sebulba_ppo_envpool.py

"""
0. multi-threaded actor
python sebulba_ppo_envpool.py --actor-device-ids 0 --num-actor-threads 2 --learner-device-ids 1 --params-queue-timeout 0.02 --profile --test-actor-learner-throughput --total-timesteps 500000 --track
python sebulba_ppo_envpool.py --actor-device-ids 0 --learner-device-ids 1 --params-queue-timeout 0.02 --profile --test-actor-learner-throughput --total-timesteps 500000 --track

🔥 core settings:

* test throughput
    * python sebulba_ppo_envpool.py --exp-name sebula_thpt_a0_l1_timeout --actor-device-ids 0 --learner-device-ids 1 --params-queue-timeout 0.02 --profile --test-actor-learner-throughput --total-timesteps 500000 --track
    * python sebulba_ppo_envpool.py --exp-name sebula_thpt_a0_l12_timeout --actor-device-ids 0 --learner-device-ids 1 2 --params-queue-timeout 0.02 --profile --test-actor-learner-throughput --total-timesteps 500000 --track
        * this will help us diagnose the throughput issue
    * python sebulba_ppo_envpool.py --exp-name sebula_thpt_a0_l1 --actor-device-ids 0 --learner-device-ids 1 --profile --total-timesteps 500000 --track
    * python sebulba_ppo_envpool.py --exp-name sebula_thpt_a0_l12 --actor-device-ids 0 --learner-device-ids 1 2  --profile --total-timesteps 500000 --track
    * python sebulba_ppo_envpool.py --actor-device-ids 0 --learner-device-ids 1 2 3 4 --num-actor-threads 2 --track
* Best performance so far
    * python sebulba_ppo_envpool.py --exp-name sebulba_ppo_envpool_a0_l01_rollout_is_faster --actor-device-ids 0 --learner-device-ids 0 1 --total-timesteps 500000 --track
    * python sebulba_ppo_envpool.py --actor-device-ids 0 --learner-device-ids 1 2 3 4 --params-queue-timeout 0.02 --track

# 1. rollout is faster than training

## throughput
python sebulba_ppo_envpool.py --exp-name sebulba_ppo_envpool_thpt_rollout_is_faster --actor-device-ids 0 --learner-device-ids 1 --params-queue-timeout 0.02 --profile --test-actor-learner-throughput --total-timesteps 500000 --track

## shared: actor on GPU0 and learner on GPU0
python sebulba_ppo_envpool.py --exp-name sebulba_ppo_envpool_1gpu_rollout_is_faster --actor-device-ids 0 --learner-device-ids 0 --total-timesteps 500000 --track

## separate: actor on GPU0 and learner on GPU1
python sebulba_ppo_envpool.py --exp-name sebulba_ppo_envpool_a0_l1_rollout_is_faster --actor-device-ids 0 --learner-device-ids 1 --total-timesteps 500000 --track

## shared: actor on GPU0 and learner on GPU0,1
python sebulba_ppo_envpool.py --exp-name sebulba_ppo_envpool_a0_l01_rollout_is_faster --actor-device-ids 0 --learner-device-ids 0 1 --total-timesteps 500000 --track

## separate: actor on GPU0 and learner on GPU1,2
python sebulba_ppo_envpool.py --exp-name sebulba_ppo_envpool_a0_l12_rollout_is_faster --actor-device-ids 0 --learner-device-ids 1 2 --total-timesteps 500000 --track


# 1.1 rollout is faster than training w/ timeout

## shared: actor on GPU0 and learner on GPU0
python sebulba_ppo_envpool.py --exp-name sebulba_ppo_envpool_1gpu_rollout_is_faster_timeout --actor-device-ids 0 --learner-device-ids 0 --params-queue-timeout 0.02 --total-timesteps 500000 --track

## separate: actor on GPU0 and learner on GPU1
python sebulba_ppo_envpool.py --exp-name sebulba_ppo_envpool_a0_l1_rollout_is_faster_timeout --actor-device-ids 0 --learner-device-ids 1 --params-queue-timeout 0.02 --total-timesteps 500000 --track

## shared: actor on GPU0 and learner on GPU0,1
python sebulba_ppo_envpool.py --exp-name sebulba_ppo_envpool_a0_l01_rollout_is_faster_timeout --actor-device-ids 0 --learner-device-ids 0 1 --params-queue-timeout 0.02 --total-timesteps 500000 --track

## separate: actor on GPU0 and learner on GPU1,2
python sebulba_ppo_envpool.py --exp-name sebulba_ppo_envpool_a0_l12_rollout_is_faster_timeout --actor-device-ids 0 --learner-device-ids 1 2 --params-queue-timeout 0.02 --total-timesteps 500000 --track

# 1.2. rollout is much faster than training w/ timeout

## throughput
python sebulba_ppo_envpool.py --exp-name sebulba_ppo_envpool_thpt_rollout_is_much_faster_timeout --actor-device-ids 0 --learner-device-ids 1 --update-epochs 8 --params-queue-timeout 0.02 --profile --test-actor-learner-throughput --total-timesteps 500000 --track

