I’m getting a Runtime error that only shows up when I use torch.compile() (in both nightly and stable releases). In eager mode the script runs without any error.
RuntimeError: Trying to backward through the graph a second time (or directly access saved tensors after they have already been freed). Saved intermediate values of the graph are freed when you call .backward() or autograd.grad(). Specify retain_graph=True if you need to backward through the graph a second time or if you need to access saved tensors after calling backward.
Here is a minimal script that reproduces the error. If you comment out the line net = torch.compile(net) then it will run without error:
import torch
from torch import nn
import torch.nn.functional as F
def device():
return "cuda"
class WidthBuffer:
def __init__(self):
self._values = None
def size(self):
if self._values is None:
return 0
return self._values.shape[3]
def add(self, new_values):
if self._values is None:
self._values = new_values
else:
self._values = torch.cat((self._values, new_values), 3)
def remove(self, size):
assert size <= self._values.shape[3]
removed = self._values[:, :, :, :size]
self._values = self._values[:, :, :, size:]
return removed
def detach(self):
if self._values is not None:
self._values = self._values.detach()
class FinalConv(nn.Module):
def __init__(self):
super().__init__()
self.buffer = WidthBuffer()
self.conv = nn.Conv2d(1, 4, 3, 1, 1, device=device())
def detach(self):
self.buffer.detach()
def forward(self, inputs, higher):
if higher is not None:
self.buffer.add(higher)
x = self.buffer.remove(inputs.shape[3])
x = self.conv(x)
return x
class Layer(nn.Module):
def __init__(self):
super().__init__()
self.values = None
self.conv_1 = nn.Conv2d(4, 1, 3, 2, device=device(), padding=1)
self.conv_2 = nn.ConvTranspose2d(1, 1, 2, 2, padding=0, device=device())
self.buffer = WidthBuffer()
self.input_buffer = WidthBuffer()
self.input_shape = None
def compute1(self, inputs):
self.input_buffer.add(inputs)
self.input_shape = inputs.shape
size = self.input_buffer.size()
remove_count = (size // 2) * 2
if remove_count:
inputs = self.input_buffer.remove(remove_count)
self.values = self.conv_1(inputs)
else:
empty_values_shape = (inputs.shape[0], 1, inputs.shape[2] // 2, 0)
self.values = torch.zeros(empty_values_shape, device=inputs.device)
def compute2(self):
out = self.conv_2(self.values)
self.values = None
return out
def detach(self):
self.buffer.detach()
self.input_buffer.detach()
class Model(nn.Module):
def __init__(self):
super().__init__()
self.layers = nn.ModuleList([Layer()])
self.final_conv = FinalConv()
self.requires_start_init = True
def forward(self, input_idxs, process_size):
if process_size and input_idxs.shape[3] != process_size:
padding = process_size - input_idxs.shape[3]
input_idxs = F.pad(input_idxs, (0, padding))
values = original_inputs = F.one_hot(input_idxs.squeeze(1), 4).permute(0, 3, 1, 2).float()
if self.requires_start_init:
self.start_init(values.shape, values.device)
self.layers[0].compute1(values)
x = self.layers[0].compute2()
x = self.final_conv(original_inputs, x)
return x
def start_init(self, input_shape, device):
inputs = values = torch.zeros(input_shape[:3] + (1,), device=device)
for layer in self.layers:
values = layer.compute1(values)
layer.values = torch.zeros(layer.values.shape[:3] + (1,), device=device)
self.final_conv(inputs, self.layers[0].compute2())
self.requires_start_init = False
def detach(self):
self.layers[0].detach()
self.final_conv.detach()
def train_net(net):
optimizer = torch.optim.Adam(net.parameters())
process_size = 10
for train_idx in range(10):
x = torch.rand((1, 6, 20), device=device())
for width_idx in range(0, x.shape[-1], process_size):
inputs = x[:, :, width_idx: width_idx + process_size].unsqueeze(1).long()
out = net(inputs, process_size)
loss = out.sum()
loss.backward()
optimizer.step()
optimizer.zero_grad()
net.detach()
print(f"Loss={loss.item():0.2f}")
net = Model()
net = torch.compile(net) # Works when this line is commented out
train_net(net)
My environment:
$ python3 -m torch.utils.collect_env
Collecting environment information...
