Fairseq v0.10.2 compatible

README.md

@@ -16,17 +16,17 @@ Below shows the generation speed gain by using FastSeq.
 
 | Model            | W/O FastSeq (in samples/s) | W/ FastSeq (in samples/s) | Speedup |
 |------------------|:--------------------------:|:-------------------------:|:-----:|
-| [ProphetNet](examples/prophetnet/README.md)       | 2.8 | 10.7  | 3.8x  |
-| [Bart (`fs`)](examples/bart/README.md)              | 2.4  | 25.3 | 10.5x  |
+| [ProphetNet](examples/prophetnet/README.md)       | 2.8 | 11.9  | 4.3  |
+| [Bart (`fs`)](examples/bart/README.md)              | 3.3  | 25.1 | 7.7x  |
 | [Bart (`hf`)](examples/bart/README.md#speedup-bart-huggingface-transformers-version-by-using-fastseq) | 2.5 | 12.4 | 5.0x  |
 | [DistilBart (`hf`)](examples/distilbart/README.md)    | 3.4  | 18.5  | 5.4x  |
 | [T5 (`hf`)](examples/t5/README.md)                  | 8.7  | 31.3  | 3.6x  |
-| [WMT16 En-De (`fs`)](examples/wmt/README.md)        | 96.0   | 417.0  | 4.3x  |
+| [WMT16 En-De (`fs`)](examples/wmt/README.md)        | 144.5   | 422.8  | 2.9x  |
 | [GPT2 (`hf`)](examples/gpt2/README.md)        | 3.0   | 16.7  | 5.5x  |
 | [UniLM (`hf`)](examples/unilm/README.md)        | 1.7   | 16.4  | 9.6x  |
 
 - All benchmarking experiments run on NVIDIA-V100-16GB with [docker](docker/Dockerfile). Highest speed recorded for each model by tuning batch size. For parameter setting details, click link of corresponding model.
-- `fs` stands for [Fairseq](https://github.com/pytorch/fairseq) 0.9.0 version, `hf` stands for [Huggingface Transformers](https://github.com/huggingface/transformers) 3.0.2 version.
+- `fs` stands for [Fairseq](https://github.com/pytorch/fairseq) 0.10.2 version, `hf` stands for [Huggingface Transformers](https://github.com/huggingface/transformers) 3.0.2 version.
 - Optimizations were automatically applied to all generation/sequence models in Fairseq & Huggingface Transformers. Above only lists a subset of them.
 
 ## How it works?

azure-pipelines.yml

@@ -15,13 +15,29 @@ jobs:
       demands:
         - agent.name -equals gpu3
     container:
-      image: adsbrainwestus2.azurecr.io/fastseq:dev-py3
+      image: adsbrainwestus2.azurecr.io/fastseq:dev-py3 
       endpoint: fastseq-acr
       options: --gpus device=3
     steps:
     - script: |
         #install fastseq
-        pip install --editable .[transformers,fairseq]
+        which pip
+        which python 
+
+        echo "******* Installing fairseq *******"
+        pip install fairseq==0.10.2
+        pip show fairseq
+
+        echo "******* Installing transformers *******"
+        pip install transformers
+        pip show transformers
+
+        echo "******* Installing fastseq *******"
+        pip install --editable .
+        pip show fastseq
+
+        echo "******* Adding local bin to path *******"
+        export PATH="$HOME/bin:$HOME/.local/bin:$PATH"
 
         echo "******* Running fastseq unittests *******"
         pip install pytorch-transformers==1.0.0

benchmarks/benchmark_fs.sh

@@ -112,6 +112,7 @@ for bs in "${bs_list[@]}"; do
             --no-repeat-ngram-size 3 \
             --lenpen 2.0 \
             --use-el-attn \
+            --required-seq-len-multiple 8 \
 	    `#--print-alignment` \
             `#--print-step      # KeyError: steps` \
             --skip-invalid-size-inputs-valid-test $* \
@@ -132,6 +133,7 @@ for bs in "${bs_list[@]}"; do
             --max-len-b 140 \
             --no-repeat-ngram-size 3 \
             --lenpen 2.0 \
+            --required-seq-len-multiple 8 \
 	    `#--print-alignment` \
 	    `#--print-step	# KeyError: steps` \
             --skip-invalid-size-inputs-valid-test $* \
@@ -140,7 +142,7 @@ for bs in "${bs_list[@]}"; do
     ret=$?
     end=`date +%s`
     runtime=$(($end-$start))
-    tail=`tail -2 $STDOUT_FILE`
+    tail=`tail -3 $STDOUT_FILE`
     if [[ $ret -eq 0 &&  $tail == *$mark1* ]]; then
         samples=`echo $tail | sed 's/.*Translated \([0-9]*\) sentences.*/\1/'`
         tokens=`echo $tail | sed 's/.*Translated .* sentences (\([0-9]*\) tokens).*/\1/'`

