Merge branch 'megatron_v0.11.0' of...

Merge branch 'megatron_v0.11.0' of http://developer.sourcefind.cn/codes/OpenDAS/dcu_megatron into megatron_v0.11.0

dcu_megatron/legacy/model/rms_norm.py

+# Copyright (c) 2023, NVIDIA CORPORATION. All rights reserved.
+
+import torch
+from torch import nn
+
+class RMSNorm(torch.nn.Module):
+
+    def __init__(self,
+                 dim: int,
+                 eps: float = 1e-6,
+                 sequence_parallel: bool = False,
+                 config: dict = None):
+        """RMS Normaliation module
+
+        Args:
+            dim (int): The width of input, i.e. hidden size
+            eps (float): epsilon to use for the norm, default to 1e-6
+            sequence_parallel (bool): Set to true if sequence parallelism is being used,
+              this marks the weights as needing to be allreduced.
+        """
+        super().__init__()
+        self.eps = eps
+        self.weight = nn.Parameter(torch.ones(dim))
+
+        setattr(self.weight, 'sequence_parallel', sequence_parallel)
+
+    def _norm(self, x):
+        return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
+
+    @torch.compile(mode="max-autotune-no-cudagraphs")
+    def forward(self, x):
+        output = self._norm(x.float()).type_as(x)
+        return output * self.weight
+
 import torch
 
 from typing import Any, Callable, Dict, Optional, Tuple, Union

dcu_megatron/legacy/model/utils.py

 from megatron.training import get_args
 from megatron.legacy.model import LayerNorm
-from .rms_norm import LightopRMSNorm
+from .rms_norm import RMSNorm, LightopRMSNorm
 
 
 def get_norm(config):
@@ -15,8 +15,12 @@ def get_norm(config):
     elif args.normalization == "RMSNorm":
         if args.apply_layernorm_1p:
             raise NotImplementedError('RMSNorm does not currently support the layernorm_1p formulation.')
-
+    
+        return RMSNorm(dim=config.hidden_size,
+                       eps=config.layernorm_epsilon,
+                       sequence_parallel=config.sequence_parallel)
+    elif args.normalization == "LightopRMSNorm":
         return LightopRMSNorm(dim=config.hidden_size,
-                              eps=config.layernorm_epsilon)
+                       eps=config.layernorm_epsilon)
     else:
         raise Exception(f"unsupported norm type '{args.normalization}'.")

dcu_megatron/training/arguments.py

@@ -51,6 +51,7 @@ def parse_args(extra_args_provider=None, ignore_unknown_args=False):
 
     # Standard arguments.
     parser = _add_network_size_args(parser)
+    parser = _add_extra_network_size_args(parser)
     parser = _add_regularization_args(parser)
     parser = _add_training_args(parser)
     parser = _add_extra_training_args(parser)
@@ -106,6 +107,18 @@ def parse_args(extra_args_provider=None, ignore_unknown_args=False):
     return args
 
 
+def _add_extra_network_size_args(parser):
+    # 删除原参数
+    remove_original_params(parser, ["normalization"])
+
+    # 重定义参数
+    group = parser.add_argument_group(title='extra network size args')
+    group.add_argument('--normalization', default='LayerNorm',
+                       choices=['LayerNorm', 'RMSNorm', 'LightopRMSNorm'],
+                       help='Which normalization technique to use.')
+    return parser
+
+
 def _add_extra_distributed_args(parser):
     group = parser.add_argument_group(title='extra distributed args')
     group.add_argument('--rank', default=-1, type=int,

examples/llama/Llama2_70b.sh

+#!/bin/bash
+# set -eux
+
+#export FLASH_ATTENTION_PRINT_PARAM=1
+# Runs the "7B" parameter model
+export HSA_FORCE_FINE_GRAIN_PCIE=1
+export OMP_NUM_THREADS=1
+export NCCL_P2P_LEVEL=PXB # SYS
+
+#export HIP_ALLOC_INITIALIZE=0
+# export GPU_MAX_HW_QUEUES=10
+
+export NCCL_ALGO=Ring
+export NCCL_NCHANNELS_PER_PEER=16
+export NCCL_MIN_NCHANNELS=32 # 20
+export NCCL_MAX_NCHANNELS=32 # 20
+export NCCL_IB_TIMEOUT=22
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+
+export NCCL_IB_HCA=mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_6:1,mlx5_7:1,mlx5_8:1,mlx5_9:1
+export NCCL_NET_GDR_LEVEL=7
+export NCCL_NET_GDR_READ=1
+export RCCL_SDMA_COPY_ENABLE=0
+export NCCL_TOPO_FILE="/public/home/wangxj/Projects/rccl-test/rccl-tests-0204/topo-input.xml"
+# export NCCL_TOPO_FILE="/workspace/rccl-test/rccl-tests-0204/topo-input.xml"
+export GLOG_minloglevel=3 # 打印error级别的nccl日志
+source /opt/dtk/env.sh
+# 导入hipblaslt库
+# export LD_LIBRARY_PATH=/data/hipblaslt-install-0904/lib:$LD_LIBRARY_PATH 
+# export LD_LIBRARY_PATH=/data/hipblaslt-install-dtk-25.04-0212/lib:$LD_LIBRARY_PATH 
+export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/blas/hipblaslt-install-dtk-25.04-0212/lib:$LD_LIBRARY_PATH 
+# 更新rocblas
+# export LD_LIBRARY_PATH=/data/rocblas-install_qwen1211/lib:$LD_LIBRARY_PATH
+# export LD_LIBRARY_PATH=/data/rocblas-install_qwen1228/lib:$LD_LIBRARY_PATH
+# export LD_LIBRARY_PATH=/data/rocblas-install-0118-bf16/lib:$LD_LIBRARY_PATH
+# export LD_LIBRARY_PATH=/data/rocblas-install-0203-release/lib:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/blas/rocblas-install-0331-release/lib:$LD_LIBRARY_PATH
+
+# torch控制多流转单流
+export ALLREDUCE_STREAM_WITH_COMPUTE=1
+export SENDRECV_STREAM_WITH_COMPUTE=1 
+
+# prof采集添加同步, 避免卡顿
+# export GPU_FLUSH_ON_EXECUTION=1
+# export HIP_DIRECT_DISPATCH=0
+
+# 采集rocblas size
+# export ROCBLAS_LAYER=3
+# export HIPBLASLT_LOG_LEVEL=3
+# 采集 fa size
+# export FLASH_ATTENTION_PRINT_PARAM=1
+
+#增加编译缓存
+export cache_size_limit=64
+
+# lightop算子库
+export PYTORCH_ROCM_ARCH='gfx906;gfx926;gfx936'
+
+# CHECKPOINT_PATH=./Llama-2-7b-hf-to-meg-tp1-pp2 #CHECKPOINT_PATH=./tmp_7b # 
