- 
                Notifications
    
You must be signed in to change notification settings  - Fork 1k
 
Open
Labels
Description
Summary
When using variable-length RNN input with flow.nn.utils.rnn.pack_padded_sequence → nn.LSTM → flow.nn.utils.rnn.pad_packed_sequence inside nn.Graph, graph compilation fails with a generic:
RuntimeError: This is a oneflow bug, please submit an issue...
This happens during Graph build, not at runtime. (In eager mode this pipeline is expected to work.)
Code to reproduce bug
import oneflow as flow
import oneflow.nn as nn
class M(nn.Module):
    def __init__(self):
        super().__init__()
        self.lstm = nn.LSTM(256, 128, batch_first=True)
    def forward(self, x, xlen):
        packed = flow.nn.utils.rnn.pack_padded_sequence(
            x, xlen, batch_first=True, enforce_sorted=True
        )
        out, _ = self.lstm(packed)
        y, _ = flow.nn.utils.rnn.pad_packed_sequence(out, batch_first=True)
        return y
class G(nn.Graph):
    def __init__(self, m): super().__init__(); self.m = m
    def build(self, x, xlen): return self.m(x, xlen)
x = flow.randn(2, 12, 256)
xlen = flow.tensor([12, 10], dtype=flow.int64)
g = G(M())
g(x, xlen)  # -> RuntimeError: "This is a oneflow bug..."Output
[ERROR](GRAPH:G_0:G) building graph got error.
...
packed = flow.nn.utils.rnn.pack_padded_sequence(...)
  File ".../oneflow/nn/utils/rnn.py", line 256, in pack_padded_sequence
    data, batch_sizes = flow._C.pack_padded_sequence(input, lengths, batch_first)
RuntimeError: Error: RuntimeError : This is a oneflow bug, please submit an issue ...
System Information
- OS: Ubuntu 22.04.4 LTS (x86_64)
 - OneFlow version : 1.0.0.dev20250921+cpu
 - Python version: 3.10.16