## shared: actor on GPU0 and learner on GPU0,1
python sebulba_ppo_envpool.py --exp-name sebulba_ppo_envpool_a0_l01_rollout_is_much_faster_timeout --actor-device-ids 0 --learner-device-ids 0 1 --update-epochs 8 --params-queue-timeout 0.02 --total-timesteps 500000 --track

## separate: actor on GPU0 and learner on GPU1,2
python sebulba_ppo_envpool.py --exp-name sebulba_ppo_envpool_a0_l12_rollout_is_much_faster_timeout --actor-device-ids 0 --learner-device-ids 1 2 --update-epochs 8 --params-queue-timeout 0.02 --total-timesteps 500000 --track

# 2. training is faster than rollout

## throughput
python sebulba_ppo_envpool.py --exp-name sebulba_ppo_envpool_thpt_training_is_faster --update-epochs 1 --async-batch-size 64 --actor-device-ids 0 --learner-device-ids 1 --params-queue-timeout 0.02 --profile --test-actor-learner-throughput --total-timesteps 500000 --track

## shared: actor on GPU0 and learner on GPU0
python sebulba_ppo_envpool.py --exp-name sebulba_ppo_envpool_1gpu_training_is_faster --update-epochs 1 --async-batch-size 64 --actor-device-ids 0 --learner-device-ids 0 --total-timesteps 500000 --track

## separate: actor on GPU0 and learner on GPU1
python sebulba_ppo_envpool.py --exp-name sebulba_ppo_envpool_a0_l1_training_is_faster --update-epochs 1 --async-batch-size 64 --actor-device-ids 0 --learner-device-ids 1 --total-timesteps 500000 --track

## shared: actor on GPU0 and learner on GPU0,1
python sebulba_ppo_envpool.py --exp-name sebulba_ppo_envpool_a0_l01_training_is_faster --update-epochs 1 --async-batch-size 64 --actor-device-ids 0 --learner-device-ids 0 1 --total-timesteps 500000 --track

## separate: actor on GPU0 and learner on GPU1,2
python sebulba_ppo_envpool.py --exp-name sebulba_ppo_envpool_a0_l12_training_is_faster --update-epochs 1 --async-batch-size 64 --actor-device-ids 0 --learner-device-ids 1 2 --total-timesteps 500000 --track

"""
# docs and experiment results can be found at https://docs.cleanrl.dev/rl-algorithms/ppo/#ppo_atari_envpool_async_jax_scan_impalanet_machadopy
# https://gregorygundersen.com/blog/2020/02/09/log-sum-exp/
import argparse
import os
import random
import time
import uuid
from collections import deque
from distutils.util import strtobool
from functools import partial
from typing import Sequence

os.environ[
    "XLA_PYTHON_CLIENT_MEM_FRACTION"
] = "0.6"  # see https://github.com/google/jax/discussions/6332#discussioncomment-1279991
os.environ["XLA_FLAGS"] = "--xla_cpu_multi_thread_eigen=false " "intra_op_parallelism_threads=1"
import multiprocessing as mp
import queue
import threading

import envpool
import flax
import flax.linen as nn
import gym
import jax
import jax.numpy as jnp
import numpy as np
import optax
from flax.linen.initializers import constant, orthogonal
from flax.training.train_state import TrainState
from torch.utils.tensorboard import SummaryWriter


def parse_args():
    # fmt: off
    parser = argparse.ArgumentParser()
    parser.add_argument("--exp-name", type=str, default=os.path.basename(__file__).rstrip(".py"),
        help="the name of this experiment")
    parser.add_argument("--seed", type=int, default=1,
        help="seed of the experiment")
    parser.add_argument("--torch-deterministic", type=lambda x: bool(strtobool(x)), default=True, nargs="?", const=True,
        help="if toggled, `torch.backends.cudnn.deterministic=False`")
    parser.add_argument("--cuda", type=lambda x: bool(strtobool(x)), default=True, nargs="?", const=True,
        help="if toggled, cuda will be enabled by default")
    parser.add_argument("--track", type=lambda x: bool(strtobool(x)), default=False, nargs="?", const=True,
        help="if toggled, this experiment will be tracked with Weights and Biases")
    parser.add_argument("--wandb-project-name", type=str, default="cleanRL",
        help="the wandb's project name")
    parser.add_argument("--wandb-entity", type=str, default=None,
        help="the entity (team) of wandb's project")
    parser.add_argument("--capture-video", type=lambda x: bool(strtobool(x)), default=False, nargs="?", const=True,
        help="weather to capture videos of the agent performances (check out `videos` folder)")
    parser.add_argument("--save-model", type=lambda x: bool(strtobool(x)), default=False, nargs="?", const=True,
        help="whether to save model into the `runs/{run_name}` folder")
    parser.add_argument("--upload-model", type=lambda x: bool(strtobool(x)), default=False, nargs="?", const=True,
        help="whether to upload the saved model to huggingface")
    parser.add_argument("--hf-entity", type=str, default="",
        help="the user or org name of the model repository from the Hugging Face Hub")

