Update Llama_pretraining.sh

Llama_pretraining.sh

@@ -8,7 +8,7 @@ export OMP_NUM_THREADS=1
 export NCCL_P2P_LEVEL=PXB # SYS
 
 #export HIP_ALLOC_INITIALIZE=0
-#export GPU_MAX_HW_QUEUES=20
+#export GPU_MAX_HW_QUEUES=20 # sglang空泡
 
 export NCCL_ALGO=Ring
 export NCCL_NCHANNELS_PER_PEER=16
@@ -17,54 +17,72 @@ export NCCL_IB_TIMEOUT=22
 export CUDA_DEVICE_MAX_CONNECTIONS=1
 
 export NCCL_IB_HCA=mlx5_1,mlx5_2
+# export NCCL_IB_HCA=mlx5_0,mlx5_1,mlx5_2,mlx5_3,,mlx5_4,,mlx5_5,,mlx5_6,,mlx5_7
 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
+
+# export TORCH_COMPILE_DEBUG=1 # 查看编译后的图
+
 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
+# export LD_LIBRARY_PATH=/data/rocblas-install_0107_trans/lib:$LD_LIBRARY_PATH
+
+# torch控制多流转单流
+# export ALLREDUCE_STREAM_WITH_COMPUTE=1
 
-# # prof采集添加同步
+# 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" 
 
 GPT_MODEL_ARGS=(
-    --num-layers 6
+    --num-layers 32
     --hidden-size 4096
     --ffn-hidden-size 11008 
     --num-attention-heads 32
-    --seq-length 4096 #4096
     --max-position-embeddings 4096
+
+    --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
 )
 
+# GPT_MODEL_ARGS=(
+#     --num-layers 40
+#     --hidden-size 5120
+#     --ffn-hidden-size 13824 
+#     --num-attention-heads 40
+#     --max-position-embeddings 4096
+
+#     --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
+# --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 64 #240 #60 #512 #64
     --train-iters 10
     --weight-decay 0.1 
     --adam-beta1 0.9 
@@ -72,34 +90,30 @@ 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
+    --no-gradient-accumulation-fusion # 开启后精度不对, apex更新后可以开启
     --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
-    --recompute-granularity full
-    --recompute-num-layers 5 #0 #
-    --recompute-method block
-    --overlap-grad-reduce
-    --use-flash-attn-triton
+    --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-cutlass
 )
-# --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
@@ -109,13 +123,10 @@ MODEL_PARALLEL_ARGS=(
 
 DATA_ARGS=(
     --data-path $DATA_PATH 
+    --seq-length 4096 #4096
     --split 949,50,1
-    --untie-embeddings-and-output-weights
-    --use-rotary-position-embeddings 
-    --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 +145,7 @@ 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 +153,7 @@ 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}
@@ -158,33 +169,35 @@ 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}
-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}
-  ;;
+#   # 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}
@@ -205,4 +218,4 @@ case ${LOCAL_RANK} in
 #   ${APP}
 #   # numactl --cpunodebind=0 --membind=0 ${APP}
 #   ;;
-esac
+# esac