update nccl and launch with binding

README.md

@@ -73,27 +73,11 @@ def unpermute(
 ### 使用方式
 在使用时，进入到examples目录下，有相关模型执行脚本，所用数据集请自行下载：https://r0ddbu55vzx.feishu.cn/drive/folder/ZxHHfCoX4lg75td2hTqcmiAin3g
 ```
-examples
+examples/
+├── deepseek_v3
 ├── gpt3
-│   ├── hostfile_gpt_567B
-│   ├── README.md
-│   ├── run_gpt_567B_1nodes.sh
-│   ├── run_gpt_567B_multinodes.sh
-│   ├── topo-input.xml
-│   ├── train_gpt_567B_1nodes.sh
-│   └── train_gpt_567B_multinodes.sh
-└── mixtral
-    ├── hostfile_mixtral_8x22B
-    ├── hostfile_mixtral_8x7B
-    ├── README.md
-    ├── run_mixtral_8x22B_1nodes.sh
-    ├── run_mixtral_8x22B_multinodes.sh
-    ├── run_mixtral_8x7B_1nodes.sh
-    ├── run_mixtral_8x7B_multinodes.sh
-    ├── topo-input.xml
-    ├── train_mixtral_8x22B_1nodes.sh
-    ├── train_mixtral_8x22B_multinodes.sh
-    ├── train_mixtral_8x7B_1nodes.sh
-    └── train_mixtral_8x7B_multinodes.sh
+├── llama
+├── mixtral
+└── qwen
 ```
 

requirements/launch_with_binding.sh

 #!/bin/bash
 
 export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
+numa_map=(0 1 2 3 4 5 6 7)
 
 LOCAL_RANK=$1
 shift
 
-numa_map=(0 1 2 3 4 5 6 7)
 NUMA_ID=${numa_map[$LOCAL_RANK]}
-numactl --cpunodebind=${NUMA_ID} --membind=${NUMA_ID} "$@"
+numactl --cpunodebind=${NUMA_ID} --membind=${NUMA_ID} "$@"
\ No newline at end of file

requirements/nccl_zz/env.sh

@@ -2,15 +2,23 @@
 CURRENT_DIR="$( cd "$( dirname "$0" )" && pwd )"
 MEGATRON_PATH=$( dirname $( dirname ${CURRENT_DIR}))
 export NCCL_ALGO=Ring
-export NCCL_MAX_NCHANNELS=32
-export NCCL_MIN_NCHANNELS=32
-export NCCL_NET_GDR_LEVEL=4
+export NCCL_MAX_NCHANNELS=16
+export NCCL_MIN_NCHANNELS=16
+export NCCL_NCHANNELS_PER_PEER=16
+export NCCL_MIN_P2P_NCHANNELS=16
+export NCCL_MAX_P2P_NCHANNELS=16
+export NCCL_NET_GDR_LEVEL=7
 export NCCL_NET_GDR_READ=1
 export RCCL_SDMA_COPY_ENABLE=0
 export NCCL_IB_HCA=shca_0:1,shca_1:1,shca_2:1,shca_3:1
 export NCCL_TOPO_FILE=${MEGATRON_PATH}/requirements/nccl_zz/topo-input.xml
 export NCCL_IB_PCI_RELAXED_ORDERING=0
 export NCCL_PLUGIN_P2P=ucx
-export NCCL_PXN_DISABLE=0
 export NCCL_SOCKET_IFNAME=eno1
+export SHCA_DEBUG_MASK=0
+export SHCA_CMR_LOG_LEVEL=1
+export SHCA_SHUT_UP_FWB=1
+export SHCA_UCT_CQ_SIZE_INC=5
+export UCX_RNDV_PUT_FORCE_FLUSH=y
+export NCCL_PXN_DISABLE=0
 export LD_LIBRARY_PATH=${MEGATRON_PATH}/requirements/nccl_zz/lib-v8:$LD_LIBRARY_PATH
\ No newline at end of file