{"id":528,"date":"2024-01-18T13:01:50","date_gmt":"2024-01-18T04:01:50","guid":{"rendered":"https:\/\/elosove.com\/?p=528"},"modified":"2024-01-18T13:01:50","modified_gmt":"2024-01-18T04:01:50","slug":"deepspeedv0-11-2%e3%82%92windows%e3%81%ab%e3%82%a4%e3%83%b3%e3%82%b9%e3%83%88%e3%83%bc%e3%83%ab%e3%81%99%e3%82%8b%e6%96%b9%e6%b3%95","status":"publish","type":"post","link":"https:\/\/elosove.com\/?p=528","title":{"rendered":"DeepSpeed[v0.11.2]\u3092Windows\u306b\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\u3059\u308b\u65b9\u6cd5"},"content":{"rendered":"\n

\u4eca\u56de\u306fDeepSpeed\u306ev0.11.2\u3092\u30d3\u30eb\u30c9&\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\u3059\u308b\u65b9\u6cd5\u3092\u7d39\u4ecb\u3057\u307e\u3059\u3002<\/p>\n\n\n\n

\u30d3\u30eb\u30c9\u304c\u3081\u3093\u3069\u304f\u3055\u3044\u4eba\u306fv0.11.1\u306ewhl\u30d5\u30a1\u30a4\u30eb\u304c\u516c\u958b\u3055\u308c\u3066\u3044\u308b\u306e\u3067DL\u3057\u3066pip install\u3057\u3066\u304f\u3060\u3055\u3044\u3002[deepspeed-0.11.1+e9503fe-cp311-cp311-win_amd64.rar.zip<\/a>]<\/p>\n\n\n\n

(\u53c2\u8003\u60c5\u5831 : Installing DeepSpeed on Windows<\/a>)<\/p>\n\n\n\n

(v0.8.3\u306e\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\u65b9\u6cd5\u306f\u3053\u3061\u3089<\/a>)<\/p>\n\n\n\n

\u30c6\u30b9\u30c8\u74b0\u5883<\/h2>\n\n\n\n
\n

OS : Windows 10<\/p>\n\n\n\n

CUDA : 11.8<\/p>\n\n\n\n

Visual C++ build tools : Visual Studio 2022 community<\/p>\n\n\n\n

Python : 3.11.7 (3.10\u7cfb\u3067\u3082\u52d5\u304f\u305d\u3046, Anaconda\u74b0\u5883\u4e0b)<\/p>\n\n\n\n

PyTroch : 2.1.1<\/p>\n<\/blockquote>\n\n\n\n

\u203bCUDA\u7cfb\u306e\u30d1\u30b9(CUDA_HOME \/ CUDA_PATH)\u306f\u901a\u3063\u3066\u308b\u524d\u63d0\u3067\u3059\u3002<\/p>\n\n\n\n

\u3000<\/p>\n\n\n\n

1 : DeepSpeed\u306eclone<\/h3>\n\n\n\n
\n

git clone –branch v0.11.2<\/strong> https:\/\/github.com\/microsoft\/DeepSpeed.git<\/p>\n<\/blockquote>\n\n\n\n

\u3000<\/p>\n\n\n\n

2 : build_win.bat\u306e\u7de8\u96c6<\/h3>\n\n\n\n
\n

set DS_BUILD_EVOFORMER_ATTN=0<\/p>\n<\/blockquote>\n\n\n\n

\u3092”set DS_BUILD_SPARSE_ATTN=0″\u306e\u5f8c\u308d\u306b\u8ffd\u52a0<\/p>\n\n\n\n

