Open alexisVallet opened 4 months ago
same problem with cuda 12.2
If that may help other people encountering this issue. I encounter this issue when training using DistributedDataParallel
when using bfloat16
weights. The best workaround I have found is to convert the gradients to float32
before all_reduce
, using code like this:
import torch
import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel
# Based on code for:
# https://pytorch.org/docs/stable/ddp_comm_hooks.html#torch.distributed.algorithms.ddp_comm_hooks.default_hooks.fp16_compress_hook
def convert_grad_float32(
process_group: Optional[dist.ProcessGroup], bucket: dist.GradBucket
) -> torch.futures.Future[torch.Tensor]:
group_to_use = (
process_group if process_group is not None else dist.group.WORLD
)
world_size = group_to_use.size()
tensor = bucket.buffer()
orig_dtype = tensor.dtype
if orig_dtype == torch.bfloat16:
tensor = tensor.to(torch.float32)
tensor.div_(world_size)
fut = dist.all_reduce(
tensor, group=group_to_use, async_op=True
).get_future()
def decode(fut):
out_tensor = bucket.buffer()
out_tensor.copy_(fut.value()[0])
return out_tensor
return fut.then(decode)
model_dist = DistributedDataParallel(model)
model_dist.register_comm_hook(None, convert_grad_float32)
same issue here, version: cuda==12.2 torch==2.2.2 , only occurs when training with DDP and 4 or 8 or more GPUs
so I guess if there is some alignment issues for bf16 and specific input shapes, like this issue perhaps: https://github.com/Dao-AILab/flash-attention/issues/289 since DDP training with bf16 is a very common user case but this issue is reported not so often
Same problem. It probably has something to do with hardware, since I met this problem when switching from A100 to H100 with code unchanged.
🐛 Describe the bug
Hi! I am encountering the following error when using
torch.distributed.all_reduce
on bfloat16 tensors of a certain size using NCCL:RuntimeError: CUDA error: misaligned address
.I can only reproduce this with a large enough number of GPUs - in my environment this occurs with 8 GPUs but not with 2 GPUs for instance. This issue seems environment specific, as I could not reproduce it everywhere I tried, though I can reproduce it on multiple nodes of the same cluster. I believe this is likely a NCCL bug, but I could not reproduce it consistently with nccl-tests for instance, only with Pytorch. EDIT: after further testing, I could also reproduce this consistently with nccl-test.
Some more potentially relevant environment information that
collect_env.py
didn't pick up:I also tried various driver versions, cuda versions, pytorch versions, but this error always occurs with some bfloat16 tensor size (though not necessarily the one in this example).
Minimal example to reproduce on my environment:
Running it with:
Results in this log:
Versions
$ python3 collect_env.py Collecting environment information... PyTorch version: 2.2.0+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 3.22.1 Libc version: glibc-2.35
Python version: 3.10.12 (main, Nov 20 2023, 15:14:05) [GCC 11.4.0] (64-bit runtime) Python platform: Linux-5.15.0-92-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 H100 80GB HBM3 GPU 1: NVIDIA H100 80GB HBM3 GPU 2: NVIDIA H100 80GB HBM3 GPU 3: NVIDIA H100 80GB HBM3 GPU 4: NVIDIA H100 80GB HBM3 GPU 5: NVIDIA H100 80GB HBM3 GPU 6: NVIDIA H100 80GB HBM3 GPU 7: NVIDIA H100 80GB HBM3
Nvidia driver version: 545.23.08 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, 57 bits virtual Byte Order: Little Endian CPU(s): 112 On-line CPU(s) list: 0-111 Vendor ID: GenuineIntel Model name: Intel(R) Xeon(R) Platinum 8480+ CPU family: 6 Model: 143 Thread(s) per core: 1 Core(s) per socket: 56 Socket(s): 2 Stepping: 8 Frequency boost: enabled CPU max MHz: 2001.0000 CPU min MHz: 800.0000 BogoMIPS: 4000.00 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 pcid dca sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand lahf_lm abm 3dnowprefetch cpuid_fault epb cat_l3 cat_l2 cdp_l3 invpcid_single intel_ppin cdp_l2 ssbd mba ibrs ibpb stibp ibrs_enhanced tpr_shadow vnmi flexpriority ept vpid ept_ad fsgsbase tsc_adjust bmi1 avx2 smep bmi2 erms invpcid cqm rdt_a avx512f avx512dq rdseed adx smap avx512ifma clflushopt clwb intel_pt avx512cd sha_ni avx512bw avx512vl xsaveopt xsavec xgetbv1 xsaves cqm_llc cqm_occup_llc cqm_mbm_total cqm_mbm_local split_lock_detect avx_vnni avx512_bf16 wbnoinvd dtherm ida arat pln pts avx512vbmi umip pku ospke waitpkg avx512_vbmi2 gfni vaes vpclmulqdq avx512_vnni avx512_bitalg tme avx512_vpopcntdq la57 rdpid bus_lock_detect cldemote movdiri movdir64b enqcmd fsrm md_clear serialize tsxldtrk pconfig arch_lbr amx_bf16 avx512_fp16 amx_tile amx_int8 flush_l1d arch_capabilities Virtualization: VT-x L1d cache: 5.3 MiB (112 instances) L1i cache: 3.5 MiB (112 instances) L2 cache: 224 MiB (112 instances) L3 cache: 210 MiB (2 instances) NUMA node(s): 2 NUMA node0 CPU(s): 0-55 NUMA node1 CPU(s): 56-111 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
Versions of relevant libraries: [pip3] numpy==1.26.4 [pip3] torch==2.2.0 [pip3] torchvision==0.17.0 [pip3] triton==2.2.0 [conda] Could not collect
cc @ptrblck