FSDP2 + TP does not work

[ChenchaoZhao]

🐛 Describe the bug

I implemented FSDP2 + TP model according to torchtitan implementations. But I used a regular optimizer AdamW. I used gradient norm clipping but I got following error
NotImplementedError: aten._foreach_norm.Scalar: DTensor does not support cross-mesh operation yet! Got meshes: DeviceMesh('cuda', [[3]], mesh_dim_names=('dp', 'fsdp')) DeviceMesh('cuda', [[[0, 1, 2, 3, 4, 5, 6, 7]]], mesh_dim_names=('dp', 'fsdp', 'tp'))

I also tried the norm clipping function in torchtitan which gathers DTensor before clipping the norm but still it did not work for me.

If I remove norm clipping, the backward step also complained about DTensor does not support cross-mesh operation yet!

Versions

PyTorch version: 2.6.0+cu124
Is debug build: False
CUDA used to build PyTorch: 12.4
ROCM used to build PyTorch: N/A

OS: Amazon Linux 2 (x86_64)
GCC version: (GCC) 7.3.1 20180712 (Red Hat 7.3.1-17)
Clang version: Could not collect
CMake version: Could not collect
Libc version: glibc-2.26

Python version: 3.11.10 (main, Oct  3 2024, 07:29:13) [GCC 11.2.0] (64-bit runtime)
Python platform: Linux-5.10.237-211.949.amzn2int.x86_64-x86_64-with-glibc2.26
Is CUDA available: True
CUDA runtime version: Could not collect
CUDA_MODULE_LOADING set to: LAZY
GPU models and configuration: 
GPU 0: Tesla V100-SXM2-16GB
GPU 1: Tesla V100-SXM2-16GB
GPU 2: Tesla V100-SXM2-16GB
GPU 3: Tesla V100-SXM2-16GB
GPU 4: Tesla V100-SXM2-16GB
GPU 5: Tesla V100-SXM2-16GB
GPU 6: Tesla V100-SXM2-16GB
GPU 7: Tesla V100-SXM2-16GB

Nvidia driver version: 550.54.15
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
CPU(s):              64
On-line CPU(s) list: 0-63
Thread(s) per core:  2
Core(s) per socket:  16
Socket(s):           2
NUMA node(s):        2
Vendor ID:           GenuineIntel
CPU family:          6
Model:               79
Model name:          Intel(R) Xeon(R) CPU E5-2686 v4 @ 2.30GHz
Stepping:            1
CPU MHz:             2493.709
CPU max MHz:         3000.0000
CPU min MHz:         1200.0000
BogoMIPS:            4600.01
Hypervisor vendor:   Xen
Virtualization type: full
L1d cache:           32K
L1i cache:           32K
L2 cache:            256K
L3 cache:            46080K
NUMA node0 CPU(s):   0-15,32-47
NUMA node1 CPU(s):   16-31,48-63
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 pdpe1gb rdtscp lm constant_tsc arch_perfmon rep_good nopl xtopology nonstop_tsc cpuid aperfmperf tsc_known_freq pni pclmulqdq monitor est ssse3 fma cx16 pcid sse4_1 sse4_2 x2apic movbe popcnt tsc_deadline_timer aes xsave avx f16c rdrand hypervisor lahf_lm abm 3dnowprefetch cpuid_fault invpcid_single pti fsgsbase bmi1 hle avx2 smep bmi2 erms invpcid rtm rdseed adx xsaveopt ida

