1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725from __future__ import annotations
import math
import os
import random
import time
from dataclasses import dataclass
from typing import Tuple
import gymnasium as gym
import numpy as np
import torch
import torch.nn as nn
import torch.optim as optim
import tqdm
import tyro
import wandb
from torch.distributions.normal import Normal
from rustoracerpy import RustoracerEnv
# ━━━━━━━━━━━━━━━━━━━━ hyper-parameters ━━━━━━━━━━━━━━━━━━━━━━━━
@dataclass
class Args:
exp_name: str = os.path.basename(__file__)[: -len(".py")]
"""the name of this experiment"""
seed: int = 42
"""seed of the experiment"""
torch_deterministic: bool = True
"""if toggled, `torch.backends.cudnn.deterministic=False`"""
cuda: bool = True
"""if toggled, cuda will be enabled by default"""
capture_video: bool = True
"""whether to capture videos of the agent performances"""
# Algorithm specific arguments
yaml: str = "maps/berlin.yaml"
"""path to the RustoracerEnv YAML map file"""
total_timesteps: int = 50_000_000
"""total timesteps of the experiments"""
learning_rate: float = 3e-4
"""the learning rate of the optimizer"""
num_envs: int = 1024
"""the number of parallel game environments"""
num_steps: int = 128
"""the number of steps to run in each environment per policy rollout"""
max_ep_steps: int = 10_000
"""maximum steps per episode"""
anneal_lr: bool = True
"""Toggle learning rate annealing for policy and value networks"""
gamma: float = 0.99
"""the discount factor gamma"""
gae_lambda: float = 0.95
"""the lambda for the general advantage estimation"""
num_minibatches: int = 32
"""the number of mini-batches"""
update_epochs: int = 4
"""the K epochs to update the policy"""
norm_adv: bool = True
"""Toggles advantages normalization"""
clip_coef: float = 0.2
"""the surrogate clipping coefficient"""
clip_vloss: bool = True
"""Toggles whether or not to use a clipped loss for the value function"""
ent_coef: float = 0.01
"""coefficient of the entropy"""
vf_coef: float = 0.5
"""coefficient of the value function"""
max_grad_norm: float = 0.5
"""the maximum norm for the gradient clipping"""
target_kl: float = 0.02
"""the target KL divergence threshold"""
# Network
hidden: int = 256
"""hidden layer size"""
# to be filled in runtime
batch_size: int = 0
"""the batch size (computed in runtime)"""
minibatch_size: int = 0
"""the mini-batch size (computed in runtime)"""
num_iterations: int = 0
"""the number of iterations (computed in runtime)"""
video_interval: int = 25
"""record an evaluation video every N iterations (0 to disable)"""
video_max_steps: int = 600
"""max steps per evaluation video episode"""
save_interval: int = 50
"""save checkpoint every N iterations"""
save_dir: str = "checkpoints"
"""directory to save checkpoints"""
compile: bool = False
"""whether to use torch.compile"""
cudagraphs: bool = False
"""whether to use cudagraphs on top of compile"""
# ━━━━━━━━━━━━ GPU running mean / std (Welford, f32 accumulators) ━━━━━━━━━━━
class GPURunningMeanStd:
"""Welford running stats kept on device."""
def __init__(self, shape: tuple[int, ...], device: torch.device) -> None:
self.mean = torch.zeros(shape, dtype=torch.float32, device=device)
self.var = torch.ones(shape, dtype=torch.float32, device=device)
self.count: float = 1e-4
self.device = device
# Cached for fast normalize
self._inv_std = torch.ones(shape, dtype=torch.float32, device=device)
def _refresh_cache(self) -> None:
self._inv_std = torch.rsqrt(self.var + 1e-8)
def update(self, batch: torch.Tensor) -> None:
batch = batch.reshape(-1, *self.mean.shape).float()
bm = batch.mean(0)
bv = batch.var(0, correction=0)
bc = batch.shape[0]
delta = bm - self.mean
tot = self.count + bc
new_mean = self.mean + delta * (bc / tot)
m2 = self.var * self.count + bv * bc + delta**2 * (self.count * bc / tot)
self.mean = new_mean
self.var = m2 / tot
self.count = tot
self._refresh_cache()
def normalize(self, x: torch.Tensor, clip: float = 10.0) -> torch.Tensor:
return ((x - self.mean) * self._inv_std).clamp(-clip, clip)
class GPURunningReturnStd:
"""Track discounted return variance for reward normalisation."""