PyTorch version: 2.1.0.dev20230429+cu118
Is debug build: False
CUDA used to build PyTorch: 11.8
ROCM used to build PyTorch: N/A
OS: Ubuntu 22.04.2 LTS (x86_64)
GCC version: (Ubuntu 11.3.0-1ubuntu1~22.04) 11.3.0
Clang version: Could not collect
CMake version: version 3.26.0
Libc version: glibc-2.35
Python version: 3.10.6 (main, Mar 10 2023, 10:55:28) [GCC 11.3.0] (64-bit runtime)
Python platform: Linux-5.19.0-41-generic-x86_64-with-glibc2.35
Is CUDA available: True
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration:
GPU 0: NVIDIA RTX A6000
GPU 1: NVIDIA GeForce GTX TITAN X
Nvidia driver version: 525.105.17
cuDNN version: Could not collect
HIP runtime version: N/A
MIOpen runtime version: N/A
Is XNNPACK available: True
CPU:
Architecture: x86_64
CPU op-mode(s): 32-bit, 64-bit
Address sizes: 48 bits physical, 48 bits virtual
Byte Order: Little Endian
CPU(s): 32
On-line CPU(s) list: 0-31
Vendor ID: AuthenticAMD
Model name: AMD Ryzen Threadripper PRO 5955WX 16-Cores
CPU family: 25
Model: 8
Thread(s) per core: 2
Core(s) per socket: 16
Socket(s): 1
Stepping: 2
Frequency boost: enabled
CPU max MHz: 7031.2500
CPU min MHz: 1800.0000
BogoMIPS: 7985.23
Flags: fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush mmx fxsr sse sse2 ht syscall nx mmxext fxsr_opt pdpe1gb rdtscp lm constant_tsc rep_good nopl nonstop_tsc cpuid extd_apicid aperfmperf rapl pni pclmulqdq monitor ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt aes xsave avx f16c rdrand lahf_lm cmp_legacy svm extapic cr8_legacy abm sse4a misalignsse 3dnowprefetch osvw ibs skinit wdt tce topoext perfctr_core perfctr_nb bpext perfctr_llc mwaitx cpb cat_l3 cdp_l3 invpcid_single hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a rdseed adx smap clflushopt clwb sha_ni xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local clzero irperf xsaveerptr rdpru wbnoinvd amd_ppin brs arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip pku ospke vaes vpclmulqdq rdpid overflow_recov succor smca fsrm
Virtualization: AMD-V
L1d cache: 512 KiB (16 instances)
L1i cache: 512 KiB (16 instances)
L2 cache: 8 MiB (16 instances)
L3 cache: 64 MiB (2 instances)
NUMA node(s): 1
NUMA node0 CPU(s): 0-31
Vulnerability Itlb multihit: Not affected
Vulnerability L1tf: Not affected
Vulnerability Mds: Not affected
Vulnerability Meltdown: Not affected
Vulnerability Mmio stale data: Not affected
Vulnerability Retbleed: Not affected
Vulnerability Spec store bypass: Mitigation; Speculative Store Bypass disabled via prctl
Vulnerability Spectre v1: Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2: Mitigation; Retpolines, IBPB conditional, IBRS_FW, STIBP always-on, RSB filling, PBRSB-eIBRS Not affected
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected
Versions of relevant libraries:
[pip3] numpy==1.24.2
[pip3] pytorch-triton==2.1.0+7d1a95b046
[pip3] torch==2.1.0.dev20230429+cu118
[pip3] torchaudio==2.1.0.dev20230429+cu118
[pip3] torchvision==0.16.0.dev20230429+cu118
[pip3] triton==2.0.0
[conda] Could not collect