class ExportFormat: """Describes the format to export the trial Trainable. This may correspond to different file formats based on the Trainable implementation. """ CHECKPOINT = "checkpoint" MODEL = "model" @staticmethod def validate(export_formats): """Validates export_formats. Raises: ValueError if the format is unknown. """ for i in range(len(export_formats)): export_formats[i] = export_formats[i].strip().lower() if export_formats[i] not in [ ExportFormat.CHECKPOINT, ExportFormat.MODEL ]: raise TuneError("Unsupported export format: " + export_formats[i]) def checkpoint_deleter(trial_id, runner): """Returns a checkpoint deleter callback for a runner.""" if not runner: return lambda checkpoint: None def delete(checkpoint): """Requests checkpoint deletion asynchronously. Args: checkpoint (Checkpoint): Checkpoint to delete. """ if checkpoint.storage == Checkpoint.PERSISTENT and checkpoint.value: logger.debug("Trial %s: Deleting checkpoint %s", trial_id, checkpoint.value) checkpoint_path = checkpoint.value # Delete local copy, if any exists. if os.path.exists(checkpoint_path): try: checkpoint_dir = TrainableUtil.find_checkpoint_dir( checkpoint_path) shutil.rmtree(checkpoint_dir) except FileNotFoundError: logger.warning("Checkpoint dir not found during deletion.") # TODO(ujvl): Batch remote deletes. runner.delete_checkpoint.remote(checkpoint.value) return delete class Trial: """A trial object holds the state for one model training run. Trials are themselves managed by the TrialRunner class, which implements the event loop for submitting trial runs to a Ray cluster. Trials start in the PENDING state, and transition to RUNNING once started. On error it transitions to ERROR, otherwise TERMINATED on success. """ PENDING = "PENDING" RUNNING = "RUNNING" PAUSED = "PAUSED" TERMINATED = "TERMINATED" ERROR = "ERROR" def __init__(self, trainable_name, config=None, trial_id=None, local_dir=DEFAULT_RESULTS_DIR, evaluated_params=None, experiment_tag="", resources=None, stopping_criterion=None, remote_checkpoint_dir=None, checkpoint_freq=0, checkpoint_at_end=False, sync_on_checkpoint=True, keep_checkpoints_num=None, checkpoint_score_attr=TRAINING_ITERATION, export_formats=None, restore_path=None, trial_name_creator=None, loggers=None, sync_to_driver_fn=None, max_failures=0): """Initialize a new trial. The args here take the same meaning as the command line flags defined in ray.tune.config_parser. """ validate_trainable(trainable_name) # Trial config self.trainable_name = trainable_name self.trial_id = Trial.generate_id() if trial_id is None else trial_id self.config = config or {} self.local_dir = local_dir # This remains unexpanded for syncing. #: Parameters that Tune varies across searches. self.evaluated_params = evaluated_params or {} self.experiment_tag = experiment_tag trainable_cls = self.get_trainable_cls() if trainable_cls and hasattr(trainable_cls, "default_resource_request"): default_resources = trainable_cls.default_resource_request( self.config) if default_resources: if resources: raise ValueError( "Resources for {} have been automatically set to {} " "by its `default_resource_request()` method. Please " "clear the `resources_per_trial` option.".format( trainable_cls, default_resources)) resources = default_resources self.location = Location() self.resources = resources or Resources(cpu=1, gpu=0) self.stopping_criterion = stopping_criterion or {} self.loggers = loggers self.sync_to_driver_fn = sync_to_driver_fn self.verbose = True self.max_failures = max_failures # Local trial state that is updated during the run self.last_result = {} self.last_update_time = -float("inf") # stores in memory max/min/last result for each metric by trial self.metric_analysis = {} self.export_formats = export_formats self.status = Trial.PENDING self.start_time = None self.logdir = None self.runner = None self.result_logger = None self.last_debug = 0 self.error_file = None self.error_msg = None self.custom_trial_name = None # Checkpointing fields if remote_checkpoint_dir: self.remote_checkpoint_dir_prefix = remote_checkpoint_dir else: self.remote_checkpoint_dir_prefix = None self.checkpoint_freq = checkpoint_freq self.checkpoint_at_end = checkpoint_at_end self.sync_on_checkpoint = sync_on_checkpoint newest_checkpoint = Checkpoint(Checkpoint.PERSISTENT, restore_path) self.checkpoint_manager = CheckpointManager( keep_checkpoints_num, checkpoint_score_attr, checkpoint_deleter(str(self), self.runner)) self.checkpoint_manager.newest_checkpoint = newest_checkpoint # Restoration fields self.restoring_from = None self.num_failures = 0 self.num_consecutive_start_attempts = 0 # AutoML fields self.results = None self.best_result = None self.param_config = None self.extra_arg = None self._nonjson_fields = [ "loggers", "sync_to_driver_fn", "results", "best_result", "param_config", "extra_arg", ] if trial_name_creator: self.custom_trial_name = trial_name_creator(self) @property def node_ip(self): return self.location.hostname @property def checkpoint(self): return self.checkpoint_manager.newest_checkpoint @classmethod def generate_id(cls): return str(uuid.uuid1().hex)[:8] @property def remote_checkpoint_dir(self): assert self.logdir, "Trial {}: logdir not initialized.".format(self) if not self.remote_checkpoint_dir_prefix: return None logdir_name = os.path.basename(self.logdir) return os.path.join(self.remote_checkpoint_dir_prefix, logdir_name) @classmethod def create_logdir(cls, identifier, local_dir): local_dir = os.path.expanduser(local_dir) os.makedirs(local_dir, exist_ok=True) return tempfile.mkdtemp( prefix="{}_{}".format(identifier[:MAX_LEN_IDENTIFIER], date_str()), dir=local_dir) def init_logger(self): """Init logger.""" if not self.result_logger: if not self.logdir: self.logdir = Trial.create_logdir(str(self), self.local_dir) else: os.makedirs(self.logdir, exist_ok=True) self.result_logger = UnifiedLogger( self.config, self.logdir, trial=self, loggers=self.loggers, sync_function=self.sync_to_driver_fn) def update_resources(self, cpu, gpu, **kwargs): """EXPERIMENTAL: Updates the resource requirements. Should only be called when the trial is not running. Raises: ValueError if trial status is running. """ if self.status is Trial.RUNNING: raise ValueError("Cannot update resources while Trial is running.") self.resources = Resources(cpu, gpu, **kwargs) def set_runner(self, runner): self.runner = runner self.checkpoint_manager.delete = checkpoint_deleter(str(self), runner) def set_location(self, location): """Sets the location of the trial.""" self.location = location def set_status(self, status): """Sets the status of the trial.""" self.status = status if status == Trial.RUNNING: if self.start_time is None: self.start_time = time.time() def close_logger(self): """Closes logger.""" if self.result_logger: self.result_logger.close() self.result_logger = None def write_error_log(self, error_msg): if error_msg and self.logdir: self.num_failures += 1 self.error_file = os.path.join(self.logdir, "error.txt") with open(self.error_file, "a+") as f: f.write("Failure # {} (occurred at {})\n".format( self.num_failures, date_str())) f.write(error_msg + "\n") self.error_msg = error_msg def should_stop(self, result): """Whether the given result meets this trial's stopping criteria.""" if result.get(DONE): return True if callable(self.stopping_criterion): return self.stopping_criterion(self.trial_id, result) for criteria, stop_value in self.stopping_criterion.items(): if criteria not in result: raise TuneError( "Stopping criteria {} not provided in result {}.".format( criteria, result)) elif isinstance(criteria, dict): raise ValueError( "Stopping criteria is now flattened by default. " "Use forward slashes to nest values `key1/key2/key3`.") elif result[criteria] >= stop_value: return True return False def should_checkpoint(self): """Whether this trial is due for checkpointing.""" result = self.last_result or {} if result.get(DONE) and self.checkpoint_at_end: return True return (self.checkpoint_freq and result.get(TRAINING_ITERATION, 0) % self.checkpoint_freq == 0) def has_checkpoint(self): return self.checkpoint.value is not None def clear_checkpoint(self): self.checkpoint.value = None self.restoring_from = None def on_checkpoint(self, checkpoint): """Hook for handling checkpoints taken by the Trainable. Args: checkpoint (Checkpoint): Checkpoint taken. """ if checkpoint.storage == Checkpoint.MEMORY: # TODO(ujvl): Handle this separately to avoid restoration failure. self.checkpoint_manager.on_checkpoint(checkpoint) return if self.sync_on_checkpoint: try: # Wait for any other syncs to finish. We need to sync again # after this to handle checkpoints taken mid-sync. self.result_logger.wait() except TuneError as e: # Errors occurring during this wait are not fatal for this # checkpoint, so it should just be logged. logger.error( "Trial %s: An error occurred during the " "checkpoint pre-sync wait.", str(e)) # Force sync down and wait before tracking the new checkpoint. try: if self.result_logger.sync_down(): self.result_logger.wait() else: logger.error( "Trial %s: Checkpoint sync skipped. " "This should not happen.", self) except TuneError as e: if issubclass(self.get_trainable_cls(), DurableTrainable): # Even though rsync failed the trainable can restore # from remote durable storage. logger.error("Trial %s: Sync error - %s", self, str(e)) else: # If the trainable didn't have remote storage to upload # to then this checkpoint may have been lost, so we # shouldn't track it with the checkpoint_manager. raise e if not issubclass(self.get_trainable_cls(), DurableTrainable): if not os.path.exists(checkpoint.value): raise TuneError("Trial {}: Checkpoint path {} not " "found after successful sync down.".format( self, checkpoint.value)) self.checkpoint_manager.on_checkpoint(checkpoint) def on_restore(self): """Handles restoration completion.""" assert self.is_restoring self.last_result = self.restoring_from.result self.restoring_from = None def should_recover(self): """Returns whether the trial qualifies for retrying. This is if the trial has not failed more than max_failures. Note this may return true even when there is no checkpoint, either because `self.checkpoint_freq` is `0` or because the trial failed before a checkpoint has been made. """ return self.num_failures < self.max_failures or self.max_failures < 0 def update_last_result(self, result, terminate=False): result.update(trial_id=self.trial_id, done=terminate) if self.experiment_tag: result.update(experiment_tag=self.experiment_tag) if self.verbose and (terminate or time.time() - self.last_debug > DEBUG_PRINT_INTERVAL): print("Result for {}:".format(self)) print(" {}".format(pretty_print(result).replace("\n", "\n "))) self.last_debug = time.time() self.set_location(Location(result.get("node_ip"), result.get("pid"))) self.last_result = result self.last_update_time = time.time() self.result_logger.on_result(self.last_result) for metric, value in flatten_dict(result).items(): if isinstance(value, Number): if metric not in self.metric_analysis: self.metric_analysis[metric] = { "max": value, "min": value, "last": value } else: self.metric_analysis[metric]["max"] = max( value, self.metric_analysis[metric]["max"]) self.metric_analysis[metric]["min"] = min( value, self.metric_analysis[metric]["min"]) self.metric_analysis[metric]["last"] = value def get_trainable_cls(self): return get_trainable_cls(self.trainable_name) def set_verbose(self, verbose): self.verbose = verbose def is_finished(self): return self.status in [Trial.ERROR, Trial.TERMINATED] @property def is_restoring(self): return self.restoring_from is not None def __repr__(self): return str(self) def __str__(self): """Combines ``env`` with ``trainable_name`` and ``trial_id``. Can be overridden with a custom string creator. """ if self.custom_trial_name: return self.custom_trial_name if "env" in self.config: env = self.config["env"] if isinstance(env, type): env = env.__name__ identifier = "{}_{}".format(self.trainable_name, env) else: identifier = self.trainable_name identifier += "_" + self.trial_id return identifier.replace("/", "_") def __getstate__(self): """Memento generator for Trial. Sets RUNNING trials to PENDING, and flushes the result logger. Note this can only occur if the trial holds a PERSISTENT checkpoint. """ assert self.checkpoint.storage == Checkpoint.PERSISTENT, ( "Checkpoint must not be in-memory.") state = self.__dict__.copy() state["resources"] = resources_to_json(self.resources) for key in self._nonjson_fields: state[key] = binary_to_hex(cloudpickle.dumps(state.get(key))) state["runner"] = None state["result_logger"] = None if self.result_logger: self.result_logger.flush(sync_down=False) state["__logger_started__"] = True else: state["__logger_started__"] = False return copy.deepcopy(state) def __setstate__(self, state): logger_started = state.pop("__logger_started__") state["resources"] = json_to_resources(state["resources"]) if state["status"] == Trial.RUNNING: state["status"] = Trial.PENDING for key in self._nonjson_fields: state[key] = cloudpickle.loads(hex_to_binary(state[key])) self.__dict__.update(state) validate_trainable(self.trainable_name) if logger_started: self.init_logger()
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/trial_executor.py
Python
# coding: utf-8 import logging from ray.tune.trial import Trial, Checkpoint from ray.tune.error import TuneError logger = logging.getLogger(__name__) class TrialExecutor: """Manages platform-specific details such as resource handling and starting/stopping trials. """ def __init__(self, queue_trials=False): """Initializes a new TrialExecutor. Args: queue_trials (bool): Whether to queue trials when the cluster does not currently have enough resources to launch one. This should be set to True when running on an autoscaling cluster to enable automatic scale-up. """ self._queue_trials = queue_trials self._cached_trial_state = {} def set_status(self, trial, status): """Sets status and checkpoints metadata if needed. Only checkpoints metadata if trial status is a terminal condition. PENDING, PAUSED, and RUNNING switches have checkpoints taken care of in the TrialRunner. Args: trial (Trial): Trial to checkpoint. status (Trial.status): Status to set trial to. """ if trial.status == status: logger.debug("Trial %s: Status %s unchanged.", trial, trial.status) else: logger.debug("Trial %s: Changing status from %s to %s.", trial, trial.status, status) trial.set_status(status) if status in [Trial.TERMINATED, Trial.ERROR]: self.try_checkpoint_metadata(trial) def try_checkpoint_metadata(self, trial): """Checkpoints metadata. Args: trial (Trial): Trial to checkpoint. """ if trial.checkpoint.storage == Checkpoint.MEMORY: logger.debug("Trial %s: Not saving data for memory checkpoint.", trial) return try: logger.debug("Trial %s: Saving trial metadata.", trial) self._cached_trial_state[trial.trial_id] = trial.__getstate__() except Exception: logger.exception("Trial %s: Error checkpointing trial metadata.", trial) def get_checkpoints(self): """Returns a copy of mapping of the trial ID to pickled metadata.""" return self._cached_trial_state.copy() def has_resources(self, resources): """Returns whether this runner has at least the specified resources.""" raise NotImplementedError("Subclasses of TrialExecutor must provide " "has_resources() method") def start_trial(self, trial, checkpoint=None): """Starts the trial restoring from checkpoint if checkpoint is provided. Args: trial (Trial): Trial to be started. checkpoint(Checkpoint): A Python object or path storing the state of trial. """ raise NotImplementedError("Subclasses of TrialExecutor must provide " "start_trial() method") def stop_trial(self, trial, error=False, error_msg=None, stop_logger=True): """Stops the trial. Stops this trial, releasing all allocating resources. If stopping the trial fails, the run will be marked as terminated in error, but no exception will be thrown. Args: error (bool): Whether to mark this trial as terminated in error. error_msg (str): Optional error message. stop_logger (bool): Whether to shut down the trial logger. """ raise NotImplementedError("Subclasses of TrialExecutor must provide " "stop_trial() method") def continue_training(self, trial): """Continues the training of this trial.""" pass def pause_trial(self, trial): """Pauses the trial. We want to release resources (specifically GPUs) when pausing an experiment. This results in PAUSED state that similar to TERMINATED. """ assert trial.status == Trial.RUNNING, trial.status try: self.save(trial, Checkpoint.MEMORY) self.stop_trial(trial, stop_logger=False) self.set_status(trial, Trial.PAUSED) except Exception: logger.exception("Error pausing runner.") self.set_status(trial, Trial.ERROR) def unpause_trial(self, trial): """Sets PAUSED trial to pending to allow scheduler to start.""" assert trial.status == Trial.PAUSED, trial.status self.set_status(trial, Trial.PENDING) def resume_trial(self, trial): """Resumes PAUSED trials. This is a blocking call.""" assert trial.status == Trial.PAUSED, trial.status self.start_trial(trial) def reset_trial(self, trial, new_config, new_experiment_tag): """Tries to invoke `Trainable.reset_config()` to reset trial. Args: trial (Trial): Trial to be reset. new_config (dict): New configuration for Trial trainable. new_experiment_tag (str): New experiment name for trial. Returns: True if `reset_config` is successful else False. """ raise NotImplementedError def get_running_trials(self): """Returns all running trials.""" raise NotImplementedError("Subclasses of TrialExecutor must provide " "get_running_trials() method") def on_step_begin(self, trial_runner): """A hook called before running one step of the trial event loop.""" pass def on_step_end(self, trial_runner): """A hook called after running one step of the trial event loop.""" pass def on_no_available_trials(self, trial_runner): if self._queue_trials: return for trial in trial_runner.get_trials(): if trial.status == Trial.PENDING: if not self.has_resources(trial.resources): raise TuneError( ("Insufficient cluster resources to launch trial: " "trial requested {} but the cluster has only {}. " "Pass `queue_trials=True` in " "ray.tune.run() or on the command " "line to queue trials until the cluster scales " "up or resources become available. {}").format( trial.resources.summary_string(), self.resource_string(), trial.get_trainable_cls().resource_help( trial.config))) elif trial.status == Trial.PAUSED: raise TuneError("There are paused trials, but no more pending " "trials with sufficient resources.") def get_next_available_trial(self): """Blocking call that waits until one result is ready. Returns: Trial object that is ready for intermediate processing. """ raise NotImplementedError def get_next_failed_trial(self): """Non-blocking call that detects and returns one failed trial. Returns: A Trial object that is ready for failure processing. None if no failure detected. """ raise NotImplementedError def fetch_result(self, trial): """Fetches one result for the trial. Assumes the trial is running. Return: Result object for the trial. """ raise NotImplementedError def debug_string(self): """Returns a human readable message for printing to the console.""" raise NotImplementedError def resource_string(self): """Returns a string describing the total resources available.""" raise NotImplementedError def restore(self, trial, checkpoint=None): """Restores training state from a checkpoint. If checkpoint is None, try to restore from trial.checkpoint. If restoring fails, the trial status will be set to ERROR. Args: trial (Trial): Trial to be restored. checkpoint (Checkpoint): Checkpoint to restore from. Return: False if error occurred, otherwise return True. """ raise NotImplementedError("Subclasses of TrialExecutor must provide " "restore() method") def save(self, trial, storage=Checkpoint.PERSISTENT, result=None): """Saves training state of this trial to a checkpoint. If result is None, this trial's last result will be used. Args: trial (Trial): The state of this trial to be saved. storage (str): Where to store the checkpoint. Defaults to PERSISTENT. result (dict): The state of this trial as a dictionary to be saved. Return: A Python object if storage==Checkpoint.MEMORY otherwise a path to the checkpoint. """ raise NotImplementedError("Subclasses of TrialExecutor must provide " "save() method") def export_trial_if_needed(self, trial): """Exports model of this trial based on trial.export_formats. Args: trial (Trial): The state of this trial to be saved. Return: A dict that maps ExportFormats to successfully exported models. """ raise NotImplementedError("Subclasses of TrialExecutor must provide " "export_trial_if_needed() method") def has_gpus(self): """Returns True if GPUs are detected on the cluster.""" return None
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/trial_runner.py
Python
import click from datetime import datetime import json import logging import os import time import traceback import types import ray.cloudpickle as cloudpickle from ray.tune import TuneError from ray.tune.progress_reporter import trial_progress_str from ray.tune.ray_trial_executor import RayTrialExecutor from ray.tune.result import (TIME_THIS_ITER_S, RESULT_DUPLICATE, SHOULD_CHECKPOINT) from ray.tune.syncer import get_cloud_syncer from ray.tune.trial import Checkpoint, Trial from ray.tune.schedulers import FIFOScheduler, TrialScheduler from ray.tune.suggest import BasicVariantGenerator from ray.tune.utils import warn_if_slow, flatten_dict from ray.tune.web_server import TuneServer from ray.utils import binary_to_hex, hex_to_binary MAX_DEBUG_TRIALS = 20 logger = logging.getLogger(__name__) def _find_newest_ckpt(ckpt_dir): """Returns path to most recently modified checkpoint.""" full_paths = [ os.path.join(ckpt_dir, fname) for fname in os.listdir(ckpt_dir) if fname.startswith("experiment_state") and fname.endswith(".json") ] return max(full_paths) class _TuneFunctionEncoder(json.JSONEncoder): def default(self, obj): if isinstance(obj, types.FunctionType): return self._to_cloudpickle(obj) try: return super(_TuneFunctionEncoder, self).default(obj) except Exception: logger.debug("Unable to encode. Falling back to cloudpickle.") return self._to_cloudpickle(obj) def _to_cloudpickle(self, obj): return { "_type": "CLOUDPICKLE_FALLBACK", "value": binary_to_hex(cloudpickle.dumps(obj)) } class _TuneFunctionDecoder(json.JSONDecoder): def __init__(self, *args, **kwargs): json.JSONDecoder.__init__( self, object_hook=self.object_hook, *args, **kwargs) def object_hook(self, obj): if obj.get("_type") == "CLOUDPICKLE_FALLBACK": return self._from_cloudpickle(obj) return obj def _from_cloudpickle(self, obj): return cloudpickle.loads(hex_to_binary(obj["value"])) class TrialRunner: """A TrialRunner implements the event loop for scheduling trials on Ray. Example: runner = TrialRunner() runner.add_trial(Trial(...)) runner.add_trial(Trial(...)) while not runner.is_finished(): runner.step() print(runner.debug_string()) The main job of TrialRunner is scheduling trials to efficiently use cluster resources, without overloading the cluster. While Ray itself provides resource management for tasks and actors, this is not sufficient when scheduling trials that may instantiate multiple actors. This is because if insufficient resources are available, concurrent trials could deadlock waiting for new resources to become available. Furthermore, oversubscribing the cluster could degrade training performance, leading to misleading benchmark results. """ CKPT_FILE_TMPL = "experiment_state-{}.json" VALID_RESUME_TYPES = [True, "LOCAL", "REMOTE", "PROMPT"] def __init__(self, search_alg=None, scheduler=None, launch_web_server=False, local_checkpoint_dir=None, remote_checkpoint_dir=None, sync_to_cloud=None, resume=False, server_port=TuneServer.DEFAULT_PORT, verbose=True, checkpoint_period=10, trial_executor=None): """Initializes a new TrialRunner. Args: search_alg (SearchAlgorithm): SearchAlgorithm for generating Trial objects. scheduler (TrialScheduler): Defaults to FIFOScheduler. launch_web_server (bool): Flag for starting TuneServer local_checkpoint_dir (str): Path where global checkpoints are stored and restored from. remote_checkpoint_dir (str): Remote path where global checkpoints are stored and restored from. Used if `resume` == REMOTE. resume (str|False): see `tune.py:run`. sync_to_cloud (func|str): See `tune.py:run`. server_port (int): Port number for launching TuneServer. verbose (bool): Flag for verbosity. If False, trial results will not be output. trial_executor (TrialExecutor): Defaults to RayTrialExecutor. """ self._search_alg = search_alg or BasicVariantGenerator() self._scheduler_alg = scheduler or FIFOScheduler() self.trial_executor = trial_executor or RayTrialExecutor() # For debugging, it may be useful to halt trials after some time has # elapsed. TODO(ekl) consider exposing this in the API. self._global_time_limit = float( os.environ.get("TRIALRUNNER_WALLTIME_LIMIT", float("inf"))) self._total_time = 0 self._iteration = 0 self._verbose = verbose self._server = None self._server_port = server_port if launch_web_server: self._server = TuneServer(self, self._server_port) self._trials = [] self._stop_queue = [] self._local_checkpoint_dir = local_checkpoint_dir if self._local_checkpoint_dir: os.makedirs(self._local_checkpoint_dir, exist_ok=True) self._remote_checkpoint_dir = remote_checkpoint_dir self._syncer = get_cloud_syncer(local_checkpoint_dir, remote_checkpoint_dir, sync_to_cloud) self._resumed = False if self._validate_resume(resume_type=resume): try: self.resume() logger.info("Resuming trial.") self._resumed = True except Exception: logger.exception( "Runner restore failed. Restarting experiment.") else: logger.debug("Starting a new experiment.") self._start_time = time.time() self._last_checkpoint_time = -float("inf") self._checkpoint_period = checkpoint_period self._session_str = datetime.fromtimestamp( self._start_time).strftime("%Y-%m-%d_%H-%M-%S") self.checkpoint_file = None if self._local_checkpoint_dir: self.checkpoint_file = os.path.join( self._local_checkpoint_dir, TrialRunner.CKPT_FILE_TMPL.format(self._session_str)) @property def scheduler_alg(self): return self._scheduler_alg def _validate_resume(self, resume_type): """Checks whether to resume experiment. Args: resume_type: One of True, "REMOTE", "LOCAL", "PROMPT". """ if not resume_type: return False assert resume_type in self.VALID_RESUME_TYPES, ( "resume_type {} is not one of {}".format(resume_type, self.VALID_RESUME_TYPES)) # Not clear if we need this assertion, since we should always have a # local checkpoint dir. assert self._local_checkpoint_dir or self._remote_checkpoint_dir if resume_type in [True, "LOCAL", "PROMPT"]: if not self.checkpoint_exists(self._local_checkpoint_dir): raise ValueError("Called resume when no checkpoint exists " "in local directory.") elif resume_type == "PROMPT": if click.confirm("Resume from local directory?"): return True if resume_type in ["REMOTE", "PROMPT"]: if resume_type == "PROMPT" and not click.confirm( "Try downloading from remote directory?"): return False if not self._remote_checkpoint_dir: raise ValueError( "Called resume from remote without remote directory.") # Try syncing down the upload directory. logger.info("Downloading from %s", self._remote_checkpoint_dir) # TODO(ujvl): Note that this syncs down the entire directory, # which may also contain trial checkpoints. We should selectively # sync the necessary files instead. self._syncer.sync_down_if_needed() if not self.checkpoint_exists(self._local_checkpoint_dir): raise ValueError("Called resume when no checkpoint exists " "in remote or local directory.") return True @classmethod def checkpoint_exists(cls, directory): if not os.path.exists(directory): return False return any( (fname.startswith("experiment_state") and fname.endswith(".json")) for fname in os.listdir(directory)) def add_experiment(self, experiment): if not self._resumed: self._search_alg.add_configurations([experiment]) else: logger.info("TrialRunner resumed, ignoring new add_experiment.") def checkpoint(self, force=False): """Saves execution state to `self._local_checkpoint_dir`. Overwrites the current session checkpoint, which starts when self is instantiated. Throttle depends on self._checkpoint_period. Args: force (bool): Forces a checkpoint despite checkpoint_period. """ if not self._local_checkpoint_dir: return now = time.time() if now - self._last_checkpoint_time < self._checkpoint_period and ( not force): return self._last_checkpoint_time = now runner_state = { "checkpoints": list( self.trial_executor.get_checkpoints().values()), "runner_data": self.__getstate__(), "stats": { "start_time": self._start_time, "timestamp": self._last_checkpoint_time } } tmp_file_name = os.path.join(self._local_checkpoint_dir, ".tmp_checkpoint") with open(tmp_file_name, "w") as f: json.dump(runner_state, f, indent=2, cls=_TuneFunctionEncoder) os.rename(tmp_file_name, self.checkpoint_file) if force: self._syncer.sync_up() else: self._syncer.sync_up_if_needed() return self._local_checkpoint_dir def resume(self): """Resumes all checkpointed trials from previous run. Requires user to manually re-register their objects. Also stops all ongoing trials. """ newest_ckpt_path = _find_newest_ckpt(self._local_checkpoint_dir) with open(newest_ckpt_path, "r") as f: runner_state = json.load(f, cls=_TuneFunctionDecoder) self.checkpoint_file = newest_ckpt_path logger.warning("".join([ "Attempting to resume experiment from {}. ".format( self._local_checkpoint_dir), "This feature is experimental, " "and may not work with all search algorithms. ", "This will ignore any new changes to the specification." ])) self.__setstate__(runner_state["runner_data"]) trials = [] for trial_cp in runner_state["checkpoints"]: new_trial = Trial(trial_cp["trainable_name"]) new_trial.__setstate__(trial_cp) trials += [new_trial] for trial in sorted( trials, key=lambda t: t.last_update_time, reverse=True): self.add_trial(trial) def is_finished(self): """Returns whether all trials have finished running.""" if self._total_time > self._global_time_limit: logger.warning("Exceeded global time limit {} / {}".format( self._total_time, self._global_time_limit)) return True trials_done = all(trial.is_finished() for trial in self._trials) return trials_done and self._search_alg.is_finished() def step(self): """Runs one step of the trial event loop. Callers should typically run this method repeatedly in a loop. They may inspect or modify the runner's state in between calls to step(). """ if self.is_finished(): raise TuneError("Called step when all trials finished?") with warn_if_slow("on_step_begin"): self.trial_executor.on_step_begin(self) next_trial = self._get_next_trial() # blocking if next_trial is not None: with warn_if_slow("start_trial"): self.trial_executor.start_trial(next_trial) elif self.trial_executor.get_running_trials(): self._process_events() # blocking else: self.trial_executor.on_no_available_trials(self) try: with warn_if_slow("experiment_checkpoint"): self.checkpoint() except Exception: logger.exception("Trial Runner checkpointing failed.") self._iteration += 1 if self._server: with warn_if_slow("server"): self._process_requests() if self.is_finished(): self._server.shutdown() with warn_if_slow("on_step_end"): self.trial_executor.on_step_end(self) def get_trial(self, tid): trial = [t for t in self._trials if t.trial_id == tid] return trial[0] if trial else None def get_trials(self): """Returns the list of trials managed by this TrialRunner. Note that the caller usually should not mutate trial state directly. """ return self._trials def add_trial(self, trial): """Adds a new trial to this TrialRunner. Trials may be added at any time. Args: trial (Trial): Trial to queue. """ trial.set_verbose(self._verbose) self._trials.append(trial) with warn_if_slow("scheduler.on_trial_add"): self._scheduler_alg.on_trial_add(self, trial) self.trial_executor.try_checkpoint_metadata(trial) def debug_string(self, delim="\n"): messages = [ self._scheduler_alg.debug_string(), self.trial_executor.debug_string(), trial_progress_str(self.get_trials()), ] return delim.join(messages) def has_resources(self, resources): """Returns whether this runner has at least the specified resources.""" return self.trial_executor.has_resources(resources) def _get_next_trial(self): """Replenishes queue. Blocks if all trials queued have finished, but search algorithm is still not finished. """ trials_done = all(trial.is_finished() for trial in self._trials) wait_for_trial = trials_done and not self._search_alg.is_finished() self._update_trial_queue(blocking=wait_for_trial) with warn_if_slow("choose_trial_to_run"): trial = self._scheduler_alg.choose_trial_to_run(self) return trial def _process_events(self): failed_trial = self.trial_executor.get_next_failed_trial() if failed_trial: error_msg = ( "{} (IP: {}) detected as stale. This is likely because the " "node was lost").format(failed_trial, failed_trial.node_ip) logger.info(error_msg) with warn_if_slow("process_failed_trial"): self._process_trial_failure(failed_trial, error_msg=error_msg) else: # TODO(ujvl): Consider combining get_next_available_trial and # fetch_result functionality so that we don't timeout on fetch. trial = self.trial_executor.get_next_available_trial() # blocking if trial.is_restoring: with warn_if_slow("process_trial_restore"): self._process_trial_restore(trial) else: with warn_if_slow("process_trial"): self._process_trial(trial) def _process_trial(self, trial): """Processes a trial result.""" try: result = self.trial_executor.fetch_result(trial) is_duplicate = RESULT_DUPLICATE in result # TrialScheduler and SearchAlgorithm still receive a # notification because there may be special handling for # the `on_trial_complete` hook. if is_duplicate: logger.debug("Trial finished without logging 'done'.") result = trial.last_result result.update(done=True) self._total_time += result.get(TIME_THIS_ITER_S, 0) flat_result = flatten_dict(result) if trial.should_stop(flat_result): # Hook into scheduler self._scheduler_alg.on_trial_complete(self, trial, flat_result) self._search_alg.on_trial_complete( trial.trial_id, result=flat_result) decision = TrialScheduler.STOP else: with warn_if_slow("scheduler.on_trial_result"): decision = self._scheduler_alg.on_trial_result( self, trial, flat_result) with warn_if_slow("search_alg.on_trial_result"): self._search_alg.on_trial_result(trial.trial_id, flat_result) if decision == TrialScheduler.STOP: with warn_if_slow("search_alg.on_trial_complete"): self._search_alg.on_trial_complete( trial.trial_id, result=flat_result, early_terminated=True) if not is_duplicate: trial.update_last_result( result, terminate=(decision == TrialScheduler.STOP)) # Checkpoints to disk. This should be checked even if # the scheduler decision is STOP or PAUSE. Note that # PAUSE only checkpoints to memory and does not update # the global checkpoint state. self._checkpoint_trial_if_needed( trial, force=result.get(SHOULD_CHECKPOINT, False)) if decision == TrialScheduler.CONTINUE: self.trial_executor.continue_training(trial) elif decision == TrialScheduler.PAUSE: self.trial_executor.pause_trial(trial) elif decision == TrialScheduler.STOP: self.trial_executor.export_trial_if_needed(trial) self.trial_executor.stop_trial(trial) else: assert False, "Invalid scheduling decision: {}".format( decision) except Exception: logger.exception("Trial %s: Error processing event.", trial) self._process_trial_failure(trial, traceback.format_exc()) def _process_trial_restore(self, trial): """Processes a trial restore. Args: trial: Trial being restored. """ logger.debug("Trial %s: Processing trial restore.", trial) try: self.trial_executor.fetch_result(trial) trial.on_restore() logger.debug("Trial %s: Restore processed successfully", trial) self.trial_executor.set_status(trial, Trial.RUNNING) self.trial_executor.continue_training(trial) except Exception: logger.exception("Trial %s: Error processing restore.", trial) self._process_trial_failure(trial, traceback.format_exc()) def _process_trial_failure(self, trial, error_msg): """Handle trial failure. Attempt trial recovery if possible, clean up state otherwise. Args: trial (Trial): Failed trial. error_msg (str): Error message prior to invoking this method. """ if trial.status == Trial.RUNNING: if trial.should_recover(): self._try_recover(trial, error_msg) else: self._scheduler_alg.on_trial_error(self, trial) self._search_alg.on_trial_complete(trial.trial_id, error=True) self.trial_executor.stop_trial( trial, error=True, error_msg=error_msg) def _checkpoint_trial_if_needed(self, trial, force=False): """Checkpoints trial based off trial.last_result.""" if trial.should_checkpoint() or force: # Save trial runtime if possible if trial.runner: self.trial_executor.save(trial, storage=Checkpoint.PERSISTENT) self.trial_executor.try_checkpoint_metadata(trial) def _try_recover(self, trial, error_msg): """Tries to recover trial. Notifies SearchAlgorithm and Scheduler if failure to recover. Args: trial (Trial): Trial to recover. error_msg (str): Error message from prior to invoking this method. """ if trial.is_restoring: # Restore was unsuccessful, try again without checkpoint. trial.clear_checkpoint() self.trial_executor.stop_trial( trial, error=error_msg is not None, error_msg=error_msg, stop_logger=False) trial.result_logger.flush() if self.trial_executor.has_resources(trial.resources): logger.info( "Trial %s: Attempting to restore " "trial state from last checkpoint.", trial) self.trial_executor.start_trial(trial) if trial.status == Trial.ERROR: logger.exception( "Trial %s: Error restoring trial from checkpoint, abort.", trial) self._scheduler_alg.on_trial_error(self, trial) self._search_alg.on_trial_complete(trial.trial_id, error=True) else: logger.debug("Trial %s: Restore dispatched correctly.", trial) else: logger.debug("Trial %s: Notifying Scheduler and requeueing.", trial) self._requeue_trial(trial) def _requeue_trial(self, trial): """Notification to TrialScheduler and requeue trial. This does not notify the SearchAlgorithm because the function evaluation is still in progress. """ self._scheduler_alg.on_trial_error(self, trial) self.trial_executor.set_status(trial, Trial.PENDING) # TODO(rliaw): Right now, this pushes the trial to the end of queue # because restoration can be expensive. However, this is not # ideal since it just hides the issue - a better fix would # be to use an actor table to detect the IP of the Trainable # and rsync the files there. # See https://github.com/ray-project/ray/issues/5168 self._trials.pop(self._trials.index(trial)) self._trials.append(trial) with warn_if_slow("scheduler.on_trial_add"): self._scheduler_alg.on_trial_add(self, trial) def _update_trial_queue(self, blocking=False, timeout=600): """Adds next trials to queue if possible. Note that the timeout is currently unexposed to the user. Args: blocking (bool): Blocks until either a trial is available or is_finished (timeout or search algorithm finishes). timeout (int): Seconds before blocking times out. """ trials = self._search_alg.next_trials() if blocking and not trials: start = time.time() # Checking `is_finished` instead of _search_alg.is_finished # is fine because blocking only occurs if all trials are # finished and search_algorithm is not yet finished while (not trials and not self.is_finished() and time.time() - start < timeout): logger.info("Blocking for next trial...") trials = self._search_alg.next_trials() time.sleep(1) for trial in trials: self.add_trial(trial) def request_stop_trial(self, trial): self._stop_queue.append(trial) def _process_requests(self): while self._stop_queue: t = self._stop_queue.pop() self.stop_trial(t) def stop_trial(self, trial): """Stops trial. Trials may be stopped at any time. If trial is in state PENDING or PAUSED, calls `on_trial_remove` for scheduler and `on_trial_complete(..., early_terminated=True) for search_alg. Otherwise waits for result for the trial and calls `on_trial_complete` for scheduler and search_alg if RUNNING. """ error = False error_msg = None if trial.status in [Trial.ERROR, Trial.TERMINATED]: return elif trial.status in [Trial.PENDING, Trial.PAUSED]: self._scheduler_alg.on_trial_remove(self, trial) self._search_alg.on_trial_complete( trial.trial_id, early_terminated=True) elif trial.status is Trial.RUNNING: try: result = self.trial_executor.fetch_result(trial) trial.update_last_result(result, terminate=True) self._scheduler_alg.on_trial_complete(self, trial, result) self._search_alg.on_trial_complete( trial.trial_id, result=result) except Exception: error_msg = traceback.format_exc() logger.exception("Error processing event.") self._scheduler_alg.on_trial_error(self, trial) self._search_alg.on_trial_complete(trial.trial_id, error=True) error = True self.trial_executor.stop_trial(trial, error=error, error_msg=error_msg) def __getstate__(self): """Gets state for trial. Note that this is not used as a pickling override as does not have all fields. """ state = self.__dict__.copy() for k in [ "_trials", "_stop_queue", "_server", "_search_alg", "_scheduler_alg", "trial_executor", "_syncer", ]: del state[k] state["launch_web_server"] = bool(self._server) return state def __setstate__(self, state): launch_web_server = state.pop("launch_web_server") # Use session_str from previous checkpoint if does not exist session_str = state.pop("_session_str") self.__dict__.setdefault("_session_str", session_str) # Use start_time from previous checkpoint if does not exist start_time = state.pop("_start_time") self.__dict__.setdefault("_start_time", start_time) self.__dict__.update(state) if launch_web_server: self._server = TuneServer(self, self._server_port)
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/tune.py
Python