Versions of relevant libraries:
[pip3] mypy==1.13.0
[pip3] mypy-extensions==1.0.0
[pip3] numpy==1.26.4
[pip3] nvidia-cublas-cu12==12.4.5.8
[pip3] nvidia-cuda-cupti-cu12==12.4.127
[pip3] nvidia-cuda-nvrtc-cu12==12.4.127
[pip3] nvidia-cuda-runtime-cu12==12.4.127
[pip3] nvidia-cudnn-cu12==9.1.0.70
[pip3] nvidia-cufft-cu12==11.2.1.3
[pip3] nvidia-curand-cu12==10.3.5.147
[pip3] nvidia-cusolver-cu12==11.6.1.9
[pip3] nvidia-cusparse-cu12==12.3.1.170
[pip3] nvidia-cusparselt-cu12==0.6.2
[pip3] nvidia-nccl-cu12==2.21.5
[pip3] nvidia-nvjitlink-cu12==12.4.127
[pip3] nvidia-nvtx-cu12==12.4.127
[pip3] torch==2.6.0
[pip3] torchprofile==0.0.4
[pip3] torchtext==0.18.0
[pip3] torchvision==0.21.0
[pip3] triton==3.2.0
[conda] numpy                     1.26.4                   pypi_0    pypi
[conda] nvidia-cublas-cu12        12.4.5.8                 pypi_0    pypi
[conda] nvidia-cuda-cupti-cu12    12.4.127                 pypi_0    pypi
[conda] nvidia-cuda-nvrtc-cu12    12.4.127                 pypi_0    pypi
[conda] nvidia-cuda-runtime-cu12  12.4.127                 pypi_0    pypi
[conda] nvidia-cudnn-cu12         9.1.0.70                 pypi_0    pypi
[conda] nvidia-cufft-cu12         11.2.1.3                 pypi_0    pypi
[conda] nvidia-curand-cu12        10.3.5.147               pypi_0    pypi
[conda] nvidia-cusolver-cu12      11.6.1.9                 pypi_0    pypi
[conda] nvidia-cusparse-cu12      12.3.1.170               pypi_0    pypi
[conda] nvidia-cusparselt-cu12    0.6.2                    pypi_0    pypi
[conda] nvidia-nccl-cu12          2.21.5                   pypi_0    pypi
[conda] nvidia-nvjitlink-cu12     12.4.127                 pypi_0    pypi
[conda] nvidia-nvtx-cu12          12.4.127                 pypi_0    pypi
[conda] torch                     2.6.0                    pypi_0    pypi
[conda] torchprofile              0.0.4                    pypi_0    pypi
[conda] torchtext                 0.18.0                   pypi_0    pypi
[conda] torchvision               0.21.0                   pypi_0    pypi
[conda] triton                    3.2.0                    pypi_0    pypi
`

cc @H-Huang @awgu @wanchaol @fegin @fduwjj @wz337 @wconstab @d4l3k

[weifengpy]

DeviceMesh('cuda', [[3]], mesh_dim_names=('dp', 'fsdp')) DeviceMesh('cuda', [[[0, 1, 2, 3, 4, 5, 6, 7]]], mesh_dim_names=('dp', 'fsdp', 'tp'))

Hi @ChenchaoZhao , thanks for the report. do you mind share code on how FSDP/TP are applied? I want to understand why there are both mesh_dim_names=('dp', 'fsdp') and mesh_dim_names=('dp', 'fsdp', 'tp')). Is it applying TP to some submodules but the rest does not have TP? also want ot understand how 'dp' differs from 'fsdp' dim

I also tried the norm clipping function in torchtitan which gathers DTensor before clipping the norm but still it did not work for me.

this is explainable as ('dp', 'fsdp') and ('dp', 'fsdp', 'tp') cannot be added

If I remove norm clipping, the backward step also complained about DTensor does not support cross-mesh operation yet!

that's not expected. thus I still want to see the code on how FSDP/TP are applied

[ChenchaoZhao]

Thank you for the quick response! Sure I can add some minimal code later.

That's correct I only applied TP to some modules. I have a small vocab size so I didn't apply TP to embedding and output heads. I also had some issues with sequence parallel so I didn't apply TP to norm layers either.

Should I apply TP to every module but use replicate instead of shard for those modules that don't need sharding?

Thanks!

[weifengpy]

Should I apply TP to every module but use replicate instead of shard for those modules that don't need sharding?

@ChenchaoZhao yes, that would be the quickest way to get unblocked, because every model.parameters() will be in ('dp', 'fsdp', 'tp')

@tianyu-l Are you hitting the same issue for EP? I can imagine EP modules having ('fsdp', 'tp', 'ep') and other modules having ('fsdp', 'tp'). we have to split them into two groups for gradient clipping and optimizer

[ChenchaoZhao]

Cool! I will try it out and report back. Thanks!

[tianyu-l]

@ChenchaoZhao
For unblock selective TP, I used a NoParallel basically doing what you said https://github.com/pytorch/torchtitan/blob/main/torchtitan/experiments/llama4/infra/parallelize.py#L171

For EP it's a bit different IMO where in nature experts / non-experts are on different meshes, so I had to split them into groups.