OOM Error with `roi_align` in PyTorch 2.1.1 but fine in PyTorch 2.0.1

davidjaw commented 8 months ago

🐛 Describe the bug

Description

I am encountering an Out of Memory (OOM) error when using the roi_align function from PyTorch version 2.1.1 with torchvision 0.16.1. This issue does not occur with PyTorch version 2.0.1 and torchvision 0.15.2. The error happens regardless of the GPU used (tested on NVIDIA A2000 and RTX 4090). Note that when I downgrade the PyTorch and torchvision back to 2.0.1 and 0.15.2, the function can work properly. I am seeking assistance in understanding why this OOM error occurs in the newer versions of PyTorch and torchvision and whether this is a bug or a change in how roi_align manages memory.

Background

Hooking feature maps from YOLO model and using RoI-align to crop out object features.

Function

The object_roi_align function crops feature maps based on YOLO's object detection labels and uses RoI align to extract features of the object. The function accepts feature maps, YOLO detection labels, and several optional parameters for noise and class constraints.

def object_roi_align(feature, targets, nc=None, target_size=7, return_cls_feature=False,
                 noise_ratio=None, constrained_cls=None, constrained_cls_ratio=None):
"""
This function will crop feature map based on the label, and use RoI align to get the feature of the object
Args:
    feature: feature maps in shape [batch_size, channels, height, width]
    targets: [batch_idx, cls_index, cx, cy, w, h] (coords are normalized)
    target_size: size of the roi-align output (will be one of [7, 14, 28])
    noise_ratio: if not None, add noise to the target box (0 to 2)
    constrained_cls: if not None, limit the ratio of noise applied to the box of this class
    constrained_cls_ratio: if not None, limit the ratio of noise applied to the box of this class
    return_cls_feature: if True, return a list of feature maps for each class
"""
scale = feature.shape[-1] - 1
bidx, cidx, *box = [targets[:, i] for i in range(6)]
box = torch.stack(box, dim=1)
if noise_ratio is not None:
    # randomly add noise to the box coordinates to provide context
    box_wh = box[:, 2:]
    box_wh_ratio = torch.rand_like(box_wh, device=box_wh.device) * noise_ratio
    if constrained_cls is not None:
        # if targets[:, 1] == 7 (board), ratio is limited to max of 0.3
        box_wh_ratio[targets[:, 1] == constrained_cls] = torch.clamp(
            box_wh_ratio[targets[:, 1] == constrained_cls], max=constrained_cls_ratio)
    box_wh_noise = box_wh_ratio * box_wh
    box = torch.cat([box[:, :2], box_wh + box_wh_noise], dim=1)
# RoI align function takes real coordinates instead of normalized coordinates
box = box_convert(box, 'cxcywh', 'xyxy')
box = torch.clamp(box, min=0, max=1) * scale
box = torch.cat([bidx.unsqueeze(-1), box], dim=1).to(feature.device)
roi_features = roi_align(feature, box, output_size=target_size, aligned=True)
if return_cls_feature and nc:
    return split_feature_by_cls(roi_features, targets, nc)

return roi_features

Error messages (A2000)

  File "/home/davidjaw/Desktop/yolov5/custom_func.py", line 322, in <listcomp>
    x = [object_roi_align(x[i], targets, self.nc, target_size=self.base_resolution[i],
  File "/home/davidjaw/Desktop/yolov5/custom_func.py", line 486, in object_roi_align
    roi_features = roi_align(feature, box, output_size=target_size, aligned=True)
  File "/home/davidjaw/miniconda3/envs/torch210/lib/python3.10/site-packages/torchvision/ops/roi_align.py", line 236, in roi_align
    return _roi_align(input, rois, spatial_scale, output_size[0], output_size[1], sampling_ratio, aligned)
  File "/home/davidjaw/miniconda3/envs/torch210/lib/python3.10/site-packages/torchvision/ops/roi_align.py", line 168, in _roi_align
    val = _bilinear_interpolate(input, roi_batch_ind, y, x, ymask, xmask)  # [K, C, PH, PW, IY, IX]
  File "/home/davidjaw/miniconda3/envs/torch210/lib/python3.10/site-packages/torchvision/ops/roi_align.py", line 62, in _bilinear_interpolate
    v1 = masked_index(y_low, x_low)
  File "/home/davidjaw/miniconda3/envs/torch210/lib/python3.10/site-packages/torchvision/ops/roi_align.py", line 55, in masked_index
    return input[
torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 14.96 GiB. GPU 0 has a total capacty of 11.75 GiB of which 3.89 GiB is free. Including non-PyTorch memory, this process has 7.38 GiB memory in use. Of the allocated memory 7.01 GiB is allocated by PyTorch, and 219.61 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting max_split_size_mb to avoid fragmentation.  See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF

Error message (RTX 4090)

  File "/home/jdway/Desktop/yolov5/custom_func.py", line 322, in <listcomp>
    x = [object_roi_align(x[i], targets, self.nc, target_size=self.base_resolution[i],
  File "/home/jdway/Desktop/yolov5/custom_func.py", line 486, in object_roi_align
    roi_features = roi_align(feature, box, output_size=target_size, aligned=True)
  File "/home/jdway/miniconda3/envs/torch210/lib/python3.10/site-packages/torchvision/ops/roi_align.py", line 236, in roi_align
    return _roi_align(input, rois, spatial_scale, output_size[0], output_size[1], sampling_ratio, aligned)
  File "/home/jdway/miniconda3/envs/torch210/lib/python3.10/site-packages/torchvision/ops/roi_align.py", line 168, in _roi_align
    val = _bilinear_interpolate(input, roi_batch_ind, y, x, ymask, xmask)  # [K, C, PH, PW, IY, IX]
  File "/home/jdway/miniconda3/envs/torch210/lib/python3.10/site-packages/torchvision/ops/roi_align.py", line 62, in _bilinear_interpolate
    v1 = masked_index(y_low, x_low)
  File "/home/jdway/miniconda3/envs/torch210/lib/python3.10/site-packages/torchvision/ops/roi_align.py", line 55, in masked_index
    return input[
torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 14.96 GiB. GPU 0 has a total capacty of 23.64 GiB of which 14.67 GiB is free. Including non-PyTorch memory, this process has 8.72 GiB memory in use. Of the allocated memory 6.98 GiB is allocated by PyTorch, and 224.38 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting max_split_size_mb to avoid fragmentation.  See documentation for Memory Management and PYTORCH_CUDA_ALLOC_CONF
R

Versions

Versions (RTX 4090)

PyTorch version: 2.1.1+cu121
Is debug build: False
CUDA used to build PyTorch: 12.1
ROCM used to build PyTorch: N/A

OS: Ubuntu 20.04.6 LTS (x86_64)
GCC version: (Ubuntu 9.4.0-1ubuntu1~20.04.2) 9.4.0
Clang version: Could not collect
CMake version: Could not collect
Libc version: glibc-2.31

Python version: 3.10.13 | packaged by conda-forge | (main, Oct 26 2023, 18:07:37) [GCC 12.3.0] (64-bit runtime)
Python platform: Linux-5.15.0-91-generic-x86_64-with-glibc2.31
Is CUDA available: True
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: GPU 0: NVIDIA GeForce RTX 4090
Nvidia driver version: 530.30.02
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
Byte Order:                         Little Endian
Address sizes:                      46 bits physical, 48 bits virtual
CPU(s):                             24
On-line CPU(s) list:                0-23
Thread(s) per core:                 1
Core(s) per socket:                 16
Socket(s):                          1
NUMA node(s):                       1
Vendor ID:                          GenuineIntel
CPU family:                         6
Model:                              151
Model name:                         12th Gen Intel(R) Core(TM) i9-12900
Stepping:                           2
CPU MHz:                            2400.000
CPU max MHz:                        5100.0000
CPU min MHz:                        800.0000
BogoMIPS:                           4838.40
Virtualization:                     VT-x
L1d cache:                          384 KiB
L1i cache:                          256 KiB
L2 cache:                           10 MiB
NUMA node0 CPU(s):                  0-23
Vulnerability Gather data sampling: Not affected
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 rstack overflow: Not affected
Vulnerability Spec store bypass:    Mitigation; Speculative Store Bypass disabled via prctl and seccomp
Vulnerability Spectre v1:           Mitigation; usercopy/swapgs barriers and __user pointer sanitization
Vulnerability Spectre v2:           Mitigation; Enhanced IBRS, IBPB conditional, RSB filling, PBRSB-eIBRS SW sequence
Vulnerability Srbds:                Not affected
Vulnerability Tsx async abort:      Not affected
Flags:                              fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l2 cdp_l2 ssbd ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid rdt_a rdseed adx smap clflushopt clwb intel_pt sha_ni xsaveopt xsavec xgetbv1 xsaves split_lock_detect avx_vnni dtherm ida arat pln pts hwp hwp_notify hwp_act_window hwp_epp hwp_pkg_req umip pku ospke waitpkg gfni vaes vpclmulqdq tme rdpid movdiri movdir64b fsrm md_clear serialize pconfig arch_lbr flush_l1d arch_capabilities

