search

NVIDIA / Megatron-LM

Ongoing research training transformer models at scale
https://docs.nvidia.com/megatron-core/developer-guide/latest/user-guide/index.html#quick-start
Other
10.66k stars 2.39k forks source link

[BUG] Context parallel gives NCCL error #1151

Open YJHMITWEB opened 2 months ago

YJHMITWEB commented 2 months ago

Describe the bug I am using the train_gpt3_175b_distributed.sh script to launch training on a single node with 4 A100 80GB GPUs. Training goes well if I use tensor parallel or pipeline parallel, but fails if I enable context parallel. The following is my script:

#!/bin/bash

# Runs the "175B" parameter model

export CUDA_DEVICE_MAX_CONNECTIONS=1

GPUS_PER_NODE=4
# Change for multinode config
MASTER_ADDR=localhost
MASTER_PORT=6000
NUM_NODES=1
NODE_RANK=0
WORLD_SIZE=$(($GPUS_PER_NODE*$NUM_NODES))

CHECKPOINT_PATH=$1 #<Specify path>
TENSORBOARD_LOGS_PATH=$2 #<Specify path>
VOCAB_FILE=$3 #<Specify path to file>/gpt2-vocab.json
MERGE_FILE=$4 #<Specify path to file>/gpt2-merges.txt
DATA_PATH=$5 #<Specify path and file prefix>_text_document

DISTRIBUTED_ARGS=(
    --nproc_per_node $GPUS_PER_NODE 
    --nnodes $NUM_NODES 
    --master_addr $MASTER_ADDR 
    --master_port $MASTER_PORT
)

GPT_MODEL_ARGS=(
    --num-layers 32
    --hidden-size 2048
    --ffn-hidden-size 8192 
    --num-attention-heads 16 
    --seq-length 2048
    --max-position-embeddings 2048
)

TRAINING_ARGS=(
    --micro-batch-size 1 
    --global-batch-size 1 
    --train-iters 500000 
    --weight-decay 0.1 
    --adam-beta1 0.9 
    --adam-beta2 0.95 
    --init-method-std 0.006 
    --clip-grad 1.0 
    --fp16
    --lr 6.0e-5 
    --lr-decay-style cosine 
    --min-lr 6.0e-6
    --lr-warmup-fraction .001 
    --lr-decay-iters 430000 
    --recompute-activations
)

MODEL_PARALLEL_ARGS=(
    --tensor-model-parallel-size 1 
    --pipeline-model-parallel-size 1
    --context-parallel-size 4 
)

DATA_ARGS=(
    --data-path "/temp_document"
    --vocab-file "/gpt2-vocab.json"
    --merge-file "/gpt2-merges.txt"
    --split 949,50,1
)

EVAL_AND_LOGGING_ARGS=(
    --log-interval 100
    --save-interval 10000 
    --eval-interval 1000 
    --save "/temp/"
    --load "/temp/"
    --eval-iters 10
    --tensorboard-dir "/Megatron-LM/examples/gpt3"
)

torchrun ${DISTRIBUTED_ARGS[@]} pretrain_gpt.py \
    ${GPT_MODEL_ARGS[@]} \
    ${TRAINING_ARGS[@]} \
    ${MODEL_PARALLEL_ARGS[@]} \
    ${DATA_ARGS[@]} \
    ${EVAL_AND_LOGGING_ARGS[@]}

The output log is:

building GPT model ...
 > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 1718685696
 > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 1718685696
 > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 1718685696
 > number of parameters on (tensor, pipeline) model parallel rank (0, 0): 1718685696
WARNING: could not find the metadata file /temp/latest_checkpointed_iteration.txt
    will not load any checkpoints and will start from random
/pytorch/torch/distributed/c10d_logger.py:83: FutureWarning: `torch.distributed._all_gather_base` is a private function and will be deprecated. Please use `torch.distributed.all_gather_into_tensor` instead.
  return func(*args, **kwargs)
/pytorch/torch/distributed/c10d_logger.py:83: FutureWarning: `torch.distributed._all_gather_base` is a private function and will be deprecated. Please use `torch.distributed.all_gather_into_tensor` instead.
  return func(*args, **kwargs)
/pytorch/torch/distributed/c10d_logger.py:83: FutureWarning: `torch.distributed._all_gather_base` is a private function and will be deprecated. Please use `torch.distributed.all_gather_into_tensor` instead.
  return func(*args, **kwargs)
/pytorch/torch/distributed/c10d_logger.py:83: FutureWarning: `torch.distributed._all_gather_base` is a private function and will be deprecated. Please use `torch.distributed.all_gather_into_tensor` instead.
  return func(*args, **kwargs)
(min, max) time across ranks (ms):
    load-checkpoint ................................: (0.73, 0.78)
[after model, optimizer, and learning rate scheduler are built] datetime: 2024-09-19 15:56:43 
> building train, validation, and test datasets ...
 > datasets target sizes (minimum size):
    train:      500000
    validation: 5010
    test:       10
> building train, validation, and test datasets for GPT ...
> finished creating GPT datasets ...
[after dataloaders are built] datetime: 2024-09-19 15:56:43 
done with setup ...
training ...
(min, max) time across ranks (ms):
    model-and-optimizer-setup ......................: (300.04, 309.25)
    train/valid/test-data-iterators-setup ..........: (288.56, 334.41)
[before the start of training step] datetime: 2024-09-19 15:56:43 
WARNING:megatron.core.utils:NCCL Error 5: invalid usage (run with NCCL_DEBUG=WARN for details)
['Traceback (most recent call last):\n', '  File "/Megatron-LM/pretrain_gpt.py", line 192, in forward_step\n    output_tensor = model(tokens, position_ids, attention_mask,\n', '  File "/pytorch/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl\n    return self._call_impl(*args, **kwargs)\n', '  File "/pytorch/torch/nn/modules/module.py", line 1747, in _call_impl\n    return forward_call(*args, **kwargs)\n', '  File "/Megatron-LM/megatron/core/distributed/distributed_data_parallel.py", line 305, in forward\n    return self.module(*inputs, **kwargs)\n', '  File "/pytorch/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl\n    return self._call_impl(*args, **kwargs)\n', '  File "/pytorch/torch/nn/modules/module.py", line 1747, in _call_impl\n    return forward_call(*args, **kwargs)\n', '  File "/Megatron-LM/megatron/legacy/model/module.py", line 189, in forward\n    outputs = self.module(*inputs, **kwargs)\n', '  File "/pytorch/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl\n    return self._call_impl(*args, **kwargs)\n', '  File "/pytorch/torch/nn/modules/module.py", line 1747, in _call_impl\n    return forward_call(*args, **kwargs)\n', '  File "/Megatron-LM/megatron/core/models/gpt/gpt_model.py", line 217, in forward\n    hidden_states = self.decoder(\n', '  File "/pytorch/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl\n    return self._call_impl(*args, **kwargs)\n', '  File "/pytorch/torch/nn/modules/module.py", line 1747, in _call_impl\n    return forward_call(*args, **kwargs)\n', '  File "/Megatron-LM/megatron/core/transformer/transformer_block.py", line 496, in forward\n    hidden_states, context = layer(\n', '  File "/Megatron-LM/megatron/core/transformer/transformer_layer.py", line 377, in __call__\n    return super(MegatronModule, self).__call__(*args, **kwargs)\n', '  File "/pytorch/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl\n    return self._call_impl(*args, **kwargs)\n', '  File "/pytorch/torch/nn/modules/module.py", line 1747, in _call_impl\n    return forward_call(*args, **kwargs)\n', '  File "/Megatron-LM/megatron/core/transformer/transformer_layer.py", line 281, in forward\n    attention_output_with_bias = self.self_attention(\n', '  File "/pytorch/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl\n    return self._call_impl(*args, **kwargs)\n', '  File "/pytorch/torch/nn/modules/module.py", line 1747, in _call_impl\n    return forward_call(*args, **kwargs)\n', '  File "/Megatron-LM/megatron/core/transformer/attention.py", line 291, in forward\n    core_attn_out = self._checkpointed_attention_forward(\n', '  File "/Megatron-LM/megatron/core/transformer/attention.py", line 143, in _checkpointed_attention_forward\n    hidden_states = tensor_parallel.checkpoint(\n', '  File "/Megatron-LM/megatron/core/tensor_parallel/random.py", line 308, in checkpoint\n    return CheckpointFunction.apply(function, distribute_saved_activations, *args)\n', '  File "/pytorch/torch/autograd/function.py", line 575, in apply\n    return super().apply(*args, **kwargs)  # type: ignore[misc]\n', '  File "/Megatron-LM/megatron/core/tensor_parallel/random.py", line 247, in forward\n    outputs = run_function(*args)\n', '  File "/Megatron-LM/megatron/core/transformer/attention.py", line 130, in custom_forward\n    output_ = self.core_attention(\n', '  File "/pytorch/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl\n    return self._call_impl(*args, **kwargs)\n', '  File "/pytorch/torch/nn/modules/module.py", line 1747, in _call_impl\n    return forward_call(*args, **kwargs)\n', '  File "/Megatron-LM/megatron/core/extensions/transformer_engine.py", line 589, in forward\n    core_attn_out = super().forward(\n', '  File "/miniconda3/envs/megatron_lm/lib/python3.10/site-packages/transformer_engine/pytorch/attention.py", line 6854, in forward\n    return self.flash_attention(\n', '  File "/pytorch/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl\n    return self._call_impl(*args, **kwargs)\n', '  File "/pytorch/torch/nn/modules/module.py", line 1747, in _call_impl\n    return forward_call(*args, **kwargs)\n', '  File "/miniconda3/envs/megatron_lm/lib/python3.10/site-packages/transformer_engine/pytorch/attention.py", line 4328, in forward\n    output = attn_forward_func_with_cp(\n', '  File "/miniconda3/envs/megatron_lm/lib/python3.10/site-packages/transformer_engine/pytorch/attention.py", line 3379, in attn_forward_func_with_cp\n    out = AttnFuncWithCPAndKVP2P.apply(\n', '  File "/pytorch/torch/autograd/function.py", line 575, in apply\n    return super().apply(*args, **kwargs)  # type: ignore[misc]\n', '  File "/miniconda3/envs/megatron_lm/lib/python3.10/site-packages/transformer_engine/pytorch/attention.py", line 1473, in forward\n    send_recv_reqs[i % 2] = flash_attn_p2p_communicate(\n', '  File "/miniconda3/envs/megatron_lm/lib/python3.10/site-packages/transformer_engine/pytorch/attention.py", line 1249, in flash_attn_p2p_communicate\n    send_op = torch.distributed.isend(send_tensor, send_dst, cp_group)\n', '  File "/pytorch/torch/distributed/distributed_c10d.py", line 2071, in isend\n    return pg.send([tensor], dst, tag)\n', 'RuntimeError: NCCL Error 5: invalid usage (run with NCCL_DEBUG=WARN for details)\n']
[rank2]: Traceback (most recent call last):
[rank2]:   File "/Megatron-LM/pretrain_gpt.py", line 264, in <module>
[rank2]:     pretrain(
[rank2]:   File "/Megatron-LM/megatron/training/training.py", line 355, in pretrain
[rank2]:     iteration, num_floating_point_operations_so_far = train(
[rank2]:   File "/Megatron-LM/megatron/training/training.py", line 1234, in train
[rank2]:     train_step(forward_step_func,
[rank2]:   File "/Megatron-LM/megatron/training/training.py", line 718, in train_step
[rank2]:     losses_reduced = forward_backward_func(
[rank2]:   File "/Megatron-LM/megatron/core/pipeline_parallel/schedules.py", line 468, in forward_backward_no_pipelining
[rank2]:     output_tensor, num_tokens = forward_step(
[rank2]:   File "/Megatron-LM/megatron/core/pipeline_parallel/schedules.py", line 273, in forward_step
[rank2]:     output_tensor, loss_func = forward_step_func(data_iterator, model)
[rank2]:   File "/Megatron-LM/pretrain_gpt.py", line 192, in forward_step
[rank2]:     output_tensor = model(tokens, position_ids, attention_mask,
[rank2]:   File "/pytorch/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
[rank2]:     return self._call_impl(*args, **kwargs)
[rank2]:   File "/pytorch/torch/nn/modules/module.py", line 1747, in _call_impl
[rank2]:     return forward_call(*args, **kwargs)
[rank2]:   File "/Megatron-LM/megatron/core/distributed/distributed_data_parallel.py", line 305, in forward
[rank2]:     return self.module(*inputs, **kwargs)
[rank2]:   File "/pytorch/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
[rank2]:     return self._call_impl(*args, **kwargs)
[rank2]:   File "/pytorch/torch/nn/modules/module.py", line 1747, in _call_impl
[rank2]:     return forward_call(*args, **kwargs)
[rank2]:   File "/Megatron-LM/megatron/legacy/model/module.py", line 189, in forward
[rank2]:     outputs = self.module(*inputs, **kwargs)
[rank2]:   File "/pytorch/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
[rank2]:     return self._call_impl(*args, **kwargs)
[rank2]:   File "/pytorch/torch/nn/modules/module.py", line 1747, in _call_impl
[rank2]:     return forward_call(*args, **kwargs)
[rank2]:   File "/Megatron-LM/megatron/core/models/gpt/gpt_model.py", line 217, in forward
[rank2]:     hidden_states = self.decoder(
