[cambricon] support llava1.5_7b & llama3_70b & mixtral_8*7b with

training/benchmarks/llama3_70B_continuetrain/flagscale/run_pretraining.py

@@ -164,8 +164,8 @@ def replace_yamls(scale_home, config_module, args):
         args.log_dir, "outputs_llama3", "logs", "host_" +
         str(timestamp_log_noderank) + "_" + timestamp_log_host + ".output")
 
-    info_line = []
     while True:
+        info_line = []
         try:
             with open(timestamp_log_file, 'r') as f:
                 lines = f.readlines()

training/cambricon/docker_image/flagscale_2409/Dockerfile

@@ -0,0 +1,14 @@
+FROM cambricon_flagscale_2409_pt21
+#shell
+SHELL ["/bin/bash", "-c"]
+RUN apt-get update && DEBIAN_FRONTEND=noninteractive apt-get install -yq --no-install-recommends tzdata && apt-get install -y openssh-server && mkdir -p /run/sshd
+RUN apt update -y && apt install -y sudo dmidecode ipmitool sysstat net-tools sshpass
+# modify ~/.bashrc file
+RUN sed -i '/\[ -z "\$PS1" \] \&\& return/s/^/#/' ~/.bashrc
+RUN echo -e "\n# Add environment variables\n\
+export NEUWARE_HOME=/usr/local/neuware\n\
+export LD_LIBRARY_PATH=/usr/local/neuware/lib64:/usr/local/openmpi/lib:${LD_LIBRARY_PATH}\n\
+export PATH=/torch/venv3/pytorch/bin:/usr/local/neuware/bin:/usr/local/openmpi/bin:/usr/local/sbin:/usr/local/bin:/usr/sbin:/usr/bin:/sbin:/bin:${PATH}\n\
+export CNCL_MLULINK_OVER_ROCE_DISABLE=1\n\
+export CNCL_MLULINK_CROSS_HOSTS_ENABLE=0\n\
+export CNCL_MLU_DIRECT_LEVEL=1" >> ~/.bashrc

training/cambricon/docker_image/flagscale_2409/flagscale_2409_install.sh

@@ -0,0 +1,7 @@
+#!/bin/bash
+#!/bin/bash
+set -xe
+pip3 config set global.index-url https://pypi.tuna.tsinghua.edu.cn/simple
+pip3 install regex==2024.5.15 schedule==1.2.2 accelerate==0.31.0 transformers==4.40.1 protobuf==3.20.0
+pip3 install pybind11 hydra-core s3fs braceexpand webdataset wandb loguru sentencepiece datasets
+pip3 install megatron-energon==2.2.0

training/cambricon/llama3_70B_continuetrain-flagscale/README.md

@@ -0,0 +1 @@
+此测例为FlagScale相关项目测例

training/cambricon/llama3_70B_continuetrain-flagscale/config/config_MLUx4x8.py

@@ -0,0 +1,34 @@
+# scale_parent must under FlagPerf/ or data_dir/, otherwise you cannot mount it into baremetal, therefore cannot use shared storage
+scale_parent = "/share/project/zhaodeming/data_dir"
+scale_home = f"{scale_parent}/FlagScale/build/cambricon_MLU/FlagScale"
+
+# this cmd should install scale at <scale_home>. <scale_home> is set by flagperf.training.benchmarks.llava1.5_7b.flagscale.run_pretraining.py
+scale_download_cmd = f"cd {scale_parent}"
+
+# NV need nothing because all requirements have been established in base docker image. vendor can do anything related here
+scale_install_cmd = ""
+
+scale_conf_dir = f"{scale_home}/examples/llama/conf"
+configyaml = f"{scale_conf_dir}/config.yaml"
+trainyaml = f"{scale_conf_dir}/train/train_llama3_70b_finetune.yaml"
+dataset = f"{scale_parent}/SAMPLE50B/llama3/llama3_dataset"
+tokenizer = f"{scale_parent}/SAMPLE50B/llama3/llama3_tokenizer"
+ckpt = f"{scale_parent}/llama3_ckpt"
+
+#cmds = {"before_start": "source /root/miniconda3/bin/activate flagscale"}
+cmds = {}
+# flagscale's requirements
+flagscale_chip_type = "MLU"
+flagscale_ssh_port = 55623
+
+# 请在Flagperf目录下，创建MLU_FP16_FLOPS.py，
+# 文件包含MLU硬件算力值，示例如下：
+# MLU_FP16_FLOPS=1.0
+FLOPS_DIR='../../../../'
+import sys
+sys.path.append(FLOPS_DIR)
+from MLU_FP16_FLOPS import MLU_FP16_FLOPS
+flops = float(MLU_FP16_FLOPS)
+
+# for llama's algorithm
+steps = 500

