update perf (#1457)

Author: smallv0221

docs/imgs/bart_perf.png

@@ -64,42 +64,42 @@ torch version 1.10.0+cu113, transformers version 4.12.5
 
 | Model Size | Decode Strategy| FasterGeneration(FP32)<br>(ms) | FasterGeneration(FP16)<br>(ms) | HF generate<br>(ms) | Speed Up Rate<br>(Faster32/HF) | Speed Up Rate<br>(Faster16/HF) |
 |-----|----|---|---|---|---|---|
-|num_layers = 6<br>num_attention_heads = 12<br>hidden_size = 768<br>(bart-base)|top_k = 1|30.08|27.95|166.90|5.55|5.97
-| |top_k = 4    |30.82|30.01|184.58|5.99|6.15 |
-| |top_k = 8    |32.06|31.05|183.44|5.72|5.91|
-| |top_k = 16   |32.66|32.35|187.14|5.73|5.78|
-| |top_p = 0.4  |37.99|30.25|208.33|5.48|6.89|
-| |num_beams = 4|45.99|37.51|285.01|5.43|7.6|
-| |num_beams = 8|50.12|37.82|316.56|6.32|8.37|
-| |num_beams = 16|67.66|40.98|467.76|6.91|11.41|
-|num_layers = 12<br>num_attention_heads = 16<br>hidden_size = 1024<br>(bart-large)|top_k = 1|50.23|39.08|222.59|4.43|5.7|
-| |top_k = 4|60.59|48.32|307.76|5.08|6.37|
-| |top_k = 8|59.67|49.65|310.49|5.20|6.25|
-| |top_k = 16|59.15|52.68|333.75|5.64|6.34|
-| |top_p = 0.4|61.36|50.83|340.74|5.55|6.7|
-| |num_beams = 4|65.60|53.24|336.28|5.12|6.32|
-| |num_beams = 8|76.20|54.13|396.62|5.20|7.33|
-| |num_beams = 16|102.04|61.11|531.92|5.21|8.7|
+|num_layers = 6<br>num_attention_heads = 12<br>hidden_size = 768<br>(bart-base)|top_k = 1|31.1|27.4|139.46|4.48|5.09
+| |top_k = 4    |32.13|29.06|149.81|4.66|5.16|
+| |top_k = 8    |31.7|28.36|154.3|4.87|5.44|
+| |top_k = 16   |32.93|28.66|145.85|4.43|5.09|
+| |top_p = 0.4  |33.35|29.01|173.18|5.19|5.97|
+| |num_beams = 4|47.55|38.02|252.71|5.31|6.65|
+| |num_beams = 8|52.19|41.39|282.3|5.41|6.82|
+| |num_beams = 16|67.18|45.82|441.59|6.57|9.64|
+|num_layers = 12<br>num_attention_heads = 16<br>hidden_size = 1024<br>(bart-large)|top_k = 1|45.8|37.43|173.08|3.78|4.62|
+| |top_k = 4|51.11|48.28|246.27|4.82|5.1|
+| |top_k = 8|61.61|50.67|246.19|4.0|4.86|
+| |top_k = 16|63.81|48.33|272.93|4.28|5.65|
+| |top_p = 0.4|63.0|50.05|288.76|4.58|5.77|
+| |num_beams = 4|65.54|48.58|273.84|4.18|5.64|
+| |num_beams = 8|75.68|52.59|340.86|4.5|6.48|
+| |num_beams = 16|102.87|62.25|477.97|4.65|7.68|
 
 **GPT:**
 
 | Model Size | Decode Strategy| FasterGeneration(FP32)<br>(ms) | FasterGeneration(FP16)<br>(ms) | HF generate<br>(ms) | Speed Up Rate<br>(Faster32/HF) | Speed Up Rate<br>(Faster16/HF) |
 |-----|----|---|---|---|---|---|
-|num_layers = 12<br>num_attention_heads = 12<br>hidden_size = 768<br>(gpt2)|top_k = 1|49.75|40.15|483.02|9.71|12.03|
-| |top_k = 4|49.70|41.69|496.63|9.99|11.91|
-| |top_k = 8|51.81|40.81|485.77|9.38|11.9|
-| |top_k = 16|50.36|42.88|488.38|9.70|11.39|
-| |top_p = 0.4|68.30|53.58|544.53|7.97|10.16|
-|num_layers = 24<br>num_attention_heads = 16<br>hidden_size = 1024<br>(gpt2-medium)|top_k = 1|109.86|76.88|936.02|8.52|12.18|
-| |top_k = 4|109.69|78.70|943.71|8.60|11.99|
-| |top_k = 8|109.70|78.39|963.73|8.79|12.29|
-| |top_k = 16|111.18|79.05|945.27|8.50|11.96|
-| |top_p = 0.4|127.54|89.76|999.28|7.83|11.13|
-|num_layers = 36<br>num_attention_heads = 20<br>hidden_size = 1280<br>(gpt2-large)|top_k = 1|205.92|142.85|1368.78|6.65|9.58|
-| |top_k = 4|205.43|140.40|1374.83|6.69|9.79|
-| |top_k = 8|205.62|139.47|1406.42|6.84|10.08|
-| |top_k = 16|205.16|139.77|1392.37|6.79|9.96|
-| |top_p = 0.4|221.06|152.35|1452.07|6.57|9.53|
+|num_layers = 12<br>num_attention_heads = 12<br>hidden_size = 768<br>(gpt2)|top_k = 1|50.84|40.37|399.58|7.86|9.9|
+| |top_k = 4|50.38|38.81|419.55|8.33|10.81|
+| |top_k = 8|51.23|36.78|411.7|8.04|11.19|
+| |top_k = 16|51.03|38.76|408.36|8.0|10.54|
+| |top_p = 0.4|68.55|48.04|489.45|7.14|10.19|
+|num_layers = 24<br>num_attention_heads = 16<br>hidden_size = 1024<br>(gpt2-medium)|top_k = 1|111.37|79.73|753.11|6.76|9.45|
+| |top_k = 4|110.53|80.48|767.48|6.94|9.54|
+| |top_k = 8|109.87|78.92|754.99|6.87|9.57|
+| |top_k = 16|110.61|85.26|764.16|6.91|8.96|
+| |top_p = 0.4|127.51|87.72|830.24|6.51|9.46|
+|num_layers = 36<br>num_attention_heads = 20<br>hidden_size = 1280<br>(gpt2-large)|top_k = 1|203.76|142.85|1108.26|5.44|7.76|
+| |top_k = 4|204.18|139.49|1230.63|6.03|8.82|
+| |top_k = 8|204.22|139.14|1238.96|6.07|8.9|
+| |top_k = 16|204.11|140.04|1148.05|5.62|8.2|
+| |top_p = 0.4|222.12|150.68|1248.75|5.62|8.29|
 
 
 ## 测试方法