import logging import time import six from ray.tune.error import TuneError from ray.tune.experiment import convert_to_experiment_list, Experiment from ray.tune.analysis import ExperimentAnalysis from ray.tune.suggest import BasicVariantGenerator from ray.tune.trial import Trial, DEBUG_PRINT_INTERVAL from ray.tune.trainable import Trainable from ray.tune.ray_trial_executor import RayTrialExecutor from ray.tune.registry import get_trainable_cls from ray.tune.syncer import wait_for_sync from ray.tune.trial_runner import TrialRunner from ray.tune.progress_reporter import CLIReporter, JupyterNotebookReporter from ray.tune.schedulers import (HyperBandScheduler, AsyncHyperBandScheduler, FIFOScheduler, MedianStoppingRule) from ray.tune.web_server import TuneServer logger = logging.getLogger(__name__) _SCHEDULERS = { "FIFO": FIFOScheduler, "MedianStopping": MedianStoppingRule, "HyperBand": HyperBandScheduler, "AsyncHyperBand": AsyncHyperBandScheduler, } try: class_name = get_ipython().__class__.__name__ IS_NOTEBOOK = True if "Terminal" not in class_name else False except NameError: IS_NOTEBOOK = False def _make_scheduler(args): if args.scheduler in _SCHEDULERS: return _SCHEDULERS[args.scheduler](**args.scheduler_config) else: raise TuneError("Unknown scheduler: {}, should be one of {}".format( args.scheduler, _SCHEDULERS.keys())) def _check_default_resources_override(run_identifier): if not isinstance(run_identifier, six.string_types): # If obscure dtype, assume it is overriden. return True trainable_cls = get_trainable_cls(run_identifier) return hasattr(trainable_cls, "default_resource_request") and ( trainable_cls.default_resource_request.__code__ != Trainable.default_resource_request.__code__) def run(run_or_experiment, name=None, stop=None, config=None, resources_per_trial=None, num_samples=1, local_dir=None, upload_dir=None, trial_name_creator=None, loggers=None, sync_to_cloud=None, sync_to_driver=None, checkpoint_freq=0, checkpoint_at_end=False, sync_on_checkpoint=True, keep_checkpoints_num=None, checkpoint_score_attr=None, global_checkpoint_period=10, export_formats=None, max_failures=0, restore=None, search_alg=None, scheduler=None, with_server=False, server_port=TuneServer.DEFAULT_PORT, verbose=2, resume=False, queue_trials=False, reuse_actors=False, trial_executor=None, raise_on_failed_trial=True, return_trials=False, ray_auto_init=True, sync_function=None): """Executes training. Args: run_or_experiment (function|class|str|Experiment): If function|class|str, this is the algorithm or model to train. This may refer to the name of a built-on algorithm (e.g. RLLib's DQN or PPO), a user-defined trainable function or class, or the string identifier of a trainable function or class registered in the tune registry. If Experiment, then Tune will execute training based on Experiment.spec. name (str): Name of experiment. stop (dict|func): The stopping criteria. If dict, the keys may be any field in the return result of 'train()', whichever is reached first. If function, it must take (trial_id, result) as arguments and return a boolean (True if trial should be stopped, False otherwise). config (dict): Algorithm-specific configuration for Tune variant generation (e.g. env, hyperparams). Defaults to empty dict. Custom search algorithms may ignore this. resources_per_trial (dict): Machine resources to allocate per trial, e.g. ``{"cpu": 64, "gpu": 8}``. Note that GPUs will not be assigned unless you specify them here. Defaults to 1 CPU and 0 GPUs in ``Trainable.default_resource_request()``. num_samples (int): Number of times to sample from the hyperparameter space. Defaults to 1. If `grid_search` is provided as an argument, the grid will be repeated `num_samples` of times. local_dir (str): Local dir to save training results to. Defaults to ``~/ray_results``. upload_dir (str): Optional URI to sync training results and checkpoints to (e.g. ``s3://bucket`` or ``gs://bucket``). trial_name_creator (func): Optional function for generating the trial string representation. loggers (list): List of logger creators to be used with each Trial. If None, defaults to ray.tune.logger.DEFAULT_LOGGERS. See `ray/tune/logger.py`. sync_to_cloud (func|str): Function for syncing the local_dir to and from upload_dir. If string, then it must be a string template that includes `{source}` and `{target}` for the syncer to run. If not provided, the sync command defaults to standard S3 or gsutil sync commands. sync_to_driver (func|str|bool): Function for syncing trial logdir from remote node to local. If string, then it must be a string template that includes `{source}` and `{target}` for the syncer to run. If True or not provided, it defaults to using rsync. If False, syncing to driver is disabled. checkpoint_freq (int): How many training iterations between checkpoints. A value of 0 (default) disables checkpointing. checkpoint_at_end (bool): Whether to checkpoint at the end of the experiment regardless of the checkpoint_freq. Default is False. sync_on_checkpoint (bool): Force sync-down of trial checkpoint to driver. If set to False, checkpoint syncing from worker to driver is asynchronous and best-effort. This does not affect persistent storage syncing. Defaults to True. keep_checkpoints_num (int): Number of checkpoints to keep. A value of `None` keeps all checkpoints. Defaults to `None`. If set, need to provide `checkpoint_score_attr`. checkpoint_score_attr (str): Specifies by which attribute to rank the best checkpoint. Default is increasing order. If attribute starts with `min-` it will rank attribute in decreasing order, i.e. `min-validation_loss`. global_checkpoint_period (int): Seconds between global checkpointing. This does not affect `checkpoint_freq`, which specifies frequency for individual trials. export_formats (list): List of formats that exported at the end of the experiment. Default is None. max_failures (int): Try to recover a trial at least this many times. Ray will recover from the latest checkpoint if present. Setting to -1 will lead to infinite recovery retries. Setting to 0 will disable retries. Defaults to 3. restore (str): Path to checkpoint. Only makes sense to set if running 1 trial. Defaults to None. search_alg (SearchAlgorithm): Search Algorithm. Defaults to BasicVariantGenerator. scheduler (TrialScheduler): Scheduler for executing the experiment. Choose among FIFO (default), MedianStopping, AsyncHyperBand, HyperBand and PopulationBasedTraining. Refer to ray.tune.schedulers for more options. with_server (bool): Starts a background Tune server. Needed for using the Client API. server_port (int): Port number for launching TuneServer. verbose (int): 0, 1, or 2. Verbosity mode. 0 = silent, 1 = only status updates, 2 = status and trial results. resume (str|bool): One of "LOCAL", "REMOTE", "PROMPT", or bool. LOCAL/True restores the checkpoint from the local_checkpoint_dir. REMOTE restores the checkpoint from remote_checkpoint_dir. PROMPT provides CLI feedback. False forces a new experiment. If resume is set but checkpoint does not exist, ValueError will be thrown. queue_trials (bool): Whether to queue trials when the cluster does not currently have enough resources to launch one. This should be set to True when running on an autoscaling cluster to enable automatic scale-up. reuse_actors (bool): Whether to reuse actors between different trials when possible. This can drastically speed up experiments that start and stop actors often (e.g., PBT in time-multiplexing mode). This requires trials to have the same resource requirements. trial_executor (TrialExecutor): Manage the execution of trials. raise_on_failed_trial (bool): Raise TuneError if there exists failed trial (of ERROR state) when the experiments complete. ray_auto_init (bool): Automatically starts a local Ray cluster if using a RayTrialExecutor (which is the default) and if Ray is not initialized. Defaults to True. sync_function: Deprecated. See `sync_to_cloud` and `sync_to_driver`. Returns: List of Trial objects. Raises: TuneError if any trials failed and `raise_on_failed_trial` is True. Examples: >>> tune.run(mytrainable, scheduler=PopulationBasedTraining()) >>> tune.run(mytrainable, num_samples=5, reuse_actors=True) >>> tune.run( >>> "PG", >>> num_samples=5, >>> config={ >>> "env": "CartPole-v0", >>> "lr": tune.sample_from(lambda _: np.random.rand()) >>> } >>> ) """ trial_executor = trial_executor or RayTrialExecutor( queue_trials=queue_trials, reuse_actors=reuse_actors, ray_auto_init=ray_auto_init) if isinstance(run_or_experiment, list): experiments = run_or_experiment else: experiments = [run_or_experiment] if len(experiments) > 1: logger.info( "Running multiple concurrent experiments is experimental and may " "not work with certain features.") for i, exp in enumerate(experiments): if not isinstance(exp, Experiment): run_identifier = Experiment.register_if_needed(exp) experiments[i] = Experiment( name=name, run=run_identifier, stop=stop, config=config, resources_per_trial=resources_per_trial, num_samples=num_samples, local_dir=local_dir, upload_dir=upload_dir, sync_to_driver=sync_to_driver, trial_name_creator=trial_name_creator, loggers=loggers, checkpoint_freq=checkpoint_freq, checkpoint_at_end=checkpoint_at_end, sync_on_checkpoint=sync_on_checkpoint, keep_checkpoints_num=keep_checkpoints_num, checkpoint_score_attr=checkpoint_score_attr, export_formats=export_formats, max_failures=max_failures, restore=restore, sync_function=sync_function) else: logger.debug("Ignoring some parameters passed into tune.run.") if sync_to_cloud: for exp in experiments: assert exp.remote_checkpoint_dir, ( "Need `upload_dir` if `sync_to_cloud` given.") runner = TrialRunner( search_alg=search_alg or BasicVariantGenerator(), scheduler=scheduler or FIFOScheduler(), local_checkpoint_dir=experiments[0].checkpoint_dir, remote_checkpoint_dir=experiments[0].remote_checkpoint_dir, sync_to_cloud=sync_to_cloud, checkpoint_period=global_checkpoint_period, resume=resume, launch_web_server=with_server, server_port=server_port, verbose=bool(verbose > 1), trial_executor=trial_executor) for exp in experiments: runner.add_experiment(exp) if IS_NOTEBOOK: reporter = JupyterNotebookReporter(overwrite=verbose < 2) else: reporter = CLIReporter() # User Warning for GPUs if trial_executor.has_gpus(): if isinstance(resources_per_trial, dict) and "gpu" in resources_per_trial: # "gpu" is manually set. pass elif _check_default_resources_override(experiments[0].run_identifier): # "default_resources" is manually overriden. pass else: logger.warning("Tune detects GPUs, but no trials are using GPUs. " "To enable trials to use GPUs, set " "tune.run(resources_per_trial={'gpu': 1}...) " "which allows Tune to expose 1 GPU to each trial. " "You can also override " "`Trainable.default_resource_request` if using the " "Trainable API.") last_debug = 0 while not runner.is_finished(): runner.step() if time.time() - last_debug > DEBUG_PRINT_INTERVAL: if verbose: reporter.report(runner) last_debug = time.time() try: runner.checkpoint(force=True) except Exception: logger.exception("Trial Runner checkpointing failed.") if verbose: reporter.report(runner) wait_for_sync() errored_trials = [] for trial in runner.get_trials(): if trial.status != Trial.TERMINATED: errored_trials += [trial] if errored_trials: if raise_on_failed_trial: raise TuneError("Trials did not complete", errored_trials) else: logger.error("Trials did not complete: %s", errored_trials) trials = runner.get_trials() if return_trials: return trials logger.info("Returning an analysis object by default. You can call " "`analysis.trials` to retrieve a list of trials. " "This message will be removed in future versions of Tune.") return ExperimentAnalysis(runner.checkpoint_file, trials=trials) def run_experiments(experiments, search_alg=None, scheduler=None, with_server=False, server_port=TuneServer.DEFAULT_PORT, verbose=2, resume=False, queue_trials=False, reuse_actors=False, trial_executor=None, raise_on_failed_trial=True, concurrent=False): """Runs and blocks until all trials finish. Examples: >>> experiment_spec = Experiment("experiment", my_func) >>> run_experiments(experiments=experiment_spec) >>> experiment_spec = {"experiment": {"run": my_func}} >>> run_experiments(experiments=experiment_spec) >>> run_experiments( >>> experiments=experiment_spec, >>> scheduler=MedianStoppingRule(...)) >>> run_experiments( >>> experiments=experiment_spec, >>> search_alg=SearchAlgorithm(), >>> scheduler=MedianStoppingRule(...)) Returns: List of Trial objects, holding data for each executed trial. """ # This is important to do this here # because it schematize the experiments # and it conducts the implicit registration. experiments = convert_to_experiment_list(experiments) if concurrent: return run( experiments, search_alg=search_alg, scheduler=scheduler, with_server=with_server, server_port=server_port, verbose=verbose, resume=resume, queue_trials=queue_trials, reuse_actors=reuse_actors, trial_executor=trial_executor, raise_on_failed_trial=raise_on_failed_trial, return_trials=True) else: trials = [] for exp in experiments: trials += run( exp, search_alg=search_alg, scheduler=scheduler, with_server=with_server, server_port=server_port, verbose=verbose, resume=resume, queue_trials=queue_trials, reuse_actors=reuse_actors, trial_executor=trial_executor, raise_on_failed_trial=raise_on_failed_trial, return_trials=True) return trials
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/utils/__init__.py
Python
from ray.tune.utils.util import (deep_update, flatten_dict, get_pinned_object, merge_dicts, pin_in_object_store, UtilMonitor, validate_save_restore, warn_if_slow) __all__ = [ "deep_update", "flatten_dict", "get_pinned_object", "merge_dicts", "pin_in_object_store", "UtilMonitor", "validate_save_restore", "warn_if_slow" ]
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/utils/mock.py
Python
import os from ray.rllib.agents.mock import _MockTrainer from ray.tune import DurableTrainable from ray.tune.sync_client import get_sync_client from ray.tune.syncer import NodeSyncer MOCK_REMOTE_DIR = "/tmp/mock-tune-remote/" # Sync and delete templates that operate on local directories. LOCAL_SYNC_TEMPLATE = "mkdir -p {target} && rsync -avz {source}/ {target}/" LOCAL_DELETE_TEMPLATE = "rm -rf {target}" def mock_storage_client(): """Mocks storage client that treats a local dir as durable storage.""" return get_sync_client(LOCAL_SYNC_TEMPLATE, LOCAL_DELETE_TEMPLATE) class MockNodeSyncer(NodeSyncer): """Mock NodeSyncer that syncs to and from /tmp""" def has_remote_target(self): return True @property def _remote_path(self): if self._remote_dir.startswith("/"): self._remote_dir = self._remote_dir[1:] return os.path.join(MOCK_REMOTE_DIR, self._remote_dir) class MockRemoteTrainer(_MockTrainer): """Mock Trainable that saves at tmp for simulated clusters.""" def __init__(self, *args, **kwargs): super(MockRemoteTrainer, self).__init__(*args, **kwargs) if self._logdir.startswith("/"): self._logdir = self._logdir[1:] self._logdir = os.path.join(MOCK_REMOTE_DIR, self._logdir) if not os.path.exists(self._logdir): os.makedirs(self._logdir) class MockDurableTrainer(DurableTrainable, _MockTrainer): """Mock DurableTrainable that saves at tmp for simulated clusters.""" # TODO(ujvl): This class uses multiple inheritance; it should be cleaned # up once the durable training API converges. def __init__(self, remote_checkpoint_dir, *args, **kwargs): _MockTrainer.__init__(self, *args, **kwargs) DurableTrainable.__init__(self, remote_checkpoint_dir, *args, **kwargs) def _create_storage_client(self): return mock_storage_client()
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/utils/util.py
Python
import copy import logging import threading import time from collections import defaultdict from threading import Thread import numpy as np import ray logger = logging.getLogger(__name__) try: import psutil except ImportError: psutil = None try: import GPUtil except ImportError: GPUtil = None _pinned_objects = [] PINNED_OBJECT_PREFIX = "ray.tune.PinnedObject:" START_OF_TIME = time.time() class UtilMonitor(Thread): """Class for system usage utilization monitoring. It keeps track of CPU, RAM, GPU, VRAM usage (each gpu separately) by pinging for information every x seconds in a separate thread. Requires psutil and GPUtil to be installed. Can be enabled with tune.run(config={"log_sys_usage": True}). """ def __init__(self, start=True, delay=0.7): self.stopped = True if GPUtil is None and start: logger.warning("Install gputil for GPU system monitoring.") if psutil is None and start: logger.warning("Install psutil to monitor system performance.") if GPUtil is None and psutil is None: return super(UtilMonitor, self).__init__() self.delay = delay # Time between calls to GPUtil self.values = defaultdict(list) self.lock = threading.Lock() self.daemon = True if start: self.start() def _read_utilization(self): with self.lock: if psutil is not None: self.values["cpu_util_percent"].append( float(psutil.cpu_percent(interval=None))) self.values["ram_util_percent"].append( float(getattr(psutil.virtual_memory(), "percent"))) if GPUtil is not None: for gpu in GPUtil.getGPUs(): self.values["gpu_util_percent" + str(gpu.id)].append( float(gpu.load)) self.values["vram_util_percent" + str(gpu.id)].append( float(gpu.memoryUtil)) def get_data(self): if self.stopped: return {} with self.lock: ret_values = copy.deepcopy(self.values) for key, val in self.values.items(): del val[:] return { "perf": { k: np.mean(v) for k, v in ret_values.items() if len(v) > 0 } } def run(self): self.stopped = False while not self.stopped: self._read_utilization() time.sleep(self.delay) def stop(self): self.stopped = True def pin_in_object_store(obj): """Deprecated, use ray.put(value, weakref=False) instead.""" obj_id = ray.put(obj, weakref=False) _pinned_objects.append(obj_id) return obj_id def get_pinned_object(pinned_id): """Deprecated.""" return ray.get(pinned_id) class warn_if_slow: """Prints a warning if a given operation is slower than 100ms. Example: >>> with warn_if_slow("some_operation"): ... ray.get(something) """ DEFAULT_THRESHOLD = 0.5 def __init__(self, name, threshold=None): self.name = name self.threshold = threshold or self.DEFAULT_THRESHOLD self.too_slow = False def __enter__(self): self.start = time.time() return self def __exit__(self, type, value, traceback): now = time.time() if now - self.start > self.threshold and now - START_OF_TIME > 60.0: self.too_slow = True logger.warning( "The `%s` operation took %s seconds to complete, " "which may be a performance bottleneck.", self.name, now - self.start) def merge_dicts(d1, d2): """Returns a new dict that is d1 and d2 deep merged.""" merged = copy.deepcopy(d1) deep_update(merged, d2, True, []) return merged def deep_update(original, new_dict, new_keys_allowed, whitelist): """Updates original dict with values from new_dict recursively. If new key is introduced in new_dict, then if new_keys_allowed is not True, an error will be thrown. Further, for sub-dicts, if the key is in the whitelist, then new subkeys can be introduced. Args: original (dict): Dictionary with default values. new_dict (dict): Dictionary with values to be updated new_keys_allowed (bool): Whether new keys are allowed. whitelist (list): List of keys that correspond to dict values where new subkeys can be introduced. This is only at the top level. """ for k, value in new_dict.items(): if k not in original: if not new_keys_allowed: raise Exception("Unknown config parameter `{}` ".format(k)) if isinstance(original.get(k), dict): if k in whitelist: deep_update(original[k], value, True, []) else: deep_update(original[k], value, new_keys_allowed, []) else: original[k] = value return original def flatten_dict(dt, delimiter="/"): dt = copy.deepcopy(dt) while any(isinstance(v, dict) for v in dt.values()): remove = [] add = {} for key, value in dt.items(): if isinstance(value, dict): for subkey, v in value.items(): add[delimiter.join([key, subkey])] = v remove.append(key) dt.update(add) for k in remove: del dt[k] return dt def _to_pinnable(obj): """Converts obj to a form that can be pinned in object store memory. Currently only numpy arrays are pinned in memory, if you have a strong reference to the array value. """ return (obj, np.zeros(1)) def _from_pinnable(obj): """Retrieve from _to_pinnable format.""" return obj[0] def validate_save_restore(trainable_cls, config=None, num_gpus=0, use_object_store=False): """Helper method to check if your Trainable class will resume correctly. Args: trainable_cls: Trainable class for evaluation. config (dict): Config to pass to Trainable when testing. num_gpus (int): GPU resources to allocate when testing. use_object_store (bool): Whether to save and restore to Ray's object store. Recommended to set this to True if planning to use algorithms that pause training (i.e., PBT, HyperBand). """ assert ray.is_initialized(), "Need Ray to be initialized." remote_cls = ray.remote(num_gpus=num_gpus)(trainable_cls) trainable_1 = remote_cls.remote(config=config) trainable_2 = remote_cls.remote(config=config) from ray.tune.result import TRAINING_ITERATION for _ in range(3): res = ray.get(trainable_1.train.remote()) assert res.get(TRAINING_ITERATION), ( "Validation will not pass because it requires `training_iteration` " "to be returned.") if use_object_store: restore_check = trainable_2.restore_from_object.remote( trainable_1.save_to_object.remote()) ray.get(restore_check) else: restore_check = ray.get( trainable_2.restore.remote(trainable_1.save.remote())) res = ray.get(trainable_2.train.remote()) assert res[TRAINING_ITERATION] == 4 res = ray.get(trainable_2.train.remote()) assert res[TRAINING_ITERATION] == 5 return True if __name__ == "__main__": ray.init() X = pin_in_object_store("hello") print(X) result = get_pinned_object(X) print(result)
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/utils/visual_utils.py
Python
import pandas as pd from pandas.api.types import is_string_dtype, is_numeric_dtype import logging import os import os.path as osp import numpy as np import json from ray.tune.utils import flatten_dict logger = logging.getLogger(__name__) logger.warning("This module will be deprecated in a future version of Tune.") def _parse_results(res_path): res_dict = {} try: with open(res_path) as f: # Get last line in file for line in f: pass res_dict = flatten_dict(json.loads(line.strip())) except Exception: logger.exception("Importing %s failed...Perhaps empty?" % res_path) return res_dict def _parse_configs(cfg_path): try: with open(cfg_path) as f: cfg_dict = flatten_dict(json.load(f)) except Exception: logger.exception("Config parsing failed.") return cfg_dict def _resolve(directory, result_fname): try: resultp = osp.join(directory, result_fname) res_dict = _parse_results(resultp) cfgp = osp.join(directory, "params.json") cfg_dict = _parse_configs(cfgp) cfg_dict.update(res_dict) return cfg_dict except Exception: return None def load_results_to_df(directory, result_name="result.json"): exp_directories = [ dirpath for dirpath, dirs, files in os.walk(directory) for f in files if f == result_name ] data = [_resolve(d, result_name) for d in exp_directories] data = [d for d in data if d] return pd.DataFrame(data) def generate_plotly_dim_dict(df, field): dim_dict = {} dim_dict["label"] = field column = df[field] if is_numeric_dtype(column): dim_dict["values"] = column elif is_string_dtype(column): texts = column.unique() dim_dict["values"] = [ np.argwhere(texts == x).flatten()[0] for x in column ] dim_dict["tickvals"] = list(range(len(texts))) dim_dict["ticktext"] = texts else: raise Exception("Unidentifiable Type") return dim_dict
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/tune/web_server.py
Python
import json import logging import threading from urllib.parse import urljoin, urlparse from http.server import SimpleHTTPRequestHandler, HTTPServer import ray.cloudpickle as cloudpickle from ray.tune import TuneError from ray.tune.suggest import BasicVariantGenerator from ray.utils import binary_to_hex, hex_to_binary logger = logging.getLogger(__name__) try: import requests # `requests` is not part of stdlib. except ImportError: requests = None logger.exception("Couldn't import `requests` library. " "Be sure to install it on the client side.") class TuneClient: """Client to interact with an ongoing Tune experiment. Requires a TuneServer to have started running. Attributes: tune_address (str): Address of running TuneServer port_forward (int): Port number of running TuneServer """ def __init__(self, tune_address, port_forward): self._tune_address = tune_address self._port_forward = port_forward self._path = "http://{}:{}".format(tune_address, port_forward) def get_all_trials(self): """Returns a list of all trials' information.""" response = requests.get(urljoin(self._path, "trials")) return self._deserialize(response) def get_trial(self, trial_id): """Returns trial information by trial_id.""" response = requests.get( urljoin(self._path, "trials/{}".format(trial_id))) return self._deserialize(response) def add_trial(self, name, specification): """Adds a trial by name and specification (dict).""" payload = {"name": name, "spec": specification} response = requests.post(urljoin(self._path, "trials"), json=payload) return self._deserialize(response) def stop_trial(self, trial_id): """Requests to stop trial by trial_id.""" response = requests.put( urljoin(self._path, "trials/{}".format(trial_id))) return self._deserialize(response) @property def server_address(self): return self._tune_address @property def server_port(self): return self._port_forward def _load_trial_info(self, trial_info): trial_info["config"] = cloudpickle.loads( hex_to_binary(trial_info["config"])) trial_info["result"] = cloudpickle.loads( hex_to_binary(trial_info["result"])) def _deserialize(self, response): parsed = response.json() if "trial" in parsed: self._load_trial_info(parsed["trial"]) elif "trials" in parsed: for trial_info in parsed["trials"]: self._load_trial_info(trial_info) return parsed def RunnerHandler(runner): class Handler(SimpleHTTPRequestHandler): """A Handler is a custom handler for TuneServer. Handles all requests and responses coming into and from the TuneServer. """ def _do_header(self, response_code=200, headers=None): """Sends the header portion of the HTTP response. Parameters: response_code (int): Standard HTTP response code headers (list[tuples]): Standard HTTP response headers """ if headers is None: headers = [("Content-type", "application/json")] self.send_response(response_code) for key, value in headers: self.send_header(key, value) self.end_headers() def do_HEAD(self): """HTTP HEAD handler method.""" self._do_header() def do_GET(self): """HTTP GET handler method.""" response_code = 200 message = "" try: result = self._get_trial_by_url(self.path) resource = {} if result: if isinstance(result, list): infos = [self._trial_info(t) for t in result] resource["trials"] = infos else: resource["trial"] = self._trial_info(result) message = json.dumps(resource) except TuneError as e: response_code = 404 message = str(e) self._do_header(response_code=response_code) self.wfile.write(message.encode()) def do_PUT(self): """HTTP PUT handler method.""" response_code = 200 message = "" try: result = self._get_trial_by_url(self.path) resource = {} if result: if isinstance(result, list): infos = [self._trial_info(t) for t in result] resource["trials"] = infos for t in result: runner.request_stop_trial(t) else: resource["trial"] = self._trial_info(result) runner.request_stop_trial(result) message = json.dumps(resource) except TuneError as e: response_code = 404 message = str(e) self._do_header(response_code=response_code) self.wfile.write(message.encode()) def do_POST(self): """HTTP POST handler method.""" response_code = 201 content_len = int(self.headers.get("Content-Length"), 0) raw_body = self.rfile.read(content_len) parsed_input = json.loads(raw_body.decode()) resource = self._add_trials(parsed_input["name"], parsed_input["spec"]) headers = [("Content-type", "application/json"), ("Location", "/trials/")] self._do_header(response_code=response_code, headers=headers) self.wfile.write(json.dumps(resource).encode()) def _trial_info(self, trial): """Returns trial information as JSON.""" if trial.last_result: result = trial.last_result.copy() else: result = None info_dict = { "id": trial.trial_id, "trainable_name": trial.trainable_name, "config": binary_to_hex(cloudpickle.dumps(trial.config)), "status": trial.status, "result": binary_to_hex(cloudpickle.dumps(result)) } return info_dict def _get_trial_by_url(self, url): """Parses url to get either all trials or trial by trial_id.""" parts = urlparse(url) path = parts.path if path == "/trials": return list(runner.get_trials()) else: trial_id = path.split("/")[-1] return runner.get_trial(trial_id) def _add_trials(self, name, spec): """Add trial by invoking TrialRunner.""" resource = {} resource["trials"] = [] trial_generator = BasicVariantGenerator() trial_generator.add_configurations({name: spec}) for trial in trial_generator.next_trials(): runner.add_trial(trial) resource["trials"].append(self._trial_info(trial)) return resource return Handler class TuneServer(threading.Thread): """A TuneServer is a thread that initializes and runs a HTTPServer. The server handles requests from a TuneClient. Attributes: runner (TrialRunner): Runner that modifies and accesses trials. port_forward (int): Port number of TuneServer. """ DEFAULT_PORT = 4321 def __init__(self, runner, port=None): """Initialize HTTPServer and serve forever by invoking self.run()""" threading.Thread.__init__(self) self._port = port if port else self.DEFAULT_PORT address = ("localhost", self._port) logger.info("Starting Tune Server...") self._server = HTTPServer(address, RunnerHandler(runner)) self.daemon = True self.start() def run(self): self._server.serve_forever() def shutdown(self): """Shutdown the underlying server.""" self._server.shutdown()
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/utils.py
Python
import binascii import errno import hashlib import inspect import logging import numpy as np import os import six import subprocess import sys import threading import time import uuid import ray.gcs_utils import ray.ray_constants as ray_constants def _random_string(): id_hash = hashlib.sha1() id_hash.update(uuid.uuid4().bytes) id_bytes = id_hash.digest() assert len(id_bytes) == ray_constants.ID_SIZE return id_bytes def format_error_message(exception_message, task_exception=False): """Improve the formatting of an exception thrown by a remote function. This method takes a traceback from an exception and makes it nicer by removing a few uninformative lines and adding some space to indent the remaining lines nicely. Args: exception_message (str): A message generated by traceback.format_exc(). Returns: A string of the formatted exception message. """ lines = exception_message.split("\n") if task_exception: # For errors that occur inside of tasks, remove lines 1 and 2 which are # always the same, they just contain information about the worker code. lines = lines[0:1] + lines[3:] pass return "\n".join(lines) def push_error_to_driver(worker, error_type, message, job_id=None): """Push an error message to the driver to be printed in the background. Args: worker: The worker to use. error_type (str): The type of the error. message (str): The message that will be printed in the background on the driver. job_id: The ID of the driver to push the error message to. If this is None, then the message will be pushed to all drivers. """ if job_id is None: job_id = ray.JobID.nil() assert isinstance(job_id, ray.JobID) worker.raylet_client.push_error(job_id, error_type, message, time.time()) def push_error_to_driver_through_redis(redis_client, error_type, message, job_id=None): """Push an error message to the driver to be printed in the background. Normally the push_error_to_driver function should be used. However, in some instances, the raylet client is not available, e.g., because the error happens in Python before the driver or worker has connected to the backend processes. Args: redis_client: The redis client to use. error_type (str): The type of the error. message (str): The message that will be printed in the background on the driver. job_id: The ID of the driver to push the error message to. If this is None, then the message will be pushed to all drivers. """ if job_id is None: job_id = ray.JobID.nil() assert isinstance(job_id, ray.JobID) # Do everything in Python and through the Python Redis client instead # of through the raylet. error_data = ray.gcs_utils.construct_error_message(job_id, error_type, message, time.time()) redis_client.execute_command( "RAY.TABLE_APPEND", ray.gcs_utils.TablePrefix.Value("ERROR_INFO"), ray.gcs_utils.TablePubsub.Value("ERROR_INFO_PUBSUB"), job_id.binary(), error_data) def is_cython(obj): """Check if an object is a Cython function or method""" # TODO(suo): We could split these into two functions, one for Cython # functions and another for Cython methods. # TODO(suo): There doesn't appear to be a Cython function 'type' we can # check against via isinstance. Please correct me if I'm wrong. def check_cython(x): return type(x).__name__ == "cython_function_or_method" # Check if function or method, respectively return check_cython(obj) or \ (hasattr(obj, "__func__") and check_cython(obj.__func__)) def is_function_or_method(obj): """Check if an object is a function or method. Args: obj: The Python object in question. Returns: True if the object is an function or method. """ return inspect.isfunction(obj) or inspect.ismethod(obj) or is_cython(obj) def is_class_method(f): """Returns whether the given method is a class_method.""" return hasattr(f, "__self__") and f.__self__ is not None def is_static_method(cls, f_name): """Returns whether the class has a static method with the given name. Args: cls: The Python class (i.e. object of type `type`) to search for the method in. f_name: The name of the method to look up in this class and check whether or not it is static. """ for cls in inspect.getmro(cls): if f_name in cls.__dict__: return isinstance(cls.__dict__[f_name], staticmethod) return False def random_string(): """Generate a random string to use as an ID. Note that users may seed numpy, which could cause this function to generate duplicate IDs. Therefore, we need to seed numpy ourselves, but we can't interfere with the state of the user's random number generator, so we extract the state of the random number generator and reset it after we are done. TODO(rkn): If we want to later guarantee that these are generated in a deterministic manner, then we will need to make some changes here. Returns: A random byte string of length ray_constants.ID_SIZE. """ # Get the state of the numpy random number generator. numpy_state = np.random.get_state() # Try to use true randomness. np.random.seed(None) # Generate the random ID. random_id = np.random.bytes(ray_constants.ID_SIZE) # Reset the state of the numpy random number generator. np.random.set_state(numpy_state) return random_id def decode(byte_str, allow_none=False): """Make this unicode in Python 3, otherwise leave it as bytes. Args: byte_str: The byte string to decode. allow_none: If true, then we will allow byte_str to be None in which case we will return an empty string. TODO(rkn): Remove this flag. This is only here to simplify upgrading to flatbuffers 1.10.0. Returns: A byte string in Python 2 and a unicode string in Python 3. """ if byte_str is None and allow_none: return "" if not isinstance(byte_str, bytes): raise ValueError( "The argument {} must be a bytes object.".format(byte_str)) if sys.version_info >= (3, 0): return byte_str.decode("ascii") else: return byte_str def ensure_str(s, encoding="utf-8", errors="strict"): """Coerce *s* to `str`. To keep six with lower version, see Issue 4169, we copy this function from six == 1.12.0. TODO(yuhguo): remove this function when six >= 1.12.0. For Python 2: - `unicode` -> encoded to `str` - `str` -> `str` For Python 3: - `str` -> `str` - `bytes` -> decoded to `str` """ if six.PY3: text_type = str binary_type = bytes else: text_type = unicode # noqa: F821 binary_type = str if not isinstance(s, (text_type, binary_type)): raise TypeError("not expecting type '%s'" % type(s)) if six.PY2 and isinstance(s, text_type): s = s.encode(encoding, errors) elif six.PY3 and isinstance(s, binary_type): s = s.decode(encoding, errors) return s def binary_to_object_id(binary_object_id): return ray.ObjectID(binary_object_id) def binary_to_task_id(binary_task_id): return ray.TaskID(binary_task_id) def binary_to_hex(identifier): hex_identifier = binascii.hexlify(identifier) if sys.version_info >= (3, 0): hex_identifier = hex_identifier.decode() return hex_identifier def hex_to_binary(hex_identifier): return binascii.unhexlify(hex_identifier) # TODO(qwang): Remove these hepler functions # once we separate `WorkerID` from `UniqueID`. def compute_job_id_from_driver(driver_id): assert isinstance(driver_id, ray.WorkerID) return ray.JobID(driver_id.binary()[0:ray.JobID.size()]) def compute_driver_id_from_job(job_id): assert isinstance(job_id, ray.JobID) rest_length = ray_constants.ID_SIZE - job_id.size() driver_id_str = job_id.binary() + (rest_length * b"\xff") return ray.WorkerID(driver_id_str) def get_cuda_visible_devices(): """Get the device IDs in the CUDA_VISIBLE_DEVICES environment variable. Returns: if CUDA_VISIBLE_DEVICES is set, this returns a list of integers with the IDs of the GPUs. If it is not set, this returns None. """ gpu_ids_str = os.environ.get("CUDA_VISIBLE_DEVICES", None) if gpu_ids_str is None: return None if gpu_ids_str == "": return [] return [int(i) for i in gpu_ids_str.split(",")] last_set_gpu_ids = None def set_cuda_visible_devices(gpu_ids): """Set the CUDA_VISIBLE_DEVICES environment variable. Args: gpu_ids: This is a list of integers representing GPU IDs. """ global last_set_gpu_ids if last_set_gpu_ids == gpu_ids: return # optimization: already set os.environ["CUDA_VISIBLE_DEVICES"] = ",".join([str(i) for i in gpu_ids]) last_set_gpu_ids = gpu_ids def resources_from_resource_arguments( default_num_cpus, default_num_gpus, default_memory, default_object_store_memory, default_resources, runtime_num_cpus, runtime_num_gpus, runtime_memory, runtime_object_store_memory, runtime_resources): """Determine a task's resource requirements. Args: default_num_cpus: The default number of CPUs required by this function or actor method. default_num_gpus: The default number of GPUs required by this function or actor method. default_memory: The default heap memory required by this function or actor method. default_object_store_memory: The default object store memory required by this function or actor method. default_resources: The default custom resources required by this function or actor method. runtime_num_cpus: The number of CPUs requested when the task was invoked. runtime_num_gpus: The number of GPUs requested when the task was invoked. runtime_memory: The heap memory requested when the task was invoked. runtime_object_store_memory: The object store memory requested when the task was invoked. runtime_resources: The custom resources requested when the task was invoked. Returns: A dictionary of the resource requirements for the task. """ if runtime_resources is not None: resources = runtime_resources.copy() elif default_resources is not None: resources = default_resources.copy() else: resources = {} if "CPU" in resources or "GPU" in resources: raise ValueError("The resources dictionary must not " "contain the key 'CPU' or 'GPU'") elif "memory" in resources or "object_store_memory" in resources: raise ValueError("The resources dictionary must not " "contain the key 'memory' or 'object_store_memory'") assert default_num_cpus is not None resources["CPU"] = (default_num_cpus if runtime_num_cpus is None else runtime_num_cpus) if runtime_num_gpus is not None: resources["GPU"] = runtime_num_gpus elif default_num_gpus is not None: resources["GPU"] = default_num_gpus memory = default_memory or runtime_memory object_store_memory = (default_object_store_memory or runtime_object_store_memory) if memory is not None: resources["memory"] = ray_constants.to_memory_units( memory, round_up=True) if object_store_memory is not None: resources["object_store_memory"] = ray_constants.to_memory_units( object_store_memory, round_up=True) return resources _default_handler = None def setup_logger(logging_level, logging_format): """Setup default logging for ray.""" logger = logging.getLogger("ray") if type(logging_level) is str: logging_level = logging.getLevelName(logging_level.upper()) logger.setLevel(logging_level) global _default_handler if _default_handler is None: _default_handler = logging.StreamHandler() logger.addHandler(_default_handler) _default_handler.setFormatter(logging.Formatter(logging_format)) logger.propagate = False # This function is copied and modified from # https://github.com/giampaolo/psutil/blob/5bd44f8afcecbfb0db479ce230c790fc2c56569a/psutil/tests/test_linux.py#L132-L138 # noqa: E501 def vmstat(stat): """Run vmstat and get a particular statistic. Args: stat: The statistic that we are interested in retrieving. Returns: The parsed output. """ out = subprocess.check_output(["vmstat", "-s"]) stat = stat.encode("ascii") for line in out.split(b"\n"): line = line.strip() if stat in line: return int(line.split(b" ")[0]) raise ValueError("Can't find {} in 'vmstat' output.".format(stat)) # This function is copied and modified from # https://github.com/giampaolo/psutil/blob/5e90b0a7f3fccb177445a186cc4fac62cfadb510/psutil/tests/test_osx.py#L29-L38 # noqa: E501 def sysctl(command): """Run a sysctl command and parse the output. Args: command: A sysctl command with an argument, for example, ["sysctl", "hw.memsize"]. Returns: The parsed output. """ out = subprocess.check_output(command) result = out.split(b" ")[1] try: return int(result) except ValueError: return result def get_system_memory(): """Return the total amount of system memory in bytes. Returns: The total amount of system memory in bytes. """ # Try to accurately figure out the memory limit if we are in a docker # container. Note that this file is not specific to Docker and its value is # often much larger than the actual amount of memory. docker_limit = None memory_limit_filename = "/sys/fs/cgroup/memory/memory.limit_in_bytes" if os.path.exists(memory_limit_filename): with open(memory_limit_filename, "r") as f: docker_limit = int(f.read()) # Use psutil if it is available. psutil_memory_in_bytes = None try: import psutil psutil_memory_in_bytes = psutil.virtual_memory().total except ImportError: pass if psutil_memory_in_bytes is not None: memory_in_bytes = psutil_memory_in_bytes elif sys.platform == "linux" or sys.platform == "linux2": # Handle Linux. bytes_in_kilobyte = 1024 memory_in_bytes = vmstat("total memory") * bytes_in_kilobyte else: # Handle MacOS. memory_in_bytes = sysctl(["sysctl", "hw.memsize"]) if docker_limit is not None: return min(docker_limit, memory_in_bytes) else: return memory_in_bytes def estimate_available_memory(): """Return the currently available amount of system memory in bytes. Returns: The total amount of available memory in bytes. It may be an overestimate if psutil is not installed. """ # Use psutil if it is available. try: import psutil return psutil.virtual_memory().available except ImportError: pass # Handle Linux. if sys.platform == "linux" or sys.platform == "linux2": bytes_in_kilobyte = 1024 return ( vmstat("total memory") - vmstat("used memory")) * bytes_in_kilobyte # Give up return get_system_memory() def get_shared_memory_bytes(): """Get the size of the shared memory file system. Returns: The size of the shared memory file system in bytes. """ # Make sure this is only called on Linux. assert sys.platform == "linux" or sys.platform == "linux2" shm_fd = os.open("/dev/shm", os.O_RDONLY) try: shm_fs_stats = os.fstatvfs(shm_fd) # The value shm_fs_stats.f_bsize is the block size and the # value shm_fs_stats.f_bavail is the number of available # blocks. shm_avail = shm_fs_stats.f_bsize * shm_fs_stats.f_bavail finally: os.close(shm_fd) return shm_avail def check_oversized_pickle(pickled, name, obj_type, worker): """Send a warning message if the pickled object is too large. Args: pickled: the pickled object. name: name of the pickled object. obj_type: type of the pickled object, can be 'function', 'remote function', 'actor', or 'object'. worker: the worker used to send warning message. """ length = len(pickled) if length <= ray_constants.PICKLE_OBJECT_WARNING_SIZE: return warning_message = ( "Warning: The {} {} has size {} when pickled. " "It will be stored in Redis, which could cause memory issues. " "This may mean that its definition uses a large array or other object." ).format(obj_type, name, length) push_error_to_driver( worker, ray_constants.PICKLING_LARGE_OBJECT_PUSH_ERROR, warning_message, job_id=worker.current_job_id) def is_main_thread(): return threading.current_thread().getName() == "MainThread" def try_make_directory_shared(directory_path): try: os.chmod(directory_path, 0o0777) except OSError as e: # Silently suppress the PermissionError that is thrown by the chmod. # This is done because the user attempting to change the permissions # on a directory may not own it. The chmod is attempted whether the # directory is new or not to avoid race conditions. # ray-project/ray/#3591 if e.errno in [errno.EACCES, errno.EPERM]: pass else: raise def try_to_create_directory(directory_path): """Attempt to create a directory that is globally readable/writable. Args: directory_path: The path of the directory to create. """ directory_path = os.path.expanduser(directory_path) os.makedirs(directory_path, exist_ok=True) # Change the log directory permissions so others can use it. This is # important when multiple people are using the same machine. try_make_directory_shared(directory_path) def try_to_symlink(symlink_path, target_path): """Attempt to create a symlink. If the symlink path exists and isn't a symlink, the symlink will not be created. If a symlink exists in the path, it will be attempted to be removed and replaced. Args: symlink_path: The path at which to create the symlink. target_path: The path the symlink should point to. """ symlink_path = os.path.expanduser(symlink_path) target_path = os.path.expanduser(target_path) if os.path.exists(symlink_path): if os.path.islink(symlink_path): # Try to remove existing symlink. try: os.remove(symlink_path) except OSError: return else: # There's an existing non-symlink file, don't overwrite it. return try: os.symlink(target_path, symlink_path) except OSError: return
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/worker.py
Python
