From 455b634bfeee0221845ab95d029f0a4475d35dbc Mon Sep 17 00:00:00 2001
From: Your Name <you@example.com>
Date: Tue, 2 Jul 2024 02:45:55 +0000
Subject: [PATCH] update add device and id_rsa

---
 finetune/docker-compose_pytorch2.3_device.yml | 69 +++++++++++++++++++
 1 file changed, 69 insertions(+)
 create mode 100644 finetune/docker-compose_pytorch2.3_device.yml

diff --git a/finetune/docker-compose_pytorch2.3_device.yml b/finetune/docker-compose_pytorch2.3_device.yml
new file mode 100644
index 0000000..5211bca
--- /dev/null
+++ b/finetune/docker-compose_pytorch2.3_device.yml
@@ -0,0 +1,69 @@
+version: '3.8'
+
+# DeepSpeed支持多种C++/CUDA扩展（ops），这些ops旨在优化深度学习的训练和推理过程。以下是一些主要的DeepSpeed ops及其功能：
+
+# FusedAdam - 提供融合优化的Adam优化器，适用于GPU。
+# FusedLamb - 类似FusedAdam，针对LAMB优化器，适用于大规模分布式训练。
+# SparseAttention - 用于高效计算稀疏注意力机制。
+# Transformer - 提供Transformer模型的高效实现。
+# TransformerInference - 专门用于Transformer模型的推理优化。
+# CPUAdam - 针对CPU优化的Adam优化器。
+# CPULion - 针对CPU的Lion优化器。
+# Quantizer - 提供量化支持，以减少模型大小和提高推理速度。
+# RandomLTD - 用于随机层裁剪的优化器。
+# StochasticTransformer - 支持随机Transformer模型的训练和推理。
+
+services:
+  ubuntu-finetune:
+    build: 
+      context: .
+      dockerfile: Dockerfile
+      args: # PyTorch版本、Python版本与pytorch_lightning版本的对应关系表 https://blog.csdn.net/qq_41813454/article/details/137421822
+        PYTHON_VERSION: "3.10"
+        CUDA_VERSION: "12.1.0"
+        PYTORCH_VERSION: "2.3.0"
+        TORCHVISION_VERSION: "0.18.0"
+        TORCHAUDIO_VERSION: "2.3.0"
+        DS_BUILD_OPS: 1
+        USE_CUDA: 1
+        USE_ROCM: 0
+        USE_XPU: 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: "80;86;89;90;90a" # 90a for H100 GPU 89:GeForce RTX 4090 
+        DEEPSPEED_VERSION: "master"
+        DEEPSPEED_INSTALL_FLAGS: "--allow_sudo"
+    volumes:
+      - ./src:/bbtft
+      - ./id_rsa_finetune:/root/.ssh/id_rsa
+      - ./id_rsa.pub:/root/.ssh/id_rsa.pub
+    container_name: ubuntu-finetune
+    pull_policy: if_not_present
+    # tty: true
+    restart: unless-stopped
+    image: hotwa/deepspeed:pt23
+    shm_size: '32gb'
+    ports:
+      - 3228:22
+    environment:
+      - NVIDIA_VISIBLE_DEVICES=all
+      - NVIDIA_DRIVER_CAPABILITIES=compute,utility
+      - TMPDIR=/var/tmp
+    networks:
+      - network_finetune
+    deploy:
+      resources:
+        reservations:
+          devices:
+            - driver: nvidia
+              count: all
+              capabilities: [gpu]
+    cap_add:
+      - IPC_LOCK
+    devices:
+      - /dev/infiniband:/dev/infiniband
+
+networks:
+  network_finetune:
+    name: network_finetune