update modelwriter and operator doc

Tencent · Sep 2, 2024 · 03a186c · 03a186c
1 parent b5a3053
commit 03a186c
Show file tree

Hide file tree

Showing 2 changed files with 13 additions and 0 deletions.
diff --git a/docs/developer-guide/operators.md b/docs/developer-guide/operators.md
@@ -837,11 +837,13 @@ y = embedding(x)
 | 1         | input_dim     | int   | 0         |                   |
 | 2         | bias_term     | int   | 0         |                   |
 | 3         | weight_data_size | int | 0        |                   |
+| 18        | int8_scale_term| int  | 0         |                   |
 
 | weight        | type  | shape                 |
 | ------------- | ----- | --------------------- |
 | weight_data   | float | [weight_data_size]    |
 | bias_term     | float | [num_output]          |
+| weight_data_int8_scales| float | [1]          |
 
 # Exp
 ```

diff --git a/tools/modelwriter.h b/tools/modelwriter.h
@@ -1676,9 +1676,20 @@ int ModelWriter::save(const char* parampath, const char* binpath)
             fprintf_param_value(" 1=%d", input_dim)
             fprintf_param_value(" 2=%d", bias_term)
             fprintf_param_value(" 3=%d", weight_data_size)
+            fprintf_param_value(" 18=%d", int8_scale_term)
 
             fwrite_weight_tag_data(op->weight_data, bp);
             fwrite_weight_data(op->bias_data, bp);
+
+#if NCNN_INT8
+            // write int8_scale data
+            if (op->int8_scale_term)
+            {
+                ncnn::Mat weight_data_int8_scales(1);
+                weight_data_int8_scales[0] = op->weight_data_int8_scale;
+                fwrite_weight_data(weight_data_int8_scales, bp, 90, 100);
+            }
+#endif // NCNN_INT8
         }
         else if (layer->type == "Exp")
         {