from contextlib import contextmanager import colorama import atexit import faulthandler import hashlib import inspect import io import json import logging import os import redis import signal from six.moves import queue import sys import threading import time import traceback import random # Ray modules import ray.cloudpickle as pickle import ray.gcs_utils import ray.memory_monitor as memory_monitor import ray.node import ray.parameter import ray.ray_constants as ray_constants import ray.remote_function import ray.serialization as serialization import ray.services as services import ray.signature import ray.state from ray import ( ActorID, JobID, ObjectID, ) from ray import import_thread from ray import profiling from ray.exceptions import ( RayError, RayTaskError, ObjectStoreFullError, ) from ray.function_manager import FunctionActorManager from ray.utils import ( _random_string, check_oversized_pickle, is_cython, setup_logger, ) from ray.local_mode_manager import LocalModeManager SCRIPT_MODE = 0 WORKER_MODE = 1 LOCAL_MODE = 2 ERROR_KEY_PREFIX = b"Error:" # Logger for this module. It should be configured at the entry point # into the program using Ray. Ray provides a default configuration at # entry/init points. logger = logging.getLogger(__name__) try: import setproctitle except ImportError: setproctitle = None class ActorCheckpointInfo: """Information used to maintain actor checkpoints.""" __slots__ = [ # Number of tasks executed since last checkpoint. "num_tasks_since_last_checkpoint", # Timestamp of the last checkpoint, in milliseconds. "last_checkpoint_timestamp", # IDs of the previous checkpoints. "checkpoint_ids", ] def __init__(self, num_tasks_since_last_checkpoint, last_checkpoint_timestamp, checkpoint_ids): self.num_tasks_since_last_checkpoint = num_tasks_since_last_checkpoint self.last_checkpoint_timestamp = last_checkpoint_timestamp self.checkpoint_ids = checkpoint_ids class Worker: """A class used to define the control flow of a worker process. Note: The methods in this class are considered unexposed to the user. The functions outside of this class are considered exposed. Attributes: connected (bool): True if Ray has been started and False otherwise. node (ray.node.Node): The node this worker is attached to. mode: The mode of the worker. One of SCRIPT_MODE, LOCAL_MODE, and WORKER_MODE. cached_functions_to_run (List): A list of functions to run on all of the workers that should be exported as soon as connect is called. """ def __init__(self): """Initialize a Worker object.""" self.node = None self.mode = None self.cached_functions_to_run = [] self.actor_init_error = None self.make_actor = None self.actors = {} # Information used to maintain actor checkpoints. self.actor_checkpoint_info = {} self.actor_task_counter = 0 # When the worker is constructed. Record the original value of the # CUDA_VISIBLE_DEVICES environment variable. self.original_gpu_ids = ray.utils.get_cuda_visible_devices() self.memory_monitor = memory_monitor.MemoryMonitor() # A dictionary that maps from driver id to SerializationContext # TODO: clean up the SerializationContext once the job finished. self.serialization_context_map = {} self.function_actor_manager = FunctionActorManager(self) # This event is checked regularly by all of the threads so that they # know when to exit. self.threads_stopped = threading.Event() # Index of the current session. This number will # increment every time when `ray.shutdown` is called. self._session_index = 0 # Functions to run to process the values returned by ray.get. Each # postprocessor must take two arguments ("object_ids", and "values"). self._post_get_hooks = [] @property def connected(self): return self.node is not None @property def node_ip_address(self): self.check_connected() return self.node.node_ip_address @property def load_code_from_local(self): self.check_connected() return self.node.load_code_from_local @property def use_pickle(self): self.check_connected() return self.node.use_pickle @property def current_job_id(self): if hasattr(self, "core_worker"): return self.core_worker.get_current_job_id() return JobID.nil() @property def actor_id(self): if hasattr(self, "core_worker"): return self.core_worker.get_actor_id() return ActorID.nil() @property def current_task_id(self): return self.core_worker.get_current_task_id() @property def current_session_and_job(self): """Get the current session index and job id as pair.""" assert isinstance(self._session_index, int) assert isinstance(self.current_job_id, ray.JobID) return self._session_index, self.current_job_id def mark_actor_init_failed(self, error): """Called to mark this actor as failed during initialization.""" self.actor_init_error = error def reraise_actor_init_error(self): """Raises any previous actor initialization error.""" if self.actor_init_error is not None: raise self.actor_init_error def get_serialization_context(self, job_id=None): """Get the SerializationContext of the job that this worker is processing. Args: job_id: The ID of the job that indicates which job to get the serialization context for. Returns: The serialization context of the given job. """ # This function needs to be protected by a lock, because it will be # called by`register_class_for_serialization`, as well as the import # thread, from different threads. Also, this function will recursively # call itself, so we use RLock here. if job_id is None: job_id = self.current_job_id with self.lock: if job_id not in self.serialization_context_map: self.serialization_context_map[ job_id] = serialization.SerializationContext(self) self.serialization_context_map[job_id].initialize() return self.serialization_context_map[job_id] def check_connected(self): """Check if the worker is connected. Raises: Exception: An exception is raised if the worker is not connected. """ if not self.connected: raise RayConnectionError("Ray has not been started yet. You can " "start Ray with 'ray.init()'.") def set_mode(self, mode): """Set the mode of the worker. The mode SCRIPT_MODE should be used if this Worker is a driver that is being run as a Python script or interactively in a shell. It will print information about task failures. The mode WORKER_MODE should be used if this Worker is not a driver. It will not print information about tasks. The mode LOCAL_MODE should be used if this Worker is a driver and if you want to run the driver in a manner equivalent to serial Python for debugging purposes. It will not send remote function calls to the scheduler and will instead execute them in a blocking fashion. Args: mode: One of SCRIPT_MODE, WORKER_MODE, and LOCAL_MODE. """ self.mode = mode def put_object(self, value, object_id=None, pin_object=True): """Put value in the local object store with object id `objectid`. This assumes that the value for `objectid` has not yet been placed in the local object store. If the plasma store is full, the worker will automatically retry up to DEFAULT_PUT_OBJECT_RETRIES times. Each retry will delay for an exponentially doubling amount of time, starting with DEFAULT_PUT_OBJECT_DELAY. After this, exception will be raised. Args: value: The value to put in the object store. object_id (object_id.ObjectID): The object ID of the value to be put. If None, one will be generated. pin_object: If set, the object will be pinned at the raylet. Returns: object_id.ObjectID: The object ID the object was put under. Raises: ray.exceptions.ObjectStoreFullError: This is raised if the attempt to store the object fails because the object store is full even after multiple retries. """ # Make sure that the value is not an object ID. if isinstance(value, ObjectID): raise TypeError( "Calling 'put' on an ray.ObjectID is not allowed " "(similarly, returning an ray.ObjectID from a remote " "function is not allowed). If you really want to " "do this, you can wrap the ray.ObjectID in a list and " "call 'put' on it (or return it).") serialized_value = self.get_serialization_context().serialize(value) return self.core_worker.put_serialized_object( serialized_value, object_id=object_id, pin_object=pin_object) def deserialize_objects(self, data_metadata_pairs, object_ids, error_timeout=10): context = self.get_serialization_context() return context.deserialize_objects(data_metadata_pairs, object_ids, error_timeout) def get_objects(self, object_ids, timeout=None): """Get the values in the object store associated with the IDs. Return the values from the local object store for object_ids. This will block until all the values for object_ids have been written to the local object store. Args: object_ids (List[object_id.ObjectID]): A list of the object IDs whose values should be retrieved. timeout (float): timeout (float): The maximum amount of time in seconds to wait before returning. Raises: Exception if running in LOCAL_MODE and any of the object IDs do not exist in the emulated object store. """ # Make sure that the values are object IDs. for object_id in object_ids: if not isinstance(object_id, ObjectID): raise TypeError( "Attempting to call `get` on the value {}, " "which is not an ray.ObjectID.".format(object_id)) if self.mode == LOCAL_MODE: return self.local_mode_manager.get_objects(object_ids) timeout_ms = int(timeout * 1000) if timeout else -1 data_metadata_pairs = self.core_worker.get_objects( object_ids, self.current_task_id, timeout_ms) return self.deserialize_objects(data_metadata_pairs, object_ids) def run_function_on_all_workers(self, function, run_on_other_drivers=False): """Run arbitrary code on all of the workers. This function will first be run on the driver, and then it will be exported to all of the workers to be run. It will also be run on any new workers that register later. If ray.init has not been called yet, then cache the function and export it later. Args: function (Callable): The function to run on all of the workers. It takes only one argument, a worker info dict. If it returns anything, its return values will not be used. run_on_other_drivers: The boolean that indicates whether we want to run this function on other drivers. One case is we may need to share objects across drivers. """ # If ray.init has not been called yet, then cache the function and # export it when connect is called. Otherwise, run the function on all # workers. if self.mode is None: self.cached_functions_to_run.append(function) else: # Attempt to pickle the function before we need it. This could # fail, and it is more convenient if the failure happens before we # actually run the function locally. pickled_function = pickle.dumps(function) function_to_run_id = hashlib.sha1(pickled_function).digest() key = b"FunctionsToRun:" + function_to_run_id # First run the function on the driver. # We always run the task locally. function({"worker": self}) # Check if the function has already been put into redis. function_exported = self.redis_client.setnx(b"Lock:" + key, 1) if not function_exported: # In this case, the function has already been exported, so # we don't need to export it again. return check_oversized_pickle(pickled_function, function.__name__, "function", self) # Run the function on all workers. self.redis_client.hmset( key, { "job_id": self.current_job_id.binary(), "function_id": function_to_run_id, "function": pickled_function, "run_on_other_drivers": str(run_on_other_drivers) }) self.redis_client.rpush("Exports", key) # TODO(rkn): If the worker fails after it calls setnx and before it # successfully completes the hmset and rpush, then the program will # most likely hang. This could be fixed by making these three # operations into a transaction (or by implementing a custom # command that does all three things). def _get_arguments_for_execution(self, function_name, serialized_args): """Retrieve the arguments for the remote function. This retrieves the values for the arguments to the remote function that were passed in as object IDs. Arguments that were passed by value are not changed. This is called by the worker that is executing the remote function. Args: function_name (str): The name of the remote function whose arguments are being retrieved. serialized_args (List): The arguments to the function. These are either strings representing serialized objects passed by value or they are ray.ObjectIDs. Returns: The retrieved arguments in addition to the arguments that were passed by value. Raises: RayError: This exception is raised if a task that created one of the arguments failed. """ arguments = [None] * len(serialized_args) object_ids = [] object_indices = [] for (i, arg) in enumerate(serialized_args): if isinstance(arg, ObjectID): object_ids.append(arg) object_indices.append(i) else: # pass the argument by value arguments[i] = arg # Get the objects from the local object store. if len(object_ids) > 0: values = self.get_objects(object_ids) for i, value in enumerate(values): if isinstance(value, RayError): raise value else: arguments[object_indices[i]] = value return ray.signature.recover_args(arguments) def main_loop(self): """The main loop a worker runs to receive and execute tasks.""" def sigterm_handler(signum, frame): shutdown(True) sys.exit(1) signal.signal(signal.SIGTERM, sigterm_handler) self.core_worker.run_task_loop() sys.exit(0) def get_gpu_ids(): """Get the IDs of the GPUs that are available to the worker. If the CUDA_VISIBLE_DEVICES environment variable was set when the worker started up, then the IDs returned by this method will be a subset of the IDs in CUDA_VISIBLE_DEVICES. If not, the IDs will fall in the range [0, NUM_GPUS - 1], where NUM_GPUS is the number of GPUs that the node has. Returns: A list of GPU IDs. """ if _mode() == LOCAL_MODE: raise Exception("ray.get_gpu_ids() currently does not work in LOCAL " "MODE.") all_resource_ids = global_worker.core_worker.resource_ids() assigned_ids = [ resource_id for resource_id, _ in all_resource_ids.get("GPU", []) ] # If the user had already set CUDA_VISIBLE_DEVICES, then respect that (in # the sense that only GPU IDs that appear in CUDA_VISIBLE_DEVICES should be # returned). if global_worker.original_gpu_ids is not None: assigned_ids = [ global_worker.original_gpu_ids[gpu_id] for gpu_id in assigned_ids ] return assigned_ids def get_resource_ids(): """Get the IDs of the resources that are available to the worker. Returns: A dictionary mapping the name of a resource to a list of pairs, where each pair consists of the ID of a resource and the fraction of that resource reserved for this worker. """ if _mode() == LOCAL_MODE: raise Exception( "ray.get_resource_ids() currently does not work in LOCAL " "MODE.") return global_worker.core_worker.resource_ids() def get_webui_url(): """Get the URL to access the web UI. Note that the URL does not specify which node the web UI is on. Returns: The URL of the web UI as a string. """ if _global_node is None: raise Exception("Ray has not been initialized/connected.") return _global_node.webui_url global_worker = Worker() """Worker: The global Worker object for this worker process. We use a global Worker object to ensure that there is a single worker object per worker process. """ _global_node = None """ray.node.Node: The global node object that is created by ray.init().""" class RayConnectionError(Exception): pass def print_failed_task(task_status): """Print information about failed tasks. Args: task_status (Dict): A dictionary containing the name, operationid, and error message for a failed task. """ logger.error(""" Error: Task failed Function Name: {} Task ID: {} Error Message: \n{} """.format(task_status["function_name"], task_status["operationid"], task_status["error_message"])) def init(address=None, redis_address=None, num_cpus=None, num_gpus=None, memory=None, object_store_memory=None, resources=None, driver_object_store_memory=None, redis_max_memory=None, log_to_driver=True, node_ip_address=ray_constants.NODE_DEFAULT_IP, object_id_seed=None, local_mode=False, redirect_worker_output=None, redirect_output=None, ignore_reinit_error=False, num_redis_shards=None, redis_max_clients=None, redis_password=ray_constants.REDIS_DEFAULT_PASSWORD, plasma_directory=None, huge_pages=False, include_webui=None, webui_host="localhost", job_id=None, configure_logging=True, logging_level=logging.INFO, logging_format=ray_constants.LOGGER_FORMAT, plasma_store_socket_name=None, raylet_socket_name=None, temp_dir=None, load_code_from_local=False, use_pickle=ray.cloudpickle.FAST_CLOUDPICKLE_USED, _internal_config=None): """Connect to an existing Ray cluster or start one and connect to it. This method handles two cases. Either a Ray cluster already exists and we just attach this driver to it, or we start all of the processes associated with a Ray cluster and attach to the newly started cluster. To start Ray and all of the relevant processes, use this as follows: .. code-block:: python ray.init() To connect to an existing Ray cluster, use this as follows (substituting in the appropriate address): .. code-block:: python ray.init(address="123.45.67.89:6379") Args: address (str): The address of the Ray cluster to connect to. If this address is not provided, then this command will start Redis, a raylet, a plasma store, a plasma manager, and some workers. It will also kill these processes when Python exits. redis_address (str): Deprecated; same as address. num_cpus (int): Number of cpus the user wishes all raylets to be configured with. num_gpus (int): Number of gpus the user wishes all raylets to be configured with. resources: A dictionary mapping the name of a resource to the quantity of that resource available. memory: The amount of memory (in bytes) that is available for use by workers requesting memory resources. By default, this is autoset based on available system memory. object_store_memory: The amount of memory (in bytes) to start the object store with. By default, this is autoset based on available system memory, subject to a 20GB cap. redis_max_memory: The max amount of memory (in bytes) to allow each redis shard to use. Once the limit is exceeded, redis will start LRU eviction of entries. This only applies to the sharded redis tables (task, object, and profile tables). By default, this is autoset based on available system memory, subject to a 10GB cap. log_to_driver (bool): If true, then output from all of the worker processes on all nodes will be directed to the driver. node_ip_address (str): The IP address of the node that we are on. object_id_seed (int): Used to seed the deterministic generation of object IDs. The same value can be used across multiple runs of the same driver in order to generate the object IDs in a consistent manner. However, the same ID should not be used for different drivers. local_mode (bool): True if the code should be executed serially without Ray. This is useful for debugging. driver_object_store_memory (int): Limit the amount of memory the driver can use in the object store for creating objects. By default, this is autoset based on available system memory, subject to a 20GB cap. ignore_reinit_error: True if we should suppress errors from calling ray.init() a second time. num_redis_shards: The number of Redis shards to start in addition to the primary Redis shard. redis_max_clients: If provided, attempt to configure Redis with this maxclients number. redis_password (str): Prevents external clients without the password from connecting to Redis if provided. plasma_directory: A directory where the Plasma memory mapped files will be created. huge_pages: Boolean flag indicating whether to start the Object Store with hugetlbfs support. Requires plasma_directory. include_webui: Boolean flag indicating whether to start the web UI, which displays the status of the Ray cluster. If this argument is None, then the UI will be started if the relevant dependencies are present. webui_host: The host to bind the web UI server to. Can either be localhost (127.0.0.1) or 0.0.0.0 (available from all interfaces). By default, this is set to localhost to prevent access from external machines. job_id: The ID of this job. configure_logging: True if allow the logging cofiguration here. Otherwise, the users may want to configure it by their own. logging_level: Logging level, default will be logging.INFO. logging_format: Logging format, default contains a timestamp, filename, line number, and message. See ray_constants.py. plasma_store_socket_name (str): If provided, it will specify the socket name used by the plasma store. raylet_socket_name (str): If provided, it will specify the socket path used by the raylet process. temp_dir (str): If provided, it will specify the root temporary directory for the Ray process. load_code_from_local: Whether code should be loaded from a local module or from the GCS. use_pickle: Whether data objects should be serialized with cloudpickle. _internal_config (str): JSON configuration for overriding RayConfig defaults. For testing purposes ONLY. Returns: Address information about the started processes. Raises: Exception: An exception is raised if an inappropriate combination of arguments is passed in. """ if redis_address is not None: raise DeprecationWarning("The redis_address argument is deprecated. " "Please use address instead.") if redis_address is not None or address is not None: redis_address, _, _ = services.validate_redis_address( address, redis_address) if configure_logging: setup_logger(logging_level, logging_format) if local_mode: driver_mode = LOCAL_MODE else: driver_mode = SCRIPT_MODE if "OMP_NUM_THREADS" in os.environ: logger.warning("OMP_NUM_THREADS={} is set, this may impact " "object transfer performance.".format( os.environ["OMP_NUM_THREADS"])) if setproctitle is None: logger.warning( "WARNING: Not updating worker name since `setproctitle` is not " "installed. Install this with `pip install setproctitle` " "(or ray[debug]) to enable monitoring of worker processes.") if global_worker.connected: if ignore_reinit_error: logger.error("Calling ray.init() again after it has already been " "called.") return else: raise Exception("Perhaps you called ray.init twice by accident? " "This error can be suppressed by passing in " "'ignore_reinit_error=True' or by calling " "'ray.shutdown()' prior to 'ray.init()'.") # Convert hostnames to numerical IP address. if node_ip_address is not None: node_ip_address = services.address_to_ip(node_ip_address) global _global_node if driver_mode == LOCAL_MODE: # If starting Ray in LOCAL_MODE, don't start any other processes. _global_node = ray.node.LocalNode() elif redis_address is None: # In this case, we need to start a new cluster. ray_params = ray.parameter.RayParams( redis_address=redis_address, node_ip_address=node_ip_address, object_id_seed=object_id_seed, local_mode=local_mode, driver_mode=driver_mode, redirect_worker_output=redirect_worker_output, redirect_output=redirect_output, num_cpus=num_cpus, num_gpus=num_gpus, resources=resources, num_redis_shards=num_redis_shards, redis_max_clients=redis_max_clients, redis_password=redis_password, plasma_directory=plasma_directory, huge_pages=huge_pages, include_webui=include_webui, webui_host=webui_host, memory=memory, object_store_memory=object_store_memory, redis_max_memory=redis_max_memory, plasma_store_socket_name=plasma_store_socket_name, raylet_socket_name=raylet_socket_name, temp_dir=temp_dir, load_code_from_local=load_code_from_local, use_pickle=use_pickle, _internal_config=_internal_config, ) # Start the Ray processes. We set shutdown_at_exit=False because we # shutdown the node in the ray.shutdown call that happens in the atexit # handler. We still spawn a reaper process in case the atexit handler # isn't called. _global_node = ray.node.Node( head=True, shutdown_at_exit=False, spawn_reaper=True, ray_params=ray_params) else: # In this case, we are connecting to an existing cluster. if num_cpus is not None or num_gpus is not None: raise Exception("When connecting to an existing cluster, num_cpus " "and num_gpus must not be provided.") if resources is not None: raise Exception("When connecting to an existing cluster, " "resources must not be provided.") if num_redis_shards is not None: raise Exception("When connecting to an existing cluster, " "num_redis_shards must not be provided.") if redis_max_clients is not None: raise Exception("When connecting to an existing cluster, " "redis_max_clients must not be provided.") if memory is not None: raise Exception("When connecting to an existing cluster, " "memory must not be provided.") if object_store_memory is not None: raise Exception("When connecting to an existing cluster, " "object_store_memory must not be provided.") if redis_max_memory is not None: raise Exception("When connecting to an existing cluster, " "redis_max_memory must not be provided.") if plasma_directory is not None: raise Exception("When connecting to an existing cluster, " "plasma_directory must not be provided.") if huge_pages: raise Exception("When connecting to an existing cluster, " "huge_pages must not be provided.") if temp_dir is not None: raise Exception("When connecting to an existing cluster, " "temp_dir must not be provided.") if plasma_store_socket_name is not None: raise Exception("When connecting to an existing cluster, " "plasma_store_socket_name must not be provided.") if raylet_socket_name is not None: raise Exception("When connecting to an existing cluster, " "raylet_socket_name must not be provided.") if _internal_config is not None: raise Exception("When connecting to an existing cluster, " "_internal_config must not be provided.") # In this case, we only need to connect the node. ray_params = ray.parameter.RayParams( node_ip_address=node_ip_address, redis_address=redis_address, redis_password=redis_password, object_id_seed=object_id_seed, temp_dir=temp_dir, load_code_from_local=load_code_from_local, use_pickle=use_pickle) _global_node = ray.node.Node( ray_params, head=False, shutdown_at_exit=False, spawn_reaper=False, connect_only=True) connect( _global_node, mode=driver_mode, log_to_driver=log_to_driver, worker=global_worker, driver_object_store_memory=driver_object_store_memory, job_id=job_id, internal_config=json.loads(_internal_config) if _internal_config else {}) for hook in _post_init_hooks: hook() return _global_node.address_info # Functions to run as callback after a successful ray init. _post_init_hooks = [] def shutdown(exiting_interpreter=False): """Disconnect the worker, and terminate processes started by ray.init(). This will automatically run at the end when a Python process that uses Ray exits. It is ok to run this twice in a row. The primary use case for this function is to cleanup state between tests. Note that this will clear any remote function definitions, actor definitions, and existing actors, so if you wish to use any previously defined remote functions or actors after calling ray.shutdown(), then you need to redefine them. If they were defined in an imported module, then you will need to reload the module. Args: exiting_interpreter (bool): True if this is called by the atexit hook and false otherwise. If we are exiting the interpreter, we will wait a little while to print any extra error messages. """ if exiting_interpreter and global_worker.mode == SCRIPT_MODE: # This is a duration to sleep before shutting down everything in order # to make sure that log messages finish printing. time.sleep(0.5) disconnect(exiting_interpreter) # We need to destruct the core worker here because after this function, # we will tear down any processes spawned by ray.init() and the background # IO thread in the core worker doesn't currently handle that gracefully. if hasattr(global_worker, "core_worker"): del global_worker.core_worker # Disconnect global state from GCS. ray.state.state.disconnect() # Shut down the Ray processes. global _global_node if _global_node is not None: _global_node.kill_all_processes(check_alive=False, allow_graceful=True) _global_node = None # TODO(rkn): Instead of manually resetting some of the worker fields, we # should simply set "global_worker" to equal "None" or something like that. global_worker.set_mode(None) global_worker._post_get_hooks = [] atexit.register(shutdown, True) def sigterm_handler(signum, frame): sys.exit(signal.SIGTERM) try: signal.signal(signal.SIGTERM, sigterm_handler) except ValueError: logger.warning("Failed to set SIGTERM handler, processes might" "not be cleaned up properly on exit.") # Define a custom excepthook so that if the driver exits with an exception, we # can push that exception to Redis. normal_excepthook = sys.excepthook def custom_excepthook(type, value, tb): # If this is a driver, push the exception to redis. if global_worker.mode == SCRIPT_MODE: error_message = "".join(traceback.format_tb(tb)) try: global_worker.redis_client.hmset( b"Drivers:" + global_worker.worker_id, {"exception": error_message}) except (ConnectionRefusedError, redis.exceptions.ConnectionError): logger.warning("Could not push exception to redis.") # Call the normal excepthook. normal_excepthook(type, value, tb) sys.excepthook = custom_excepthook # The last time we raised a TaskError in this process. We use this value to # suppress redundant error messages pushed from the workers. last_task_error_raise_time = 0 # The max amount of seconds to wait before printing out an uncaught error. UNCAUGHT_ERROR_GRACE_PERIOD = 5 def print_logs(redis_client, threads_stopped): """Prints log messages from workers on all of the nodes. Args: redis_client: A client to the primary Redis shard. threads_stopped (threading.Event): A threading event used to signal to the thread that it should exit. """ pubsub_client = redis_client.pubsub(ignore_subscribe_messages=True) pubsub_client.subscribe(ray.gcs_utils.LOG_FILE_CHANNEL) localhost = services.get_node_ip_address() try: # Keep track of the number of consecutive log messages that have been # received with no break in between. If this number grows continually, # then the worker is probably not able to process the log messages as # rapidly as they are coming in. num_consecutive_messages_received = 0 while True: # Exit if we received a signal that we should stop. if threads_stopped.is_set(): return msg = pubsub_client.get_message() if msg is None: num_consecutive_messages_received = 0 threads_stopped.wait(timeout=0.01) continue num_consecutive_messages_received += 1 data = json.loads(ray.utils.decode(msg["data"])) def color_for(data): if data["pid"] == "raylet": return colorama.Fore.YELLOW else: return colorama.Fore.CYAN if data["ip"] == localhost: for line in data["lines"]: print("{}{}(pid={}){} {}".format( colorama.Style.DIM, color_for(data), data["pid"], colorama.Style.RESET_ALL, line)) else: for line in data["lines"]: print("{}{}(pid={}, ip={}){} {}".format( colorama.Style.DIM, color_for(data), data["pid"], data["ip"], colorama.Style.RESET_ALL, line)) if (num_consecutive_messages_received % 100 == 0 and num_consecutive_messages_received > 0): logger.warning( "The driver may not be able to keep up with the " "stdout/stderr of the workers. To avoid forwarding logs " "to the driver, use 'ray.init(log_to_driver=False)'.") except (OSError, redis.exceptions.ConnectionError) as e: logger.error("print_logs: {}".format(e)) finally: # Close the pubsub client to avoid leaking file descriptors. pubsub_client.close() def print_error_messages_raylet(task_error_queue, threads_stopped): """Prints message received in the given output queue. This checks periodically if any un-raised errors occured in the background. Args: task_error_queue (queue.Queue): A queue used to receive errors from the thread that listens to Redis. threads_stopped (threading.Event): A threading event used to signal to the thread that it should exit. """ while True: # Exit if we received a signal that we should stop. if threads_stopped.is_set(): return try: error, t = task_error_queue.get(block=False) except queue.Empty: threads_stopped.wait(timeout=0.01) continue # Delay errors a little bit of time to attempt to suppress redundant # messages originating from the worker. while t + UNCAUGHT_ERROR_GRACE_PERIOD > time.time(): threads_stopped.wait(timeout=1) if threads_stopped.is_set(): break if t < last_task_error_raise_time + UNCAUGHT_ERROR_GRACE_PERIOD: logger.debug("Suppressing error from worker: {}".format(error)) else: logger.error( "Possible unhandled error from worker: {}".format(error)) def listen_error_messages_raylet(worker, task_error_queue, threads_stopped): """Listen to error messages in the background on the driver. This runs in a separate thread on the driver and pushes (error, time) tuples to the output queue. Args: worker: The worker class that this thread belongs to. task_error_queue (queue.Queue): A queue used to communicate with the thread that prints the errors found by this thread. threads_stopped (threading.Event): A threading event used to signal to the thread that it should exit. """ worker.error_message_pubsub_client = worker.redis_client.pubsub( ignore_subscribe_messages=True) # Exports that are published after the call to # error_message_pubsub_client.subscribe and before the call to # error_message_pubsub_client.listen will still be processed in the loop. # Really we should just subscribe to the errors for this specific job. # However, currently all errors seem to be published on the same channel. error_pubsub_channel = str( ray.gcs_utils.TablePubsub.Value("ERROR_INFO_PUBSUB")).encode("ascii") worker.error_message_pubsub_client.subscribe(error_pubsub_channel) # worker.error_message_pubsub_client.psubscribe("*") try: # Get the errors that occurred before the call to subscribe. error_messages = ray.errors() for error_message in error_messages: logger.error(error_message) while True: # Exit if we received a signal that we should stop. if threads_stopped.is_set(): return msg = worker.error_message_pubsub_client.get_message() if msg is None: threads_stopped.wait(timeout=0.01) continue gcs_entry = ray.gcs_utils.GcsEntry.FromString(msg["data"]) assert len(gcs_entry.entries) == 1 error_data = ray.gcs_utils.ErrorTableData.FromString( gcs_entry.entries[0]) job_id = error_data.job_id if job_id not in [ worker.current_job_id.binary(), JobID.nil().binary() ]: continue error_message = error_data.error_message if (error_data.type == ray_constants.TASK_PUSH_ERROR): # Delay it a bit to see if we can suppress it task_error_queue.put((error_message, time.time())) else: logger.warning(error_message) except (OSError, redis.exceptions.ConnectionError) as e: logger.error("listen_error_messages_raylet: {}".format(e)) finally: # Close the pubsub client to avoid leaking file descriptors. worker.error_message_pubsub_client.close() def is_initialized(): """Check if ray.init has been called yet. Returns: True if ray.init has already been called and false otherwise. """ return ray.worker.global_worker.connected def connect(node, mode=WORKER_MODE, log_to_driver=False, worker=global_worker, driver_object_store_memory=None, job_id=None, internal_config=None): """Connect this worker to the raylet, to Plasma, and to Redis. Args: node (ray.node.Node): The node to connect. mode: The mode of the worker. One of SCRIPT_MODE, WORKER_MODE, and LOCAL_MODE. log_to_driver (bool): If true, then output from all of the worker processes on all nodes will be directed to the driver. worker: The ray.Worker instance. driver_object_store_memory: Limit the amount of memory the driver can use in the object store when creating objects. job_id: The ID of job. If it's None, then we will generate one. internal_config: Dictionary of (str,str) containing internal config options to override the defaults. """ # Do some basic checking to make sure we didn't call ray.init twice. error_message = "Perhaps you called ray.init twice by accident?" assert not worker.connected, error_message assert worker.cached_functions_to_run is not None, error_message # Enable nice stack traces on SIGSEGV etc. try: if not faulthandler.is_enabled(): faulthandler.enable(all_threads=False) except io.UnsupportedOperation: pass # ignore ray._raylet.set_internal_config(internal_config) if mode is not LOCAL_MODE: # Create a Redis client to primary. # The Redis client can safely be shared between threads. However, # that is not true of Redis pubsub clients. See the documentation at # https://github.com/andymccurdy/redis-py#thread-safety. worker.redis_client = node.create_redis_client() # Initialize some fields. if mode is WORKER_MODE: # We should not specify the job_id if it's `WORKER_MODE`. assert job_id is None job_id = JobID.nil() # TODO(qwang): Rename this to `worker_id_str` or type to `WorkerID` worker.worker_id = _random_string() if setproctitle: setproctitle.setproctitle("ray::IDLE") elif mode is LOCAL_MODE: if job_id is None: job_id = JobID.from_int(random.randint(1, 65535)) worker.worker_id = ray.utils.compute_driver_id_from_job( job_id).binary() else: # This is the code path of driver mode. if job_id is None: # TODO(qwang): use `GcsClient::GenerateJobId()` here. job_id = JobID.from_int( int(worker.redis_client.incr("JobCounter"))) # When tasks are executed on remote workers in the context of multiple # drivers, the current job ID is used to keep track of which job is # responsible for the task so that error messages will be propagated to # the correct driver. worker.worker_id = ray.utils.compute_driver_id_from_job( job_id).binary() if not isinstance(job_id, JobID): raise TypeError("The type of given job id must be JobID.") # All workers start out as non-actors. A worker can be turned into an actor # after it is created. worker.node = node worker.set_mode(mode) # If running Ray in LOCAL_MODE, there is no need to create call # create_worker or to start the worker service. if mode == LOCAL_MODE: worker.local_mode_manager = LocalModeManager() return # For driver's check that the version information matches the version # information that the Ray cluster was started with. try: ray.services.check_version_info(worker.redis_client) except Exception as e: if mode == SCRIPT_MODE: raise e elif mode == WORKER_MODE: traceback_str = traceback.format_exc() ray.utils.push_error_to_driver_through_redis( worker.redis_client, ray_constants.VERSION_MISMATCH_PUSH_ERROR, traceback_str, job_id=None) worker.lock = threading.RLock() # Create an object for interfacing with the global state. ray.state.state._initialize_global_state( node.redis_address, redis_password=node.redis_password) # Register the worker with Redis. if mode == SCRIPT_MODE: # The concept of a driver is the same as the concept of a "job". # Register the driver/job with Redis here. import __main__ as main driver_info = { "node_ip_address": node.node_ip_address, "driver_id": worker.worker_id, "start_time": time.time(), "plasma_store_socket": node.plasma_store_socket_name, "raylet_socket": node.raylet_socket_name, "name": (main.__file__ if hasattr(main, "__file__") else "INTERACTIVE MODE") } worker.redis_client.hmset(b"Drivers:" + worker.worker_id, driver_info) elif mode == WORKER_MODE: # Register the worker with Redis. worker_dict = { "node_ip_address": node.node_ip_address, "plasma_store_socket": node.plasma_store_socket_name, } # Check the RedirectOutput key in Redis and based on its value redirect # worker output and error to their own files. # This key is set in services.py when Redis is started. redirect_worker_output_val = worker.redis_client.get("RedirectOutput") if (redirect_worker_output_val is not None and int(redirect_worker_output_val) == 1): log_stdout_file, log_stderr_file = ( node.new_worker_redirected_log_file(worker.worker_id)) # Redirect stdout/stderr at the file descriptor level. If we simply # set sys.stdout and sys.stderr, then logging from C++ can fail to # be redirected. os.dup2(log_stdout_file.fileno(), sys.stdout.fileno()) os.dup2(log_stderr_file.fileno(), sys.stderr.fileno()) # We also manually set sys.stdout and sys.stderr because that seems # to have an affect on the output buffering. Without doing this, # stdout and stderr are heavily buffered resulting in seemingly # lost logging statements. sys.stdout = log_stdout_file sys.stderr = log_stderr_file # This should always be the first message to appear in the worker's # stdout and stderr log files. The string "Ray worker pid:" is # parsed in the log monitor process. print("Ray worker pid: {}".format(os.getpid())) print("Ray worker pid: {}".format(os.getpid()), file=sys.stderr) sys.stdout.flush() sys.stderr.flush() worker_dict["stdout_file"] = os.path.abspath(log_stdout_file.name) worker_dict["stderr_file"] = os.path.abspath(log_stderr_file.name) worker.redis_client.hmset(b"Workers:" + worker.worker_id, worker_dict) else: raise ValueError("Invalid worker mode. Expected DRIVER or WORKER.") redis_address, redis_port = node.redis_address.split(":") gcs_options = ray._raylet.GcsClientOptions( redis_address, int(redis_port), node.redis_password, ) worker.core_worker = ray._raylet.CoreWorker( (mode == SCRIPT_MODE), node.plasma_store_socket_name, node.raylet_socket_name, job_id, gcs_options, node.get_logs_dir_path(), node.node_ip_address, node.node_manager_port, ) worker.raylet_client = ray._raylet.RayletClient(worker.core_worker) if driver_object_store_memory is not None: worker.core_worker.set_object_store_client_options( "ray_driver_{}".format(os.getpid()), driver_object_store_memory) # Put something in the plasma store so that subsequent plasma store # accesses will be faster. Currently the first access is always slow, and # we don't want the user to experience this. temporary_object_id = ray.ObjectID.from_random() worker.put_object(1, object_id=temporary_object_id) ray.internal.free([temporary_object_id]) # Start the import thread worker.import_thread = import_thread.ImportThread(worker, mode, worker.threads_stopped) worker.import_thread.start() # If this is a driver running in SCRIPT_MODE, start a thread to print error # messages asynchronously in the background. Ideally the scheduler would # push messages to the driver's worker service, but we ran into bugs when # trying to properly shutdown the driver's worker service, so we are # temporarily using this implementation which constantly queries the # scheduler for new error messages. if mode == SCRIPT_MODE: q = queue.Queue() worker.listener_thread = threading.Thread( target=listen_error_messages_raylet, name="ray_listen_error_messages", args=(worker, q, worker.threads_stopped)) worker.printer_thread = threading.Thread( target=print_error_messages_raylet, name="ray_print_error_messages", args=(q, worker.threads_stopped)) worker.listener_thread.daemon = True worker.listener_thread.start() worker.printer_thread.daemon = True worker.printer_thread.start() if log_to_driver: worker.logger_thread = threading.Thread( target=print_logs, name="ray_print_logs", args=(worker.redis_client, worker.threads_stopped)) worker.logger_thread.daemon = True worker.logger_thread.start() if mode == SCRIPT_MODE: # Add the directory containing the script that is running to the Python # paths of the workers. Also add the current directory. Note that this # assumes that the directory structures on the machines in the clusters # are the same. script_directory = os.path.abspath(os.path.dirname(sys.argv[0])) current_directory = os.path.abspath(os.path.curdir) worker.run_function_on_all_workers( lambda worker_info: sys.path.insert(1, script_directory)) worker.run_function_on_all_workers( lambda worker_info: sys.path.insert(1, current_directory)) # TODO(rkn): Here we first export functions to run, then remote # functions. The order matters. For example, one of the functions to # run may set the Python path, which is needed to import a module used # to define a remote function. We may want to change the order to # simply be the order in which the exports were defined on the driver. # In addition, we will need to retain the ability to decide what the # first few exports are (mostly to set the Python path). Additionally, # note that the first exports to be defined on the driver will be the # ones defined in separate modules that are imported by the driver. # Export cached functions_to_run. for function in worker.cached_functions_to_run: worker.run_function_on_all_workers(function) worker.cached_functions_to_run = None def disconnect(exiting_interpreter=False): """Disconnect this worker from the raylet and object store.""" # Reset the list of cached remote functions and actors so that if more # remote functions or actors are defined and then connect is called again, # the remote functions will be exported. This is mostly relevant for the # tests. worker = global_worker if worker.connected: # Shutdown all of the threads that we've started. TODO(rkn): This # should be handled cleanly in the worker object's destructor and not # in this disconnect method. worker.threads_stopped.set() if hasattr(worker, "import_thread"): worker.import_thread.join_import_thread() if hasattr(worker, "listener_thread"): worker.listener_thread.join() if hasattr(worker, "printer_thread"): worker.printer_thread.join() if hasattr(worker, "logger_thread"): worker.logger_thread.join() worker.threads_stopped.clear() worker._session_index += 1 worker.node = None # Disconnect the worker from the node. worker.cached_functions_to_run = [] worker.serialization_context_map.clear() @contextmanager def _changeproctitle(title, next_title): if setproctitle: setproctitle.setproctitle(title) yield if setproctitle: setproctitle.setproctitle(next_title) def register_custom_serializer(cls, serializer=None, deserializer=None, use_pickle=False, use_dict=False, local=None, job_id=None, class_id=None): """Registers custom functions for efficient object serialization. The serializer and deserializer are used when transferring objects of `cls` across processes and nodes. This can be significantly faster than the Ray default fallbacks. Wraps `register_custom_serializer` underneath. `use_pickle` tells Ray to automatically use cloudpickle for serialization, and `use_dict` automatically uses `cls.