def __init__(self, num_envs: int, gamma: float, device: torch.device) -> None:
self.gamma = gamma
self.returns = torch.zeros(num_envs, dtype=torch.float32, device=device)
self.rms = GPURunningMeanStd(shape=(), device=device)
def update(self, rewards: torch.Tensor, dones: torch.Tensor) -> None:
self.returns = (
self.returns * self.gamma * (1.0 - dones.float()) + rewards.float()
)
self.rms.update(self.returns)
def normalize(self, rewards: torch.Tensor) -> torch.Tensor:
return rewards * self.rms._inv_std
# ━━━━━━━━━━ CPU running mean / std (for eval video only) ━━━━━━━━━━━
class CPURunningMeanStd:
def __init__(self, shape: tuple[int, ...]) -> None:
self.mean = np.zeros(shape, np.float64)
self.var = np.ones(shape, np.float64)
self.count: float = 1e-4
def normalize(self, x: np.ndarray, clip: float = 10.0) -> np.ndarray:
return np.clip((x - self.mean) / np.sqrt(self.var + 1e-8), -clip, clip)
# ━━━━━━━━━━━━━━━━━━ actor-critic network ━━━━━━━━━━━━━━━━━━━━━━
LOG_STD_MIN = -5.0
LOG_STD_MAX = 0.5
def layer_init(
layer: nn.Linear, std: float = float(np.sqrt(2)), bias: float = 0.0
) -> nn.Linear:
nn.init.orthogonal_(layer.weight, std) # ← needs QR, CPU-only on MPS
nn.init.constant_(layer.bias, bias)
return layer
class Agent(nn.Module):
def __init__(self, obs_dim: int, act_dim: int, hidden: int = 256) -> None:
super().__init__()
# Build + orthogonal-init on CPU (avoids MPS linalg_qr gap),
# caller does .to(device) afterward.
self.critic = nn.Sequential(
layer_init(nn.Linear(obs_dim, hidden)),
nn.Tanh(),
layer_init(nn.Linear(hidden, hidden)),
nn.Tanh(),
layer_init(nn.Linear(hidden, 1), std=1.0),
)
self.actor_mean = nn.Sequential(
layer_init(nn.Linear(obs_dim, hidden)),
nn.Tanh(),
layer_init(nn.Linear(hidden, hidden)),
nn.Tanh(),
layer_init(nn.Linear(hidden, act_dim), std=0.01),
)
self.actor_logstd = nn.Parameter(torch.full((1, act_dim), -0.5))
def get_value(self, x: torch.Tensor) -> torch.Tensor:
return self.critic(x)
def get_action_and_value(
self,
x: torch.Tensor,
action: torch.Tensor | None = None,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
mean = self.actor_mean(x)
log_std = self.actor_logstd.expand_as(mean).clamp(LOG_STD_MIN, LOG_STD_MAX)
std = log_std.exp()
dist = Normal(mean, std)
if action is None:
action = mean + std * torch.randn_like(mean)
return (
action,
dist.log_prob(action).sum(-1),
dist.entropy().sum(-1),
self.critic(x).squeeze(-1),
)
# ─────────────────────────────────────────────────────────────
# Video recording helper
# ─────────────────────────────────────────────────────────────
@torch.no_grad()
def record_eval_video(
yaml: str,
agent: Agent,
obs_rms_cpu: CPURunningMeanStd,
device: torch.device,
max_steps: int = 600,
max_ep_steps: int = 10_000,
) -> Tuple["wandb.Video | None", float]:
eval_env = RustoracerEnv(
yaml=yaml, num_envs=1, max_steps=max_ep_steps, render_mode="rgb_array"
)
raw_obs, _ = eval_env.reset(seed=42)
frames: list[np.ndarray] = []
total_reward = 0.0
for _ in range(max_steps):
frame = eval_env.render()
if frame is not None:
frames.append(frame)
obs_norm = obs_rms_cpu.normalize(raw_obs)
obs_t = torch.tensor(obs_norm, device=device, dtype=torch.float32)
action_mean = agent.actor_mean(obs_t)
action_np = action_mean.cpu().numpy().astype(np.float64).clip(-1.0, 1.0)
raw_obs, reward, terminated, truncated, _ = eval_env.step(action_np)
total_reward += float(reward[0])
if terminated[0] or truncated[0]:
break
eval_env.close()
if not frames:
return None, total_reward
