Update Llama_pretraining.sh 最优化超参数环境变量设置

Llama_pretraining.sh

@@ -17,8 +17,8 @@ export NCCL_IB_TIMEOUT=22
 export CUDA_DEVICE_MAX_CONNECTIONS=1
 
 export NCCL_IB_HCA=mlx5_1,mlx5_2
-export NCCL_NET_GDR_LEVEL=SYS
-export NCCL_NET_GDR_READ=0
+export NCCL_NET_GDR_LEVEL=7
+export NCCL_NET_GDR_READ=1
 export GLOG_minloglevel=3 # 打印error级别的nccl日志
 source /opt/dtk/env.sh
 # 导入hipblaslt库
@@ -26,9 +26,11 @@ source /opt/dtk/env.sh
 # 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
 
 # torch控制多流转单流
-# export ALLREDUCE_STREAM_WITH_COMPUTE=1
+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
@@ -39,6 +41,9 @@ source /opt/dtk/env.sh
 # 采集 fa size
 # export FLASH_ATTENTION_PRINT_PARAM=1
 
+#增加编译缓存
+export cache_size_limit=64
+
 CHECKPOINT_PATH=./tmp_7b #$1 #<Specify path>
 TENSORBOARD_LOGS_PATH=./tmp_7b  #$2 #<Specify path>
 DATA_PATH="/data/datasets/nemo_pretrain/oscar-1GB/oscar-1GB-llama_text_document" #<Specify path and file prefix>_text_document
@@ -91,8 +96,8 @@ TRAINING_ARGS=(
     # --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
+    # --tp-comm-overlap # tensor parallel comm和gemm重叠
+    # --tp-comm-overlap-rs-dgrad # reduce-scatter和dgrad gemm重叠
     --use-flash-attn-triton
 )
 # --use-flash-attn-cutlass # cutlass fa
@@ -100,7 +105,7 @@ TRAINING_ARGS=(
 
 MODEL_PARALLEL_ARGS=(
     --sequence-parallel
-	--tensor-model-parallel-size 2
+	--tensor-model-parallel-size 1
 	--pipeline-model-parallel-size 2
 )
 
@@ -157,48 +162,50 @@ APP="python -u pretrain_gpt.py \
 # 开启profile
 # ${PROFILE_ARGS[@]} \
 
-export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 #  # 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 CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-#   ${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}
-#   ;;
-# [2])
-# #   export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
-#   ${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}
-#   ;;
-# [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
+
+# 使用numactl绑定
+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