__dict__`. When calling this function, you can only provide one of the following: 1. serializer and deserializer 2. `use_pickle` 3. `use_dict` Args: cls (type): The class that ray should use this custom serializer for. serializer: The custom serializer that takes in a cls instance and outputs a serialized representation. use_pickle and use_dict must be False if provided. deserializer: The custom deserializer that takes in a serialized representation of the cls and outputs a cls instance. use_pickle and use_dict must be False if provided. use_pickle (bool): If true, objects of this class will be serialized using pickle. Must be False if use_dict is true. use_dict (bool): If true, objects of this class be serialized turning their __dict__ fields into a dictionary. Must be False if use_pickle is true. local: Deprecated. job_id: Deprecated. class_id (str): Unique ID of the class. Autogenerated if None. """ if job_id: raise DeprecationWarning( "`job_id` is no longer a valid parameter and will be removed in " "future versions of Ray. If this breaks your application, " "see `SerializationContext.register_custom_serializer`.") if local: raise DeprecationWarning( "`local` is no longer a valid parameter and will be removed in " "future versions of Ray. If this breaks your application, " "see `SerializationContext.register_custom_serializer`.") context = global_worker.get_serialization_context() context.register_custom_serializer( cls, use_pickle=use_pickle, use_dict=use_dict, serializer=serializer, deserializer=deserializer, class_id=class_id) def show_in_webui(message): """Display message in dashboard. Display message for the current task or actor in the dashboard. For example, this can be used to display the status of a long-running computation. Args: message (str): Message to be displayed. """ worker = global_worker worker.check_connected() worker.core_worker.set_webui_display(message.encode()) def get(object_ids, timeout=None): """Get a remote object or a list of remote objects from the object store. This method blocks until the object corresponding to the object ID is available in the local object store. If this object is not in the local object store, it will be shipped from an object store that has it (once the object has been created). If object_ids is a list, then the objects corresponding to each object in the list will be returned. Args: object_ids: Object ID of the object to get or a list of object IDs to get. timeout (Optional[float]): The maximum amount of time in seconds to wait before returning. Returns: A Python object or a list of Python objects. Raises: RayTimeoutError: A RayTimeoutError is raised if a timeout is set and the get takes longer than timeout to return. Exception: An exception is raised if the task that created the object or that created one of the objects raised an exception. """ worker = global_worker worker.check_connected() if hasattr( worker, "core_worker") and worker.core_worker.current_actor_is_asyncio(): raise RayError("Using blocking ray.get inside async actor. " "This blocks the event loop. Please " "use `await` on object id with asyncio.gather.") with profiling.profile("ray.get"): is_individual_id = isinstance(object_ids, ray.ObjectID) if is_individual_id: object_ids = [object_ids] if not isinstance(object_ids, list): raise ValueError("'object_ids' must either be an object ID " "or a list of object IDs.") global last_task_error_raise_time # TODO(ujvl): Consider how to allow user to retrieve the ready objects. values = worker.get_objects(object_ids, timeout=timeout) for i, value in enumerate(values): if isinstance(value, RayError): last_task_error_raise_time = time.time() if isinstance(value, ray.exceptions.UnreconstructableError): worker.core_worker.dump_object_store_memory_usage() if isinstance(value, RayTaskError): raise value.as_instanceof_cause() else: raise value # Run post processors. for post_processor in worker._post_get_hooks: values = post_processor(object_ids, values) if is_individual_id: values = values[0] return values def put(value, weakref=False): """Store an object in the object store. The object may not be evicted while a reference to the returned ID exists. Note that this pinning only applies to the particular object ID returned by put, not object IDs in general. Args: value: The Python object to be stored. weakref: If set, allows the object to be evicted while a reference to the returned ID exists. You might want to set this if putting a lot of objects that you might not need in the future. Returns: The object ID assigned to this value. """ worker = global_worker worker.check_connected() with profiling.profile("ray.put"): if worker.mode == LOCAL_MODE: object_id = worker.local_mode_manager.put_object(value) else: try: object_id = worker.put_object(value, pin_object=not weakref) except ObjectStoreFullError: logger.info( "Put failed since the value was either too large or the " "store was full of pinned objects. If you are putting " "and holding references to a lot of object ids, consider " "ray.put(value, weakref=True) to allow object data to " "be evicted early.") raise return object_id def wait(object_ids, num_returns=1, timeout=None): """Return a list of IDs that are ready and a list of IDs that are not. .. warning:: The **timeout** argument used to be in **milliseconds** (up through ``ray==0.6.1``) and now it is in **seconds**. If timeout is set, the function returns either when the requested number of IDs are ready or when the timeout is reached, whichever occurs first. If it is not set, the function simply waits until that number of objects is ready and returns that exact number of object IDs. This method returns two lists. The first list consists of object IDs that correspond to objects that are available in the object store. The second list corresponds to the rest of the object IDs (which may or may not be ready). Ordering of the input list of object IDs is preserved. That is, if A precedes B in the input list, and both are in the ready list, then A will precede B in the ready list. This also holds true if A and B are both in the remaining list. Args: object_ids (List[ObjectID]): List of object IDs for objects that may or may not be ready. Note that these IDs must be unique. num_returns (int): The number of object IDs that should be returned. timeout (float): The maximum amount of time in seconds to wait before returning. Returns: A list of object IDs that are ready and a list of the remaining object IDs. """ worker = global_worker if hasattr( worker, "core_worker") and worker.core_worker.current_actor_is_asyncio(): raise RayError("Using blocking ray.wait inside async method. " "This blocks the event loop. Please use `await` " "on object id with asyncio.wait. ") if isinstance(object_ids, ObjectID): raise TypeError( "wait() expected a list of ray.ObjectID, got a single ray.ObjectID" ) if not isinstance(object_ids, list): raise TypeError( "wait() expected a list of ray.ObjectID, got {}".format( type(object_ids))) if isinstance(timeout, int) and timeout != 0: logger.warning("The 'timeout' argument now requires seconds instead " "of milliseconds. This message can be suppressed by " "passing in a float.") if timeout is not None and timeout < 0: raise ValueError("The 'timeout' argument must be nonnegative. " "Received {}".format(timeout)) for object_id in object_ids: if not isinstance(object_id, ObjectID): raise TypeError("wait() expected a list of ray.ObjectID, " "got list containing {}".format(type(object_id))) worker.check_connected() # TODO(swang): Check main thread. with profiling.profile("ray.wait"): # When Ray is run in LOCAL_MODE, all functions are run immediately, # so all objects in object_id are ready. if worker.mode == LOCAL_MODE: return object_ids[:num_returns], object_ids[num_returns:] # TODO(rkn): This is a temporary workaround for # https://github.com/ray-project/ray/issues/997. However, it should be # fixed in Arrow instead of here. if len(object_ids) == 0: return [], [] if len(object_ids) != len(set(object_ids)): raise Exception("Wait requires a list of unique object IDs.") if num_returns <= 0: raise Exception( "Invalid number of objects to return %d." % num_returns) if num_returns > len(object_ids): raise Exception("num_returns cannot be greater than the number " "of objects provided to ray.wait.") timeout = timeout if timeout is not None else 10**6 timeout_milliseconds = int(timeout * 1000) ready_ids, remaining_ids = worker.core_worker.wait( object_ids, num_returns, timeout_milliseconds, worker.current_task_id, ) return ready_ids, remaining_ids def _mode(worker=global_worker): """This is a wrapper around worker.mode. We use this wrapper so that in the remote decorator, we can call _mode() instead of worker.mode. The difference is that when we attempt to serialize remote functions, we don't attempt to serialize the worker object, which cannot be serialized. """ return worker.mode def get_global_worker(): return global_worker def make_decorator(num_return_vals=None, num_cpus=None, num_gpus=None, memory=None, object_store_memory=None, resources=None, max_calls=None, max_retries=None, max_reconstructions=None, worker=None): def decorator(function_or_class): if (inspect.isfunction(function_or_class) or is_cython(function_or_class)): # Set the remote function default resources. if max_reconstructions is not None: raise Exception("The keyword 'max_reconstructions' is not " "allowed for remote functions.") return ray.remote_function.RemoteFunction( function_or_class, num_cpus, num_gpus, memory, object_store_memory, resources, num_return_vals, max_calls, max_retries) if inspect.isclass(function_or_class): if num_return_vals is not None: raise Exception("The keyword 'num_return_vals' is not allowed " "for actors.") if max_calls is not None: raise Exception("The keyword 'max_calls' is not allowed for " "actors.") return worker.make_actor(function_or_class, num_cpus, num_gpus, memory, object_store_memory, resources, max_reconstructions) raise Exception("The @ray.remote decorator must be applied to " "either a function or to a class.") return decorator def remote(*args, **kwargs): """Define a remote function or an actor class. This can be used with no arguments to define a remote function or actor as follows: .. code-block:: python @ray.remote def f(): return 1 @ray.remote class Foo: def method(self): return 1 It can also be used with specific keyword arguments: * **num_return_vals:** This is only for *remote functions*. It specifies the number of object IDs returned by the remote function invocation. * **num_cpus:** The quantity of CPU cores to reserve for this task or for the lifetime of the actor. * **num_gpus:** The quantity of GPUs to reserve for this task or for the lifetime of the actor. * **resources:** The quantity of various custom resources to reserve for this task or for the lifetime of the actor. This is a dictionary mapping strings (resource names) to numbers. * **max_calls:** Only for *remote functions*. This specifies the maximum number of times that a given worker can execute the given remote function before it must exit (this can be used to address memory leaks in third-party libraries or to reclaim resources that cannot easily be released, e.g., GPU memory that was acquired by TensorFlow). By default this is infinite. * **max_reconstructions**: Only for *actors*. This specifies the maximum number of times that the actor should be reconstructed when it dies unexpectedly. The minimum valid value is 0 (default), which indicates that the actor doesn't need to be reconstructed. And the maximum valid value is ray.ray_constants.INFINITE_RECONSTRUCTION. * **max_retries**: Only for *remote functions*. This specifies the maximum number of times that the remote function should be rerun when the worker process executing it crashes unexpectedly. The minimum valid value is 0, the default is 4 (default), and the maximum valid value is ray.ray_constants.INFINITE_RECONSTRUCTION. This can be done as follows: .. code-block:: python @ray.remote(num_gpus=1, max_calls=1, num_return_vals=2) def f(): return 1, 2 @ray.remote(num_cpus=2, resources={"CustomResource": 1}) class Foo: def method(self): return 1 Remote task and actor objects returned by @ray.remote can also be dynamically modified with the same arguments as above using ``.options()`` as follows: .. code-block:: python @ray.remote(num_gpus=1, max_calls=1, num_return_vals=2) def f(): return 1, 2 g = f.options(num_gpus=2, max_calls=None) @ray.remote(num_cpus=2, resources={"CustomResource": 1}) class Foo: def method(self): return 1 Bar = Foo.options(num_cpus=1, resources=None) Running remote actors will be terminated when the actor handle to them in Python is deleted, which will cause them to complete any outstanding work and then shut down. If you want to kill them immediately, you can also call ``actor_handle.__ray_kill__()``. """ worker = get_global_worker() if len(args) == 1 and len(kwargs) == 0 and callable(args[0]): # This is the case where the decorator is just @ray.remote. return make_decorator(worker=worker)(args[0]) # Parse the keyword arguments from the decorator. error_string = ("The @ray.remote decorator must be applied either " "with no arguments and no parentheses, for example " "'@ray.remote', or it must be applied using some of " "the arguments 'num_return_vals', 'num_cpus', 'num_gpus', " "'memory', 'object_store_memory', 'resources', " "'max_calls', or 'max_reconstructions', like " "'@ray.remote(num_return_vals=2, " "resources={\"CustomResource\": 1})'.") assert len(args) == 0 and len(kwargs) > 0, error_string for key in kwargs: assert key in [ "num_return_vals", "num_cpus", "num_gpus", "memory", "object_store_memory", "resources", "max_calls", "max_reconstructions", "max_retries", ], error_string num_cpus = kwargs["num_cpus"] if "num_cpus" in kwargs else None num_gpus = kwargs["num_gpus"] if "num_gpus" in kwargs else None resources = kwargs.get("resources") if not isinstance(resources, dict) and resources is not None: raise Exception("The 'resources' keyword argument must be a " "dictionary, but received type {}.".format( type(resources))) if resources is not None: assert "CPU" not in resources, "Use the 'num_cpus' argument." assert "GPU" not in resources, "Use the 'num_gpus' argument." # Handle other arguments. num_return_vals = kwargs.get("num_return_vals") max_calls = kwargs.get("max_calls") max_reconstructions = kwargs.get("max_reconstructions") memory = kwargs.get("memory") object_store_memory = kwargs.get("object_store_memory") max_retries = kwargs.get("max_retries") return make_decorator( num_return_vals=num_return_vals, num_cpus=num_cpus, num_gpus=num_gpus, memory=memory, object_store_memory=object_store_memory, resources=resources, max_calls=max_calls, max_reconstructions=max_reconstructions, max_retries=max_retries, worker=worker)
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/ray/workers/default_worker.py
Python
import argparse import json import ray import ray.actor import ray.node import ray.ray_constants as ray_constants import ray.utils from ray.parameter import RayParams parser = argparse.ArgumentParser( description=("Parse addresses for the worker " "to connect to.")) parser.add_argument( "--node-ip-address", required=True, type=str, help="the ip address of the worker's node") parser.add_argument( "--node-manager-port", required=True, type=int, help="the port of the worker's node") parser.add_argument( "--redis-address", required=True, type=str, help="the address to use for Redis") parser.add_argument( "--redis-password", required=False, type=str, default=None, help="the password to use for Redis") parser.add_argument( "--object-store-name", required=True, type=str, help="the object store's name") parser.add_argument( "--raylet-name", required=False, type=str, help="the raylet's name") parser.add_argument( "--logging-level", required=False, type=str, default=ray_constants.LOGGER_LEVEL, choices=ray_constants.LOGGER_LEVEL_CHOICES, help=ray_constants.LOGGER_LEVEL_HELP) parser.add_argument( "--logging-format", required=False, type=str, default=ray_constants.LOGGER_FORMAT, help=ray_constants.LOGGER_FORMAT_HELP) parser.add_argument( "--config-list", required=False, type=str, default=None, help="Override internal config options for the worker process.") parser.add_argument( "--temp-dir", required=False, type=str, default=None, help="Specify the path of the temporary directory use by Ray process.") parser.add_argument( "--load-code-from-local", default=False, action="store_true", help="True if code is loaded from local files, as opposed to the GCS.") parser.add_argument( "--use-pickle", default=False, action="store_true", help="True if cloudpickle should be used for serialization.") if __name__ == "__main__": args = parser.parse_args() ray.utils.setup_logger(args.logging_level, args.logging_format) internal_config = {} if args.config_list is not None: config_list = args.config_list.split(",") if len(config_list) > 1: i = 0 while i < len(config_list): internal_config[config_list[i]] = config_list[i + 1] i += 2 ray_params = RayParams( node_ip_address=args.node_ip_address, node_manager_port=args.node_manager_port, redis_address=args.redis_address, redis_password=args.redis_password, plasma_store_socket_name=args.object_store_name, raylet_socket_name=args.raylet_name, temp_dir=args.temp_dir, load_code_from_local=args.load_code_from_local, use_pickle=args.use_pickle, _internal_config=json.dumps(internal_config), ) node = ray.node.Node( ray_params, head=False, shutdown_at_exit=False, spawn_reaper=False, connect_only=True) ray.worker._global_node = node ray.worker.connect( node, mode=ray.WORKER_MODE, internal_config=internal_config) ray.worker.global_worker.main_loop()
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
python/setup.py
Python
from itertools import chain import os import re import shutil import subprocess import sys from setuptools import setup, find_packages, Distribution import setuptools.command.build_ext as _build_ext # Ideally, we could include these files by putting them in a # MANIFEST.in or using the package_data argument to setup, but the # MANIFEST.in gets applied at the very beginning when setup.py runs # before these files have been created, so we have to move the files # manually. # NOTE: The lists below must be kept in sync with ray/BUILD.bazel. ray_files = [ "ray/core/src/ray/thirdparty/redis/src/redis-server", "ray/core/src/ray/gcs/redis_module/libray_redis_module.so", "ray/core/src/plasma/plasma_store_server", "ray/_raylet.so", "ray/core/src/ray/raylet/raylet_monitor", "ray/core/src/ray/raylet/raylet", "ray/dashboard/dashboard.py", "ray/streaming/_streaming.so", ] build_java = os.getenv("RAY_INSTALL_JAVA") == "1" if build_java: ray_files.append("ray/jars/ray_dist.jar") # These are the directories where automatically generated Python protobuf # bindings are created. generated_python_directories = [ "ray/core/generated", "ray/streaming/generated", ] optional_ray_files = [] ray_autoscaler_files = [ "ray/autoscaler/aws/example-full.yaml", "ray/autoscaler/gcp/example-full.yaml", "ray/autoscaler/local/example-full.yaml", "ray/autoscaler/kubernetes/example-full.yaml", "ray/autoscaler/kubernetes/kubectl-rsync.sh", ] ray_project_files = [ "ray/projects/schema.json", "ray/projects/templates/cluster_template.yaml", "ray/projects/templates/project_template.yaml", "ray/projects/templates/requirements.txt" ] ray_dashboard_files = [ os.path.join(dirpath, filename) for dirpath, dirnames, filenames in os.walk("ray/dashboard/client/build") for filename in filenames ] optional_ray_files += ray_autoscaler_files optional_ray_files += ray_project_files optional_ray_files += ray_dashboard_files if "RAY_USE_NEW_GCS" in os.environ and os.environ["RAY_USE_NEW_GCS"] == "on": ray_files += [ "ray/core/src/credis/build/src/libmember.so", "ray/core/src/credis/build/src/libmaster.so", "ray/core/src/credis/redis/src/redis-server" ] extras = { "debug": ["psutil", "setproctitle", "py-spy >= 0.2.0"], "dashboard": ["aiohttp", "google", "grpcio", "psutil", "setproctitle"], "serve": ["uvicorn", "pygments", "werkzeug", "flask", "pandas", "blist"], "tune": ["tabulate", "tensorboardX"], } extras["rllib"] = extras["tune"] + [ "pyyaml", "gym[atari]", "opencv-python-headless", "lz4", "scipy", ] extras["all"] = list(set(chain.from_iterable(extras.values()))) class build_ext(_build_ext.build_ext): def run(self): # Note: We are passing in sys.executable so that we use the same # version of Python to build pyarrow inside the build.sh script. Note # that certain flags will not be passed along such as --user or sudo. # TODO(rkn): Fix this. command = ["../build.sh", "-p", sys.executable] if build_java: # Also build binaries for Java if the above env variable exists. command += ["-l", "python,java"] subprocess.check_call(command) # We also need to install pyarrow along with Ray, so make sure that the # relevant non-Python pyarrow files get copied. pyarrow_files = [] for (root, dirs, filenames) in os.walk("./ray/pyarrow_files/pyarrow"): for name in filenames: pyarrow_files.append(os.path.join(root, name)) # We also need to install pickle5 along with Ray, so make sure that the # relevant non-Python pickle5 files get copied. pickle5_files = [] for (root, dirs, filenames) in os.walk("./ray/pickle5_files/pickle5"): for name in filenames: pickle5_files.append(os.path.join(root, name)) files_to_include = ray_files + pyarrow_files + pickle5_files # Copy over the autogenerated protobuf Python bindings. for directory in generated_python_directories: for filename in os.listdir(directory): if filename[-3:] == ".py": files_to_include.append(os.path.join(directory, filename)) for filename in files_to_include: self.move_file(filename) # Try to copy over the optional files. for filename in optional_ray_files: try: self.move_file(filename) except Exception: print("Failed to copy optional file {}. This is ok." .format(filename)) def move_file(self, filename): # TODO(rkn): This feels very brittle. It may not handle all cases. See # https://github.com/apache/arrow/blob/master/python/setup.py for an # example. source = filename destination = os.path.join(self.build_lib, filename) # Create the target directory if it doesn't already exist. parent_directory = os.path.dirname(destination) if not os.path.exists(parent_directory): os.makedirs(parent_directory) if not os.path.exists(destination): print("Copying {} to {}.".format(source, destination)) shutil.copy(source, destination, follow_symlinks=True) class BinaryDistribution(Distribution): def has_ext_modules(self): return True def find_version(*filepath): # Extract version information from filepath here = os.path.abspath(os.path.dirname(__file__)) with open(os.path.join(here, *filepath)) as fp: version_match = re.search(r"^__version__ = ['\"]([^'\"]*)['\"]", fp.read(), re.M) if version_match: return version_match.group(1) raise RuntimeError("Unable to find version string.") requires = [ "numpy >= 1.16", "filelock", "jsonschema", "funcsigs", "click", "colorama", "packaging", "pytest", "pyyaml", "redis>=3.3.2", # NOTE: Don't upgrade the version of six! Doing so causes installation # problems. See https://github.com/ray-project/ray/issues/4169. "six >= 1.0.0", "faulthandler;python_version<'3.3'", "protobuf >= 3.8.0", "cloudpickle", ] setup( name="ray", version=find_version("ray", "__init__.py"), author="Ray Team", author_email="ray-dev@googlegroups.com", description=("A system for parallel and distributed Python that unifies " "the ML ecosystem."), long_description=open("../README.rst").read(), url="https://github.com/ray-project/ray", keywords=("ray distributed parallel machine-learning " "reinforcement-learning deep-learning python"), packages=find_packages(), cmdclass={"build_ext": build_ext}, # The BinaryDistribution argument triggers build_ext. distclass=BinaryDistribution, install_requires=requires, setup_requires=["cython >= 0.29"], extras_require=extras, entry_points={ "console_scripts": [ "ray=ray.scripts.scripts:main", "rllib=ray.rllib.scripts:cli [rllib]", "tune=ray.tune.scripts:cli" ] }, include_package_data=True, zip_safe=False, license="Apache 2.0")
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/__init__.py
Python
import logging # Note: do not introduce unnecessary library dependencies here, e.g. gym. # This file is imported from the tune module in order to register RLlib agents. from ray.rllib.env.base_env import BaseEnv from ray.rllib.env.external_env import ExternalEnv from ray.rllib.env.multi_agent_env import MultiAgentEnv from ray.rllib.env.vector_env import VectorEnv from ray.rllib.evaluation.policy_graph import PolicyGraph from ray.rllib.evaluation.rollout_worker import RolloutWorker from ray.rllib.evaluation.tf_policy_graph import TFPolicyGraph from ray.rllib.policy.policy import Policy from ray.rllib.policy.sample_batch import SampleBatch from ray.rllib.policy.tf_policy import TFPolicy from ray.tune.registry import register_trainable def _setup_logger(): logger = logging.getLogger("ray.rllib") handler = logging.StreamHandler() handler.setFormatter( logging.Formatter( "%(asctime)s\t%(levelname)s %(filename)s:%(lineno)s -- %(message)s" )) logger.addHandler(handler) logger.propagate = False def _register_all(): from ray.rllib.agents.trainer import Trainer, with_common_config from ray.rllib.agents.registry import ALGORITHMS, get_agent_class from ray.rllib.contrib.registry import CONTRIBUTED_ALGORITHMS for key in list(ALGORITHMS.keys()) + list(CONTRIBUTED_ALGORITHMS.keys( )) + ["__fake", "__sigmoid_fake_data", "__parameter_tuning"]: register_trainable(key, get_agent_class(key)) def _see_contrib(name): """Returns dummy agent class warning algo is in contrib/.""" class _SeeContrib(Trainer): _name = "SeeContrib" _default_config = with_common_config({}) def _setup(self, config): raise NameError( "Please run `contrib/{}` instead.".format(name)) return _SeeContrib # also register the aliases minus contrib/ to give a good error message for key in list(CONTRIBUTED_ALGORITHMS.keys()): assert key.startswith("contrib/") alias = key.split("/", 1)[1] register_trainable(alias, _see_contrib(alias)) _setup_logger() _register_all() __all__ = [ "Policy", "PolicyGraph", "TFPolicy", "TFPolicyGraph", "RolloutWorker", "PolicyEvaluator", "SampleBatch", "BaseEnv", "MultiAgentEnv", "VectorEnv", "ExternalEnv", ]
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/__init__.py
Python
from ray.rllib.agents.trainer import Trainer, with_common_config from ray.rllib.agents.agent import Agent __all__ = ["Agent", "Trainer", "with_common_config"]
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/a3c/__init__.py
Python
from ray.rllib.agents.a3c.a3c import A3CTrainer, DEFAULT_CONFIG from ray.rllib.agents.a3c.a2c import A2CTrainer from ray.rllib.utils import renamed_agent A2CAgent = renamed_agent(A2CTrainer) A3CAgent = renamed_agent(A3CTrainer) __all__ = [ "A2CAgent", "A3CAgent", "A2CTrainer", "A3CTrainer", "DEFAULT_CONFIG" ]
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/a3c/a2c.py
Python
from ray.rllib.agents.a3c.a3c import DEFAULT_CONFIG as A3C_CONFIG, \ validate_config, get_policy_class from ray.rllib.optimizers import SyncSamplesOptimizer, MicrobatchOptimizer from ray.rllib.agents.a3c.a3c_tf_policy import A3CTFPolicy from ray.rllib.agents.trainer_template import build_trainer from ray.rllib.utils import merge_dicts A2C_DEFAULT_CONFIG = merge_dicts( A3C_CONFIG, { "sample_batch_size": 20, "min_iter_time_s": 10, "sample_async": False, # A2C supports microbatching, in which we accumulate gradients over # batch of this size until the train batch size is reached. This allows # training with batch sizes much larger than can fit in GPU memory. # To enable, set this to a value less than the train batch size. "microbatch_size": None, }, ) def choose_policy_optimizer(workers, config): if config["microbatch_size"]: return MicrobatchOptimizer( workers, train_batch_size=config["train_batch_size"], microbatch_size=config["microbatch_size"]) else: return SyncSamplesOptimizer( workers, train_batch_size=config["train_batch_size"]) A2CTrainer = build_trainer( name="A2C", default_config=A2C_DEFAULT_CONFIG, default_policy=A3CTFPolicy, get_policy_class=get_policy_class, make_policy_optimizer=choose_policy_optimizer, validate_config=validate_config)
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/a3c/a3c.py
Python
from ray.rllib.agents.a3c.a3c_tf_policy import A3CTFPolicy from ray.rllib.agents.trainer import with_common_config from ray.rllib.agents.trainer_template import build_trainer from ray.rllib.optimizers import AsyncGradientsOptimizer # yapf: disable # __sphinx_doc_begin__ DEFAULT_CONFIG = with_common_config({ # Size of rollout batch "sample_batch_size": 10, # GAE(gamma) parameter "lambda": 1.0, # Max global norm for each gradient calculated by worker "grad_clip": 40.0, # Learning rate "lr": 0.0001, # Learning rate schedule "lr_schedule": None, # Value Function Loss coefficient "vf_loss_coeff": 0.5, # Entropy coefficient "entropy_coeff": 0.01, # Min time per iteration "min_iter_time_s": 5, # Workers sample async. Note that this increases the effective # sample_batch_size by up to 5x due to async buffering of batches. "sample_async": True, }) # __sphinx_doc_end__ # yapf: enable def get_policy_class(config): if config["use_pytorch"]: from ray.rllib.agents.a3c.a3c_torch_policy import \ A3CTorchPolicy return A3CTorchPolicy else: return A3CTFPolicy def validate_config(config): if config["entropy_coeff"] < 0: raise DeprecationWarning("entropy_coeff must be >= 0") if config["sample_async"] and config["use_pytorch"]: raise ValueError( "The sample_async option is not supported with use_pytorch: " "Multithreading can be lead to crashes if used with pytorch.") def make_async_optimizer(workers, config): return AsyncGradientsOptimizer(workers, **config["optimizer"]) A3CTrainer = build_trainer( name="A3C", default_config=DEFAULT_CONFIG, default_policy=A3CTFPolicy, get_policy_class=get_policy_class, validate_config=validate_config, make_policy_optimizer=make_async_optimizer)
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/a3c/a3c_tf_policy.py
Python
"""Note: Keep in sync with changes to VTraceTFPolicy.""" import ray from ray.rllib.policy.sample_batch import SampleBatch from ray.rllib.utils.explained_variance import explained_variance from ray.rllib.evaluation.postprocessing import compute_advantages, \ Postprocessing from ray.rllib.policy.tf_policy_template import build_tf_policy from ray.rllib.policy.tf_policy import LearningRateSchedule from ray.rllib.utils.tf_ops import make_tf_callable from ray.rllib.utils import try_import_tf tf = try_import_tf() class A3CLoss: def __init__(self, action_dist, actions, advantages, v_target, vf, vf_loss_coeff=0.5, entropy_coeff=0.01): log_prob = action_dist.logp(actions) # The "policy gradients" loss self.pi_loss = -tf.reduce_sum(log_prob * advantages) delta = vf - v_target self.vf_loss = 0.5 * tf.reduce_sum(tf.square(delta)) self.entropy = tf.reduce_sum(action_dist.entropy()) self.total_loss = (self.pi_loss + self.vf_loss * vf_loss_coeff - self.entropy * entropy_coeff) def actor_critic_loss(policy, model, dist_class, train_batch): model_out, _ = model.from_batch(train_batch) action_dist = dist_class(model_out, model) policy.loss = A3CLoss(action_dist, train_batch[SampleBatch.ACTIONS], train_batch[Postprocessing.ADVANTAGES], train_batch[Postprocessing.VALUE_TARGETS], model.value_function(), policy.config["vf_loss_coeff"], policy.config["entropy_coeff"]) return policy.loss.total_loss def postprocess_advantages(policy, sample_batch, other_agent_batches=None, episode=None): completed = sample_batch[SampleBatch.DONES][-1] if completed: last_r = 0.0 else: next_state = [] for i in range(policy.num_state_tensors()): next_state.append([sample_batch["state_out_{}".format(i)][-1]]) last_r = policy._value(sample_batch[SampleBatch.NEXT_OBS][-1], sample_batch[SampleBatch.ACTIONS][-1], sample_batch[SampleBatch.REWARDS][-1], *next_state) return compute_advantages(sample_batch, last_r, policy.config["gamma"], policy.config["lambda"]) def add_value_function_fetch(policy): return {SampleBatch.VF_PREDS: policy.model.value_function()} class ValueNetworkMixin: def __init__(self): @make_tf_callable(self.get_session()) def value(ob, prev_action, prev_reward, *state): model_out, _ = self.model({ SampleBatch.CUR_OBS: tf.convert_to_tensor([ob]), SampleBatch.PREV_ACTIONS: tf.convert_to_tensor([prev_action]), SampleBatch.PREV_REWARDS: tf.convert_to_tensor([prev_reward]), "is_training": tf.convert_to_tensor(False), }, [tf.convert_to_tensor([s]) for s in state], tf.convert_to_tensor([1])) return self.model.value_function()[0] self._value = value def stats(policy, train_batch): return { "cur_lr": tf.cast(policy.cur_lr, tf.float64), "policy_loss": policy.loss.pi_loss, "policy_entropy": policy.loss.entropy, "var_gnorm": tf.global_norm(list(policy.model.trainable_variables())), "vf_loss": policy.loss.vf_loss, } def grad_stats(policy, train_batch, grads): return { "grad_gnorm": tf.global_norm(grads), "vf_explained_var": explained_variance( train_batch[Postprocessing.VALUE_TARGETS], policy.model.value_function()), } def clip_gradients(policy, optimizer, loss): grads_and_vars = optimizer.compute_gradients( loss, policy.model.trainable_variables()) grads = [g for (g, v) in grads_and_vars] grads, _ = tf.clip_by_global_norm(grads, policy.config["grad_clip"]) clipped_grads = list(zip(grads, policy.model.trainable_variables())) return clipped_grads def setup_mixins(policy, obs_space, action_space, config): ValueNetworkMixin.__init__(policy) LearningRateSchedule.__init__(policy, config["lr"], config["lr_schedule"]) A3CTFPolicy = build_tf_policy( name="A3CTFPolicy", get_default_config=lambda: ray.rllib.agents.a3c.a3c.DEFAULT_CONFIG, loss_fn=actor_critic_loss, stats_fn=stats, grad_stats_fn=grad_stats, gradients_fn=clip_gradients, postprocess_fn=postprocess_advantages, extra_action_fetches_fn=add_value_function_fetch, before_loss_init=setup_mixins, mixins=[ValueNetworkMixin, LearningRateSchedule])
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/a3c/a3c_torch_policy.py
Python
import ray from ray.rllib.evaluation.postprocessing import compute_advantages, \ Postprocessing from ray.rllib.policy.sample_batch import SampleBatch from ray.rllib.policy.torch_policy_template import build_torch_policy from ray.rllib.utils.framework import try_import_torch torch, nn = try_import_torch() F = nn.functional def actor_critic_loss(policy, model, dist_class, train_batch): logits, _ = model.from_batch(train_batch) values = model.value_function() dist = dist_class(logits, model) log_probs = dist.logp(train_batch[SampleBatch.ACTIONS]) policy.entropy = dist.entropy().mean() policy.pi_err = -train_batch[Postprocessing.ADVANTAGES].dot( log_probs.reshape(-1)) policy.value_err = F.mse_loss( values.reshape(-1), train_batch[Postprocessing.VALUE_TARGETS]) overall_err = sum([ policy.pi_err, policy.config["vf_loss_coeff"] * policy.value_err, -policy.config["entropy_coeff"] * policy.entropy, ]) return overall_err def loss_and_entropy_stats(policy, train_batch): return { "policy_entropy": policy.entropy.item(), "policy_loss": policy.pi_err.item(), "vf_loss": policy.value_err.item(), } def add_advantages(policy, sample_batch, other_agent_batches=None, episode=None): completed = sample_batch[SampleBatch.DONES][-1] if completed: last_r = 0.0 else: last_r = policy._value(sample_batch[SampleBatch.NEXT_OBS][-1]) return compute_advantages(sample_batch, last_r, policy.config["gamma"], policy.config["lambda"]) def model_value_predictions(policy, input_dict, state_batches, model, action_dist): return {SampleBatch.VF_PREDS: model.value_function().cpu().numpy()} def apply_grad_clipping(policy): info = {} if policy.config["grad_clip"]: total_norm = nn.utils.clip_grad_norm_(policy.model.parameters(), policy.config["grad_clip"]) info["grad_gnorm"] = total_norm return info def torch_optimizer(policy, config): return torch.optim.Adam(policy.model.parameters(), lr=config["lr"]) class ValueNetworkMixin: def _value(self, obs): obs = torch.from_numpy(obs).float().unsqueeze(0).to(self.device) _ = self.model({"obs": obs}, [], [1]) return self.model.value_function().detach().cpu().numpy().squeeze() A3CTorchPolicy = build_torch_policy( name="A3CTorchPolicy", get_default_config=lambda: ray.rllib.agents.a3c.a3c.DEFAULT_CONFIG, loss_fn=actor_critic_loss, stats_fn=loss_and_entropy_stats, postprocess_fn=add_advantages, extra_action_out_fn=model_value_predictions, extra_grad_process_fn=apply_grad_clipping, optimizer_fn=torch_optimizer, mixins=[ValueNetworkMixin])
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/agent.py
Python
from ray.rllib.agents.trainer import Trainer from ray.rllib.utils import renamed_agent Agent = renamed_agent(Trainer)
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/ars/__init__.py
Python
from ray.rllib.agents.ars.ars import (ARSTrainer, DEFAULT_CONFIG) from ray.rllib.utils import renamed_agent ARSAgent = renamed_agent(ARSTrainer) __all__ = ["ARSAgent", "ARSTrainer", "DEFAULT_CONFIG"]
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/ars/ars.py
Python
# Code in this file is copied and adapted from # https://github.com/openai/evolution-strategies-starter and from # https://github.com/modestyachts/ARS from collections import namedtuple import logging import numpy as np import time import ray from ray.rllib.agents import Trainer, with_common_config from ray.rllib.agents.ars import optimizers from ray.rllib.agents.ars import policies from ray.rllib.agents.ars import utils from ray.rllib.policy.sample_batch import DEFAULT_POLICY_ID from ray.rllib.utils.annotations import override from ray.rllib.utils.memory import ray_get_and_free from ray.rllib.utils import FilterManager logger = logging.getLogger(__name__) Result = namedtuple("Result", [ "noise_indices", "noisy_returns", "sign_noisy_returns", "noisy_lengths", "eval_returns", "eval_lengths" ]) # yapf: disable # __sphinx_doc_begin__ DEFAULT_CONFIG = with_common_config({ "noise_stdev": 0.02, # std deviation of parameter noise "num_rollouts": 32, # number of perturbs to try "rollouts_used": 32, # number of perturbs to keep in gradient estimate "num_workers": 2, "sgd_stepsize": 0.01, # sgd step-size "observation_filter": "MeanStdFilter", "noise_size": 250000000, "eval_prob": 0.03, # probability of evaluating the parameter rewards "report_length": 10, # how many of the last rewards we average over "offset": 0, }) # __sphinx_doc_end__ # yapf: enable @ray.remote def create_shared_noise(count): """Create a large array of noise to be shared by all workers.""" seed = 123 noise = np.random.RandomState(seed).randn(count).astype(np.float32) return noise class SharedNoiseTable: def __init__(self, noise): self.noise = noise assert self.noise.dtype == np.float32 def get(self, i, dim): return self.noise[i:i + dim] def sample_index(self, dim): return np.random.randint(0, len(self.noise) - dim + 1) def get_delta(self, dim): idx = self.sample_index(dim) return idx, self.get(idx, dim) @ray.remote class Worker: def __init__(self, config, env_creator, noise, min_task_runtime=0.2): self.min_task_runtime = min_task_runtime self.config = config self.noise = SharedNoiseTable(noise) self.env = env_creator(config["env_config"]) from ray.rllib import models self.preprocessor = models.ModelCatalog.get_preprocessor(self.env) self.sess = utils.make_session(single_threaded=True) self.policy = policies.GenericPolicy( self.sess, self.env.action_space, self.env.observation_space, self.preprocessor, config["observation_filter"], config["model"]) @property def filters(self): return {DEFAULT_POLICY_ID: self.policy.get_filter()} def sync_filters(self, new_filters): for k in self.filters: self.filters[k].sync(new_filters[k]) def get_filters(self, flush_after=False): return_filters = {} for k, f in self.filters.items(): return_filters[k] = f.as_serializable() if flush_after: f.clear_buffer() return return_filters def rollout(self, timestep_limit, add_noise=False): rollout_rewards, rollout_length = policies.rollout( self.policy, self.env, timestep_limit=timestep_limit, add_noise=add_noise, offset=self.config["offset"]) return rollout_rewards, rollout_length def do_rollouts(self, params, timestep_limit=None): # Set the network weights. self.policy.set_weights(params) noise_indices, returns, sign_returns, lengths = [], [], [], [] eval_returns, eval_lengths = [], [] # Perform some rollouts with noise. while (len(noise_indices) == 0): if np.random.uniform() < self.config["eval_prob"]: # Do an evaluation run with no perturbation. self.policy.set_weights(params) rewards, length = self.rollout(timestep_limit, add_noise=False) eval_returns.append(rewards.sum()) eval_lengths.append(length) else: # Do a regular run with parameter perturbations. noise_index = self.noise.sample_index(self.policy.num_params) perturbation = self.config["noise_stdev"] * self.noise.get( noise_index, self.policy.num_params) # These two sampling steps could be done in parallel on # different actors letting us update twice as frequently. self.policy.set_weights(params + perturbation) rewards_pos, lengths_pos = self.rollout(timestep_limit) self.policy.set_weights(params - perturbation) rewards_neg, lengths_neg = self.rollout(timestep_limit) noise_indices.append(noise_index) returns.append([rewards_pos.sum(), rewards_neg.sum()]) sign_returns.append( [np.sign(rewards_pos).sum(), np.sign(rewards_neg).sum()]) lengths.append([lengths_pos, lengths_neg]) return Result( noise_indices=noise_indices, noisy_returns=returns, sign_noisy_returns=sign_returns, noisy_lengths=lengths, eval_returns=eval_returns, eval_lengths=eval_lengths) class ARSTrainer(Trainer): """Large-scale implementation of Augmented Random Search in Ray.""" _name = "ARS" _default_config = DEFAULT_CONFIG @override(Trainer) def _init(self, config, env_creator): # PyTorch check. if config["use_pytorch"]: raise ValueError( "ARS does not support PyTorch yet! Use tf instead." ) env = env_creator(config["env_config"]) from ray.rllib import models preprocessor = models.ModelCatalog.get_preprocessor(env) self.sess = utils.make_session(single_threaded=False) self.policy = policies.GenericPolicy( self.sess, env.action_space, env.observation_space, preprocessor, config["observation_filter"], config["model"]) self.optimizer = optimizers.SGD(self.policy, config["sgd_stepsize"]) self.rollouts_used = config["rollouts_used"] self.num_rollouts = config["num_rollouts"] self.report_length = config["report_length"] # Create the shared noise table. logger.info("Creating shared noise table.") noise_id = create_shared_noise.remote(config["noise_size"]) self.noise = SharedNoiseTable(ray.get(noise_id)) # Create the actors. logger.info("Creating actors.") self.workers = [ Worker.remote(config, env_creator, noise_id) for _ in range(config["num_workers"]) ] self.episodes_so_far = 0 self.reward_list = [] self.tstart = time.time() @override(Trainer) def _train(self): config = self.config theta = self.policy.get_weights() assert theta.dtype == np.float32 # Put the current policy weights in the object store. theta_id = ray.put(theta) # Use the actors to do rollouts, note that we pass in the ID of the # policy weights. results, num_episodes, num_timesteps = self._collect_results( theta_id, config["num_rollouts"]) all_noise_indices = [] all_training_returns = [] all_training_lengths = [] all_eval_returns = [] all_eval_lengths = [] # Loop over the results. for result in results: all_eval_returns += result.eval_returns all_eval_lengths += result.eval_lengths all_noise_indices += result.noise_indices all_training_returns += result.noisy_returns all_training_lengths += result.noisy_lengths assert len(all_eval_returns) == len(all_eval_lengths) assert (len(all_noise_indices) == len(all_training_returns) == len(all_training_lengths)) self.episodes_so_far += num_episodes # Assemble the results. eval_returns = np.array(all_eval_returns) eval_lengths = np.array(all_eval_lengths) noise_indices = np.array(all_noise_indices) noisy_returns = np.array(all_training_returns) noisy_lengths = np.array(all_training_lengths) # keep only the best returns # select top performing directions if rollouts_used < num_rollouts max_rewards = np.max(noisy_returns, axis=1) if self.rollouts_used > self.num_rollouts: self.rollouts_used = self.num_rollouts percentile = 100 * (1 - (self.rollouts_used / self.num_rollouts)) idx = np.arange(max_rewards.size)[ max_rewards >= np.percentile(max_rewards, percentile)] noise_idx = noise_indices[idx] noisy_returns = noisy_returns[idx, :] # Compute and take a step. g, count = utils.batched_weighted_sum( noisy_returns[:, 0] - noisy_returns[:, 1], (self.noise.get(index, self.policy.num_params) for index in noise_idx), batch_size=min(500, noisy_returns[:, 0].size)) g /= noise_idx.size # scale the returns by their standard deviation if not np.isclose(np.std(noisy_returns), 0.0): g /= np.std(noisy_returns) assert (g.shape == (self.policy.num_params, ) and g.dtype == np.float32) # Compute the new weights theta. theta, update_ratio = self.optimizer.update(-g) # Set the new weights in the local copy of the policy. self.policy.set_weights(theta) # update the reward list if len(all_eval_returns) > 0: self.reward_list.append(eval_returns.mean()) # Now sync the filters FilterManager.synchronize({ DEFAULT_POLICY_ID: self.policy.get_filter() }, self.workers) info = { "weights_norm": np.square(theta).sum(), "weights_std": np.std(theta), "grad_norm": np.square(g).sum(), "update_ratio": update_ratio, "episodes_this_iter": noisy_lengths.size, "episodes_so_far": self.episodes_so_far, } result = dict( episode_reward_mean=np.mean( self.reward_list[-self.report_length:]), episode_len_mean=eval_lengths.mean(), timesteps_this_iter=noisy_lengths.sum(), info=info) return result @override(Trainer) def _stop(self): # workaround for https://github.com/ray-project/ray/issues/1516 for w in self.workers: w.__ray_terminate__.remote() @override(Trainer) def compute_action(self, observation): return self.policy.compute(observation, update=True)[0] def _collect_results(self, theta_id, min_episodes): num_episodes, num_timesteps = 0, 0 results = [] while num_episodes < min_episodes: logger.debug( "Collected {} episodes {} timesteps so far this iter".format( num_episodes, num_timesteps)) rollout_ids = [ worker.do_rollouts.remote(theta_id) for worker in self.workers ] # Get the results of the rollouts. for result in ray_get_and_free(rollout_ids): results.append(result) # Update the number of episodes and the number of timesteps # keeping in mind that result.noisy_lengths is a list of lists, # where the inner lists have length 2. num_episodes += sum(len(pair) for pair in result.noisy_lengths) num_timesteps += sum( sum(pair) for pair in result.noisy_lengths) return results, num_episodes, num_timesteps def __getstate__(self): return { "weights": self.policy.get_weights(), "filter": self.policy.get_filter(), "episodes_so_far": self.episodes_so_far, } def __setstate__(self, state): self.episodes_so_far = state["episodes_so_far"] self.policy.set_weights(state["weights"]) self.policy.set_filter(state["filter"]) FilterManager.synchronize({ DEFAULT_POLICY_ID: self.policy.get_filter() }, self.workers)
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/ars/optimizers.py