video_np = np.stack(frames).transpose(0, 3, 1, 2)
return wandb.Video(video_np, fps=60, format="mp4"), total_reward
# ─────────────────────────────────────────────────────────────
# Helpers
# ─────────────────────────────────────────────────────────────
def gpu_rms_to_cpu(gpu_rms: GPURunningMeanStd) -> CPURunningMeanStd:
cpu_rms = CPURunningMeanStd(shape=tuple(gpu_rms.mean.shape))
cpu_rms.mean = gpu_rms.mean.cpu().numpy()
cpu_rms.var = gpu_rms.var.cpu().numpy()
cpu_rms.count = gpu_rms.count
return cpu_rms
def save_checkpoint(
args: Args,
agent: Agent,
optimizer: optim.Adam,
obs_rms: GPURunningMeanStd,
ret_rms: GPURunningReturnStd,
global_step: int,
update: int,
name: str | None = None,
) -> None:
if name is None:
name = f"agent_{global_step}.pt"
path = os.path.join(args.save_dir, name)
torch.save(
{
"model": agent.state_dict(),
"optimizer": optimizer.state_dict(),
"obs_rms_mean": obs_rms.mean.cpu(),
"obs_rms_var": obs_rms.var.cpu(),
"obs_rms_count": obs_rms.count,
"ret_rms_var": ret_rms.rms.var.cpu(),
"ret_rms_count": ret_rms.rms.count,
"global_step": global_step,
"update": update,
},
path,
)
print(f" 💾 {path}")
# ─────────────────────────────────────────────────────────────
# Main training loop
# ─────────────────────────────────────────────────────────────
if __name__ == "__main__":
args = tyro.cli(Args)
batch_size = args.num_envs * args.num_steps
args.minibatch_size = batch_size // args.num_minibatches
args.batch_size = batch_size
args.num_iterations = args.total_timesteps // batch_size
run_name = f"Rustoracer__{args.exp_name}__{args.seed}"
print(f"batch_size = {args.batch_size:,}")
print(f"minibatch_size = {args.minibatch_size:,}")
print(f"num_iterations = {args.num_iterations}")
print(f"hidden = {args.hidden}")
print(f"update_epochs = {args.update_epochs}")
print(f"ent_coef = {args.ent_coef}")
print(f"target_kl = {args.target_kl}")
print(f"max_ep_steps = {args.max_ep_steps}")
print(f"compile = {args.compile}")
print(f"cudagraphs = {args.cudagraphs}")
wandb.init(
project="ppo_continuous_action",
name=f"{os.path.splitext(os.path.basename(__file__))[0]}-{run_name}",
config=vars(args),
save_code=True,
)
# ── seeding ──────────────────────────────────────────────────
random.seed(args.seed)
np.random.seed(args.seed)
torch.manual_seed(args.seed)
torch.backends.cudnn.deterministic = args.torch_deterministic
# ── device ───────────────────────────────────────────────────
if args.cuda and torch.cuda.is_available():
device = torch.device("cuda")
elif torch.backends.mps.is_available():
device = torch.device("mps")
else:
device = torch.device("cpu")
print(f"Device : {device}")
is_cuda = device.type == "cuda"
# ── environment ──────────────────────────────────────────────
env = RustoracerEnv(
yaml=args.yaml, num_envs=args.num_envs, max_steps=args.max_ep_steps
)
obs_dim: int = env.single_observation_space.shape[0]
act_dim: int = env.single_action_space.shape[0]
print(f"obs={obs_dim} act={act_dim} envs={args.num_envs}")
# ── network + optimiser ──────────────────────────────────────
agent = Agent(obs_dim, act_dim, args.hidden).to(device)
optimizer = optim.Adam(agent.parameters(), lr=args.learning_rate, eps=1e-5)
print(f"Parameters : {sum(p.numel() for p in agent.parameters()):,}")
# ── GPU normalisers ──────────────────────────────────────────
obs_rms = GPURunningMeanStd(shape=(obs_dim,), device=device)
ret_rms = GPURunningReturnStd(args.num_envs, args.gamma, device=device)
# ── GPU rollout buffers (float32) ────────────────────────────
obs_buf = torch.zeros(