    # Algorithm specific arguments
    parser.add_argument("--env-id", type=str, default="Breakout-v5",
        help="the id of the environment")
    parser.add_argument("--total-timesteps", type=int, default=50000000,
        help="total timesteps of the experiments")
    parser.add_argument("--learning-rate", type=float, default=2.5e-4,
        help="the learning rate of the optimizer")
    parser.add_argument("--num-envs", type=int, default=64,
        help="the number of parallel game environments")
    parser.add_argument("--async-batch-size", type=int, default=16,
        help="the envpool's batch size in the async mode")
    parser.add_argument("--num-steps", type=int, default=128,
        help="the number of steps to run in each environment per policy rollout")
    parser.add_argument("--anneal-lr", type=lambda x: bool(strtobool(x)), default=True, nargs="?", const=True,
        help="Toggle learning rate annealing for policy and value networks")
    parser.add_argument("--gamma", type=float, default=0.99,
        help="the discount factor gamma")
    parser.add_argument("--gae-lambda", type=float, default=0.95,
        help="the lambda for the general advantage estimation")
    parser.add_argument("--num-minibatches", type=int, default=4,
        help="the number of mini-batches")
    parser.add_argument("--update-epochs", type=int, default=4,
        help="the K epochs to update the policy")
    parser.add_argument("--norm-adv", type=lambda x: bool(strtobool(x)), default=True, nargs="?", const=True,
        help="Toggles advantages normalization")
    parser.add_argument("--clip-coef", type=float, default=0.1,
        help="the surrogate clipping coefficient")
    parser.add_argument("--ent-coef", type=float, default=0.01,
        help="coefficient of the entropy")
    parser.add_argument("--vf-coef", type=float, default=0.5,
        help="coefficient of the value function")
    parser.add_argument("--max-grad-norm", type=float, default=0.5,
        help="the maximum norm for the gradient clipping")
    parser.add_argument("--target-kl", type=float, default=None,
        help="the target KL divergence threshold")

    parser.add_argument("--actor-device-ids", type=int, nargs="+", default=[0], # type is actually List[int]
        help="the device ids that actor workers will use")
    parser.add_argument("--learner-device-ids", type=int, nargs="+", default=[0], # type is actually List[int]
        help="the device ids that actor workers will use")
    parser.add_argument("--num-actor-threads", type=int, default=1,
        help="the number of actor threads")
    parser.add_argument("--profile", type=lambda x: bool(strtobool(x)), default=False, nargs="?", const=True,
        help="whether to call block_until_ready() for profiling")
    parser.add_argument("--test-actor-learner-throughput", type=lambda x: bool(strtobool(x)), default=False, nargs="?", const=True,
        help="whether to test actor-learner throughput by removing the actor-learner communication")
    parser.add_argument("--params-queue-timeout", type=float, default=None,
        help="the timeout for the `params_queue.get()` operation in the actor thread to pull params;" + \
             "by default it's `None`; if you set a timeout, it will likely make the actor run faster but will introduce some side effects," + \
             "such as the actor will not be able to pull the latest params from the learner and will use the old params instead")
    args = parser.parse_args()
    args.batch_size = int(args.num_envs * args.num_steps)
    args.minibatch_size = int(args.batch_size // args.num_minibatches)
    args.num_updates = args.total_timesteps // args.batch_size
    args.async_update = int(args.num_envs / args.async_batch_size)
    assert len(args.actor_device_ids) == 1, "only 1 actor_device_ids is supported now"
    # fmt: on
    return args


LEARNER_WARMUP_TIME = 10  # seconds


def make_env(env_id, seed, num_envs, async_batch_size=1, num_threads=None, thread_affinity_offset=-1):
    def thunk():
        envs = envpool.make(
            env_id,
            env_type="gym",
            num_envs=num_envs,
            num_threads=num_threads if num_threads is not None else async_batch_size,
            thread_affinity_offset=thread_affinity_offset,
            batch_size=async_batch_size,
            episodic_life=False,  # Machado et al. 2017 (Revisitng ALE: Eval protocols) p. 6
            repeat_action_probability=0.25,  # Machado et al. 2017 (Revisitng ALE: Eval protocols) p. 12
            noop_max=1,  # Machado et al. 2017 (Revisitng ALE: Eval protocols) p. 12 (no-op is deprecated in favor of sticky action, right?)
            full_action_space=True,  # Machado et al. 2017 (Revisitng ALE: Eval protocols) Tab. 5
            max_episode_steps=int(108000 / 4),  # Hessel et al. 2018 (Rainbow DQN), Table 3, Max frames per episode
            reward_clip=True,
            seed=seed,
        )
        envs.num_envs = num_envs
        envs.single_action_space = envs.action_space
        envs.single_observation_space = envs.observation_space
        envs.is_vector_env = True
        return envs

    return thunk


class ResidualBlock(nn.Module):
    channels: int

    @nn.compact
    def __call__(self, x):
        inputs = x
        x = nn.relu(x)
        x = nn.Conv(
            self.channels,
            kernel_size=(3, 3),
        )(x)
        x = nn.relu(x)
        x = nn.Conv(
            self.channels,
            kernel_size=(3, 3),
        )(x)
        return x + inputs


class ConvSequence(nn.Module):
    channels: int

    @nn.compact
    def __call__(self, x):
        x = nn.Conv(
            self.channels,
            kernel_size=(3, 3),
        )(x)
        x = nn.max_pool(x, window_shape=(3, 3), strides=(2, 2), padding="SAME")
        x = ResidualBlock(self.channels)(x)
        x = ResidualBlock(self.channels)(x)
        return x


class Network(nn.Module):
    channelss: Sequence[int] = (16, 32, 32)