Versions of relevant libraries:
[pip3] numpy==1.26.2
[pip3] torch==2.1.1
[pip3] torchaudio==2.1.1
[pip3] torchdata==0.7.1
[pip3] torchtext==0.16.1
[pip3] torchvision==0.16.1
[pip3] triton==2.1.0
[conda] libopenvino-pytorch-frontend 2023.1.0             h59595ed_2    conda-forge
[conda] numpy                     1.26.2          py310hb13e2d6_0    conda-forge
[conda] torch                     2.1.1                    pypi_0    pypi
[conda] torchaudio                2.1.1                    pypi_0    pypi
[conda] torchdata                 0.7.1                    pypi_0    pypi
[conda] torchtext                 0.16.1                   pypi_0    pypi
[conda] torchvision               0.16.1                   pypi_0    pypi
[conda] triton                    2.1.0                    pypi_0    pypi

Versions (A2000)

Collecting environment information...
PyTorch version: 2.1.1+cu121
Is debug build: False
CUDA used to build PyTorch: 12.1
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.1 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: Could not collect
CMake version: Could not collect
Libc version: glibc-2.35

Python version: 3.10.13 (main, Sep 11 2023, 13:44:35) [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-6.2.0-37-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 A2000 12GB
Nvidia driver version: 535.129.03
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:                      46 bits physical, 48 bits virtual
Byte Order:                         Little Endian
CPU(s):                             20
On-line CPU(s) list:                0-19
Vendor ID:                          GenuineIntel
Model name:                         12th Gen Intel(R) Core(TM) i7-12700
CPU family:                         6
Model:                              151
Thread(s) per core:                 2
Core(s) per socket:                 12
Socket(s):                          1
Stepping:                           2
CPU max MHz:                        4900.0000
CPU min MHz:                        800.0000
BogoMIPS:                           4224.00
Flags:                              fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm constant_tsc art arch_perfmon pebs bts rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 sdbg fma cx16 xtpr pdcm sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb ssbd ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid rdseed adx smap clflushopt clwb intel_pt sha_ni xsaveopt xsavec xgetbv1 xsaves split_lock_detect avx_vnni dtherm ida arat pln pts hwp hwp_notify hwp_act_window hwp_epp hwp_pkg_req hfi umip pku ospke waitpkg gfni vaes vpclmulqdq tme rdpid movdiri movdir64b fsrm md_clear serialize pconfig arch_lbr ibt flush_l1d arch_capabilities
Virtualization:                     VT-x
L1d cache:                          512 KiB (12 instances)
L1i cache:                          512 KiB (12 instances)
L2 cache:                           12 MiB (9 instances)
L3 cache:                           25 MiB (1 instance)
NUMA node(s):                       1
NUMA node0 CPU(s):                  0-19
Vulnerability Gather data sampling: Not affected
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 rstack overflow: 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; Enhanced IBRS, IBPB conditional, RSB filling, PBRSB-eIBRS SW sequence
Vulnerability Srbds:                Not affected
Vulnerability Tsx async abort:      Not affected

Versions of relevant libraries:
[pip3] numpy==1.26.2
[pip3] torch==2.1.1
[pip3] torchaudio==2.1.1
[pip3] torchdata==0.7.1
[pip3] torchtext==0.16.1
[pip3] torchvision==0.16.1
[pip3] triton==2.1.0
[conda] numpy                     1.26.2                   pypi_0    pypi
[conda] torch                     2.1.1                    pypi_0    pypi
[conda] torchaudio                2.1.1                    pypi_0    pypi
[conda] torchdata                 0.7.1                    pypi_0    pypi
[conda] torchtext                 0.16.1                   pypi_0    pypi
[conda] torchvision               0.16.1                   pypi_0    pypi
[conda] triton                    2.1.0                    pypi_0    pypi

cc @ezyang @gchanan @zou3519 @kadeng @ptrblck

malfet commented 8 months ago

We need to understand where regression is coming from, but sounds a bit like a torchvision problem, isn't it?

Also, I wonder if this is CUDA-11.8 vs CUDA-12.1 regression (2.0.1 was shipped with 11.8 by default, but 2.1 with 12.1)

davidjaw commented 8 months ago

Hello,

I've created a minimal toy example to demonstrate the issue in detail. You can find it here: https://gist.github.com/davidjaw/40bcbcf44cb3db01fd9178e193edb0de

This example relies on the ultralytics library. For context, the code runs as expected when using PyTorch version 2.0.1 and Torchvision version 0.15.2+cu118, and OOM when PyTorch 2.1.1. I believe this setup aligns with the requirements mentioned in the original issue.

Please take a look at the gist, and let me know if you need any more information or if there's anything else I can do to assist in resolving this issue.

