删除用于dcu的te的batchgemm, 添加nv上的llama和qwen脚本

dcu_megatron/adaptor/megatron_adaptor.py

@@ -140,13 +140,13 @@ class CoreAdaptation(MegatronAdaptationABC):
         from ..core.extensions.transformer_engine import TEDotProductAttentionPatch
         from megatron.core.extensions.transformer_engine import TEGroupedLinear
 
-        if not is_te_min_version("1.10.0"):
-            # kv channels, te_min_version 1.10.0 -> 1.9.0
-            MegatronAdaptation.register('megatron.core.extensions.transformer_engine.TEDotProductAttention.__init__',
-                                        TEDotProductAttentionPatch.__init__)
+        # if not is_te_min_version("1.10.0"):
+        #     # kv channels, te_min_version 1.10.0 -> 1.9.0
+        #     MegatronAdaptation.register('megatron.core.extensions.transformer_engine.TEDotProductAttention.__init__',
+        #                                 TEDotProductAttentionPatch.__init__)
 
-        if int(os.getenv("GROUPED_GEMM_BatchLinear", '0')):
-            TEGroupedLinear.__bases__ = (te.pytorch.BatchedLinear if is_te_min_version("2.3.0.dev0") else te.pytorch.BatchLinear,)
+        # if int(os.getenv("GROUPED_GEMM_BatchLinear", '0')):
+        #     TEGroupedLinear.__bases__ = (te.pytorch.BatchedLinear if is_te_min_version("2.3.0.dev0") else te.pytorch.BatchLinear,)
 
     def patch_tensor_parallel(self):
         from ..core.tensor_parallel.cross_entropy import VocabParallelCrossEntropy

examples/llama/hostfile

+a121 slots=8
+a124 slots=8
\ No newline at end of file

examples/llama/run_llama2_13B.sh

+for para in $*
+do
+    if [[ $para == --profiling* ]];then
+        profiling=${para#*=}
+    fi
+done
+
+# Those variables need to modify
+GPUS="16"                 # how many gpus to use
+HOST="a121"                 # hostname
+PORT="11452"                 # port id
+DATA_PATH="/data/datasets/oscar-1GB_head-llama2_text_document"            # path to oscar-1GB_head-llama2_text_document
+TOKENIZER_MODEL_PATH="/data/models/llama2/tokenizer.model" # path to tokenizer.model
+CHECKPOINT_PATH="./ckpt"      # path to ckpt
+
+# Runs Llama2 7B model
+mpirun -np ${GPUS}  --hostfile hostfile \
+                    --allow-run-as-root \
+                    --bind-to none \
+                    --mca plm_rsh_no_tree_spawn 1 \
+                    --mca plm_rsh_args "-p 11451" \
+                    bash -c "
+                    ./train_llama2_13b_2nodes.sh \
+                    ${HOST} \
+                    ${PORT} \
+                    --data_path=$DATA_PATH \
+                    --tokenizer_path=$TOKENIZER_MODEL_PATH \
+                    --checkpoint_path=$CHECKPOINT_PATH \
+                    --profiling=$profiling" > ./logs/log-$((${GPUS} / 8))nodes-`date +%F-%H%M`.log 2>&1
+
+wait
\ No newline at end of file

examples/llama/run_llama2_7B.sh

@@ -6,9 +6,9 @@ do
 done
 
 # Those variables need to modify
-GPUS="4"                 # how many gpus to use
-HOST="localhost"                 # hostname
-PORT="11451"                 # port id
+GPUS="16"                 # how many gpus to use
+HOST="a121"                 # hostname
+PORT="11452"                 # port id
 DATA_PATH="/data/datasets/oscar-1GB_head-llama2_text_document"            # path to oscar-1GB_head-llama2_text_document
 TOKENIZER_MODEL_PATH="/data/models/llama2/tokenizer.model" # path to tokenizer.model
 CHECKPOINT_PATH="./ckpt"      # path to ckpt
@@ -18,13 +18,14 @@ mpirun -np ${GPUS}  --hostfile hostfile \
                     --allow-run-as-root \
                     --bind-to none \
                     --mca plm_rsh_no_tree_spawn 1 \
+                    --mca plm_rsh_args "-p 11451" \
                     bash -c "
-                    ./train_llama2_7b_1nodes.sh \
+                    ./train_llama2_7b_2nodes.sh \
                     ${HOST} \
                     ${PORT} \
                     --data_path=$DATA_PATH \
                     --tokenizer_path=$TOKENIZER_MODEL_PATH \
                     --checkpoint_path=$CHECKPOINT_PATH \
-                    --profiling=$profiling" > log-$((${GPUS} / 8))nodes-`date +%F-%H%M`.log 2>&1
+                    --profiling=$profiling" > ./logs/log-$((${GPUS} / 8))nodes-`date +%F-%H%M`.log 2>&1
 
