TransformerEngine v1.2.1 throws CuDNN frontend error on H100 GPU (AWS p5.48xlarge instance)

[sirutBuasai]

Hi, we are currently running into TransformerEngine related error when running GPT model on H100 GPU (AWS p5.48xlarge). Below is the error log

Error:

RuntimeErrorRuntimeError: RuntimeError/fsx/sbuasai/test_te/deps/TransformerEngine/transformer_engine/common/fused_attn/fused_attn_f16_arbitrary_seqlen.cu:227 in function operator(): cuDNN Error: [cudnn_frontend] Error: No execution plans built successfully.. For more information, enable cuDNN error logging by setting CUDNN_LOGERR_DBG=1 and CUDNN_LOGDEST_DBG=stderr in the environment.RuntimeError
: : : /fsx/sbuasai/test_te/deps/TransformerEngine/transformer_engine/common/fused_attn/fused_attn_f16_arbitrary_seqlen.cu:227 in function operator(): cuDNN Error: [cudnn_frontend] Error: No execution plans built successfully.. For more information, enable cuDNN error logging by setting CUDNN_LOGERR_DBG=1 and CUDNN_LOGDEST_DBG=stderr in the environment./fsx/sbuasai/test_te/deps/TransformerEngine/transformer_engine/common/fused_attn/fused_attn_f16_arbitrary_seqlen.cu:227 in function operator(): cuDNN Error: [cudnn_frontend] Error: No execution plans built successfully.. For more information, enable cuDNN error logging by setting CUDNN_LOGERR_DBG=1 and CUDNN_LOGDEST_DBG=stderr in the environment.    /fsx/sbuasai/test_te/deps/TransformerEngine/transformer_engine/common/fused_attn/fused_attn_f16_arbitrary_seqlen.cu:227 in function operator(): cuDNN Error: [cudnn_frontend] Error: No execution plans built successfully.. For more information, enable cuDNN error logging by setting CUDNN_LOGERR_DBG=1 and CUDNN_LOGDEST_DBG=stderr in the environment.

output_tensors = tex.fused_attn_fwd(

return fn(*args, **kwargs)
  File "/home/ubuntu/.conda/envs/megatron_bench/lib/python3.10/site-packages/transformer_engine/pytorch/attention.py", line 1835, in forward
RuntimeError: /fsx/sbuasai/test_te/deps/TransformerEngine/transformer_engine/common/fused_attn/fused_attn_f16_arbitrary_seqlen.cu:227 in function operator(): cuDNN Error: [cudnn_frontend] Error: No execution plans built successfully.. For more information, enable cuDNN error logging by setting CUDNN_LOGERR_DBG=1 and CUDNN_LOGDEST_DBG=stderr in the environment.
    out, aux_ctx_tensors = fused_attn_fwd(
  File "/home/ubuntu/.conda/envs/megatron_bench/lib/python3.10/site-packages/transformer_engine/pytorch/cpp_extensions/fused_attn.py", line 811, in fused_attn_fwd
    output = FusedAttnFunc.apply(
  File "/home/ubuntu/.conda/envs/megatron_bench/lib/python3.10/site-packages/torch/autograd/function.py", line 539, in apply
    output_tensors = tex.fused_attn_fwd(
RuntimeError: /fsx/sbuasai/test_te/deps/TransformerEngine/transformer_engine/common/fused_attn/fused_attn_f16_arbitrary_seqlen.cu:227 in function operator(): cuDNN Error: [cudnn_frontend] Error: No execution plans built successfully.. For more information, enable cuDNN error logging by setting CUDNN_LOGERR_DBG=1 and CUDNN_LOGDEST_DBG=stderr in the environment.
    return super().apply(*args, **kwargs)  # type: ignore[misc]
  File "/home/ubuntu/.conda/envs/megatron_bench/lib/python3.10/site-packages/transformer_engine/pytorch/attention.py", line 1616, in forward
    output_tensors = tex.fused_attn_fwd(
RuntimeError: /fsx/sbuasai/test_te/deps/TransformerEngine/transformer_engine/common/fused_attn/fused_attn_f16_arbitrary_seqlen.cu:227 in function operator(): cuDNN Error: [cudnn_frontend] Error: No execution plans built successfully.. For more information, enable cuDNN error logging by setting CUDNN_LOGERR_DBG=1 and CUDNN_LOGDEST_DBG=stderr in the environment.
    out, aux_ctx_tensors = fused_attn_fwd(
  File "/home/ubuntu/.conda/envs/megatron_bench/lib/python3.10/site-packages/transformer_engine/pytorch/cpp_extensions/fused_attn.py", line 811, in fused_attn_fwd
    output_tensors = tex.fused_attn_fwd(
RuntimeError: /fsx/sbuasai/test_te/deps/TransformerEngine/transformer_engine/common/fused_attn/fused_attn_f16_arbitrary_seqlen.cu:227 in function operator(): cuDNN Error: [cudnn_frontend] Error: No execution plans built successfully.. For more information, enable cuDNN error logging by setting CUDNN_LOGERR_DBG=1 and CUDNN_LOGDEST_DBG=stderr in the environment.
[2024-02-02 01:08:10,718] torch.distributed.elastic.multiprocessing.api: [WARNING] Sending process 24721 closing signal SIGTERM
[2024-02-02 01:08:10,750] torch.distributed.elastic.multiprocessing.api: [ERROR] failed (exitcode: 1) local_rank: 0 (pid: 24718) of binary: /home/ubuntu/.conda/envs/megatron_bench/bin/python