Thank you!

FabianSchuetze commented 7 months ago

I just want to chime in an mention that I have the same problem. A very large memory allocation is attempted both on the GPU and the CPU. I observe the problem in the following environment:

Collecting environment information...
PyTorch version: 2.1.0+cu118
Is debug build: False
CUDA used to build PyTorch: 11.8
ROCM used to build PyTorch: N/A

But everything works with:

Collecting environment information...
PyTorch version: 2.1.0.dev20230714+cu121
Is debug build: False
CUDA used to build PyTorch: 12.1
ROCM used to build PyTorch: N/A

Below is a little snippet which leads to the OOM error with 2.1.0+cu118:

import torchvision
import torch
inp = torch.rand((1, 256, 48, 64))
bbox = torch.tensor([[0, 0, 0, 128, 96]]).float()
output_size = (48, 64)
scale = 12 / 384
aligned=True
torch.use_deterministic_algorithms(True)
out = torchvision.ops.roi_align(inp.cuda(), bbox.cuda(), output_size, scale, aligned=aligned)

Can someone else replicate this?

ezyang commented 7 months ago

Oh you know what, it's probably because of use deterministic algorithms. We added a deterministic implementation but it is very memory hungry

JohannesTheo commented 3 months ago

Hey, I just came across this issue. Is there any update?

I understand the appeal of a deterministic implementation but the caveat very memory hungry is an understatement :D

When I call the problematic _bilinear_interpolate -> masked_index part manually it will allocate ~30GB VRAM for a single input of size 400x400. Essentially, this breaks any Mask R-CNN model when using torch.use_deterministic_algorithms(True).

Has this been actually tested or run in a benchmark? If so, how? I fail to see how this is the intended behavior unless I'm missing something fundamental 😅 thx for any help.

ezyang commented 3 months ago

The implementation doesn't OOM if we torch.compile it. So I think I will fix it by making torch.compile on it mandatory.

FabianSchuetze commented 3 months ago

The implementation doesn't OOM if we torch.compile it. So I think I will fix it by making torch.compile on it mandatory.

Jesus, how is that possible? Sounds great.

JohannesTheo commented 3 months ago

Thank you for the quick response and the fix @ezyang 🚀

I applied your patch manually in my system and can confirm that it does eliminate the OOM issue! A Mask R-CNN ResNet-50 FPN now consumes ~5500 MB for batch_size 2 on COCO and about ~38000 MB for batch_size 16.

I ran some quick tests using the torchvision reference implementation and can further confirm that we now have deterministic training (see below). In addition, I append some timings in case this helps moving forward.

Test Setup

torchvision reference implementation (with some tweaks for reproducible determinism)
Python 3.11 with pre-build (pypi) torch 2.1.2+cu121 and torchvision 0.16.2+cu121
Nvidia RTX A6000 GPU(s)
maskrcnn_resnet50_fpn, coco, batch_size 2, hflip (+ internal *ShortestEdgeResize*(800,1333))

For reproducible determinism I set:

torch.backends.cudnn.benchmark = False
torch.backends.cudnn.deterministic = True
torch.use_deterministic_algorithms(True)
os.environ['CUBLAS_WORKSPACE_CONFIG'] = ":4096:8"

# seed for main process/thread (torch, random, numpy)
seed=617971023

# seed for sampler and dataloader generators (num_workers=4)
seed_data=0

Deterministic Training works :)

# nn.Module run 1
Epoch: [0]  [ 0/58633]  eta: 80 days, 11:30:44  lr: 0.000000  loss: 6.0988 (6.0988)  loss_classifier: 4.5028 (4.5028)  loss_box_reg: 0.0175 (0.0175)  loss_mask: 0.7927 (0.7927)  loss_objectness: 0.6928 (0.6928)  loss_rpn_box_reg: 0.0930 (0.0930)  time: 118.5927  data: 0.2771  max mem: 4346
Epoch: [0]  [20/58633]  eta: 12 days, 17:24:42  lr: 0.000002  loss: 6.0132 (6.0607)  loss_classifier: 4.4406 (4.4362)  loss_box_reg: 0.0319 (0.0412)  loss_mask: 0.7437 (0.7650)  loss_objectness: 0.6933 (0.6934)  loss_rpn_box_reg: 0.0742 (0.1249)  time: 13.7666  data: 0.0025  max mem: 5533
Epoch: [0]  [40/58633]  eta: 11 days, 12:51:42  lr: 0.000004  loss: 5.5885 (5.7992)  loss_classifier: 3.9741 (4.1683)  loss_box_reg: 0.0536 (0.0510)  loss_mask: 0.7486 (0.7632)  loss_objectness: 0.6906 (0.6920)  loss_rpn_box_reg: 0.0687 (0.1248)  time: 15.1755  data: 0.0028  max mem: 5533
Epoch: [0]  [60/58633]  eta: 11 days, 21:56:55  lr: 0.000006  loss: 3.4103 (4.9821)  loss_classifier: 1.5982 (3.3080)  loss_box_reg: 0.0917 (0.0753)  loss_mask: 0.8951 (0.8050)  loss_objectness: 0.6612 (0.6780)  loss_rpn_box_reg: 0.0843 (0.1159)  time: 18.7317  data: 0.0029  max mem: 5533