 wait
\ No newline at end of file

examples/llama/train_llama2_13b_1nodes.sh

+#!/bin/bash
+
+INITIALIZATION_ARGS=( --num-workers 2)
+
+for para in $*
+do
+    if [[ $para == --data_path* ]];then
+        data_path=${para#*=}
+    elif [[ $para == --tokenizer_path* ]];then
+        tokenizer_path=${para#*=}
+    elif [[ $para == --checkpoint_path* ]];then
+        checkpoint_path=${para#*=}
+    elif [[ $para == --profiling* ]];then
+        profiling=${para#*=}
+    elif [[ $para == --reproduce* ]];then
+        INITIALIZATION_ARGS=( --reproduce --num-workers 0)
+        export MIOPEN_DEBUG_CONVOLUTION_DETERMINISTIC=1  # miopen 确定算法打开
+        export ROCBLAS_ATOMICS_MOD=0                     # rocblas 关闭原子操作
+        # 关闭miopen中的atomic操作算法, 只保留gemm算法
+        export MIOPEN_DEBUG_CONV_FFT=0
+        export MIOPEN_DEBUG_CONV_DIRECT=0
+        export MIOPEN_DEBUG_CONV_GEMM=1
+        export MIOPEN_DEBUG_CONV_WINOGRAD=0
+        export MIOPEN_DEBUG_CONV_IMPLICIT_GEMM=0
+    fi
+done
+
+# data path
+DATA_PATH=${data_path}
+TOKENIZER_MODEL_PATH=${tokenizer_path}
+CHECKPOINT_PATH=${checkpoint_path}
+
+# default env
+DIST_URL=${1}
+DIST_PORT=${2}
+RANK=$OMPI_COMM_WORLD_RANK
+LOCAL_RANK=$OMPI_COMM_WORLD_LOCAL_RANK
+WORLD_SIZE=$OMPI_COMM_WORLD_SIZE
+CURRENT_DIR="$( cd "$( dirname "$0" )" && pwd )"
+MEGATRON_PATH=$( dirname $( dirname ${CURRENT_DIR}))
+export GLOG_minloglevel=3
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+export HSA_FORCE_FINE_GRAIN_PCIE=1
+export OMP_NUM_THREADS=1
+export GPU_MAX_HW_QUEUES=10
+export PYTHONPATH=${MEGATRON_PATH}/Megatron-LM:$PYTHONPATH
+
+# 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
+
+DISTRIBUTED_ARGS=(
+    --rank ${RANK}
+    --world-size ${WORLD_SIZE}
+    --local-rank ${LOCAL_RANK}
+    --dist-url tcp://${DIST_URL}:${DIST_PORT}
+)
+
+GPT_MODEL_ARGS=(
+    --seq-length 4096
+    --num-layers 40
+    --hidden-size 5120
+    --ffn-hidden-size 13824 
+    --num-attention-heads 40
+    --max-position-embeddings 4096
+    --normalization RMSNorm
+    --position-embedding-type rope
+    --untie-embeddings-and-output-weights
+)
+
+TRAINING_ARGS=(
+    --transformer-impl local
+    --use-legacy-models 
+    --micro-batch-size 1
+    --global-batch-size 256
+    --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
+    --disable-bias-linear
+    --attention-dropout 0
+    --hidden-dropout 0
+    --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
+    --overlap-grad-reduce
+    --use-flash-attn
+)
+
+MODEL_PARALLEL_ARGS=(
+    --tensor-model-parallel-size 2
+    --pipeline-model-parallel-size 2
+    --context-parallel-size 1
+    --use-distributed-optimizer 
+    --sequence-parallel
+)
+
+DATA_ARGS=(
+    --tokenizer-type Llama2Tokenizer
+    --tokenizer-model ${TOKENIZER_MODEL_PATH}
+    --data-path ${DATA_PATH} 
+    --split 949,50,1
+)
+
+EVAL_AND_LOGGING_ARGS=(
+    --log-throughput
+    --eval-iters 5
+    --log-interval 1
+    --save-interval 1000 
+    --eval-interval 1000 
+    --save $CHECKPOINT_PATH
+    --load $CHECKPOINT_PATH
+    --tensorboard-dir "${CHECKPOINT_PATH}/tensorboard" 
+)
+
+TORCH_PROFIE_ARGS=(
+    --profile
+    --profile-ranks 0 1 2 3 4 5 6 7
+    --profile-step-start 3
+    --profile-step-end 4
+    --profile-dir torch_prof_llama_1nodes_tp1-pp2-cp1
+    --use-pytorch-profiler
+)
+
+HIP_PROFIE_ARGS=(
+    --profile
+    --profile-ranks 0 1 2 3 4 5 6 7
+    --profile-step-start 4
+    --profile-step-end 5
+    --use-hip-profiler
+)
+
+APP="python -u ${MEGATRON_PATH}/pretrain_gpt.py \
+    ${GPT_MODEL_ARGS[@]} \
+    ${TRAINING_ARGS[@]} \
+    ${MODEL_PARALLEL_ARGS[@]} \
+    ${DATA_ARGS[@]} \
+    ${EVAL_AND_LOGGING_ARGS[@]} \
+    ${DISTRIBUTED_ARGS[@]} \
+    ${INITIALIZATION_ARGS[@]} \
+    "
+
+if [[ $profiling == "torch" ]]; then
+    APP+=" ${TORCH_PROFIE_ARGS[@]}"
+elif [[ $profiling == "hip" ]]; then
+    mkdir -p hip_prof_data
+    APP+=" ${HIP_PROFIE_ARGS[@]}"
+    APP="hipprof -d hip_prof_data --hip-trace --trace-off ${APP}"
+fi
+
+#for hygon cpu
+export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+${APP}
\ No newline at end of file