benchmarks/models/fs_bart.sh

@@ -34,21 +34,21 @@ grep "bart.large.cnn cnn_dm/len-1024.bin valid " perf \
 	| awk '{if($8!="NA"){c+=1;s+=$8}}END{print s/c}' \
 	| ./range.sh 17.9 18
 # Speed on V100 16GB 250W
-grep -E "fairseq_v0.9.0 bart.large.cnn cnn_dm/len-1024.bin valid 32 " perf \
+grep -E "fairseq_v0.10.2 bart.large.cnn cnn_dm/len-1024.bin valid 32 " perf \
 	| awk '{s+=$13}END{if(NR==0) print -1; else print s/NR}' \
-	| ./range.sh 2.1 2.7
-grep -E "fairseq_v0.9.0\+fastseq_v.* bart.large.cnn cnn_dm/len-1024.bin valid 32 " perf \
+	| ./range.sh 3.1 3.7
+grep -E "fairseq_v0.10.2\+fastseq_v.* bart.large.cnn cnn_dm/len-1024.bin valid 32 " perf \
 	| awk '{s+=$13}END{print s/NR}' \
 	| ./range.sh 7.8 100
-grep -E "fairseq_v0.9.0\+fastseq_v.* bart.large.cnn cnn_dm/len-1024.bin valid 64 " perf \
+grep -E "fairseq_v0.10.2\+fastseq_v.* bart.large.cnn cnn_dm/len-1024.bin valid 64 " perf \
 	| awk '{s+=$13}END{print s/NR}' \
 	| ./range.sh 13.0 100
-grep -E "fairseq_v0.9.0\+fastseq_v.* bart.large.cnn cnn_dm/len-1024.bin valid 128 " perf \
+grep -E "fairseq_v0.10.2\+fastseq_v.* bart.large.cnn cnn_dm/len-1024.bin valid 128 " perf \
 	| awk '{s+=$13}END{print s/NR}' \
 	| ./range.sh 18.1 100
-grep -E "fairseq_v0.9.0\+fastseq_v.* bart.large.cnn cnn_dm/len-1024.bin valid 256 " perf \
+grep -E "fairseq_v0.10.2\+fastseq_v.* bart.large.cnn cnn_dm/len-1024.bin valid 256 " perf \
 	| awk '{s+=$13}END{print s/NR}' \
 	| ./range.sh 19 100
-grep -E "fairseq_v0.9.0\+fastseq_v.* bart.large.cnn cnn_dm/len-1024.bin valid 320 " perf \
+grep -E "fairseq_v0.10.2\+fastseq_v.* bart.large.cnn cnn_dm/len-1024.bin valid 320 " perf \
         | awk '{s+=$13}END{print s/NR}' \
-        | ./range.sh 25 100
+        | ./range.sh 24.5 100