(args.num_steps, args.num_envs, obs_dim), dtype=torch.float32, device=device
)
act_buf = torch.zeros(
(args.num_steps, args.num_envs, act_dim), dtype=torch.float32, device=device
)
logp_buf = torch.zeros(
(args.num_steps, args.num_envs), dtype=torch.float32, device=device
)
rew_buf = torch.zeros(
(args.num_steps, args.num_envs), dtype=torch.float32, device=device
)
done_buf = torch.zeros(
(args.num_steps, args.num_envs), dtype=torch.float32, device=device
)
val_buf = torch.zeros(
(args.num_steps, args.num_envs), dtype=torch.float32, device=device
)
# ── Pinned CPU staging buffers for fast async transfer ───────
if is_cuda:
pin_obs = torch.zeros(
(args.num_envs, obs_dim), dtype=torch.float32
).pin_memory()
pin_rew = torch.zeros((args.num_envs,), dtype=torch.float32).pin_memory()
pin_done = torch.zeros((args.num_envs,), dtype=torch.float32).pin_memory()
else:
pin_obs = pin_rew = pin_done = None
# ── episode trackers (CPU — avoids GPU syncs) ────────────────
ep_ret_cpu = np.zeros(args.num_envs, dtype=np.float64)
ep_len_cpu = np.zeros(args.num_envs, dtype=np.int64)
recent_returns: list[float] = []
recent_lengths: list[int] = []
# ── Reusable CPU buffer for clipped actions ──────────────────
act_np_buf = np.empty((args.num_envs, act_dim), dtype=np.float64)
# ── first reset ──────────────────────────────────────────────
next_obs_raw_np, _ = env.reset(seed=args.seed)
if is_cuda:
pin_obs.numpy()[:] = next_obs_raw_np.astype(np.float32)
next_obs_raw = pin_obs.to(device, non_blocking=True).clone()
else:
next_obs_raw = torch.as_tensor(
next_obs_raw_np, dtype=torch.float32, device=device
)
obs_rms.update(next_obs_raw)
next_obs_n = obs_rms.normalize(next_obs_raw)
next_done = torch.zeros(args.num_envs, dtype=torch.float32, device=device)
os.makedirs(args.save_dir, exist_ok=True)
global_step = 0
t0 = time.time()
# ── compilable functions ─────────────────────────────────────
_gamma = args.gamma
_gae_lambda = args.gae_lambda
_num_steps = args.num_steps
_clip_coef = args.clip_coef
_vf_coef = args.vf_coef
_ent_coef = args.ent_coef
_max_grad_norm = args.max_grad_norm
_clip_vloss = args.clip_vloss
_norm_adv = args.norm_adv
def compute_gae(
next_val: torch.Tensor,
next_done_t: torch.Tensor,
rew_buf: torch.Tensor,
done_buf: torch.Tensor,
val_buf: torch.Tensor,
) -> tuple[torch.Tensor, torch.Tensor]:
advantages = torch.zeros_like(rew_buf)
lastgae = torch.zeros_like(next_val)
for t in range(_num_steps - 1, -1, -1):
if t == _num_steps - 1:
nnt = 1.0 - next_done_t
nv = next_val
else:
nnt = 1.0 - done_buf[t + 1]
nv = val_buf[t + 1]
delta = rew_buf[t] + _gamma * nv * nnt - val_buf[t]
lastgae = delta + _gamma * _gae_lambda * nnt * lastgae
advantages[t] = lastgae
returns = advantages + val_buf
return advantages, returns
def ppo_update_step(
b_obs: torch.Tensor,
b_act: torch.Tensor,
b_logp: torch.Tensor,
b_adv: torch.Tensor,
b_ret: torch.Tensor,
b_val: torch.Tensor,
mb_inds: torch.Tensor,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor, torch.Tensor]:
_, newlogp, entropy, newval = agent.get_action_and_value(
b_obs[mb_inds], b_act[mb_inds]
)
logratio = newlogp - b_logp[mb_inds]
ratio = logratio.exp()
with torch.no_grad():
approx_kl = ((ratio - 1) - logratio).mean()
clipfrac = ((ratio - 1.0).abs() > _clip_coef).float().mean()
mb_adv = b_adv[mb_inds]
if _norm_adv:
mb_adv = (mb_adv - mb_adv.mean()) / (mb_adv.std() + 1e-8)
# Policy loss
pg1 = -mb_adv * ratio
pg2 = -mb_adv * ratio.clamp(1 - _clip_coef, 1 + _clip_coef)
