!1759 新增Qwen2-7B/Qwen1.5-4B mcore全参微调脚本

Merge pull request !1759 from 徐源徽/master

examples/mcore/qwen15/chat_qwen15_4b_ptd.sh

@@ -0,0 +1,59 @@
+#!/bin/bash
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+
+# please fill these path configurations
+CHECKPOINT="your model ckpt path"
+TOKENIZER_PATH="your tokenizer path"
+
+
+# Change for multinode config
+MASTER_ADDR=localhost
+MASTER_PORT=6003
+NNODES=1
+NODE_RANK=0
+NPUS_PER_NODE=8
+WORLD_SIZE=$(($NPUS_PER_NODE*$NNODES))
+
+TP=1
+PP=8
+
+DISTRIBUTED_ARGS="
+    --nproc_per_node $NPUS_PER_NODE \
+    --nnodes $NNODES \
+    --node_rank $NODE_RANK \
+    --master_addr $MASTER_ADDR \
+    --master_port $MASTER_PORT
+"
+
+torchrun $DISTRIBUTED_ARGS inference.py \
+       --prompt-type qwen \
+       --use-mcore-models \
+       --tensor-model-parallel-size ${TP} \
+       --pipeline-model-parallel-size ${PP} \
+       --num-layers 40 \
+       --hidden-size 2560 \
+       --num-attention-heads 20 \
+       --ffn-hidden-size 6912 \
+       --max-position-embeddings 8192 \
+       --seq-length 8192 \
+       --make-vocab-size-divisible-by 1 \
+       --untie-embeddings-and-output-weights \
+       --micro-batch-size 1 \
+       --swiglu \
+       --disable-bias-linear \
+       --tokenizer-type PretrainedFromHF \
+       --tokenizer-name-or-path ${TOKENIZER_PATH} \
+       --load ${CHECKPOINT} \
+       --normalization RMSNorm \
+       --position-embedding-type rope \
+       --norm-epsilon 1e-6 \
+       --hidden-dropout 0 \
+       --attention-dropout 0 \
+       --tokenizer-not-use-fast \
+       --add-qkv-bias \
+       --max-new-tokens 256 \
+       --seed 1234 \
+       --bf16 \
+       --rotary-base 1000000.0 \
+       --padded-vocab-size 151936 \
+       | tee logs/generate_mcore_qwen15_4b_full.log

examples/mcore/qwen15/data_convert_qwen15_instruction.sh

@@ -0,0 +1,14 @@
+# 请按照您的真实环境修改 set_env.sh 路径
+source /usr/local/Ascend/ascend-toolkit/set_env.sh
+mkdir ./finetune_dataset
+
+python ./preprocess_data.py \
+   --input ./dataset/train-00000-of-00001-a09b74b3ef9c3b56.parquet \
+   --tokenizer-name-or-path ./model_from_hf/qwen15_hf/ \
+   --output-prefix ./finetune_dataset/alpaca \
+   --handler-name AlpacaStyleInstructionHandler \
+   --tokenizer-type PretrainedFromHF \
+   --workers 4 \
+   --log-interval 1000  \
+   --prompt-type qwen
+   # --map-keys '{"prompt":"instruction","query":"input","response":"output"}' # 默认值，可不传