# nn.Module run 2
Epoch: [0]  [ 0/58633]  eta: 79 days, 04:24:00  lr: 0.000000  loss: 6.0988 (6.0988)  loss_classifier: 4.5028 (4.5028)  loss_box_reg: 0.0175 (0.0175)  loss_mask: 0.7927 (0.7927)  loss_objectness: 0.6928 (0.6928)  loss_rpn_box_reg: 0.0930 (0.0930)  time: 116.6824  data: 0.2564  max mem: 4346
Epoch: [0]  [20/58633]  eta: 12 days, 14:39:22  lr: 0.000002  loss: 6.0132 (6.0607)  loss_classifier: 4.4406 (4.4362)  loss_box_reg: 0.0319 (0.0412)  loss_mask: 0.7437 (0.7650)  loss_objectness: 0.6933 (0.6934)  loss_rpn_box_reg: 0.0742 (0.1249)  time: 13.6844  data: 0.0023  max mem: 5400
Epoch: [0]  [40/58633]  eta: 11 days, 11:22:22  lr: 0.000004  loss: 5.5885 (5.7992)  loss_classifier: 3.9741 (4.1683)  loss_box_reg: 0.0536 (0.0510)  loss_mask: 0.7486 (0.7632)  loss_objectness: 0.6906 (0.6920)  loss_rpn_box_reg: 0.0687 (0.1248)  time: 15.1657  data: 0.0026  max mem: 5400
Epoch: [0]  [60/58633]  eta: 11 days, 21:05:58  lr: 0.000006  loss: 3.4103 (4.9821)  loss_classifier: 1.5982 (3.3080)  loss_box_reg: 0.0917 (0.0753)  loss_mask: 0.8951 (0.8050)  loss_objectness: 0.6612 (0.6780)  loss_rpn_box_reg: 0.0843 (0.1159)  time: 18.7601  data: 0.0027  max mem: 5400

# DDP (world_size 1) run 1
Epoch: [0]  [ 0/58633]  eta: 70 days, 18:22:51  lr: 0.000000  loss: 6.0988 (6.0988)  loss_classifier: 4.5028 (4.5028)  loss_box_reg: 0.0175 (0.0175)  loss_mask: 0.7927 (0.7927)  loss_objectness: 0.6928 (0.6928)  loss_rpn_box_reg: 0.0930 (0.0930)  time: 104.2787  data: 0.4824  max mem: 4515
Epoch: [0]  [20/58633]  eta: 12 days, 02:47:31  lr: 0.000002  loss: 6.0132 (6.0607)  loss_classifier: 4.4406 (4.4362)  loss_box_reg: 0.0319 (0.0412)  loss_mask: 0.7437 (0.7650)  loss_objectness: 0.6933 (0.6934)  loss_rpn_box_reg: 0.0742 (0.1249)  time: 13.5395  data: 0.0021  max mem: 5572
Epoch: [0]  [40/58633]  eta: 11 days, 04:49:32  lr: 0.000004  loss: 5.5885 (5.7992)  loss_classifier: 3.9741 (4.1683)  loss_box_reg: 0.0536 (0.0510)  loss_mask: 0.7486 (0.7632)  loss_objectness: 0.6906 (0.6920)  loss_rpn_box_reg: 0.0687 (0.1248)  time: 15.1061  data: 0.0027  max mem: 5572
Epoch: [0]  [60/58633]  eta: 11 days, 16:31:04  lr: 0.000006  loss: 3.4103 (4.9821)  loss_classifier: 1.5982 (3.3080)  loss_box_reg: 0.0917 (0.0753)  loss_mask: 0.8951 (0.8050)  loss_objectness: 0.6612 (0.6780)  loss_rpn_box_reg: 0.0843 (0.1159)  time: 18.7259  data: 0.0028  max mem: 5572