training/cambricon/llava1.5_7b_continuetrain-flagscale/README.md

@@ -0,0 +1 @@
+此测例为FlagScale相关项目测例

training/cambricon/llava1.5_7b_continuetrain-flagscale/config/config_MLUx4x8.py

@@ -0,0 +1,35 @@
+# scale_parent must under FlagPerf/ or data_dir/, otherwise you cannot mount it into baremetal, therefore cannot use shared storage
+scale_parent = "/share/project/zhaodeming/data_dir"
+scale_home = f"{scale_parent}/FlagScale/build/cambricon_MLU/FlagScale"
+
+# this cmd should install scale at <scale_home>. <scale_home> is set by flagperf.training.benchmarks.llava1.5_7b.flagscale.run_pretraining.py
+scale_download_cmd = f"cd {scale_parent}"
+
+# NV need nothing because all requirements have been established in base docker image. vendor can do anything related here
+scale_install_cmd = ""
+
+# locate energon. the copy from energon_install_path to flagscale/megatron/ is done by flagperf...run_pretraining.py
+energon_locate_cmd = r"pip show megatron-energon | grep Location | awk -F: '{print $2}' | xargs"
+
+scale_conf_dir = f"{scale_home}/examples/llava/conf"
+configyaml = f"{scale_conf_dir}/config.yaml"
+trainyaml = f"{scale_conf_dir}/train/train_llava1.5_7b.yaml"
+datasetyaml = f"{scale_home}/megatron/examples/multimodal/pretrain_dataset.yaml"
+prompt = f"{scale_home}/megatron/examples/multimodal/manual_prompts.json"
+
+cmds = {}
+# flagscale's requirements
+flagscale_chip_type = "MLU"
+flagscale_ssh_port = 55623
+
+# 请在Flagperf目录下，创建MLU_FP16_FLOPS.py，
+# 文件包含MLU硬件算力值，示例如下：
+# MLU_FP16_FLOPS=1.0
+FLOPS_DIR='../../../../'
+import sys
+sys.path.append(FLOPS_DIR)
+from MLU_FP16_FLOPS import MLU_FP16_FLOPS
+flops = float(MLU_FP16_FLOPS)
+
+# for llava's algorithm
+steps = 5000

training/cambricon/llava1.5_7b_continuetrain-flagscale/config/requirements.txt

@@ -0,0 +1 @@
+megatron-energon==2.2.0

training/cambricon/mixtral_8x7B_continuetrain-flagscale/README.md

@@ -0,0 +1 @@
+此测例为FlagScale相关项目测例

training/cambricon/mixtral_8x7B_continuetrain-flagscale/config/config_MLUx4x8.py

@@ -0,0 +1,33 @@
+# scale_parent must under FlagPerf/ or data_dir/, otherwise you cannot mount it into baremetal, therefore cannot use shared storage
+scale_parent = "/share/project/zhaodeming/data_dir"
+scale_home = f"{scale_parent}/FlagScale/build/cambricon_MLU/FlagScale"
+
+# this cmd should install scale at <scale_home>. <scale_home> is set by flagperf.training.benchmarks.llava1.5_7b.flagscale.run_pretraining.py
+scale_download_cmd = f"cd {scale_parent}"
+
+# NV need nothing because all requirements have been established in base docker image. vendor can do anything related here
+scale_install_cmd = ""
+
+scale_conf_dir = f"{scale_home}/examples/mixtral/conf"
+configyaml = f"{scale_conf_dir}/config.yaml"
+trainyaml = f"{scale_conf_dir}/train/train_mixtral_8x7b.yaml"
+dataset = f"{scale_parent}/SAMPLE50B/mixtral/mixtral_dataset"
+tokenizer = f"{scale_parent}/SAMPLE50B/mixtral/mixtral_tokenizer"
+ckpt = f"{scale_parent}/mixtral_tp2_pp4_ep4_latest"
+
+cmds = {}
+# flagscale's requirements
+flagscale_chip_type = "MLU"
+flagscale_ssh_port = 55623
+
+# 请在Flagperf目录下，创建MLU_FP16_FLOPS.py，
+# 文件包含MLU硬件算力值，示例如下：
+# MLU_FP16_FLOPS=1.0
+FLOPS_DIR='../../../../'
+import sys
+sys.path.append(FLOPS_DIR)
+from MLU_FP16_FLOPS import MLU_FP16_FLOPS
+flops = float(MLU_FP16_FLOPS)
+
+# for mixtral's algorithm
+steps = 1000