From af4cc28de2b6dfb4d234fd41715fc8c256b55df8 Mon Sep 17 00:00:00 2001 From: hailin Date: Fri, 8 Aug 2025 21:43:40 +0800 Subject: [PATCH] . --- .../install_deepspeed_src_mamba_single_gpu.sh | 41 +++++++++++-------- 1 file changed, 23 insertions(+), 18 deletions(-) diff --git a/dspeed/install_deepspeed_src_mamba_single_gpu.sh b/dspeed/install_deepspeed_src_mamba_single_gpu.sh index f9772e7..476a951 100755 --- a/dspeed/install_deepspeed_src_mamba_single_gpu.sh +++ b/dspeed/install_deepspeed_src_mamba_single_gpu.sh @@ -173,32 +173,37 @@ log "Build & install DeepSpeed" cd "${REPO_PATH}" pip uninstall -y deepspeed >/dev/null 2>&1 || true rm -rf build + +# —— 编译前置(强烈建议)—— +python - <<'PY' +import sys, subprocess +subprocess.check_call([sys.executable, "-m", "pip", "install", "-U", "pip", "setuptools", "wheel", "ninja"]) +PY + + +# 并行编译提速(按 CPU 核心数调) +export MAX_JOBS=${MAX_JOBS:-$(nproc)} + +# —— DeepSpeed 功能开关 —— +# 编译核心 CUDA/C++ 扩展(必须) export DS_BUILD_OPS=1 -export DS_BUILD_AIO=1 -export DS_BUILD_FUSED_ADAM=1 -export DS_BUILD_CPU_ADAM=1 -export DS_BUILD_EVOFORMER_ATTN=0 -export DS_BUILD_AIO=1 +# 训练常用加速内核 +export DS_BUILD_TRANSFORMER=1 +export DS_BUILD_SPARSE_ATTN=1 +export DS_BUILD_FLASH_ATTN=1 + +# 常用优化器(强烈建议开) export DS_BUILD_FUSED_ADAM=1 export DS_BUILD_CPU_ADAM=1 -# 全部关掉和推理/transformer/CUTLASS/flash/sparse 有关的东西 -export DS_BUILD_EVOFORMER_ATTN=0 -export DS_BUILD_INFERENCE=0 -export DS_BUILD_TRANSFORMER=0 -export DS_BUILD_SPARSE_ATTN=0 -export DS_BUILD_SDP_ATTN=0 -export DS_BUILD_FLASH_ATTN=0 -export DS_BUILD_QKV_GEMM=0 -export DS_BUILD_CUTLASS=0 -unset CUTLASS_PATH +# 可选:异步 IO(Linux 才生效;Windows 不支持 AIO/GDS) +export DS_BUILD_AIO=1 -# 如需推理内核:export DS_BUILD_TRANSFORMER=1 - -# 提供更详细日志便于排错 +# 安装(带详细日志) pip install -v . + # ============================= # 7. 验证安装 # =============================