examples/mcore/qwen15/evaluate_qwen15_4b_full_ptd.sh

@@ -0,0 +1,63 @@
+#!/bin/bash
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+
+# Change for multinode config
+MASTER_ADDR=localhost
+MASTER_PORT=6001
+NNODES=1
+NODE_RANK=0
+NPUS_PER_NODE=8
+WORLD_SIZE=$(($NPUS_PER_NODE*$NNODES))
+
+# please fill these path configurations
+CHECKPOINT="your model ckpt path"
+TOKENIZER_PATH="your tokenizer path"
+DATA_PATH="./mmlu/test"
+TASK="mmlu"
+
+TP=1
+PP=8
+
+DISTRIBUTED_ARGS="
+    --nproc_per_node $NPUS_PER_NODE \
+    --nnodes $NNODES \
+    --node_rank $NODE_RANK \
+    --master_addr $MASTER_ADDR \
+    --master_port $MASTER_PORT
+"
+
+# Different task needs different max_new_tokens value, please follow the instruction in readme.
+torchrun $DISTRIBUTED_ARGS evaluation.py   \
+       --prompt-type qwen \
+       --use-mcore-models \
+       --task-data-path $DATA_PATH \
+       --task ${TASK} \
+       --seq-length 8192 \
+       --max-new-tokens 1 \
+       --max-position-embeddings 8192 \
+       --tensor-model-parallel-size ${TP} \
+       --pipeline-model-parallel-size ${PP} \
+       --num-layers 40  \
+       --hidden-size 2560  \
+       --ffn-hidden-size 6912 \
+       --num-attention-heads 20  \
+       --disable-bias-linear \
+       --swiglu \
+       --position-embedding-type rope \
+       --load $CHECKPOINT \
+       --normalization RMSNorm \
+       --tokenizer-type PretrainedFromHF \
+       --tokenizer-name-or-path ${TOKENIZER_PATH} \
+       --tokenizer-not-use-fast \
+       --bf16 \
+       --micro-batch-size 1  \
+       --exit-on-missing-checkpoint \
+       --no-load-rng \
+       --no-load-optim \
+       --untie-embeddings-and-output-weights \
+       --add-qkv-bias \
+       --make-vocab-size-divisible-by 1 \
+       --seed 1234 \
+       --rotary-base 1000000.0 \
+       --padded-vocab-size 151936 \
+       | tee logs/eval_mcore_qwen15_4b_full_${TASK}.log

examples/mcore/qwen15/tune_qwen15_4b_full_ptd.sh

@@ -0,0 +1,101 @@
+#!/bin/bash
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+
+NPUS_PER_NODE=8
+MASTER_ADDR=localhost
+MASTER_PORT=6005
+NNODES=1
+NODE_RANK=0
+WORLD_SIZE=$(($NPUS_PER_NODE*$NNODES))
+
+# please fill these path configurations
+CKPT_SAVE_DIR="your model save ckpt path"
+DATA_PATH="your data path"
+TOKENIZER_PATH="your tokenizer path"
+CKPT_LOAD_DIR="your model ckpt path"
+
+TP=1
+PP=8
+
+DISTRIBUTED_ARGS="
+    --nproc_per_node $NPUS_PER_NODE \
+    --nnodes $NNODES \
+    --node_rank $NODE_RANK \
+    --master_addr $MASTER_ADDR \
+    --master_port $MASTER_PORT
+"
+
+GPT_ARGS="
+    --finetune \
+    --is-instruction-dataset \
+    --variable-seq-lengths \
+    --tokenizer-not-use-fast \
+    --prompt-type qwen \
+    --use-mcore-models \
+    --tensor-model-parallel-size ${TP} \
+    --pipeline-model-parallel-size ${PP} \
+    --num-layers 40 \
+    --hidden-size 2560 \
+    --ffn-hidden-size 6912 \
+    --num-attention-heads 20 \
+    --tokenizer-type PretrainedFromHF \
+    --load ${CKPT_LOAD_DIR} \
+    --tokenizer-name-or-path ${TOKENIZER_PATH} \
+    --seq-length 8192 \
+    --max-position-embeddings 32768 \
+    --micro-batch-size 1 \
+    --global-batch-size 64 \
+    --make-vocab-size-divisible-by 1 \
+    --lr 1.25e-6 \
+    --min-lr 1.25e-7 \
+    --train-iters 2000 \
+    --lr-decay-style cosine \
+    --untie-embeddings-and-output-weights \
+    --disable-bias-linear \
+    --attention-dropout 0.0 \
+    --init-method-std 0.02 \
+    --hidden-dropout 0.0 \
+    --position-embedding-type rope \
+    --normalization RMSNorm \
+    --swiglu \
+    --use-flash-attn \
+    --use-fused-swiglu \
+    --use-fused-rmsnorm \
+    --use-fused-rotary-pos-emb \
+    --use-distributed-optimizer \
+    --use-rotary-position-embeddings \
+    --no-masked-softmax-fusion \
+    --attention-softmax-in-fp32 \
+    --weight-decay 0.0 \
+    --clip-grad 1.0 \
+    --adam-beta1 0.9 \
+    --adam-beta2 0.999 \
+    --rotary-base 1000000.0 \
+    --add-qkv-bias \
+    --initial-loss-scale 8192 \
+    --no-gradient-accumulation-fusion \
+    --lr-warmup-fraction 0.01 \
+    --no-load-optim \
+    --no-load-rng \
+    --padded-vocab-size 151936 \
+    --bf16 \
+    --seed 1234
+"
+
+DATA_ARGS="
+    --data-path $DATA_PATH \
+    --split 100,0,0
+"
+
+OUTPUT_ARGS="
+    --log-interval 1 \
+    --save-interval 2000 \
+"
+
+torchrun $DISTRIBUTED_ARGS pretrain_gpt.py \
+    $GPT_ARGS \
+    $DATA_ARGS \
+    $OUTPUT_ARGS \
+    --distributed-backend nccl \
+    --save ${CKPT_SAVE_DIR} \
+    | tee logs/tune_mcore_qwen15_4b_full_2000.log

