!2196 DeepSeek-R1-Distill-Qwen-1.5B and LLaMA-8B、70B

Merge pull request !2196 from guoxinjie/distill-qwen

README.md

@@ -19,6 +19,27 @@ MindSpeed-LLM是基于昇腾生态的大语言模型分布式训练框架，旨
 
 ---
 
+## NEWS !!!
+🚀🚀🚀**DeepSeek-R1-Distill** 系列模型已上线！！🚀🚀🚀
+
+**[DeepSeek-R1-Distill-Qwen](./examples/mcore/deepseek_r1_distill_qwen/)** 😊
+
+**[DeepSeek-R1-Distill-LLaMA](./examples/mcore/deepseek_r1_distill_llama/)** 😊
+
+包含数据处理、权重转换、在线推理、全参微调
+
+## COMING !!!
+
+🚀🚀 On 2.16 will release **DeepSeekV3-671B**: Pretrain、Lora、QLora 👋👋
+
+🚀🚀 On 2.16 will release **DeepSeek-R1-ZERO** 👋👋
+
+🚀🚀 On 2.23 will release **DeepSeek-R1**：SFT、GRPO、Reward 👋👋
+
+🚀🚀 On 3.30 will release **[MindSpeed-RL](https://gitee.com/ascend/MindSpeed-RL)** 👋👋
+
+---
+
 ## 版本配套表
 
 MindSpeed-LLM的以来配套如下表，安装步骤参考[安装指导](./docs/features/install_guide.md)。

docs/models/dense_model.md

@@ -726,9 +726,17 @@
       <td>【Test】</td>
     </tr>
     <tr>
-      <td rowspan="3"><a href="https://huggingface.co/deepseek-ai">DeepSeek-R1-Distill-Qwen</a></td>
+      <td rowspan="4"><a href="https://huggingface.co/deepseek-ai">DeepSeek-R1-Distill-Qwen</a></td>
+      <td><a href="https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B">1.5B</a></td>
+      <td rowspan="4"><a href="../../examples/mcore/deepseek_r1_distill_qwen">deepseek_r1_distill_qwen</a></td>
+      <td> 4K </td>
+      <th>Mcore</th>
+      <td> 1x8 </td>
+      <td>【Ascend】</td>
+      <td>【Test】</td>
+    </tr>
+    <tr>
       <td><a href="https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B">7B</a></td>
-      <td rowspan="3"><a href="../../examples/mcore/deepseek_r1_distill_qwen">deepseek_r1_distill_qwen</a></td>
       <td> 4K </td>
       <th>Mcore</th>
       <td> 1x8 </td>
@@ -751,6 +759,24 @@
       <td>【Ascend】</td>
       <td>【Test】</td>
     </tr>
+    <tr>
+      <td rowspan="2"><a href="https://huggingface.co/deepseek-ai">DeepSeek-R1-Distill-LLaMA</a></td>
+      <td><a href="https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B">8B</a></td>
+      <td rowspan="2"><a href="../../examples/mcore/deepseek_r1_distill_llama/">deepseek_r1_distill_llama</a></td>
+      <td> 8K </td>
+      <th>Mcore</th>
+      <td> 1x8 </td>
+      <td>【Ascend】</td>
+      <td>【Test】</td>
+    </tr>
+    <tr>
+      <td><a href="https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B">70B</a></td>
+      <td> 8K </td>
+      <th>Mcore</th>
+      <td> 4x8 </td>
+      <td>【Ascend】</td>
+      <td>【Test】</td>
+    </tr>
   </tbody>
 </table>
 

examples/mcore/deepseek_r1_distill_llama/ckpt_convert_distill_llama_hf2mcore.sh

@@ -0,0 +1,16 @@
+# 修改 ascend-toolkit 路径
+source /usr/local/Ascend/ascend-toolkit/set_env.sh
+
+# 权重格式转换
+python convert_ckpt.py \
+   --use-mcore-models \
+   --model-type-hf llama2 \
+   --model-type GPT \
+   --load-model-type hf \
+   --save-model-type mg \
+   --params-dtype bf16 \
+   --target-tensor-parallel-size 1 \
+   --target-pipeline-parallel-size 1 \
+   --load-dir ./model_from_hf/llama3_distill/ \
+   --save-dir ./model_weights/llama3_mcore/ \
+   --tokenizer-model ./model_from_hf/llama3_distill/tokenizer.json  # --num-layer-list 17,20,22,21 等参数根据模型需求添加

examples/mcore/deepseek_r1_distill_llama/data_convert_distill_llama_instruction.sh