examples/llama/train_llama2_13b_2nodes.sh

+#!/bin/bash
+
+INITIALIZATION_ARGS=( --num-workers 2)
+
+for para in $*
+do
+    if [[ $para == --data_path* ]];then
+        data_path=${para#*=}
+    elif [[ $para == --tokenizer_path* ]];then
+        tokenizer_path=${para#*=}
+    elif [[ $para == --checkpoint_path* ]];then
+        checkpoint_path=${para#*=}
+    elif [[ $para == --profiling* ]];then
+        profiling=${para#*=}
+    elif [[ $para == --reproduce* ]];then
+        INITIALIZATION_ARGS=( --reproduce --num-workers 0)
+        export MIOPEN_DEBUG_CONVOLUTION_DETERMINISTIC=1  # miopen 确定算法打开
+        export ROCBLAS_ATOMICS_MOD=0                     # rocblas 关闭原子操作
+        # 关闭miopen中的atomic操作算法, 只保留gemm算法
+        export MIOPEN_DEBUG_CONV_FFT=0
+        export MIOPEN_DEBUG_CONV_DIRECT=0
+        export MIOPEN_DEBUG_CONV_GEMM=1
+        export MIOPEN_DEBUG_CONV_WINOGRAD=0
+        export MIOPEN_DEBUG_CONV_IMPLICIT_GEMM=0
+    fi
+done
+
+# data path
+DATA_PATH=${data_path}
+TOKENIZER_MODEL_PATH=${tokenizer_path}
+CHECKPOINT_PATH=${checkpoint_path}
+
+# default env
+DIST_URL=${1}
+DIST_PORT=${2}
+RANK=$OMPI_COMM_WORLD_RANK
+LOCAL_RANK=$OMPI_COMM_WORLD_LOCAL_RANK
+WORLD_SIZE=$OMPI_COMM_WORLD_SIZE
+CURRENT_DIR="$( cd "$( dirname "$0" )" && pwd )"
+MEGATRON_PATH=$( dirname $( dirname ${CURRENT_DIR}))
+export GLOG_minloglevel=3
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+export HSA_FORCE_FINE_GRAIN_PCIE=1
+export OMP_NUM_THREADS=1
+export GPU_MAX_HW_QUEUES=10
+export PYTHONPATH=${MEGATRON_PATH}/Megatron-LM:$PYTHONPATH
+
+# 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
+
+DISTRIBUTED_ARGS=(
+    --rank ${RANK}
+    --world-size ${WORLD_SIZE}
+    --local-rank ${LOCAL_RANK}
+    --dist-url tcp://${DIST_URL}:${DIST_PORT}
+)
+
+GPT_MODEL_ARGS=(
+    --seq-length 4096
+    --num-layers 40
+    --hidden-size 5120
+    --ffn-hidden-size 13824 
+    --num-attention-heads 40
+    --max-position-embeddings 4096
+    --normalization RMSNorm
+    --position-embedding-type rope
+    --untie-embeddings-and-output-weights
+)
+
+TRAINING_ARGS=(
+    --transformer-impl local
+    --use-legacy-models 
+    --micro-batch-size 1
+    --global-batch-size 256
+    --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
+    --disable-bias-linear
+    --attention-dropout 0
+    --hidden-dropout 0
+    --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
+    --overlap-grad-reduce
+    --use-flash-attn
+)
+
+MODEL_PARALLEL_ARGS=(
+    --tensor-model-parallel-size 2
+    --pipeline-model-parallel-size 2
+    --context-parallel-size 1
+    --use-distributed-optimizer 
+    --sequence-parallel
+)
+
+DATA_ARGS=(
+    --tokenizer-type Llama2Tokenizer
+    --tokenizer-model ${TOKENIZER_MODEL_PATH}
+    --data-path ${DATA_PATH} 
+    --split 949,50,1
+)
+
+EVAL_AND_LOGGING_ARGS=(
+    --log-throughput
+    --eval-iters 5
+    --log-interval 1
+    --save-interval 1000 
+    --eval-interval 1000 
+    --save $CHECKPOINT_PATH
+    --load $CHECKPOINT_PATH
+    --tensorboard-dir "${CHECKPOINT_PATH}/tensorboard" 
+)
+
+TORCH_PROFIE_ARGS=(
+    --profile
+    --profile-ranks 0 1 2 3 4 5 6 7
+    --profile-step-start 3
+    --profile-step-end 4
+    --profile-dir torch_prof_llama_1nodes_tp1-pp2-cp1
+    --use-pytorch-profiler
+)
+
+HIP_PROFIE_ARGS=(
+    --profile
+    --profile-ranks 0 1 2 3 4 5 6 7
+    --profile-step-start 4
+    --profile-step-end 5
+    --use-hip-profiler
+)
+
+APP="python -u ${MEGATRON_PATH}/pretrain_gpt.py \
+    ${GPT_MODEL_ARGS[@]} \
+    ${TRAINING_ARGS[@]} \
+    ${MODEL_PARALLEL_ARGS[@]} \
+    ${DATA_ARGS[@]} \
+    ${EVAL_AND_LOGGING_ARGS[@]} \
+    ${DISTRIBUTED_ARGS[@]} \
+    ${INITIALIZATION_ARGS[@]} \
+    "
+
+if [[ $profiling == "torch" ]]; then
+    APP+=" ${TORCH_PROFIE_ARGS[@]}"
+elif [[ $profiling == "hip" ]]; then
+    mkdir -p hip_prof_data
+    APP+=" ${HIP_PROFIE_ARGS[@]}"
+    APP="hipprof -d hip_prof_data --hip-trace --trace-off ${APP}"
+fi
+
+#for hygon cpu
+export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+${APP}
\ No newline at end of file

examples/llama/train_llama2_7b_1nodes.sh

@@ -72,8 +72,8 @@ GPT_MODEL_ARGS=(
 )
 
