添加legacy的优化项和llama示例脚本

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-0227/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
+    --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
\ No newline at end of file

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

megatron/legacy/model/transformer.py

@@ -712,7 +712,7 @@ class ParallelAttention(MegatronModule):
                 dim=3)
 
             # [sq, b, ng, np/ng * hn] -> [sq, b, np, hn] -
-            query_layer = query_layer.view(query_layer.size(0), query_layer.size(1), -1, self.hidden_size_per_attention_head)
+            query_layer = query_layer.contiguous().view(query_layer.size(0), query_layer.size(1), -1, self.hidden_size_per_attention_head)
         else:
             # Attention heads [sk, b, h] --> [sk, b, (np * 2 * hn)]
             mixed_kv_layer, _ = self.key_value(encoder_output)

megatron/training/arguments.py

@@ -2026,6 +2026,8 @@ def _add_tokenizer_args(parser):
                                 'GPTSentencePieceTokenizer',
                                 'HuggingFaceTokenizer',
                                 'Llama2Tokenizer',
+                                'Llama3Tokenizer',
+                                'QwenTokenizer',
                                 'TikTokenizer',
                                 'MultimodalTokenizer',
                                 'NullTokenizer'],

megatron/training/tokenizer/tokenizer.py

@@ -15,6 +15,7 @@ from megatron.core.datasets.megatron_tokenizer import MegatronTokenizer
 from .bert_tokenization import FullTokenizer as FullBertTokenizer
 from .gpt2_tokenization import GPT2Tokenizer
 from megatron.training.tokenizer.multimodal_tokenizer import MultimodalTokenizer
+from transformers import Qwen2Tokenizer
 
 
 def build_tokenizer(args, **kwargs):
@@ -50,6 +51,11 @@ def build_tokenizer(args, **kwargs):
     elif args.tokenizer_type == 'Llama2Tokenizer':
         assert args.tokenizer_model is not None
         tokenizer = _Llama2Tokenizer(args.tokenizer_model)
+    elif args.tokenizer_type == 'Llama3Tokenizer':
+        assert args.tokenizer_model is not None
+        tokenizer = _Llama3Tokenizer(args.tokenizer_model)
+    elif args.tokenizer_type == 'QwenTokenizer':
+        tokenizer = _Qwen2Tokenizer(args.vocab_file, args.merge_file)
     elif args.tokenizer_type == 'TikTokenizer':
         assert args.tokenizer_model is not None
         assert args.tiktoken_pattern is not None
@@ -605,6 +611,94 @@ class _Llama2Tokenizer(_SentencePieceTokenizer):
     def additional_special_tokens_ids(self):
         return None
 
+class _Llama3Tokenizer(MegatronTokenizer):
+    """tiktokenTokenizer-Megatron llama3 改写"""
+    # https://github.com/meta-llama/llama3/blob/main/llama/tokenizer.py
+
+    def __init__(self, model_file):
+        super().__init__(model_file)
+        from pathlib import Path
+        import tiktoken
+        from tiktoken.load import load_tiktoken_bpe
+        tokenizer_path=model_file
+        special_tokens = [
+            "<|begin_of_text|>",
+            "<|end_of_text|>",
+            "<|reserved_special_token_0|>",
+            "<|reserved_special_token_1|>",
+            "<|reserved_special_token_2|>",
+            "<|reserved_special_token_3|>",
+            "<|start_header_id|>",
+            "<|end_header_id|>",
+            "<|reserved_special_token_4|>",
+            "<|eot_id|>",  # end of turn
+            ] + [f"<|reserved_special_token_{i}|>" for i in range (5, 256 - 5)] 
+        mergeable_ranks = load_tiktoken_bpe(tokenizer_path)
+        self.tokenizer = tiktoken.Encoding(tokenizer_path,
+            pat_str = r"(?i:'s|'t|'re|'ve|'m|'ll|'d)|[^\r\n\p{L}\p{N}]?\p{L}+|\p{N}{1,3}| ?[^\s\p{L}\p{N}]+[\r\n]*|\s*[\r\n]+|\s+(?!\S)|\s+",
+            mergeable_ranks=mergeable_ranks,
+            special_tokens={token: len (mergeable_ranks) + i for i, token in enumerate (special_tokens)},
+            )
+
+        self.eod_id = self.tokenizer.encode("<|end_of_text|>", allowed_special="all")[0]
+    @property
+    def vocab_size(self):
+        return self.tokenizer.n_vocab
+
+    @property
+    def vocab(self):
+        return self.tokenizer.encode
+
+    @property
+    def inv_vocab(self):
+        return self.tokenizer.encode
+
+    def tokenize(self, text):
+        return self.tokenizer.encode(text)
+
+    def detokenize(self, token_ids):
+        return self.tokenizer.encode(token_ids)
+
+    @property
+    def eod(self):
+        return self.eod_id
+
+class _Qwen2Tokenizer(MegatronTokenizer):
+    def __init__(self, vocab_file, merge_file,extra_vocab_size=0):
+        super().__init__(vocab_file, merge_file)
+        self.tokenizer = Qwen2Tokenizer(vocab_file, merge_file)
+        self.extra_vocab_size = extra_vocab_size
+        self.tokenizer.add_special_tokens(special_tokens_dict=dict(pad_token="<|extra_0|>"))
+
+    @property
+    def vocab_size(self):
+        return len(self.tokenizer.encoder) + self.extra_vocab_size
+
+    @property
+    def vocab(self):
+        return self.tokenizer.encoder
+
+    @property
+    def inv_vocab(self):
+        return self.tokenizer.decoder
+
+    def tokenize(self, text):
+        return self.tokenizer.encode(text)
+
+    def detokenize(self, token_ids):
+        return self.tokenizer.decode(token_ids)
+
+    @property
+    def eod(self):
+        return self.tokenizer.eos_token_id
+
+    @property
+    def eos_token(self):
+        return self.tokenizer.eos_token
+
+    @property
+    def pad_token_id(self):
+        return self.tokenizer.pad_token_id
 
 def reload_mergeable_ranks(path: str, max_vocab: Optional[int] = None) -> Dict[bytes, int]:
     """