source /mnt/zhengcf3/env/lmp/bin/activate
/mnt/zhengcf3/nvidia/nsight-systems/2025.6.1/bin/nsys profile --force-overwrite true -o report1.nsys-rep python generate.py > generate_multi.log
/mnt/zhengcf3/nvidia/nsight-systems/2025.6.1/bin/nsys profile --force-overwrite true -o report2.nsys-rep python test_normal.py > test_normal.log
/mnt/zhengcf3/nvidia/nsight-systems/2025.6.1/bin/nsys profile --force-overwrite true -o report3.nsys-rep python test_sllm_store.py > test_sllm_store_b2.log
/mnt/zhengcf3/nvidia/nsight-systems/2025.6.1/bin/nsys profile --force-overwrite true -o report4_testmp.nsys-rep python test_init_meta_manager_mp_shared.py > test_init_meta_manager_mp_shared.log
/mnt/zhengcf3/nvidia/nsight-systems/2025.6.1/bin/nsys profile --force-overwrite true -o report5.nsys-rep python test_device_mp.py > test_device_mp.log
/mnt/zhengcf3/nvidia/nsight-systems/2025.6.1/bin/nsys profile --force-overwrite true -o report6.nsys-rep python test_cpu_mp.py > test_cpu_mp.log
cd /mnt/zhengcf3/lmp/src/sllm_store && python -m grpc_tools.protoc --proto_path=proto --python_out=sllm_store --grpc_python_out=sllm_store proto/storage.proto
pip install -e . python setup.py build_ext --inplace
Deepseek sllm-store start --storage-path /mnt/zhengcf3/models/sllm_models --num-thread 8 --mem-pool-size 32GB --use-shared-memory True --chunk-size 1056MB
QWEN30B sllm-store start --storage-path /mnt/zhengcf3/models/sllm_models --num-thread 8 --mem-pool-size 64GB --use-shared-memory True --chunk-size 1152MB
Gemma4 sllm-store start --storage-path /mnt/zhengcf3/models/sllm_models --num-thread 8 --mem-pool-size 64GB --use-shared-memory True --chunk-size 1452MB
numactl --cpunodebind=0 --membind=0 sllm-store start --storage-path /mnt/zhengcf3/models/sllm_models --num-thread 8 --mem-pool-size 64GB --use-shared-memory True --chunk-size 1GB
sllm-store start --storage-path /mnt/zhengcf3/models/sllm_models --num-thread 4 --mem-pool-size 64GB --use-shared-memory True --chunk-size 1GB
Gemma4 vllm sllm-store start --storage-path /mnt/zhengcf3/models/vllm_sllm_models --num-thread 8 --mem-pool-size 60GB --use-shared-memory True --chunk-size 1GB --port 8074
QWEN-1.5 sllm-store start --storage-path /mnt/zhengcf3/models/sllm_models --num-thread 8 --mem-pool-size 32GB --use-shared-memory True --chunk-size 990MB --port 8074
sllm-store start --storage-path /mnt/zhengcf3/models/sllm_models --num-thread 8 --mem-pool-size 95GB --use-shared-memory True --chunk-size 2688MB
python /mnt/zhengcf3/lmp/examples/test_sllm_store.py
python /mnt/zhengcf3/lpllm/lpllm/resize_index.py /mnt/zhengcf3/models/sllm_models/Qwen3-30B-A3B/tensor_index.json
conda create -n lpllm python=3.10 -y & apt-get install gpustat -y
export PATH=/mnt/huwf5/conda-envs/sida39/bin:$PATH
watch -n 1 gpustat
python /mnt/zhengcf3/lpllm/lpllm/resize_index.py /mnt/zhengcf3/models/sllm_models/Mixtral-8x7B/tensor_index.json > /mnt/zhengcf3/models/sllm_models/Mixtral-8x7B/chunk_size
python setup.py build_ext --inplace python -m pip install -e . --no-build-isolation
pip install "git+https://github.com/huggingface/transformers.git@v5.5.4" pip install -i https://pypi.org/simple --upgrade "transformers==5.5.3"
source /mnt/zhengcf3/lmp_env/vllm/bin/activate
source /root/miniconda3/etc/profile.d/conda.sh conda create -y -p /mnt/zhengcf3/conda_envs/sllm_vllm python=3.10 conda activate /mnt/zhengcf3/conda_envs/sllm_vllm
source /mnt/zhengcf3/use_mnt_conda.sh
conda create -y -n myenv python=3.10 conda activate myenv
conda create -y -p /mnt/zhengcf3/conda_envs/myenv python=3.10 conda activate /mnt/zhengcf3/conda_envs/myenv
/mnt/zhengcf3/miniconda3/bin/conda create -y -p /mnt/zhengcf3/conda_envs/myenv -c conda-forge python=3.10 /mnt/zhengcf3/miniconda3/bin/conda run -p /mnt/zhengcf3/conda_envs/myenv python -V /mnt/zhengcf3/miniconda3/bin/conda run -p /mnt/zhengcf3/conda_envs/myenv pip install -U pip eval "$(/mnt/zhengcf3/miniconda3/bin/conda shell.bash hook)" conda activate /mnt/zhengcf3/conda_envs/myenv