    @nn.compact
    def __call__(self, x):
        x = jnp.transpose(x, (0, 2, 3, 1))
        x = x / (255.0)
        for channels in self.channelss:
            x = ConvSequence(channels)(x)
        x = nn.relu(x)
        x = x.reshape((x.shape[0], -1))
        x = nn.Dense(256, kernel_init=orthogonal(np.sqrt(2)), bias_init=constant(0.0))(x)
        x = nn.relu(x)
        return x


class Critic(nn.Module):
    @nn.compact
    def __call__(self, x):
        return nn.Dense(1, kernel_init=orthogonal(1), bias_init=constant(0.0))(x)


class Actor(nn.Module):
    action_dim: int

    @nn.compact
    def __call__(self, x):
        return nn.Dense(self.action_dim, kernel_init=orthogonal(0.01), bias_init=constant(0.0))(x)


@flax.struct.dataclass
class AgentParams:
    network_params: flax.core.FrozenDict
    actor_params: flax.core.FrozenDict
    critic_params: flax.core.FrozenDict


@partial(jax.jit, static_argnums=(3))
def get_action_and_value(
    params: TrainState,
    next_obs: np.ndarray,
    key: jax.random.PRNGKey,
    action_dim: int,
):
    hidden = Network().apply(params.network_params, next_obs)
    logits = Actor(action_dim).apply(params.actor_params, hidden)
    # sample action: Gumbel-softmax trick
    # see https://stats.stackexchange.com/questions/359442/sampling-from-a-categorical-distribution
    key, subkey = jax.random.split(key)
    u = jax.random.uniform(subkey, shape=logits.shape)
    action = jnp.argmax(logits - jnp.log(-jnp.log(u)), axis=1)
    logprob = jax.nn.log_softmax(logits)[jnp.arange(action.shape[0]), action]
    value = Critic().apply(params.critic_params, hidden)
    return action, logprob, value.squeeze(), key


@jax.jit
def prepare_data(
    obs: list,
    dones: list,
    values: list,
    actions: list,
    logprobs: list,
    env_ids: list,
    rewards: list,
):
    obs = jnp.asarray(obs)
    dones = jnp.asarray(dones)
    values = jnp.asarray(values)
    actions = jnp.asarray(actions)
    logprobs = jnp.asarray(logprobs)
    env_ids = jnp.asarray(env_ids)
    rewards = jnp.asarray(rewards)

    # TODO: in an unlikely event, one of the envs might have not stepped at all, which may results in unexpected behavior
    T, B = env_ids.shape
    index_ranges = jnp.arange(T * B, dtype=jnp.int32)
    next_index_ranges = jnp.zeros_like(index_ranges, dtype=jnp.int32)
    last_env_ids = jnp.zeros(args.num_envs, dtype=jnp.int32) - 1

    def f(carry, x):
        last_env_ids, next_index_ranges = carry
        env_id, index_range = x
        next_index_ranges = next_index_ranges.at[last_env_ids[env_id]].set(
            jnp.where(last_env_ids[env_id] != -1, index_range, next_index_ranges[last_env_ids[env_id]])
        )
        last_env_ids = last_env_ids.at[env_id].set(index_range)
        return (last_env_ids, next_index_ranges), None

    (last_env_ids, next_index_ranges), _ = jax.lax.scan(
        f,
        (last_env_ids, next_index_ranges),
        (env_ids.reshape(-1), index_ranges),
    )

    # rewards is off by one time step
    rewards = rewards.reshape(-1)[next_index_ranges].reshape((args.num_steps) * args.async_update, args.async_batch_size)
    advantages, returns, _, final_env_ids = compute_gae(env_ids, rewards, values, dones)
    # b_inds = jnp.nonzero(final_env_ids.reshape(-1), size=(args.num_steps) * args.async_update * args.async_batch_size)[0] # useful for debugging
    b_obs = obs.reshape((-1,) + obs.shape[2:])
    b_actions = actions.reshape(-1)
    b_logprobs = logprobs.reshape(-1)
    b_advantages = advantages.reshape(-1)
    b_returns = returns.reshape(-1)
    return b_obs, b_actions, b_logprobs, b_advantages, b_returns


def rollout(
    i,
    num_threads,  # =None,
    thread_affinity_offset,  # =-1,
    key: jax.random.PRNGKey,
    args,
    rollout_queue,
    params_queue: queue.Queue,
    writer,
    learner_devices,
):
    envs = make_env(args.env_id, args.seed, args.num_envs, args.async_batch_size, num_threads, thread_affinity_offset)()
    len_actor_device_ids = len(args.actor_device_ids)
    global_step = 0
    # TRY NOT TO MODIFY: start the game
    start_time = time.time()