Python
# Code in this file is copied and adapted from # https://github.com/openai/evolution-strategies-starter. import numpy as np class Optimizer: def __init__(self, policy): self.policy = policy self.dim = policy.num_params self.t = 0 def update(self, globalg): self.t += 1 step = self._compute_step(globalg) theta = self.policy.get_weights() ratio = np.linalg.norm(step) / np.linalg.norm(theta) return theta + step, ratio def _compute_step(self, globalg): raise NotImplementedError class SGD(Optimizer): def __init__(self, policy, stepsize, momentum=0.0): Optimizer.__init__(self, policy) self.v = np.zeros(self.dim, dtype=np.float32) self.stepsize, self.momentum = stepsize, momentum def _compute_step(self, globalg): self.v = self.momentum * self.v + (1. - self.momentum) * globalg step = -self.stepsize * self.v return step class Adam(Optimizer): def __init__(self, policy, stepsize, beta1=0.9, beta2=0.999, epsilon=1e-08): Optimizer.__init__(self, policy) self.stepsize = stepsize self.beta1 = beta1 self.beta2 = beta2 self.epsilon = epsilon self.m = np.zeros(self.dim, dtype=np.float32) self.v = np.zeros(self.dim, dtype=np.float32) def _compute_step(self, globalg): a = self.stepsize * (np.sqrt(1 - self.beta2**self.t) / (1 - self.beta1**self.t)) self.m = self.beta1 * self.m + (1 - self.beta1) * globalg self.v = self.beta2 * self.v + (1 - self.beta2) * (globalg * globalg) step = -a * self.m / (np.sqrt(self.v) + self.epsilon) return step
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/ars/policies.py
Python
# Code in this file is copied and adapted from # https://github.com/openai/evolution-strategies-starter. import gym import numpy as np import ray import ray.experimental.tf_utils from ray.rllib.evaluation.sampler import _unbatch_tuple_actions from ray.rllib.utils.filter import get_filter from ray.rllib.models import ModelCatalog from ray.rllib.utils import try_import_tf tf = try_import_tf() def rollout(policy, env, timestep_limit=None, add_noise=False, offset=0): """Do a rollout. If add_noise is True, the rollout will take noisy actions with noise drawn from that stream. Otherwise, no action noise will be added. Parameters ---------- policy: tf object policy from which to draw actions env: GymEnv environment from which to draw rewards, done, and next state timestep_limit: int, optional steps after which to end the rollout add_noise: bool, optional indicates whether exploratory action noise should be added offset: int, optional value to subtract from the reward. For example, survival bonus from humanoid """ env_timestep_limit = env.spec.max_episode_steps timestep_limit = (env_timestep_limit if timestep_limit is None else min( timestep_limit, env_timestep_limit)) rews = [] t = 0 observation = env.reset() for _ in range(timestep_limit or 999999): ac = policy.compute(observation, add_noise=add_noise, update=True)[0] observation, rew, done, _ = env.step(ac) rew -= np.abs(offset) rews.append(rew) t += 1 if done: break rews = np.array(rews, dtype=np.float32) return rews, t class GenericPolicy: def __init__(self, sess, action_space, obs_space, preprocessor, observation_filter, model_config, action_noise_std=0.0): self.sess = sess self.action_space = action_space self.action_noise_std = action_noise_std self.preprocessor = preprocessor self.observation_filter = get_filter(observation_filter, self.preprocessor.shape) self.inputs = tf.placeholder(tf.float32, [None] + list(self.preprocessor.shape)) # Policy network. dist_class, dist_dim = ModelCatalog.get_action_dist( action_space, model_config, dist_type="deterministic") model = ModelCatalog.get_model({ "obs": self.inputs }, obs_space, action_space, dist_dim, model_config) dist = dist_class(model.outputs, model) self.sampler = dist.sample() self.variables = ray.experimental.tf_utils.TensorFlowVariables( model.outputs, self.sess) self.num_params = sum( np.prod(variable.shape.as_list()) for _, variable in self.variables.variables.items()) self.sess.run(tf.global_variables_initializer()) def compute(self, observation, add_noise=False, update=True): observation = self.preprocessor.transform(observation) observation = self.observation_filter(observation[None], update=update) action = self.sess.run( self.sampler, feed_dict={self.inputs: observation}) action = _unbatch_tuple_actions(action) if add_noise and isinstance(self.action_space, gym.spaces.Box): action += np.random.randn(*action.shape) * self.action_noise_std return action def set_weights(self, x): self.variables.set_flat(x) def set_filter(self, obs_filter): self.observation_filter = obs_filter def get_filter(self): return self.observation_filter def get_weights(self): return self.variables.get_flat()
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/ars/utils.py
Python
# Code in this file is copied and adapted from # https://github.com/openai/evolution-strategies-starter. import numpy as np from ray.rllib.utils import try_import_tf tf = try_import_tf() def compute_ranks(x): """Returns ranks in [0, len(x)) Note: This is different from scipy.stats.rankdata, which returns ranks in [1, len(x)]. """ assert x.ndim == 1 ranks = np.empty(len(x), dtype=int) ranks[x.argsort()] = np.arange(len(x)) return ranks def compute_centered_ranks(x): y = compute_ranks(x.ravel()).reshape(x.shape).astype(np.float32) y /= (x.size - 1) y -= 0.5 return y def make_session(single_threaded): if not single_threaded: return tf.Session() return tf.Session( config=tf.ConfigProto( inter_op_parallelism_threads=1, intra_op_parallelism_threads=1)) def itergroups(items, group_size): assert group_size >= 1 group = [] for x in items: group.append(x) if len(group) == group_size: yield tuple(group) del group[:] if group: yield tuple(group) def batched_weighted_sum(weights, vecs, batch_size): total = 0 num_items_summed = 0 for batch_weights, batch_vecs in zip( itergroups(weights, batch_size), itergroups(vecs, batch_size)): assert len(batch_weights) == len(batch_vecs) <= batch_size total += np.dot( np.asarray(batch_weights, dtype=np.float32), np.asarray(batch_vecs, dtype=np.float32)) num_items_summed += len(batch_weights) return total, num_items_summed
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/ddpg/__init__.py
Python
from ray.rllib.agents.ddpg.apex import ApexDDPGTrainer from ray.rllib.agents.ddpg.ddpg import DDPGTrainer, DEFAULT_CONFIG from ray.rllib.agents.ddpg.td3 import TD3Trainer from ray.rllib.utils import renamed_agent ApexDDPGAgent = renamed_agent(ApexDDPGTrainer) DDPGAgent = renamed_agent(DDPGTrainer) __all__ = [ "DDPGAgent", "ApexDDPGAgent", "DDPGTrainer", "ApexDDPGTrainer", "TD3Trainer", "DEFAULT_CONFIG" ]
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/ddpg/apex.py
Python
from ray.rllib.agents.dqn.apex import APEX_TRAINER_PROPERTIES from ray.rllib.agents.ddpg.ddpg import DDPGTrainer, \ DEFAULT_CONFIG as DDPG_CONFIG from ray.rllib.utils import merge_dicts APEX_DDPG_DEFAULT_CONFIG = merge_dicts( DDPG_CONFIG, # see also the options in ddpg.py, which are also supported { "optimizer": merge_dicts( DDPG_CONFIG["optimizer"], { "max_weight_sync_delay": 400, "num_replay_buffer_shards": 4, "debug": False }), "n_step": 3, "num_gpus": 0, "num_workers": 32, "buffer_size": 2000000, "learning_starts": 50000, "train_batch_size": 512, "sample_batch_size": 50, "target_network_update_freq": 500000, "timesteps_per_iteration": 25000, "per_worker_exploration": True, "worker_side_prioritization": True, "min_iter_time_s": 30, }, ) ApexDDPGTrainer = DDPGTrainer.with_updates( name="APEX_DDPG", default_config=APEX_DDPG_DEFAULT_CONFIG, **APEX_TRAINER_PROPERTIES)
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/ddpg/ddpg.py
Python
from ray.rllib.agents.trainer import with_common_config from ray.rllib.agents.dqn.dqn import GenericOffPolicyTrainer, \ update_worker_explorations from ray.rllib.agents.ddpg.ddpg_policy import DDPGTFPolicy from ray.rllib.utils.schedules import ConstantSchedule, LinearSchedule # yapf: disable # __sphinx_doc_begin__ DEFAULT_CONFIG = with_common_config({ # === Twin Delayed DDPG (TD3) and Soft Actor-Critic (SAC) tricks === # TD3: https://spinningup.openai.com/en/latest/algorithms/td3.html # In addition to settings below, you can use "exploration_noise_type" and # "exploration_gauss_act_noise" to get IID Gaussian exploration noise # instead of OU exploration noise. # twin Q-net "twin_q": False, # delayed policy update "policy_delay": 1, # target policy smoothing # (this also replaces OU exploration noise with IID Gaussian exploration # noise, for now) "smooth_target_policy": False, # gaussian stddev of target action noise for smoothing "target_noise": 0.2, # target noise limit (bound) "target_noise_clip": 0.5, # === Evaluation === # Evaluate with epsilon=0 every `evaluation_interval` training iterations. # The evaluation stats will be reported under the "evaluation" metric key. # Note that evaluation is currently not parallelized, and that for Ape-X # metrics are already only reported for the lowest epsilon workers. "evaluation_interval": None, # Number of episodes to run per evaluation period. "evaluation_num_episodes": 10, # === Model === # Apply a state preprocessor with spec given by the "model" config option # (like other RL algorithms). This is mostly useful if you have a weird # observation shape, like an image. Disabled by default. "use_state_preprocessor": False, # Postprocess the policy network model output with these hidden layers. If # use_state_preprocessor is False, then these will be the *only* hidden # layers in the network. "actor_hiddens": [400, 300], # Hidden layers activation of the postprocessing stage of the policy # network "actor_hidden_activation": "relu", # Postprocess the critic network model output with these hidden layers; # again, if use_state_preprocessor is True, then the state will be # preprocessed by the model specified with the "model" config option first. "critic_hiddens": [400, 300], # Hidden layers activation of the postprocessing state of the critic. "critic_hidden_activation": "relu", # N-step Q learning "n_step": 1, # === Exploration === # Turns on annealing schedule for exploration noise. Exploration is # annealed from 1.0 to exploration_final_eps over schedule_max_timesteps # scaled by exploration_fraction. Original DDPG and TD3 papers do not # anneal noise, so this is False by default. "exploration_should_anneal": False, # Max num timesteps for annealing schedules. "schedule_max_timesteps": 100000, # Number of env steps to optimize for before returning "timesteps_per_iteration": 1000, # Fraction of entire training period over which the exploration rate is # annealed "exploration_fraction": 0.1, # Final scaling multiplier for action noise (initial is 1.0) "exploration_final_scale": 0.02, # valid values: "ou" (time-correlated, like original DDPG paper), # "gaussian" (IID, like TD3 paper) "exploration_noise_type": "ou", # OU-noise scale; this can be used to scale down magnitude of OU noise # before adding to actions (requires "exploration_noise_type" to be "ou") "exploration_ou_noise_scale": 0.1, # theta for OU "exploration_ou_theta": 0.15, # sigma for OU "exploration_ou_sigma": 0.2, # gaussian stddev of act noise for exploration (requires # "exploration_noise_type" to be "gaussian") "exploration_gaussian_sigma": 0.1, # If True parameter space noise will be used for exploration # See https://blog.openai.com/better-exploration-with-parameter-noise/ "parameter_noise": False, # Until this many timesteps have elapsed, the agent's policy will be # ignored & it will instead take uniform random actions. Can be used in # conjunction with learning_starts (which controls when the first # optimization step happens) to decrease dependence of exploration & # optimization on initial policy parameters. Note that this will be # disabled when the action noise scale is set to 0 (e.g during evaluation). "pure_exploration_steps": 1000, # Extra configuration that disables exploration. "evaluation_config": { "exploration_fraction": 0, "exploration_final_eps": 0, }, # === Replay buffer === # Size of the replay buffer. Note that if async_updates is set, then # each worker will have a replay buffer of this size. "buffer_size": 50000, # If True prioritized replay buffer will be used. "prioritized_replay": True, # Alpha parameter for prioritized replay buffer. "prioritized_replay_alpha": 0.6, # Beta parameter for sampling from prioritized replay buffer. "prioritized_replay_beta": 0.4, # Fraction of entire training period over which the beta parameter is # annealed "beta_annealing_fraction": 0.2, # Final value of beta "final_prioritized_replay_beta": 0.4, # Epsilon to add to the TD errors when updating priorities. "prioritized_replay_eps": 1e-6, # Whether to LZ4 compress observations "compress_observations": False, # === Optimization === # Learning rate for the critic (Q-function) optimizer. "critic_lr": 1e-3, # Learning rate for the actor (policy) optimizer. "actor_lr": 1e-3, # Update the target network every `target_network_update_freq` steps. "target_network_update_freq": 0, # Update the target by \tau * policy + (1-\tau) * target_policy "tau": 0.002, # If True, use huber loss instead of squared loss for critic network # Conventionally, no need to clip gradients if using a huber loss "use_huber": False, # Threshold of a huber loss "huber_threshold": 1.0, # Weights for L2 regularization "l2_reg": 1e-6, # If not None, clip gradients during optimization at this value "grad_norm_clipping": None, # How many steps of the model to sample before learning starts. "learning_starts": 1500, # Update the replay buffer with this many samples at once. Note that this # setting applies per-worker if num_workers > 1. "sample_batch_size": 1, # Size of a batched sampled from replay buffer for training. Note that # if async_updates is set, then each worker returns gradients for a # batch of this size. "train_batch_size": 256, # === Parallelism === # Number of workers for collecting samples with. This only makes sense # to increase if your environment is particularly slow to sample, or if # you're using the Async or Ape-X optimizers. "num_workers": 0, # Whether to use a distribution of epsilons across workers for exploration. "per_worker_exploration": False, # Whether to compute priorities on workers. "worker_side_prioritization": False, # Prevent iterations from going lower than this time span "min_iter_time_s": 1, }) # __sphinx_doc_end__ # yapf: enable def make_exploration_schedule(config, worker_index): # Modification of DQN's schedule to take into account # `exploration_ou_noise_scale` if config["per_worker_exploration"]: assert config["num_workers"] > 1, "This requires multiple workers" if worker_index >= 0: # FIXME: what do magic constants mean? (0.4, 7) max_index = float(config["num_workers"] - 1) exponent = 1 + worker_index / max_index * 7 return ConstantSchedule(0.4**exponent) else: # local ev should have zero exploration so that eval rollouts # run properly return ConstantSchedule(0.0) elif config["exploration_should_anneal"]: return LinearSchedule( schedule_timesteps=int(config["exploration_fraction"] * config["schedule_max_timesteps"]), initial_p=1.0, final_p=config["exploration_final_scale"]) else: # *always* add exploration noise return ConstantSchedule(1.0) def setup_ddpg_exploration(trainer): trainer.exploration0 = make_exploration_schedule(trainer.config, -1) trainer.explorations = [ make_exploration_schedule(trainer.config, i) for i in range(trainer.config["num_workers"]) ] def add_pure_exploration_phase(trainer): global_timestep = trainer.optimizer.num_steps_sampled pure_expl_steps = trainer.config["pure_exploration_steps"] if pure_expl_steps: # tell workers whether they should do pure exploration only_explore = global_timestep < pure_expl_steps trainer.workers.local_worker().foreach_trainable_policy( lambda p, _: p.set_pure_exploration_phase(only_explore)) for e in trainer.workers.remote_workers(): e.foreach_trainable_policy.remote( lambda p, _: p.set_pure_exploration_phase(only_explore)) update_worker_explorations(trainer) def validate_config(config): # PyTorch check. if config["use_pytorch"]: raise ValueError("DDPG does not support PyTorch yet! Use tf instead.") DDPGTrainer = GenericOffPolicyTrainer.with_updates( name="DDPG", default_config=DEFAULT_CONFIG, default_policy=DDPGTFPolicy, validate_config=validate_config, before_init=setup_ddpg_exploration, before_train_step=add_pure_exploration_phase)
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/ddpg/ddpg_policy.py
Python
from gym.spaces import Box import numpy as np import ray import ray.experimental.tf_utils from ray.rllib.agents.dqn.dqn_policy import _postprocess_dqn from ray.rllib.policy.sample_batch import SampleBatch from ray.rllib.evaluation.metrics import LEARNER_STATS_KEY from ray.rllib.models import ModelCatalog from ray.rllib.utils.annotations import override from ray.rllib.utils.error import UnsupportedSpaceException from ray.rllib.policy.policy import Policy from ray.rllib.policy.tf_policy import TFPolicy from ray.rllib.utils import try_import_tf from ray.rllib.utils.tf_ops import huber_loss, minimize_and_clip, scope_vars tf = try_import_tf() ACTION_SCOPE = "action" POLICY_SCOPE = "policy" POLICY_TARGET_SCOPE = "target_policy" Q_SCOPE = "critic" Q_TARGET_SCOPE = "target_critic" TWIN_Q_SCOPE = "twin_critic" TWIN_Q_TARGET_SCOPE = "twin_target_critic" # Importance sampling weights for prioritized replay PRIO_WEIGHTS = "weights" class DDPGPostprocessing: """Implements n-step learning and param noise adjustments.""" @override(Policy) def postprocess_trajectory(self, sample_batch, other_agent_batches=None, episode=None): if self.config["parameter_noise"]: # adjust the sigma of parameter space noise states, noisy_actions = [ list(x) for x in sample_batch.columns( [SampleBatch.CUR_OBS, SampleBatch.ACTIONS]) ] self.sess.run(self.remove_noise_op) clean_actions = self.sess.run( self.output_actions, feed_dict={ self.cur_observations: states, self.stochastic: False, self.noise_scale: .0, self.pure_exploration_phase: False, }) distance_in_action_space = np.sqrt( np.mean(np.square(clean_actions - noisy_actions))) self.pi_distance = distance_in_action_space if distance_in_action_space < \ self.config["exploration_ou_sigma"] * self.cur_noise_scale: # multiplying the sampled OU noise by noise scale is # equivalent to multiplying the sigma of OU by noise scale self.parameter_noise_sigma_val *= 1.01 else: self.parameter_noise_sigma_val /= 1.01 self.parameter_noise_sigma.load( self.parameter_noise_sigma_val, session=self.sess) return _postprocess_dqn(self, sample_batch) class DDPGTFPolicy(DDPGPostprocessing, TFPolicy): def __init__(self, observation_space, action_space, config): config = dict(ray.rllib.agents.ddpg.ddpg.DEFAULT_CONFIG, **config) if not isinstance(action_space, Box): raise UnsupportedSpaceException( "Action space {} is not supported for DDPG.".format( action_space)) if len(action_space.shape) > 1: raise UnsupportedSpaceException( "Action space has multiple dimensions " "{}. ".format(action_space.shape) + "Consider reshaping this into a single dimension, " "using a Tuple action space, or the multi-agent API.") self.config = config self.cur_noise_scale = 1.0 self.cur_pure_exploration_phase = False self.dim_actions = action_space.shape[0] self.low_action = action_space.low self.high_action = action_space.high # create global step for counting the number of update operations self.global_step = tf.train.get_or_create_global_step() # use separate optimizers for actor & critic self._actor_optimizer = tf.train.AdamOptimizer( learning_rate=self.config["actor_lr"]) self._critic_optimizer = tf.train.AdamOptimizer( learning_rate=self.config["critic_lr"]) # Action inputs self.stochastic = tf.placeholder(tf.bool, (), name="stochastic") self.noise_scale = tf.placeholder(tf.float32, (), name="noise_scale") self.pure_exploration_phase = tf.placeholder( tf.bool, (), name="pure_exploration_phase") self.cur_observations = tf.placeholder( tf.float32, shape=(None, ) + observation_space.shape, name="cur_obs") with tf.variable_scope(POLICY_SCOPE) as scope: policy_out, self.policy_model = self._build_policy_network( self.cur_observations, observation_space, action_space) self.policy_vars = scope_vars(scope.name) # Noise vars for P network except for layer normalization vars if self.config["parameter_noise"]: self._build_parameter_noise([ var for var in self.policy_vars if "LayerNorm" not in var.name ]) # Action outputs with tf.variable_scope(ACTION_SCOPE): self.output_actions = self._add_exploration_noise( policy_out, self.stochastic, self.noise_scale, self.pure_exploration_phase, action_space) if self.config["smooth_target_policy"]: self.reset_noise_op = tf.no_op() else: with tf.variable_scope(ACTION_SCOPE, reuse=True): exploration_sample = tf.get_variable(name="ornstein_uhlenbeck") self.reset_noise_op = tf.assign(exploration_sample, self.dim_actions * [.0]) # Replay inputs self.obs_t = tf.placeholder( tf.float32, shape=(None, ) + observation_space.shape, name="observation") self.act_t = tf.placeholder( tf.float32, shape=(None, ) + action_space.shape, name="action") self.rew_t = tf.placeholder(tf.float32, [None], name="reward") self.obs_tp1 = tf.placeholder( tf.float32, shape=(None, ) + observation_space.shape) self.done_mask = tf.placeholder(tf.float32, [None], name="done") self.importance_weights = tf.placeholder( tf.float32, [None], name="weight") # policy network evaluation with tf.variable_scope(POLICY_SCOPE, reuse=True) as scope: prev_update_ops = set(tf.get_collection(tf.GraphKeys.UPDATE_OPS)) self.policy_t, _ = self._build_policy_network( self.obs_t, observation_space, action_space) policy_batchnorm_update_ops = list( set(tf.get_collection(tf.GraphKeys.UPDATE_OPS)) - prev_update_ops) # target policy network evaluation with tf.variable_scope(POLICY_TARGET_SCOPE) as scope: policy_tp1, _ = self._build_policy_network( self.obs_tp1, observation_space, action_space) target_policy_vars = scope_vars(scope.name) # Action outputs with tf.variable_scope(ACTION_SCOPE, reuse=True): if config["smooth_target_policy"]: target_noise_clip = self.config["target_noise_clip"] clipped_normal_sample = tf.clip_by_value( tf.random_normal( tf.shape(policy_tp1), stddev=self.config["target_noise"]), -target_noise_clip, target_noise_clip) policy_tp1_smoothed = tf.clip_by_value( policy_tp1 + clipped_normal_sample, action_space.low * tf.ones_like(policy_tp1), action_space.high * tf.ones_like(policy_tp1)) else: # no smoothing, just use deterministic actions policy_tp1_smoothed = policy_tp1 # q network evaluation prev_update_ops = set(tf.get_collection(tf.GraphKeys.UPDATE_OPS)) with tf.variable_scope(Q_SCOPE) as scope: # Q-values for given actions & observations in given current q_t, self.q_model = self._build_q_network( self.obs_t, observation_space, action_space, self.act_t) self.q_func_vars = scope_vars(scope.name) self.stats = { "mean_q": tf.reduce_mean(q_t), "max_q": tf.reduce_max(q_t), "min_q": tf.reduce_min(q_t), } with tf.variable_scope(Q_SCOPE, reuse=True): # Q-values for current policy (no noise) in given current state q_t_det_policy, _ = self._build_q_network( self.obs_t, observation_space, action_space, self.policy_t) if self.config["twin_q"]: with tf.variable_scope(TWIN_Q_SCOPE) as scope: twin_q_t, self.twin_q_model = self._build_q_network( self.obs_t, observation_space, action_space, self.act_t) self.twin_q_func_vars = scope_vars(scope.name) q_batchnorm_update_ops = list( set(tf.get_collection(tf.GraphKeys.UPDATE_OPS)) - prev_update_ops) # target q network evaluation with tf.variable_scope(Q_TARGET_SCOPE) as scope: q_tp1, _ = self._build_q_network(self.obs_tp1, observation_space, action_space, policy_tp1_smoothed) target_q_func_vars = scope_vars(scope.name) if self.config["twin_q"]: with tf.variable_scope(TWIN_Q_TARGET_SCOPE) as scope: twin_q_tp1, _ = self._build_q_network( self.obs_tp1, observation_space, action_space, policy_tp1_smoothed) twin_target_q_func_vars = scope_vars(scope.name) if self.config["twin_q"]: self.critic_loss, self.actor_loss, self.td_error \ = self._build_actor_critic_loss( q_t, q_tp1, q_t_det_policy, twin_q_t=twin_q_t, twin_q_tp1=twin_q_tp1) else: self.critic_loss, self.actor_loss, self.td_error \ = self._build_actor_critic_loss( q_t, q_tp1, q_t_det_policy) if config["l2_reg"] is not None: for var in self.policy_vars: if "bias" not in var.name: self.actor_loss += (config["l2_reg"] * tf.nn.l2_loss(var)) for var in self.q_func_vars: if "bias" not in var.name: self.critic_loss += (config["l2_reg"] * tf.nn.l2_loss(var)) if self.config["twin_q"]: for var in self.twin_q_func_vars: if "bias" not in var.name: self.critic_loss += ( config["l2_reg"] * tf.nn.l2_loss(var)) # update_target_fn will be called periodically to copy Q network to # target Q network self.tau_value = config.get("tau") self.tau = tf.placeholder(tf.float32, (), name="tau") update_target_expr = [] for var, var_target in zip( sorted(self.q_func_vars, key=lambda v: v.name), sorted(target_q_func_vars, key=lambda v: v.name)): update_target_expr.append( var_target.assign(self.tau * var + (1.0 - self.tau) * var_target)) if self.config["twin_q"]: for var, var_target in zip( sorted(self.twin_q_func_vars, key=lambda v: v.name), sorted(twin_target_q_func_vars, key=lambda v: v.name)): update_target_expr.append( var_target.assign(self.tau * var + (1.0 - self.tau) * var_target)) for var, var_target in zip( sorted(self.policy_vars, key=lambda v: v.name), sorted(target_policy_vars, key=lambda v: v.name)): update_target_expr.append( var_target.assign(self.tau * var + (1.0 - self.tau) * var_target)) self.update_target_expr = tf.group(*update_target_expr) self.sess = tf.get_default_session() self.loss_inputs = [ (SampleBatch.CUR_OBS, self.obs_t), (SampleBatch.ACTIONS, self.act_t), (SampleBatch.REWARDS, self.rew_t), (SampleBatch.NEXT_OBS, self.obs_tp1), (SampleBatch.DONES, self.done_mask), (PRIO_WEIGHTS, self.importance_weights), ] input_dict = dict(self.loss_inputs) if self.config["use_state_preprocessor"]: # Model self-supervised losses self.actor_loss = self.policy_model.custom_loss( self.actor_loss, input_dict) self.critic_loss = self.q_model.custom_loss( self.critic_loss, input_dict) if self.config["twin_q"]: self.critic_loss = self.twin_q_model.custom_loss( self.critic_loss, input_dict) TFPolicy.__init__( self, observation_space, action_space, self.config, self.sess, obs_input=self.cur_observations, action_sampler=self.output_actions, loss=self.actor_loss + self.critic_loss, loss_inputs=self.loss_inputs, update_ops=q_batchnorm_update_ops + policy_batchnorm_update_ops) self.sess.run(tf.global_variables_initializer()) # Note that this encompasses both the policy and Q-value networks and # their corresponding target networks self.variables = ray.experimental.tf_utils.TensorFlowVariables( tf.group(q_t_det_policy, q_tp1), self.sess) # Hard initial update self.update_target(tau=1.0) @override(TFPolicy) def optimizer(self): # we don't use this because we have two separate optimisers return None @override(TFPolicy) def build_apply_op(self, optimizer, grads_and_vars): # for policy gradient, update policy net one time v.s. # update critic net `policy_delay` time(s) should_apply_actor_opt = tf.equal( tf.mod(self.global_step, self.config["policy_delay"]), 0) def make_apply_op(): return self._actor_optimizer.apply_gradients( self._actor_grads_and_vars) actor_op = tf.cond( should_apply_actor_opt, true_fn=make_apply_op, false_fn=lambda: tf.no_op()) critic_op = self._critic_optimizer.apply_gradients( self._critic_grads_and_vars) # increment global step & apply ops with tf.control_dependencies([tf.assign_add(self.global_step, 1)]): return tf.group(actor_op, critic_op) @override(TFPolicy) def gradients(self, optimizer, loss): if self.config["grad_norm_clipping"] is not None: actor_grads_and_vars = minimize_and_clip( self._actor_optimizer, self.actor_loss, var_list=self.policy_vars, clip_val=self.config["grad_norm_clipping"]) critic_grads_and_vars = minimize_and_clip( self._critic_optimizer, self.critic_loss, var_list=self.q_func_vars + self.twin_q_func_vars if self.config["twin_q"] else self.q_func_vars, clip_val=self.config["grad_norm_clipping"]) else: actor_grads_and_vars = self._actor_optimizer.compute_gradients( self.actor_loss, var_list=self.policy_vars) if self.config["twin_q"]: critic_vars = self.q_func_vars + self.twin_q_func_vars else: critic_vars = self.q_func_vars critic_grads_and_vars = self._critic_optimizer.compute_gradients( self.critic_loss, var_list=critic_vars) # save these for later use in build_apply_op self._actor_grads_and_vars = [(g, v) for (g, v) in actor_grads_and_vars if g is not None] self._critic_grads_and_vars = [(g, v) for (g, v) in critic_grads_and_vars if g is not None] grads_and_vars = self._actor_grads_and_vars \ + self._critic_grads_and_vars return grads_and_vars @override(TFPolicy) def extra_compute_action_feed_dict(self): return { # FIXME: what about turning off exploration? Isn't that a good # idea? self.stochastic: True, self.noise_scale: self.cur_noise_scale, self.pure_exploration_phase: self.cur_pure_exploration_phase, } @override(TFPolicy) def extra_compute_grad_fetches(self): return { "td_error": self.td_error, LEARNER_STATS_KEY: self.stats, } @override(TFPolicy) def get_weights(self): return self.variables.get_weights() @override(TFPolicy) def set_weights(self, weights): self.variables.set_weights(weights) @override(Policy) def get_state(self): return [ TFPolicy.get_state(self), self.cur_noise_scale, self.cur_pure_exploration_phase ] @override(Policy) def set_state(self, state): TFPolicy.set_state(self, state[0]) self.set_epsilon(state[1]) self.set_pure_exploration_phase(state[2]) def _build_q_network(self, obs, obs_space, action_space, actions): if self.config["use_state_preprocessor"]: q_model = ModelCatalog.get_model({ "obs": obs, "is_training": self._get_is_training_placeholder(), }, obs_space, action_space, 1, self.config["model"]) q_out = tf.concat([q_model.last_layer, actions], axis=1) else: q_model = None q_out = tf.concat([obs, actions], axis=1) activation = getattr(tf.nn, self.config["critic_hidden_activation"]) for hidden in self.config["critic_hiddens"]: q_out = tf.layers.dense(q_out, units=hidden, activation=activation) q_values = tf.layers.dense(q_out, units=1, activation=None) return q_values, q_model def _build_policy_network(self, obs, obs_space, action_space): if self.config["use_state_preprocessor"]: model = ModelCatalog.get_model({ "obs": obs, "is_training": self._get_is_training_placeholder(), }, obs_space, action_space, 1, self.config["model"]) action_out = model.last_layer else: model = None action_out = obs activation = getattr(tf.nn, self.config["actor_hidden_activation"]) for hidden in self.config["actor_hiddens"]: if self.config["parameter_noise"]: import tensorflow.contrib.layers as layers action_out = layers.fully_connected( action_out, num_outputs=hidden, activation_fn=activation, normalizer_fn=layers.layer_norm) else: action_out = tf.layers.dense( action_out, units=hidden, activation=activation) action_out = tf.layers.dense( action_out, units=self.dim_actions, activation=None) # Use sigmoid to scale to [0,1], but also double magnitude of input to # emulate behaviour of tanh activation used in DDPG and TD3 papers. sigmoid_out = tf.nn.sigmoid(2 * action_out) # Rescale to actual env policy scale # (shape of sigmoid_out is [batch_size, dim_actions], so we reshape to # get same dims) action_range = (action_space.high - action_space.low)[None] low_action = action_space.low[None] actions = action_range * sigmoid_out + low_action return actions, model def _add_exploration_noise(self, deterministic_actions, should_be_stochastic, noise_scale, enable_pure_exploration, action_space): noise_type = self.config["exploration_noise_type"] action_low = action_space.low action_high = action_space.high action_range = action_space.high - action_low def compute_stochastic_actions(): def make_noisy_actions(): # shape of deterministic_actions is [None, dim_action] if noise_type == "gaussian": # add IID Gaussian noise for exploration, TD3-style normal_sample = noise_scale * tf.random_normal( tf.shape(deterministic_actions), stddev=self.config["exploration_gaussian_sigma"]) stochastic_actions = tf.clip_by_value( deterministic_actions + normal_sample, action_low * tf.ones_like(deterministic_actions), action_high * tf.ones_like(deterministic_actions)) elif noise_type == "ou": # add OU noise for exploration, DDPG-style zero_acts = action_low.size * [.0] exploration_sample = tf.get_variable( name="ornstein_uhlenbeck", dtype=tf.float32, initializer=zero_acts, trainable=False) normal_sample = tf.random_normal( shape=[action_low.size], mean=0.0, stddev=1.0) ou_new = self.config["exploration_ou_theta"] \ * -exploration_sample \ + self.config["exploration_ou_sigma"] * normal_sample exploration_value = tf.assign_add(exploration_sample, ou_new) base_scale = self.config["exploration_ou_noise_scale"] noise = noise_scale * base_scale \ * exploration_value * action_range stochastic_actions = tf.clip_by_value( deterministic_actions + noise, action_low * tf.ones_like(deterministic_actions), action_high * tf.ones_like(deterministic_actions)) else: raise ValueError( "Unknown noise type '%s' (try 'ou' or 'gaussian')" % noise_type) return stochastic_actions def make_uniform_random_actions(): # pure random exploration option uniform_random_actions = tf.random_uniform( tf.shape(deterministic_actions)) # rescale uniform random actions according to action range tf_range = tf.constant(action_range[None], dtype="float32") tf_low = tf.constant(action_low[None], dtype="float32") uniform_random_actions = uniform_random_actions * tf_range \ + tf_low return uniform_random_actions stochastic_actions = tf.cond( # need to condition on noise_scale > 0 because zeroing # noise_scale is how a worker signals no noise should be used # (this is ugly and should be fixed by adding an "eval_mode" # config flag or something) tf.logical_and(enable_pure_exploration, noise_scale > 0), true_fn=make_uniform_random_actions, false_fn=make_noisy_actions) return stochastic_actions enable_stochastic = tf.logical_and(should_be_stochastic, not self.config["parameter_noise"]) actions = tf.cond(enable_stochastic, compute_stochastic_actions, lambda: deterministic_actions) return actions def _build_actor_critic_loss(self, q_t, q_tp1, q_t_det_policy, twin_q_t=None, twin_q_tp1=None): twin_q = self.config["twin_q"] gamma = self.config["gamma"] n_step = self.config["n_step"] use_huber = self.config["use_huber"] huber_threshold = self.config["huber_threshold"] q_t_selected = tf.squeeze(q_t, axis=len(q_t.shape) - 1) if twin_q: twin_q_t_selected = tf.squeeze(twin_q_t, axis=len(q_t.shape) - 1) q_tp1 = tf.minimum(q_tp1, twin_q_tp1) q_tp1_best = tf.squeeze(input=q_tp1, axis=len(q_tp1.shape) - 1) q_tp1_best_masked = (1.0 - self.done_mask) * q_tp1_best # compute RHS of bellman equation q_t_selected_target = tf.stop_gradient( self.rew_t + gamma**n_step * q_tp1_best_masked) # compute the error (potentially clipped) if twin_q: td_error = q_t_selected - q_t_selected_target twin_td_error = twin_q_t_selected - q_t_selected_target td_error = td_error + twin_td_error if use_huber: errors = huber_loss(td_error, huber_threshold) \ + huber_loss(twin_td_error, huber_threshold) else: errors = 0.5 * tf.square(td_error) + 0.5 * tf.square( twin_td_error) else: td_error = q_t_selected - q_t_selected_target if use_huber: errors = huber_loss(td_error, huber_threshold) else: errors = 0.5 * tf.square(td_error) critic_loss = tf.reduce_mean(self.importance_weights * errors) actor_loss = -tf.reduce_mean(q_t_det_policy) return critic_loss, actor_loss, td_error def _build_parameter_noise(self, pnet_params): self.parameter_noise_sigma_val = self.config["exploration_ou_sigma"] self.parameter_noise_sigma = tf.get_variable( initializer=tf.constant_initializer( self.parameter_noise_sigma_val), name="parameter_noise_sigma", shape=(), trainable=False, dtype=tf.float32) self.parameter_noise = list() # No need to add any noise on LayerNorm parameters for var in pnet_params: noise_var = tf.get_variable( name=var.name.split(":")[0] + "_noise", shape=var.shape, initializer=tf.constant_initializer(.0), trainable=False) self.parameter_noise.append(noise_var) remove_noise_ops = list() for var, var_noise in zip(pnet_params, self.parameter_noise): remove_noise_ops.append(tf.assign_add(var, -var_noise)) self.remove_noise_op = tf.group(*tuple(remove_noise_ops)) generate_noise_ops = list() for var_noise in self.parameter_noise: generate_noise_ops.append( tf.assign( var_noise, tf.random_normal( shape=var_noise.shape, stddev=self.parameter_noise_sigma))) with tf.control_dependencies(generate_noise_ops): add_noise_ops = list() for var, var_noise in zip(pnet_params, self.parameter_noise): add_noise_ops.append(tf.assign_add(var, var_noise)) self.add_noise_op = tf.group(*tuple(add_noise_ops)) self.pi_distance = None def compute_td_error(self, obs_t, act_t, rew_t, obs_tp1, done_mask, importance_weights): td_err = self.sess.run( self.td_error, feed_dict={ self.obs_t: [np.array(ob) for ob in obs_t], self.act_t: act_t, self.rew_t: rew_t, self.obs_tp1: [np.array(ob) for ob in obs_tp1], self.done_mask: done_mask, self.importance_weights: importance_weights }) return td_err def reset_noise(self, sess): sess.run(self.reset_noise_op) def add_parameter_noise(self): if self.config["parameter_noise"]: self.sess.run(self.add_noise_op) # support both hard and soft sync def update_target(self, tau=None): tau = tau or self.tau_value return self.sess.run( self.update_target_expr, feed_dict={self.tau: tau}) def set_epsilon(self, epsilon): # set_epsilon is called by optimizer to anneal exploration as # necessary, and to turn it off during evaluation. The "epsilon" part # is a carry-over from DQN, which uses epsilon-greedy exploration # rather than adding action noise to the output of a policy network. self.cur_noise_scale = epsilon def set_pure_exploration_phase(self, pure_exploration_phase): self.cur_pure_exploration_phase = pure_exploration_phase
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/ddpg/noop_model.py
Python
from ray.rllib.models.tf.tf_modelv2 import TFModelV2 from ray.rllib.utils.annotations import override from ray.rllib.utils import try_import_tf tf = try_import_tf() class NoopModel(TFModelV2): """Trivial model that just returns the obs flattened. This is the model used if use_state_preprocessor=False.""" @override(TFModelV2) def forward(self, input_dict, state, seq_lens): return tf.cast(input_dict["obs_flat"], tf.float32), state
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/ddpg/td3.py
Python
"""A more stable successor to TD3. By default, this uses a near-identical configuration to that reported in the TD3 paper. """ from ray.rllib.agents.ddpg.ddpg import DDPGTrainer, \ DEFAULT_CONFIG as DDPG_CONFIG from ray.rllib.utils import merge_dicts TD3_DEFAULT_CONFIG = merge_dicts( DDPG_CONFIG, { # largest changes: twin Q functions, delayed policy updates, and target # smoothing "twin_q": True, "policy_delay": 2, "smooth_target_policy": True, "target_noise": 0.2, "target_noise_clip": 0.5, # other changes & things we want to keep fixed: IID Gaussian # exploration noise, larger actor learning rate, no l2 regularisation, # no Huber loss, etc. "exploration_should_anneal": False, "exploration_noise_type": "gaussian", "exploration_gaussian_sigma": 0.1, "learning_starts": 10000, "pure_exploration_steps": 10000, "actor_hiddens": [400, 300], "critic_hiddens": [400, 300], "n_step": 1, "gamma": 0.99, "actor_lr": 1e-3, "critic_lr": 1e-3, "l2_reg": 0.0, "tau": 5e-3, "train_batch_size": 100, "use_huber": False, "target_network_update_freq": 0, "num_workers": 0, "num_gpus_per_worker": 0, "per_worker_exploration": False, "worker_side_prioritization": False, "buffer_size": 1000000, "prioritized_replay": False, "clip_rewards": False, "use_state_preprocessor": False, }, ) TD3Trainer = DDPGTrainer.with_updates( name="TD3", default_config=TD3_DEFAULT_CONFIG)
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/dqn/__init__.py
Python
from ray.rllib.agents.dqn.apex import ApexTrainer from ray.rllib.agents.dqn.dqn import DQNTrainer, SimpleQTrainer, DEFAULT_CONFIG from ray.rllib.utils import renamed_agent DQNAgent = renamed_agent(DQNTrainer) ApexAgent = renamed_agent(ApexTrainer) __all__ = [ "DQNAgent", "ApexAgent", "ApexTrainer", "DQNTrainer", "DEFAULT_CONFIG", "SimpleQTrainer" ]
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/dqn/apex.py
Python
from ray.rllib.agents.dqn.dqn import DQNTrainer, DEFAULT_CONFIG as DQN_CONFIG from ray.rllib.optimizers import AsyncReplayOptimizer from ray.rllib.utils import merge_dicts # yapf: disable # __sphinx_doc_begin__ APEX_DEFAULT_CONFIG = merge_dicts( DQN_CONFIG, # see also the options in dqn.py, which are also supported { "optimizer": merge_dicts( DQN_CONFIG["optimizer"], { "max_weight_sync_delay": 400, "num_replay_buffer_shards": 4, "debug": False }), "n_step": 3, "num_gpus": 1, "num_workers": 32, "buffer_size": 2000000, "learning_starts": 50000, "train_batch_size": 512, "sample_batch_size": 50, "target_network_update_freq": 500000, "timesteps_per_iteration": 25000, "per_worker_exploration": True, "worker_side_prioritization": True, "min_iter_time_s": 30, }, ) # __sphinx_doc_end__ # yapf: enable def defer_make_workers(trainer, env_creator, policy, config): # Hack to workaround https://github.com/ray-project/ray/issues/2541 # The workers will be created later, after the optimizer is created return trainer._make_workers(env_creator, policy, config, 0) def make_async_optimizer(workers, config): assert len(workers.remote_workers()) == 0 extra_config = config["optimizer"].copy() for key in [ "prioritized_replay", "prioritized_replay_alpha", "prioritized_replay_beta", "prioritized_replay_eps" ]: if key in config: extra_config[key] = config[key] opt = AsyncReplayOptimizer( workers, learning_starts=config["learning_starts"], buffer_size=config["buffer_size"], train_batch_size=config["train_batch_size"], sample_batch_size=config["sample_batch_size"], **extra_config) workers.add_workers(config["num_workers"]) opt._set_workers(workers.remote_workers()) return opt def update_target_based_on_num_steps_trained(trainer, fetches): # Ape-X updates based on num steps trained, not sampled if (trainer.optimizer.num_steps_trained - trainer.state["last_target_update_ts"] > trainer.config["target_network_update_freq"]): trainer.workers.local_worker().foreach_trainable_policy( lambda p, _: p.update_target()) trainer.state["last_target_update_ts"] = ( trainer.optimizer.num_steps_trained) trainer.state["num_target_updates"] += 1 APEX_TRAINER_PROPERTIES = { "make_workers": defer_make_workers, "make_policy_optimizer": make_async_optimizer, "after_optimizer_step": update_target_based_on_num_steps_trained, } ApexTrainer = DQNTrainer.with_updates( name="APEX", default_config=APEX_DEFAULT_CONFIG, **APEX_TRAINER_PROPERTIES)
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/dqn/distributional_q_model.py
Python