pip install -i https://pypi.org/simple datasets
source /mnt/zhengcf3/lmp_env/fslmp/bin/activate
conda create -n lpllm python=3.10 -y
python3 /mnt/zhengcf3/lmp/scripts/split_gemma4_tensor_index_experts.py
--index-in /mnt/zhengcf3/models/sllm_models/gemma4-26B-A4B/tensor_index.json
--config /mnt/zhengcf3/models/sllm_models/gemma4-26B-A4B/config.json
--index-out /mnt/zhengcf3/models/sllm_models/gemma4-26B-A4B/tensor_index_resize_per_expert.json
version 0.19.1 patch for serverlessllm 0.8.0
source /mnt/zhengcf3/lmp_env/fslmp/bin/activate
source /mnt/zhengcf3/lmp_env/lmp/bin/activate
website https://serverlessllm.github.io/docs/store/quickstart
save load
python3 /mnt/zhengcf3/ServerlessLLM/sllm_store/examples/save_vllm_model.py --model-name gemma4-26B-A4B --storage-path /mnt/zhengcf3/models/vllm_sllm_models --tensor-parallel-size 4
本机为 3×GPU 时：tensor_parallel_size=4 不可用；tensor_parallel_size=3 与 ERNIE / Qwen 的 head 数不整除。已用 TP=2、--dtype bfloat16、--gpu-memory-utilization 0.95，并在运行前 export VLLM_USE_FLASHINFER_MOE_FP16=1 VLLM_FLASHINFER_MOE_BACKEND=latency 跑通下列四个权重（save_vllm_model.py 内已设 disable_custom_all_reduce=True）。一键顺序导出：bash /mnt/zhengcf3/lmp/scripts/run_save_vllm_models_four.sh（日志在 lmp/logs/save_vllm_models/）。
python3 /mnt/zhengcf3/ServerlessLLM/sllm_store/examples/load_vllm_model.py --model-name gemma4-26B-A4B --storage-path /mnt/zhengcf3/models/vllm_sllm_models --tensor-parallel-size 4
` python3 /mnt/zhengcf3/ServerlessLLM/sllm_store/examples/save_vllm_model.py --model-name Qwen3.5-35B --local-model-path /mnt/zhengcf3/models/Qwen3.5-35B --storage-path /mnt/zhengcf3/models/vllm_sllm_models --tensor-parallel-size 4
python3 /mnt/zhengcf3/ServerlessLLM/sllm_store/examples/save_vllm_model.py --model-name gemma4-26B-A4B --local-model-path /mnt/zhengcf3/models/gemma4-26B-A4B --storage-path /mnt/zhengcf3/models/vllm_sllm_models --tensor-parallel-size 4
python3 /mnt/zhengcf3/ServerlessLLM/sllm_store/examples/save_vllm_model.py --model-name Qwen3-30B-A3B --local-model-path /mnt/zhengcf3/models/Qwen3-30B-A3B --storage-path /mnt/zhengcf3/models/vllm_sllm_models --tensor-parallel-size 4
python3 /mnt/zhengcf3/ServerlessLLM/sllm_store/examples/save_vllm_model.py --model-name ERNIE-4.5-VL-28B-A3B-Thinking --local-model-path /mnt/zhengcf3/models/ERNIE-4.5-VL-28B-A3B-Thinking --storage-path /mnt/zhengcf3/models/vllm_sllm_models --tensor-parallel-size 4
python3 /mnt/zhengcf3/ServerlessLLM/sllm_store/examples/save_vllm_model.py --model-name ERNIE-4.5-21B-A3B-Thinking --local-model-path /mnt/zhengcf3/models/ERNIE-4.5-21B-A3B-Thinking --storage-path /mnt/zhengcf3/models/vllm_sllm_models --tensor-parallel-size 4
python3 /mnt/zhengcf3/ServerlessLLM/sllm_store/examples/save_vllm_model.py --model-name Qwen1.5-MoE-A2.7B --local-model-path /mnt/zhengcf3/models/Qwen1.5-MoE-A2.7B --storage-path /mnt/zhengcf3/models/vllm_sllm_models --tensor-parallel-size 4
python3 /mnt/zhengcf3/ServerlessLLM/sllm_store/examples/save_vllm_model.py --model-name gpt-oss-20b --local-model-path /mnt/zhengcf3/models/gpt-oss-20b --storage-path /mnt/zhengcf3/models/vllm_sllm_models --tensor-parallel-size 4
python3 /mnt/zhengcf3/ServerlessLLM/sllm_store/examples/save_vllm_model.py --model-name deepseek-moe-16b-base --local-model-path /mnt/zhengcf3/models/deepseek-moe-16b-base --storage-path /mnt/zhengcf3/models/vllm_sllm_models --tensor-parallel-size 4
python3 /mnt/zhengcf3/ServerlessLLM/sllm_store/examples/save_vllm_model.py --model-name DeepSeek-V2-Lite --local-model-path /mnt/zhengcf3/models/DeepSeek-V2-Lite --storage-path /mnt/zhengcf3/models/vllm_sllm_models --tensor-parallel-size 4 `
python3 /mnt/zhengcf3/lmp/scripts/analyze_prefill_log.py /mnt/zhengcf3/lmp/examples/generate_cpu_sanityspread0.6_12.log