\u3000<\/p>\n\n\n\n

3 : \u30b3\u30fc\u30c9\u306e\u4fee\u6b63<\/h3>\n\n\n\n

\u30fbDeepSpeed\/csrc\/quantization\/pt_binding.cpp<\/strong><\/p>\n\n\n\n

244-250\u884c\u76ee\u3092\u4e0b\u8a18\u306b\u4fee\u6b63<\/p>\n\n\n\n

<\/circle><\/circle><\/circle><\/g><\/svg><\/span><\/path><\/path><\/svg><\/span>
    <\/span>std<\/span>::<\/span>vector<\/span><<\/span>int64_t<\/span>> <\/span>sz_vector<\/span>(<\/span>input_vals<\/span>.<\/span>sizes<\/span>().<\/span>begin<\/span>(), <\/span>input_vals<\/span>.<\/span>sizes<\/span>().<\/span>end<\/span>());<\/span><\/span>\n    <\/span>sz_vector<\/span>[<\/span>sz_vector<\/span>.<\/span>size<\/span>() - <\/span>1<\/span>] = <\/span>sz_vector<\/span>.<\/span>back<\/span>() \/ devices_per_node;<\/span>  \/\/ num of GPU per nodes<\/span><\/span>\n    <\/span>at<\/span>::<\/span>IntArrayRef<\/span> <\/span>sz<\/span>(<\/span>sz_vector<\/span>);<\/span><\/span>\n    <\/span>auto<\/span> output = <\/span>torch<\/span>::<\/span>empty<\/span>(sz, output_options);<\/span><\/span>\n<\/span>\n    <\/span>const<\/span> <\/span>int<\/span> elems_per_in_tensor = <\/span>at<\/span>::<\/span>numel<\/span>(input_vals) \/ devices_per_node;<\/span><\/span>\n    <\/span>const<\/span> <\/span>int<\/span> elems_per_in_group = elems_per_in_tensor \/ (in_groups \/ devices_per_node);<\/span><\/span>\n    <\/span>const<\/span> <\/span>int<\/span> elems_per_out_group = elems_per_in_tensor \/ out_groups;<\/span><\/span><\/code><\/pre><\/div>\n\n\n\n
\u3000<\/p>\n\n\n\n
\u3000<\/p>\n\n\n\n
\u30fbDeepSpeed\/csrc\/transformer\/inference\/csrc\/pt_binding.cpp<\/strong><\/p>\n\n\n\n
541-542\u884c\u76ee\u3092\u4e0b\u8a18\u306b\u4fee\u6b63<\/p>\n\n\n\n
<\/circle><\/circle><\/circle><\/g><\/svg><\/span><\/path><\/path><\/svg><\/span>
\t\t\t\t\t\t\t\t\t {<\/span>static_cast<\/span><<\/span>unsigned<\/span>>(hidden_dim * <\/span>InferenceContext<\/span>::<\/span>Instance<\/span>().<\/span>GetMaxTokenLength<\/span>()),<\/span><\/span>\n\t\t\t\t\t\t\t\t\t  <\/span>static_cast<\/span><<\/span>unsigned<\/span>>(k * <\/span>InferenceContext<\/span>::<\/span>Instance<\/span>().<\/span>GetMaxTokenLength<\/span>()),<\/span><\/span><\/code><\/pre><\/div>\n\n\n\n
550-551\u884c\u76ee\u3092\u4e0b\u8a18\u306b\u4fee\u6b63<\/p>\n\n\n\n
<\/circle><\/circle><\/circle><\/g><\/svg><\/span><\/path><\/path><\/svg><\/span>
\t\t\t\t\t\t {<\/span>static_cast<\/span><<\/span>unsigned<\/span>>(hidden_dim * <\/span>InferenceContext<\/span>::<\/span>Instance<\/span>().<\/span>GetMaxTokenLength<\/span>()),<\/span><\/span>\n\t\t\t\t\t\t  <\/span>static_cast<\/span><<\/span>unsigned<\/span>>(k * <\/span>InferenceContext<\/span>::<\/span>Instance<\/span>().<\/span>GetMaxTokenLength<\/span>()),<\/span><\/span><\/code><\/pre><\/div>\n\n\n\n
1581\u884c\u76ee\u3092\u4e0b\u8a18\u306b\u4fee\u6b63<\/p>\n\n\n\n
<\/circle><\/circle><\/circle><\/g><\/svg><\/span><\/path><\/path><\/svg><\/span>
\t\t<\/span>at<\/span>::<\/span>from_blob<\/span>(intermediate_ptr, {<\/span>input<\/span>.<\/span>size<\/span>(<\/span>0<\/span>), <\/span>input<\/span>.<\/span>size<\/span>(<\/span>1<\/span>), <\/span>static_cast<\/span><<\/span>int64_t<\/span>>(mlp_1_out_neurons)}, options);<\/span><\/span><\/code><\/pre><\/div>\n\n\n\n
\u3000<\/p>\n\n\n\n
\u30fbDeepSpeed\/deepspeed\/env_report.py<\/strong><\/p>\n\n\n\n
10\u884c\u76ee\u306b\u4e0b\u8a18\u3092\u8ffd\u52a0<\/p>\n\n\n\n
<\/circle><\/circle><\/circle><\/g><\/svg><\/span><\/path><\/path><\/svg><\/span>
import<\/span> psutil<\/span><\/span><\/code><\/pre><\/div>\n\n\n\n
\u00a083-100\u884c\u76ee\u306e\u95a2\u6570\u3092\u4e0b\u8a18\u3067\u7f6e\u304d\u63db\u3048(10\u884c\u76ee\u306b\u8ffd\u52a0\u3055\u308c\u3066\u3044\u308b\u306e\u3067\u3001\u884c\u6570\u30ba\u30ec\u306b\u6ce8\u610f)<\/p>\n\n\n\n
<\/circle><\/circle><\/circle><\/g><\/svg><\/span><\/path><\/path><\/svg><\/span>