+SAVE_PATH=./tmp_7b
+TENSORBOARD_LOGS_PATH=./tmp_7b  #$2 #<Specify path>
+DATA_PATH="/public/home/gmhtest_tmp/RedPajama-Data-1T-Sample/redpajama_text_document" #<Specify path and file prefix>_text_document
+# DATA_PATH="/data/datasets/oscar-1GB-head/oscar-1GB_head-llama3.2_text_document" #<Specify path and file prefix>_text_document
+
+
+GPT_MODEL_ARGS=(
+    --num-layers 80 #80 #80 #40 # 20 # 
+    --hidden-size 8192
+    --ffn-hidden-size 22016 # 28672
+    --num-attention-heads 64
+    --max-position-embeddings 8192
+    --group-query-attention
+    --num-query-groups 8
+
+    --normalization RMSNorm 
+    --position-embedding-type rope
+    --untie-embeddings-and-output-weights # 分开处理embed和输出权重, 增加灵活性
+)
+
+export NVTE_FLASH_ATTN=1 # 走cutlass
+# export NVTE_FLASH_ATTN_TRITON=1 # 走triton_fa
+# --transformer-impl transformer_engine # 走core用这两组参数
+    # --use-mcore-models
+    # --transformer-impl local # 走legacy用这两组参数
+    # --use-legacy-models 
+TRAINING_ARGS=(
+    --transformer-impl local # 走legacy用这两组参数
+    --use-legacy-models 
+    --micro-batch-size 1
+    --global-batch-size 512 #32 #512 #256 # 64 #240 #60 #512 #64
+    --train-iters 300
+    --weight-decay 0.1 
+    --adam-beta1 0.9 
+    --adam-beta2 0.95 
+    --init-method-std 0.006 
+    --clip-grad 1.0 
+    --bf16
+    # --fp16 # 开启fp16需要指定loss-scale
+    # --loss-scale 1024
+    --use-distributed-optimizer 
+    --disable-bias-linear
+    --attention-dropout 0
+    --hidden-dropout 0
+    # --no-gradient-accumulation-fusion
+    # --no-check-for-nan-in-loss-and-grad
+    --swiglu
+    --lr 3.0e-5 
+    --lr-decay-style cosine 
+    --min-lr 3.0e-6
+    --lr-warmup-iters 1
+    --ckpt-format torch
+    --ddp-average-in-collective # 在dp阶段通信中, 梯度或参数将被直接平均, 而不是先求和(到一个设备)再平均
+    # --recompute-activations
+    # --recompute-granularity full # 开启重计算降低显存增加耗时
+    # --recompute-num-layers 1 #0 #
+    # --recompute-method block
+    --overlap-grad-reduce # 重叠ddp grad reduce
+    # --tp-comm-overlap # tensor parallel comm和gemm重叠, 启动core
+    # --tp-comm-overlap-rs-dgrad # reduce-scatter和dgrad gemm重叠, 启动core
+    --use-flash-attn
+)
+# export TORCHINDUCTOR_COORDINATE_DESCENT_TUNING=1
+# export TORCHINDUCTOR_BENCHMARK_FUSION=1
+# export TORCHINDUCTOR_BENCHMARK_MULTI_TEMPLATES=1
+# export TORCHINDUCTOR_MAX_AUTOTUNE=1
+# export TORCHINDUCTOR_CACHE_DIR=./cache
+# --use-flash-attn-cutlass # cutlass fa
+# --use-flash-attn-triton # triton fa
+# --use-flash-attn-torch # torch fa
+
+MODEL_PARALLEL_ARGS=(
+    --sequence-parallel
+	--tensor-model-parallel-size 4
+	--pipeline-model-parallel-size 8
+  --context-parallel-size 1
+    # --num-layers-per-virtual-pipeline-stage 1
+  # --microbatch-group-size-per-virtual-pipeline-stage 5
+  # --no-overlap-p2p-communication # 开启后
+)
+
+DATA_ARGS=(
+    --data-path $DATA_PATH 
+    --seq-length 4096 #8192 #4096
+    --split 949,50,1
+    --tokenizer-type Llama2Tokenizer
+    --tokenizer-model /public/home/gmhtest_tmp/RedPajama-Data-1T-Sample/tokenizer.model
+    # --tokenizer-model /data/model_weights/llama3.2/tokenizer.model
+)
+
+EVAL_AND_LOGGING_ARGS=(
+    --log-interval 1
+    --log-throughput
+    --save-interval 500 
+    --eval-interval 50 
+    --eval-iters 3
+    --save $SAVE_PATH 
+    --load $SAVE_PATH 
+    --tensorboard-dir $TENSORBOARD_LOGS_PATH 
+)
+
+# FINETUNE_ARGS=(
+#     # --finetune
+#     # --pretrained-checkpoint $CHECKPOINT_PATH
+#     --load $CHECKPOINT_PATH
+#     --no-load-optim
+#     --no-load-rng
+# )
+
+PROFILE_ARGS=(
+    --profile
+    --profile-step-start 4
+    --profile-step-end 5
+    --use-pytorch-profiler
+    --profile-ranks 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
+    --profile-dir prof_data
+)
+
+RANK=$OMPI_COMM_WORLD_RANK
+LOCAL_RANK=$OMPI_COMM_WORLD_LOCAL_RANK
+WORLD_SIZE=$OMPI_COMM_WORLD_SIZE
+DIST_URL=${1}
+DIST_PORT=34577
+
+DISTRIBUTED_ARGS=(
+    --rank ${RANK}
+    --world-size ${WORLD_SIZE}
+    --local-rank ${LOCAL_RANK}
+    --dist-url tcp://${DIST_URL}:${DIST_PORT}
+)
+
+APP="python -u ../../pretrain_gpt.py \
+        ${GPT_MODEL_ARGS[@]} \
+        ${TRAINING_ARGS[@]} \
+        ${MODEL_PARALLEL_ARGS[@]} \
+        ${DATA_ARGS[@]} \
+        ${EVAL_AND_LOGGING_ARGS[@]} \
+        ${DISTRIBUTED_ARGS[@]} \
+        
+"
+# 开启profile
+# ${PROFILE_ARGS[@]} \
+
+# export HIP_VISIBLE_DEVICES=0,7 #  # 4,5,6,7 #,
+export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 #  # 4,5,6,7 #,
+# export CUDA_VISIBLE_DEVICES=4,5,6,7 # 0,1,2,3,
+# ${APP}
+case ${LOCAL_RANK} in
+[0])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[1])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  numactl --cpunodebind=1 --membind=1 ${APP}
+  ;;
+[2])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  numactl --cpunodebind=2 --membind=2 ${APP}
+  ;;
+[3])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  numactl --cpunodebind=3 --membind=3 ${APP}
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[4])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  numactl --cpunodebind=4 --membind=4 ${APP}
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[5])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  numactl --cpunodebind=5 --membind=5 ${APP}
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[6])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  numactl --cpunodebind=6 --membind=6 ${APP}
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[7])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  numactl --cpunodebind=7 --membind=7 ${APP}
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+esac
\ No newline at end of file