examples/mcore/qwen2/chat_qwen2_7b_ptd.sh

@@ -0,0 +1,64 @@
+#!/bin/bash
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+
+# please fill these path configurations
+TOKENIZER_PATH="your tokenizer directory path"
+CHECKPOINT="your model directory path"
+
+# Change for multinode config
+MASTER_ADDR=localhost
+MASTER_PORT=6014
+NNODES=1
+NODE_RANK=0
+NPUS_PER_NODE=8
+WORLD_SIZE=$(($NPUS_PER_NODE*$NNODES))
+
+TP=1
+PP=8
+SEQ_LENGTH=4096
+
+DISTRIBUTED_ARGS="
+    --nproc_per_node $NPUS_PER_NODE \
+    --nnodes $NNODES \
+    --node_rank $NODE_RANK \
+    --master_addr $MASTER_ADDR \
+    --master_port $MASTER_PORT
+"
+
+torchrun $DISTRIBUTED_ARGS inference.py \
+       --prompt-type qwen \
+       --use-mcore-models \
+       --tensor-model-parallel-size ${TP} \
+       --pipeline-model-parallel-size ${PP} \
+       --load ${CHECKPOINT} \
+       --num-layers 28 \
+       --num-layer-list 4,4,4,4,3,3,3,3 \
+       --hidden-size 3584  \
+       --num-attention-heads 28  \
+       --ffn-hidden-size 18944 \
+       --max-position-embeddings ${SEQ_LENGTH} \
+       --seq-length ${SEQ_LENGTH} \
+       --make-vocab-size-divisible-by 1 \
+       --padded-vocab-size 152064 \
+       --rotary-base 1000000.0 \
+       --untie-embeddings-and-output-weights \
+       --micro-batch-size 1 \
+       --swiglu \
+       --disable-bias-linear \
+       --tokenizer-type PretrainedFromHF  \
+       --tokenizer-name-or-path ${TOKENIZER_PATH} \
+       --normalization RMSNorm \
+       --position-embedding-type rope \
+       --norm-epsilon 1e-6 \
+       --hidden-dropout 0 \
+       --attention-dropout 0 \
+       --tokenizer-not-use-fast \
+       --add-qkv-bias \
+       --max-new-tokens 256 \
+       --no-gradient-accumulation-fusion \
+       --exit-on-missing-checkpoint \
+       --attention-softmax-in-fp32 \
+       --seed 1234 \
+       --group-query-attention \
+       --num-query-groups 4 \
+       | tee logs/generate_mcore_qwen2_7b_full.log