    # put data in the last index
    episode_returns = np.zeros((args.num_envs,), dtype=np.float32)
    returned_episode_returns = np.zeros((args.num_envs,), dtype=np.float32)
    episode_lengths = np.zeros((args.num_envs,), dtype=np.float32)
    returned_episode_lengths = np.zeros((args.num_envs,), dtype=np.float32)
    envs.async_reset()

    params_queue_get_time = deque(maxlen=10)
    rollout_time = deque(maxlen=10)
    data_transfer_time = deque(maxlen=10)
    rollout_queue_put_time = deque(maxlen=10)
    params_timeout_count = 0
    for update in range(1, args.num_updates + 2):
        update_time_start = time.time()
        obs = []
        dones = []
        actions = []
        logprobs = []
        values = []
        env_ids = []
        rewards = []
        truncations = []
        terminations = []
        env_recv_time = 0
        inference_time = 0
        storage_time = 0
        env_send_time = 0

        # NOTE: This is a major difference from the sync version:
        # at the end of the rollout phase, the sync version will have the next observation
        # ready for the value bootstrap, but the async version will not have it.
        # for this reason we do `num_steps + 1`` to get the extra states for value bootstrapping.
        # but note that the extra states are not used for the loss computation in the next iteration,
        # while the sync version will use the extra state for the loss computation.
        params_queue_get_time_start = time.time()
        try:
            params = params_queue.get(timeout=args.params_queue_timeout)
        except queue.Empty:
            # print("params_queue.get timeout triggered")
            params_timeout_count += 1
        params_queue_get_time.append(time.time() - params_queue_get_time_start)
        writer.add_scalar("stats/params_queue_get_time", np.mean(params_queue_get_time), global_step)
        writer.add_scalar("stats/params_queue_timeout_count", params_timeout_count, global_step)
        rollout_time_start = time.time()
        for _ in range(
            args.async_update, (args.num_steps + 1) * args.async_update
        ):  # num_steps + 1 to get the states for value bootstrapping.
            env_recv_time_start = time.time()
            next_obs, next_reward, next_done, info = envs.recv()
            env_recv_time += time.time() - env_recv_time_start
            global_step += len(next_done) * args.num_actor_threads * len_actor_device_ids
            env_id = info["env_id"]

            inference_time_start = time.time()
            action, logprob, value, key = get_action_and_value(params, next_obs, key, envs.single_action_space.n)
            inference_time += time.time() - inference_time_start

            env_send_time_start = time.time()
            envs.send(np.array(action), env_id)
            env_send_time += time.time() - env_send_time_start
            storage_time_start = time.time()
            obs.append(next_obs)
            dones.append(next_done)
            values.append(value)
            actions.append(action)
            logprobs.append(logprob)
            env_ids.append(env_id)
            rewards.append(next_reward)
            truncations.append(info["TimeLimit.truncated"])
            terminations.append(info["terminated"])
            episode_returns[env_id] += info["reward"]
            returned_episode_returns[env_id] = np.where(
                info["terminated"] + info["TimeLimit.truncated"], episode_returns[env_id], returned_episode_returns[env_id]
            )
            episode_returns[env_id] *= (1 - info["terminated"]) * (1 - info["TimeLimit.truncated"])
            episode_lengths[env_id] += 1
            returned_episode_lengths[env_id] = np.where(
                info["terminated"] + info["TimeLimit.truncated"], episode_lengths[env_id], returned_episode_lengths[env_id]
            )
            episode_lengths[env_id] *= (1 - info["terminated"]) * (1 - info["TimeLimit.truncated"])
            storage_time += time.time() - storage_time_start
        if args.profile:
            action.block_until_ready()
        rollout_time.append(time.time() - rollout_time_start)
        writer.add_scalar("stats/rollout_time", np.mean(rollout_time), global_step)

        avg_episodic_return = np.mean(returned_episode_returns)
        writer.add_scalar("charts/avg_episodic_return", avg_episodic_return, global_step)
        writer.add_scalar("charts/avg_episodic_length", np.mean(returned_episode_lengths), global_step)
        if i == 0:
            print(f"global_step={global_step}, avg_episodic_return={avg_episodic_return}")
            print("SPS:", int(global_step / (time.time() - start_time)))
        writer.add_scalar("charts/SPS", int(global_step / (time.time() - start_time)), global_step)

        writer.add_scalar("stats/truncations", np.sum(truncations), global_step)
        writer.add_scalar("stats/terminations", np.sum(terminations), global_step)
        writer.add_scalar("stats/env_recv_time", env_recv_time, global_step)
        writer.add_scalar("stats/inference_time", inference_time, global_step)
        writer.add_scalar("stats/storage_time", storage_time, global_step)
        writer.add_scalar("stats/env_send_time", env_send_time, global_step)

        data_transfer_time_start = time.time()
        b_obs, b_actions, b_logprobs, b_advantages, b_returns = prepare_data(
            obs,
            dones,
            values,
            actions,
            logprobs,
            env_ids,
            rewards,
        )
        payload = (
            global_step,
            update,
            jnp.array_split(b_obs, len(learner_devices)),
            jnp.array_split(b_actions, len(learner_devices)),
            jnp.array_split(b_logprobs, len(learner_devices)),
            jnp.array_split(b_advantages, len(learner_devices)),
            jnp.array_split(b_returns, len(learner_devices)),
        )
        if args.profile:
            payload[2][0].block_until_ready()
        data_transfer_time.append(time.time() - data_transfer_time_start)
        writer.add_scalar("stats/data_transfer_time", np.mean(data_transfer_time), global_step)
        if update == 1 or not args.test_actor_learner_throughput:
            rollout_queue_put_time_start = time.time()
            rollout_queue.put(payload)
            rollout_queue_put_time.append(time.time() - rollout_queue_put_time_start)
            writer.add_scalar("stats/rollout_queue_put_time", np.mean(rollout_queue_put_time), global_step)