examples/llama/Llama2_7b.sh

+#!/bin/bash
+# set -eux
+
+for para in $*
+do
+    if [[ $para == --profiling* ]];then
+        profiling=${para#*=}
+    fi
+done
+
+CURRENT_DIR="$( cd "$( dirname "$0" )" && pwd )"
+MEGATRON_PATH=$( dirname $( dirname ${CURRENT_DIR}))
+
+
+#default env
+#export FLASH_ATTENTION_PRINT_PARAM=1
+export HSA_FORCE_FINE_GRAIN_PCIE=1
+export OMP_NUM_THREADS=1
+export NCCL_P2P_LEVEL=PXB # SYS
+# export GPU_MAX_HW_QUEUES=10
+#export HIP_ALLOC_INITIALIZE=0
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+
+# nccl env
+export NCCL_ALGO=Ring
+export NCCL_NCHANNELS_PER_PEER=16
+export NCCL_MIN_NCHANNELS=32 # 20
+export NCCL_MAX_NCHANNELS=32 # 20
+export NCCL_IB_TIMEOUT=22
+export NCCL_NET_GDR_LEVEL=7
+export NCCL_NET_GDR_READ=1
+export RCCL_SDMA_COPY_ENABLE=0
+export NCCL_IB_HCA=mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_6:1,mlx5_7:1,mlx5_8:1,mlx5_9:1
+export NCCL_TOPO_FILE="/workspace/rccl-test/rccl-tests-0204/topo-input.xml"
+export GLOG_minloglevel=3 # 打印error级别的nccl日志
+source /opt/dtk/env.sh
+
+# hipblaslt库
+export LD_LIBRARY_PATH=/data/blas/hipblaslt-install-dtk-25.04-0212/lib:$LD_LIBRARY_PATH 
+
+# rocblas
+export LD_LIBRARY_PATH=/data/blas/rocblas-install-0331-release/lib:$LD_LIBRARY_PATH
+
+# torch控制多流转单流
+export ALLREDUCE_STREAM_WITH_COMPUTE=1
+export SENDRECV_STREAM_WITH_COMPUTE=1 
+
+#增加编译缓存
+export cache_size_limit=64
+
+# CHECKPOINT_PATH=./Llama-2-7b-hf-to-meg-tp1-pp2 #CHECKPOINT_PATH=./tmp_7b # 
+SAVE_PATH=./tmp_7b
+TENSORBOARD_LOGS_PATH=./tmp_7b  #$2 #<Specify path>
+DATA_PATH="/data/datasets/oscar-1GB/oscar-1GB-llama2_text_document" #<Specify path and file prefix>_text_document
+
+GPT_MODEL_ARGS=(
+    --num-layers 32
+    --hidden-size 4096
+    --ffn-hidden-size 11008 
+    --num-attention-heads 32
+    --max-position-embeddings 4096
+
+    --normalization RMSNorm # LightopRMSNorm
+    --position-embedding-type rope # none # 
+    --untie-embeddings-and-output-weights # 分开处理embed和输出权重, 增加灵活性
+)
+
+export NVTE_FLASH_ATTN=1 # 走cutlass
+# export NVTE_FLASH_ATTN_TRITON=1 # 走triton_fa
+# --transformer-impl transformer_engine # 走core用这两组参数
+    # --use-mcore-models
+    # --transformer-impl local # 走legacy用这两组参数
+    # --use-legacy-models 
+TRAINING_ARGS=(
+    --transformer-impl local # 走legacy用这两组参数
+    --use-legacy-models 
+    --micro-batch-size 1
+    --global-batch-size 256 #256 #240 #60 #512 #64
+    --train-iters 50
+    --weight-decay 0.1 
+    --adam-beta1 0.9 
+    --adam-beta2 0.95 
+    --init-method-std 0.006 
+    --clip-grad 1.0 
+    --bf16
+    # --fp16 # 开启fp16需要指定loss-scale
+    # --loss-scale 1024
+    --use-distributed-optimizer 
+    --disable-bias-linear
+    --attention-dropout 0
+    --hidden-dropout 0
+    # --no-gradient-accumulation-fusion
+    --swiglu
+    --lr 3.0e-5 
+    --lr-decay-style cosine 
+    --min-lr 3.0e-6
+    --lr-warmup-iters 1
+    --ckpt-format torch
+    --ddp-average-in-collective # 在dp阶段通信中, 梯度或参数将被直接平均, 而不是先求和(到一个设备)再平均
+    # --recompute-granularity full # 开启重计算降低显存增加耗时
+    # --recompute-num-layers 5 #0 #
+    # --recompute-method block
+    --overlap-grad-reduce # 重叠ddp grad reduce
+    # --tp-comm-overlap # tensor parallel comm和gemm重叠, 优化项未适配
+    # --tp-comm-overlap-rs-dgrad # reduce-scatter和dgrad gemm重叠
+    --use-flash-attn
+)
+# 使用torch fa的环境变量
+# export TORCHINDUCTOR_COORDINATE_DESCENT_TUNING=1
+# export TORCHINDUCTOR_BENCHMARK_FUSION=1
+# export TORCHINDUCTOR_BENCHMARK_MULTI_TEMPLATES=1
+# export TORCHINDUCTOR_MAX_AUTOTUNE=1
+# export TORCHINDUCTOR_CACHE_DIR=./cache
+# --use-flash-attn-cutlass # cutlass fa
+# --use-flash-attn-triton # triton fa
+# --use-flash-attn-torch # torch fa
+
+MODEL_PARALLEL_ARGS=(
+    --sequence-parallel
+	--tensor-model-parallel-size 1
+	--pipeline-model-parallel-size 2
+  # --context-parallel-size 2
+  # --num-layers-per-virtual-pipeline-stage 4
+  # --microbatch-group-size-per-virtual-pipeline-stage 1
+  # --no-overlap-p2p-communication # 开启后
+)
+
+DATA_ARGS=(
+    --data-path $DATA_PATH 
+    --seq-length 4096 #4096
+    --split 949,50,1
+    --tokenizer-type Llama2Tokenizer
+    --tokenizer-model /data/model_weights/llama2_7b_hf/tokenizer.model
+)
+
+EVAL_AND_LOGGING_ARGS=(
+    --log-throughput
+    --eval-iters 50
+    --log-interval 1
+    --save-interval 1000 
+    --eval-interval 1000 
+    --save $SAVE_PATH 
+    --load $SAVE_PATH 
+    --tensorboard-dir $TENSORBOARD_LOGS_PATH 
+)
+
+# FINETUNE_ARGS=(
+#     # --finetune
+#     # --pretrained-checkpoint $CHECKPOINT_PATH
+#     --load $CHECKPOINT_PATH
+#     --no-load-optim
+#     --no-load-rng
+# )
+
+PROFILE_ARGS=(
+    --profile
+    --profile-step-start 4
+    --profile-step-end 5
+    --use-pytorch-profiler
+    --profile-ranks 0 1 2 3 4 5 6 7
+    --profile-dir prof_data
+)
+
+RANK=$OMPI_COMM_WORLD_RANK
+LOCAL_RANK=$OMPI_COMM_WORLD_LOCAL_RANK
+WORLD_SIZE=$OMPI_COMM_WORLD_SIZE
+DIST_URL=${1}
+DIST_PORT=34577
+
+DISTRIBUTED_ARGS=(
+    --rank ${RANK}
+    --world-size ${WORLD_SIZE}
+    --local-rank ${LOCAL_RANK}
+    --dist-url tcp://${DIST_URL}:${DIST_PORT}
+)
+
+APP="python -u ${MEGATRON_PATH}/pretrain_gpt.py \
+        ${GPT_MODEL_ARGS[@]} \
+        ${TRAINING_ARGS[@]} \
+        ${MODEL_PARALLEL_ARGS[@]} \
+        ${DATA_ARGS[@]} \
+        ${EVAL_AND_LOGGING_ARGS[@]} \
+        ${DISTRIBUTED_ARGS[@]} \
+        
+"
+# 开启profile
+# ${PROFILE_ARGS[@]} \
+
+# export HIP_VISIBLE_DEVICES=0,7 #  # 4,5,6,7 #,
+export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 #  # 4,5,6,7 #,
+# export CUDA_VISIBLE_DEVICES=4,5,6,7 # 0,1,2,3,
+# ${APP}
+case ${LOCAL_RANK} in
+[0])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[1])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  numactl --cpunodebind=1 --membind=1 ${APP}
+  ;;
+[2])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  numactl --cpunodebind=2 --membind=2 ${APP}
+  ;;
+[3])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  numactl --cpunodebind=3 --membind=3 ${APP}
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[4])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  numactl --cpunodebind=4 --membind=4 ${APP}
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[5])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  numactl --cpunodebind=5 --membind=5 ${APP}
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[6])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  numactl --cpunodebind=6 --membind=6 ${APP}
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[7])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  numactl --cpunodebind=7 --membind=7 ${APP}
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+esac
\ No newline at end of file

