Update Llama_pretraining.sh

40ea1bd3 · wxj · 18c8cf95 · 40ea1bd3
Commit 40ea1bd3 authored Jan 10, 2025 by wxj
Show whitespace changes
Inline Side-by-side

Showing with 63 additions and 69 deletions

Llama_pretraining.sh Llama_pretraining.sh +63 -69

No files found.
--- a/Llama_pretraining.sh
+++ b/Llama_pretraining.sh
@@ -21,32 +21,30 @@ export NCCL_NET_GDR_LEVEL=SYS
 export NCCL_NET_GDR_READ=0
 export GLOG_minloglevel=3 # 打印error级别的nccl日志
 source /opt/dtk/env.sh
-# te调用gemm需要导入hipblaslt库
+# 导入hipblaslt库
 # export LD_LIBRARY_PATH=/data/hipblaslt-install-0904/lib:$LD_LIBRARY_PATH 
 # 更新rocblas
-export LD_LIBRARY_PATH=/data/rocblas-install/lib:$LD_LIBRARY_PATH
+# export LD_LIBRARY_PATH=/data/rocblas-install_qwen1211/lib:$LD_LIBRARY_PATH
+# export LD_LIBRARY_PATH=/data/rocblas-install_qwen1228/lib:$LD_LIBRARY_PATH

-# # prof采集添加同步
+# torch控制多流转单流
+# export ALLREDUCE_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
+
 CHECKPOINT_PATH=./tmp_7b #$1 #<Specify path>
 TENSORBOARD_LOGS_PATH=./tmp_7b  #$2 #<Specify path>
-DATA_PATH="/public/home/wangxj3/Downloads/datasets/nemo_pretrain/oscar-1GB/oscar-1GB-llama_text_document" #<Specify path and file prefix>_text_document
-
-# GPT_MODEL_ARGS=(
-#     --num-layers 32
-#     --hidden-size 5120
-#     --ffn-hidden-size 13824
-#     --num-attention-heads 40
-#     --seq-length 4096 #4096
-#     --max-position-embeddings 32768 #4096
-#     --num-query-groups 40
-#     --group-query-attention
-# )
+DATA_PATH="/data/datasets/nemo_pretrain/oscar-1GB/oscar-1GB-llama_text_document" #<Specify path and file prefix>_text_document

 GPT_MODEL_ARGS=(
-    --num-layers 6
+    --num-layers 32
    --hidden-size 4096
    --ffn-hidden-size 11008 
    --num-attention-heads 32
@@ -56,15 +54,15 @@ GPT_MODEL_ARGS=(

 # export NVTE_FLASH_ATTN=1 # 走cutlass
 export NVTE_FLASH_ATTN_TRITON=1 # 走triton_fa
-# --transformer-impl transformer_engine
+# --transformer-impl transformer_engine # 走core用这两组参数
    # --use-mcore-models
-    # --transformer-impl local
+    # --transformer-impl local # 走legacy用这两组参数
    # --use-legacy-models 
 TRAINING_ARGS=(
-    --transformer-impl transformer_engine
-    --use-mcore-models
+    --transformer-impl local # 走legacy用这两组参数
+    --use-legacy-models 
    --micro-batch-size 1
-    --global-batch-size 6 #240 #60 #512 #64
+    --global-batch-size 60 #240 #60 #512 #64
    --train-iters 10
    --weight-decay 0.1 
    --adam-beta1 0.9 
@@ -72,12 +70,13 @@ TRAINING_ARGS=(
    --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
-    --add-qkv-bias
    --swiglu
    --lr 3.0e-5 
    --lr-decay-style cosine 
@@ -85,26 +84,19 @@ TRAINING_ARGS=(
    --lr-warmup-iters 1
    --ckpt-format torch
    --ddp-average-in-collective
-    --recompute-granularity full
-    --recompute-num-layers 5 #0 #
-    --recompute-method block
+    # --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
    --use-flash-attn-triton
 )
-# --add-qkv-bias # qwen
-# --ckpt-format torch
-#     --ddp-average-in-collective
-#     --recompute-granularity full
-#     --recompute-num-layers 5
-#     --recompute-method block
-#     --overlap-grad-reduce
-# --use-flash-attn-cutlass
-# --use-flash-attn-triton
+# --use-flash-attn-cutlass # cutlass fa
+# --use-flash-attn-triton # triton fa