        if update == 1 or update == 2 or update == 3:
            time.sleep(LEARNER_WARMUP_TIME)  # makes sure the actor does to fill the rollout_queue at the get go

        writer.add_scalar(
            "charts/SPS_update",
            int(
                args.num_envs
                * args.num_steps
                * args.num_actor_threads
                * len_actor_device_ids
                / (time.time() - update_time_start)
            ),
            global_step,
        )


@partial(jax.jit, static_argnums=(3))
def get_action_and_value2(
    params: flax.core.FrozenDict,
    x: np.ndarray,
    action: np.ndarray,
    action_dim: int,
):
    hidden = Network().apply(params.network_params, x)
    logits = Actor(action_dim).apply(params.actor_params, hidden)
    logprob = jax.nn.log_softmax(logits)[jnp.arange(action.shape[0]), action]
    logits = logits - jax.scipy.special.logsumexp(logits, axis=-1, keepdims=True)
    logits = logits.clip(min=jnp.finfo(logits.dtype).min)
    p_log_p = logits * jax.nn.softmax(logits)
    entropy = -p_log_p.sum(-1)
    value = Critic().apply(params.critic_params, hidden).squeeze()
    return logprob, entropy, value


@jax.jit
def compute_gae(
    env_ids: np.ndarray,
    rewards: np.ndarray,
    values: np.ndarray,
    dones: np.ndarray,
):
    dones = jnp.asarray(dones)
    values = jnp.asarray(values)
    env_ids = jnp.asarray(env_ids)
    rewards = jnp.asarray(rewards)

    _, B = env_ids.shape
    final_env_id_checked = jnp.zeros(args.num_envs, jnp.int32) - 1
    final_env_ids = jnp.zeros(B, jnp.int32)
    advantages = jnp.zeros(B)
    lastgaelam = jnp.zeros(args.num_envs)
    lastdones = jnp.zeros(args.num_envs) + 1
    lastvalues = jnp.zeros(args.num_envs)

    def compute_gae_once(carry, x):
        lastvalues, lastdones, advantages, lastgaelam, final_env_ids, final_env_id_checked = carry
        (
            done,
            value,
            eid,
            reward,
        ) = x
        nextnonterminal = 1.0 - lastdones[eid]
        nextvalues = lastvalues[eid]
        delta = jnp.where(final_env_id_checked[eid] == -1, 0, reward + args.gamma * nextvalues * nextnonterminal - value)
        advantages = delta + args.gamma * args.gae_lambda * nextnonterminal * lastgaelam[eid]
        final_env_ids = jnp.where(final_env_id_checked[eid] == 1, 1, 0)
        final_env_id_checked = final_env_id_checked.at[eid].set(
            jnp.where(final_env_id_checked[eid] == -1, 1, final_env_id_checked[eid])
        )

        # the last_ variables keeps track of the actual `num_steps`
        lastgaelam = lastgaelam.at[eid].set(advantages)
        lastdones = lastdones.at[eid].set(done)
        lastvalues = lastvalues.at[eid].set(value)
        return (lastvalues, lastdones, advantages, lastgaelam, final_env_ids, final_env_id_checked), (
            advantages,
            final_env_ids,
        )

    (_, _, _, _, final_env_ids, final_env_id_checked), (advantages, final_env_ids) = jax.lax.scan(
        compute_gae_once,
        (
            lastvalues,
            lastdones,
            advantages,
            lastgaelam,
            final_env_ids,
            final_env_id_checked,
        ),
        (
            dones,
            values,
            env_ids,
            rewards,
        ),
        reverse=True,
    )
    return advantages, advantages + values, final_env_id_checked, final_env_ids


def ppo_loss(params, x, a, logp, mb_advantages, mb_returns, action_dim):
    newlogprob, entropy, newvalue = get_action_and_value2(params, x, a, action_dim)
    logratio = newlogprob - logp
    ratio = jnp.exp(logratio)
    approx_kl = ((ratio - 1) - logratio).mean()

    if args.norm_adv:
        mb_advantages = (mb_advantages - mb_advantages.mean()) / (mb_advantages.std() + 1e-8)

    # Policy loss
    pg_loss1 = -mb_advantages * ratio
    pg_loss2 = -mb_advantages * jnp.clip(ratio, 1 - args.clip_coef, 1 + args.clip_coef)
    pg_loss = jnp.maximum(pg_loss1, pg_loss2).mean()

    # Value loss
    v_loss = 0.5 * ((newvalue - mb_returns) ** 2).mean()

    entropy_loss = entropy.mean()
    loss = pg_loss - args.ent_coef * entropy_loss + v_loss * args.vf_coef
    return loss, (pg_loss, v_loss, entropy_loss, jax.lax.stop_gradient(approx_kl))