examples/llama/Llama3_405b.sh

+#!/bin/bash
+# set -eux
+
+#export FLASH_ATTENTION_PRINT_PARAM=1
+# Runs the "7B" parameter model
+export HSA_FORCE_FINE_GRAIN_PCIE=1
+export OMP_NUM_THREADS=1
+export NCCL_P2P_LEVEL=PXB # SYS
+
+#export HIP_ALLOC_INITIALIZE=0
+# export GPU_MAX_HW_QUEUES=10
+
+export NCCL_ALGO=Ring
+export NCCL_NCHANNELS_PER_PEER=16
+export NCCL_MIN_NCHANNELS=32 # 20
+export NCCL_MAX_NCHANNELS=32 # 20
+export NCCL_IB_TIMEOUT=22
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+
+export NCCL_IB_HCA=mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_6:1,mlx5_7:1,mlx5_8:1,mlx5_9:1
+export NCCL_NET_GDR_LEVEL=7
+export NCCL_NET_GDR_READ=1
+export RCCL_SDMA_COPY_ENABLE=0
+export NCCL_TOPO_FILE="/public/home/wangxj/Projects/rccl-test/rccl-tests-0204/topo-input.xml"
+# export NCCL_TOPO_FILE="/workspace/rccl-test/rccl-tests-0204/topo-input.xml"
+export GLOG_minloglevel=3 # 打印error级别的nccl日志
+source /opt/dtk/env.sh
+# 导入hipblaslt库
+# export LD_LIBRARY_PATH=/data/hipblaslt-install-0904/lib:$LD_LIBRARY_PATH 
+# export LD_LIBRARY_PATH=/data/hipblaslt-install-dtk-25.04-0212/lib:$LD_LIBRARY_PATH 
+export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/blas/hipblaslt-install-dtk-25.04-0212/lib:$LD_LIBRARY_PATH 
+# 更新rocblas
+# export LD_LIBRARY_PATH=/data/rocblas-install_qwen1211/lib:$LD_LIBRARY_PATH
+# export LD_LIBRARY_PATH=/data/rocblas-install_qwen1228/lib:$LD_LIBRARY_PATH
+# export LD_LIBRARY_PATH=/data/rocblas-install-0118-bf16/lib:$LD_LIBRARY_PATH
+# export LD_LIBRARY_PATH=/data/rocblas-install-0203-release/lib:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/blas/rocblas-install-0203-release/lib:$LD_LIBRARY_PATH
+
+# torch控制多流转单流
+export ALLREDUCE_STREAM_WITH_COMPUTE=1
+export SENDRECV_STREAM_WITH_COMPUTE=1 
+
+# prof采集添加同步, 避免卡顿
+# export GPU_FLUSH_ON_EXECUTION=1
+# export HIP_DIRECT_DISPATCH=0
+
+# 采集rocblas size
+# export ROCBLAS_LAYER=3
+# 采集 fa size
+# export FLASH_ATTENTION_PRINT_PARAM=1
+
+#增加编译缓存
+export cache_size_limit=64
+
+# CHECKPOINT_PATH=./Llama-2-7b-hf-to-meg-tp1-pp2 #CHECKPOINT_PATH=./tmp_7b # 
+SAVE_PATH=./tmp_7b
+TENSORBOARD_LOGS_PATH=./tmp_7b  #$2 #<Specify path>
+DATA_PATH="/public/home/wangxj/Downloads/datasets/oscar-1GB-head/oscar-1GB_head-llama3.2_text_document" #<Specify path and file prefix>_text_document
+# DATA_PATH="/data/datasets/oscar-1GB-head/oscar-1GB_head-llama3.2_text_document" #<Specify path and file prefix>_text_document
+
+GPT_MODEL_ARGS=(
+    --num-layers 126 #96 #8 # 126
+    --hidden-size 16384
+    --ffn-hidden-size 53248 
+    --num-attention-heads 128
+    --max-position-embeddings 16384
+    --group-query-attention
+    --num-query-groups 16
+
+    --normalization RMSNorm 
+    --position-embedding-type rope
+    --untie-embeddings-and-output-weights # 分开处理embed和输出权重, 增加灵活性
+)
+
+export NVTE_FLASH_ATTN=1 # 走cutlass
+# export NVTE_FLASH_ATTN_TRITON=1 # 走triton_fa
+# --transformer-impl transformer_engine # 走core用这两组参数
+    # --use-mcore-models
+    # --transformer-impl local # 走legacy用这两组参数
+    # --use-legacy-models 
+TRAINING_ARGS=(
+    --transformer-impl transformer_engine # 走core用这两组参数
+    --use-mcore-models
+    --micro-batch-size 1
+    --global-batch-size 6912 # 252 #32 # 64 #240 #60 #512 #64
+    --train-iters 100
+    --weight-decay 0.1 
+    --adam-beta1 0.9 
+    --adam-beta2 0.95 
+    --init-method-std 0.006 
+    --clip-grad 1.0 
+    --bf16
+    # --fp16 # 开启fp16需要指定loss-scale
+    # --loss-scale 1024
+    --use-distributed-optimizer 
+    --disable-bias-linear
+    --attention-dropout 0
+    --hidden-dropout 0
+    # --no-gradient-accumulation-fusion
+    --swiglu
+    --lr 3.0e-5 
+    --lr-decay-style cosine 
+    --min-lr 3.0e-6
+    --lr-warmup-iters 1
+    --ckpt-format torch
+    --ddp-average-in-collective # 在dp阶段通信中, 梯度或参数将被直接平均, 而不是先求和(到一个设备)再平均
+    # --recompute-granularity full # 开启重计算降低显存增加耗时
+    # --recompute-num-layers 5 #0 #
+    # --recompute-method block
+    --overlap-grad-reduce # 重叠ddp grad reduce
+    # --tp-comm-overlap # tensor parallel comm和gemm重叠, 优化项未适配
+    # --tp-comm-overlap-rs-dgrad # reduce-scatter和dgrad gemm重叠, 优化项未适配
+    --use-flash-attn-cutlass
+)
+# export TORCHINDUCTOR_COORDINATE_DESCENT_TUNING=1
+# export TORCHINDUCTOR_BENCHMARK_FUSION=1
+# export TORCHINDUCTOR_BENCHMARK_MULTI_TEMPLATES=1
+# export TORCHINDUCTOR_MAX_AUTOTUNE=1
+# export TORCHINDUCTOR_CACHE_DIR=./cache
+# --use-flash-attn-cutlass # cutlass fa
+# --use-flash-attn-triton # triton fa
+# --use-flash-attn-torch # torch fa
+
+MODEL_PARALLEL_ARGS=(
+    --sequence-parallel
+	--tensor-model-parallel-size 8
+	--pipeline-model-parallel-size 18 # 7 layer/gpu
+    --context-parallel-size 2
+)
+
+DATA_ARGS=(
+    --data-path $DATA_PATH 
+    --seq-length 4096 #4096
+    --split 949,50,1
+    --tokenizer-type Llama3Tokenizer
+    --tokenizer-model /public/home/wangxj/Downloads/model_weights/llama3.2/tokenizer.model
+    # --tokenizer-model /data/model_weights/llama3.2/tokenizer.model
+)
+
+EVAL_AND_LOGGING_ARGS=(
+    --log-interval 1
+    --log-throughput
+    --save-interval 1000 
+    --eval-interval 1000 
+    --save $SAVE_PATH 
+    --load $SAVE_PATH 
+    --eval-iters 10
+    --tensorboard-dir $TENSORBOARD_LOGS_PATH 
+)
+
+# FINETUNE_ARGS=(
+#     # --finetune
+#     # --pretrained-checkpoint $CHECKPOINT_PATH
+#     --load $CHECKPOINT_PATH
+#     --no-load-optim
+#     --no-load-rng
+# )
+
+PROFILE_ARGS=(
+    --profile
+    --profile-step-start 4
+    --profile-step-end 5
+    --use-pytorch-profiler
+    --profile-ranks 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32
+    --profile-dir prof_data
+)
+
+RANK=$OMPI_COMM_WORLD_RANK
+LOCAL_RANK=$OMPI_COMM_WORLD_LOCAL_RANK
+WORLD_SIZE=$OMPI_COMM_WORLD_SIZE
+DIST_URL=${1}
+DIST_PORT=34577
+
+DISTRIBUTED_ARGS=(
+    --rank ${RANK}
+    --world-size ${WORLD_SIZE}
+    --local-rank ${LOCAL_RANK}
+    --dist-url tcp://${DIST_URL}:${DIST_PORT}
+)
+
+APP="python -u pretrain_gpt.py \
+        ${GPT_MODEL_ARGS[@]} \
+        ${TRAINING_ARGS[@]} \
+        ${MODEL_PARALLEL_ARGS[@]} \
+        ${DATA_ARGS[@]} \
+        ${EVAL_AND_LOGGING_ARGS[@]} \
+        ${DISTRIBUTED_ARGS[@]} \
+        
+"
+# 开启profile
+# ${PROFILE_ARGS[@]} \
+
+# export HIP_VISIBLE_DEVICES=0,7 #  # 4,5,6,7 #,
+export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 #  # 4,5,6,7 #,
+# export CUDA_VISIBLE_DEVICES=4,5,6,7 # 0,1,2,3,
+${APP}
+# case ${LOCAL_RANK} in
+# [0])
+#   export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+#   # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+#   numactl --cpunodebind=0 --membind=0 ${APP}
+#   ;;
+# [1])
+#   export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+#   # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+#   numactl --cpunodebind=1 --membind=1 ${APP}
+#   ;;
+# [2])
+#   export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+#   # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+#   numactl --cpunodebind=2 --membind=2 ${APP}
+#   ;;
+# [3])
+#   export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+#   numactl --cpunodebind=3 --membind=3 ${APP}
+#   # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+#   ;;
+# [4])
+#   export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+#   numactl --cpunodebind=4 --membind=4 ${APP}
+#   # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+#   ;;
+# [5])
+#   export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+#   numactl --cpunodebind=5 --membind=5 ${APP}
+#   # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+#   ;;
+# [6])
+#   export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+#   numactl --cpunodebind=6 --membind=6 ${APP}
+#   # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+#   ;;
+# [7])
+#   export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+#   numactl --cpunodebind=7 --membind=7 ${APP}
+#   # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+#   ;;
+# esac
\ No newline at end of file