pg_loss = torch.max(pg1, pg2).mean()
# Value loss
if _clip_vloss:
v_unclipped = (newval - b_ret[mb_inds]) ** 2
v_clipped = b_val[mb_inds] + (newval - b_val[mb_inds]).clamp(
-_clip_coef, _clip_coef
)
v_loss = (
0.5 * torch.max(v_unclipped, (v_clipped - b_ret[mb_inds]) ** 2).mean()
)
else:
v_loss = 0.5 * ((newval - b_ret[mb_inds]) ** 2).mean()
ent_loss = entropy.mean()
loss = pg_loss - _ent_coef * ent_loss + _vf_coef * v_loss
optimizer.zero_grad(set_to_none=True)
loss.backward()
nn.utils.clip_grad_norm_(agent.parameters(), _max_grad_norm)
optimizer.step()
return (
pg_loss.detach(),
v_loss.detach(),
ent_loss.detach(),
approx_kl.detach(),
clipfrac.detach(),
)
if args.compile:
compute_gae = torch.compile(compute_gae, fullgraph=True)
ppo_update_step = torch.compile(ppo_update_step)
# ── video iteration schedule ─────────────────────────────────
if args.capture_video and args.video_interval > 0:
video_iters = set(range(1, args.num_iterations + 1, args.video_interval))
video_iters.add(args.num_iterations)
else:
video_iters = set()
pbar = tqdm.tqdm(range(1, args.num_iterations + 1))
# ═══════════════════ outer loop ════════════════════════════
for update in pbar:
# ── learning-rate annealing ──────────────────────────────
if args.anneal_lr:
frac = 1.0 - (update - 1) / args.num_iterations
optimizer.param_groups[0]["lr"] = args.learning_rate * frac
# ── rollout collection ───────────────────────────────────
for step in range(args.num_steps):
global_step += args.num_envs
obs_buf[step] = next_obs_n
done_buf[step] = next_done
with torch.inference_mode():
action, logprob, _, value = agent.get_action_and_value(next_obs_n)
val_buf[step] = value
logp_buf[step] = logprob
act_buf[step] = action
# Transfer action to CPU for env.step (unavoidable sync)
np.copyto(act_np_buf, action.cpu().numpy().clip(-1.0, 1.0))
next_obs_raw_np, reward_np, terminated_np, truncated_np, _ = env.step(
act_np_buf
)
done_np = np.logical_or(terminated_np, truncated_np).astype(np.float32)
# Fast CPU→GPU via pinned memory
if is_cuda:
pin_obs.numpy()[:] = next_obs_raw_np.astype(np.float32)
next_obs_raw = pin_obs.to(device, non_blocking=True)
pin_rew.numpy()[:] = reward_np.astype(np.float32)
reward_t = pin_rew.to(device, non_blocking=True)
pin_done.numpy()[:] = done_np
done_t = pin_done.to(device, non_blocking=True)
else:
next_obs_raw = torch.as_tensor(
next_obs_raw_np.astype(np.float32), device=device
)
reward_t = torch.as_tensor(reward_np.astype(np.float32), device=device)
done_t = torch.as_tensor(done_np, device=device)
# Reward normalisation
ret_rms.update(reward_t, done_t)
rew_buf[step] = ret_rms.normalize(reward_t)
next_done = done_t
# Episode bookkeeping (CPU — no GPU syncs)
ep_ret_cpu += reward_np
ep_len_cpu += 1
done_idx = np.where(done_np > 0.5)[0]
for i in done_idx:
recent_returns.append(float(ep_ret_cpu[i]))
recent_lengths.append(int(ep_len_cpu[i]))
ep_ret_cpu[i] = 0.0
ep_len_cpu[i] = 0
# Update obs normaliser
obs_rms.update(next_obs_raw)
next_obs_n = obs_rms.normalize(next_obs_raw)
# ── GAE ──────────────────────────────────────────────────
with torch.inference_mode():
next_val = agent.get_value(next_obs_n).squeeze(-1)
advantages, returns = compute_gae(
next_val, next_done, rew_buf, done_buf, val_buf
)
# ── flatten ──────────────────────────────────────────────
b_obs = obs_buf.reshape(batch_size, obs_dim)
b_act = act_buf.reshape(batch_size, act_dim)
b_logp = logp_buf.reshape(batch_size)
b_adv = advantages.reshape(batch_size)
b_ret = returns.reshape(batch_size)
