Update README.md
Browse files
README.md
CHANGED
|
@@ -19,15 +19,35 @@ This is an INT8 PyTorch model quantized with [huggingface/optimum-intel](https:
|
|
| 19 |
|
| 20 |
The original fp32 model comes from the fine-tuned model [sysresearch101/t5-large-finetuned-xsum-cnn](https://huggingface.co/sysresearch101/t5-large-finetuned-xsum-cnn).
|
| 21 |
|
| 22 |
-
Below linear modules are fallbacked to fp32 for less than 1% relative accuracy loss:
|
| 23 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24 |
|
| 25 |
### Evaluation result
|
| 26 |
|
| 27 |
| |INT8|FP32|
|
| 28 |
|---|:---:|:---:|
|
| 29 |
-
| **Accuracy (eval-rougeLsum)** | 41.
|
| 30 |
-
| **Model size** |
|
| 31 |
|
| 32 |
### Load with optimum:
|
| 33 |
|
|
|
|
| 19 |
|
| 20 |
The original fp32 model comes from the fine-tuned model [sysresearch101/t5-large-finetuned-xsum-cnn](https://huggingface.co/sysresearch101/t5-large-finetuned-xsum-cnn).
|
| 21 |
|
| 22 |
+
Below linear modules (40/193) are fallbacked to fp32 for less than 1% relative accuracy loss:
|
| 23 |
+
|
| 24 |
+
**'model.decoder.layers.10.fc1'**, **'model.decoder.layers.0.fc2'**,
|
| 25 |
+
**'model.decoder.layers.4.fc2'**, **'model.decoder.layers.1.fc2'**,
|
| 26 |
+
**'model.decoder.layers.6.fc2'**, **'model.decoder.layers.2.fc2'**,
|
| 27 |
+
**'model.decoder.layers.3.fc2'**, **'model.encoder.layers.11.fc2'**,
|
| 28 |
+
**'model.decoder.layers.9.fc1'**, **'model.decoder.layers.5.fc2'**,
|
| 29 |
+
**'model.decoder.layers.7.fc1'**, **'model.decoder.layers.8.fc1'**,
|
| 30 |
+
**'model.encoder.layers.0.fc2'**, **'model.decoder.layers.11.fc1'**,
|
| 31 |
+
**'model.encoder.layers.8.fc2'**, **'model.encoder.layers.11.fc1'**,
|
| 32 |
+
**'model.decoder.layers.8.fc2'**, **'model.decoder.layers.2.fc1'**,
|
| 33 |
+
**'model.decoder.layers.11.self_attn.v_proj'**, **'model.encoder.layers.9.fc1'**,
|
| 34 |
+
**'model.decoder.layers.9.fc2'**, **'model.decoder.layers.7.fc2'**,
|
| 35 |
+
**'model.decoder.layers.6.fc1'**, **'model.decoder.layers.0.fc1'**,
|
| 36 |
+
**'model.decoder.layers.1.self_attn.v_proj'**, **'model.encoder.layers.3.fc1'**,
|
| 37 |
+
**'model.encoder.layers.2.fc2'**, **'model.encoder.layers.7.fc2'**,
|
| 38 |
+
**'model.decoder.layers.3.fc1'**, **'model.encoder.layers.1.fc2'**,
|
| 39 |
+
**'model.encoder.layers.10.fc2'**, **'model.encoder.layers.8.fc1'**,
|
| 40 |
+
**'lm_head'**, **'model.decoder.layers.6.self_attn.v_proj'**,
|
| 41 |
+
**'model.decoder.layers.11.self_attn.out_proj'**, **'model.decoder.layers.11.encoder_attn.v_proj'**,
|
| 42 |
+
**'model.encoder.layers.10.fc1'**, **'model.encoder.layers.6.fc1'**,
|
| 43 |
+
**'model.decoder.layers.4.fc1'**, **'model.decoder.layers.1.fc1'**
|
| 44 |
|
| 45 |
### Evaluation result
|
| 46 |
|
| 47 |
| |INT8|FP32|
|
| 48 |
|---|:---:|:---:|
|
| 49 |
+
| **Accuracy (eval-rougeLsum)** | 41.2224 | 41.5274 |
|
| 50 |
+
| **Model size** |625M|1669M|
|
| 51 |
|
| 52 |
### Load with optimum:
|
| 53 |
|