 TRAINING_ARGS=(
-    --transformer-impl local
-    --use-legacy-models 
+    --transformer-impl transformer_engine
+    --use-mcore-models 
     --micro-batch-size 1
     --global-batch-size 64
     --train-iters 50
@@ -159,5 +159,5 @@ elif [[ $profiling == "hip" ]]; then
 fi
 
 #for hygon cpu
-export CUDA_VISIBLE_DEVICES=4,5,6,7
+export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
 ${APP}
\ No newline at end of file

examples/llama/train_llama2_7b_2nodes.sh

+#!/bin/bash
+
+INITIALIZATION_ARGS=( --num-workers 2)
+
+for para in $*
+do
+    if [[ $para == --data_path* ]];then
+        data_path=${para#*=}
+    elif [[ $para == --tokenizer_path* ]];then
+        tokenizer_path=${para#*=}
+    elif [[ $para == --checkpoint_path* ]];then
+        checkpoint_path=${para#*=}
+    elif [[ $para == --profiling* ]];then
+        profiling=${para#*=}
+    elif [[ $para == --reproduce* ]];then
+        INITIALIZATION_ARGS=( --reproduce --num-workers 0)
+        export MIOPEN_DEBUG_CONVOLUTION_DETERMINISTIC=1  # miopen 确定算法打开
+        export ROCBLAS_ATOMICS_MOD=0                     # rocblas 关闭原子操作
+        # 关闭miopen中的atomic操作算法, 只保留gemm算法
+        export MIOPEN_DEBUG_CONV_FFT=0
+        export MIOPEN_DEBUG_CONV_DIRECT=0
+        export MIOPEN_DEBUG_CONV_GEMM=1
+        export MIOPEN_DEBUG_CONV_WINOGRAD=0
+        export MIOPEN_DEBUG_CONV_IMPLICIT_GEMM=0
+    fi
+done
+
+# data path
+DATA_PATH=${data_path}
+TOKENIZER_MODEL_PATH=${tokenizer_path}
+CHECKPOINT_PATH=${checkpoint_path}
+
+# default env
+DIST_URL=${1}
+DIST_PORT=${2}
+RANK=$OMPI_COMM_WORLD_RANK
+LOCAL_RANK=$OMPI_COMM_WORLD_LOCAL_RANK
+WORLD_SIZE=$OMPI_COMM_WORLD_SIZE
+CURRENT_DIR="$( cd "$( dirname "$0" )" && pwd )"
+MEGATRON_PATH=$( dirname $( dirname ${CURRENT_DIR}))
+export GLOG_minloglevel=3
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+export HSA_FORCE_FINE_GRAIN_PCIE=1
+export OMP_NUM_THREADS=1
+export GPU_MAX_HW_QUEUES=10
+export PYTHONPATH=${MEGATRON_PATH}/Megatron-LM:$PYTHONPATH
+
+# 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
+
+DISTRIBUTED_ARGS=(
+    --rank ${RANK}
+    --world-size ${WORLD_SIZE}
+    --local-rank ${LOCAL_RANK}
+    --dist-url tcp://${DIST_URL}:${DIST_PORT}
+)
+
+GPT_MODEL_ARGS=(
+    --seq-length 4096
+    --num-layers 32
+    --hidden-size 4096
+    --ffn-hidden-size 11008 
+    --num-attention-heads 32
+    --max-position-embeddings 4096
+    --normalization RMSNorm
+    --position-embedding-type rope
+    --untie-embeddings-and-output-weights
+)
+
+TRAINING_ARGS=(
+    --transformer-impl local
+    --use-legacy-models 
+    --micro-batch-size 1
+    --global-batch-size 256
+    --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
+    --disable-bias-linear
+    --attention-dropout 0
+    --hidden-dropout 0
+    --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
+    --overlap-grad-reduce
+    --use-flash-attn
+)
+
+MODEL_PARALLEL_ARGS=(
+    --tensor-model-parallel-size 1
+    --pipeline-model-parallel-size 2
+    --context-parallel-size 1
+    --use-distributed-optimizer 
+    --sequence-parallel
+)
+
+DATA_ARGS=(
+    --tokenizer-type Llama2Tokenizer
+    --tokenizer-model ${TOKENIZER_MODEL_PATH}
+    --data-path ${DATA_PATH} 
+    --split 949,50,1
+)
+
+EVAL_AND_LOGGING_ARGS=(
+    --log-throughput
+    --eval-iters 5
+    --log-interval 1
+    --save-interval 1000 
+    --eval-interval 1000 
+    --save $CHECKPOINT_PATH
+    --load $CHECKPOINT_PATH
+    --tensorboard-dir "${CHECKPOINT_PATH}/tensorboard" 
+)
+
+TORCH_PROFIE_ARGS=(
+    --profile
+    --profile-ranks 0 1 2 3 4 5 6 7
+    --profile-step-start 3
+    --profile-step-end 4
+    --profile-dir torch_prof_llama_1nodes_tp1-pp2-cp1
+    --use-pytorch-profiler
+)
+
+HIP_PROFIE_ARGS=(
+    --profile
+    --profile-ranks 0 1 2 3 4 5 6 7
+    --profile-step-start 4
+    --profile-step-end 5
+    --use-hip-profiler
+)
+
+APP="python -u ${MEGATRON_PATH}/pretrain_gpt.py \
+    ${GPT_MODEL_ARGS[@]} \
+    ${TRAINING_ARGS[@]} \
+    ${MODEL_PARALLEL_ARGS[@]} \
+    ${DATA_ARGS[@]} \
+    ${EVAL_AND_LOGGING_ARGS[@]} \
+    ${DISTRIBUTED_ARGS[@]} \
+    ${INITIALIZATION_ARGS[@]} \
+    "
+
+if [[ $profiling == "torch" ]]; then
+    APP+=" ${TORCH_PROFIE_ARGS[@]}"
+elif [[ $profiling == "hip" ]]; then
+    mkdir -p hip_prof_data
+    APP+=" ${HIP_PROFIE_ARGS[@]}"
+    APP="hipprof -d hip_prof_data --hip-trace --trace-off ${APP}"
+fi
+
+#for hygon cpu
+export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+${APP}
\ No newline at end of file