@@ -0,0 +1,16 @@
+# 请根据 examples/README.md 下 “数据集准备及处理” 章节下载 Alpaca 数据集
+# 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/llama3-distill/ \
+    --output-prefix ./finetune_dataset/alpaca \
+    --workers 4 \
+    --log-interval 1000 \
+    --tokenizer-type PretrainedFromHF \
+    --tokenizer-not-use-fast \
+    --handler-name AlpacaStyleInstructionHandler \
+    --prompt-type llama3 \
+    --seq-length 8192 \

examples/mcore/deepseek_r1_distill_llama/generate_distill_llama_70b.sh

@@ -0,0 +1,55 @@
+#!/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=6001
+NNODES=1
+NODE_RANK=0
+NPUS_PER_NODE=4
+WORLD_SIZE=$(($NPUS_PER_NODE*$NNODES))
+
+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 \
+       --tensor-model-parallel-size 4 \
+       --pipeline-model-parallel-size 1 \
+       --use-mcore-models \
+       --num-layers 80 \
+       --hidden-size 8192  \
+       --ffn-hidden-size 28672 \
+       --position-embedding-type rope \
+       --rotary-base 500000 \
+       --num-attention-heads 64 \
+       --group-query-attention \
+       --num-query-groups 8 \
+       --swiglu \
+       --normalization RMSNorm \
+       --norm-epsilon 1e-5 \
+       --use-fused-swiglu \
+       --use-fused-rmsnorm \
+       --use-fused-rotary-pos-emb \
+       --use-rotary-position-embeddings \
+       --load ${CHECKPOINT} \
+       --tokenizer-type PretrainedFromHF \
+       --tokenizer-name-or-path ${TOKENIZER_PATH} \
+       --tokenizer-not-use-fast \
+       --seq-length 8192 \
+       --max-position-embeddings 8192 \
+       --max-new-tokens 256 \
+       --micro-batch-size 1 \
+       --hidden-dropout 0 \
+       --attention-dropout 0 \
+       --untie-embeddings-and-output-weights \
+       --disable-bias-linear \
+       --attention-softmax-in-fp32 \
+       --exit-on-missing-checkpoint \
+       --make-vocab-size-divisible-by 16032 \
+       --bf16 \
+       --seed 42 \
+       | tee logs/generate_mcore_llama_distill_70b.log
+

examples/mcore/deepseek_r1_distill_llama/generate_distill_llama_8b.sh

@@ -0,0 +1,54 @@
+#!/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=6001
+NNODES=1
+NODE_RANK=0
+NPUS_PER_NODE=1
+WORLD_SIZE=$(($NPUS_PER_NODE*$NNODES))
+
+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 \
+       --tensor-model-parallel-size 1 \
+       --pipeline-model-parallel-size 1 \
+       --use-mcore-models \
+       --num-layers 32 \
+       --hidden-size 4096 \
+       --ffn-hidden-size 14336 \
+       --position-embedding-type rope \
+       --rotary-base 500000 \
+       --num-attention-heads 32 \
+       --group-query-attention \
+       --num-query-groups 8 \
+       --swiglu \
+       --normalization RMSNorm \
+       --norm-epsilon 1e-5 \
+       --seq-length 8192 \
+       --max-position-embeddings 8192 \
+       --max-new-tokens 256 \
+       --micro-batch-size 1 \
+       --use-fused-swiglu \
+       --use-fused-rmsnorm \
+       --use-fused-rotary-pos-emb \
+       --load ${CHECKPOINT} \
+       --tokenizer-type PretrainedFromHF \
+       --tokenizer-name-or-path ${TOKENIZER_PATH} \
+       --tokenizer-not-use-fast \
+       --hidden-dropout 0 \
+       --attention-dropout 0 \
+       --untie-embeddings-and-output-weights \
+       --disable-bias-linear \
+       --attention-softmax-in-fp32 \
+       --exit-on-missing-checkpoint \
+       --make-vocab-size-divisible-by 16032 \
+       --bf16 \
+       --seed 42 \
+       | tee logs/generate__mcore_llama_distill_8b.log
+