import numpy as np from ray.rllib.models.tf.tf_modelv2 import TFModelV2 from ray.rllib.utils import try_import_tf tf = try_import_tf() class DistributionalQModel(TFModelV2): """Extension of standard TFModel to provide distributional Q values. It also supports options for noisy nets and parameter space noise. Data flow: obs -> forward() -> model_out model_out -> get_q_value_distributions() -> Q(s, a) atoms model_out -> get_state_value() -> V(s) Note that this class by itself is not a valid model unless you implement forward() in a subclass.""" def __init__(self, obs_space, action_space, num_outputs, model_config, name, q_hiddens=(256, ), dueling=False, num_atoms=1, use_noisy=False, v_min=-10.0, v_max=10.0, sigma0=0.5, parameter_noise=False): """Initialize variables of this model. Extra model kwargs: q_hiddens (list): defines size of hidden layers for the q head. These will be used to postprocess the model output for the purposes of computing Q values. dueling (bool): whether to build the state value head for DDQN num_atoms (int): if >1, enables distributional DQN use_noisy (bool): use noisy nets v_min (float): min value support for distributional DQN v_max (float): max value support for distributional DQN sigma0 (float): initial value of noisy nets parameter_noise (bool): enable layer norm for param noise Note that the core layers for forward() are not defined here, this only defines the layers for the Q head. Those layers for forward() should be defined in subclasses of DistributionalQModel. """ super(DistributionalQModel, self).__init__( obs_space, action_space, num_outputs, model_config, name) # setup the Q head output (i.e., model for get_q_values) self.model_out = tf.keras.layers.Input( shape=(num_outputs, ), name="model_out") def build_action_value(model_out): if q_hiddens: action_out = model_out for i in range(len(q_hiddens)): if use_noisy: action_out = self._noisy_layer( "hidden_%d" % i, action_out, q_hiddens[i], sigma0) elif parameter_noise: action_out = tf.keras.layers.Dense( units=q_hiddens[i], activation_fn=tf.nn.relu, normalizer_fn=tf.keras.layers.LayerNormalization )(action_out) else: action_out = tf.keras.layers.Dense( units=q_hiddens[i], activation=tf.nn.relu, name="hidden_%d" % i )(action_out) else: # Avoid postprocessing the outputs. This enables custom models # to be used for parametric action DQN. action_out = model_out if use_noisy: action_scores = self._noisy_layer( "output", action_out, self.action_space.n * num_atoms, sigma0, non_linear=False) elif q_hiddens: action_scores = tf.keras.layers.Dense( units=self.action_space.n * num_atoms, activation=None )(action_out) else: action_scores = model_out if num_atoms > 1: # Distributional Q-learning uses a discrete support z # to represent the action value distribution z = tf.range(num_atoms, dtype=tf.float32) z = v_min + z * (v_max - v_min) / float(num_atoms - 1) support_logits_per_action = tf.reshape( tensor=action_scores, shape=(-1, self.action_space.n, num_atoms)) support_prob_per_action = tf.nn.softmax( logits=support_logits_per_action) action_scores = tf.reduce_sum( input_tensor=z * support_prob_per_action, axis=-1) logits = support_logits_per_action dist = support_prob_per_action return [ action_scores, z, support_logits_per_action, logits, dist ] else: logits = tf.expand_dims(tf.ones_like(action_scores), -1) dist = tf.expand_dims(tf.ones_like(action_scores), -1) return [action_scores, logits, dist] def build_state_score(model_out): state_out = model_out for i in range(len(q_hiddens)): if use_noisy: state_out = self._noisy_layer("dueling_hidden_%d" % i, state_out, q_hiddens[i], sigma0) elif parameter_noise: state_out = tf.keras.layers.Dense( units=q_hiddens[i], activation_fn=tf.nn.relu, normalizer_fn=tf.contrib.layers.layer_norm )(state_out) else: state_out = tf.keras.layers.Dense( units=q_hiddens[i], activation=tf.nn.relu )(state_out) if use_noisy: state_score = self._noisy_layer( "dueling_output", state_out, num_atoms, sigma0, non_linear=False) else: state_score = tf.keras.layers.Dense( units=num_atoms, activation=None )(state_out) return state_score if tf.executing_eagerly(): from tensorflow.python.ops import variable_scope # Have to use a variable store to reuse variables in eager mode store = variable_scope.EagerVariableStore() # Save the scope objects, since in eager we will execute this # path repeatedly and there is no guarantee it will always be run # in the same original scope. with tf.variable_scope(name + "/action_value") as action_scope: pass with tf.variable_scope(name + "/state_value") as state_scope: pass def build_action_value_in_scope(model_out): with store.as_default(): with tf.variable_scope(action_scope, reuse=tf.AUTO_REUSE): return build_action_value(model_out) def build_state_score_in_scope(model_out): with store.as_default(): with tf.variable_scope(state_scope, reuse=tf.AUTO_REUSE): return build_state_score(model_out) else: def build_action_value_in_scope(model_out): with tf.variable_scope( name + "/action_value", reuse=tf.AUTO_REUSE): return build_action_value(model_out) def build_state_score_in_scope(model_out): with tf.variable_scope( name + "/state_value", reuse=tf.AUTO_REUSE): return build_state_score(model_out) q_out = build_action_value_in_scope(self.model_out) self.q_value_head = tf.keras.Model(self.model_out, q_out) self.register_variables(self.q_value_head.variables) if dueling: state_out = build_state_score_in_scope( self.model_out) self.state_value_head = tf.keras.Model(self.model_out, state_out) self.register_variables(self.state_value_head.variables) def get_q_value_distributions(self, model_out): """Returns distributional values for Q(s, a) given a state embedding. Override this in your custom model to customize the Q output head. Arguments: model_out (Tensor): embedding from the model layers Returns: (action_scores, logits, dist) if num_atoms == 1, otherwise (action_scores, z, support_logits_per_action, logits, dist) """ return self.q_value_head(model_out) def get_state_value(self, model_out): """Returns the state value prediction for the given state embedding.""" return self.state_value_head(model_out) def _noisy_layer(self, prefix, action_in, out_size, sigma0, non_linear=True): """ a common dense layer: y = w^{T}x + b a noisy layer: y = (w + \epsilon_w*\sigma_w)^{T}x + (b+\epsilon_b*\sigma_b) where \epsilon are random variables sampled from factorized normal distributions and \sigma are trainable variables which are expected to vanish along the training procedure """ in_size = int(action_in.shape[1]) epsilon_in = tf.random_normal(shape=[in_size]) epsilon_out = tf.random_normal(shape=[out_size]) epsilon_in = self._f_epsilon(epsilon_in) epsilon_out = self._f_epsilon(epsilon_out) epsilon_w = tf.matmul( a=tf.expand_dims(epsilon_in, -1), b=tf.expand_dims(epsilon_out, 0)) epsilon_b = epsilon_out sigma_w = tf.get_variable( name=prefix + "_sigma_w", shape=[in_size, out_size], dtype=tf.float32, initializer=tf.random_uniform_initializer( minval=-1.0 / np.sqrt(float(in_size)), maxval=1.0 / np.sqrt(float(in_size)))) # TF noise generation can be unreliable on GPU # If generating the noise on the CPU, # lowering sigma0 to 0.1 may be helpful sigma_b = tf.get_variable( name=prefix + "_sigma_b", shape=[out_size], dtype=tf.float32, # 0.5~GPU, 0.1~CPU initializer=tf.constant_initializer( sigma0 / np.sqrt(float(in_size)))) w = tf.get_variable( name=prefix + "_fc_w", shape=[in_size, out_size], dtype=tf.float32, initializer=tf.initializers.GlorotUniform()) b = tf.get_variable( name=prefix + "_fc_b", shape=[out_size], dtype=tf.float32, initializer=tf.zeros_initializer()) action_activation = tf.nn.xw_plus_b(action_in, w + sigma_w * epsilon_w, b + sigma_b * epsilon_b) if not non_linear: return action_activation return tf.nn.relu(action_activation) def _f_epsilon(self, x): return tf.sign(x) * tf.sqrt(tf.abs(x))
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/dqn/dqn.py
Python
import logging from ray.rllib.agents.trainer import with_common_config from ray.rllib.agents.trainer_template import build_trainer from ray.rllib.agents.dqn.dqn_policy import DQNTFPolicy from ray.rllib.agents.dqn.simple_q_policy import SimpleQPolicy from ray.rllib.optimizers import SyncReplayOptimizer from ray.rllib.policy.sample_batch import DEFAULT_POLICY_ID from ray.rllib.utils.schedules import ConstantSchedule, LinearSchedule logger = logging.getLogger(__name__) # yapf: disable # __sphinx_doc_begin__ DEFAULT_CONFIG = with_common_config({ # === Model === # Number of atoms for representing the distribution of return. When # this is greater than 1, distributional Q-learning is used. # the discrete supports are bounded by v_min and v_max "num_atoms": 1, "v_min": -10.0, "v_max": 10.0, # Whether to use noisy network "noisy": False, # control the initial value of noisy nets "sigma0": 0.5, # Whether to use dueling dqn "dueling": True, # Whether to use double dqn "double_q": True, # Postprocess model outputs with these hidden layers to compute the # state and action values. See also the model config in catalog.py. "hiddens": [256], # N-step Q learning "n_step": 1, # === Exploration === # Max num timesteps for annealing schedules. Exploration is annealed from # 1.0 to exploration_fraction over this number of timesteps scaled by # exploration_fraction "schedule_max_timesteps": 100000, # Minimum env steps to optimize for per train call. This value does # not affect learning, only the length of iterations. "timesteps_per_iteration": 1000, # Fraction of entire training period over which the exploration rate is # annealed "exploration_fraction": 0.1, # Final value of random action probability "exploration_final_eps": 0.02, # Update the target network every `target_network_update_freq` steps. "target_network_update_freq": 500, # Use softmax for sampling actions. Required for off policy estimation. "soft_q": False, # Softmax temperature. Q values are divided by this value prior to softmax. # Softmax approaches argmax as the temperature drops to zero. "softmax_temp": 1.0, # If True parameter space noise will be used for exploration # See https://blog.openai.com/better-exploration-with-parameter-noise/ "parameter_noise": False, # Extra configuration that disables exploration. "evaluation_config": { "exploration_fraction": 0, "exploration_final_eps": 0, }, # === Replay buffer === # Size of the replay buffer. Note that if async_updates is set, then # each worker will have a replay buffer of this size. "buffer_size": 50000, # If True prioritized replay buffer will be used. "prioritized_replay": True, # Alpha parameter for prioritized replay buffer. "prioritized_replay_alpha": 0.6, # Beta parameter for sampling from prioritized replay buffer. "prioritized_replay_beta": 0.4, # Fraction of entire training period over which the beta parameter is # annealed "beta_annealing_fraction": 0.2, # Final value of beta "final_prioritized_replay_beta": 0.4, # Epsilon to add to the TD errors when updating priorities. "prioritized_replay_eps": 1e-6, # Whether to LZ4 compress observations "compress_observations": True, # === Optimization === # Learning rate for adam optimizer "lr": 5e-4, # Learning rate schedule "lr_schedule": None, # Adam epsilon hyper parameter "adam_epsilon": 1e-8, # If not None, clip gradients during optimization at this value "grad_norm_clipping": 40, # How many steps of the model to sample before learning starts. "learning_starts": 1000, # Update the replay buffer with this many samples at once. Note that # this setting applies per-worker if num_workers > 1. "sample_batch_size": 4, # Size of a batched sampled from replay buffer for training. Note that # if async_updates is set, then each worker returns gradients for a # batch of this size. "train_batch_size": 32, # === Parallelism === # Number of workers for collecting samples with. This only makes sense # to increase if your environment is particularly slow to sample, or if # you"re using the Async or Ape-X optimizers. "num_workers": 0, # Whether to use a distribution of epsilons across workers for exploration. "per_worker_exploration": False, # Whether to compute priorities on workers. "worker_side_prioritization": False, # Prevent iterations from going lower than this time span "min_iter_time_s": 1, }) # __sphinx_doc_end__ # yapf: enable def make_optimizer(workers, config): return SyncReplayOptimizer( workers, learning_starts=config["learning_starts"], buffer_size=config["buffer_size"], prioritized_replay=config["prioritized_replay"], prioritized_replay_alpha=config["prioritized_replay_alpha"], prioritized_replay_beta=config["prioritized_replay_beta"], schedule_max_timesteps=config["schedule_max_timesteps"], beta_annealing_fraction=config["beta_annealing_fraction"], final_prioritized_replay_beta=config["final_prioritized_replay_beta"], prioritized_replay_eps=config["prioritized_replay_eps"], train_batch_size=config["train_batch_size"], sample_batch_size=config["sample_batch_size"], **config["optimizer"]) def check_config_and_setup_param_noise(config): """Update the config based on settings. Rewrites sample_batch_size to take into account n_step truncation, and also adds the necessary callbacks to support parameter space noise exploration. """ # PyTorch check. if config["use_pytorch"]: raise ValueError("DQN does not support PyTorch yet! Use tf instead.") # Update effective batch size to include n-step adjusted_batch_size = max(config["sample_batch_size"], config.get("n_step", 1)) config["sample_batch_size"] = adjusted_batch_size if config.get("parameter_noise", False): if config["batch_mode"] != "complete_episodes": raise ValueError("Exploration with parameter space noise requires " "batch_mode to be complete_episodes.") if config.get("noisy", False): raise ValueError( "Exploration with parameter space noise and noisy network " "cannot be used at the same time.") if config["callbacks"]["on_episode_start"]: start_callback = config["callbacks"]["on_episode_start"] else: start_callback = None def on_episode_start(info): # as a callback function to sample and pose parameter space # noise on the parameters of network policies = info["policy"] for pi in policies.values(): pi.add_parameter_noise() if start_callback: start_callback(info) config["callbacks"]["on_episode_start"] = on_episode_start if config["callbacks"]["on_episode_end"]: end_callback = config["callbacks"]["on_episode_end"] else: end_callback = None def on_episode_end(info): # as a callback function to monitor the distance # between noisy policy and original policy policies = info["policy"] episode = info["episode"] model = policies[DEFAULT_POLICY_ID].model if hasattr(model, "pi_distance"): episode.custom_metrics["policy_distance"] = model.pi_distance if end_callback: end_callback(info) config["callbacks"]["on_episode_end"] = on_episode_end def get_initial_state(config): return { "last_target_update_ts": 0, "num_target_updates": 0, } def make_exploration_schedule(config, worker_index): # Use either a different `eps` per worker, or a linear schedule. if config["per_worker_exploration"]: assert config["num_workers"] > 1, \ "This requires multiple workers" if worker_index >= 0: # Exploration constants from the Ape-X paper exponent = ( 1 + worker_index / float(config["num_workers"] - 1) * 7) return ConstantSchedule(0.4**exponent) else: # local ev should have zero exploration so that eval rollouts # run properly return ConstantSchedule(0.0) return LinearSchedule( schedule_timesteps=int( config["exploration_fraction"] * config["schedule_max_timesteps"]), initial_p=1.0, final_p=config["exploration_final_eps"]) def setup_exploration(trainer): trainer.exploration0 = make_exploration_schedule(trainer.config, -1) trainer.explorations = [ make_exploration_schedule(trainer.config, i) for i in range(trainer.config["num_workers"]) ] def update_worker_explorations(trainer): global_timestep = trainer.optimizer.num_steps_sampled exp_vals = [trainer.exploration0.value(global_timestep)] trainer.workers.local_worker().foreach_trainable_policy( lambda p, _: p.set_epsilon(exp_vals[0])) for i, e in enumerate(trainer.workers.remote_workers()): exp_val = trainer.explorations[i].value(global_timestep) e.foreach_trainable_policy.remote(lambda p, _: p.set_epsilon(exp_val)) exp_vals.append(exp_val) trainer.train_start_timestep = global_timestep trainer.cur_exp_vals = exp_vals def add_trainer_metrics(trainer, result): global_timestep = trainer.optimizer.num_steps_sampled result.update( timesteps_this_iter=global_timestep - trainer.train_start_timestep, info=dict({ "min_exploration": min(trainer.cur_exp_vals), "max_exploration": max(trainer.cur_exp_vals), "num_target_updates": trainer.state["num_target_updates"], }, **trainer.optimizer.stats())) def update_target_if_needed(trainer, fetches): global_timestep = trainer.optimizer.num_steps_sampled if global_timestep - trainer.state["last_target_update_ts"] > \ trainer.config["target_network_update_freq"]: trainer.workers.local_worker().foreach_trainable_policy( lambda p, _: p.update_target()) trainer.state["last_target_update_ts"] = global_timestep trainer.state["num_target_updates"] += 1 def collect_metrics(trainer): if trainer.config["per_worker_exploration"]: # Only collect metrics from the third of workers with lowest eps result = trainer.collect_metrics( selected_workers=trainer.workers.remote_workers()[ -len(trainer.workers.remote_workers()) // 3:]) else: result = trainer.collect_metrics() return result def disable_exploration(trainer): trainer.evaluation_workers.local_worker().foreach_trainable_policy( lambda p, _: p.set_epsilon(0)) GenericOffPolicyTrainer = build_trainer( name="GenericOffPolicyAlgorithm", default_policy=None, default_config=DEFAULT_CONFIG, validate_config=check_config_and_setup_param_noise, get_initial_state=get_initial_state, make_policy_optimizer=make_optimizer, before_init=setup_exploration, before_train_step=update_worker_explorations, after_optimizer_step=update_target_if_needed, after_train_result=add_trainer_metrics, collect_metrics_fn=collect_metrics, before_evaluate_fn=disable_exploration) DQNTrainer = GenericOffPolicyTrainer.with_updates( name="DQN", default_policy=DQNTFPolicy, default_config=DEFAULT_CONFIG) SimpleQTrainer = DQNTrainer.with_updates(default_policy=SimpleQPolicy)
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/dqn/dqn_policy.py
Python
from gym.spaces import Discrete import numpy as np from scipy.stats import entropy import ray from ray.rllib.agents.dqn.distributional_q_model import DistributionalQModel from ray.rllib.agents.dqn.simple_q_policy import ExplorationStateMixin, \ TargetNetworkMixin from ray.rllib.policy.sample_batch import SampleBatch from ray.rllib.models import ModelCatalog from ray.rllib.models.tf.tf_action_dist import Categorical from ray.rllib.utils.error import UnsupportedSpaceException from ray.rllib.policy.tf_policy import LearningRateSchedule from ray.rllib.policy.tf_policy_template import build_tf_policy from ray.rllib.utils.tf_ops import huber_loss, reduce_mean_ignore_inf, \ minimize_and_clip from ray.rllib.utils import try_import_tf from ray.rllib.utils.tf_ops import make_tf_callable tf = try_import_tf() Q_SCOPE = "q_func" Q_TARGET_SCOPE = "target_q_func" # Importance sampling weights for prioritized replay PRIO_WEIGHTS = "weights" class QLoss: def __init__(self, q_t_selected, q_logits_t_selected, q_tp1_best, q_dist_tp1_best, importance_weights, rewards, done_mask, gamma=0.99, n_step=1, num_atoms=1, v_min=-10.0, v_max=10.0): if num_atoms > 1: # Distributional Q-learning which corresponds to an entropy loss z = tf.range(num_atoms, dtype=tf.float32) z = v_min + z * (v_max - v_min) / float(num_atoms - 1) # (batch_size, 1) * (1, num_atoms) = (batch_size, num_atoms) r_tau = tf.expand_dims( rewards, -1) + gamma**n_step * tf.expand_dims( 1.0 - done_mask, -1) * tf.expand_dims(z, 0) r_tau = tf.clip_by_value(r_tau, v_min, v_max) b = (r_tau - v_min) / ((v_max - v_min) / float(num_atoms - 1)) lb = tf.floor(b) ub = tf.ceil(b) # indispensable judgement which is missed in most implementations # when b happens to be an integer, lb == ub, so pr_j(s', a*) will # be discarded because (ub-b) == (b-lb) == 0 floor_equal_ceil = tf.to_float(tf.less(ub - lb, 0.5)) l_project = tf.one_hot( tf.cast(lb, dtype=tf.int32), num_atoms) # (batch_size, num_atoms, num_atoms) u_project = tf.one_hot( tf.cast(ub, dtype=tf.int32), num_atoms) # (batch_size, num_atoms, num_atoms) ml_delta = q_dist_tp1_best * (ub - b + floor_equal_ceil) mu_delta = q_dist_tp1_best * (b - lb) ml_delta = tf.reduce_sum( l_project * tf.expand_dims(ml_delta, -1), axis=1) mu_delta = tf.reduce_sum( u_project * tf.expand_dims(mu_delta, -1), axis=1) m = ml_delta + mu_delta # Rainbow paper claims that using this cross entropy loss for # priority is robust and insensitive to `prioritized_replay_alpha` self.td_error = tf.nn.softmax_cross_entropy_with_logits( labels=m, logits=q_logits_t_selected) self.loss = tf.reduce_mean(self.td_error * importance_weights) self.stats = { # TODO: better Q stats for dist dqn "mean_td_error": tf.reduce_mean(self.td_error), } else: q_tp1_best_masked = (1.0 - done_mask) * q_tp1_best # compute RHS of bellman equation q_t_selected_target = rewards + gamma**n_step * q_tp1_best_masked # compute the error (potentially clipped) self.td_error = ( q_t_selected - tf.stop_gradient(q_t_selected_target)) self.loss = tf.reduce_mean( tf.cast(importance_weights, tf.float32) * huber_loss( self.td_error)) self.stats = { "mean_q": tf.reduce_mean(q_t_selected), "min_q": tf.reduce_min(q_t_selected), "max_q": tf.reduce_max(q_t_selected), "mean_td_error": tf.reduce_mean(self.td_error), } class QValuePolicy: def __init__(self, q_values, observations, num_actions, cur_epsilon, softmax, softmax_temp, model_config): if softmax: action_dist = Categorical(q_values / softmax_temp) self.action = action_dist.sample() self.action_prob = tf.exp(action_dist.sampled_action_logp()) return deterministic_actions = tf.argmax(q_values, axis=1) batch_size = tf.shape(observations)[0] # Special case masked out actions (q_value ~= -inf) so that we don't # even consider them for exploration. random_valid_action_logits = tf.where( tf.equal(q_values, tf.float32.min), tf.ones_like(q_values) * tf.float32.min, tf.ones_like(q_values)) random_actions = tf.squeeze( tf.multinomial(random_valid_action_logits, 1), axis=1) chose_random = tf.random_uniform( tf.stack([batch_size]), minval=0, maxval=1, dtype=tf.float32) < cur_epsilon self.action = tf.where(chose_random, random_actions, deterministic_actions) self.action_prob = None class ComputeTDErrorMixin: def __init__(self): @make_tf_callable(self.get_session(), dynamic_shape=True) def compute_td_error(obs_t, act_t, rew_t, obs_tp1, done_mask, importance_weights): # Do forward pass on loss to update td error attribute build_q_losses( self, self.model, None, { SampleBatch.CUR_OBS: tf.convert_to_tensor(obs_t), SampleBatch.ACTIONS: tf.convert_to_tensor(act_t), SampleBatch.REWARDS: tf.convert_to_tensor(rew_t), SampleBatch.NEXT_OBS: tf.convert_to_tensor(obs_tp1), SampleBatch.DONES: tf.convert_to_tensor(done_mask), PRIO_WEIGHTS: tf.convert_to_tensor(importance_weights), }) return self.q_loss.td_error self.compute_td_error = compute_td_error def postprocess_trajectory(policy, sample_batch, other_agent_batches=None, episode=None): if policy.config["parameter_noise"]: # adjust the sigma of parameter space noise states = [list(x) for x in sample_batch.columns(["obs"])][0] noisy_action_distribution = policy.get_session().run( policy.action_probs, feed_dict={policy.cur_observations: states}) policy.get_session().run(policy.remove_noise_op) clean_action_distribution = policy.get_session().run( policy.action_probs, feed_dict={policy.cur_observations: states}) distance_in_action_space = np.mean( entropy(clean_action_distribution.T, noisy_action_distribution.T)) policy.pi_distance = distance_in_action_space if (distance_in_action_space < -np.log(1 - policy.cur_epsilon_value + policy.cur_epsilon_value / policy.num_actions)): policy.parameter_noise_sigma_val *= 1.01 else: policy.parameter_noise_sigma_val /= 1.01 policy.parameter_noise_sigma.load( policy.parameter_noise_sigma_val, session=policy.get_session()) return _postprocess_dqn(policy, sample_batch) def build_q_model(policy, obs_space, action_space, config): if not isinstance(action_space, Discrete): raise UnsupportedSpaceException( "Action space {} is not supported for DQN.".format(action_space)) if config["hiddens"]: # try to infer the last layer size, otherwise fall back to 256 num_outputs = ([256] + config["model"]["fcnet_hiddens"])[-1] config["model"]["no_final_linear"] = True else: num_outputs = action_space.n policy.q_model = ModelCatalog.get_model_v2( obs_space, action_space, num_outputs, config["model"], framework="tf", model_interface=DistributionalQModel, name=Q_SCOPE, num_atoms=config["num_atoms"], q_hiddens=config["hiddens"], dueling=config["dueling"], use_noisy=config["noisy"], v_min=config["v_min"], v_max=config["v_max"], sigma0=config["sigma0"], parameter_noise=config["parameter_noise"]) policy.target_q_model = ModelCatalog.get_model_v2( obs_space, action_space, num_outputs, config["model"], framework="tf", model_interface=DistributionalQModel, name=Q_TARGET_SCOPE, num_atoms=config["num_atoms"], q_hiddens=config["hiddens"], dueling=config["dueling"], use_noisy=config["noisy"], v_min=config["v_min"], v_max=config["v_max"], sigma0=config["sigma0"], parameter_noise=config["parameter_noise"]) return policy.q_model def build_q_networks(policy, q_model, input_dict, obs_space, action_space, config): # Action Q network q_values, q_logits, q_dist = _compute_q_values( policy, q_model, input_dict[SampleBatch.CUR_OBS], obs_space, action_space) policy.q_values = q_values policy.q_func_vars = q_model.variables() # Noise vars for Q network except for layer normalization vars if config["parameter_noise"]: _build_parameter_noise( policy, [var for var in policy.q_func_vars if "LayerNorm" not in var.name]) policy.action_probs = tf.nn.softmax(policy.q_values) # Action outputs qvp = QValuePolicy(q_values, input_dict[SampleBatch.CUR_OBS], action_space.n, policy.cur_epsilon, config["soft_q"], config["softmax_temp"], config["model"]) policy.output_actions, policy.action_prob = qvp.action, qvp.action_prob actions = policy.output_actions action_prob = (tf.log(policy.action_prob) if policy.action_prob is not None else None) return actions, action_prob def _build_parameter_noise(policy, pnet_params): policy.parameter_noise_sigma_val = 1.0 policy.parameter_noise_sigma = tf.get_variable( initializer=tf.constant_initializer(policy.parameter_noise_sigma_val), name="parameter_noise_sigma", shape=(), trainable=False, dtype=tf.float32) policy.parameter_noise = list() # No need to add any noise on LayerNorm parameters for var in pnet_params: noise_var = tf.get_variable( name=var.name.split(":")[0] + "_noise", shape=var.shape, initializer=tf.constant_initializer(.0), trainable=False) policy.parameter_noise.append(noise_var) remove_noise_ops = list() for var, var_noise in zip(pnet_params, policy.parameter_noise): remove_noise_ops.append(tf.assign_add(var, -var_noise)) policy.remove_noise_op = tf.group(*tuple(remove_noise_ops)) generate_noise_ops = list() for var_noise in policy.parameter_noise: generate_noise_ops.append( tf.assign( var_noise, tf.random_normal( shape=var_noise.shape, stddev=policy.parameter_noise_sigma))) with tf.control_dependencies(generate_noise_ops): add_noise_ops = list() for var, var_noise in zip(pnet_params, policy.parameter_noise): add_noise_ops.append(tf.assign_add(var, var_noise)) policy.add_noise_op = tf.group(*tuple(add_noise_ops)) policy.pi_distance = None def build_q_losses(policy, model, _, train_batch): config = policy.config # q network evaluation q_t, q_logits_t, q_dist_t = _compute_q_values( policy, policy.q_model, train_batch[SampleBatch.CUR_OBS], policy.observation_space, policy.action_space) # target q network evalution q_tp1, q_logits_tp1, q_dist_tp1 = _compute_q_values( policy, policy.target_q_model, train_batch[SampleBatch.NEXT_OBS], policy.observation_space, policy.action_space) policy.target_q_func_vars = policy.target_q_model.variables() # q scores for actions which we know were selected in the given state. one_hot_selection = tf.one_hot( tf.cast(train_batch[SampleBatch.ACTIONS], tf.int32), policy.action_space.n) q_t_selected = tf.reduce_sum(q_t * one_hot_selection, 1) q_logits_t_selected = tf.reduce_sum( q_logits_t * tf.expand_dims(one_hot_selection, -1), 1) # compute estimate of best possible value starting from state at t + 1 if config["double_q"]: q_tp1_using_online_net, q_logits_tp1_using_online_net, \ q_dist_tp1_using_online_net = _compute_q_values( policy, policy.q_model, train_batch[SampleBatch.NEXT_OBS], policy.observation_space, policy.action_space) q_tp1_best_using_online_net = tf.argmax(q_tp1_using_online_net, 1) q_tp1_best_one_hot_selection = tf.one_hot(q_tp1_best_using_online_net, policy.action_space.n) q_tp1_best = tf.reduce_sum(q_tp1 * q_tp1_best_one_hot_selection, 1) q_dist_tp1_best = tf.reduce_sum( q_dist_tp1 * tf.expand_dims(q_tp1_best_one_hot_selection, -1), 1) else: q_tp1_best_one_hot_selection = tf.one_hot( tf.argmax(q_tp1, 1), policy.action_space.n) q_tp1_best = tf.reduce_sum(q_tp1 * q_tp1_best_one_hot_selection, 1) q_dist_tp1_best = tf.reduce_sum( q_dist_tp1 * tf.expand_dims(q_tp1_best_one_hot_selection, -1), 1) policy.q_loss = QLoss( q_t_selected, q_logits_t_selected, q_tp1_best, q_dist_tp1_best, train_batch[PRIO_WEIGHTS], train_batch[SampleBatch.REWARDS], tf.cast(train_batch[SampleBatch.DONES], tf.float32), config["gamma"], config["n_step"], config["num_atoms"], config["v_min"], config["v_max"]) return policy.q_loss.loss def adam_optimizer(policy, config): return tf.train.AdamOptimizer( learning_rate=policy.cur_lr, epsilon=config["adam_epsilon"]) def clip_gradients(policy, optimizer, loss): if policy.config["grad_norm_clipping"] is not None: grads_and_vars = minimize_and_clip( optimizer, loss, var_list=policy.q_func_vars, clip_val=policy.config["grad_norm_clipping"]) else: grads_and_vars = optimizer.compute_gradients( loss, var_list=policy.q_func_vars) grads_and_vars = [(g, v) for (g, v) in grads_and_vars if g is not None] return grads_and_vars def build_q_stats(policy, batch): return dict({ "cur_lr": tf.cast(policy.cur_lr, tf.float64), }, **policy.q_loss.stats) def setup_early_mixins(policy, obs_space, action_space, config): LearningRateSchedule.__init__(policy, config["lr"], config["lr_schedule"]) ExplorationStateMixin.__init__(policy, obs_space, action_space, config) def setup_mid_mixins(policy, obs_space, action_space, config): ComputeTDErrorMixin.__init__(policy) def setup_late_mixins(policy, obs_space, action_space, config): TargetNetworkMixin.__init__(policy, obs_space, action_space, config) def _compute_q_values(policy, model, obs, obs_space, action_space): config = policy.config model_out, state = model({ "obs": obs, "is_training": policy._get_is_training_placeholder(), }, [], None) if config["num_atoms"] > 1: (action_scores, z, support_logits_per_action, logits, dist) = model.get_q_value_distributions(model_out) else: (action_scores, logits, dist) = model.get_q_value_distributions(model_out) if config["dueling"]: state_score = model.get_state_value(model_out) if config["num_atoms"] > 1: support_logits_per_action_mean = tf.reduce_mean( support_logits_per_action, 1) support_logits_per_action_centered = ( support_logits_per_action - tf.expand_dims( support_logits_per_action_mean, 1)) support_logits_per_action = tf.expand_dims( state_score, 1) + support_logits_per_action_centered support_prob_per_action = tf.nn.softmax( logits=support_logits_per_action) value = tf.reduce_sum( input_tensor=z * support_prob_per_action, axis=-1) logits = support_logits_per_action dist = support_prob_per_action else: action_scores_mean = reduce_mean_ignore_inf(action_scores, 1) action_scores_centered = action_scores - tf.expand_dims( action_scores_mean, 1) value = state_score + action_scores_centered else: value = action_scores return value, logits, dist def _adjust_nstep(n_step, gamma, obs, actions, rewards, new_obs, dones): """Rewrites the given trajectory fragments to encode n-step rewards. reward[i] = ( reward[i] * gamma**0 + reward[i+1] * gamma**1 + ... + reward[i+n_step-1] * gamma**(n_step-1)) The ith new_obs is also adjusted to point to the (i+n_step-1)'th new obs. At the end of the trajectory, n is truncated to fit in the traj length. """ assert not any(dones[:-1]), "Unexpected done in middle of trajectory" traj_length = len(rewards) for i in range(traj_length): for j in range(1, n_step): if i + j < traj_length: new_obs[i] = new_obs[i + j] dones[i] = dones[i + j] rewards[i] += gamma**j * rewards[i + j] def _postprocess_dqn(policy, batch): # N-step Q adjustments if policy.config["n_step"] > 1: _adjust_nstep(policy.config["n_step"], policy.config["gamma"], batch[SampleBatch.CUR_OBS], batch[SampleBatch.ACTIONS], batch[SampleBatch.REWARDS], batch[SampleBatch.NEXT_OBS], batch[SampleBatch.DONES]) if PRIO_WEIGHTS not in batch: batch[PRIO_WEIGHTS] = np.ones_like(batch[SampleBatch.REWARDS]) # Prioritize on the worker side if batch.count > 0 and policy.config["worker_side_prioritization"]: td_errors = policy.compute_td_error( batch[SampleBatch.CUR_OBS], batch[SampleBatch.ACTIONS], batch[SampleBatch.REWARDS], batch[SampleBatch.NEXT_OBS], batch[SampleBatch.DONES], batch[PRIO_WEIGHTS]) new_priorities = ( np.abs(td_errors) + policy.config["prioritized_replay_eps"]) batch.data[PRIO_WEIGHTS] = new_priorities return batch DQNTFPolicy = build_tf_policy( name="DQNTFPolicy", get_default_config=lambda: ray.rllib.agents.dqn.dqn.DEFAULT_CONFIG, make_model=build_q_model, action_sampler_fn=build_q_networks, loss_fn=build_q_losses, stats_fn=build_q_stats, postprocess_fn=postprocess_trajectory, optimizer_fn=adam_optimizer, gradients_fn=clip_gradients, extra_action_fetches_fn=lambda policy: {"q_values": policy.q_values}, extra_learn_fetches_fn=lambda policy: {"td_error": policy.q_loss.td_error}, before_init=setup_early_mixins, before_loss_init=setup_mid_mixins, after_init=setup_late_mixins, obs_include_prev_action_reward=False, mixins=[ ExplorationStateMixin, TargetNetworkMixin, ComputeTDErrorMixin, LearningRateSchedule, ])
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/dqn/simple_q_model.py
Python
from ray.rllib.models.tf.tf_modelv2 import TFModelV2 from ray.rllib.utils import try_import_tf tf = try_import_tf() class SimpleQModel(TFModelV2): """Extension of standard TFModel to provide Q values. Data flow: obs -> forward() -> model_out model_out -> get_q_values() -> Q(s, a) Note that this class by itself is not a valid model unless you implement forward() in a subclass.""" def __init__(self, obs_space, action_space, num_outputs, model_config, name, q_hiddens=(256, )): """Initialize variables of this model. Extra model kwargs: q_hiddens (list): defines size of hidden layers for the q head. These will be used to postprocess the model output for the purposes of computing Q values. Note that the core layers for forward() are not defined here, this only defines the layers for the Q head. Those layers for forward() should be defined in subclasses of SimpleQModel. """ super(SimpleQModel, self).__init__(obs_space, action_space, num_outputs, model_config, name) # setup the Q head output (i.e., model for get_q_values) self.model_out = tf.keras.layers.Input( shape=(num_outputs, ), name="model_out") if q_hiddens: last_layer = self.model_out for i, n in enumerate(q_hiddens): last_layer = tf.keras.layers.Dense( n, name="q_hidden_{}".format(i), activation=tf.nn.relu)(last_layer) q_out = tf.keras.layers.Dense( action_space.n, activation=None, name="q_out")(last_layer) else: q_out = self.model_out self.q_value_head = tf.keras.Model(self.model_out, q_out) self.register_variables(self.q_value_head.variables) def get_q_values(self, model_out): """Returns Q(s, a) given a feature tensor for the state. Override this in your custom model to customize the Q output head. Arguments: model_out (Tensor): embedding from the model layers Returns: action scores Q(s, a) for each action, shape [None, action_space.n] """ return self.q_value_head(model_out)
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/dqn/simple_q_policy.py
Python
"""Basic example of a DQN policy without any optimizations.""" from gym.spaces import Discrete import logging import ray from ray.rllib.agents.dqn.simple_q_model import SimpleQModel from ray.rllib.policy.policy import Policy from ray.rllib.policy.sample_batch import SampleBatch from ray.rllib.models import ModelCatalog from ray.rllib.utils.annotations import override from ray.rllib.utils.error import UnsupportedSpaceException from ray.rllib.policy.tf_policy import TFPolicy from ray.rllib.policy.tf_policy_template import build_tf_policy from ray.rllib.utils import try_import_tf from ray.rllib.utils.tf_ops import huber_loss, make_tf_callable tf = try_import_tf() logger = logging.getLogger(__name__) Q_SCOPE = "q_func" Q_TARGET_SCOPE = "target_q_func" class ExplorationStateMixin: def __init__(self, obs_space, action_space, config): # Python value, should always be same as the TF variable self.cur_epsilon_value = 1.0 self.cur_epsilon = tf.get_variable( initializer=tf.constant_initializer(self.cur_epsilon_value), name="eps", shape=(), trainable=False, dtype=tf.float32) def add_parameter_noise(self): if self.config["parameter_noise"]: self.sess.run(self.add_noise_op) def set_epsilon(self, epsilon): self.cur_epsilon_value = epsilon self.cur_epsilon.load( self.cur_epsilon_value, session=self.get_session()) @override(Policy) def get_state(self): return [TFPolicy.get_state(self), self.cur_epsilon_value] @override(Policy) def set_state(self, state): TFPolicy.set_state(self, state[0]) self.set_epsilon(state[1]) class TargetNetworkMixin: def __init__(self, obs_space, action_space, config): @make_tf_callable(self.get_session()) def do_update(): # update_target_fn will be called periodically to copy Q network to # target Q network update_target_expr = [] assert len(self.q_func_vars) == len(self.target_q_func_vars), \ (self.q_func_vars, self.target_q_func_vars) for var, var_target in zip(self.q_func_vars, self.target_q_func_vars): update_target_expr.append(var_target.assign(var)) logger.debug("Update target op {}".format(var_target)) return tf.group(*update_target_expr) self.update_target = do_update @override(TFPolicy) def variables(self): return self.q_func_vars + self.target_q_func_vars def build_q_models(policy, obs_space, action_space, config): if not isinstance(action_space, Discrete): raise UnsupportedSpaceException( "Action space {} is not supported for DQN.".format(action_space)) if config["hiddens"]: num_outputs = 256 config["model"]["no_final_linear"] = True else: num_outputs = action_space.n policy.q_model = ModelCatalog.get_model_v2( obs_space, action_space, num_outputs, config["model"], framework="tf", name=Q_SCOPE, model_interface=SimpleQModel, q_hiddens=config["hiddens"]) policy.target_q_model = ModelCatalog.get_model_v2( obs_space, action_space, num_outputs, config["model"], framework="tf", name=Q_TARGET_SCOPE, model_interface=SimpleQModel, q_hiddens=config["hiddens"]) return policy.q_model def build_action_sampler(policy, q_model, input_dict, obs_space, action_space, config): # Action Q network q_values = _compute_q_values(policy, q_model, input_dict[SampleBatch.CUR_OBS], obs_space, action_space) policy.q_values = q_values policy.q_func_vars = q_model.variables() # Action outputs deterministic_actions = tf.argmax(q_values, axis=1) batch_size = tf.shape(input_dict[SampleBatch.CUR_OBS])[0] # Special case masked out actions (q_value ~= -inf) so that we don't # even consider them for exploration. random_valid_action_logits = tf.where( tf.equal(q_values, tf.float32.min), tf.ones_like(q_values) * tf.float32.min, tf.ones_like(q_values)) random_actions = tf.squeeze( tf.multinomial(random_valid_action_logits, 1), axis=1) chose_random = tf.random_uniform( tf.stack([batch_size]), minval=0, maxval=1, dtype=tf.float32) < policy.cur_epsilon stochastic_actions = tf.where(chose_random, random_actions, deterministic_actions) action_logp = None return stochastic_actions, action_logp def build_q_losses(policy, model, dist_class, train_batch): # q network evaluation q_t = _compute_q_values(policy, policy.q_model, train_batch[SampleBatch.CUR_OBS], policy.observation_space, policy.action_space) # target q network evalution q_tp1 = _compute_q_values(policy, policy.target_q_model, train_batch[SampleBatch.NEXT_OBS], policy.observation_space, policy.action_space) policy.target_q_func_vars = policy.target_q_model.variables() # q scores for actions which we know were selected in the given state. one_hot_selection = tf.one_hot( tf.cast(train_batch[SampleBatch.ACTIONS], tf.int32), policy.action_space.n) q_t_selected = tf.reduce_sum(q_t * one_hot_selection, 1) # compute estimate of best possible value starting from state at t + 1 dones = tf.cast(train_batch[SampleBatch.DONES], tf.float32) q_tp1_best_one_hot_selection = tf.one_hot( tf.argmax(q_tp1, 1), policy.action_space.n) q_tp1_best = tf.reduce_sum(q_tp1 * q_tp1_best_one_hot_selection, 1) q_tp1_best_masked = (1.0 - dones) * q_tp1_best # compute RHS of bellman equation q_t_selected_target = (train_batch[SampleBatch.REWARDS] + policy.config["gamma"] * q_tp1_best_masked) # compute the error (potentially clipped) td_error = q_t_selected - tf.stop_gradient(q_t_selected_target) loss = tf.reduce_mean(huber_loss(td_error)) # save TD error as an attribute for outside access policy.td_error = td_error return loss def _compute_q_values(policy, model, obs, obs_space, action_space): input_dict = { "obs": obs, "is_training": policy._get_is_training_placeholder(), } model_out, _ = model(input_dict, [], None) return model.get_q_values(model_out) def setup_early_mixins(policy, obs_space, action_space, config): ExplorationStateMixin.__init__(policy, obs_space, action_space, config) def setup_late_mixins(policy, obs_space, action_space, config): TargetNetworkMixin.__init__(policy, obs_space, action_space, config) SimpleQPolicy = build_tf_policy( name="SimpleQPolicy", get_default_config=lambda: ray.rllib.agents.dqn.dqn.DEFAULT_CONFIG, make_model=build_q_models, action_sampler_fn=build_action_sampler, loss_fn=build_q_losses, extra_action_fetches_fn=lambda policy: {"q_values": policy.q_values}, extra_learn_fetches_fn=lambda policy: {"td_error": policy.td_error}, before_init=setup_early_mixins, after_init=setup_late_mixins, obs_include_prev_action_reward=False, mixins=[ ExplorationStateMixin, TargetNetworkMixin, ])
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/es/__init__.py
Python
from ray.rllib.agents.es.es import (ESTrainer, DEFAULT_CONFIG) from ray.rllib.utils import renamed_agent ESAgent = renamed_agent(ESTrainer) __all__ = ["ESAgent", "ESTrainer", "DEFAULT_CONFIG"]
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/es/es.py
Python