examples/qwen/hostfile

+a121 slots=8
+a124 slots=8
\ No newline at end of file

examples/qwen/run_qwen1.5_14B.sh

+for para in $*
+do
+    if [[ $para == --profiling* ]];then
+        profiling=${para#*=}
+    fi
+done
+
+# Those variables need to modify
+GPUS="16"                 # how many gpus to use
+HOST="a121"                 # hostname
+PORT="11452"                 # port id
+DATA_PATH="/data/datasets/oscar-1GB_head-qwen_text_document"            # path to oscar-1GB_head-llama2_text_document
+TOKENIZER_MODEL_PATH="/data/models/qwen1.5" # path to tokenizer.model
+CHECKPOINT_PATH="./ckpt"      # path to ckpt
+
+# Runs Llama2 7B model
+mpirun -np ${GPUS}  --hostfile hostfile \
+                    --allow-run-as-root \
+                    --bind-to none \
+                    --mca plm_rsh_no_tree_spawn 1 \
+                    --mca plm_rsh_args "-p 11451" \
+                    bash -c "
+                    ./train_qwen1.5_14b_2nodes.sh \
+                    ${HOST} \
+                    ${PORT} \
+                    --data_path=$DATA_PATH \
+                    --tokenizer_path=$TOKENIZER_MODEL_PATH \
+                    --checkpoint_path=$CHECKPOINT_PATH \
+                    --profiling=$profiling" > ./logs/log-$((${GPUS} / 8))nodes-`date +%F-%H%M`.log 2>&1
+
+wait
\ No newline at end of file

examples/qwen/run_qwen1.5_32B.sh

+for para in $*
+do
+    if [[ $para == --profiling* ]];then
+        profiling=${para#*=}
+    fi
+done
+
+# Those variables need to modify
+GPUS="16"                 # how many gpus to use
+HOST="a121"                 # hostname
+PORT="11452"                 # port id
+DATA_PATH="/data/datasets/oscar-1GB_head-qwen_text_document"            # path to oscar-1GB_head-llama2_text_document
+TOKENIZER_MODEL_PATH="/data/models/qwen1.5" # path to tokenizer.model
+CHECKPOINT_PATH="./ckpt"      # path to ckpt
+
+# Runs Llama2 7B model
+mpirun -np ${GPUS}  --hostfile hostfile \
+                    --allow-run-as-root \
+                    --bind-to none \
+                    --mca plm_rsh_no_tree_spawn 1 \
+                    --mca plm_rsh_args "-p 11451" \
+                    bash -c "
+                    ./train_qwen1.5_14b_1nodes.sh \
+                    ${HOST} \
+                    ${PORT} \
+                    --data_path=$DATA_PATH \
+                    --tokenizer_path=$TOKENIZER_MODEL_PATH \
+                    --checkpoint_path=$CHECKPOINT_PATH \
+                    --profiling=$profiling" > ./logs/log-$((${GPUS} / 8))nodes-`date +%F-%H%M`.log 2>&1
+
+wait
\ No newline at end of file