Steps to reproduce:

1. Create conda env with conda env create -f megatron_bench.yml and conda activate megatron_bench
2. Install flash-attnetion, TransformerEngine, apex, and Megatron-LM from source declared in install_deps.sh.
3. Update the path to the data in train.sh.
4. Run training using script using train.sh.

megatron_bench.yml:

name: megatron_bench
channels:
  - pytorch
  - nvidia
dependencies:
  - python=3.10
  - pip
  - conda:
    - python=3.10
    - pytorch=2.1.2
    - pytorch-cuda=12.1
    - torchvision
    - torchaudio

install_deps.sh:

#!/bin/bash

set -e

# ===================================================
# Set dependencies pin
# ===================================================
FLASH_ATTN_BRANCH='v2.0.4'
TE_BRANCH='v1.2.1'
APEX_HASH='6c8f384b40a596bbed960f5e8d9a808ebd0e93d8'
MEGATRON_LM_HASH='2c3468a49ed51324ae9b442e0d88416f1b29422b'

# ===================================================
# Install megatron python dependencies
# ===================================================
conda install -y regex astunparse ninja pyyaml mkl mkl-include setuptools cmake cffi typing_extensions future six requests libcurl dataclasses packaging
pip install six regex tensorboardX daal4py deepspeed pyarrow pybind11 numpy==1.23.5

# ===================================================
# Install flash-attention
# ===================================================
cd $DEPS_DIR
if [ ! -d "$DEPS_DIR/flash-attention" ]; then
  git clone -b ${FLASH_ATTN_BRANCH} https://github.com/Dao-AILab/flash-attention.git
  cd flash-attention
  python setup.py install
fi

# ===================================================
# Install TransformerEngine
# ===================================================
cd $DEPS_DIR
if [ ! -d "$DEPS_DIR/TransformerEngine" ]; then
  git clone --branch stable --recursive https://github.com/NVIDIA/TransformerEngine.git
  cd TransformerEngine
  git checkout ${TE_BRANCH}
  git submodule update --init --recursive
  export NVTE_FRAMEWORK="pytorch"
  export CUDNN_PATH=/usr/local/cuda-12.1
  export CUDNN_INCLUDE_DIR=/usr/local/cuda-12.1/include
  pip install .
fi

# ===================================================
# Install apex
# ===================================================
cd $DEPS_DIR
if [ ! -d "$DEPS_DIR/apex" ]; then
  git clone https://github.com/NVIDIA/apex.git
  cd apex
  git checkout ${APEX_HASH}
  pip install -v --disable-pip-version-check --no-cache-dir --no-build-isolation \
    --config-settings "--global-option=--cpp_ext" \
    --config-settings "--global-option=--cuda_ext" \
    ./
fi

# ===================================================
# Clone Megatron-LM for scripts
# ===================================================
cd $DEPS_DIR
if [ ! -d "$DEPS_DIR/Megatron-LM" ]; then
  git clone https://github.com/NVIDIA/Megatron-LM.git
  cd Megatron-LM
  git checkout ${MEGATRON_LM_HASH}
  cd $DEPS_DIR
fi

train.sh:

#!/bin/bash

DEPS_DIR="$(pwd)/deps"
DATASET_DIR="<DECLARE YOUR DATASET DIRECTORY HERE>"

bash install_deps.sh

export GPT_HOME="${DATASET_DIR}"
export DATASET="${DATASET_DIR}/my-gpt2_text_document/my-gpt2_text_document"
export CHECKPOINT_PATH="${DATASET_DIR}/checkpoints/gptmodel"
export VOCAB_FILE="${DATASET_DIR}/gpt2-vocab.json"
export MERGES_FILE="${DATASET_DIR}/gpt2-merges.txt"
export DATA_PATH="${DATASET_DIR}/my-gpt2_text_document/my-gpt2_text_document"
export NVTE_BIAS_GELU_NVFUSION=0
export PYTORCH_CUDA_ALLOC_CONF=max_split_size_mb:512
export CUDA_DEVICE_MAX_CONNECTIONS=1
export NCCL_DEBUG=INFO
export NCCL_PROTO=LL,LL128,Simple
export FI_PROVIDER=efa
export FI_EFA_USE_DEVICE_RDMA=1
export RDMAV_FORK_SAFE=1

# remove previous checkpoints
rm -rf ${DATASET_DIR}/checkpoints/

torchrun --nproc-per-node 8 --nnodes 1 \
  ${DEPS_DIR}/Megatron-LM/pretrain_gpt.py \
  --tensor-model-parallel-size 1 \
  --pipeline-model-parallel-size 1 \
  --sequence-parallel \
  --num-layers 24 \
  --hidden-size 1024 \
  --num-attention-heads 16 \
  --micro-batch-size 1 \
  --global-batch-size 8 \
  --seq-length 2048 \
  --max-position-embeddings 2048 \
  --train-iters 1200 \
  --lr-decay-iters 320000 \
  --save ${CHECKPOINT_PATH} \
  --load ${CHECKPOINT_PATH} \
  --data-path ${DATA_PATH} \
  --vocab-file ${VOCAB_FILE} \
  --merge-file ${MERGES_FILE} \
  --split 949,50,1 \
  --distributed-backend nccl \
  --lr 0.00015 \
  --lr-decay-style cosine \
  --min-lr 1.0e-5 \
  --weight-decay 1e-2 \
  --clip-grad 1.0 \
  --lr-warmup-fraction .01 \
  --log-interval 100 \
  --save-interval 10000 \
  --eval-interval 1000 \
  --eval-iters 10 \
  --adam-beta1 0.9 \
  --adam-beta2 0.95 \
  --init-method-std 0.006 \
  --bf16 \
  --transformer-impl transformer_engine \
  --attention-softmax-in-fp32

[ptrendx]

Hi @sirutBuasai, what is the cuDNN version you are using?

[sirutBuasai]

CuDNN installed with torch==2.1.2 is 8.9.2

(megatron_bench) ubuntu@ip-10-0-0-88:~$ python -c "import torch;print(torch.backends.cudnn.version())"
8902

[cyanguwa]

Hi @sirutBuasai , could you try upgrading to cuDNN 8.9.7+ please?

[sirutBuasai]

Will do, in the meantime, is there a TE version that is built with CuDNN 8.9.2?

[cyanguwa]

I think it's probably v0.10, but I'd rather you roll forward with cuDNN than backward with TE. There's been a lot of development in the last year or so. If it's easier, you can use the NGC pytorch container, which has the latest TE (1.3) and cuDNN (9.0): nvcr.io/nvidia/pytorch:24.01-py3

[ptrendx]

@cyanguwa I think we still should catch this error from cuDNN Frontend and just disable cuDNN's implementation of attention in this case.

[liu21yd]

@sirutBuasai Was your problem solved? Could you tell me the solution. I meet the same problem.

[sirutBuasai]

@liu21yd, We ended up using TE v0.10 but it is pretty old. I haven't tried upgrading CuDNN and TE together but that would be a place to start.

[odashi]

Recently we observed similar issues with any combinations of TE 1.4/1.7 and cuDNN 8.9.4/8.9.7. In our cases, the fused_attn test in this repository also fails, as well as the frontend toolkit (Megatron-LM) doesn't work.

Note that our operating system is Rocky, not Debian-ish ones.

For a workaround we eventually set NVTE_FUSED_ATTN=0 to disable fused attention kernels, then the issue went away.