[rank2]:   File "/pytorch/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
[rank2]:     return self._call_impl(*args, **kwargs)
[rank2]:   File "/pytorch/torch/nn/modules/module.py", line 1747, in _call_impl
[rank2]:     return forward_call(*args, **kwargs)
[rank2]:   File "/Megatron-LM/megatron/core/transformer/transformer_block.py", line 496, in forward
[rank2]:     hidden_states, context = layer(
[rank2]:   File "/Megatron-LM/megatron/core/transformer/transformer_layer.py", line 377, in __call__
[rank2]:     return super(MegatronModule, self).__call__(*args, **kwargs)
[rank2]:   File "/pytorch/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
[rank2]:     return self._call_impl(*args, **kwargs)
[rank2]:   File "/pytorch/torch/nn/modules/module.py", line 1747, in _call_impl
[rank2]:     return forward_call(*args, **kwargs)
[rank2]:   File "/Megatron-LM/megatron/core/transformer/transformer_layer.py", line 281, in forward
[rank2]:     attention_output_with_bias = self.self_attention(
[rank2]:   File "/pytorch/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
[rank2]:     return self._call_impl(*args, **kwargs)
[rank2]:   File "/pytorch/torch/nn/modules/module.py", line 1747, in _call_impl
[rank2]:     return forward_call(*args, **kwargs)
[rank2]:   File "/Megatron-LM/megatron/core/transformer/attention.py", line 291, in forward
[rank2]:     core_attn_out = self._checkpointed_attention_forward(
[rank2]:   File "/Megatron-LM/megatron/core/transformer/attention.py", line 143, in _checkpointed_attention_forward
[rank2]:     hidden_states = tensor_parallel.checkpoint(
[rank2]:   File "/Megatron-LM/megatron/core/tensor_parallel/random.py", line 308, in checkpoint
[rank2]:     return CheckpointFunction.apply(function, distribute_saved_activations, *args)
[rank2]:   File "/pytorch/torch/autograd/function.py", line 575, in apply
[rank2]:     return super().apply(*args, **kwargs)  # type: ignore[misc]
[rank2]:   File "/Megatron-LM/megatron/core/tensor_parallel/random.py", line 247, in forward
[rank2]:     outputs = run_function(*args)
[rank2]:   File "/Megatron-LM/megatron/core/transformer/attention.py", line 130, in custom_forward
[rank2]:     output_ = self.core_attention(
[rank2]:   File "/pytorch/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
[rank2]:     return self._call_impl(*args, **kwargs)
[rank2]:   File "/pytorch/torch/nn/modules/module.py", line 1747, in _call_impl
[rank2]:     return forward_call(*args, **kwargs)
[rank2]:   File "/Megatron-LM/megatron/core/extensions/transformer_engine.py", line 589, in forward
[rank2]:     core_attn_out = super().forward(
[rank2]:   File "/miniconda3/envs/megatron_lm/lib/python3.10/site-packages/transformer_engine/pytorch/attention.py", line 6854, in forward
[rank2]:     return self.flash_attention(
[rank2]:   File "/pytorch/torch/nn/modules/module.py", line 1736, in _wrapped_call_impl
[rank2]:     return self._call_impl(*args, **kwargs)
[rank2]:   File "/pytorch/torch/nn/modules/module.py", line 1747, in _call_impl
[rank2]:     return forward_call(*args, **kwargs)
[rank2]:   File "/miniconda3/envs/megatron_lm/lib/python3.10/site-packages/transformer_engine/pytorch/attention.py", line 4328, in forward
[rank2]:     output = attn_forward_func_with_cp(
[rank2]:   File "/miniconda3/envs/megatron_lm/lib/python3.10/site-packages/transformer_engine/pytorch/attention.py", line 3379, in attn_forward_func_with_cp
[rank2]:     out = AttnFuncWithCPAndKVP2P.apply(
[rank2]:   File "/pytorch/torch/autograd/function.py", line 575, in apply
[rank2]:     return super().apply(*args, **kwargs)  # type: ignore[misc]
[rank2]:   File "/miniconda3/envs/megatron_lm/lib/python3.10/site-packages/transformer_engine/pytorch/attention.py", line 1473, in forward
[rank2]:     send_recv_reqs[i % 2] = flash_attn_p2p_communicate(
[rank2]:   File "/miniconda3/envs/megatron_lm/lib/python3.10/site-packages/transformer_engine/pytorch/attention.py", line 1249, in flash_attn_p2p_communicate
[rank2]:     send_op = torch.distributed.isend(send_tensor, send_dst, cp_group)
[rank2]:   File "/pytorch/torch/distributed/distributed_c10d.py", line 2071, in isend
[rank2]:     return pg.send([tensor], dst, tag)
[rank2]: RuntimeError: NCCL Error 5: invalid usage (run with NCCL_DEBUG=WARN for details)

To Reproduce Steps to reproduce the behavior. The easier it is to reproduce the faster it will get maintainer attention.

Expected behavior A clear and concise description of what you expected to happen.

Stack trace/logs If applicable, add the stack trace or logs from the time of the error.

Environment (please complete the following information):

Proposed fix If you have a proposal for how to fix the issue state it here or link to a PR.

Additional context Add any other context about the problem here.

github-actions[bot] commented 4 days ago

Marking as stale. No activity in 60 days.