# DDP (world_size 1) run 2
Epoch: [0]  [ 0/58633]  eta: 71 days, 13:13:18  lr: 0.000000  loss: 6.0988 (6.0988)  loss_classifier: 4.5028 (4.5028)  loss_box_reg: 0.0175 (0.0175)  loss_mask: 0.7927 (0.7927)  loss_objectness: 0.6928 (0.6928)  loss_rpn_box_reg: 0.0930 (0.0930)  time: 105.4355  data: 0.4228  max mem: 4515
Epoch: [0]  [20/58633]  eta: 12 days, 03:27:48  lr: 0.000002  loss: 6.0132 (6.0607)  loss_classifier: 4.4406 (4.4362)  loss_box_reg: 0.0319 (0.0412)  loss_mask: 0.7437 (0.7650)  loss_objectness: 0.6933 (0.6934)  loss_rpn_box_reg: 0.0742 (0.1249)  time: 13.5249  data: 0.0024  max mem: 5566
Epoch: [0]  [40/58633]  eta: 11 days, 05:10:02  lr: 0.000004  loss: 5.5885 (5.7992)  loss_classifier: 3.9741 (4.1683)  loss_box_reg: 0.0536 (0.0510)  loss_mask: 0.7486 (0.7632)  loss_objectness: 0.6906 (0.6920)  loss_rpn_box_reg: 0.0687 (0.1248)  time: 15.1059  data: 0.0027  max mem: 5566
Epoch: [0]  [60/58633]  eta: 11 days, 16:41:03  lr: 0.000006  loss: 3.4103 (4.9821)  loss_classifier: 1.5982 (3.3080)  loss_box_reg: 0.0917 (0.0753)  loss_mask: 0.8951 (0.8050)  loss_objectness: 0.6612 (0.6780)  loss_rpn_box_reg: 0.0843 (0.1159)  time: 18.7140  data: 0.0026  max mem: 5566

Time and memory overhead for DDP models

To anyone who randomly stumbles across this, please NOTE:

The following measurements are NOT definitive. It's only a quick check of a not yet merged fix that I have applied manually! So unless you read the context of the whole conversation, please don't draw any strong conclusion from this or worse, use it to claim anything about PyTorch/Torchvision speed/mem, thx.

The following values are for DDP models with batch_size 2 per gpu (world_size) and depict the avg. time per batch and max_mem after 20 batches (measured with MetricLogger from reference implementation).

world_size	deterministic	avg. time/batch	max_mem
1	False	0.1548	2929
1	True	13.0091 (~84x)	5377 (~1.83x)
4	False	0.1909	3094
4	True	24.5909 (~139x)	5645 (~1.82x)
8	False	0.2002	3094
8	True	29.8613 (~149x)	5547 (~1.79x)

Misc

I get the following deprecation warnings with torch 2.1.2+cu121, torchvision 0.16.2+cu121

UserWarning: 'has_cuda' is deprecated, please use 'torch.backends.cuda.is_built()'
UserWarning: 'has_cudnn' is deprecated, please use 'torch.backends.cudnn.is_available()'
UserWarning: 'has_mps' is deprecated, please use 'torch.backends.mps.is_built()'
UserWarning: 'has_mkldnn' is deprecated, please use 'torch.backends.mkldnn.is_available()'

ezyang commented 3 months ago

By the way, I think Inductor can potentially do a lot better codegen on this to bring down the time/memory overhead, just need some concerted elbow grease on it.

JohannesTheo commented 3 months ago

I'd love to help but have to admit that this part of the code base is a little over my head 😅 However, if I can support you with some isolated testing (that does not require in-depth knowledge) let me know!

JohannesTheo commented 3 months ago

Quick update: I ran some more tests to see if newer torch/torchvision versions will improve things but, it appears that I've been lucky with torch 2.1.2+cu121 and torchvision 0.16.2+cu121 and this needs more testing 😅 .

With torch 2.2.2+cu121 and torchvision 0.17.2+cu121 it runs OOM again and with torch 2.3.0+cu121 and torchvision 0.18.0+cu121 I get an assert error, see below.

Both are the pre-build versions from pypi. DDM models on 1 GPU (but same errors with nn.Module). I replaced my env path with ... for brevity. Hope this helps to narrow things down.

OOM with `torch 2.2.2+cu121` and `torchvision 0.17.2+cu121`

skipping cudagraphs due to deterministic index put. Found from :
   File ".../site-packages/torchvision/ops/roi_align.py", line 185, in _roi_align
    val = _bilinear_interpolate(input, roi_batch_ind, y, x, ymask, xmask)  # [K, C, PH, PW, IY, IX]
  File ".../site-packages/torchvision/ops/roi_align.py", line 78, in _bilinear_interpolate
    v1 = masked_index(y_low, x_low)
  File ".../site-packages/torchvision/ops/roi_align.py", line 71, in masked_index
    return input[
...
3 x more times the same skipping cudagraphs message, not shown here.
...