examples/llama/Llama3_70b.sh

+#!/bin/bash
+# set -eux
+
+#export FLASH_ATTENTION_PRINT_PARAM=1
+# Runs the "7B" parameter model
+export HSA_FORCE_FINE_GRAIN_PCIE=1
+export OMP_NUM_THREADS=1
+export NCCL_P2P_LEVEL=PXB # SYS
+
+#export HIP_ALLOC_INITIALIZE=0
+# export GPU_MAX_HW_QUEUES=10
+
+export NCCL_ALGO=Ring
+export NCCL_NCHANNELS_PER_PEER=16
+export NCCL_MIN_NCHANNELS=32 # 20
+export NCCL_MAX_NCHANNELS=32 # 20
+export NCCL_IB_TIMEOUT=22
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+
+export NCCL_IB_HCA=mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_6:1,mlx5_7:1,mlx5_8:1,mlx5_9:1
+export NCCL_NET_GDR_LEVEL=7
+export NCCL_NET_GDR_READ=1
+export RCCL_SDMA_COPY_ENABLE=0
+export NCCL_TOPO_FILE="/public/home/wangxj/Projects/rccl-test/rccl-tests-0204/topo-input.xml"
+# export NCCL_TOPO_FILE="/workspace/rccl-test/rccl-tests-0204/topo-input.xml"
+export GLOG_minloglevel=3 # 打印error级别的nccl日志
+source /opt/dtk/env.sh
+# 导入hipblaslt库
+# export LD_LIBRARY_PATH=/data/hipblaslt-install-0904/lib:$LD_LIBRARY_PATH 
+# export LD_LIBRARY_PATH=/data/hipblaslt-install-dtk-25.04-0212/lib:$LD_LIBRARY_PATH 
+export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/hipblaslt-install-dtk-25.04-0212/lib:$LD_LIBRARY_PATH 
+# 更新rocblas
+# export LD_LIBRARY_PATH=/data/rocblas-install_qwen1211/lib:$LD_LIBRARY_PATH
+# export LD_LIBRARY_PATH=/data/rocblas-install_qwen1228/lib:$LD_LIBRARY_PATH
+# export LD_LIBRARY_PATH=/data/rocblas-install-0118-bf16/lib:$LD_LIBRARY_PATH
+# export LD_LIBRARY_PATH=/data/rocblas-install-0203-release/lib:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/rocblas-install-0203-release/lib:$LD_LIBRARY_PATH
+
+# torch控制多流转单流
+export ALLREDUCE_STREAM_WITH_COMPUTE=1
+export SENDRECV_STREAM_WITH_COMPUTE=1 
+
+# prof采集添加同步, 避免卡顿
+# export GPU_FLUSH_ON_EXECUTION=1
+# export HIP_DIRECT_DISPATCH=0
+
+# 采集rocblas size
+# export ROCBLAS_LAYER=3
+# export HIPBLASLT_LOG_LEVEL=3
+# 采集 fa size
+# export FLASH_ATTENTION_PRINT_PARAM=1
+
+#增加编译缓存
+export cache_size_limit=64
+
+# CHECKPOINT_PATH=./Llama-2-7b-hf-to-meg-tp1-pp2 #CHECKPOINT_PATH=./tmp_7b # 
+SAVE_PATH=./tmp_7b
+TENSORBOARD_LOGS_PATH=./tmp_7b  #$2 #<Specify path>
+DATA_PATH="/public/home/wangxj/Downloads/datasets/oscar-1GB-head/oscar-1GB_head-llama3.2_text_document" #<Specify path and file prefix>_text_document
+# DATA_PATH="/data/datasets/oscar-1GB-head/oscar-1GB_head-llama3.2_text_document" #<Specify path and file prefix>_text_document
+
+
+GPT_MODEL_ARGS=(
+    --num-layers 80 #80 #80 #40 # 20 # 
+    --hidden-size 8192
+    --ffn-hidden-size 28672 
+    --num-attention-heads 64
+    --max-position-embeddings 8192
+    --group-query-attention
+    --num-query-groups 8
+
+    --normalization RMSNorm 
+    --position-embedding-type rope
+    --untie-embeddings-and-output-weights # 分开处理embed和输出权重, 增加灵活性
+)
+
+export NVTE_FLASH_ATTN=1 # 走cutlass
+# export NVTE_FLASH_ATTN_TRITON=1 # 走triton_fa
+# --transformer-impl transformer_engine # 走core用这两组参数
+    # --use-mcore-models
+    # --transformer-impl local # 走legacy用这两组参数
+    # --use-legacy-models 
+TRAINING_ARGS=(
+    --transformer-impl local # 走legacy用这两组参数
+    --use-legacy-models 
+    --micro-batch-size 1
+    --global-batch-size 960 #32 #512 #256 # 64 #240 #60 #512 #64
+    --train-iters 100
+    --weight-decay 0.1 
+    --adam-beta1 0.9 
+    --adam-beta2 0.95 
+    --init-method-std 0.006 
+    --clip-grad 1.0 
+    --bf16
+    # --fp16 # 开启fp16需要指定loss-scale
+    # --loss-scale 1024
+    --use-distributed-optimizer 
+    --disable-bias-linear
+    --attention-dropout 0
+    --hidden-dropout 0
+    # --no-gradient-accumulation-fusion
+    --swiglu
+    --lr 3.0e-5 
+    --lr-decay-style cosine 
+    --min-lr 3.0e-6
+    --lr-warmup-iters 1
+    --ckpt-format torch
+    --ddp-average-in-collective # 在dp阶段通信中, 梯度或参数将被直接平均, 而不是先求和(到一个设备)再平均
+    # --recompute-activations
+    # --recompute-granularity full # 开启重计算降低显存增加耗时
+    # --recompute-num-layers 1 #0 #
+    # --recompute-method block
+    --overlap-grad-reduce # 重叠ddp grad reduce
+    # --tp-comm-overlap # tensor parallel comm和gemm重叠, 启动core
+    # --tp-comm-overlap-rs-dgrad # reduce-scatter和dgrad gemm重叠, 启动core
+    --use-flash-attn-cutlass
+)
+# export TORCHINDUCTOR_COORDINATE_DESCENT_TUNING=1
+# export TORCHINDUCTOR_BENCHMARK_FUSION=1
+# export TORCHINDUCTOR_BENCHMARK_MULTI_TEMPLATES=1
+# export TORCHINDUCTOR_MAX_AUTOTUNE=1
+# export TORCHINDUCTOR_CACHE_DIR=./cache
+# --use-flash-attn-cutlass # cutlass fa
+# --use-flash-attn-triton # triton fa
+# --use-flash-attn-torch # torch fa
+
+MODEL_PARALLEL_ARGS=(
+    --sequence-parallel
+	--tensor-model-parallel-size 4
+	--pipeline-model-parallel-size 8
+  # --context-parallel-size 2
+    # --num-layers-per-virtual-pipeline-stage 5
+  # --microbatch-group-size-per-virtual-pipeline-stage 1
+  # --no-overlap-p2p-communication # 开启后
+)
+
+DATA_ARGS=(
+    --data-path $DATA_PATH 
+    --seq-length 4096 #8192 #4096
+    --split 949,50,1
+    --tokenizer-type Llama3Tokenizer
+    --tokenizer-model /public/home/wangxj/Downloads/model_weights/llama3.2/tokenizer.model
+    # --tokenizer-model /data/model_weights/llama3.2/tokenizer.model
+)
+
+EVAL_AND_LOGGING_ARGS=(
+    --log-interval 1
+    --log-throughput
+    --save-interval 1000 
+    --eval-interval 1000 
+    --save $SAVE_PATH 
+    --load $SAVE_PATH 
+    --eval-iters 10
+    --tensorboard-dir $TENSORBOARD_LOGS_PATH 
+)
+
+# FINETUNE_ARGS=(
+#     # --finetune
+#     # --pretrained-checkpoint $CHECKPOINT_PATH
+#     --load $CHECKPOINT_PATH
+#     --no-load-optim
+#     --no-load-rng
+# )
+
+PROFILE_ARGS=(
+    --profile
+    --profile-step-start 4
+    --profile-step-end 5
+    --use-pytorch-profiler
+    --profile-ranks 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
+    --profile-dir prof_data
+)
+
+RANK=$OMPI_COMM_WORLD_RANK
+LOCAL_RANK=$OMPI_COMM_WORLD_LOCAL_RANK
+WORLD_SIZE=$OMPI_COMM_WORLD_SIZE
+DIST_URL=${1}
+DIST_PORT=34577
+
+DISTRIBUTED_ARGS=(
+    --rank ${RANK}
+    --world-size ${WORLD_SIZE}
+    --local-rank ${LOCAL_RANK}
+    --dist-url tcp://${DIST_URL}:${DIST_PORT}
+)
+
+APP="python -u pretrain_gpt.py \
+        ${GPT_MODEL_ARGS[@]} \
+        ${TRAINING_ARGS[@]} \
+        ${MODEL_PARALLEL_ARGS[@]} \
+        ${DATA_ARGS[@]} \
+        ${EVAL_AND_LOGGING_ARGS[@]} \
+        ${DISTRIBUTED_ARGS[@]} \
+        
+"
+# 开启profile
+# ${PROFILE_ARGS[@]} \
+
+# export HIP_VISIBLE_DEVICES=0,7 #  # 4,5,6,7 #,
+export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 #  # 4,5,6,7 #,
+# export CUDA_VISIBLE_DEVICES=4,5,6,7 # 0,1,2,3,
+# ${APP}
+case ${LOCAL_RANK} in
+[0])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[1])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  numactl --cpunodebind=1 --membind=1 ${APP}
+  ;;
+[2])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  numactl --cpunodebind=2 --membind=2 ${APP}
+  ;;
+[3])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  numactl --cpunodebind=3 --membind=3 ${APP}
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[4])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  numactl --cpunodebind=4 --membind=4 ${APP}
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[5])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  numactl --cpunodebind=5 --membind=5 ${APP}
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[6])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  numactl --cpunodebind=6 --membind=6 ${APP}
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[7])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  numactl --cpunodebind=7 --membind=7 ${APP}
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+esac
\ No newline at end of file