examples/mcore/deepseek_r1_distill_llama/tune_distill_llama_70b_full.sh

@@ -0,0 +1,108 @@
+#!/bin/bash
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+
+NPUS_PER_NODE=8
+MASTER_ADDR="your master IP Address"
+MASTER_PORT="your master IP Port, can be changed if current port is occupied"
+NNODES=4
+NODE_RANK=0
+WORLD_SIZE=$(($NPUS_PER_NODE*$NNODES))
+
+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=8
+PP=4
+
+DISTRIBUTED_ARGS="
+    --nproc_per_node $NPUS_PER_NODE \
+    --nnodes $NNODES \
+    --node_rank $NODE_RANK \
+    --master_addr $MASTER_ADDR \
+    --master_port $MASTER_PORT
+"
+
+TUNE_ARGS="
+    --finetune \
+    --stage sft \
+    --is-instruction-dataset \
+    --variable-seq-lengths \
+    --prompt-type llama3 \
+"
+
+GPT_ARGS="
+    --tensor-model-parallel-size ${TP} \
+    --pipeline-model-parallel-size ${PP} \
+    --num-layers-per-virtual-pipeline-stage 2 \
+    --use-mcore-models \
+    --micro-batch-size 1 \
+    --global-batch-size 128 \
+    --sequence-parallel \
+    --use-fused-rotary-pos-emb \
+    --use-fused-rmsnorm \
+    --use-fused-swiglu \
+    --tokenizer-type PretrainedFromHF \
+    --tokenizer-name-or-path ${TOKENIZER_PATH} \
+    --num-layers 80 \
+    --hidden-size 8192 \
+    --ffn-hidden-size 28672 \
+    --num-attention-heads 64 \
+    --group-query-attention \
+    --num-query-groups 8 \
+    --position-embedding-type rope \
+    --rotary-base 500000 \
+    --normalization RMSNorm \
+    --norm-epsilon 1e-5 \
+    --swiglu \
+    --seq-length 8192 \
+    --max-position-embeddings 8192 \
+    --make-vocab-size-divisible-by 16032 \
+    --untie-embeddings-and-output-weights \
+    --disable-bias-linear \
+    --attention-dropout 0.0 \
+    --init-method-std 0.01 \
+    --hidden-dropout 0.0 \
+    --no-masked-softmax-fusion \
+    --attention-softmax-in-fp32 \
+    --lr 1.25e-6 \
+    --train-iters 2000 \
+    --lr-decay-style cosine \
+    --min-lr 1.25e-7 \
+    --weight-decay 1e-1 \
+    --lr-warmup-fraction 0.01 \
+    --clip-grad 1.0 \
+    --adam-beta1 0.9 \
+    --adam-beta2 0.95 \
+    --initial-loss-scale 4096 \
+    --no-gradient-accumulation-fusion \
+    --no-shared-storage \
+    --no-load-optim \
+    --no-load-rng \
+    --no-save-optim \
+    --no-save-rng \
+    --bf16
+"
+
+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 posttrain_gpt.py \
+    $TUNE_ARGS \
+    $GPT_ARGS \
+    $DATA_ARGS \
+    $OUTPUT_ARGS \
+    $ACCELERATE_ARGS \
+    --distributed-backend nccl \
+    --load ${CKPT_LOAD_DIR} \
+    --save ${CKPT_SAVE_DIR} \
+    | tee logs/tune_mcore_distill_llama_70b_full_8k.log