examples/qwen/train_qwen1.5_14b_1nodes.sh

+#!/bin/bash
+
+INITIALIZATION_ARGS=( --num-workers 2)
+
+for para in $*
+do
+    if [[ $para == --data_path* ]];then
+        data_path=${para#*=}
+    elif [[ $para == --tokenizer_path* ]];then
+        tokenizer_path=${para#*=}
+    elif [[ $para == --checkpoint_path* ]];then
+        checkpoint_path=${para#*=}
+    elif [[ $para == --profiling* ]];then
+        profiling=${para#*=}
+    elif [[ $para == --reproduce* ]];then
+        INITIALIZATION_ARGS=( --reproduce --num-workers 0)
+        export MIOPEN_DEBUG_CONVOLUTION_DETERMINISTIC=1  # miopen 确定算法打开
+        export ROCBLAS_ATOMICS_MOD=0                     # rocblas 关闭原子操作
+        # 关闭miopen中的atomic操作算法, 只保留gemm算法
+        export MIOPEN_DEBUG_CONV_FFT=0
+        export MIOPEN_DEBUG_CONV_DIRECT=0
+        export MIOPEN_DEBUG_CONV_GEMM=1
+        export MIOPEN_DEBUG_CONV_WINOGRAD=0
+        export MIOPEN_DEBUG_CONV_IMPLICIT_GEMM=0
+    fi
+done
+
+# data path
+DATA_PATH=${data_path}
+TOKENIZER_MODEL_PATH=${tokenizer_path}
+CHECKPOINT_PATH=${checkpoint_path}
+
+# default env
+DIST_URL=${1}
+DIST_PORT=${2}
+RANK=$OMPI_COMM_WORLD_RANK
+LOCAL_RANK=$OMPI_COMM_WORLD_LOCAL_RANK
+WORLD_SIZE=$OMPI_COMM_WORLD_SIZE
+CURRENT_DIR="$( cd "$( dirname "$0" )" && pwd )"
+MEGATRON_PATH=$( dirname $( dirname ${CURRENT_DIR}))
+export GLOG_minloglevel=3
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+export HSA_FORCE_FINE_GRAIN_PCIE=1
+export OMP_NUM_THREADS=1
+export GPU_MAX_HW_QUEUES=10
+export PYTHONPATH=${MEGATRON_PATH}/Megatron-LM:$PYTHONPATH
+
+# 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
+
+DISTRIBUTED_ARGS=(
+    --rank ${RANK}
+    --world-size ${WORLD_SIZE}
+    --local-rank ${LOCAL_RANK}
+    --dist-url tcp://${DIST_URL}:${DIST_PORT}
+)
+
+GPT_MODEL_ARGS=(
+    --seq-length 4096
+    --num-layers 40
+    --hidden-size 5120
+    --ffn-hidden-size 13696 
+    --num-attention-heads 40
+    --max-position-embeddings 32768
+    --normalization RMSNorm
+    --position-embedding-type rope
+    --untie-embeddings-and-output-weights
+)
+
+TRAINING_ARGS=(
+    --transformer-impl transformer_engine
+    --use-mcore-models 
+    --micro-batch-size 1
+    --global-batch-size 256
+    --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
+    --disable-bias-linear
+    --attention-dropout 0
+    --hidden-dropout 0
+    --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
+    --overlap-grad-reduce
+    --use-flash-attn
+)
+
+MODEL_PARALLEL_ARGS=(
+    --tensor-model-parallel-size 2
+    --pipeline-model-parallel-size 4
+    --context-parallel-size 1
+    --use-distributed-optimizer 
+    --sequence-parallel
+)
+
+DATA_ARGS=(
+    --tokenizer-type QwenTokenizer
+    --merge-file ${TOKENIZER_MODEL_PATH}/merges.txt
+    --vocab-file ${TOKENIZER_MODEL_PATH}/vocab.json
+    --data-path ${DATA_PATH} 
+    --split 949,50,1
+)
+
+EVAL_AND_LOGGING_ARGS=(
+    --log-throughput
+    --eval-iters 5
+    --log-interval 1
+    --save-interval 1000 
+    --eval-interval 1000 
+    --save $CHECKPOINT_PATH
+    --load $CHECKPOINT_PATH
+    --tensorboard-dir "${CHECKPOINT_PATH}/tensorboard" 
+)
+
+TORCH_PROFIE_ARGS=(
+    --profile
+    --profile-ranks 0 1 2 3 4 5 6 7
+    --profile-step-start 3
+    --profile-step-end 4
+    --profile-dir torch_prof_llama_1nodes_tp1-pp2-cp1
+    --use-pytorch-profiler
+)
+
+HIP_PROFIE_ARGS=(
+    --profile
+    --profile-ranks 0 1 2 3 4 5 6 7
+    --profile-step-start 4
+    --profile-step-end 5
+    --use-hip-profiler
+)
+
+APP="python -u ${MEGATRON_PATH}/pretrain_gpt.py \
+    ${GPT_MODEL_ARGS[@]} \
+    ${TRAINING_ARGS[@]} \
+    ${MODEL_PARALLEL_ARGS[@]} \
+    ${DATA_ARGS[@]} \
+    ${EVAL_AND_LOGGING_ARGS[@]} \
+    ${DISTRIBUTED_ARGS[@]} \
+    ${INITIALIZATION_ARGS[@]} \
+    "
+
+if [[ $profiling == "torch" ]]; then
+    APP+=" ${TORCH_PROFIE_ARGS[@]}"
+elif [[ $profiling == "hip" ]]; then
+    mkdir -p hip_prof_data
+    APP+=" ${HIP_PROFIE_ARGS[@]}"
+    APP="hipprof -d hip_prof_data --hip-trace --trace-off ${APP}"
+fi
+
+#for hygon cpu
+export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+${APP}
\ No newline at end of file