examples/llama/Llama3_8b.sh

+#!/bin/bash
+# set -eux
+
+#export FLASH_ATTENTION_PRINT_PARAM=1
+# Runs the "7B" parameter model
+export HSA_FORCE_FINE_GRAIN_PCIE=1
+export OMP_NUM_THREADS=1
+export NCCL_P2P_LEVEL=PXB # SYS
+
+#export HIP_ALLOC_INITIALIZE=0
+# export GPU_MAX_HW_QUEUES=10
+
+export NCCL_ALGO=Ring
+export NCCL_NCHANNELS_PER_PEER=16
+export NCCL_MIN_NCHANNELS=32 # 20
+export NCCL_MAX_NCHANNELS=32 # 20
+export NCCL_IB_TIMEOUT=22
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+
+export NCCL_IB_HCA=mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_6:1,mlx5_7:1,mlx5_8:1,mlx5_9:1
+export NCCL_NET_GDR_LEVEL=7
+export NCCL_NET_GDR_READ=1
+export RCCL_SDMA_COPY_ENABLE=0
+export NCCL_TOPO_FILE="/public/home/wangxj/Projects/rccl-test/rccl-tests-0204/topo-input.xml"
+# export NCCL_TOPO_FILE="/workspace/rccl-test/rccl-tests-0204/topo-input.xml"
+export GLOG_minloglevel=3 # 打印error级别的nccl日志
+source /opt/dtk/env.sh
+# 导入hipblaslt库
+# export LD_LIBRARY_PATH=/data/hipblaslt-install-0904/lib:$LD_LIBRARY_PATH 
+export LD_LIBRARY_PATH=/data/hipblaslt-install-0904/lib:$LD_LIBRARY_PATH 
+# 更新rocblas
+# export LD_LIBRARY_PATH=/data/rocblas-install_qwen1211/lib:$LD_LIBRARY_PATH
+# export LD_LIBRARY_PATH=/data/rocblas-install_qwen1228/lib:$LD_LIBRARY_PATH
+# export LD_LIBRARY_PATH=/data/rocblas-install-0118-bf16/lib:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/data/rocblas-install-0203-release/lib:$LD_LIBRARY_PATH
+
+# torch控制多流转单流
+export ALLREDUCE_STREAM_WITH_COMPUTE=1
+export SENDRECV_STREAM_WITH_COMPUTE=1 
+
+# prof采集添加同步, 避免卡顿
+# export GPU_FLUSH_ON_EXECUTION=1
+# export HIP_DIRECT_DISPATCH=0
+
+# 采集rocblas size
+# export ROCBLAS_LAYER=3
+# 采集 fa size
+# export FLASH_ATTENTION_PRINT_PARAM=1
+
+#增加编译缓存
+export cache_size_limit=64
+
+# CHECKPOINT_PATH=./Llama-2-7b-hf-to-meg-tp1-pp2 #CHECKPOINT_PATH=./tmp_7b # 
+SAVE_PATH=./tmp_7b
+TENSORBOARD_LOGS_PATH=./tmp_7b  #$2 #<Specify path>
+DATA_PATH="/public/home/wangxj/Downloads/datasets/oscar-1GB-head/oscar-1GB_head-llama3.2_text_document" #<Specify path and file prefix>_text_document
+
+GPT_MODEL_ARGS=(
+    --num-layers 32
+    --hidden-size 4096
+    --ffn-hidden-size 14336 
+    --num-attention-heads 32
+    --max-position-embeddings 8192
+    --group-query-attention
+    --num-query-groups 8
+
+    --swiglu
+    --normalization RMSNorm 
+    --position-embedding-type rope
+    --untie-embeddings-and-output-weights # 分开处理embed和输出权重, 增加灵活性
+)
+
+export NVTE_FLASH_ATTN=1 # 走cutlass
+# export NVTE_FLASH_ATTN_TRITON=1 # 走triton_fa
+# --transformer-impl transformer_engine # 走core用这两组参数
+    # --use-mcore-models
+    # --transformer-impl local # 走legacy用这两组参数
+    # --use-legacy-models 
+TRAINING_ARGS=(
+    --transformer-impl local # 走legacy用这两组参数
+    --use-legacy-models 
+    --micro-batch-size 1
+    --global-batch-size 64 #240 #60 #512 #64
+    --train-iters 10
+    --weight-decay 0.1 
+    --adam-beta1 0.9 
+    --adam-beta2 0.95 
+    --init-method-std 0.006 
+    --clip-grad 1.0 
+    --bf16
+    # --fp16 # 开启fp16需要指定loss-scale
+    # --loss-scale 1024
+    --use-distributed-optimizer 
+    --disable-bias-linear
+    --attention-dropout 0
+    --hidden-dropout 0
+    # --no-gradient-accumulation-fusion
+    --lr 3.0e-5 
+    --lr-decay-style cosine 
+    --min-lr 3.0e-6
+    --lr-warmup-iters 1
+    --ckpt-format torch
+    --ddp-average-in-collective # 在dp阶段通信中, 梯度或参数将被直接平均, 而不是先求和(到一个设备)再平均
+    # --recompute-granularity full # 开启重计算降低显存增加耗时
+    # --recompute-num-layers 5 #0 #
+    # --recompute-method block
+    --overlap-grad-reduce # 重叠ddp grad reduce
+    # --tp-comm-overlap # tensor parallel comm和gemm重叠, 优化项未适配
+    # --tp-comm-overlap-rs-dgrad # reduce-scatter和dgrad gemm重叠, 优化项未适配
+    --use-flash-attn
+)
+# export TORCHINDUCTOR_COORDINATE_DESCENT_TUNING=1
+# export TORCHINDUCTOR_BENCHMARK_FUSION=1
+# export TORCHINDUCTOR_BENCHMARK_MULTI_TEMPLATES=1
+# export TORCHINDUCTOR_MAX_AUTOTUNE=1
+# export TORCHINDUCTOR_CACHE_DIR=./cache
+# --use-flash-attn-cutlass # cutlass fa
+# --use-flash-attn-triton # triton fa
+# --use-flash-attn-torch # torch fa
+
+MODEL_PARALLEL_ARGS=(
+    --sequence-parallel
+	--tensor-model-parallel-size 2
+	--pipeline-model-parallel-size 2
+)
+
+DATA_ARGS=(
+    --data-path $DATA_PATH 
+    --seq-length 4096 #4096
+    --split 949,50,1
+    --tokenizer-type Llama3Tokenizer
+    --tokenizer-model /public/home/wangxj/Downloads/model_weights/llama3.2/tokenizer.model
+)
+
+EVAL_AND_LOGGING_ARGS=(
+    --log-interval 1
+    --log-throughput
+    --save-interval 1000 
+    --eval-interval 1000 
+    --save $SAVE_PATH 
+    --load $SAVE_PATH 
+    --eval-iters 10
+    --tensorboard-dir $TENSORBOARD_LOGS_PATH 
+)
+
+# FINETUNE_ARGS=(
+#     # --finetune
+#     # --pretrained-checkpoint $CHECKPOINT_PATH
+#     --load $CHECKPOINT_PATH
+#     --no-load-optim
+#     --no-load-rng
+# )
+
+PROFILE_ARGS=(
+    --profile
+    --profile-step-start 4
+    --profile-step-end 5
+    --use-pytorch-profiler
+    --profile-ranks 0 1 2 3 4 5 6 7
+    --profile-dir prof_data
+)
+
+RANK=$OMPI_COMM_WORLD_RANK
+LOCAL_RANK=$OMPI_COMM_WORLD_LOCAL_RANK
+WORLD_SIZE=$OMPI_COMM_WORLD_SIZE
+DIST_URL=${1}
+DIST_PORT=34577
+
+DISTRIBUTED_ARGS=(
+    --rank ${RANK}
+    --world-size ${WORLD_SIZE}
+    --local-rank ${LOCAL_RANK}
+    --dist-url tcp://${DIST_URL}:${DIST_PORT}
+)
+
+APP="python -u ../../pretrain_gpt.py \
+        ${GPT_MODEL_ARGS[@]} \
+        ${TRAINING_ARGS[@]} \
+        ${MODEL_PARALLEL_ARGS[@]} \
+        ${DATA_ARGS[@]} \
+        ${EVAL_AND_LOGGING_ARGS[@]} \
+        ${DISTRIBUTED_ARGS[@]} \
+        
+"
+# 开启profile
+# ${PROFILE_ARGS[@]} \
+
+# export HIP_VISIBLE_DEVICES=0,7 #  # 4,5,6,7 #,
+export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 #  # 4,5,6,7 #,
+# export CUDA_VISIBLE_DEVICES=4,5,6,7 # 0,1,2,3,
+${APP}
+# case ${LOCAL_RANK} in
+# [0])
+#   export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+#   # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+#   numactl --cpunodebind=0 --membind=0 ${APP}
+#   ;;
+# [1])
+#   export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+#   # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+#   numactl --cpunodebind=1 --membind=1 ${APP}
+#   ;;
+# [2])
+#   export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+#   # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+#   numactl --cpunodebind=2 --membind=2 ${APP}
+#   ;;
+# [3])
+#   export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+#   numactl --cpunodebind=3 --membind=3 ${APP}
+#   # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+#   ;;
+# [4])
+#   export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+#   numactl --cpunodebind=4 --membind=4 ${APP}
+#   # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+#   ;;
+# [5])
+#   export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+#   numactl --cpunodebind=5 --membind=5 ${APP}
+#   # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+#   ;;
+# [6])
+#   export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+#   numactl --cpunodebind=6 --membind=6 ${APP}
+#   # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+#   ;;
+# [7])
+#   export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+#   numactl --cpunodebind=7 --membind=7 ${APP}
+#   # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+#   ;;
+# esac
\ No newline at end of file