def<\/span> <\/span>get_shm_size<\/span>():<\/span><\/span>\n    <\/span>try<\/span>:<\/span><\/span>\n        temp_dir = os.getenv(<\/span>'TEMP'<\/span>) <\/span>or<\/span> os.getenv(<\/span>'TMP'<\/span>) <\/span>or<\/span> os.path.join(os.path.expanduser(<\/span>'~'<\/span>), <\/span>'tmp'<\/span>)<\/span><\/span>\n        shm_stats = psutil.disk_usage(temp_dir)<\/span><\/span>\n        shm_size = shm_stats.total<\/span><\/span>\n        shm_hbytes = human_readable_size(shm_size)<\/span><\/span>\n        warn = []<\/span><\/span>\n        <\/span>if<\/span> shm_size < <\/span>512<\/span> * <\/span>1024<\/span>**<\/span>2<\/span>:<\/span><\/span>\n            warn.append(<\/span><\/span>\n                <\/span>f<\/span>" <\/span>{<\/span>YELLOW<\/span>}<\/span> [WARNING] Shared memory size might be too small, consider increasing it. <\/span>{<\/span>END<\/span>}<\/span>"<\/span><\/span>\n            )<\/span><\/span>\n            <\/span># Add additional warnings specific to your use case if needed.<\/span><\/span>\n        <\/span>return<\/span> shm_hbytes, warn<\/span><\/span>\n    <\/span>except<\/span> <\/span>Exception<\/span> <\/span>as<\/span> e:<\/span><\/span>\n        <\/span>return<\/span> <\/span>"UNKNOWN"<\/span>, [<\/span>f<\/span>"Error getting shared memory size: <\/span>{<\/span>e<\/span>}<\/span>"<\/span>]<\/span><\/span><\/code><\/pre><\/div>\n\n\n\n
\u3000<\/p>\n\n\n\n
4 : \u30d3\u30eb\u30c9\u5b9f\u884c<\/h3>\n\n\n\n
\u7ba1\u7406\u8005\u6a29\u9650\u3067\u30b3\u30de\u30f3\u30c9\u30d7\u30ed\u30f3\u30d7\u30c8\u3092\u958b\u304d\u3001build_win.bat\u3092\u5b9f\u884c\u3002(\u7d50\u69cb\u5f85\u3064)<\/p>\n\n\n\n
\"\"<\/figure>\n\n\n\n
\u3000<\/p>\n\n\n\n
5 : install<\/h3>\n\n\n\n
\u30d3\u30eb\u30c9\u306b\u6210\u529f\u3059\u308b\u3068dist\u30d5\u30a9\u30eb\u30c0\u306bwhl\u30d5\u30a1\u30a4\u30eb\u304c\u751f\u6210\u3055\u308c\u308b\u306e\u3067\u3001pip install\u3059\u308b\u3002<\/p>\n\n\n\n
\n
cd dist<\/p>\n\n\n\n
pip install deepspeed-0.11.2+f0604078-cp311-cp311-win_amd64.whll<\/p>\n<\/blockquote>\n\n\n\n
\u203b.whl\u306e\u30d5\u30a1\u30a4\u30eb\u540d\u306f\u74b0\u5883\u306b\u3088\u3063\u3066\u9055\u3044\u307e\u3059\u3002<\/p>\n","protected":false},"excerpt":{"rendered":"
\u4eca\u56de\u306fDeepSpeed\u306ev0.11.2\u3092\u30d3\u30eb\u30c9&\u30a4\u30f3\u30b9\u30c8\u30fc\u30eb\u3059\u308b\u65b9\u6cd5\u3092\u7d39\u4ecb\u3057\u307e\u3059\u3002 \u30d3\u30eb\u30c9\u304c\u3081\u3093\u3069\u304f\u3055\u3044\u4eba\u306fv0.11.1\u306ewhl\u30d5\u30a1\u30a4\u30eb\u304c\u516c\u958b\u3055\u308c\u3066\u3044\u308b\u306e\u3067DL\u3057\u3066pip install\u3057\u3066\u304f\u3060\u3055\u3044\u3002[dee […]<\/p>\n","protected":false},"author":1,"featured_media":526,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[4],"tags":[19],"class_list":["post-528","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-tech","tag-deepspeed"],"_links":{"self":[{"href":"https:\/\/elosove.com\/index.php?rest_route=\/wp\/v2\/posts\/528","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/elosove.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/elosove.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/elosove.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/elosove.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=528"}],"version-history":[{"count":3,"href":"https:\/\/elosove.com\/index.php?rest_route=\/wp\/v2\/posts\/528\/revisions"}],"predecessor-version":[{"id":534,"href":"https:\/\/elosove.com\/index.php?rest_route=\/wp\/v2\/posts\/528\/revisions\/534"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/elosove.com\/index.php?rest_route=\/wp\/v2\/media\/526"}],"wp:attachment":[{"href":"https:\/\/elosove.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=528"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/elosove.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=528"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/elosove.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=528"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}