examples/qwen/train_qwen1.5_14b_2nodes.sh

+#!/bin/bash
+
+INITIALIZATION_ARGS=( --num-workers 2)
+
+for para in $*
+do
+    if [[ $para == --data_path* ]];then
+        data_path=${para#*=}
+    elif [[ $para == --tokenizer_path* ]];then
+        tokenizer_path=${para#*=}
+    elif [[ $para == --checkpoint_path* ]];then
+        checkpoint_path=${para#*=}
+    elif [[ $para == --profiling* ]];then
+        profiling=${para#*=}
+    elif [[ $para == --reproduce* ]];then
+        INITIALIZATION_ARGS=( --reproduce --num-workers 0)
+        export MIOPEN_DEBUG_CONVOLUTION_DETERMINISTIC=1  # miopen 确定算法打开
+        export ROCBLAS_ATOMICS_MOD=0                     # rocblas 关闭原子操作
+        # 关闭miopen中的atomic操作算法, 只保留gemm算法
+        export MIOPEN_DEBUG_CONV_FFT=0
+        export MIOPEN_DEBUG_CONV_DIRECT=0
+        export MIOPEN_DEBUG_CONV_GEMM=1
+        export MIOPEN_DEBUG_CONV_WINOGRAD=0
+        export MIOPEN_DEBUG_CONV_IMPLICIT_GEMM=0
+    fi
+done
+
+# data path
+DATA_PATH=${data_path}
+TOKENIZER_MODEL_PATH=${tokenizer_path}
+CHECKPOINT_PATH=${checkpoint_path}
+
+# default env
+DIST_URL=${1}
+DIST_PORT=${2}
+RANK=$OMPI_COMM_WORLD_RANK
+LOCAL_RANK=$OMPI_COMM_WORLD_LOCAL_RANK
+WORLD_SIZE=$OMPI_COMM_WORLD_SIZE
+CURRENT_DIR="$( cd "$( dirname "$0" )" && pwd )"
+MEGATRON_PATH=$( dirname $( dirname ${CURRENT_DIR}))
+export GLOG_minloglevel=3
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+export HSA_FORCE_FINE_GRAIN_PCIE=1
+export OMP_NUM_THREADS=1
+export GPU_MAX_HW_QUEUES=10
+export PYTHONPATH=${MEGATRON_PATH}/Megatron-LM:$PYTHONPATH
+
+# 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
+
+DISTRIBUTED_ARGS=(
+    --rank ${RANK}
+    --world-size ${WORLD_SIZE}
+    --local-rank ${LOCAL_RANK}
+    --dist-url tcp://${DIST_URL}:${DIST_PORT}
+)
+
+GPT_MODEL_ARGS=(
+    --seq-length 4096
+    --num-layers 40
+    --hidden-size 5120
+    --ffn-hidden-size 13696 
+    --num-attention-heads 40
+    --max-position-embeddings 32768
+    --normalization RMSNorm
+    --position-embedding-type rope
+    --untie-embeddings-and-output-weights
+)
+
+TRAINING_ARGS=(
+    --transformer-impl transformer_engine
+    --use-mcore-models 
+    --micro-batch-size 1
+    --global-batch-size 256
+    --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
+    --disable-bias-linear
+    --attention-dropout 0
+    --hidden-dropout 0
+    --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
+    --overlap-grad-reduce
+    --use-flash-attn
+)
+
+MODEL_PARALLEL_ARGS=(
+    --tensor-model-parallel-size 2
+    --pipeline-model-parallel-size 2
+    --context-parallel-size 1
+    --use-distributed-optimizer 
+    --sequence-parallel
+)
+
+DATA_ARGS=(
+    --tokenizer-type QwenTokenizer
+    --merge-file ${TOKENIZER_MODEL_PATH}/merges.txt
+    --vocab-file ${TOKENIZER_MODEL_PATH}/vocab.json
+    --data-path ${DATA_PATH} 
+    --split 949,50,1
+)
+
+EVAL_AND_LOGGING_ARGS=(
+    --log-throughput
+    --eval-iters 5
+    --log-interval 1
+    --save-interval 1000 
+    --eval-interval 1000 
+    --save $CHECKPOINT_PATH
+    --load $CHECKPOINT_PATH
+    --tensorboard-dir "${CHECKPOINT_PATH}/tensorboard" 
+)
+
+TORCH_PROFIE_ARGS=(
+    --profile
+    --profile-ranks 0 1 2 3 4 5 6 7
+    --profile-step-start 3
+    --profile-step-end 4
+    --profile-dir torch_prof_llama_1nodes_tp1-pp2-cp1
+    --use-pytorch-profiler
+)
+
+HIP_PROFIE_ARGS=(
+    --profile
+    --profile-ranks 0 1 2 3 4 5 6 7
+    --profile-step-start 4
+    --profile-step-end 5
+    --use-hip-profiler
+)
+
+APP="python -u ${MEGATRON_PATH}/pretrain_gpt.py \
+    ${GPT_MODEL_ARGS[@]} \
+    ${TRAINING_ARGS[@]} \
+    ${MODEL_PARALLEL_ARGS[@]} \
+    ${DATA_ARGS[@]} \
+    ${EVAL_AND_LOGGING_ARGS[@]} \
+    ${DISTRIBUTED_ARGS[@]} \
+    ${INITIALIZATION_ARGS[@]} \
+    "
+
+if [[ $profiling == "torch" ]]; then
+    APP+=" ${TORCH_PROFIE_ARGS[@]}"
+elif [[ $profiling == "hip" ]]; then
+    mkdir -p hip_prof_data
+    APP+=" ${HIP_PROFIE_ARGS[@]}"
+    APP="hipprof -d hip_prof_data --hip-trace --trace-off ${APP}"
+fi
+
+#for hygon cpu
+export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+${APP}
\ No newline at end of file

