Merge pull request bytedance#68 from bytedance/move-inference-code

separate inference and training

Author: godweiyang

.gitignore

@@ -141,3 +141,6 @@ cython_debug/
 test_case/
 build/
 *.pb
+.vscode
+*.bin
+*.idx

CMakeLists.txt

@@ -47,11 +47,11 @@ if(DEBUG_MODE)
     message(STATUS "Debug computation result")
 endif()
 
-add_subdirectory(kernels)
-add_subdirectory(tools)
-add_subdirectory(proto)
-add_subdirectory(model)
-add_subdirectory(example)
+add_subdirectory(lightseq/inference/kernels)
+add_subdirectory(lightseq/inference/tools)
+add_subdirectory(lightseq/inference/proto)
+add_subdirectory(lightseq/inference/model)
+add_subdirectory(examples/inference/cpp)
 add_subdirectory(3rdparty/pybind11)
-add_subdirectory(pywrapper)
-add_subdirectory(server)
+add_subdirectory(lightseq/inference/pywrapper)
+add_subdirectory(lightseq/inference/server)

README.md

@@ -8,8 +8,8 @@
 from fairseq import tasks, distributed_utils
 from fairseq.logging import metrics
 
-from ds_fairseq_data import BatchIterator
-from ds_fairseq_argument import gen_ds_fairseq_arg
+from examples.training.deepspeed.ds_fairseq_data import BatchIterator
+from examples.training.deepspeed.ds_fairseq_argument import gen_ds_fairseq_arg
 
 
 best_bleu = 0.0

docs/build.md renamed to docs/inference/build.md

@@ -1,17 +1,17 @@
 #!/usr/bin/env bash
 set -ex
 THIS_DIR=$(dirname $(readlink -f $0))
-cd $THIS_DIR/../..
+cd $THIS_DIR/../../..
 
-if [ ! -d "wmt14_en_de" ]; then
+if [ ! -d "/tmp/wmt14_en_de" ]; then
     echo "Downloading dataset"
-    wget http://sf3-ttcdn-tos.pstatp.com/obj/nlp-opensource/lightseq/wmt_data/databin_wmt14_en_de.tar.gz
-    tar -zxvf databin_wmt14_en_de.tar.gz && rm databin_wmt14_en_de.tar.gz
+    wget http://sf3-ttcdn-tos.pstatp.com/obj/nlp-opensource/lightseq/wmt_data/databin_wmt14_en_de.tar.gz -P /tmp
+    tar -zxvf /tmp/databin_wmt14_en_de.tar.gz -C /tmp && rm /tmp/databin_wmt14_en_de.tar.gz
 fi
 
 deepspeed ${THIS_DIR}/ds_fairseq.py \
-    ./wmt14_en_de/ \
-    --user-dir ./examples/fairseq/fs_modules \
+    /tmp/wmt14_en_de/ \
+    --user-dir  ${THIS_DIR}/../fairseq/fs_modules \
     --arch ls_transformer_wmt_en_de_big_t2t --share-decoder-input-output-embed \
     --optimizer ls_adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 \
     --lr 5e-4 --lr-scheduler inverse_sqrt --warmup-updates 4000 \

docs/export_model.md renamed to docs/inference/export_model.md

@@ -1,15 +1,15 @@
 #!/usr/bin/env bash
 set -ex
 THIS_DIR=$(dirname $(readlink -f $0))
-cd $THIS_DIR/../../
+cd $THIS_DIR/../../..
 
-if [ ! -d "wmt14_en_de" ]; then
+if [ ! -d "/tmp/wmt14_en_de" ]; then
     echo "Downloading dataset"
-    wget http://sf3-ttcdn-tos.pstatp.com/obj/nlp-opensource/lightseq/wmt_data/databin_wmt14_en_de.tar.gz
-    tar -zxvf databin_wmt14_en_de.tar.gz && rm databin_wmt14_en_de.tar.gz
+    wget http://sf3-ttcdn-tos.pstatp.com/obj/nlp-opensource/lightseq/wmt_data/databin_wmt14_en_de.tar.gz -P /tmp
+    tar -zxvf /tmp/databin_wmt14_en_de.tar.gz -C /tmp && rm /tmp/databin_wmt14_en_de.tar.gz
 fi
 
-lightseq-train ./wmt14_en_de/ \
+lightseq-train /tmp/wmt14_en_de/ \
     --task translation \
     --arch ls_transformer_wmt_en_de_big_t2t --share-decoder-input-output-embed \
     --optimizer ls_adam --adam-betas '(0.9, 0.98)' \

docs/images/generation.png renamed to docs/inference/images/generation.png

@@ -1,14 +1,15 @@
 #!/usr/bin/env bash
 set -ex
 THIS_DIR=$(dirname $(readlink -f $0))
-cd $THIS_DIR/../../
+cd $THIS_DIR/../../..
 
-if [ ! -d "wmt14_en_de" ]; then
-    wget http://sf3-ttcdn-tos.pstatp.com/obj/nlp-opensource/lightseq/wmt_data/databin_wmt14_en_de.tar.gz
-    tar -zxvf databin_wmt14_en_de.tar.gz && rm databin_wmt14_en_de.tar.gz
+if [ ! -d "/tmp/wmt14_en_de" ]; then
+    echo "Downloading dataset"
+    wget http://sf3-ttcdn-tos.pstatp.com/obj/nlp-opensource/lightseq/wmt_data/databin_wmt14_en_de.tar.gz -P /tmp
+    tar -zxvf /tmp/databin_wmt14_en_de.tar.gz -C /tmp && rm /tmp/databin_wmt14_en_de.tar.gz
 fi
 
-fairseq-train ./wmt14_en_de/ \
+fairseq-train /tmp/wmt14_en_de/ \
     --arch transformer_wmt_en_de_big_t2t --share-decoder-input-output-embed \
     --optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 \
     --lr 5e-4 --lr-scheduler inverse_sqrt --warmup-updates 4000 \