@partial(jax.jit, static_argnums=(6))
def single_device_update(
    agent_state: TrainState,
    b_obs,
    b_actions,
    b_logprobs,
    b_advantages,
    b_returns,
    action_dim,
    key: jax.random.PRNGKey,
):
    ppo_loss_grad_fn = jax.value_and_grad(ppo_loss, has_aux=True)

    def update_epoch(carry, _):
        agent_state, key = carry
        key, subkey = jax.random.split(key)

        # taken from: https://github.com/google/brax/blob/main/brax/training/agents/ppo/train.py
        def convert_data(x: jnp.ndarray):
            x = jax.random.permutation(subkey, x)
            x = jnp.reshape(x, (args.num_minibatches, -1) + x.shape[1:])
            return x

        def update_minibatch(agent_state, minibatch):
            mb_obs, mb_actions, mb_logprobs, mb_advantages, mb_returns = minibatch
            (loss, (pg_loss, v_loss, entropy_loss, approx_kl)), grads = ppo_loss_grad_fn(
                agent_state.params,
                mb_obs,
                mb_actions,
                mb_logprobs,
                mb_advantages,
                mb_returns,
                action_dim,
            )
            grads = jax.lax.pmean(grads, axis_name="devices")
            agent_state = agent_state.apply_gradients(grads=grads)
            return agent_state, (loss, pg_loss, v_loss, entropy_loss, approx_kl, grads)

        agent_state, (loss, pg_loss, v_loss, entropy_loss, approx_kl, grads) = jax.lax.scan(
            update_minibatch,
            agent_state,
            (
                convert_data(b_obs),
                convert_data(b_actions),
                convert_data(b_logprobs),
                convert_data(b_advantages),
                convert_data(b_returns),
            ),
        )
        return (agent_state, key), (loss, pg_loss, v_loss, entropy_loss, approx_kl, grads)

    (agent_state, key), (loss, pg_loss, v_loss, entropy_loss, approx_kl, _) = jax.lax.scan(
        update_epoch, (agent_state, key), (), length=args.update_epochs
    )
    return agent_state, loss, pg_loss, v_loss, entropy_loss, approx_kl, key


if __name__ == "__main__":
    devices = jax.devices("gpu")
    args = parse_args()
    run_name = f"{args.env_id}__{args.exp_name}__{args.seed}__{uuid.uuid4()}"
    if args.track:
        import wandb

        wandb.init(
            project=args.wandb_project_name,
            entity=args.wandb_entity,
            sync_tensorboard=True,
            config=vars(args),
            name=run_name,
            monitor_gym=True,
            save_code=True,
        )
    writer = SummaryWriter(f"runs/{run_name}")
    writer.add_text(
        "hyperparameters",
        "|param|value|\n|-|-|\n%s" % ("\n".join([f"|{key}|{value}|" for key, value in vars(args).items()])),
    )

    # TRY NOT TO MODIFY: seeding
    random.seed(args.seed)
    np.random.seed(args.seed)
    key = jax.random.PRNGKey(args.seed)
    key, network_key, actor_key, critic_key = jax.random.split(key, 4)

    # env setup
    envs = make_env(args.env_id, args.seed, args.num_envs, args.async_batch_size)()
    assert isinstance(envs.single_action_space, gym.spaces.Discrete), "only discrete action space is supported"

    def linear_schedule(count):
        # anneal learning rate linearly after one training iteration which contains
        # (args.num_minibatches * args.update_epochs) gradient updates
        frac = 1.0 - (count // (args.num_minibatches * args.update_epochs)) / args.num_updates
        return args.learning_rate * frac

    network = Network()
    actor = Actor(action_dim=envs.single_action_space.n)
    critic = Critic()
    network_params = network.init(network_key, np.array([envs.single_observation_space.sample()]))
    agent_state = TrainState.create(
        apply_fn=None,
        params=AgentParams(
            network_params,
            actor.init(actor_key, network.apply(network_params, np.array([envs.single_observation_space.sample()]))),
            critic.init(critic_key, network.apply(network_params, np.array([envs.single_observation_space.sample()]))),
        ),
        tx=optax.chain(
            optax.clip_by_global_norm(args.max_grad_norm),
            optax.inject_hyperparams(optax.adam)(
                learning_rate=linear_schedule if args.anneal_lr else args.learning_rate, eps=1e-5
            ),
        ),
    )
    learner_devices = [devices[d_id] for d_id in args.learner_device_ids]
    actor_devices = [devices[d_id] for d_id in args.actor_device_ids]
    agent_state = flax.jax_utils.replicate(agent_state, devices=learner_devices)

    multi_device_update = jax.pmap(
        single_device_update,
        axis_name="devices",
        devices=learner_devices,
        in_axes=(0, 0, 0, 0, 0, 0, None, None),
        out_axes=(0, 0, 0, 0, 0, 0, None),
        static_broadcasted_argnums=(6),
    )

    rollout_queue = queue.Queue(maxsize=2)
    params_queues = []
    num_cpus = mp.cpu_count()
    fair_num_cpus = num_cpus // len(args.actor_device_ids)

    class DummyWriter:
        def add_scalar(self, arg0, arg1, arg3):
            pass