examples/qwen/train_qwen1.5_32b_2nodes.sh

+#!/bin/bash
+
+INITIALIZATION_ARGS=( --num-workers 2)
+
+for para in $*
+do
+    if [[ $para == --data_path* ]];then
+        data_path=${para#*=}
+    elif [[ $para == --tokenizer_path* ]];then
+        tokenizer_path=${para#*=}
+    elif [[ $para == --checkpoint_path* ]];then
+        checkpoint_path=${para#*=}
+    elif [[ $para == --profiling* ]];then
+        profiling=${para#*=}
+    elif [[ $para == --reproduce* ]];then
+        INITIALIZATION_ARGS=( --reproduce --num-workers 0)
+        export MIOPEN_DEBUG_CONVOLUTION_DETERMINISTIC=1  # miopen 确定算法打开
+        export ROCBLAS_ATOMICS_MOD=0                     # rocblas 关闭原子操作
+        # 关闭miopen中的atomic操作算法, 只保留gemm算法
+        export MIOPEN_DEBUG_CONV_FFT=0
+        export MIOPEN_DEBUG_CONV_DIRECT=0
+        export MIOPEN_DEBUG_CONV_GEMM=1
+        export MIOPEN_DEBUG_CONV_WINOGRAD=0
+        export MIOPEN_DEBUG_CONV_IMPLICIT_GEMM=0
+    fi
+done
+
+# data path
+DATA_PATH=${data_path}
+TOKENIZER_MODEL_PATH=${tokenizer_path}
+CHECKPOINT_PATH=${checkpoint_path}
+
+# default env
+DIST_URL=${1}
+DIST_PORT=${2}
+RANK=$OMPI_COMM_WORLD_RANK
+LOCAL_RANK=$OMPI_COMM_WORLD_LOCAL_RANK
+WORLD_SIZE=$OMPI_COMM_WORLD_SIZE
+CURRENT_DIR="$( cd "$( dirname "$0" )" && pwd )"
+MEGATRON_PATH=$( dirname $( dirname ${CURRENT_DIR}))
+export GLOG_minloglevel=3
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+export HSA_FORCE_FINE_GRAIN_PCIE=1
+export OMP_NUM_THREADS=1
+export GPU_MAX_HW_QUEUES=10
+export PYTHONPATH=${MEGATRON_PATH}/Megatron-LM:$PYTHONPATH
+
+# 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
+
+DISTRIBUTED_ARGS=(
+    --rank ${RANK}
+    --world-size ${WORLD_SIZE}
+    --local-rank ${LOCAL_RANK}
+    --dist-url tcp://${DIST_URL}:${DIST_PORT}
+)
+
+GPT_MODEL_ARGS=(
+    --seq-length 4096
+    --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
+    --untie-embeddings-and-output-weights
+)
+
+TRAINING_ARGS=(
+    --transformer-impl transformer_engine
+    --use-mcore-models 
+    --micro-batch-size 1
+    --global-batch-size 1024
+    --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
+    --disable-bias-linear
+    --attention-dropout 0
+    --hidden-dropout 0
+    --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
+    --overlap-grad-reduce
+    --use-flash-attn
+)
+
+MODEL_PARALLEL_ARGS=(
+    --tensor-model-parallel-size 4
+    --pipeline-model-parallel-size 4
+    --context-parallel-size 1
+    --use-distributed-optimizer 
+    --sequence-parallel
+)
+
+DATA_ARGS=(
+    --tokenizer-type QwenTokenizer
+    --merge-file ${TOKENIZER_MODEL_PATH}/merges.txt
+    --vocab-file ${TOKENIZER_MODEL_PATH}/vocab.json
+    --data-path ${DATA_PATH} 
+    --split 949,50,1
+)
+
+EVAL_AND_LOGGING_ARGS=(
+    --log-throughput
+    --eval-iters 5
+    --log-interval 1
+    --save-interval 1000 
+    --eval-interval 1000 
+    --save $CHECKPOINT_PATH
+    --load $CHECKPOINT_PATH
+    --tensorboard-dir "${CHECKPOINT_PATH}/tensorboard" 
+)
+
+TORCH_PROFIE_ARGS=(
+    --profile
+    --profile-ranks 0 1 2 3 4 5 6 7
+    --profile-step-start 3
+    --profile-step-end 4
+    --profile-dir torch_prof_llama_1nodes_tp1-pp2-cp1
+    --use-pytorch-profiler
+)
+
+HIP_PROFIE_ARGS=(
+    --profile
+    --profile-ranks 0 1 2 3 4 5 6 7
+    --profile-step-start 4
+    --profile-step-end 5
+    --use-hip-profiler
+)
+
+APP="python -u ${MEGATRON_PATH}/pretrain_gpt.py \
+    ${GPT_MODEL_ARGS[@]} \
+    ${TRAINING_ARGS[@]} \
+    ${MODEL_PARALLEL_ARGS[@]} \
+    ${DATA_ARGS[@]} \
+    ${EVAL_AND_LOGGING_ARGS[@]} \
+    ${DISTRIBUTED_ARGS[@]} \
+    ${INITIALIZATION_ARGS[@]} \
+    "
+
+if [[ $profiling == "torch" ]]; then
+    APP+=" ${TORCH_PROFIE_ARGS[@]}"
+elif [[ $profiling == "hip" ]]; then
+    mkdir -p hip_prof_data
+    APP+=" ${HIP_PROFIE_ARGS[@]}"
+    APP="hipprof -d hip_prof_data --hip-trace --trace-off ${APP}"
+fi
+
+#for hygon cpu
+export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+${APP}
\ No newline at end of file