🐛 Describe the bug
For this simple case, f2 has a break while f1 doesn't:
import torch
def f1(a):
num = a.nonzero().squeeze(-1).numel()
return torch.tensor([num]) # no break
def f2(a):
num = a.nonzero().squeeze(-1).numel()
return a.new_tensor([num]) # break
a = torch.tensor([1, 0])
torch.compile(f1, fullgraph=True)(a)
torch.compile(f2, fullgraph=True)(a)Error logs
Traceback (most recent call last):
File "test.py", line 13, in <module>
torch.compile(f2, fullgraph=True)(a)
File "pytorch/torch/_dynamo/eval_frame.py", line 1030, in compile_wrapper
raise e.with_traceback(None) from e.__cause__ # User compiler error
torch._dynamo.exc.UserError: Could not extract specialized integer from data-dependent expression u0 (unhinted: u0). (Size-like symbols: u0)
Caused by: return a.new_tensor([num]) # break # test.py:9 in f2 (_dynamo/utils.py:3906 in run_node)
For more information, run with TORCH_LOGS="dynamic"
For extended logs when we create symbols, also add TORCHDYNAMO_EXTENDED_DEBUG_CREATE_SYMBOL="u0"
If you suspect the guard was triggered from C++, add TORCHDYNAMO_EXTENDED_DEBUG_CPP=1
For more debugging help, see https://docs.google.com/document/d/1HSuTTVvYH1pTew89Rtpeu84Ht3nQEFTYhAX3Ypa_xJs/edit?usp=sharing
User Stack (most recent call last):
(snipped, see stack below for prefix)
File "test.py", line 9, in f2
return a.new_tensor([num]) # break
For C++ stack trace, run with TORCHDYNAMO_EXTENDED_DEBUG_CPP=1
For more information about this error, see: https://pytorch.org/docs/main/generated/exportdb/index.html#constrain-as-size-example
from user code:
File "test.py", line 9, in f2
return a.new_tensor([num]) # break
versions.log
Versions
To my understanding, if torch.tensor([num]) could be compiled without break, a.new_tensor([num]) should also pass?
cc @chauhang @penguinwu @ezyang @bobrenjc93 @aditvenk @laithsakka @voznesenskym @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @amjames @Lucaskabela @jataylo
🐛 Describe the bug
For this simple case, f2 has a break while f1 doesn't:
Error logs
versions.log
Versions
To my understanding, if
torch.tensor([num])could be compiled without break,a.new_tensor([num])should also pass?cc @chauhang @penguinwu @ezyang @bobrenjc93 @aditvenk @laithsakka @voznesenskym @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @kadeng @amjames @Lucaskabela @jataylo