    # lock = threading.Lock()
    # AgentParamsStore = namedtuple("AgentParamsStore", ["params", "version"])
    # agent_params_store = AgentParamsStore(agent_state.params, 0)

    dummy_writer = DummyWriter()
    for d_idx, d_id in enumerate(args.actor_device_ids):
        for j in range(args.num_actor_threads):
            params_queue = queue.Queue(maxsize=2)
            params_queue.put(jax.device_put(flax.jax_utils.unreplicate(agent_state.params), devices[d_id]))
            threading.Thread(
                target=rollout,
                args=(
                    j,
                    fair_num_cpus if args.num_actor_threads > 1 else None,
                    j * args.num_actor_threads if args.num_actor_threads > 1 else -1,
                    jax.device_put(key, devices[d_id]),
                    args,
                    rollout_queue,
                    params_queue,
                    writer if d_idx == 0 and j == 0 else dummy_writer,
                    learner_devices,
                ),
            ).start()
            params_queues.append(params_queue)

    rollout_queue_get_time = deque(maxlen=10)
    learner_update = 0
    while True:
        learner_update += 1
        if learner_update == 1 or not args.test_actor_learner_throughput:
            rollout_queue_get_time_start = time.time()
            global_step, update, b_obs, b_actions, b_logprobs, b_advantages, b_returns = rollout_queue.get()
            rollout_queue_get_time.append(time.time() - rollout_queue_get_time_start)
            writer.add_scalar("stats/rollout_queue_get_time", np.mean(rollout_queue_get_time), global_step)

        training_time_start = time.time()
        (agent_state, loss, pg_loss, v_loss, entropy_loss, approx_kl, key) = multi_device_update(
            agent_state,
            jax.device_put_sharded(b_obs, learner_devices),
            jax.device_put_sharded(b_actions, learner_devices),
            jax.device_put_sharded(b_logprobs, learner_devices),
            jax.device_put_sharded(b_advantages, learner_devices),
            jax.device_put_sharded(b_returns, learner_devices),
            envs.single_action_space.n,
            key,
        )
        if learner_update == 1 or not args.test_actor_learner_throughput:
            for d_idx, d_id in enumerate(args.actor_device_ids):
                for j in range(args.num_actor_threads):
                    params_queues[d_idx * args.num_actor_threads + j].put(
                        jax.device_put(flax.jax_utils.unreplicate(agent_state.params), devices[d_id])
                    )
        if args.profile:
            v_loss[-1, -1, -1].block_until_ready()
        writer.add_scalar("stats/training_time", time.time() - training_time_start, global_step)
        writer.add_scalar("stats/rollout_queue_size", rollout_queue.qsize(), global_step)
        writer.add_scalar("stats/params_queue_size", params_queue.qsize(), global_step)
        print(global_step, update, rollout_queue.qsize(), f"training time: {time.time() - training_time_start}s")

        # TRY NOT TO MODIFY: record rewards for plotting purposes
        writer.add_scalar("charts/learning_rate", agent_state.opt_state[1].hyperparams["learning_rate"][0].item(), global_step)
        writer.add_scalar("losses/value_loss", v_loss[-1, -1, -1].item(), global_step)
        writer.add_scalar("losses/policy_loss", pg_loss[-1, -1, -1].item(), global_step)
        writer.add_scalar("losses/entropy", entropy_loss[-1, -1, -1].item(), global_step)
        writer.add_scalar("losses/approx_kl", approx_kl[-1, -1, -1].item(), global_step)
        writer.add_scalar("losses/loss", loss[-1, -1, -1].item(), global_step)
        if update > args.num_updates:
            break

    if args.save_model:
        agent_state = flax.jax_utils.unreplicate(agent_state)
        model_path = f"runs/{run_name}/{args.exp_name}.cleanrl_model"
        with open(model_path, "wb") as f:
            f.write(
                flax.serialization.to_bytes(
                    [
                        vars(args),
                        [
                            agent_state.params.network_params,
                            agent_state.params.actor_params,
                            agent_state.params.critic_params,
                        ],
                    ]
                )
            )
        print(f"model saved to {model_path}")
        from cleanrl_utils.evals.ppo_envpool_jax_eval import evaluate

        episodic_returns = evaluate(
            model_path,
            make_env,
            args.env_id,
            eval_episodes=10,
            run_name=f"{run_name}-eval",
            Model=(Network, Actor, Critic),
        )
        for idx, episodic_return in enumerate(episodic_returns):
            writer.add_scalar("eval/episodic_return", episodic_return, idx)

        if args.upload_model:
            from cleanrl_utils.huggingface import push_to_hub

            repo_name = f"{args.env_id}-{args.exp_name}-seed{args.seed}"
            repo_id = f"{args.hf_entity}/{repo_name}" if args.hf_entity else repo_name
            push_to_hub(
                args,
                episodic_returns,
                repo_id,
                "PPO",
                f"runs/{run_name}",
                f"videos/{run_name}-eval",
                extra_dependencies=["jax", "envpool", "atari"],
            )

    envs.close()
    writer.close()