[rank0]:[2024-05-23 13:24:02,534] torch._dynamo.convert_frame: [WARNING] torch._dynamo hit config.cache_size_limit (8)
[rank0]:[2024-05-23 13:24:02,534] torch._dynamo.convert_frame: [WARNING]    function: '_roi_align' (.../site-packages/torchvision/ops/roi_align.py:114)
[rank0]:[2024-05-23 13:24:02,534] torch._dynamo.convert_frame: [WARNING]    last reason: ___check_global_state()
[rank0]:[2024-05-23 13:24:02,534] torch._dynamo.convert_frame: [WARNING] To log all recompilation reasons, use TORCH_LOGS="recompiles".
[rank0]:[2024-05-23 13:24:02,534] torch._dynamo.convert_frame: [WARNING] To diagnose recompilation issues, see https://pytorch.org/docs/master/compile/troubleshooting.html.

...

File ".../site-packages/torchvision/ops/roi_align.py", line 185, in _roi_align
    val = _bilinear_interpolate(input, roi_batch_ind, y, x, ymask, xmask)  # [K, C, PH, PW, IY, IX]
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File ".../site-packages/torchvision/ops/roi_align.py", line 32, in _bilinear_interpolate
  def _bilinear_interpolate(
File ".../site-packages/torch/_dynamo/eval_frame.py", line 489, in _fn
  return fn(*args, **kwargs)
          ^^^^^^^^^^^^^^^^^^^
File ".../site-packages/torch/_dynamo/external_utils.py", line 17, in inner
  return fn(*args, **kwargs)
          ^^^^^^^^^^^^^^^^^^^
File ".../site-packages/torch/_functorch/aot_autograd.py", line 901, in forward
  return compiled_fn(full_args)
          ^^^^^^^^^^^^^^^^^^^^^^
File ".../site-packages/torch/_functorch/_aot_autograd/utils.py", line 81, in g
  return f(*args)
          ^^^^^^^^
File ".../site-packages/torch/_functorch/_aot_autograd/runtime_wrappers.py", line 94, in runtime_wrapper
  all_outs = call_func_at_runtime_with_args(
              ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File ".../site-packages/torch/_functorch/_aot_autograd/utils.py", line 105, in call_func_at_runtime_with_args
  out = normalize_as_list(f(args))
                          ^^^^^^^
File ".../site-packages/torch/_functorch/_aot_autograd/jit_compile_runtime_wrappers.py", line 118, in rng_functionalization_wrapper
  return compiled_fw(args)
          ^^^^^^^^^^^^^^^^^
File ".../site-packages/torch/_inductor/codecache.py", line 864, in __call__
  return self.get_current_callable()(inputs)
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File ".../site-packages/torch/_inductor/compile_fx.py", line 611, in run
  return model(new_inputs)
          ^^^^^^^^^^^^^^^^^
File ".../site-packages/torch/_inductor/codecache.py", line 892, in _run_from_cache
  return compiled_graph.compiled_artifact(inputs)
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "/tmp/torchinductor_theodoridis/gf/cgf64g3izdp33vfvomwklyg4wg6lvlmuxkvmlnfriqaqqb2j6wtc.py", line 175, in call
  buf0 = empty_strided((s12, 1, s4, s7, s10, s11), (s10*s11*s4*s7, s10*s11*s12*s4*s7, s10*s11*s7, s10*s11, s11, 1), device='cuda', dtype=torch.float32)
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
torch.cuda.OutOfMemoryError: CUDA out of memory. Tried to allocate 78.57 GiB. GPU 0 has a total capacity of 47.53 GiB of which 44.87 GiB is free. Including non-PyTorch memory, this process has 2.65 GiB memory in use. Of the allocated memory 1.23 GiB is allocated by PyTorch, and 580.03 MiB is reserved by PyTorch but unallocated. If reserved but unallocated memory is large try setting PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True to avoid fragmentation.  See documentation for Memory Management  (https://pytorch.org/docs/stable/notes/cuda.html#environment-variables)

AssertionError with `torch 2.3.0+cu121` and `torchvision 0.18.0+cu121`

[rank0]: Traceback (most recent call last):
[rank0]:   File "./train.py", line 384, in <module>
[rank0]:     main(cfg=cfg, distributed=distributed, gpu_id=gpu_id)
[rank0]:   File ".../site-packages/torch/distributed/elastic/multiprocessing/errors/__init__.py", line 347, in wrapper
[rank0]:     return f(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^
[rank0]:   File "./train.py", line 286, in main
[rank0]:     train_one_epoch(model=model,
[rank0]:   File "./src/engine/engine.py", line 96, in train_one_epoch
[rank0]:     loss.backward()
[rank0]:   File ".../site-packages/torch/_tensor.py", line 525, in backward
[rank0]:     torch.autograd.backward(
[rank0]:   File ".../site-packages/torch/autograd/__init__.py", line 267, in backward
[rank0]:     _engine_run_backward(
[rank0]:   File ".../site-packages/torch/autograd/graph.py", line 744, in _engine_run_backward
[rank0]:     return Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File ".../site-packages/torch/autograd/function.py", line 301, in apply
[rank0]:     return user_fn(self, *args)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^
[rank0]:   File ".../site-packages/torch/_functorch/_aot_autograd/jit_compile_runtime_wrappers.py", line 882, in backward
[rank0]:     out = call_compiled_backward()
[rank0]:           ^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File ".../site-packages/torch/_functorch/_aot_autograd/jit_compile_runtime_wrappers.py", line 831, in call_compiled_backward
[rank0]:     out = call_func_at_runtime_with_args(
[rank0]:           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File ".../site-packages/torch/_functorch/_aot_autograd/utils.py", line 113, in call_func_at_runtime_with_args
[rank0]:     out = normalize_as_list(f(args))
[rank0]:                             ^^^^^^^
[rank0]:   File ".../site-packages/torch/_dynamo/eval_frame.py", line 451, in _fn
[rank0]:     return fn(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^
[rank0]:   File ".../site-packages/torch/_dynamo/external_utils.py", line 36, in inner
[rank0]:     return fn(*args, **kwargs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^
[rank0]:   File ".../site-packages/torch/_inductor/codecache.py", line 906, in __call__
[rank0]:     return self.get_current_callable()(inputs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File ".../site-packages/torch/_inductor/compile_fx.py", line 784, in run
[rank0]:     return model(new_inputs)
[rank0]:            ^^^^^^^^^^^^^^^^^
[rank0]:   File ".../site-packages/torch/_inductor/codecache.py", line 934, in _run_from_cache
[rank0]:     return compiled_graph.compiled_artifact(inputs)
[rank0]:            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
[rank0]:   File "/tmp/torchinductor_theodoridis/xl/cxlqkkc7ap73hke3wfa5g76sk6nbmzebnv3d5v5jyb64iie2vne5.py", line 182, in call
[rank0]:     assert_size_stride(unsqueeze_86, (s4, 1, s6, 1, s8, 1), (14, 0, 2, 0, 1, 0))
[rank0]: AssertionError: expected size 12==12, stride 28==14 at dim=0

Maybe related to mask_roi_pool output_size=14? Or this warning I get with torch.use_deterministic_algorithms(False), as discussed here?

.../site-packages/torch/autograd/graph.py:744: UserWarning: Plan failed with a cudnnException: CUDNN_BACKEND_EXECUTION_PLAN_DESCRIPTOR: cudnnFinalize Descriptor Failed cudnn_status: CUDNN_STATUS_NOT_SUPPORTED (Triggered internally at ../aten/src/ATen/native/cudnn/Conv_v8.cpp:919.)
  return Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass
.../site-packages/torch/nn/modules/conv.py:456: UserWarning: Plan failed with a cudnnException: CUDNN_BACKEND_EXECUTION_PLAN_DESCRIPTOR: cudnnFinalize Descriptor Failed cudnn_status: CUDNN_STATUS_NOT_SUPPORTED (Triggered internally at ../aten/src/ATen/native/cudnn/Conv_v8.cpp:919.)
  return F.conv2d(input, weight, bias, self.stride,

ezyang commented 3 months ago

Just for clarity on your setup, did you manually patch in the change to the prebuilt binaries of torchvision to test them?

JohannesTheo commented 3 months ago

Yes, I have three separate conda envs with the mentioned torch and torchvision versions (installed from pypi with pip), and manually patched the torchvision/ops/roi_align.py files in their site-packages to match the file of #8436. This naive approach resulted in the success and errors mentioned above. Let me know if you need more info or if this was too naive and is better tested in a different way. As mentioned, I'm not very familiar with torch.compile/dynamo/inductor.

ezyang commented 3 months ago

Reopening for torch version incompatibility

ezyang commented 2 months ago

Bah, I don't have a ready to go maskrcnn setup that I can use to easily test this

ezyang commented 2 months ago

@JohannesTheo do you have a suggested way of reproducing your problems? Alternately, if you are able to do runs with TORCH_TRACE and upload them here, that would also be greatly helpful.

JohannesTheo commented 2 months ago

Hey @ezyang, I will put something together on the WE.

pytorch / vision