reduce_scatter raises a RuntimeError

[garrett361]

Describe the bug

Calling into torch.distributed.reduce_scatter raises a RuntimeError.

Minimal code to reproduce:

import os

import torch
import torch.distributed as dist
import intel_extension_for_pytorch as ipex  # noqa
import oneccl_bindings_for_pytorch  # noqa


def main() -> None:
    world_size = int(os.environ["WORLD_SIZE"])
    local_rank = int(os.environ["LOCAL_RANK"])
    device = torch.device(f"xpu:{local_rank}")
    torch.xpu.set_device(device)

    try:
        dist.init_process_group("ccl")

        t_in = [torch.empty(1, device=device) for _ in range(world_size)]
        t_out = torch.empty(1, device=device)

        dist.reduce_scatter(t_out, t_in, op=dist.ReduceOp.SUM)

    finally:
        dist.destroy_process_group()


if __name__ == "__main__":
    main()

Result:

  File "/opt/aurora/23.275.2/frameworks/aurora_nre_models_frameworks-2024.0/lib/python3.9/site-packages/torch/distributed/distributed_c10d.py", line 3288, in reduce_scatter
    work = default_pg.reduce_scatter([output], [input_list], opts)
RuntimeError: ProcessGroupCCL does not support reduce_scatter

Also reported in the comments of #640, but raising as a separate issue since it's an independent error.

Versions

Tested against several versions, including:

Collecting environment information...
PyTorch version: 2.1.0a0+cxx11.abi
PyTorch CXX11 ABI: Yes
IPEX version: 2.1.10+xpu
IPEX commit: a12f9f650
Build type: Release

OS: SUSE Linux Enterprise Server 15 SP4 (x86_64)
GCC version: (Spack GCC) 12.2.0
Clang version: N/A
IGC version: 2024.0.0 (2024.0.0.20231017)
CMake version: version 3.27.5
Libc version: glibc-2.31

Python version: 3.9.18 (tags/v3.9.18-26-g6b320c3b2f6-dirty:6b320c3b2f6, Sep 28 2023, 00:35:27)  [GCC 13.2.0] (64-bit runtime)
Python platform: Linux-5.14.21-150400.24.55-default-x86_64-with-glibc2.31
Is XPU available: True
DPCPP runtime version: latest
MKL version: latest
GPU models and configuration: 
[0] _DeviceProperties(name='Intel(R) Data Center GPU Max 1550', platform_name='Intel(R) Level-Zero', dev_type='gpu, support_fp64=1, total_memory=65536MB, max_compute_units=448, gpu_eu_count=448)
[1] _DeviceProperties(name='Intel(R) Data Center GPU Max 1550', platform_name='Intel(R) Level-Zero', dev_type='gpu, support_fp64=1, total_memory=65536MB, max_compute_units=448, gpu_eu_count=448)
[2] _DeviceProperties(name='Intel(R) Data Center GPU Max 1550', platform_name='Intel(R) Level-Zero', dev_type='gpu, support_fp64=1, total_memory=65536MB, max_compute_units=448, gpu_eu_count=448)
[3] _DeviceProperties(name='Intel(R) Data Center GPU Max 1550', platform_name='Intel(R) Level-Zero', dev_type='gpu, support_fp64=1, total_memory=65536MB, max_compute_units=448, gpu_eu_count=448)
[4] _DeviceProperties(name='Intel(R) Data Center GPU Max 1550', platform_name='Intel(R) Level-Zero', dev_type='gpu, support_fp64=1, total_memory=65536MB, max_compute_units=448, gpu_eu_count=448)
[5] _DeviceProperties(name='Intel(R) Data Center GPU Max 1550', platform_name='Intel(R) Level-Zero', dev_type='gpu, support_fp64=1, total_memory=65536MB, max_compute_units=448, gpu_eu_count=448)
[6] _DeviceProperties(name='Intel(R) Data Center GPU Max 1550', platform_name='Intel(R) Level-Zero', dev_type='gpu, support_fp64=1, total_memory=65536MB, max_compute_units=448, gpu_eu_count=448)
[7] _DeviceProperties(name='Intel(R) Data Center GPU Max 1550', platform_name='Intel(R) Level-Zero', dev_type='gpu, support_fp64=1, total_memory=65536MB, max_compute_units=448, gpu_eu_count=448)
[8] _DeviceProperties(name='Intel(R) Data Center GPU Max 1550', platform_name='Intel(R) Level-Zero', dev_type='gpu, support_fp64=1, total_memory=65536MB, max_compute_units=448, gpu_eu_count=448)
[9] _DeviceProperties(name='Intel(R) Data Center GPU Max 1550', platform_name='Intel(R) Level-Zero', dev_type='gpu, support_fp64=1, total_memory=65536MB, max_compute_units=448, gpu_eu_count=448)
[10] _DeviceProperties(name='Intel(R) Data Center GPU Max 1550', platform_name='Intel(R) Level-Zero', dev_type='gpu, support_fp64=1, total_memory=65536MB, max_compute_units=448, gpu_eu_count=448)
[11] _DeviceProperties(name='Intel(R) Data Center GPU Max 1550', platform_name='Intel(R) Level-Zero', dev_type='gpu, support_fp64=1, total_memory=65536MB, max_compute_units=448, gpu_eu_count=448)
Intel OpenCL ICD version: N/A
Level Zero version: N/A