examples/mcore/qwen2/data_convert_qwen2_instruction.sh

@@ -0,0 +1,14 @@
+# 请按照您的真实环境修改 set_env.sh 路径
+source /usr/local/Ascend/ascend-toolkit/set_env.sh
+mkdir ./finetune_dataset
+
+python ./preprocess_data.py \
+	--input ./dataset/train-00000-of-00001-a09b74b3ef9c3b56.parquet \
+	--tokenizer-name-or-path ./model_from_hf/qwen2_hf \
+	--output-prefix ./finetune_dataset/alpaca \
+	--handler-name AlpacaStyleInstructionHandler \
+	--tokenizer-type PretrainedFromHF \
+	--workers 16 \
+	--log-interval 1000 \
+	--prompt-type qwen
+    # --map-keys '{"prompt":"instruction","query":"input","response":"output"}' # 默认值，可不传

examples/mcore/qwen2/evaluate_qwen2_7b_full_ptd.sh

@@ -0,0 +1,70 @@
+#!/bin/bash
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+
+TOKENIZER_PATH="your tokenizer directory path"
+CHECKPOINT="your model directory path"
+
+# configure task and data path
+DATA_PATH="./mmlu/test/"
+TASK="mmlu"
+
+# distributed config
+MASTER_ADDR=localhost
+MASTER_PORT=6014
+NNODES=1
+NODE_RANK=0
+NPUS_PER_NODE=8
+WORLD_SIZE=$(($NPUS_PER_NODE*$NNODES))
+
+TP=1
+PP=8
+SEQ_LENGTH=4096
+
+DISTRIBUTED_ARGS="
+    --nproc_per_node $NPUS_PER_NODE \
+    --nnodes $NNODES \
+    --node_rank $NODE_RANK \
+    --master_addr $MASTER_ADDR \
+    --master_port $MASTER_PORT
+"
+
+
+# Different task needs different max_new_tokens value, please follow the instruction in readme.
+torchrun $DISTRIBUTED_ARGS evaluation.py \
+       --prompt-type qwen \
+       --use-mcore-models \
+       --task-data-path $DATA_PATH \
+       --task ${TASK} \
+       --tensor-model-parallel-size ${TP} \
+       --pipeline-model-parallel-size ${PP} \
+       --seq-length ${SEQ_LENGTH} \
+       --max-position-embeddings ${SEQ_LENGTH} \
+       --max-new-tokens 1 \
+       --num-layers 28  \
+       --num-layer-list 4,4,4,4,3,3,3,3 \
+       --hidden-size 3584  \
+       --ffn-hidden-size 18944 \
+       --num-attention-heads 28  \
+       --disable-bias-linear \
+       --swiglu \
+       --position-embedding-type rope \
+       --load ${CHECKPOINT} \
+       --normalization RMSNorm \
+       --tokenizer-type PretrainedFromHF  \
+       --tokenizer-name-or-path ${TOKENIZER_PATH} \
+       --tokenizer-not-use-fast \
+       --micro-batch-size 1  \
+       --exit-on-missing-checkpoint \
+       --no-load-rng \
+       --no-load-optim \
+       --untie-embeddings-and-output-weights \
+       --add-qkv-bias \
+       --make-vocab-size-divisible-by 1 \
+       --padded-vocab-size 152064 \
+       --rotary-base 1000000.0 \
+       --no-gradient-accumulation-fusion \
+       --attention-softmax-in-fp32 \
+       --seed 1234 \
+       --group-query-attention \
+       --num-query-groups 4 \
+       | tee logs/evaluation_mcore_qwen2_7b_full_${TASK}.log

