Hi, a follow up on #15: I compared cpu vs mps and compile vs no compile on halfcheetah for 100k steps using SAC. It shows that mps is significantly slower than cpu, and aot_eager backend makes compile slower and much more so for cpu, tho the default inductor backend makes compile quite a bit faster for cpu but doesn't work for mps.
Code change is the following:
if args.compile:
mode = None # "reduce-overhead" if not args.cudagraphs else None
backend = "aot_eager" if device == torch.device("mps") else "inductor"
update_main = torch.compile(update_main, mode=mode, backend=backend)
update_pol = torch.compile(update_pol, mode=mode, backend=backend)
policy = torch.compile(policy, mode=mode, backend=backend)
Hi, a follow up on #15: I compared cpu vs mps and compile vs no compile on halfcheetah for 100k steps using SAC. It shows that mps is significantly slower than cpu, and
aot_eagerbackend makes compile slower and much more so for cpu, tho the defaultinductorbackend makes compile quite a bit faster for cpu but doesn't work for mps.Code change is the following: