8000 Passing device mesh to FSDP raises error: Cannot pass both process_group and device_mesh at the same time · Issue #118906 · pytorch/pytorch · GitHub
[go: up one dir, main page]
Skip to content

Passing device mesh to FSDP raises error: Cannot pass both process_group and device_mesh at the same time 8000 #118906

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom
Closed
awaelchli opened this issue Feb 1, 2024 · 0 comments
Closed

Passing device mesh to FSDP raises error: Cannot pass both process_group and device_mesh at the same time #118906

awaelchli opened this issue Feb 1, 2024 · 0 comments
Assignees
@wz337
Labels
actionable module: fsdp triaged This issue has been looked at a team member, and triaged and prioritized into an appropriate module

Comments

@awaelchli
Copy link
Contributor
awaelchli commented Feb 1, 2024 •
edited by pytorch-bot bot
Loading

🐛 Describe the bug

There is a confusing error in the input validation of FSDP when passing a device mesh in combination with auto wrap policy and sharding strategy.

import os
import torch
import torch.nn as nn
from torch.distributed.device_mesh import init_device_mesh
from torch.distributed.fsdp import FullyShardedDataParallel as FSDP, ShardingStrategy
from torch.distributed.fsdp.wrap import ModuleWrapPolicy


class ToyModel(nn.Module):
    def __init__(self):
        super(ToyModel, self).__init__()
        self.net1 = nn.Linear(10, 10)
        self.relu = nn.ReLU()
        self.net2 = nn.Linear(10, 5)

    def forward(self, x):
        return self.net2(self.relu(self.net1(x)))


policy = ModuleWrapPolicy({torch.nn.Linear})
mesh_2d = init_device_mesh("cuda", (2, 4))
model = FSDP(
    ToyModel(), 
    device_mesh=mesh_2d,
    # Either not passing sharding strategy or not passing autowrap policy fixes the issue
    sharding_strategy=ShardingStrategy.HYBRID_SHARD,
    auto_wrap_policy=policy,
    device_id=int(os.environ["LOCAL_RANK"]),
)

Error:

[rank4]: Traceback (most recent call last):
[rank4]:   File "/home/adrian/repositories/lightning/examples/mesh_torch.py", line 22, in <module>
[rank4]:     model = FSDP(
[rank4]:   File "/home/adrian/.conda/envs/lightning/lib/python3.10/site-packages/torch/distributed/fsdp/fully_sharded_data_parallel.py", line 481, in __init__
[rank4]:     _auto_wrap(
[rank4]:   File "/home/adrian/.conda/envs/lightning/lib/python3.10/site-packages/torch/distributed/fsdp/_wrap_utils.py", line 72, in _auto_wrap
[rank4]:     _post_order_apply(root_module, wrap_fn)
[rank4]:   File "/home/adrian/.conda/envs/lightning/lib/python3.10/site-packages/torch/distributed/fsdp/wrap.py", line 79, in _post_order_apply
[rank4]:     _post_order_apply_inner(root_module, "", None)
[rank4]:   File "/home/adrian/.conda/envs/lightning/lib/python3.10/site-packages/torch/distributed/fsdp/wrap.py", line 63, in _post_order_apply_inner
[rank4]:     _post_order_apply_inner(child_module, child_module_name, module)
[rank4]:   File "/home/adrian/.conda/envs/lightning/lib/python3.10/site-packages/torch/distributed/fsdp/wrap.py", line 64, in _post_order_apply_inner
[rank4]:     optional_module = fn(module)
[rank4]:   File "/home/adrian/.conda/envs/lightning/lib/python3.10/site-packages/torch/distributed/fsdp/wrap.py", line 98, in fn
[rank4]:     return fsdp_fn(module, **kwargs)
[rank4]:   File "/home/adrian/.conda/envs/lightning/lib/python3.10/site-packages/torch/distributed/fsdp/fully_sharded_data_parallel.py", line 452, in __init__
[rank4]:     _init_process_group_state(
[rank4]:   File "/home/adrian/.conda/envs/lightning/lib/python3.10/site-packages/torch/distributed/fsdp/_init_utils.py", line 108, in _init_process_group_state
[rank4]:     raise ValueError(
[rank4]: ValueError: Cannot pass both process_group and device_mesh at the same time. Please just pass only one of them.

Either not passing sharding strategy or not passing autowrap policy fixes the issue. I don't see a good reason why the combination (sharding_strategy, device_mesh, auto_wrap_policy) shouldn't be allowed.

I encountered this issue while working with the official PyTorch tutorial. There, they pass both sharding strategy and device mesh, but no auto wrap policy. I think it is pretty common to pass a wrap policy.

Versions

Collecting environment information...
PyTorch version: 2.3.0.dev20240201+cu121
Is debug build: False
CUDA used to build PyTorch: 12.1
ROCM used to build PyTorch: N/A

OS: Ubuntu 22.04.3 LTS (x86_64)
GCC version: (Ubuntu 11.4.0-1ubuntu1~22.04) 11.4.0
Clang version: Could not collect
CMake version: version 3.26.0
Libc version: glibc-2.35

Python version: 3.10.9 (main, Jan 11 2023, 15:21:40) [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-5.15.0-75-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 A100-SXM4-40GB
GPU 1: NVIDIA A100-SXM4-40GB
GPU 2: NVIDIA A100-SXM4-40GB
GPU 3: NVIDIA A100-SXM4-40GB
GPU 4: NVIDIA A100-SXM4-40GB
GPU 5: NVIDIA A100-SXM4-40GB
GPU 6: NVIDIA A100-SXM4-40GB
GPU 7: NVIDIA A100-SXM4-40GB

Nvidia driver version: 525.125.06
cuDNN version: Probably one of the following:
/usr/lib/x86_64-linux-gnu/libcudnn.so.8.9.5
/usr/lib/x86_64-linux-gnu/libcudnn_adv_infer.so.8.9.5
/usr/lib/x86_64-linux-gnu/libcudnn_adv_train.so.8.9.5
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_infer.so.8.9.5
/usr/lib/x86_64-linux-gnu/libcudnn_cnn_train.so.8.9.5
/usr/lib/x86_64-linux-gnu/libcudnn_ops_infer.so.8.9.5
/usr/lib/x86_64-linux-gnu/libcudnn_ops_train.so.8.9.5
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: 43 bits physical, 48 bits virtual
Byte Order: Little Endian
CPU(s): 256
On-line CPU(s) list: 0-254
Off-line CPU(s) list: 255
Vendor ID: AuthenticAMD
Model name: AMD EPYC 7742 64-Core Processor
CPU family: 23
Model: 49
Thread(s) per core: 2
Core(s) per socket: 64
Socket(s): 2
Stepping: 0
Frequency boost: enabled
CPU max MHz: 2250.0000
CPU min MHz: 0.0000
BogoMIPS: 4499.83
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 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 hw_pstate ssbd mba ibrs ibpb stibp vmmcall fsgsbase bmi1 avx2 smep bmi2 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 arat npt lbrv svm_lock nrip_save tsc_scale vmcb_clean flushbyasid decodeassists pausefilter pfthreshold avic v_vmsave_vmload vgif v_spec_ctrl umip rdpid overflow_recov succor smca sme sev sev_es
Virtualization: AMD-V
L1d cache: 4 MiB (128 instances)
L1i cache: 4 MiB (128 instances)
L2 cache: 64 MiB (128 instances)
L3 cache: 512 MiB (32 instances)
NUMA node(s): 2
NUMA node0 CPU(s): 0-63,128-191
NUMA node1 CPU(s): 64-127,192-254
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: Mitigation; untrained return thunk; SMT enabled with STIBP protection
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; Retpolines, IBPB conditional, STIBP always-on, RSB filling, PBRSB-eIBRS Not affected
Vulnerability Srbds: Not affected
Vulnerability Tsx async abort: Not affected

Versions of relevant libraries:
[pip3] mypy==1.4.1
[pip3] mypy-extensions==1.0.0
[pip3] numpy==1.24.2
[pip3] onnx==1.12.0
[pip3] onnxruntime==1.14.1
[pip3] pytorch-lightning==2.0.6
[pip3] pytorch-triton==3.0.0+901819d2b6
[pip3] torch==2.3.0.dev20240201+cu121
[pip3] torch-tb-profiler==0.4.3
[pip3] torchmetrics==1.3.0.post0
[pip3] torchvision==0.17.0
[pip3] triton==2.2.0
[conda] numpy 1.24.2 pypi_0 pypi
[conda] pytorch-lightning 2.0.6 pypi_0 pypi
[conda] pytorch-triton 3.0.0+901819d2b6 pypi_0 pypi
[conda] torch 2.3.0.dev20240201+cu121 pypi_0 pypi
[conda] torch-tb-profiler 0.4.3 pypi_0 pypi
[conda] torchmetrics 1.3.0.post0 pypi_0 pypi
[conda] torchvision 0.17.0 pypi_0 pypi
[conda] triton 2.2.0 pypi_0 pypi

cc @zhaojuanmao @mrshenli @rohan-varma @awgu @fegin @penguinwu @kwen2501
@awgu awgu added triaged This issue has been looked at a team member, and triaged and prioritized into an appropriate module actionable module: fsdp labels Feb 1, 2024
@awgu awgu mentioned this issue Feb 2, 2024
Closed
@wz337 wz337 self-assigned this Feb 2, 2024
@awgu awgu mentioned this issue Feb 2, 2024
Closed
pytorch-bot bot pushed a commit that referenced this issue Feb 8, 2024
@jeffdaily
[FSDP] Fixed device_mesh and auto wrap (#119064)
fce5a24 
If the user passes `device_mesh`, then we should not forward the process groups to the children during auto wrap and instead just rely on the `device_mesh` argument. This should fix #118906.

Pull Request resolved: #119064
Approved by: https://github.com/wz337
mvpatel2000 pushed a commit to mvpatel2000/pytorch that referenced this issue Feb 13, 2024
@mvpatel2000
[FSDP] Fixed device_mesh and auto wrap (pytorch#119064)
5299774 
If the user passes `device_mesh`, then we should not forward the process groups to the children during auto wrap and instead just rely on the `device_mesh` argument. This should fix pytorch#118906.

Pull Request resolved: pytorch#119064
Approved by: https://github.com/wz337
atalman pushed a commit that referenced this issue Feb 14, 2024
@mvpatel2000
HSDP + TP integration bug fixes (#119819)
0aa3fd3 
Co-authored-by: Andrew Gu <andgu@fb.com>
resolved: #112435
resolved: #118620
Fixed `device_mesh` and auto wrap (#119064)
fix #118906.
resolved: #119064
resolved: #118638
Fixes #118639.
resolved: #119481
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Assignees

@wz337 wz337

Labels
actionable module: fsdp triaged This issue has been looked at a team member, and triaged and prioritized into an appropriate module
Projects
None yet
Milestone
No milestone
Development

No branches or pull requests
3 participants
@awaelchli @awgu @wz337
0