examples/qwen/qwen1.5_14b.sh

+#!/bin/bash
+set -eux
+
+#export FLASH_ATTENTION_PRINT_PARAM=1
+# Runs the "7B" parameter model
+export HSA_FORCE_FINE_GRAIN_PCIE=1
+export OMP_NUM_THREADS=1
+export NCCL_P2P_LEVEL=PXB # SYS
+
+#export HIP_ALLOC_INITIALIZE=0
+# export GPU_MAX_HW_QUEUES=10
+
+export NCCL_ALGO=Ring
+export NCCL_NCHANNELS_PER_PEER=16
+export NCCL_MIN_NCHANNELS=32 # 20
+export NCCL_MAX_NCHANNELS=32 # 20
+export NCCL_IB_TIMEOUT=22
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+
+export NCCL_IB_HCA=mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_6:1,mlx5_7:1,mlx5_8:1,mlx5_9:1
+export NCCL_NET_GDR_LEVEL=7
+export NCCL_NET_GDR_READ=1
+export RCCL_SDMA_COPY_ENABLE=0
+export NCCL_TOPO_FILE="/public/home/wangxj/Projects/rccl-test/rccl-tests-0204/topo-input.xml"
+# export NCCL_TOPO_FILE="/workspace/rccl-test/rccl-tests-0204/topo-input.xml"
+export GLOG_minloglevel=3 # 打印error级别的nccl日志
+source /opt/dtk/env.sh
+# 导入hipblaslt库
+# export LD_LIBRARY_PATH=/data/hipblaslt-install-0904/lib:$LD_LIBRARY_PATH 
+# export LD_LIBRARY_PATH=/data/hipblaslt-install-dtk-25.04-0212/lib:$LD_LIBRARY_PATH 
+export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/hipblaslt-install-dtk-25.04-0212/lib:$LD_LIBRARY_PATH 
+
+# 更新rocblas
+# export LD_LIBRARY_PATH=/data/rocblas-install_qwen1211/lib:$LD_LIBRARY_PATH
+# export LD_LIBRARY_PATH=/data/rocblas-install_qwen1228/lib:$LD_LIBRARY_PATH
+# export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/rocblas-install-0224/lib:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/rocblas-install-0331-release/lib:$LD_LIBRARY_PATH
+# export LD_LIBRARY_PATH=/data/rocblas-install-0118-bf16/lib:$LD_LIBRARY_PATH
+
+# torch控制多流转单流
+export ALLREDUCE_STREAM_WITH_COMPUTE=1
+export SENDRECV_STREAM_WITH_COMPUTE=1 
+
+# prof采集添加同步, 避免卡顿
+# export GPU_FLUSH_ON_EXECUTION=1
+# export HIP_DIRECT_DISPATCH=0
+
+# 采集rocblas size
+# export ROCBLAS_LAYER=3
+# 采集 fa size
+# export FLASH_ATTENTION_PRINT_PARAM=1
+
+#增加编译缓存
+export cache_size_limit=64
+
+# CHECKPOINT_PATH=./Llama-2-7b-hf-to-meg-tp1-pp2 #CHECKPOINT_PATH=./tmp_7b # 
+SAVE_PATH=./tmp_7b
+TENSORBOARD_LOGS_PATH=./tmp_7b  #$2 #<Specify path>
+DATA_PATH="/public/home/wangxj/Downloads/datasets/oscar-1GB-head/oscar-1GB_head-qwen_text_document" #<Specify path and file prefix>_text_document
+# DATA_PATH="/data/datasets/oscar-1GB-head/oscar-1GB_head-llama2_text_document" #<Specify path and file prefix>_text_document
+
+GPT_MODEL_ARGS=(
+    --num-layers 40
+    --hidden-size 5120
+    --ffn-hidden-size 13696 
+    --num-attention-heads 40
+    --max-position-embeddings 32768
+    --num-query-groups 40
+    --group-query-attention
+
+    --normalization RMSNorm 
+    --position-embedding-type rope # none # 
+    --untie-embeddings-and-output-weights # 分开处理embed和输出权重, 增加灵活性
+)
+
+export NVTE_FLASH_ATTN=1 # 走cutlass
+# export NVTE_FLASH_ATTN_TRITON=1 # 走triton_fa
+# --transformer-impl transformer_engine # 走core用这两组参数
+    # --use-mcore-models
+    # --transformer-impl local # 走legacy用这两组参数
+    # --use-legacy-models 
+TRAINING_ARGS=(
+    --transformer-impl local # 走legacy用这两组参数
+    --use-legacy-models 
+    --micro-batch-size 1
+    --global-batch-size 32 #256 #240 #60 #512 #64
+    --train-iters 50
+    --weight-decay 0.1 
+    --adam-beta1 0.9 
+    --adam-beta2 0.95 
+    --init-method-std 0.006 
+    --clip-grad 1.0 
+    --bf16
+    # --fp16 # 开启fp16需要指定loss-scale
+    # --loss-scale 1024
+    --use-distributed-optimizer 
+    --disable-bias-linear
+    --attention-dropout 0
+    --hidden-dropout 0
+    # --no-gradient-accumulation-fusion
+    --swiglu
+    --add-qkv-bias
+    --lr 3.0e-5 
+    --lr-decay-style cosine 
+    --min-lr 3.0e-6
+    --lr-warmup-iters 1
+    --ckpt-format torch
+    --ddp-average-in-collective # 在dp阶段通信中, 梯度或参数将被直接平均, 而不是先求和(到一个设备)再平均
+    # --recompute-granularity full # 开启重计算降低显存增加耗时
+    # --recompute-num-layers 5 #0 #
+    # --recompute-method block
+    --overlap-grad-reduce # 重叠ddp grad reduce
+    # --tp-comm-overlap # tensor parallel comm和gemm重叠, 优化项未适配
+    # --tp-comm-overlap-rs-dgrad # reduce-scatter和dgrad gemm重叠
+    --use-flash-attn
+)
+# 使用torch fa的环境变量
+# export TORCHINDUCTOR_COORDINATE_DESCENT_TUNING=1
+# export TORCHINDUCTOR_BENCHMARK_FUSION=1
+# export TORCHINDUCTOR_BENCHMARK_MULTI_TEMPLATES=1
+# export TORCHINDUCTOR_MAX_AUTOTUNE=1
+# export TORCHINDUCTOR_CACHE_DIR=./cache
+# --use-flash-attn-cutlass # cutlass fa
+# --use-flash-attn-triton # triton fa
+# --use-flash-attn-torch # torch fa
+
+MODEL_PARALLEL_ARGS=(
+    --sequence-parallel
+	--tensor-model-parallel-size 4
+	--pipeline-model-parallel-size 1
+
+  # --num-layers-per-virtual-pipeline-stage 4
+  # --microbatch-group-size-per-virtual-pipeline-stage 1
+  # --no-overlap-p2p-communication # 开启后
+)
+
+DATA_ARGS=(
+    --data-path $DATA_PATH 
+    --seq-length 4096 #4096
+    --split 949,50,1
+    --tokenizer-type QwenTokenizer
+    --merge-file /public/home/wangxj/Downloads/model_weights/qwen1.5_14b/merges.txt
+    --vocab-file /public/home/wangxj/Downloads/model_weights/qwen1.5_14b/vocab.json
+    # --tokenizer-model /public/home/wangxj/Downloads/model_weights/llama2_7b_hf/tokenizer.model
+    # --tokenizer-model /data/model_weights/llama2_7b_hf/tokenizer.model
+)
+
+EVAL_AND_LOGGING_ARGS=(
+    --log-interval 1
+    --log-throughput
+    --save-interval 1000 
+    --eval-interval 1000 
+    --save $SAVE_PATH 
+    --load $SAVE_PATH 
+    --eval-iters 10
+    --tensorboard-dir $TENSORBOARD_LOGS_PATH 
+)
+
+# FINETUNE_ARGS=(
+#     # --finetune
+#     # --pretrained-checkpoint $CHECKPOINT_PATH
+#     --load $CHECKPOINT_PATH
+#     --no-load-optim
+#     --no-load-rng
+# )
+
+PROFILE_ARGS=(
+    --profile
+    --profile-step-start 4
+    --profile-step-end 5
+    --use-pytorch-profiler
+    --profile-ranks 0 1 2 3 4 5 6 7
+    --profile-dir prof_data
+)
+
+RANK=$OMPI_COMM_WORLD_RANK
+LOCAL_RANK=$OMPI_COMM_WORLD_LOCAL_RANK
+WORLD_SIZE=$OMPI_COMM_WORLD_SIZE
+DIST_URL=${1}
+DIST_PORT=34577
+
+DISTRIBUTED_ARGS=(
+    --rank ${RANK}
+    --world-size ${WORLD_SIZE}
+    --local-rank ${LOCAL_RANK}
+    --dist-url tcp://${DIST_URL}:${DIST_PORT}
+)
+
+# torchrun --nproc_per_node 8 pretrain_gpt.py \
+#     ${GPT_MODEL_ARGS[@]} \
+#     ${TRAINING_ARGS[@]} \
+#     ${MODEL_PARALLEL_ARGS[@]} \
+#     ${DATA_ARGS[@]} \
+#     ${EVAL_AND_LOGGING_ARGS[@]}
+
+APP="python -u ../../pretrain_gpt.py \
+        ${GPT_MODEL_ARGS[@]} \
+        ${TRAINING_ARGS[@]} \
+        ${MODEL_PARALLEL_ARGS[@]} \
+        ${DATA_ARGS[@]} \
+        ${EVAL_AND_LOGGING_ARGS[@]} \
+        ${DISTRIBUTED_ARGS[@]} \
+        
+"
+# 开启profile
+# ${PROFILE_ARGS[@]} \
+
+# export HIP_VISIBLE_DEVICES=0,7 #  # 4,5,6,7 #,
+export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 #  # 4,5,6,7 #,
+# export CUDA_VISIBLE_DEVICES=4,5,6,7 # 0,1,2,3,
+# ${APP}
+case ${LOCAL_RANK} in
+[0])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[1])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  numactl --cpunodebind=1 --membind=1 ${APP}
+  ;;
+[2])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  numactl --cpunodebind=2 --membind=2 ${APP}
+  ;;
+[3])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  numactl --cpunodebind=3 --membind=3 ${APP}
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[4])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  numactl --cpunodebind=4 --membind=4 ${APP}
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[5])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  numactl --cpunodebind=5 --membind=5 ${APP}
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[6])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  numactl --cpunodebind=6 --membind=6 ${APP}
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[7])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  numactl --cpunodebind=7 --membind=7 ${APP}
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+esac