# Code in this file is copied and adapted from # https://github.com/openai/evolution-strategies-starter. from collections import namedtuple import logging import numpy as np import time import ray from ray.rllib.agents import Trainer, with_common_config from ray.rllib.agents.es import optimizers from ray.rllib.agents.es import policies from ray.rllib.agents.es import utils from ray.rllib.policy.sample_batch import DEFAULT_POLICY_ID from ray.rllib.utils.annotations import override from ray.rllib.utils.memory import ray_get_and_free from ray.rllib.utils import FilterManager logger = logging.getLogger(__name__) Result = namedtuple("Result", [ "noise_indices", "noisy_returns", "sign_noisy_returns", "noisy_lengths", "eval_returns", "eval_lengths" ]) # yapf: disable # __sphinx_doc_begin__ DEFAULT_CONFIG = with_common_config({ "l2_coeff": 0.005, "noise_stdev": 0.02, "episodes_per_batch": 1000, "train_batch_size": 10000, "eval_prob": 0.003, "return_proc_mode": "centered_rank", "num_workers": 10, "stepsize": 0.01, "observation_filter": "MeanStdFilter", "noise_size": 250000000, "report_length": 10, }) # __sphinx_doc_end__ # yapf: enable @ray.remote def create_shared_noise(count): """Create a large array of noise to be shared by all workers.""" seed = 123 noise = np.random.RandomState(seed).randn(count).astype(np.float32) return noise class SharedNoiseTable: def __init__(self, noise): self.noise = noise assert self.noise.dtype == np.float32 def get(self, i, dim): return self.noise[i:i + dim] def sample_index(self, dim): return np.random.randint(0, len(self.noise) - dim + 1) @ray.remote class Worker: def __init__(self, config, policy_params, env_creator, noise, min_task_runtime=0.2): self.min_task_runtime = min_task_runtime self.config = config self.policy_params = policy_params self.noise = SharedNoiseTable(noise) self.env = env_creator(config["env_config"]) from ray.rllib import models self.preprocessor = models.ModelCatalog.get_preprocessor( self.env, config["model"]) self.sess = utils.make_session(single_threaded=True) self.policy = policies.GenericPolicy( self.sess, self.env.action_space, self.env.observation_space, self.preprocessor, config["observation_filter"], config["model"], **policy_params) @property def filters(self): return {DEFAULT_POLICY_ID: self.policy.get_filter()} def sync_filters(self, new_filters): for k in self.filters: self.filters[k].sync(new_filters[k]) def get_filters(self, flush_after=False): return_filters = {} for k, f in self.filters.items(): return_filters[k] = f.as_serializable() if flush_after: f.clear_buffer() return return_filters def rollout(self, timestep_limit, add_noise=True): rollout_rewards, rollout_length = policies.rollout( self.policy, self.env, timestep_limit=timestep_limit, add_noise=add_noise) return rollout_rewards, rollout_length def do_rollouts(self, params, timestep_limit=None): # Set the network weights. self.policy.set_weights(params) noise_indices, returns, sign_returns, lengths = [], [], [], [] eval_returns, eval_lengths = [], [] # Perform some rollouts with noise. task_tstart = time.time() while (len(noise_indices) == 0 or time.time() - task_tstart < self.min_task_runtime): if np.random.uniform() < self.config["eval_prob"]: # Do an evaluation run with no perturbation. self.policy.set_weights(params) rewards, length = self.rollout(timestep_limit, add_noise=False) eval_returns.append(rewards.sum()) eval_lengths.append(length) else: # Do a regular run with parameter perturbations. noise_index = self.noise.sample_index(self.policy.num_params) perturbation = self.config["noise_stdev"] * self.noise.get( noise_index, self.policy.num_params) # These two sampling steps could be done in parallel on # different actors letting us update twice as frequently. self.policy.set_weights(params + perturbation) rewards_pos, lengths_pos = self.rollout(timestep_limit) self.policy.set_weights(params - perturbation) rewards_neg, lengths_neg = self.rollout(timestep_limit) noise_indices.append(noise_index) returns.append([rewards_pos.sum(), rewards_neg.sum()]) sign_returns.append( [np.sign(rewards_pos).sum(), np.sign(rewards_neg).sum()]) lengths.append([lengths_pos, lengths_neg]) return Result( noise_indices=noise_indices, noisy_returns=returns, sign_noisy_returns=sign_returns, noisy_lengths=lengths, eval_returns=eval_returns, eval_lengths=eval_lengths) class ESTrainer(Trainer): """Large-scale implementation of Evolution Strategies in Ray.""" _name = "ES" _default_config = DEFAULT_CONFIG @override(Trainer) def _init(self, config, env_creator): # PyTorch check. if config["use_pytorch"]: raise ValueError( "ES does not support PyTorch yet! Use tf instead." ) policy_params = {"action_noise_std": 0.01} env = env_creator(config["env_config"]) from ray.rllib import models preprocessor = models.ModelCatalog.get_preprocessor(env) self.sess = utils.make_session(single_threaded=False) self.policy = policies.GenericPolicy( self.sess, env.action_space, env.observation_space, preprocessor, config["observation_filter"], config["model"], **policy_params) self.optimizer = optimizers.Adam(self.policy, config["stepsize"]) self.report_length = config["report_length"] # Create the shared noise table. logger.info("Creating shared noise table.") noise_id = create_shared_noise.remote(config["noise_size"]) self.noise = SharedNoiseTable(ray.get(noise_id)) # Create the actors. logger.info("Creating actors.") self._workers = [ Worker.remote(config, policy_params, env_creator, noise_id) for _ in range(config["num_workers"]) ] self.episodes_so_far = 0 self.reward_list = [] self.tstart = time.time() @override(Trainer) def _train(self): config = self.config theta = self.policy.get_weights() assert theta.dtype == np.float32 # Put the current policy weights in the object store. theta_id = ray.put(theta) # Use the actors to do rollouts, note that we pass in the ID of the # policy weights. results, num_episodes, num_timesteps = self._collect_results( theta_id, config["episodes_per_batch"], config["train_batch_size"]) all_noise_indices = [] all_training_returns = [] all_training_lengths = [] all_eval_returns = [] all_eval_lengths = [] # Loop over the results. for result in results: all_eval_returns += result.eval_returns all_eval_lengths += result.eval_lengths all_noise_indices += result.noise_indices all_training_returns += result.noisy_returns all_training_lengths += result.noisy_lengths assert len(all_eval_returns) == len(all_eval_lengths) assert (len(all_noise_indices) == len(all_training_returns) == len(all_training_lengths)) self.episodes_so_far += num_episodes # Assemble the results. eval_returns = np.array(all_eval_returns) eval_lengths = np.array(all_eval_lengths) noise_indices = np.array(all_noise_indices) noisy_returns = np.array(all_training_returns) noisy_lengths = np.array(all_training_lengths) # Process the returns. if config["return_proc_mode"] == "centered_rank": proc_noisy_returns = utils.compute_centered_ranks(noisy_returns) else: raise NotImplementedError(config["return_proc_mode"]) # Compute and take a step. g, count = utils.batched_weighted_sum( proc_noisy_returns[:, 0] - proc_noisy_returns[:, 1], (self.noise.get(index, self.policy.num_params) for index in noise_indices), batch_size=500) g /= noisy_returns.size assert (g.shape == (self.policy.num_params, ) and g.dtype == np.float32 and count == len(noise_indices)) # Compute the new weights theta. theta, update_ratio = self.optimizer.update(-g + config["l2_coeff"] * theta) # Set the new weights in the local copy of the policy. self.policy.set_weights(theta) # Store the rewards if len(all_eval_returns) > 0: self.reward_list.append(np.mean(eval_returns)) # Now sync the filters FilterManager.synchronize({ DEFAULT_POLICY_ID: self.policy.get_filter() }, self._workers) info = { "weights_norm": np.square(theta).sum(), "grad_norm": np.square(g).sum(), "update_ratio": update_ratio, "episodes_this_iter": noisy_lengths.size, "episodes_so_far": self.episodes_so_far, } reward_mean = np.mean(self.reward_list[-self.report_length:]) result = dict( episode_reward_mean=reward_mean, episode_len_mean=eval_lengths.mean(), timesteps_this_iter=noisy_lengths.sum(), info=info) return result @override(Trainer) def compute_action(self, observation): return self.policy.compute(observation, update=False)[0] @override(Trainer) def _stop(self): # workaround for https://github.com/ray-project/ray/issues/1516 for w in self._workers: w.__ray_terminate__.remote() def _collect_results(self, theta_id, min_episodes, min_timesteps): num_episodes, num_timesteps = 0, 0 results = [] while num_episodes < min_episodes or num_timesteps < min_timesteps: logger.info( "Collected {} episodes {} timesteps so far this iter".format( num_episodes, num_timesteps)) rollout_ids = [ worker.do_rollouts.remote(theta_id) for worker in self._workers ] # Get the results of the rollouts. for result in ray_get_and_free(rollout_ids): results.append(result) # Update the number of episodes and the number of timesteps # keeping in mind that result.noisy_lengths is a list of lists, # where the inner lists have length 2. num_episodes += sum(len(pair) for pair in result.noisy_lengths) num_timesteps += sum( sum(pair) for pair in result.noisy_lengths) return results, num_episodes, num_timesteps def __getstate__(self): return { "weights": self.policy.get_weights(), "filter": self.policy.get_filter(), "episodes_so_far": self.episodes_so_far, } def __setstate__(self, state): self.episodes_so_far = state["episodes_so_far"] self.policy.set_weights(state["weights"]) self.policy.set_filter(state["filter"]) FilterManager.synchronize({ DEFAULT_POLICY_ID: self.policy.get_filter() }, self._workers)
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/es/optimizers.py
Python
# Code in this file is copied and adapted from # https://github.com/openai/evolution-strategies-starter. import numpy as np class Optimizer: def __init__(self, pi): self.pi = pi self.dim = pi.num_params self.t = 0 def update(self, globalg): self.t += 1 step = self._compute_step(globalg) theta = self.pi.get_weights() ratio = np.linalg.norm(step) / np.linalg.norm(theta) return theta + step, ratio def _compute_step(self, globalg): raise NotImplementedError class SGD(Optimizer): def __init__(self, pi, stepsize, momentum=0.9): Optimizer.__init__(self, pi) self.v = np.zeros(self.dim, dtype=np.float32) self.stepsize, self.momentum = stepsize, momentum def _compute_step(self, globalg): self.v = self.momentum * self.v + (1. - self.momentum) * globalg step = -self.stepsize * self.v return step class Adam(Optimizer): def __init__(self, pi, stepsize, beta1=0.9, beta2=0.999, epsilon=1e-08): Optimizer.__init__(self, pi) self.stepsize = stepsize self.beta1 = beta1 self.beta2 = beta2 self.epsilon = epsilon self.m = np.zeros(self.dim, dtype=np.float32) self.v = np.zeros(self.dim, dtype=np.float32) def _compute_step(self, globalg): a = self.stepsize * (np.sqrt(1 - self.beta2**self.t) / (1 - self.beta1**self.t)) self.m = self.beta1 * self.m + (1 - self.beta1) * globalg self.v = self.beta2 * self.v + (1 - self.beta2) * (globalg * globalg) step = -a * self.m / (np.sqrt(self.v) + self.epsilon) return step
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/es/policies.py
Python
# Code in this file is copied and adapted from # https://github.com/openai/evolution-strategies-starter. import gym import numpy as np import ray import ray.experimental.tf_utils from ray.rllib.evaluation.sampler import _unbatch_tuple_actions from ray.rllib.models import ModelCatalog from ray.rllib.utils.filter import get_filter from ray.rllib.utils import try_import_tf tf = try_import_tf() def rollout(policy, env, timestep_limit=None, add_noise=False): """Do a rollout. If add_noise is True, the rollout will take noisy actions with noise drawn from that stream. Otherwise, no action noise will be added. """ env_timestep_limit = env.spec.max_episode_steps timestep_limit = (env_timestep_limit if timestep_limit is None else min( timestep_limit, env_timestep_limit)) rews = [] t = 0 observation = env.reset() for _ in range(timestep_limit or 999999): ac = policy.compute(observation, add_noise=add_noise)[0] observation, rew, done, _ = env.step(ac) rews.append(rew) t += 1 if done: break rews = np.array(rews, dtype=np.float32) return rews, t class GenericPolicy: def __init__(self, sess, action_space, obs_space, preprocessor, observation_filter, model_options, action_noise_std): self.sess = sess self.action_space = action_space self.action_noise_std = action_noise_std self.preprocessor = preprocessor self.observation_filter = get_filter(observation_filter, self.preprocessor.shape) self.inputs = tf.placeholder(tf.float32, [None] + list(self.preprocessor.shape)) # Policy network. dist_class, dist_dim = ModelCatalog.get_action_dist( self.action_space, model_options, dist_type="deterministic") model = ModelCatalog.get_model({ "obs": self.inputs }, obs_space, action_space, dist_dim, model_options) dist = dist_class(model.outputs, model) self.sampler = dist.sample() self.variables = ray.experimental.tf_utils.TensorFlowVariables( model.outputs, self.sess) self.num_params = sum( np.prod(variable.shape.as_list()) for _, variable in self.variables.variables.items()) self.sess.run(tf.global_variables_initializer()) def compute(self, observation, add_noise=False, update=True): observation = self.preprocessor.transform(observation) observation = self.observation_filter(observation[None], update=update) action = self.sess.run( self.sampler, feed_dict={self.inputs: observation}) action = _unbatch_tuple_actions(action) if add_noise and isinstance(self.action_space, gym.spaces.Box): action += np.random.randn(*action.shape) * self.action_noise_std return action def set_weights(self, x): self.variables.set_flat(x) def get_weights(self): return self.variables.get_flat() def get_filter(self): return self.observation_filter def set_filter(self, observation_filter): self.observation_filter = observation_filter
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/es/utils.py
Python
# Code in this file is copied and adapted from # https://github.com/openai/evolution-strategies-starter. import numpy as np from ray.rllib.utils import try_import_tf tf = try_import_tf() def compute_ranks(x): """Returns ranks in [0, len(x)) Note: This is different from scipy.stats.rankdata, which returns ranks in [1, len(x)]. """ assert x.ndim == 1 ranks = np.empty(len(x), dtype=int) ranks[x.argsort()] = np.arange(len(x)) return ranks def compute_centered_ranks(x): y = compute_ranks(x.ravel()).reshape(x.shape).astype(np.float32) y /= (x.size - 1) y -= 0.5 return y def make_session(single_threaded): if not single_threaded: return tf.Session() return tf.Session( config=tf.ConfigProto( inter_op_parallelism_threads=1, intra_op_parallelism_threads=1)) def itergroups(items, group_size): assert group_size >= 1 group = [] for x in items: group.append(x) if len(group) == group_size: yield tuple(group) del group[:] if group: yield tuple(group) def batched_weighted_sum(weights, vecs, batch_size): total = 0 num_items_summed = 0 for batch_weights, batch_vecs in zip( itergroups(weights, batch_size), itergroups(vecs, batch_size)): assert len(batch_weights) == len(batch_vecs) <= batch_size total += np.dot( np.asarray(batch_weights, dtype=np.float32), np.asarray(batch_vecs, dtype=np.float32)) num_items_summed += len(batch_weights) return total, num_items_summed
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/impala/__init__.py
Python
from ray.rllib.agents.impala.impala import ImpalaTrainer, DEFAULT_CONFIG from ray.rllib.utils import renamed_agent ImpalaAgent = renamed_agent(ImpalaTrainer) __all__ = ["ImpalaAgent", "ImpalaTrainer", "DEFAULT_CONFIG"]
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/impala/impala.py
Python
from ray.rllib.agents.a3c.a3c_tf_policy import A3CTFPolicy from ray.rllib.agents.impala.vtrace_policy import VTraceTFPolicy from ray.rllib.agents.trainer import Trainer, with_common_config from ray.rllib.agents.trainer_template import build_trainer from ray.rllib.optimizers import AsyncSamplesOptimizer from ray.rllib.optimizers.aso_tree_aggregator import TreeAggregator from ray.rllib.utils.annotations import override from ray.tune.trainable import Trainable from ray.tune.resources import Resources # yapf: disable # __sphinx_doc_begin__ DEFAULT_CONFIG = with_common_config({ # V-trace params (see vtrace.py). "vtrace": True, "vtrace_clip_rho_threshold": 1.0, "vtrace_clip_pg_rho_threshold": 1.0, # System params. # # == Overview of data flow in IMPALA == # 1. Policy evaluation in parallel across `num_workers` actors produces # batches of size `sample_batch_size * num_envs_per_worker`. # 2. If enabled, the replay buffer stores and produces batches of size # `sample_batch_size * num_envs_per_worker`. # 3. If enabled, the minibatch ring buffer stores and replays batches of # size `train_batch_size` up to `num_sgd_iter` times per batch. # 4. The learner thread executes data parallel SGD across `num_gpus` GPUs # on batches of size `train_batch_size`. # "sample_batch_size": 50, "train_batch_size": 500, "min_iter_time_s": 10, "num_workers": 2, # number of GPUs the learner should use. "num_gpus": 1, # set >1 to load data into GPUs in parallel. Increases GPU memory usage # proportionally with the number of buffers. "num_data_loader_buffers": 1, # how many train batches should be retained for minibatching. This conf # only has an effect if `num_sgd_iter > 1`. "minibatch_buffer_size": 1, # number of passes to make over each train batch "num_sgd_iter": 1, # set >0 to enable experience replay. Saved samples will be replayed with # a p:1 proportion to new data samples. "replay_proportion": 0.0, # number of sample batches to store for replay. The number of transitions # saved total will be (replay_buffer_num_slots * sample_batch_size). "replay_buffer_num_slots": 0, # max queue size for train batches feeding into the learner "learner_queue_size": 16, # wait for train batches to be available in minibatch buffer queue # this many seconds. This may need to be increased e.g. when training # with a slow environment "learner_queue_timeout": 300, # level of queuing for sampling. "max_sample_requests_in_flight_per_worker": 2, # max number of workers to broadcast one set of weights to "broadcast_interval": 1, # use intermediate actors for multi-level aggregation. This can make sense # if ingesting >2GB/s of samples, or if the data requires decompression. "num_aggregation_workers": 0, # Learning params. "grad_clip": 40.0, # either "adam" or "rmsprop" "opt_type": "adam", "lr": 0.0005, "lr_schedule": None, # rmsprop considered "decay": 0.99, "momentum": 0.0, "epsilon": 0.1, # balancing the three losses "vf_loss_coeff": 0.5, "entropy_coeff": 0.01, "entropy_coeff_schedule": None, # use fake (infinite speed) sampler for testing "_fake_sampler": False, }) # __sphinx_doc_end__ # yapf: enable def choose_policy(config): if config["vtrace"]: return VTraceTFPolicy else: return A3CTFPolicy def validate_config(config): # PyTorch check. if config["use_pytorch"]: raise ValueError( "IMPALA does not support PyTorch yet! Use tf instead." ) if config["entropy_coeff"] < 0: raise DeprecationWarning("entropy_coeff must be >= 0") def defer_make_workers(trainer, env_creator, policy, config): # Defer worker creation to after the optimizer has been created. return trainer._make_workers(env_creator, policy, config, 0) def make_aggregators_and_optimizer(workers, config): if config["num_aggregation_workers"] > 0: # Create co-located aggregator actors first for placement pref aggregators = TreeAggregator.precreate_aggregators( config["num_aggregation_workers"]) else: aggregators = None workers.add_workers(config["num_workers"]) optimizer = AsyncSamplesOptimizer( workers, lr=config["lr"], num_gpus=config["num_gpus"], sample_batch_size=config["sample_batch_size"], train_batch_size=config["train_batch_size"], replay_buffer_num_slots=config["replay_buffer_num_slots"], replay_proportion=config["replay_proportion"], num_data_loader_buffers=config["num_data_loader_buffers"], max_sample_requests_in_flight_per_worker=config[ "max_sample_requests_in_flight_per_worker"], broadcast_interval=config["broadcast_interval"], num_sgd_iter=config["num_sgd_iter"], minibatch_buffer_size=config["minibatch_buffer_size"], num_aggregation_workers=config["num_aggregation_workers"], learner_queue_size=config["learner_queue_size"], learner_queue_timeout=config["learner_queue_timeout"], **config["optimizer"]) if aggregators: # Assign the pre-created aggregators to the optimizer optimizer.aggregator.init(aggregators) return optimizer class OverrideDefaultResourceRequest: @classmethod @override(Trainable) def default_resource_request(cls, config): cf = dict(cls._default_config, **config) Trainer._validate_config(cf) return Resources( cpu=cf["num_cpus_for_driver"], gpu=cf["num_gpus"], memory=cf["memory"], object_store_memory=cf["object_store_memory"], extra_cpu=cf["num_cpus_per_worker"] * cf["num_workers"] + cf["num_aggregation_workers"], extra_gpu=cf["num_gpus_per_worker"] * cf["num_workers"], extra_memory=cf["memory_per_worker"] * cf["num_workers"], extra_object_store_memory=cf["object_store_memory_per_worker"] * cf["num_workers"]) ImpalaTrainer = build_trainer( name="IMPALA", default_config=DEFAULT_CONFIG, default_policy=VTraceTFPolicy, validate_config=validate_config, get_policy_class=choose_policy, make_workers=defer_make_workers, make_policy_optimizer=make_aggregators_and_optimizer, mixins=[OverrideDefaultResourceRequest])
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/impala/vtrace.py
Python
# Copyright 2018 Google LLC # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # https://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Functions to compute V-trace off-policy actor critic targets. For details and theory see: "IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures" by Espeholt, Soyer, Munos et al. See https://arxiv.org/abs/1802.01561 for the full paper. In addition to the original paper's code, changes have been made to support MultiDiscrete action spaces. behaviour_policy_logits, target_policy_logits and actions parameters in the entry point multi_from_logits method accepts lists of tensors instead of just tensors. """ import collections from ray.rllib.models.tf.tf_action_dist import Categorical from ray.rllib.utils import try_import_tf tf = try_import_tf() VTraceFromLogitsReturns = collections.namedtuple("VTraceFromLogitsReturns", [ "vs", "pg_advantages", "log_rhos", "behaviour_action_log_probs", "target_action_log_probs" ]) VTraceReturns = collections.namedtuple("VTraceReturns", "vs pg_advantages") def log_probs_from_logits_and_actions(policy_logits, actions, dist_class=Categorical, model=None): return multi_log_probs_from_logits_and_actions([policy_logits], [actions], dist_class, model)[0] def multi_log_probs_from_logits_and_actions(policy_logits, actions, dist_class, model): """Computes action log-probs from policy logits and actions. In the notation used throughout documentation and comments, T refers to the time dimension ranging from 0 to T-1. B refers to the batch size and ACTION_SPACE refers to the list of numbers each representing a number of actions. Args: policy_logits: A list with length of ACTION_SPACE of float32 tensors of shapes [T, B, ACTION_SPACE[0]], ..., [T, B, ACTION_SPACE[-1]] with un-normalized log-probabilities parameterizing a softmax policy. actions: A list with length of ACTION_SPACE of tensors of shapes [T, B, ...], ..., [T, B, ...] with actions. dist_class: Python class of the action distribution Returns: A list with length of ACTION_SPACE of float32 tensors of shapes [T, B], ..., [T, B] corresponding to the sampling log probability of the chosen action w.r.t. the policy. """ log_probs = [] for i in range(len(policy_logits)): p_shape = tf.shape(policy_logits[i]) a_shape = tf.shape(actions[i]) policy_logits_flat = tf.reshape(policy_logits[i], tf.concat([[-1], p_shape[2:]], axis=0)) actions_flat = tf.reshape(actions[i], tf.concat([[-1], a_shape[2:]], axis=0)) log_probs.append( tf.reshape( dist_class(policy_logits_flat, model).logp(actions_flat), a_shape[:2])) return log_probs def from_logits(behaviour_policy_logits, target_policy_logits, actions, discounts, rewards, values, bootstrap_value, dist_class=Categorical, model=None, clip_rho_threshold=1.0, clip_pg_rho_threshold=1.0, name="vtrace_from_logits"): """multi_from_logits wrapper used only for tests""" res = multi_from_logits( [behaviour_policy_logits], [target_policy_logits], [actions], discounts, rewards, values, bootstrap_value, dist_class, model, clip_rho_threshold=clip_rho_threshold, clip_pg_rho_threshold=clip_pg_rho_threshold, name=name) return VTraceFromLogitsReturns( vs=res.vs, pg_advantages=res.pg_advantages, log_rhos=res.log_rhos, behaviour_action_log_probs=tf.squeeze( res.behaviour_action_log_probs, axis=0), target_action_log_probs=tf.squeeze( res.target_action_log_probs, axis=0), ) def multi_from_logits(behaviour_policy_logits, target_policy_logits, actions, discounts, rewards, values, bootstrap_value, dist_class, model, behaviour_action_log_probs=None, clip_rho_threshold=1.0, clip_pg_rho_threshold=1.0, name="vtrace_from_logits"): r"""V-trace for softmax policies. Calculates V-trace actor critic targets for softmax polices as described in "IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures" by Espeholt, Soyer, Munos et al. Target policy refers to the policy we are interested in improving and behaviour policy refers to the policy that generated the given rewards and actions. In the notation used throughout documentation and comments, T refers to the time dimension ranging from 0 to T-1. B refers to the batch size and ACTION_SPACE refers to the list of numbers each representing a number of actions. Args: behaviour_policy_logits: A list with length of ACTION_SPACE of float32 tensors of shapes [T, B, ACTION_SPACE[0]], ..., [T, B, ACTION_SPACE[-1]] with un-normalized log-probabilities parameterizing the softmax behaviour policy. target_policy_logits: A list with length of ACTION_SPACE of float32 tensors of shapes [T, B, ACTION_SPACE[0]], ..., [T, B, ACTION_SPACE[-1]] with un-normalized log-probabilities parameterizing the softmax target policy. actions: A list with length of ACTION_SPACE of tensors of shapes [T, B, ...], ..., [T, B, ...] with actions sampled from the behaviour policy. discounts: A float32 tensor of shape [T, B] with the discount encountered when following the behaviour policy. rewards: A float32 tensor of shape [T, B] with the rewards generated by following the behaviour policy. values: A float32 tensor of shape [T, B] with the value function estimates wrt. the target policy. bootstrap_value: A float32 of shape [B] with the value function estimate at time T. dist_class: action distribution class for the logits. model: backing ModelV2 instance behaviour_action_log_probs: precalculated values of the behaviour actions clip_rho_threshold: A scalar float32 tensor with the clipping threshold for importance weights (rho) when calculating the baseline targets (vs). rho^bar in the paper. clip_pg_rho_threshold: A scalar float32 tensor with the clipping threshold on rho_s in \rho_s \delta log \pi(a|x) (r + \gamma v_{s+1} - V(x_s)). name: The name scope that all V-trace operations will be created in. Returns: A `VTraceFromLogitsReturns` namedtuple with the following fields: vs: A float32 tensor of shape [T, B]. Can be used as target to train a baseline (V(x_t) - vs_t)^2. pg_advantages: A float 32 tensor of shape [T, B]. Can be used as an estimate of the advantage in the calculation of policy gradients. log_rhos: A float32 tensor of shape [T, B] containing the log importance sampling weights (log rhos). behaviour_action_log_probs: A float32 tensor of shape [T, B] containing behaviour policy action log probabilities (log \mu(a_t)). target_action_log_probs: A float32 tensor of shape [T, B] containing target policy action probabilities (log \pi(a_t)). """ for i in range(len(behaviour_policy_logits)): behaviour_policy_logits[i] = tf.convert_to_tensor( behaviour_policy_logits[i], dtype=tf.float32) target_policy_logits[i] = tf.convert_to_tensor( target_policy_logits[i], dtype=tf.float32) # Make sure tensor ranks are as expected. # The rest will be checked by from_action_log_probs. behaviour_policy_logits[i].shape.assert_has_rank(3) target_policy_logits[i].shape.assert_has_rank(3) with tf.name_scope( name, values=[ behaviour_policy_logits, target_policy_logits, actions, discounts, rewards, values, bootstrap_value ]): target_action_log_probs = multi_log_probs_from_logits_and_actions( target_policy_logits, actions, dist_class, model) if (len(behaviour_policy_logits) > 1 or behaviour_action_log_probs is None): # can't use precalculated values, recompute them. Note that # recomputing won't work well for autoregressive action dists # which may have variables not captured by 'logits' behaviour_action_log_probs = ( multi_log_probs_from_logits_and_actions( behaviour_policy_logits, actions, dist_class, model)) log_rhos = get_log_rhos(target_action_log_probs, behaviour_action_log_probs) vtrace_returns = from_importance_weights( log_rhos=log_rhos, discounts=discounts, rewards=rewards, values=values, bootstrap_value=bootstrap_value, clip_rho_threshold=clip_rho_threshold, clip_pg_rho_threshold=clip_pg_rho_threshold) return VTraceFromLogitsReturns( log_rhos=log_rhos, behaviour_action_log_probs=behaviour_action_log_probs, target_action_log_probs=target_action_log_probs, **vtrace_returns._asdict()) def from_importance_weights(log_rhos, discounts, rewards, values, bootstrap_value, clip_rho_threshold=1.0, clip_pg_rho_threshold=1.0, name="vtrace_from_importance_weights"): r"""V-trace from log importance weights. Calculates V-trace actor critic targets as described in "IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures" by Espeholt, Soyer, Munos et al. In the notation used throughout documentation and comments, T refers to the time dimension ranging from 0 to T-1. B refers to the batch size. This code also supports the case where all tensors have the same number of additional dimensions, e.g., `rewards` is [T, B, C], `values` is [T, B, C], `bootstrap_value` is [B, C]. Args: log_rhos: A float32 tensor of shape [T, B] representing the log importance sampling weights, i.e. log(target_policy(a) / behaviour_policy(a)). V-trace performs operations on rhos in log-space for numerical stability. discounts: A float32 tensor of shape [T, B] with discounts encountered when following the behaviour policy. rewards: A float32 tensor of shape [T, B] containing rewards generated by following the behaviour policy. values: A float32 tensor of shape [T, B] with the value function estimates wrt. the target policy. bootstrap_value: A float32 of shape [B] with the value function estimate at time T. clip_rho_threshold: A scalar float32 tensor with the clipping threshold for importance weights (rho) when calculating the baseline targets (vs). rho^bar in the paper. If None, no clipping is applied. clip_pg_rho_threshold: A scalar float32 tensor with the clipping threshold on rho_s in \rho_s \delta log \pi(a|x) (r + \gamma v_{s+1} - V(x_s)). If None, no clipping is applied. name: The name scope that all V-trace operations will be created in. Returns: A VTraceReturns namedtuple (vs, pg_advantages) where: vs: A float32 tensor of shape [T, B]. Can be used as target to train a baseline (V(x_t) - vs_t)^2. pg_advantages: A float32 tensor of shape [T, B]. Can be used as the advantage in the calculation of policy gradients. """ log_rhos = tf.convert_to_tensor(log_rhos, dtype=tf.float32) discounts = tf.convert_to_tensor(discounts, dtype=tf.float32) rewards = tf.convert_to_tensor(rewards, dtype=tf.float32) values = tf.convert_to_tensor(values, dtype=tf.float32) bootstrap_value = tf.convert_to_tensor(bootstrap_value, dtype=tf.float32) if clip_rho_threshold is not None: clip_rho_threshold = tf.convert_to_tensor( clip_rho_threshold, dtype=tf.float32) if clip_pg_rho_threshold is not None: clip_pg_rho_threshold = tf.convert_to_tensor( clip_pg_rho_threshold, dtype=tf.float32) # Make sure tensor ranks are consistent. rho_rank = log_rhos.shape.ndims # Usually 2. values.shape.assert_has_rank(rho_rank) bootstrap_value.shape.assert_has_rank(rho_rank - 1) discounts.shape.assert_has_rank(rho_rank) rewards.shape.assert_has_rank(rho_rank) if clip_rho_threshold is not None: clip_rho_threshold.shape.assert_has_rank(0) if clip_pg_rho_threshold is not None: clip_pg_rho_threshold.shape.assert_has_rank(0) with tf.name_scope( name, values=[log_rhos, discounts, rewards, values, bootstrap_value]): rhos = tf.exp(log_rhos) if clip_rho_threshold is not None: clipped_rhos = tf.minimum( clip_rho_threshold, rhos, name="clipped_rhos") tf.summary.histogram("clipped_rhos_1000", tf.minimum(1000.0, rhos)) tf.summary.scalar( "num_of_clipped_rhos", tf.reduce_sum( tf.cast( tf.equal(clipped_rhos, clip_rho_threshold), tf.int32))) tf.summary.scalar("size_of_clipped_rhos", tf.size(clipped_rhos)) else: clipped_rhos = rhos cs = tf.minimum(1.0, rhos, name="cs") # Append bootstrapped value to get [v1, ..., v_t+1] values_t_plus_1 = tf.concat( [values[1:], tf.expand_dims(bootstrap_value, 0)], axis=0) deltas = clipped_rhos * ( rewards + discounts * values_t_plus_1 - values) # All sequences are reversed, computation starts from the back. sequences = ( tf.reverse(discounts, axis=[0]), tf.reverse(cs, axis=[0]), tf.reverse(deltas, axis=[0]), ) # V-trace vs are calculated through a scan from the back to the # beginning of the given trajectory. def scanfunc(acc, sequence_item): discount_t, c_t, delta_t = sequence_item return delta_t + discount_t * c_t * acc initial_values = tf.zeros_like(bootstrap_value) vs_minus_v_xs = tf.scan( fn=scanfunc, elems=sequences, initializer=initial_values, parallel_iterations=1, back_prop=False, name="scan") # Reverse the results back to original order. vs_minus_v_xs = tf.reverse(vs_minus_v_xs, [0], name="vs_minus_v_xs") # Add V(x_s) to get v_s. vs = tf.add(vs_minus_v_xs, values, name="vs") # Advantage for policy gradient. vs_t_plus_1 = tf.concat( [vs[1:], tf.expand_dims(bootstrap_value, 0)], axis=0) if clip_pg_rho_threshold is not None: clipped_pg_rhos = tf.minimum( clip_pg_rho_threshold, rhos, name="clipped_pg_rhos") else: clipped_pg_rhos = rhos pg_advantages = ( clipped_pg_rhos * (rewards + discounts * vs_t_plus_1 - values)) # Make sure no gradients backpropagated through the returned values. return VTraceReturns( vs=tf.stop_gradient(vs), pg_advantages=tf.stop_gradient(pg_advantages)) def get_log_rhos(target_action_log_probs, behaviour_action_log_probs): """With the selected log_probs for multi-discrete actions of behaviour and target policies we compute the log_rhos for calculating the vtrace.""" t = tf.stack(target_action_log_probs) b = tf.stack(behaviour_action_log_probs) log_rhos = tf.reduce_sum(t - b, axis=0) return log_rhos
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/impala/vtrace_policy.py
Python
"""Adapted from A3CTFPolicy to add V-trace. Keep in sync with changes to A3CTFPolicy and VtraceSurrogatePolicy.""" import numpy as np import logging import gym import ray from ray.rllib.agents.impala import vtrace from ray.rllib.models.tf.tf_action_dist import Categorical from ray.rllib.policy.sample_batch import SampleBatch from ray.rllib.policy.tf_policy_template import build_tf_policy from ray.rllib.policy.tf_policy import LearningRateSchedule, \ EntropyCoeffSchedule, ACTION_LOGP from ray.rllib.utils.explained_variance import explained_variance from ray.rllib.utils import try_import_tf tf = try_import_tf() logger = logging.getLogger(__name__) BEHAVIOUR_LOGITS = "behaviour_logits" class VTraceLoss: def __init__(self, actions, actions_logp, actions_entropy, dones, behaviour_action_logp, behaviour_logits, target_logits, discount, rewards, values, bootstrap_value, dist_class, model, valid_mask, config, vf_loss_coeff=0.5, entropy_coeff=0.01, clip_rho_threshold=1.0, clip_pg_rho_threshold=1.0): """Policy gradient loss with vtrace importance weighting. VTraceLoss takes tensors of shape [T, B, ...], where `B` is the batch_size. The reason we need to know `B` is for V-trace to properly handle episode cut boundaries. Args: actions: An int|float32 tensor of shape [T, B, ACTION_SPACE]. actions_logp: A float32 tensor of shape [T, B]. actions_entropy: A float32 tensor of shape [T, B]. dones: A bool tensor of shape [T, B]. behaviour_action_logp: Tensor of shape [T, B]. behaviour_logits: A list with length of ACTION_SPACE of float32 tensors of shapes [T, B, ACTION_SPACE[0]], ..., [T, B, ACTION_SPACE[-1]] target_logits: A list with length of ACTION_SPACE of float32 tensors of shapes [T, B, ACTION_SPACE[0]], ..., [T, B, ACTION_SPACE[-1]] discount: A float32 scalar. rewards: A float32 tensor of shape [T, B]. values: A float32 tensor of shape [T, B]. bootstrap_value: A float32 tensor of shape [B]. dist_class: action distribution class for logits. valid_mask: A bool tensor of valid RNN input elements (#2992). config: Trainer config dict. """ # Compute vtrace on the CPU for better perf. with tf.device("/cpu:0"): self.vtrace_returns = vtrace.multi_from_logits( behaviour_action_log_probs=behaviour_action_logp, behaviour_policy_logits=behaviour_logits, target_policy_logits=target_logits, actions=tf.unstack(actions, axis=2), discounts=tf.to_float(~dones) * discount, rewards=rewards, values=values, bootstrap_value=bootstrap_value, dist_class=dist_class, model=model, clip_rho_threshold=tf.cast(clip_rho_threshold, tf.float32), clip_pg_rho_threshold=tf.cast(clip_pg_rho_threshold, tf.float32)) self.value_targets = self.vtrace_returns.vs # The policy gradients loss self.pi_loss = -tf.reduce_sum( tf.boolean_mask(actions_logp * self.vtrace_returns.pg_advantages, valid_mask)) # The baseline loss delta = tf.boolean_mask(values - self.vtrace_returns.vs, valid_mask) self.vf_loss = 0.5 * tf.reduce_sum(tf.square(delta)) # The entropy loss self.entropy = tf.reduce_sum( tf.boolean_mask(actions_entropy, valid_mask)) # The summed weighted loss self.total_loss = (self.pi_loss + self.vf_loss * vf_loss_coeff - self.entropy * entropy_coeff) def _make_time_major(policy, seq_lens, tensor, drop_last=False): """Swaps batch and trajectory axis. Arguments: policy: Policy reference seq_lens: Sequence lengths if recurrent or None tensor: A tensor or list of tensors to reshape. drop_last: A bool indicating whether to drop the last trajectory item. Returns: res: A tensor with swapped axes or a list of tensors with swapped axes. """ if isinstance(tensor, list): return [ _make_time_major(policy, seq_lens, t, drop_last) for t in tensor ] if policy.is_recurrent(): B = tf.shape(seq_lens)[0] T = tf.shape(tensor)[0] // B else: # Important: chop the tensor into batches at known episode cut # boundaries. TODO(ekl) this is kind of a hack T = policy.config["sample_batch_size"] B = tf.shape(tensor)[0] // T rs = tf.reshape(tensor, tf.concat([[B, T], tf.shape(tensor)[1:]], axis=0)) # swap B and T axes res = tf.transpose( rs, [1, 0] + list(range(2, 1 + int(tf.shape(tensor).shape[0])))) if drop_last: return res[:-1] return res def build_vtrace_loss(policy, model, dist_class, train_batch): model_out, _ = model.from_batch(train_batch) action_dist = dist_class(model_out, model) if isinstance(policy.action_space, gym.spaces.Discrete): is_multidiscrete = False output_hidden_shape = [policy.action_space.n] elif isinstance(policy.action_space, gym.spaces.multi_discrete.MultiDiscrete): is_multidiscrete = True output_hidden_shape = policy.action_space.nvec.astype(np.int32) else: is_multidiscrete = False output_hidden_shape = 1 def make_time_major(*args, **kw): return _make_time_major(policy, train_batch.get("seq_lens"), *args, **kw) actions = train_batch[SampleBatch.ACTIONS] dones = train_batch[SampleBatch.DONES] rewards = train_batch[SampleBatch.REWARDS] behaviour_action_logp = train_batch[ACTION_LOGP] behaviour_logits = train_batch[BEHAVIOUR_LOGITS] unpacked_behaviour_logits = tf.split( behaviour_logits, output_hidden_shape, axis=1) unpacked_outputs = tf.split(model_out, output_hidden_shape, axis=1) values = model.value_function() if policy.is_recurrent(): max_seq_len = tf.reduce_max(train_batch["seq_lens"]) - 1 mask = tf.sequence_mask(train_batch["seq_lens"], max_seq_len) mask = tf.reshape(mask, [-1]) else: mask = tf.ones_like(rewards) # Prepare actions for loss loss_actions = actions if is_multidiscrete else tf.expand_dims( actions, axis=1) # Inputs are reshaped from [B * T] => [T - 1, B] for V-trace calc. policy.loss = VTraceLoss( actions=make_time_major(loss_actions, drop_last=True), actions_logp=make_time_major( action_dist.logp(actions), drop_last=True), actions_entropy=make_time_major( action_dist.multi_entropy(), drop_last=True), dones=make_time_major(dones, drop_last=True), behaviour_action_logp=make_time_major( behaviour_action_logp, drop_last=True), behaviour_logits=make_time_major( unpacked_behaviour_logits, drop_last=True), target_logits=make_time_major(unpacked_outputs, drop_last=True), discount=policy.config["gamma"], rewards=make_time_major(rewards, drop_last=True), values=make_time_major(values, drop_last=True), bootstrap_value=make_time_major(values)[-1], dist_class=Categorical if is_multidiscrete else dist_class, model=model, valid_mask=make_time_major(mask, drop_last=True), config=policy.config, vf_loss_coeff=policy.config["vf_loss_coeff"], entropy_coeff=policy.entropy_coeff, clip_rho_threshold=policy.config["vtrace_clip_rho_threshold"], clip_pg_rho_threshold=policy.config["vtrace_clip_pg_rho_threshold"]) return policy.loss.total_loss def stats(policy, train_batch): values_batched = _make_time_major( policy, train_batch.get("seq_lens"), policy.model.value_function(), drop_last=policy.config["vtrace"]) return { "cur_lr": tf.cast(policy.cur_lr, tf.float64), "policy_loss": policy.loss.pi_loss, "entropy": policy.loss.entropy, "entropy_coeff": tf.cast(policy.entropy_coeff, tf.float64), "var_gnorm": tf.global_norm(policy.model.trainable_variables()), "vf_loss": policy.loss.vf_loss, "vf_explained_var": explained_variance( tf.reshape(policy.loss.value_targets, [-1]), tf.reshape(values_batched, [-1])), } def grad_stats(policy, train_batch, grads): return { "grad_gnorm": tf.global_norm(grads), } def postprocess_trajectory(policy, sample_batch, other_agent_batches=None, episode=None): # not used, so save some bandwidth del sample_batch.data[SampleBatch.NEXT_OBS] return sample_batch def add_behaviour_logits(policy): return {BEHAVIOUR_LOGITS: policy.model.last_output()} def validate_config(policy, obs_space, action_space, config): if config["vtrace"]: assert config["batch_mode"] == "truncate_episodes", \ "Must use `truncate_episodes` batch mode with V-trace." def choose_optimizer(policy, config): if policy.config["opt_type"] == "adam": return tf.train.AdamOptimizer(policy.cur_lr) else: return tf.train.RMSPropOptimizer(policy.cur_lr, config["decay"], config["momentum"], config["epsilon"]) def clip_gradients(policy, optimizer, loss): grads_and_vars = optimizer.compute_gradients( loss, policy.model.trainable_variables()) grads = [g for (g, v) in grads_and_vars] policy.grads, _ = tf.clip_by_global_norm(grads, policy.config["grad_clip"]) clipped_grads = list(zip(policy.grads, policy.model.trainable_variables())) return clipped_grads def setup_mixins(policy, obs_space, action_space, config): LearningRateSchedule.__init__(policy, config["lr"], config["lr_schedule"]) EntropyCoeffSchedule.__init__(policy, config["entropy_coeff"], config["entropy_coeff_schedule"]) VTraceTFPolicy = build_tf_policy( name="VTraceTFPolicy", get_default_config=lambda: ray.rllib.agents.impala.impala.DEFAULT_CONFIG, loss_fn=build_vtrace_loss, stats_fn=stats, grad_stats_fn=grad_stats, postprocess_fn=postprocess_trajectory, optimizer_fn=choose_optimizer, gradients_fn=clip_gradients, extra_action_fetches_fn=add_behaviour_logits, before_init=validate_config, before_loss_init=setup_mixins, mixins=[LearningRateSchedule, EntropyCoeffSchedule], get_batch_divisibility_req=lambda p: p.config["sample_batch_size"])
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/impala/vtrace_test.py
Python