examples/mcore/qwen2/tune_qwen2_7b_full_ptd.sh

@@ -0,0 +1,104 @@
+#!/bin/bash
+export HCCL_CONNECT_TIMEOUT=1200
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+
+NPUS_PER_NODE=8
+MASTER_ADDR=localhost
+MASTER_PORT=6000
+NNODES=1
+NODE_RANK=0
+WORLD_SIZE=$(($NPUS_PER_NODE*$NNODES))
+
+# please fill these path configurations
+CKPT_LOAD_DIR="your model ckpt path"
+CKPT_SAVE_DIR="your model save ckpt path"
+DATA_PATH="your data path"
+TOKENIZER_PATH="your tokenizer path"
+
+TP=1
+PP=8
+SEQ_LEN=4096
+
+DISTRIBUTED_ARGS="
+    --nproc_per_node $NPUS_PER_NODE \
+    --nnodes $NNODES \
+    --node_rank $NODE_RANK \
+    --master_addr $MASTER_ADDR \
+    --master_port $MASTER_PORT
+"
+    
+GPT_ARGS="
+    --finetune \
+    --is-instruction-dataset \
+    --variable-seq-lengths \
+    --tokenizer-not-use-fast \
+    --prompt-type qwen \
+    --use-mcore-models \
+    --tensor-model-parallel-size ${TP} \
+    --pipeline-model-parallel-size ${PP} \
+    --num-layers 28 \
+    --num-layer-list 4,4,4,4,3,3,3,3 \
+    --hidden-size 3584 \
+    --ffn-hidden-size 18944 \
+    --num-attention-heads 28 \
+    --tokenizer-type PretrainedFromHF \
+    --tokenizer-name-or-path ${TOKENIZER_PATH} \
+    --seq-length ${SEQ_LEN} \
+    --max-position-embeddings ${SEQ_LEN} \
+    --micro-batch-size 1 \
+    --global-batch-size 256 \
+    --make-vocab-size-divisible-by 1 \
+    --padded-vocab-size 152064 \
+    --rotary-base 1000000.0 \
+    --lr 1.25e-6 \
+    --min-lr 1.25e-7 \
+    --train-iters 2000 \
+    --lr-decay-style cosine \
+    --untie-embeddings-and-output-weights \
+    --disable-bias-linear \
+    --attention-dropout 0.0 \
+    --init-method-std 0.02 \
+    --hidden-dropout 0.0 \
+    --position-embedding-type rope \
+    --normalization RMSNorm \
+    --swiglu \
+    --use-flash-attn \
+    --weight-decay 0.0 \
+    --use-rotary-position-embeddings \
+    --no-masked-softmax-fusion \
+    --attention-softmax-in-fp32 \
+    --clip-grad 1.0 \
+    --adam-beta1 0.9 \
+    --adam-beta2 0.999 \
+    --add-qkv-bias \
+    --initial-loss-scale 4096 \
+    --no-gradient-accumulation-fusion \
+    --no-load-optim \
+    --no-load-rng \
+    --seed 1234 \
+    --bf16 \
+    --group-query-attention \
+    --num-query-groups 4 \
+    --norm-epsilon 1e-06 \
+"
+
+DATA_ARGS="
+    --data-path $DATA_PATH \
+    --split 100,0,0
+"
+
+OUTPUT_ARGS="
+    --log-interval 1 \
+    --save-interval 2000 \
+    --eval-interval 2000 \
+    --eval-iters 0 \
+"
+
+torchrun $DISTRIBUTED_ARGS pretrain_gpt.py \
+    $GPT_ARGS \
+    $DATA_ARGS \
+    $OUTPUT_ARGS \
+    --distributed-backend nccl \
+    --load ${CKPT_LOAD_DIR} \
+    --save ${CKPT_SAVE_DIR} \
+    | tee logs/tune_mcore_qwen2_7b_full_2000.log