CPU:
Architecture:                    x86_64
CPU op-mode(s):                  32-bit, 64-bit
Address sizes:                   52 bits physical, 57 bits virtual
Byte Order:                      Little Endian
CPU(s):                          208
On-line CPU(s) list:             0-207
Vendor ID:                       GenuineIntel
Model name:                      Intel(R) Xeon(R) CPU Max 9470C
CPU family:                      6
Model:                           143
Thread(s) per core:              2
Core(s) per socket:              52
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 hle avx2 smep bmi2 erms invpcid rtm 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 hwp hwp_act_window hwp_epp hwp_pkg_req 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 avx512_fp16 amx_tile flush_l1d arch_capabilities
Virtualization:                  VT-x
L1d cache:                       4.9 MiB (104 instances)
L1i cache:                       3.3 MiB (104 instances)
L2 cache:                        208 MiB (104 instances)
L3 cache:                        210 MiB (2 instances)
NUMA node(s):                    4
NUMA node0 CPU(s):               0-51,104-155
NUMA node1 CPU(s):               52-103,156-207
NUMA node2 CPU(s):               
NUMA node3 CPU(s):               
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 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] intel-extension-for-pytorch==2.1.10+xpu
[pip3] numpy==1.24.4
[pip3] torch==2.1.0a0+cxx11.abi
[pip3] torch-tb-profiler==0.4.3
[pip3] torchvision==0.16.2
[conda] intel-extension-for-pytorch 2.1.10+xpu               pypi_0    pypi
[conda] mkl                       2024.0.0            intel_49630    intel
[conda] mkl-dpcpp                 2024.0.0            intel_49630    intel
[conda] mkl-service               2.4.0           py39h3539a15_40    intel
[conda] mkl_fft                   1.3.6           py39h1d81ff8_61    intel
[conda] mkl_random                1.2.2           py39h5a378b4_81    intel
[conda] mkl_umath                 0.1.1           py39h2b1685c_91    intel
[conda] numpy                     1.23.5                   pypi_0    pypi
[conda] onemkl-sycl-blas          2024.0.0            intel_49630    intel
[conda] onemkl-sycl-datafitting   2024.0.0            intel_49630    intel
[conda] onemkl-sycl-dft           2024.0.0            intel_49630    intel
[conda] onemkl-sycl-lapack        2024.0.0            intel_49630    intel
[conda] onemkl-sycl-rng           2024.0.0            intel_49630    intel
[conda] onemkl-sycl-sparse        2024.0.0            intel_49630    intel
[conda] onemkl-sycl-stats         2024.0.0            intel_49630    intel
[conda] onemkl-sycl-vm            2024.0.0            intel_49630    intel
[conda] torch                     2.1.0a0+cxx11.abi          pypi_0    pypi
[conda] torchvision               0.16.2                   pypi_0    pypi

[Vasud-ha]

Hi @garrett361, thanks for reporting it we will try to reproduce it at our end and return back to you.

[garrett361]

This RuntimeError is raised from in torch-ccl, where the absence of reduce_scatter is just hard-coded.

E.g. in torch-ccl 2.1.300+xpu:

c10::intrusive_ptr<C10D_Work> ProcessGroupCCL::reduce_scatter(
    std::vector<at::Tensor>& /* unused */,
    std::vector<std::vector<at::Tensor>>& /* unused */,
    const ReduceScatterOptions& /* unused */)
{
  TORCH_CHECK(false, "ProcessGroupCCL does not support reduce_scatter");  <---- Source of the exception
}

[Vasud-ha]

Hi @garrett361 , I am not able to reproduce the issue getting this key value error

[garrett361]

Hi @Vasud-ha , WORLD_SIZE, RANK, etc. are the environment variables torch expects to be set in a distributed setup. Are you using multiple GPUs? If so, how are you launching the processes? torchrun is the native pytorch launcher, but you can also use mpirun and a custom script.

You can also launch with multiprocessing as in the gist here. This is probably the easiest thing to try.

[Vasud-ha]

Thanks @garrett361, this issue is reproducible, will try to debug it and share the updates.

[garrett361]

There's not much to debug: per the above torch.distributed.reduce_scatter isn't implemented in torch-ccl. So, the implementation just needs to be written.

[Vasud-ha]

Thanks @garrett361, we will check this with development team.

[Vasud-ha]

Hi @garrett361, it's a known issue in 2.1.300, it happens as you pass a vector tensor as input which is not supported in the 2.1.300 release. You can work around this issue by passing a big tensor not a vector as input. This known feature will be fixed in the next release.

[Vasud-ha]

Hi @garrett361, let me know if I can close this issue, thank you.