No guards for frozen dataclasses and readonly configs #138226
Description
🚀 The feature, motivation and pitch
Currently, a guard will be put for a frozen dataclass:
TORCH_LOGS="+guards" python -c """from dataclasses import dataclass
import torch
@dataclass(frozen=True)
class MyDC:
x: int
@torch.compile(fullgraph=True)
def func(dc: MyDC, t):
return dc.x + t
func(MyDC(0), torch.randn(()))
"""
gives a guard for
V1017 16:30:38.479000 75647 torch/_dynamo/guards.py:2314] [0/0] [__guards] | | +- GuardManager: source=L['dc'].x, accessed_by=GetAttrGuardAccessor(x)
V1017 16:30:38.479000 75647 torch/_dynamo/guards.py:2314] [0/0] [__guards] | | | +- EQUALS_MATCH: L['dc'].x == 0 # # <string>:10 in func
This is unnecessary as dc.x cannot be changed (dc can be guarded but dc.x is frozen).
The same would apply for readonly configs from omegaconf (when #138224 is fixed, this code will show guards for x):
import torch
from omegaconf import DictConfig
@torch.compile(fullgraph=True) # this will break currently
def func(dc, t):
return dc.x + t
cfg = DictConfig({'x': 0}, flags={'readonly': True})
func(cfg, torch.randn(()))Alternatives
I guess someone could hack their way through a frozen dataclass and that could lead to some errors but is it that much of a risk?
Additional context
No response
cc @ezyang @chauhang @penguinwu @voznesenskym @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @amjames @rec