# Copyright 2018 Google LLC # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # https://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. """Tests for V-trace. For details and theory see: "IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures" by Espeholt, Soyer, Munos et al. """ from absl.testing import parameterized import numpy as np import vtrace from ray.rllib.utils import try_import_tf tf = try_import_tf() def _shaped_arange(*shape): """Runs np.arange, converts to float and reshapes.""" return np.arange(np.prod(shape), dtype=np.float32).reshape(*shape) def _softmax(logits): """Applies softmax non-linearity on inputs.""" return np.exp(logits) / np.sum(np.exp(logits), axis=-1, keepdims=True) def _ground_truth_calculation(discounts, log_rhos, rewards, values, bootstrap_value, clip_rho_threshold, clip_pg_rho_threshold): """Calculates the ground truth for V-trace in Python/Numpy.""" vs = [] seq_len = len(discounts) rhos = np.exp(log_rhos) cs = np.minimum(rhos, 1.0) clipped_rhos = rhos if clip_rho_threshold: clipped_rhos = np.minimum(rhos, clip_rho_threshold) clipped_pg_rhos = rhos if clip_pg_rho_threshold: clipped_pg_rhos = np.minimum(rhos, clip_pg_rho_threshold) # This is a very inefficient way to calculate the V-trace ground truth. # We calculate it this way because it is close to the mathematical notation # of # V-trace. # v_s = V(x_s) # + \sum^{T-1}_{t=s} \gamma^{t-s} # * \prod_{i=s}^{t-1} c_i # * \rho_t (r_t + \gamma V(x_{t+1}) - V(x_t)) # Note that when we take the product over c_i, we write `s:t` as the # notation # of the paper is inclusive of the `t-1`, but Python is exclusive. # Also note that np.prod([]) == 1. values_t_plus_1 = np.concatenate( [values, bootstrap_value[None, :]], axis=0) for s in range(seq_len): v_s = np.copy(values[s]) # Very important copy. for t in range(s, seq_len): v_s += (np.prod(discounts[s:t], axis=0) * np.prod(cs[s:t], axis=0) * clipped_rhos[t] * (rewards[t] + discounts[t] * values_t_plus_1[t + 1] - values[t])) vs.append(v_s) vs = np.stack(vs, axis=0) pg_advantages = (clipped_pg_rhos * (rewards + discounts * np.concatenate( [vs[1:], bootstrap_value[None, :]], axis=0) - values)) return vtrace.VTraceReturns(vs=vs, pg_advantages=pg_advantages) class LogProbsFromLogitsAndActionsTest(tf.test.TestCase, parameterized.TestCase): @parameterized.named_parameters(("Batch1", 1), ("Batch2", 2)) def test_log_probs_from_logits_and_actions(self, batch_size): """Tests log_probs_from_logits_and_actions.""" seq_len = 7 num_actions = 3 policy_logits = _shaped_arange(seq_len, batch_size, num_actions) + 10 actions = np.random.randint( 0, num_actions - 1, size=(seq_len, batch_size), dtype=np.int32) action_log_probs_tensor = vtrace.log_probs_from_logits_and_actions( policy_logits, actions) # Ground Truth # Using broadcasting to create a mask that indexes action logits action_index_mask = actions[..., None] == np.arange(num_actions) def index_with_mask(array, mask): return array[mask].reshape(*array.shape[:-1]) # Note: Normally log(softmax) is not a good idea because it's not # numerically stable. However, in this test we have well-behaved # values. ground_truth_v = index_with_mask( np.log(_softmax(policy_logits)), action_index_mask) with self.test_session() as session: self.assertAllClose(ground_truth_v, session.run(action_log_probs_tensor)) class VtraceTest(tf.test.TestCase, parameterized.TestCase): @parameterized.named_parameters(("Batch1", 1), ("Batch5", 5)) def test_vtrace(self, batch_size): """Tests V-trace against ground truth data calculated in python.""" seq_len = 5 # Create log_rhos such that rho will span from near-zero to above the # clipping thresholds. In particular, calculate log_rhos in # [-2.5, 2.5), # so that rho is in approx [0.08, 12.2). log_rhos = _shaped_arange(seq_len, batch_size) / (batch_size * seq_len) log_rhos = 5 * (log_rhos - 0.5) # [0.0, 1.0) -> [-2.5, 2.5). values = { "log_rhos": log_rhos, # T, B where B_i: [0.9 / (i+1)] * T "discounts": np.array([[0.9 / (b + 1) for b in range(batch_size)] for _ in range(seq_len)]), "rewards": _shaped_arange(seq_len, batch_size), "values": _shaped_arange(seq_len, batch_size) / batch_size, "bootstrap_value": _shaped_arange(batch_size) + 1.0, "clip_rho_threshold": 3.7, "clip_pg_rho_threshold": 2.2, } output = vtrace.from_importance_weights(**values) with self.test_session() as session: output_v = session.run(output) ground_truth_v = _ground_truth_calculation(**values) for a, b in zip(ground_truth_v, output_v): self.assertAllClose(a, b) @parameterized.named_parameters(("Batch1", 1), ("Batch2", 2)) def test_vtrace_from_logits(self, batch_size): """Tests V-trace calculated from logits.""" seq_len = 5 num_actions = 3 clip_rho_threshold = None # No clipping. clip_pg_rho_threshold = None # No clipping. # Intentionally leaving shapes unspecified to test if V-trace can # deal with that. placeholders = { # T, B, NUM_ACTIONS "behaviour_policy_logits": tf.placeholder( dtype=tf.float32, shape=[None, None, None]), # T, B, NUM_ACTIONS "target_policy_logits": tf.placeholder( dtype=tf.float32, shape=[None, None, None]), "actions": tf.placeholder(dtype=tf.int32, shape=[None, None]), "discounts": tf.placeholder(dtype=tf.float32, shape=[None, None]), "rewards": tf.placeholder(dtype=tf.float32, shape=[None, None]), "values": tf.placeholder(dtype=tf.float32, shape=[None, None]), "bootstrap_value": tf.placeholder(dtype=tf.float32, shape=[None]), } from_logits_output = vtrace.from_logits( clip_rho_threshold=clip_rho_threshold, clip_pg_rho_threshold=clip_pg_rho_threshold, **placeholders) target_log_probs = vtrace.log_probs_from_logits_and_actions( placeholders["target_policy_logits"], placeholders["actions"]) behaviour_log_probs = vtrace.log_probs_from_logits_and_actions( placeholders["behaviour_policy_logits"], placeholders["actions"]) log_rhos = target_log_probs - behaviour_log_probs ground_truth = (log_rhos, behaviour_log_probs, target_log_probs) values = { "behaviour_policy_logits": _shaped_arange(seq_len, batch_size, num_actions), "target_policy_logits": _shaped_arange(seq_len, batch_size, num_actions), "actions": np.random.randint( 0, num_actions - 1, size=(seq_len, batch_size)), "discounts": np.array( # T, B where B_i: [0.9 / (i+1)] * T [[0.9 / (b + 1) for b in range(batch_size)] for _ in range(seq_len)]), "rewards": _shaped_arange(seq_len, batch_size), "values": _shaped_arange(seq_len, batch_size) / batch_size, "bootstrap_value": _shaped_arange(batch_size) + 1.0, # B } feed_dict = {placeholders[k]: v for k, v in values.items()} with self.test_session() as session: from_logits_output_v = session.run( from_logits_output, feed_dict=feed_dict) (ground_truth_log_rhos, ground_truth_behaviour_action_log_probs, ground_truth_target_action_log_probs) = session.run( ground_truth, feed_dict=feed_dict) # Calculate V-trace using the ground truth logits. from_iw = vtrace.from_importance_weights( log_rhos=ground_truth_log_rhos, discounts=values["discounts"], rewards=values["rewards"], values=values["values"], bootstrap_value=values["bootstrap_value"], clip_rho_threshold=clip_rho_threshold, clip_pg_rho_threshold=clip_pg_rho_threshold) with self.test_session() as session: from_iw_v = session.run(from_iw) self.assertAllClose(from_iw_v.vs, from_logits_output_v.vs) self.assertAllClose(from_iw_v.pg_advantages, from_logits_output_v.pg_advantages) self.assertAllClose(ground_truth_behaviour_action_log_probs, from_logits_output_v.behaviour_action_log_probs) self.assertAllClose(ground_truth_target_action_log_probs, from_logits_output_v.target_action_log_probs) self.assertAllClose(ground_truth_log_rhos, from_logits_output_v.log_rhos) def test_higher_rank_inputs_for_importance_weights(self): """Checks support for additional dimensions in inputs.""" placeholders = { "log_rhos": tf.placeholder( dtype=tf.float32, shape=[None, None, 1]), "discounts": tf.placeholder( dtype=tf.float32, shape=[None, None, 1]), "rewards": tf.placeholder( dtype=tf.float32, shape=[None, None, 42]), "values": tf.placeholder(dtype=tf.float32, shape=[None, None, 42]), "bootstrap_value": tf.placeholder( dtype=tf.float32, shape=[None, 42]) } output = vtrace.from_importance_weights(**placeholders) self.assertEqual(output.vs.shape.as_list()[-1], 42) def test_inconsistent_rank_inputs_for_importance_weights(self): """Test one of many possible errors in shape of inputs.""" placeholders = { "log_rhos": tf.placeholder( dtype=tf.float32, shape=[None, None, 1]), "discounts": tf.placeholder( dtype=tf.float32, shape=[None, None, 1]), "rewards": tf.placeholder( dtype=tf.float32, shape=[None, None, 42]), "values": tf.placeholder(dtype=tf.float32, shape=[None, None, 42]), # Should be [None, 42]. "bootstrap_value": tf.placeholder(dtype=tf.float32, shape=[None]) } with self.assertRaisesRegexp(ValueError, "must have rank 2"): vtrace.from_importance_weights(**placeholders) if __name__ == "__main__": tf.test.main()
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/marwil/__init__.py
Python
from ray.rllib.agents.marwil.marwil import MARWILTrainer, DEFAULT_CONFIG __all__ = ["MARWILTrainer", "DEFAULT_CONFIG"]
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/marwil/marwil.py
Python
from ray.rllib.agents.trainer import with_common_config from ray.rllib.agents.trainer_template import build_trainer from ray.rllib.agents.marwil.marwil_policy import MARWILPolicy from ray.rllib.optimizers import SyncBatchReplayOptimizer # yapf: disable # __sphinx_doc_begin__ DEFAULT_CONFIG = with_common_config({ # You should override this to point to an offline dataset (see agent.py). "input": "sampler", # Use importance sampling estimators for reward "input_evaluation": ["is", "wis"], # Scaling of advantages in exponential terms # When beta is 0, MARWIL is reduced to imitation learning "beta": 1.0, # Balancing value estimation loss and policy optimization loss "vf_coeff": 1.0, # Whether to calculate cumulative rewards "postprocess_inputs": True, # Whether to rollout "complete_episodes" or "truncate_episodes" "batch_mode": "complete_episodes", # Learning rate for adam optimizer "lr": 1e-4, # Number of timesteps collected for each SGD round "train_batch_size": 2000, # Number of steps max to keep in the batch replay buffer "replay_buffer_size": 100000, # Number of steps to read before learning starts "learning_starts": 0, # === Parallelism === "num_workers": 0, }) # __sphinx_doc_end__ # yapf: enable def make_optimizer(workers, config): return SyncBatchReplayOptimizer( workers, learning_starts=config["learning_starts"], buffer_size=config["replay_buffer_size"], train_batch_size=config["train_batch_size"], ) def validate_config(config): # PyTorch check. if config["use_pytorch"]: raise ValueError("DDPG does not support PyTorch yet! Use tf instead.") MARWILTrainer = build_trainer( name="MARWIL", default_config=DEFAULT_CONFIG, default_policy=MARWILPolicy, validate_config=validate_config, make_policy_optimizer=make_optimizer )
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/marwil/marwil_policy.py
Python
import ray from ray.rllib.models import ModelCatalog from ray.rllib.evaluation.postprocessing import compute_advantages, \ Postprocessing from ray.rllib.policy.sample_batch import SampleBatch from ray.rllib.evaluation.metrics import LEARNER_STATS_KEY from ray.rllib.utils.annotations import override from ray.rllib.policy.policy import Policy from ray.rllib.policy.tf_policy import TFPolicy from ray.rllib.utils.explained_variance import explained_variance from ray.rllib.utils import try_import_tf from ray.rllib.utils.tf_ops import scope_vars tf = try_import_tf() POLICY_SCOPE = "p_func" VALUE_SCOPE = "v_func" class ValueLoss: def __init__(self, state_values, cumulative_rewards): self.loss = 0.5 * tf.reduce_mean( tf.square(state_values - cumulative_rewards)) class ReweightedImitationLoss: def __init__(self, state_values, cumulative_rewards, logits, actions, action_space, beta, model): ma_adv_norm = tf.get_variable( name="moving_average_of_advantage_norm", dtype=tf.float32, initializer=100.0, trainable=False) # advantage estimation adv = cumulative_rewards - state_values # update averaged advantage norm update_adv_norm = tf.assign_add( ref=ma_adv_norm, value=1e-6 * (tf.reduce_mean(tf.square(adv)) - ma_adv_norm)) # exponentially weighted advantages with tf.control_dependencies([update_adv_norm]): exp_advs = tf.exp( beta * tf.divide(adv, 1e-8 + tf.sqrt(ma_adv_norm))) # log\pi_\theta(a|s) dist_class, _ = ModelCatalog.get_action_dist(action_space, {}) action_dist = dist_class(logits, model) logprobs = action_dist.logp(actions) self.loss = -1.0 * tf.reduce_mean( tf.stop_gradient(exp_advs) * logprobs) class MARWILPostprocessing: """Adds the advantages field to the trajectory.""" @override(Policy) def postprocess_trajectory(self, sample_batch, other_agent_batches=None, episode=None): completed = sample_batch["dones"][-1] if completed: last_r = 0.0 else: raise NotImplementedError( "last done mask in a batch should be True. " "For now, we only support reading experience batches produced " "with batch_mode='complete_episodes'.", len(sample_batch[SampleBatch.DONES]), sample_batch[SampleBatch.DONES][-1]) batch = compute_advantages( sample_batch, last_r, gamma=self.config["gamma"], use_gae=False) return batch class MARWILPolicy(MARWILPostprocessing, TFPolicy): def __init__(self, observation_space, action_space, config): config = dict(ray.rllib.agents.dqn.dqn.DEFAULT_CONFIG, **config) self.config = config dist_class, logit_dim = ModelCatalog.get_action_dist( action_space, self.config["model"]) # Action inputs self.obs_t = tf.placeholder( tf.float32, shape=(None, ) + observation_space.shape) prev_actions_ph = ModelCatalog.get_action_placeholder(action_space) prev_rewards_ph = tf.placeholder( tf.float32, [None], name="prev_reward") with tf.variable_scope(POLICY_SCOPE) as scope: self.model = ModelCatalog.get_model({ "obs": self.obs_t, "prev_actions": prev_actions_ph, "prev_rewards": prev_rewards_ph, "is_training": self._get_is_training_placeholder(), }, observation_space, action_space, logit_dim, self.config["model"]) logits = self.model.outputs self.p_func_vars = scope_vars(scope.name) # Action outputs action_dist = dist_class(logits, self.model) self.output_actions = action_dist.sample() # Training inputs self.act_t = ModelCatalog.get_action_placeholder(action_space) self.cum_rew_t = tf.placeholder(tf.float32, [None], name="reward") # v network evaluation with tf.variable_scope(VALUE_SCOPE) as scope: state_values = self.model.value_function() self.v_func_vars = scope_vars(scope.name) self.v_loss = self._build_value_loss(state_values, self.cum_rew_t) self.p_loss = self._build_policy_loss(state_values, self.cum_rew_t, logits, self.act_t, action_space) # which kind of objective to optimize objective = ( self.p_loss.loss + self.config["vf_coeff"] * self.v_loss.loss) self.explained_variance = tf.reduce_mean( explained_variance(self.cum_rew_t, state_values)) # initialize TFPolicy self.sess = tf.get_default_session() self.loss_inputs = [ (SampleBatch.CUR_OBS, self.obs_t), (SampleBatch.ACTIONS, self.act_t), (Postprocessing.ADVANTAGES, self.cum_rew_t), ] TFPolicy.__init__( self, observation_space, action_space, self.config, self.sess, obs_input=self.obs_t, action_sampler=self.output_actions, action_logp=action_dist.sampled_action_logp(), loss=objective, model=self.model, loss_inputs=self.loss_inputs, state_inputs=self.model.state_in, state_outputs=self.model.state_out, prev_action_input=prev_actions_ph, prev_reward_input=prev_rewards_ph) self.sess.run(tf.global_variables_initializer()) self.stats_fetches = { "total_loss": objective, "vf_explained_var": self.explained_variance, "policy_loss": self.p_loss.loss, "vf_loss": self.v_loss.loss } def _build_value_loss(self, state_values, cum_rwds): return ValueLoss(state_values, cum_rwds) def _build_policy_loss(self, state_values, cum_rwds, logits, actions, action_space): return ReweightedImitationLoss(state_values, cum_rwds, logits, actions, action_space, self.config["beta"], self.model) @override(TFPolicy) def extra_compute_grad_fetches(self): return {LEARNER_STATS_KEY: self.stats_fetches} @override(Policy) def get_initial_state(self): return self.model.state_init
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/mock.py
Python
import os import pickle import numpy as np from ray.tune import result as tune_result from ray.rllib.agents.trainer import Trainer, with_common_config class _MockTrainer(Trainer): """Mock trainer for use in tests""" _name = "MockTrainer" _default_config = with_common_config({ "mock_error": False, "persistent_error": False, "test_variable": 1, "num_workers": 0, "user_checkpoint_freq": 0, "object_store_memory_per_worker": 0, "object_store_memory": 0, }) @classmethod def default_resource_request(cls, config): return None def _init(self, config, env_creator): self.info = None self.restored = False def _train(self): if self.config["mock_error"] and self.iteration == 1 \ and (self.config["persistent_error"] or not self.restored): raise Exception("mock error") result = dict( episode_reward_mean=10, episode_len_mean=10, timesteps_this_iter=10, info={}) if self.config["user_checkpoint_freq"] > 0 and self.iteration > 0: if self.iteration % self.config["user_checkpoint_freq"] == 0: result.update({tune_result.SHOULD_CHECKPOINT: True}) return result def _save(self, checkpoint_dir): path = os.path.join(checkpoint_dir, "mock_agent.pkl") with open(path, "wb") as f: pickle.dump(self.info, f) return path def _restore(self, checkpoint_path): with open(checkpoint_path, "rb") as f: info = pickle.load(f) self.info = info self.restored = True def _register_if_needed(self, env_object): pass def set_info(self, info): self.info = info return info def get_info(self): return self.info class _SigmoidFakeData(_MockTrainer): """Trainer that returns sigmoid learning curves. This can be helpful for evaluating early stopping algorithms.""" _name = "SigmoidFakeData" _default_config = with_common_config({ "width": 100, "height": 100, "offset": 0, "iter_time": 10, "iter_timesteps": 1, "num_workers": 0, "object_store_memory_per_worker": 0, "object_store_memory": 0, }) def _train(self): i = max(0, self.iteration - self.config["offset"]) v = np.tanh(float(i) / self.config["width"]) v *= self.config["height"] return dict( episode_reward_mean=v, episode_len_mean=v, timesteps_this_iter=self.config["iter_timesteps"], time_this_iter_s=self.config["iter_time"], info={}) class _ParameterTuningTrainer(_MockTrainer): _name = "ParameterTuningTrainer" _default_config = with_common_config({ "reward_amt": 10, "dummy_param": 10, "dummy_param2": 15, "iter_time": 10, "iter_timesteps": 1, "num_workers": 0, "object_store_memory_per_worker": 0, "object_store_memory": 0, }) def _train(self): return dict( episode_reward_mean=self.config["reward_amt"] * self.iteration, episode_len_mean=self.config["reward_amt"], timesteps_this_iter=self.config["iter_timesteps"], time_this_iter_s=self.config["iter_time"], info={}) def _agent_import_failed(trace): """Returns dummy agent class for if PyTorch etc. is not installed.""" class _AgentImportFailed(Trainer): _name = "AgentImportFailed" _default_config = with_common_config({}) def _setup(self, config): raise ImportError(trace) return _AgentImportFailed
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/pg/__init__.py
Python
from ray.rllib.agents.pg.pg import PGTrainer, DEFAULT_CONFIG from ray.rllib.agents.pg.pg_tf_policy import pg_tf_loss, \ post_process_advantages from ray.rllib.agents.pg.pg_torch_policy import pg_torch_loss __all__ = ["PGTrainer", "pg_tf_loss", "pg_torch_loss", "post_process_advantages", "DEFAULT_CONFIG"]
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/pg/pg.py
Python
from ray.rllib.agents.trainer import with_common_config from ray.rllib.agents.trainer_template import build_trainer from ray.rllib.agents.pg.pg_tf_policy import PGTFPolicy # yapf: disable # __sphinx_doc_begin__ DEFAULT_CONFIG = with_common_config({ # No remote workers by default. "num_workers": 0, # Learning rate. "lr": 0.0004, }) # __sphinx_doc_end__ # yapf: enable def get_policy_class(config): if config["use_pytorch"]: from ray.rllib.agents.pg.pg_torch_policy import PGTorchPolicy return PGTorchPolicy else: return PGTFPolicy PGTrainer = build_trainer( name="PG", default_config=DEFAULT_CONFIG, default_policy=PGTFPolicy, get_policy_class=get_policy_class)
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/pg/pg_tf_policy.py
Python
import ray from ray.rllib.evaluation.postprocessing import Postprocessing, \ compute_advantages from ray.rllib.policy.tf_policy_template import build_tf_policy from ray.rllib.policy.sample_batch import SampleBatch from ray.rllib.utils import try_import_tf tf = try_import_tf() def post_process_advantages(policy, sample_batch, other_agent_batches=None, episode=None): """This adds the "advantages" column to the sample train_batch.""" return compute_advantages(sample_batch, 0.0, policy.config["gamma"], use_gae=False) def pg_tf_loss(policy, model, dist_class, train_batch): """The basic policy gradients loss.""" logits, _ = model.from_batch(train_batch) action_dist = dist_class(logits, model) return -tf.reduce_mean(action_dist.logp(train_batch[SampleBatch.ACTIONS]) * train_batch[Postprocessing.ADVANTAGES]) PGTFPolicy = build_tf_policy( name="PGTFPolicy", get_default_config=lambda: ray.rllib.agents.pg.pg.DEFAULT_CONFIG, postprocess_fn=post_process_advantages, loss_fn=pg_tf_loss)
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/pg/pg_torch_policy.py
Python
import ray from ray.rllib.agents.pg.pg_tf_policy import post_process_advantages from ray.rllib.evaluation.postprocessing import Postprocessing from ray.rllib.policy.sample_batch import SampleBatch from ray.rllib.policy.torch_policy_template import build_torch_policy from ray.rllib.utils.framework import try_import_torch torch, _ = try_import_torch() def pg_torch_loss(policy, model, dist_class, train_batch): """The basic policy gradients loss.""" logits, _ = model.from_batch(train_batch) action_dist = dist_class(logits, model) log_probs = action_dist.logp(train_batch[SampleBatch.ACTIONS]) # Save the error in the policy object. # policy.pi_err = -train_batch[Postprocessing.ADVANTAGES].dot( # log_probs.reshape(-1)) / len(log_probs) policy.pi_err = -torch.mean( log_probs * train_batch[Postprocessing.ADVANTAGES] ) return policy.pi_err def pg_loss_stats(policy, train_batch): """ The error is recorded when computing the loss.""" return {"policy_loss": policy.pi_err.item()} PGTorchPolicy = build_torch_policy( name="PGTorchPolicy", get_default_config=lambda: ray.rllib.agents.pg.pg.DEFAULT_CONFIG, loss_fn=pg_torch_loss, stats_fn=pg_loss_stats, postprocess_fn=post_process_advantages)
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/pg/tests/test_pg.py
Python
import numpy as np import unittest import ray import ray.rllib.agents.pg as pg from ray.rllib.evaluation.postprocessing import Postprocessing from ray.rllib.models.tf.tf_action_dist import Categorical from ray.rllib.models.torch.torch_action_dist import TorchCategorical from ray.rllib.policy.sample_batch import SampleBatch from ray.rllib.utils import check, fc class TestPG(unittest.TestCase): ray.init() def test_pg_compilation(self): """Test whether a PGTrainer can be built with both frameworks.""" config = pg.DEFAULT_CONFIG.copy() config["num_workers"] = 0 # Run locally. # tf. trainer = pg.PGTrainer(config=config, env="CartPole-v0") num_iterations = 2 for i in range(num_iterations): trainer.train() # Torch. config["use_pytorch"] = True trainer = pg.PGTrainer(config=config, env="CartPole-v0") for i in range(num_iterations): trainer.train() def test_pg_loss_functions(self): """Tests the PG loss function math.""" config = pg.DEFAULT_CONFIG.copy() config["num_workers"] = 0 # Run locally. config["eager"] = True config["gamma"] = 0.99 config["model"]["fcnet_hiddens"] = [10] config["model"]["fcnet_activation"] = "linear" # Fake CartPole episode of n timesteps. train_batch = { SampleBatch.CUR_OBS: np.array([ [0.1, 0.2, 0.3, 0.4], [0.5, 0.6, 0.7, 0.8], [0.9, 1.0, 1.1, 1.2] ]), SampleBatch.ACTIONS: np.array([0, 1, 1]), SampleBatch.REWARDS: np.array([1.0, 1.0, 1.0]), SampleBatch.DONES: np.array([False, False, True]) } # tf. trainer = pg.PGTrainer(config=config, env="CartPole-v0") policy = trainer.get_policy() vars = policy.model.trainable_variables() # Post-process (calculate simple (non-GAE) advantages) and attach to # train_batch dict. # A = [0.99^2 * 1.0 + 0.99 * 1.0 + 1.0, 0.99 * 1.0 + 1.0, 1.0] = # [2.9701, 1.99, 1.0] train_batch = pg.post_process_advantages(policy, train_batch) # Check Advantage values. check(train_batch[Postprocessing.ADVANTAGES], [2.9701, 1.99, 1.0]) # Actual loss results. results = pg.pg_tf_loss( policy, policy.model, dist_class=Categorical, train_batch=train_batch ) # Calculate expected results. expected_logits = fc( fc( train_batch[SampleBatch.CUR_OBS], vars[0].numpy(), vars[1].numpy() ), vars[2].numpy(), vars[3].numpy() ) expected_logp = Categorical(expected_logits, policy.model).logp( train_batch[SampleBatch.ACTIONS] ) expected_loss = -np.mean( expected_logp * train_batch[Postprocessing.ADVANTAGES] ) check(results.numpy(), expected_loss, decimals=4) # Torch. config["use_pytorch"] = True trainer = pg.PGTrainer(config=config, env="CartPole-v0") policy = trainer.get_policy() train_batch = policy._lazy_tensor_dict(train_batch) results = pg.pg_torch_loss( policy, policy.model, dist_class=TorchCategorical, train_batch=train_batch ) expected_logits = policy.model._last_output expected_logp = TorchCategorical(expected_logits, policy.model).logp( train_batch[SampleBatch.ACTIONS] ) expected_loss = -np.mean( expected_logp.detach().numpy() * train_batch[Postprocessing.ADVANTAGES].numpy() ) check(results.detach().numpy(), expected_loss, decimals=4)
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/ppo/__init__.py
Python
from ray.rllib.agents.ppo.ppo import PPOTrainer, DEFAULT_CONFIG from ray.rllib.agents.ppo.appo import APPOTrainer from ray.rllib.utils import renamed_agent PPOAgent = renamed_agent(PPOTrainer) __all__ = ["PPOAgent", "APPOTrainer", "PPOTrainer", "DEFAULT_CONFIG"]
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/ppo/appo.py
Python
from ray.rllib.agents.ppo.appo_policy import AsyncPPOTFPolicy from ray.rllib.agents.trainer import with_base_config from ray.rllib.agents.ppo.ppo import update_kl from ray.rllib.agents import impala # yapf: disable # __sphinx_doc_begin__ DEFAULT_CONFIG = with_base_config(impala.DEFAULT_CONFIG, { # Whether to use V-trace weighted advantages. If false, PPO GAE advantages # will be used instead. "vtrace": False, # == These two options only apply if vtrace: False == # If true, use the Generalized Advantage Estimator (GAE) # with a value function, see https://arxiv.org/pdf/1506.02438.pdf. "use_gae": True, # GAE(lambda) parameter "lambda": 1.0, # == PPO surrogate loss options == "clip_param": 0.4, # == PPO KL Loss options == "use_kl_loss": False, "kl_coeff": 1.0, "kl_target": 0.01, # == IMPALA optimizer params (see documentation in impala.py) == "sample_batch_size": 50, "train_batch_size": 500, "min_iter_time_s": 10, "num_workers": 2, "num_gpus": 0, "num_data_loader_buffers": 1, "minibatch_buffer_size": 1, "num_sgd_iter": 1, "replay_proportion": 0.0, "replay_buffer_num_slots": 100, "learner_queue_size": 16, "learner_queue_timeout": 300, "max_sample_requests_in_flight_per_worker": 2, "broadcast_interval": 1, "grad_clip": 40.0, "opt_type": "adam", "lr": 0.0005, "lr_schedule": None, "decay": 0.99, "momentum": 0.0, "epsilon": 0.1, "vf_loss_coeff": 0.5, "entropy_coeff": 0.01, "entropy_coeff_schedule": None, }) # __sphinx_doc_end__ # yapf: enable def update_target_and_kl(trainer, fetches): # Update the KL coeff depending on how many steps LearnerThread has stepped # through learner_steps = trainer.optimizer.learner.num_steps if learner_steps >= trainer.target_update_frequency: # Update Target Network trainer.optimizer.learner.num_steps = 0 trainer.workers.local_worker().foreach_trainable_policy( lambda p, _: p.update_target()) # Also update KL Coeff if trainer.config["use_kl_loss"]: update_kl(trainer, trainer.optimizer.learner.stats) def initialize_target(trainer): trainer.workers.local_worker().foreach_trainable_policy( lambda p, _: p.update_target()) trainer.target_update_frequency = trainer.config["num_sgd_iter"] \ * trainer.config["minibatch_buffer_size"] APPOTrainer = impala.ImpalaTrainer.with_updates( name="APPO", default_config=DEFAULT_CONFIG, default_policy=AsyncPPOTFPolicy, get_policy_class=lambda _: AsyncPPOTFPolicy, after_init=initialize_target, after_optimizer_step=update_target_and_kl)
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/ppo/appo_policy.py
Python
"""Adapted from VTraceTFPolicy to use the PPO surrogate loss. Keep in sync with changes to VTraceTFPolicy.""" import numpy as np import logging import gym from ray.rllib.agents.impala import vtrace from ray.rllib.agents.impala.vtrace_policy import _make_time_major, \ BEHAVIOUR_LOGITS, clip_gradients, validate_config, choose_optimizer from ray.rllib.evaluation.postprocessing import Postprocessing from ray.rllib.models.tf.tf_action_dist import Categorical from ray.rllib.policy.sample_batch import SampleBatch from ray.rllib.evaluation.postprocessing import compute_advantages from ray.rllib.utils import try_import_tf from ray.rllib.policy.tf_policy_template import build_tf_policy from ray.rllib.policy.tf_policy import LearningRateSchedule, TFPolicy from ray.rllib.agents.ppo.ppo_policy import KLCoeffMixin, ValueNetworkMixin from ray.rllib.models import ModelCatalog from ray.rllib.utils.annotations import override from ray.rllib.utils.explained_variance import explained_variance from ray.rllib.utils.tf_ops import make_tf_callable tf = try_import_tf() POLICY_SCOPE = "func" TARGET_POLICY_SCOPE = "target_func" logger = logging.getLogger(__name__) class PPOSurrogateLoss: """Loss used when V-trace is disabled. Arguments: prev_actions_logp: A float32 tensor of shape [T, B]. actions_logp: A float32 tensor of shape [T, B]. action_kl: A float32 tensor of shape [T, B]. actions_entropy: A float32 tensor of shape [T, B]. values: A float32 tensor of shape [T, B]. valid_mask: A bool tensor of valid RNN input elements (#2992). advantages: A float32 tensor of shape [T, B]. value_targets: A float32 tensor of shape [T, B]. vf_loss_coeff (float): Coefficient of the value function loss. entropy_coeff (float): Coefficient of the entropy regularizer. clip_param (float): Clip parameter. cur_kl_coeff (float): Coefficient for KL loss. use_kl_loss (bool): If true, use KL loss. """ def __init__(self, prev_actions_logp, actions_logp, action_kl, actions_entropy, values, valid_mask, advantages, value_targets, vf_loss_coeff=0.5, entropy_coeff=0.01, clip_param=0.3, cur_kl_coeff=None, use_kl_loss=False): def reduce_mean_valid(t): return tf.reduce_mean(tf.boolean_mask(t, valid_mask)) logp_ratio = tf.exp(actions_logp - prev_actions_logp) surrogate_loss = tf.minimum( advantages * logp_ratio, advantages * tf.clip_by_value(logp_ratio, 1 - clip_param, 1 + clip_param)) self.mean_kl = reduce_mean_valid(action_kl) self.pi_loss = -reduce_mean_valid(surrogate_loss) # The baseline loss delta = values - value_targets self.value_targets = value_targets self.vf_loss = 0.5 * reduce_mean_valid(tf.square(delta)) # The entropy loss self.entropy = reduce_mean_valid(actions_entropy) # The summed weighted loss self.total_loss = (self.pi_loss + self.vf_loss * vf_loss_coeff - self.entropy * entropy_coeff) # Optional additional KL Loss if use_kl_loss: self.total_loss += cur_kl_coeff * self.mean_kl class VTraceSurrogateLoss: def __init__(self, actions, prev_actions_logp, actions_logp, old_policy_actions_logp, action_kl, actions_entropy, dones, behaviour_logits, old_policy_behaviour_logits, target_logits, discount, rewards, values, bootstrap_value, dist_class, model, valid_mask, vf_loss_coeff=0.5, entropy_coeff=0.01, clip_rho_threshold=1.0, clip_pg_rho_threshold=1.0, clip_param=0.3, cur_kl_coeff=None, use_kl_loss=False): """APPO Loss, with IS modifications and V-trace for Advantage Estimation VTraceLoss takes tensors of shape [T, B, ...], where `B` is the batch_size. The reason we need to know `B` is for V-trace to properly handle episode cut boundaries. Arguments: actions: An int|float32 tensor of shape [T, B, logit_dim]. prev_actions_logp: A float32 tensor of shape [T, B]. actions_logp: A float32 tensor of shape [T, B]. old_policy_actions_logp: A float32 tensor of shape [T, B]. action_kl: A float32 tensor of shape [T, B]. actions_entropy: A float32 tensor of shape [T, B]. dones: A bool tensor of shape [T, B]. behaviour_logits: A float32 tensor of shape [T, B, logit_dim]. old_policy_behaviour_logits: A float32 tensor of shape [T, B, logit_dim]. target_logits: A float32 tensor of shape [T, B, logit_dim]. discount: A float32 scalar. rewards: A float32 tensor of shape [T, B]. values: A float32 tensor of shape [T, B]. bootstrap_value: A float32 tensor of shape [B]. dist_class: action distribution class for logits. model: backing ModelV2 instance valid_mask: A bool tensor of valid RNN input elements (#2992). vf_loss_coeff (float): Coefficient of the value function loss. entropy_coeff (float): Coefficient of the entropy regularizer. clip_param (float): Clip parameter. cur_kl_coeff (float): Coefficient for KL loss. use_kl_loss (bool): If true, use KL loss. """ def reduce_mean_valid(t): return tf.reduce_mean(tf.boolean_mask(t, valid_mask)) # Compute vtrace on the CPU for better perf. with tf.device("/cpu:0"): self.vtrace_returns = vtrace.multi_from_logits( behaviour_policy_logits=behaviour_logits, target_policy_logits=old_policy_behaviour_logits, actions=tf.unstack(actions, axis=2), discounts=tf.to_float(~dones) * discount, rewards=rewards, values=values, bootstrap_value=bootstrap_value, dist_class=dist_class, model=model, clip_rho_threshold=tf.cast(clip_rho_threshold, tf.float32), clip_pg_rho_threshold=tf.cast(clip_pg_rho_threshold, tf.float32)) self.is_ratio = tf.clip_by_value( tf.exp(prev_actions_logp - old_policy_actions_logp), 0.0, 2.0) logp_ratio = self.is_ratio * tf.exp(actions_logp - prev_actions_logp) advantages = self.vtrace_returns.pg_advantages surrogate_loss = tf.minimum( advantages * logp_ratio, advantages * tf.clip_by_value(logp_ratio, 1 - clip_param, 1 + clip_param)) self.mean_kl = reduce_mean_valid(action_kl) self.pi_loss = -reduce_mean_valid(surrogate_loss) # The baseline loss delta = values - self.vtrace_returns.vs self.value_targets = self.vtrace_returns.vs self.vf_loss = 0.5 * reduce_mean_valid(tf.square(delta)) # The entropy loss self.entropy = reduce_mean_valid(actions_entropy) # The summed weighted loss self.total_loss = (self.pi_loss + self.vf_loss * vf_loss_coeff - self.entropy * entropy_coeff) # Optional additional KL Loss if use_kl_loss: self.total_loss += cur_kl_coeff * self.mean_kl def build_appo_model(policy, obs_space, action_space, config): _, logit_dim = ModelCatalog.get_action_dist(action_space, config["model"]) policy.model = ModelCatalog.get_model_v2( obs_space, action_space, logit_dim, config["model"], name=POLICY_SCOPE, framework="tf") policy.target_model = ModelCatalog.get_model_v2( obs_space, action_space, logit_dim, config["model"], name=TARGET_POLICY_SCOPE, framework="tf") return policy.model def build_appo_surrogate_loss(policy, model, dist_class, train_batch): model_out, _ = model.from_batch(train_batch) action_dist = dist_class(model_out, model) if isinstance(policy.action_space, gym.spaces.Discrete): is_multidiscrete = False output_hidden_shape = [policy.action_space.n] elif isinstance(policy.action_space, gym.spaces.multi_discrete.MultiDiscrete): is_multidiscrete = True output_hidden_shape = policy.action_space.nvec.astype(np.int32) else: is_multidiscrete = False output_hidden_shape = 1 def make_time_major(*args, **kw): return _make_time_major(policy, train_batch.get("seq_lens"), *args, **kw) actions = train_batch[SampleBatch.ACTIONS] dones = train_batch[SampleBatch.DONES] rewards = train_batch[SampleBatch.REWARDS] behaviour_logits = train_batch[BEHAVIOUR_LOGITS] target_model_out, _ = policy.target_model.from_batch(train_batch) old_policy_behaviour_logits = tf.stop_gradient(target_model_out) unpacked_behaviour_logits = tf.split( behaviour_logits, output_hidden_shape, axis=1) unpacked_old_policy_behaviour_logits = tf.split( old_policy_behaviour_logits, output_hidden_shape, axis=1) unpacked_outputs = tf.split(model_out, output_hidden_shape, axis=1) old_policy_action_dist = dist_class(old_policy_behaviour_logits, model) prev_action_dist = dist_class(behaviour_logits, policy.model) values = policy.model.value_function() policy.model_vars = policy.model.variables() policy.target_model_vars = policy.target_model.variables() if policy.is_recurrent(): max_seq_len = tf.reduce_max(train_batch["seq_lens"]) - 1 mask = tf.sequence_mask(train_batch["seq_lens"], max_seq_len) mask = tf.reshape(mask, [-1]) else: mask = tf.ones_like(rewards) if policy.config["vtrace"]: logger.debug("Using V-Trace surrogate loss (vtrace=True)") # Prepare actions for loss loss_actions = actions if is_multidiscrete else tf.expand_dims( actions, axis=1) # Prepare KL for Loss mean_kl = make_time_major( old_policy_action_dist.multi_kl(action_dist), drop_last=True) policy.loss = VTraceSurrogateLoss( actions=make_time_major(loss_actions, drop_last=True), prev_actions_logp=make_time_major( prev_action_dist.logp(actions), drop_last=True), actions_logp=make_time_major( action_dist.logp(actions), drop_last=True), old_policy_actions_logp=make_time_major( old_policy_action_dist.logp(actions), drop_last=True), action_kl=tf.reduce_mean(mean_kl, axis=0) if is_multidiscrete else mean_kl, actions_entropy=make_time_major( action_dist.multi_entropy(), drop_last=True), dones=make_time_major(dones, drop_last=True), behaviour_logits=make_time_major( unpacked_behaviour_logits, drop_last=True), old_policy_behaviour_logits=make_time_major( unpacked_old_policy_behaviour_logits, drop_last=True), target_logits=make_time_major(unpacked_outputs, drop_last=True), discount=policy.config["gamma"], rewards=make_time_major(rewards, drop_last=True), values=make_time_major(values, drop_last=True), bootstrap_value=make_time_major(values)[-1], dist_class=Categorical if is_multidiscrete else dist_class, model=policy.model, valid_mask=make_time_major(mask, drop_last=True), vf_loss_coeff=policy.config["vf_loss_coeff"], entropy_coeff=policy.config["entropy_coeff"], clip_rho_threshold=policy.config["vtrace_clip_rho_threshold"], clip_pg_rho_threshold=policy.config[ "vtrace_clip_pg_rho_threshold"], clip_param=policy.config["clip_param"], cur_kl_coeff=policy.kl_coeff, use_kl_loss=policy.config["use_kl_loss"]) else: logger.debug("Using PPO surrogate loss (vtrace=False)") # Prepare KL for Loss mean_kl = make_time_major(prev_action_dist.multi_kl(action_dist)) policy.loss = PPOSurrogateLoss( prev_actions_logp=make_time_major(prev_action_dist.logp(actions)), actions_logp=make_time_major(action_dist.logp(actions)), action_kl=tf.reduce_mean(mean_kl, axis=0) if is_multidiscrete else mean_kl, actions_entropy=make_time_major(action_dist.multi_entropy()), values=make_time_major(values), valid_mask=make_time_major(mask), advantages=make_time_major(train_batch[Postprocessing.ADVANTAGES]), value_targets=make_time_major( train_batch[Postprocessing.VALUE_TARGETS]), vf_loss_coeff=policy.config["vf_loss_coeff"], entropy_coeff=policy.config["entropy_coeff"], clip_param=policy.config["clip_param"], cur_kl_coeff=policy.kl_coeff, use_kl_loss=policy.config["use_kl_loss"]) return policy.loss.total_loss def stats(policy, train_batch): values_batched = _make_time_major( policy, train_batch.get("seq_lens"), policy.model.value_function(), drop_last=policy.config["vtrace"]) stats_dict = { "cur_lr": tf.cast(policy.cur_lr, tf.float64), "policy_loss": policy.loss.pi_loss, "entropy": policy.loss.entropy, "var_gnorm": tf.global_norm(policy.model.trainable_variables()), "vf_loss": policy.loss.vf_loss, "vf_explained_var": explained_variance( tf.reshape(policy.loss.value_targets, [-1]), tf.reshape(values_batched, [-1])), } if policy.config["vtrace"]: is_stat_mean, is_stat_var = tf.nn.moments(policy.loss.is_ratio, [0, 1]) stats_dict.update({"mean_IS": is_stat_mean}) stats_dict.update({"var_IS": is_stat_var}) if policy.config["use_kl_loss"]: stats_dict.update({"kl": policy.loss.mean_kl}) stats_dict.update({"KL_Coeff": policy.kl_coeff}) return stats_dict def postprocess_trajectory(policy, sample_batch, other_agent_batches=None, episode=None): if not policy.config["vtrace"]: completed = sample_batch["dones"][-1] if completed: last_r = 0.0 else: next_state = [] for i in range(policy.num_state_tensors()): next_state.append([sample_batch["state_out_{}".format(i)][-1]]) last_r = policy._value(sample_batch[SampleBatch.NEXT_OBS][-1], sample_batch[SampleBatch.ACTIONS][-1], sample_batch[SampleBatch.REWARDS][-1], *next_state) batch = compute_advantages( sample_batch, last_r, policy.config["gamma"], policy.config["lambda"], use_gae=policy.config["use_gae"]) else: batch = sample_batch del batch.data["new_obs"] # not used, so save some bandwidth return batch def add_values_and_logits(policy): out = {BEHAVIOUR_LOGITS: policy.model.last_output()} if not policy.config["vtrace"]: out[SampleBatch.VF_PREDS] = policy.model.value_function() return out class TargetNetworkMixin: def __init__(self, obs_space, action_space, config): """Target Network is updated by the master learner every trainer.update_target_frequency steps. All worker batches are importance sampled w.r. to the target network to ensure a more stable pi_old in PPO. """ @make_tf_callable(self.get_session()) def do_update(): assign_ops = [] assert len(self.model_vars) == len(self.target_model_vars) for var, var_target in zip(self.model_vars, self.target_model_vars): assign_ops.append(var_target.assign(var)) return tf.group(*assign_ops) self.update_target = do_update @override(TFPolicy) def variables(self): return self.model_vars + self.target_model_vars def setup_mixins(policy, obs_space, action_space, config): LearningRateSchedule.__init__(policy, config["lr"], config["lr_schedule"]) KLCoeffMixin.__init__(policy, config) ValueNetworkMixin.__init__(policy, obs_space, action_space, config) def setup_late_mixins(policy, obs_space, action_space, config): TargetNetworkMixin.__init__(policy, obs_space, action_space, config) AsyncPPOTFPolicy = build_tf_policy( name="AsyncPPOTFPolicy", make_model=build_appo_model, loss_fn=build_appo_surrogate_loss, stats_fn=stats, postprocess_fn=postprocess_trajectory, optimizer_fn=choose_optimizer, gradients_fn=clip_gradients, extra_action_fetches_fn=add_values_and_logits, before_init=validate_config, before_loss_init=setup_mixins, after_init=setup_late_mixins, mixins=[ LearningRateSchedule, KLCoeffMixin, TargetNetworkMixin, ValueNetworkMixin ], get_batch_divisibility_req=lambda p: p.config["sample_batch_size"])
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta
rllib/agents/ppo/ppo.py
Python
import logging from ray.rllib.agents import with_common_config from ray.rllib.agents.ppo.ppo_policy import PPOTFPolicy from ray.rllib.agents.trainer_template import build_trainer from ray.rllib.optimizers import SyncSamplesOptimizer, LocalMultiGPUOptimizer from ray.rllib.utils import try_import_tf tf = try_import_tf() logger = logging.getLogger(__name__) # yapf: disable # __sphinx_doc_begin__ DEFAULT_CONFIG = with_common_config({ # If true, use the Generalized Advantage Estimator (GAE) # with a value function, see https://arxiv.org/pdf/1506.02438.pdf. "use_gae": True, # The GAE(lambda) parameter. "lambda": 1.0, # Initial coefficient for KL divergence. "kl_coeff": 0.2, # Size of batches collected from each worker. "sample_batch_size": 200, # Number of timesteps collected for each SGD round. This defines the size # of each SGD epoch. "train_batch_size": 4000, # Total SGD batch size across all devices for SGD. This defines the # minibatch size within each epoch. "sgd_minibatch_size": 128, # Whether to shuffle sequences in the batch when training (recommended). "shuffle_sequences": True, # Number of SGD iterations in each outer loop (i.e., number of epochs to # execute per train batch). "num_sgd_iter": 30, # Stepsize of SGD. "lr": 5e-5, # Learning rate schedule. "lr_schedule": None, # Share layers for value function. If you set this to True, it's important # to tune vf_loss_coeff. "vf_share_layers": False, # Coefficient of the value function loss. IMPORTANT: you must tune this if # you set vf_share_layers: True. "vf_loss_coeff": 1.0, # Coefficient of the entropy regularizer. "entropy_coeff": 0.0, # Decay schedule for the entropy regularizer. "entropy_coeff_schedule": None, # PPO clip parameter. "clip_param": 0.3, # Clip param for the value function. Note that this is sensitive to the # scale of the rewards. If your expected V is large, increase this. "vf_clip_param": 10.0, # If specified, clip the global norm of gradients by this amount. "grad_clip": None, # Target value for KL divergence. "kl_target": 0.01, # Whether to rollout "complete_episodes" or "truncate_episodes". "batch_mode": "truncate_episodes", # Which observation filter to apply to the observation. "observation_filter": "NoFilter", # Uses the sync samples optimizer instead of the multi-gpu one. This is # usually slower, but you might want to try it if you run into issues with # the default optimizer. "simple_optimizer": False, }) # __sphinx_doc_end__ # yapf: enable def choose_policy_optimizer(workers, config): if config["simple_optimizer"]: return SyncSamplesOptimizer( workers, num_sgd_iter=config["num_sgd_iter"], train_batch_size=config["train_batch_size"], sgd_minibatch_size=config["sgd_minibatch_size"], standardize_fields=["advantages"]) return LocalMultiGPUOptimizer( workers, sgd_batch_size=config["sgd_minibatch_size"], num_sgd_iter=config["num_sgd_iter"], num_gpus=config["num_gpus"], sample_batch_size=config["sample_batch_size"], num_envs_per_worker=config["num_envs_per_worker"], train_batch_size=config["train_batch_size"], standardize_fields=["advantages"], shuffle_sequences=config["shuffle_sequences"]) def update_kl(trainer, fetches): if "kl" in fetches: # single-agent trainer.workers.local_worker().for_policy( lambda pi: pi.update_kl(fetches["kl"])) else: def update(pi, pi_id): if pi_id in fetches: pi.update_kl(fetches[pi_id]["kl"]) else: logger.debug("No data for {}, not updating kl".format(pi_id)) # multi-agent trainer.workers.local_worker().foreach_trainable_policy(update) def warn_about_bad_reward_scales(trainer, result): if result["policy_reward_mean"]: return # Punt on handling multiagent case. # Warn about excessively high VF loss. learner_stats = result["info"]["learner"] if "default_policy" in learner_stats: scaled_vf_loss = (trainer.config["vf_loss_coeff"] * learner_stats["default_policy"]["vf_loss"]) policy_loss = learner_stats["default_policy"]["policy_loss"] if trainer.config["vf_share_layers"] and scaled_vf_loss > 100: logger.warning( "The magnitude of your value function loss is extremely large " "({}) compared to the policy loss ({}). This can prevent the " "policy from learning. Consider scaling down the VF loss by " "reducing vf_loss_coeff, or disabling vf_share_layers.".format( scaled_vf_loss, policy_loss)) # Warn about bad clipping configs if trainer.config["vf_clip_param"] <= 0: rew_scale = float("inf") else: rew_scale = round( abs(result["episode_reward_mean"]) / trainer.config["vf_clip_param"], 0) if rew_scale > 200: logger.warning( "The magnitude of your environment rewards are more than " "{}x the scale of `vf_clip_param`. ".format(rew_scale) + "This means that it will take more than " "{} iterations for your value ".format(rew_scale) + "function to converge. If this is not intended, consider " "increasing `vf_clip_param`." ) def validate_config(config): # PyTorch check. if config["use_pytorch"]: raise ValueError("PPO does not support PyTorch yet! Use tf instead.") if config["entropy_coeff"] < 0: raise DeprecationWarning("entropy_coeff must be >= 0") if isinstance(config["entropy_coeff"], int): config["entropy_coeff"] = float(config["entropy_coeff"]) if config["sgd_minibatch_size"] > config["train_batch_size"]: raise ValueError( "Minibatch size {} must be <= train batch size {}.". format(config["sgd_minibatch_size"], config["train_batch_size"]) ) if config["batch_mode"] == "truncate_episodes" and not config["use_gae"]: raise ValueError( "Episode truncation is not supported without a value " "function. Consider setting batch_mode=complete_episodes.") if config["multiagent"]["policies"] and not config["simple_optimizer"]: logger.info( "In multi-agent mode, policies will be optimized sequentially " "by the multi-GPU optimizer. Consider setting " "simple_optimizer=True if this doesn't work for you.") if config["simple_optimizer"]: logger.warning( "Using the simple minibatch optimizer. This will significantly " "reduce performance, consider simple_optimizer=False.") elif tf and tf.executing_eagerly(): config["simple_optimizer"] = True # multi-gpu not supported PPOTrainer = build_trainer( name="PPO", default_config=DEFAULT_CONFIG, default_policy=PPOTFPolicy, make_policy_optimizer=choose_policy_optimizer, validate_config=validate_config, after_optimizer_step=update_kl, after_train_result=warn_about_bad_reward_scales)
zhuohan123/hoplite-rllib
3
Python
zhuohan123
Zhuohan Li
vLLM / Meta