examples/mcore/deepseek_r1_distill_llama/tune_distill_llama_8b_full.sh

@@ -0,0 +1,117 @@
+#!/bin/bash
+
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+export PYTORCH_NPU_ALLOC_CONF=expandable_segments:True
+
+
+GPUS_PER_NODE=8
+MASTER_ADDR=localhost
+MASTER_PORT=6024
+NNODES=1
+NODE_RANK=0
+WORLD_SIZE=$(($GPUS_PER_NODE*$NNODES))
+
+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=2
+
+DISTRIBUTED_ARGS="
+    --nproc_per_node $GPUS_PER_NODE \
+    --nnodes $NNODES \
+    --node_rank $NODE_RANK \
+    --master_addr $MASTER_ADDR \
+    --master_port $MASTER_PORT
+"
+
+TUNE_ARGS="
+    --finetune \
+    --stage sft \
+    --is-instruction-dataset \
+    --variable-seq-lengths \
+    --prompt-type llama3 \
+"
+
+GPT_ARGS="
+    --tensor-model-parallel-size ${TP} \
+    --pipeline-model-parallel-size ${PP} \
+    --use-mcore-models \
+    --num-layers 32 \
+    --hidden-size 4096 \
+    --ffn-hidden-size 14336 \
+    --num-attention-heads 32 \
+    --group-query-attention \
+    --num-query-groups 8 \
+    --position-embedding-type rope \
+    --rotary-base 500000 \
+    --normalization RMSNorm \
+    --norm-epsilon 1e-5 \
+    --swiglu \
+    --micro-batch-size 1 \
+    --global-batch-size 64 \
+    --use-flash-attn \
+    --prompt-type llama3 \
+    --rope-scaling-type llama3 \
+    --rope-scaling-factor 8.0 \
+    --rotary-percent 1.0 \
+    --low-freq-factor 1.0 \
+    --high-freq-factor 4.0 \
+    --original-max-position-embeddings 8192 \
+    --tokenizer-type PretrainedFromHF \
+    --tokenizer-name-or-path ${TOKENIZER_PATH} \
+    --tokenizer-not-use-fast \
+    --seq-length 8192 \
+    --max-position-embeddings 8192 \
+    --make-vocab-size-divisible-by 1 \
+    --padded-vocab-size 128256 \
+    --untie-embeddings-and-output-weights \
+    --disable-bias-linear \
+    --attention-dropout 0.0 \
+    --init-method-std 0.02 \
+    --hidden-dropout 0.0 \
+    --no-masked-softmax-fusion \
+    --attention-softmax-in-fp32 \
+    --lr 1e-7 \
+    --train-iters 2000 \
+    --lr-decay-style constant \
+    --weight-decay 0.0 \
+    --clip-grad 1.0 \
+    --adam-beta1 0.9 \
+    --adam-beta2 0.999 \
+    --initial-loss-scale 4096 \
+    --no-gradient-accumulation-fusion \
+    --no-load-optim \
+    --no-load-rng \
+    --no-save-optim \
+    --no-save-rng \
+    --bf16 \
+    --seed 42 \
+    --vocab-size 128256 \
+    --reuse-fp32-param \
+    --use-distributed-optimizer \
+"
+
+DATA_ARGS="
+    --data-path $DATA_PATH \
+    --split 100,0,0
+"
+
+OUTPUT_ARGS="
+    --log-interval 1 \
+    --save-interval 2000 \
+    --eval-interval 2000 \
+    --eval-iters 10 \
+    --load ${CKPT_LOAD_DIR} \
+    --save ${CKPT_SAVE_DIR} \
+"
+
+torchrun $DISTRIBUTED_ARGS posttrain_gpt.py \
+    $TUNE_ARGS \
+    $GPT_ARGS \
+    $DATA_ARGS \
+    $OUTPUT_ARGS \
+    --distributed-backend nccl \
+    | tee logs/tune_mcore_distill_llama_8b_full_8k.log

examples/mcore/deepseek_r1_distill_qwen/generate_distill_qwen_1point5b.sh

@@ -0,0 +1,64 @@
+#!/bin/bash
+export CUDA_DEVICE_MAX_CONNECTIONS=1
+
+# Change for multinode config
+MASTER_ADDR=localhost
+MASTER_PORT=6002
+NNODES=1
+NODE_RANK=0
+NPUS_PER_NODE=1
+WORLD_SIZE=$(($NPUS_PER_NODE*$NNODES))
+
+# please fill these path configurations
+CHECKPOINT="your model ckpt path"
+TOKENIZER_PATH="your tokenizer path"
+
+TP=1
+PP=1
+MBS=1
+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
+"
+
+torchrun $DISTRIBUTED_ARGS inference.py \
+       --use-mcore-models \
+       --tensor-model-parallel-size ${TP} \
+       --pipeline-model-parallel-size ${PP} \
+       --num-layers 28 \
+       --hidden-size 1536  \
+       --ffn-hidden-size 8960 \
+       --num-attention-heads 12 \
+       --group-query-attention \
+       --num-query-groups 2 \
+       --tokenizer-type PretrainedFromHF  \
+       --tokenizer-name-or-path ${TOKENIZER_PATH} \
+       --max-position-embeddings ${SEQ_LEN} \
+       --seq-length ${SEQ_LEN} \
+       --make-vocab-size-divisible-by 1 \
+       --padded-vocab-size 151936 \
+       --rotary-base 10000 \
+       --micro-batch-size ${MBS} \
+       --swiglu \
+       --add-qkv-bias \
+       --untie-embeddings-and-output-weights \
+       --disable-bias-linear \
+       --load ${CHECKPOINT}  \
+       --normalization RMSNorm \
+       --norm-epsilon 1e-6 \
+       --position-embedding-type rope \
+       --hidden-dropout 0 \
+       --attention-dropout 0 \
+       --tokenizer-not-use-fast \
+       --max-new-tokens 256 \
+       --no-gradient-accumulation-fusion \
+       --exit-on-missing-checkpoint \
+       --attention-softmax-in-fp32 \
+       --seed 42 \
+       --bf16 \
+       | tee logs/generate_mcore_distil_qwen_1point5b.log

