update to h100

finetune/Dockerfile

@@ -100,8 +100,8 @@ RUN <<EOT
 source /opt/conda/etc/profile.d/conda.sh
 conda activate ${CONDA_ENV_NAME}
 # 克隆 ninja 源码并编译
-git clone https://github.com/ninja-build/ninja.git 
-cd ninja
+git clone https://github.com/ninja-build/ninja.git ${STAGE_DIR}/ninja
+cd ${STAGE_DIR}/ninja
 # 克隆 GoogleTest 源码
 git clone https://github.com/google/googletest.git
 conda run -n ${CONDA_ENV_NAME} python ./configure.py --bootstrap

finetune/docker-compose_pytorch1.13.yml

@@ -12,12 +12,13 @@ services:
         TORCHAUDIO_VERSION: 0.13.1
         DS_BUILD_OPS: 1
         DS_BUILD_SPARSE_ATTN: 0
-        DS_BUILD_FUSED_ADAM: 0
+        DS_BUILD_FUSED_ADAM: 1
         DS_BUILD_CPU_ADAM: 0
         CUDA: cu117
-        CUDA_ARCH_LIST: "80;86" # for RTX 4090, all : "80;86;89;90"
+        CUDA_ARCH_LIST: "80;86;89;90" # for RTX 4090, all : "80;86;89;90"
         SETUPTOOLS_VERSION: "69.5.1"
         ROOT_PASSWD: "root"
+        DCUTLASS_NVCC_ARCHS: "90a" # 90a for H100 ,89:GeForce RTX 4090 
     volumes:
       - ./src:/bbtft
     container_name: ubuntu-finetune

finetune/docker-compose_pytorch2.3.yml

@@ -12,11 +12,12 @@ services:
         TORCHAUDIO_VERSION: 2.3.0
         DS_BUILD_OPS: 1
         DS_BUILD_SPARSE_ATTN: 0
-        DS_BUILD_FUSED_ADAM: 0
+        DS_BUILD_FUSED_ADAM: 1
         DS_BUILD_CPU_ADAM: 0
         CUDA: cu121
         CUDA_ARCH_LIST: "80;86;89;90" # for RTX 4090, all : "80;86;89;90"
         SETUPTOOLS_VERSION: "69.5.1"
+        DCUTLASS_NVCC_ARCHS: "90a" # 90a for H100 GPU 89:GeForce RTX 4090 
     volumes:
       - ./src:/bbtft
     container_name: ubuntu-finetune