b_val = val_buf.reshape(batch_size)
# ── PPO update epochs ────────────────────────────────────
# Accumulators on GPU — sync once at end
sum_pg = torch.zeros((), device=device)
sum_vl = torch.zeros((), device=device)
sum_ent = torch.zeros((), device=device)
sum_kl = torch.zeros((), device=device)
sum_cf = torch.zeros((), device=device)
num_updates = 0
kl_early_stopped = False
for _epoch in range(args.update_epochs):
if kl_early_stopped:
break
b_inds = torch.randperm(batch_size, device=device)
for start in range(0, batch_size, args.minibatch_size):
mb_inds = b_inds[start : start + args.minibatch_size]
pg_l, v_l, e_l, kl_l, cf_l = ppo_update_step(
b_obs, b_act, b_logp, b_adv, b_ret, b_val, mb_inds
)
sum_pg += pg_l
sum_vl += v_l
sum_ent += e_l
sum_kl += kl_l
sum_cf += cf_l
num_updates += 1
# Per-epoch KL check (only 1 sync per epoch, not per minibatch)
if kl_l.item() > args.target_kl * 1.5:
kl_early_stopped = True
# ── single GPU→CPU sync for all metrics ─────────────────
if num_updates > 0:
inv = 1.0 / num_updates
metrics = (
torch.stack([sum_pg, sum_vl, sum_ent, sum_kl, sum_cf]).cpu().numpy()
* inv
)
pg_loss_val, v_loss_val, ent_loss_val, approx_kl_val, clipfrac_val = metrics
else:
pg_loss_val = v_loss_val = ent_loss_val = approx_kl_val = clipfrac_val = 0.0
# ── logging ──────────────────────────────────────────────
elapsed = time.time() - t0
sps = global_step / elapsed
mr = float(np.mean(recent_returns[-200:])) if recent_returns else 0.0
ml = float(np.mean(recent_lengths[-200:])) if recent_lengths else 0.0
log_dict = {
"charts/episode_return": mr,
"charts/episode_length": ml,
"charts/learning_rate": optimizer.param_groups[0]["lr"],
"charts/SPS": sps,
"losses/pg_loss": pg_loss_val,
"losses/v_loss": v_loss_val,
"losses/entropy": ent_loss_val,
"losses/approx_kl": approx_kl_val,
"losses/clipfrac": clipfrac_val,
"debug/kl_early_stopped": int(kl_early_stopped),
"perf/global_step": global_step,
}
# ── video ────────────────────────────────────────────────
if update in video_iters:
print(f"\n[iter {update}] Recording evaluation video...")
agent.eval()
obs_rms_cpu = gpu_rms_to_cpu(obs_rms)
vid, eval_reward = record_eval_video(
yaml=args.yaml,
agent=agent,
obs_rms_cpu=obs_rms_cpu,
device=device,
max_steps=args.video_max_steps,
max_ep_steps=args.max_ep_steps,
)
agent.train()
if vid is not None:
log_dict["media/eval_video"] = vid
print(f"[iter {update}] Video captured, eval_return={eval_reward:.2f}")
else:
print(f"[iter {update}] WARNING: no frames captured!")
log_dict["charts/eval_return"] = eval_reward
wandb.log(log_dict, step=global_step)
pbar.set_description(
f"upd {update}/{args.num_iterations} | "
f"SPS {sps:,.0f} | "
f"ret {mr:.2f} | "
f"len {ml:.0f} | "
f"pg {pg_loss_val:.4f} | "
f"vl {v_loss_val:.4f} | "
f"ent {ent_loss_val:.3f} | "
f"kl {approx_kl_val:.4f} | "
f"lr {optimizer.param_groups[0]['lr']:.1e}"
f"{' ⚠KL-STOP' if kl_early_stopped else ''}"
)
# ── checkpoint ───────────────────────────────────────────
if update % args.save_interval == 0:
save_checkpoint(
args, agent, optimizer, obs_rms, ret_rms, global_step, update
)
# ── final save ───────────────────────────────────────────────
save_checkpoint(
args,
agent,
optimizer,
obs_rms,
ret_rms,
global_step,
args.num_iterations,
name="agent_final.pt",
)
total_time = time.time() - t0
print(f"\n✅ Training complete — {global_step:,} total steps in {total_time:.1f}s")
print(f"Average speed: {global_step / total_time:,.0f} SPS")
env.close()
wandb.finish()