 MODEL_PARALLEL_ARGS=(
    --sequence-parallel
 	--tensor-model-parallel-size 2
-	--pipeline-model-parallel-size 2
+	--pipeline-model-parallel-size 4
 )

 DATA_ARGS=(
@@ -115,7 +107,7 @@ DATA_ARGS=(
    --normalization RMSNorm 
    --no-position-embedding 
    --tokenizer-type Llama2Tokenizer
-    --tokenizer-model /public/home/wangxj3/Downloads/model_weights/llama2_7b_hf/tokenizer.model
+    --tokenizer-model /data/model_weights/llama2_7b_hf/tokenizer.model
 )

 EVAL_AND_LOGGING_ARGS=(
@@ -134,7 +126,7 @@ PROFILE_ARGS=(
    --profile-step-start 4
    --profile-step-end 5
    --use-pytorch-profiler
-    --profile-ranks 0 3
+    --profile-ranks 0 1 2 3 4 5 6 7
    --profile-dir prof_data
 )

@@ -142,7 +134,7 @@ RANK=$OMPI_COMM_WORLD_RANK
 LOCAL_RANK=$OMPI_COMM_WORLD_LOCAL_RANK
 WORLD_SIZE=$OMPI_COMM_WORLD_SIZE
 DIST_URL=${1}
-DIST_PORT=34566
+DIST_PORT=34567

 DISTRIBUTED_ARGS=(
    --rank ${RANK}
@@ -158,51 +150,53 @@ APP="python -u pretrain_gpt.py \
        ${DATA_ARGS[@]} \
        ${EVAL_AND_LOGGING_ARGS[@]} \
        ${DISTRIBUTED_ARGS[@]} \
-        ${PROFILE_ARGS[@]} \
+        
 "
+# 开启profile
+# ${PROFILE_ARGS[@]} \

-export HIP_VISIBLE_DEVICES=4,5,6,7 # 0,1,2,3 # 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 CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-  ${APP}
-  # numactl --cpunodebind=0 --membind=0 ${APP}
+  # ${APP}
+  numactl --cpunodebind=0 --membind=0 ${APP}
  ;;
 [1])
 #   export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-  ${APP}
-  # numactl --cpunodebind=0 --membind=0 ${APP}
+  # ${APP}
+  numactl --cpunodebind=0 --membind=0 ${APP}
  ;;
 [2])
 #   export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-  ${APP}
-  # numactl --cpunodebind=0 --membind=0 ${APP}
+  # ${APP}
+  numactl --cpunodebind=0 --membind=0 ${APP}
  ;;
 [3])
 #   export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-  ${APP}
-  # numactl --cpunodebind=0 --membind=0 ${APP}
+  # ${APP}
+  numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[4])
+  # export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  # ${APP}
+  numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[5])
+  # export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  # ${APP}
+  numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[6])
+  # export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  # ${APP}
+  numactl --cpunodebind=0 --membind=0 ${APP}
+  ;;
+[7])
+  # export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+  # ${APP}
+  numactl --cpunodebind=0 --membind=0 ${APP}
  ;;
-# [4])
-#   export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-#   ${APP}
-#   # numactl --cpunodebind=0 --membind=0 ${APP}
-#   ;;
-# [5])
-#   export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-#   ${APP}
-#   # numactl --cpunodebind=0 --membind=0 ${APP}
-#   ;;
-# [6])
-#   export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-#   ${APP}
-#   # numactl --cpunodebind=0 --membind=0 ${APP}
-#   ;;
-# [7])
-#   export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-#   ${APP}
-#   # numactl --cpunodebind=0 --membind=0 ${APP}
-#   ;;
 esac