benchmarks/models/fs_prophetnet.sh

@@ -8,19 +8,20 @@
 #   <batch-sizes>
 source utils.sh
 
+# TODO: update when ProphetNet is compatible with fairseq 0.10.2
 # Download ProphetNet repo as the baseline if it does not exist
-prophetnet_repo_path=$CACHE_DIR/ProphetNet
-git_clone_if_not_in_cache \
-    https://github.com/microsoft/ProphetNet.git \
-    $prophetnet_repo_path
-
-./benchmark.sh \
-    fairseq \
-    prophetnet_large_160G_cnndm_model \
-    cnn_dm_bert/len-512.bin \
-    valid \
-    64 \
-    --user-dir $prophetnet_repo_path/src/prophetnet/
+# prophetnet_repo_path=$CACHE_DIR/ProphetNet
+# git_clone_if_not_in_cache \
+#     https://github.com/microsoft/ProphetNet.git \
+#     $prophetnet_repo_path
+#
+# ./benchmark.sh \
+#     fairseq \
+#     prophetnet_large_160G_cnndm_model \
+#     cnn_dm_bert/len-512.bin \
+#     valid \
+#     64 \
+#     --user-dir $prophetnet_repo_path/src/prophetnet/
 ./benchmark.sh \
     fairseq+fastseq \
     prophetnet_large_160G_cnndm_model \
@@ -33,18 +34,19 @@ grep "prophetnet_large_160G_cnndm_model cnn_dm_bert/len-512.bin valid" perf \
 	| awk '{if($8!="NA"){c+=1;s+=$8}}END{print s/c}' \
 	| ./range.sh 19.1 19.2
 # # Speed on V100 16GB 250W
-grep -E "fairseq_v0.9.0 prophetnet_large_160G_cnndm_model cnn_dm_bert/len-512.bin valid 32 " perf \
-	| awk '{s+=$13}END{if(NR==0) print -1; else print s/NR}' \
-	| ./range.sh 2 3
-grep -E "fairseq_v0.9.0 prophetnet_large_160G_cnndm_model cnn_dm_bert/len-512.bin valid 64 " perf \
-	| awk '{s+=$13}END{if(NR==0) print -1; else print s/NR}' \
-	| ./range.sh 2 3
-grep -E "fairseq_v0.9.0\+fastseq_v.* prophetnet_large_160G_cnndm_model cnn_dm_bert/len-512.bin valid 32 " perf \
+# TODO: update when ProphetNet is compatible with fairseq 0.10.2
+# grep -E "fairseq_v0.10.2 prophetnet_large_160G_cnndm_model cnn_dm_bert/len-512.bin valid 32 " perf \
+# 	| awk '{s+=$13}END{if(NR==0) print -1; else print s/NR}' \
+# 	| ./range.sh 2 3
+# grep -E "fairseq_v0.10.2 prophetnet_large_160G_cnndm_model cnn_dm_bert/len-512.bin valid 64 " perf \
+# 	| awk '{s+=$13}END{if(NR==0) print -1; else print s/NR}' \
+# 	| ./range.sh 2 3
+grep -E "fairseq_v0.10.2\+fastseq_v.* prophetnet_large_160G_cnndm_model cnn_dm_bert/len-512.bin valid 32 " perf \
 	| awk '{s+=$13}END{if(NR==0) print -1; else print s/NR}' \
 	| ./range.sh 5.7 6.5
-grep -E "fairseq_v0.9.0\+fastseq_v.* prophetnet_large_160G_cnndm_model cnn_dm_bert/len-512.bin valid 64 " perf \
+grep -E "fairseq_v0.10.2\+fastseq_v.* prophetnet_large_160G_cnndm_model cnn_dm_bert/len-512.bin valid 64 " perf \
 	| awk '{s+=$13}END{print s/NR}' \
-	| ./range.sh 7.5 10
-grep -E "fairseq_v0.9.0\+fastseq_v.* prophetnet_large_160G_cnndm_model cnn_dm_bert/len-512.bin valid 128 " perf \
+	| ./range.sh 8 10.5
+grep -E "fairseq_v0.10.2\+fastseq_v.* prophetnet_large_160G_cnndm_model cnn_dm_bert/len-512.bin valid 128 " perf \
 	| awk '{s+=$13}END{print s/NR}'\
 	| ./range.sh 10 15

benchmarks/models/fs_wmt.sh

@@ -27,15 +27,15 @@ grep " wmt16.en.de.32k wmt16_en_de_bpe32k/bin valid " perf \
 	| awk '{if($8!="NA"){c+=1;s+=$8}}END{print s/c}' \
 	| ./range.sh 0.05 0.07
 # Speed on V100 16GB 250W
-grep -E "fairseq_v0.9.0 wmt16.en.de.32k wmt16_en_de_bpe32k/bin valid 256 " perf \
+grep -E "fairseq_v0.10.2 wmt16.en.de.32k wmt16_en_de_bpe32k/bin valid 256 " perf \
 	| awk '{s+=$13}END{if(NR==0) print -1; else print s/NR}' \
-	| ./range.sh 93 100
-grep -E "fairseq_v0.9.0\+fastseq_v.* wmt16.en.de.32k wmt16_en_de_bpe32k/bin valid 256 " perf \
+	| ./range.sh 100 150
+grep -E "fairseq_v0.10.2\+fastseq_v.* wmt16.en.de.32k wmt16_en_de_bpe32k/bin valid 256 " perf \
 	| awk '{s+=$13}END{print s/NR}' \
 	| ./range.sh 350 1000
-grep -E "fairseq_v0.9.0\+fastseq_v.* wmt16.en.de.32k wmt16_en_de_bpe32k/bin valid 512 " perf \
+grep -E "fairseq_v0.10.2\+fastseq_v.* wmt16.en.de.32k wmt16_en_de_bpe32k/bin valid 512 " perf \
 	| awk '{s+=$13}END{print s/NR}' \
 	| ./range.sh 390 1000
-grep -E "fairseq_v0.9.0\+fastseq_v.* wmt16.en.de.32k wmt16_en_de_bpe32k/bin valid 1024 " perf \
+grep -E "fairseq_v0.10.2\+fastseq_v.* wmt16.en.de.32k wmt16_en_de_bpe32k/bin valid 1024 " perf \
 	| awk '{s+=$13}END{print s/NR}' \
 	| ./range.sh 405 1000

docker/Dockerfile

@@ -49,7 +49,7 @@ RUN pip install --upgrade pip && \
 	pip install requests>=v2.24.0 && \
 	pip install gitpython>=v3.1.7 && \
 	pip install rouge_score==v0.0.4 && \
-	pip install fairseq==v0.9.0 && \
+	pip install fairseq==v0.10.2 && \
 	pip install transformers==v3.0.2 && \
 	pip install pytorch-transformers==1.0.0