examples/mcore/deepseek_r1_distill_qwen/tune_distill_qwen_1point5b_full.sh

@@ -0,0 +1,119 @@
+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=1
+SEQ_LEN=4096
+MBS=1
+GBS=64
+
+DISTRIBUTED_ARGS="
+    --nproc_per_node $NPUS_PER_NODE \
+    --nnodes $NNODES \
+    --node_rank $NODE_RANK \
+    --master_addr $MASTER_ADDR \
+    --master_port $MASTER_PORT
+"
+
+TUNE_ARGS="
+    --finetune \
+    --stage sft \
+    --is-instruction-dataset \
+    --variable-seq-lengths \
+    --prompt-type qwen \
+"
+
+GPT_ARGS="
+    --use-mcore-models \
+    --tensor-model-parallel-size ${TP} \
+    --pipeline-model-parallel-size ${PP} \
+    --num-layers 28 \
+    --hidden-size 1536 \
+    --ffn-hidden-size 8960 \
+    --num-attention-heads 12 \
+    --max-position-embeddings ${SEQ_LEN} \
+    --seq-length ${SEQ_LEN} \
+    --disable-bias-linear \
+    --add-qkv-bias \
+    --group-query-attention \
+    --num-query-groups 2 \
+    --use-flash-attn \
+    --swiglu \
+    --use-fused-swiglu \
+    --normalization RMSNorm \
+    --norm-epsilon 1e-6 \
+    --use-fused-rmsnorm \
+    --position-embedding-type rope \
+    --rotary-base 10000 \
+    --use-fused-rotary-pos-emb \
+    --untie-embeddings-and-output-weights \
+    --micro-batch-size ${MBS} \
+    --global-batch-size ${GBS} \
+    --make-vocab-size-divisible-by 1 \
+    --padded-vocab-size 151936 \
+    --tokenizer-type PretrainedFromHF \
+    --tokenizer-name-or-path ${TOKENIZER_PATH} \
+    --tokenizer-not-use-fast \
+    --attention-dropout 0.0 \
+    --hidden-dropout 0.0 \
+    --train-iters 2000 \
+    --lr 1.25e-6 \
+    --lr-decay-style cosine \
+    --min-lr 1.25e-7 \
+    --lr-warmup-fraction 0.01 \
+    --init-method-std 0.01 \
+    --weight-decay 0.0 \
+    --clip-grad 1.0 \
+    --adam-beta1 0.9 \
+    --adam-beta2 0.95 \
+    --initial-loss-scale 4096 \
+    --no-gradient-accumulation-fusion \
+    --no-masked-softmax-fusion \
+    --attention-softmax-in-fp32 \
+    --bf16
+"
+
+DATA_ARGS="
+    --data-path $DATA_PATH \
+    --split 100,0,0
+"
+
+CKPT_ARGS="
+    --no-load-optim \
+    --no-load-rng \
+    --no-save-optim \
+    --no-save-rng \
+    --seed 1234 \
+"
+
+OUTPUT_ARGS="
+    --log-interval 1 \
+    --save-interval 2000 \
+    --eval-interval 2000 \
+    --eval-iters 0 \
+    --log-throughput
+"
+
+torchrun $DISTRIBUTED_ARGS posttrain_gpt.py \
+    $GPT_ARGS \
+    $DATA_ARGS \
+    $CKPT_ARGS \
+    $OUTPUT_ARGS \
+    $TUNE_ARGS \
+    --load ${CKPT_LOAD_DIR} \
+    --save ${CKPT_SAVE_DIR} \
+    --distributed-backend nccl \
+    | tee logs/tune_mcore_distill_qwen_1.5b_full_4k.log