examples/qwen/qwen1.5_32b.sh

+#!/bin/bash
+# set -eux
+
+#export FLASH_ATTENTION_PRINT_PARAM=1
+# Runs the "7B" parameter model
+export HSA_FORCE_FINE_GRAIN_PCIE=1
+export OMP_NUM_THREADS=1
+export NCCL_P2P_LEVEL=PXB # SYS
+
+#export HIP_ALLOC_INITIALIZE=0
+# export GPU_MAX_HW_QUEUES=10
+
+export NCCL_ALGO=Ring
+export NCCL_NCHANNELS_PER_PEER=16
+export NCCL_MIN_NCHANNELS=32 # 20
+export NCCL_MAX_NCHANNELS=32 # 20
+export NCCL_IB_TIMEOUT=22
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+
+export NCCL_IB_HCA=mlx5_2:1,mlx5_3:1,mlx5_4:1,mlx5_5:1,mlx5_6:1,mlx5_7:1,mlx5_8:1,mlx5_9:1
+export NCCL_NET_GDR_LEVEL=7
+export NCCL_NET_GDR_READ=1
+export RCCL_SDMA_COPY_ENABLE=0
+export NCCL_TOPO_FILE="/public/home/wangxj/Projects/rccl-test/rccl-tests-0204/topo-input.xml"
+# export NCCL_TOPO_FILE="/workspace/rccl-test/rccl-tests-0204/topo-input.xml"
+export GLOG_minloglevel=3 # 打印error级别的nccl日志
+source /opt/dtk/env.sh
+# 导入hipblaslt库
+# export LD_LIBRARY_PATH=/data/hipblaslt-install-0904/lib:$LD_LIBRARY_PATH 
+# export LD_LIBRARY_PATH=/data/hipblaslt-install-dtk-25.04-0212/lib:$LD_LIBRARY_PATH 
+export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/hipblaslt-install-dtk-25.04-0212/lib:$LD_LIBRARY_PATH 
+
+# 更新rocblas
+# export LD_LIBRARY_PATH=/data/rocblas-install_qwen1211/lib:$LD_LIBRARY_PATH
+# export LD_LIBRARY_PATH=/data/rocblas-install_qwen1228/lib:$LD_LIBRARY_PATH
+# export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/rocblas-install-0224/lib:$LD_LIBRARY_PATH
+export LD_LIBRARY_PATH=/public/home/wangxj/Downloads/rocblas-install-0331-release/lib:$LD_LIBRARY_PATH
+# export LD_LIBRARY_PATH=/data/rocblas-install-0118-bf16/lib:$LD_LIBRARY_PATH
+
+# torch控制多流转单流
+export ALLREDUCE_STREAM_WITH_COMPUTE=1
+export SENDRECV_STREAM_WITH_COMPUTE=1 
+
+# prof采集添加同步, 避免卡顿
+# export GPU_FLUSH_ON_EXECUTION=1
+# export HIP_DIRECT_DISPATCH=0
+
+# 采集rocblas size
+# export ROCBLAS_LAYER=3
+# 采集 fa size
+# export FLASH_ATTENTION_PRINT_PARAM=1
+
+#增加编译缓存
+export cache_size_limit=64
+
+# CHECKPOINT_PATH=./Llama-2-7b-hf-to-meg-tp1-pp2 #CHECKPOINT_PATH=./tmp_7b # 
+SAVE_PATH=./tmp_7b
+TENSORBOARD_LOGS_PATH=./tmp_7b  #$2 #<Specify path>
+DATA_PATH="/public/home/wangxj/Downloads/datasets/oscar-1GB-head/oscar-1GB_head-qwen_text_document" #<Specify path and file prefix>_text_document
+# DATA_PATH="/data/datasets/oscar-1GB-head/oscar-1GB_head-llama2_text_document" #<Specify path and file prefix>_text_document
+
+GPT_MODEL_ARGS=(
+    --num-layers 64
+    --hidden-size 5120
+    --ffn-hidden-size 27392 
+    --num-attention-heads 40
+    --max-position-embeddings 32768
+    --num-query-groups 8
+    --group-query-attention
+
+    --normalization RMSNorm 
+    --position-embedding-type rope # none # 
+    --untie-embeddings-and-output-weights # 分开处理embed和输出权重, 增加灵活性
+)
+
+export NVTE_FLASH_ATTN=1 # 走cutlass
+# export NVTE_FLASH_ATTN_TRITON=1 # 走triton_fa
+# --transformer-impl transformer_engine # 走core用这两组参数
+    # --use-mcore-models
+    # --transformer-impl local # 走legacy用这两组参数
+    # --use-legacy-models 
+TRAINING_ARGS=(
+    --transformer-impl local # 走legacy用这两组参数
+    --use-legacy-models 
+    --micro-batch-size 1
+    --global-batch-size 1024 #256 #240 #60 #512 #64
+    --train-iters 50
+    --weight-decay 0.1 
+    --adam-beta1 0.9 
+    --adam-beta2 0.95 
+    --init-method-std 0.006 
+    --clip-grad 1.0 
+    --bf16
+    # --fp16 # 开启fp16需要指定loss-scale
+    # --loss-scale 1024
+    --use-distributed-optimizer 
+    --disable-bias-linear
+    --attention-dropout 0
+    --hidden-dropout 0
+    # --no-gradient-accumulation-fusion
+    --swiglu
+    --add-qkv-bias
+    --lr 3.0e-5 
+    --lr-decay-style cosine 
+    --min-lr 3.0e-6
+    --lr-warmup-iters 1
+    --ckpt-format torch
+    --ddp-average-in-collective # 在dp阶段通信中, 梯度或参数将被直接平均, 而不是先求和(到一个设备)再平均
+    # --recompute-granularity full # 开启重计算降低显存增加耗时
+    # --recompute-num-layers 5 #0 #
+    # --recompute-method block
+    --overlap-grad-reduce # 重叠ddp grad reduce
+    # --tp-comm-overlap # tensor parallel comm和gemm重叠, 优化项未适配
+    # --tp-comm-overlap-rs-dgrad # reduce-scatter和dgrad gemm重叠
+    --use-flash-attn
+)
+# 使用torch fa的环境变量
+# export TORCHINDUCTOR_COORDINATE_DESCENT_TUNING=1
+# export TORCHINDUCTOR_BENCHMARK_FUSION=1
+# export TORCHINDUCTOR_BENCHMARK_MULTI_TEMPLATES=1
+# export TORCHINDUCTOR_MAX_AUTOTUNE=1
+# export TORCHINDUCTOR_CACHE_DIR=./cache
+# --use-flash-attn-cutlass # cutlass fa
+# --use-flash-attn-triton # triton fa
+# --use-flash-attn-torch # torch fa
+
+MODEL_PARALLEL_ARGS=(
+    --sequence-parallel
+	--tensor-model-parallel-size 4
+	--pipeline-model-parallel-size 4
+
+  # --num-layers-per-virtual-pipeline-stage 4
+  # --microbatch-group-size-per-virtual-pipeline-stage 1
+  # --no-overlap-p2p-communication # 开启后
+)
+
+DATA_ARGS=(
+    --data-path $DATA_PATH 
+    --seq-length 4096 #4096
+    --split 949,50,1
+    --tokenizer-type QwenTokenizer
+    --merge-file /public/home/wangxj/Downloads/model_weights/qwen1.5_14b/merges.txt
+    --vocab-file /public/home/wangxj/Downloads/model_weights/qwen1.5_14b/vocab.json
+    # --tokenizer-model /public/home/wangxj/Downloads/model_weights/llama2_7b_hf/tokenizer.model
+    # --tokenizer-model /data/model_weights/llama2_7b_hf/tokenizer.model
+)
+
+EVAL_AND_LOGGING_ARGS=(
+    --log-interval 1
+    --log-throughput
+    --save-interval 1000 
+    --eval-interval 1000 
+    --save $SAVE_PATH 
+    --load $SAVE_PATH 
+    --eval-iters 10
+    --tensorboard-dir $TENSORBOARD_LOGS_PATH 
+)
+
+# FINETUNE_ARGS=(
+#     # --finetune
+#     # --pretrained-checkpoint $CHECKPOINT_PATH
+#     --load $CHECKPOINT_PATH
+#     --no-load-optim
+#     --no-load-rng
+# )
+
+PROFILE_ARGS=(
+    --profile
+    --profile-step-start 4
+    --profile-step-end 5
+    --use-pytorch-profiler
+    --profile-ranks 0 1 2 3 4 5 6 7
+    --profile-dir prof_data
+)
+
+RANK=$OMPI_COMM_WORLD_RANK
+LOCAL_RANK=$OMPI_COMM_WORLD_LOCAL_RANK
+WORLD_SIZE=$OMPI_COMM_WORLD_SIZE
+DIST_URL=${1}
+DIST_PORT=34577
+
+DISTRIBUTED_ARGS=(
+    --rank ${RANK}
+    --world-size ${WORLD_SIZE}
+    --local-rank ${LOCAL_RANK}
+    --dist-url tcp://${DIST_URL}:${DIST_PORT}
+)
+
+# torchrun --nproc_per_node 8 pretrain_gpt.py \
+#     ${GPT_MODEL_ARGS[@]} \
+#     ${TRAINING_ARGS[@]} \
+#     ${MODEL_PARALLEL_ARGS[@]} \
+#     ${DATA_ARGS[@]} \
+#     ${EVAL_AND_LOGGING_ARGS[@]}
+
+APP="python -u ../../pretrain_gpt.py \
+        ${GPT_MODEL_ARGS[@]} \
+        ${TRAINING_ARGS[@]} \
+        ${MODEL_PARALLEL_ARGS[@]} \
+        ${DATA_ARGS[@]} \
+        ${EVAL_AND_LOGGING_ARGS[@]} \
+        ${DISTRIBUTED_ARGS[@]} \
+        
+"
+# 开启profile
+# ${PROFILE_ARGS[@]} \
+
+# export HIP_VISIBLE_DEVICES=0,7 #  # 4,5,6,7 #,
+export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 #  # 4,5,6,7 #,
+# export CUDA_VISIBLE_DEVICES=4,5,6,7 # 0,1,2,3,
+# ${APP}
+case ${LOCAL_RANK} in
+[0])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[1])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  numactl --cpunodebind=1 --membind=1 ${APP}
+  ;;
+[2])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  numactl --cpunodebind=2 --membind=2 ${APP}
+  ;;
+[3])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  numactl --cpunodebind=3 --membind=3 ${APP}
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[4])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  numactl --cpunodebind=4 --membind=4 ${APP}
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[5])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  numactl --cpunodebind=5 --membind=5 ${APP}
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[6])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  numactl --cpunodebind=6 --membind=6 ${APP}
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[7])
+  export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  numactl --cpunodebind=7 --membind=7 ${APP}
+  # hipprof --hip-trace --trace-off numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+esac

examples/qwen/qwen_data_process.sh

+python tools/preprocess_data.py \
+    --input /public/home/wangxj/Downloads/datasets/oscar-1GB-head/oscar-1GB_head.jsonl \
+    --output-prefix /public/home/wangxj/Downloads/datasets/oscar-1GB-head/oscar-1GB_head-qwen \
+    --vocab-file /public/home/wangxj/Downloads/model_weights/qwen1.5_14b/vocab.json \
+    --tokenizer-type QwenTokenizer \
+    --merge-file /public/home/wangxj/Downloads/model_weights/qwen1.5_14b/merges.txt \
+    --append-eod \
+    --workers 8

requirements.txt

+trl
+transformers >= 4.43.0
+packaging
+six
+regex
+pyyaml
+sentencepiece
+pybind11
+blobfile
+flux
+
+# ==== test ====
+nltk
+pytest
+requests 
+wrapt
+tensorboard
+tensorboardX
+scipy
+psutil