add range

Author: tp-nan

docs/backend-reference/basic.mdx

@@ -15,6 +15,7 @@ Overview
 | [`Identity`](#identity)                         | None                             | `data/any`              | `result/any`             | `dict["result"]=dict["data"]` | Default backend |
 | [`Sequential`](#sequential)                     | `Sequential::backend`            | `data/any`              |                          |                               |                 |
 | [`Jump`](../Inter-node/graphtraversal.mdx#jump) | `jump`                           |                         |                          | Jump to other (root) node     |                 |
+| [`Range`](#range)                     | `Range::backend`;`range`            | `data/any`              |                          |                               |                 |
 
 
 
@@ -68,4 +69,15 @@ jump="||node_name_1||node_name_2"
 When forwarding, sub-backends are executed in series. If the result of a non-final sub-backend is empty, the execution is terminated and an exception is thrown. If the result of a non-final sub-backend is non-empty, the filter will be executed. The default is `swap`.
 
 ### Input Range
-If the input data range of all sub-backends is [1,1], then the range of `Sequential` is [1,1]. Otherwise, it is the union of the input ranges of sub-backends whose input ranges are not [1,1]. **Currently, only one sub-backend is allowed to satisfy `max() != 1 && max() != UINT32_MAX` at most**. In `Sequential`, currently only one sub-backend with non-trivial input range is allowed at most, and this sub-backend can be regarded as the main functional implementer of the entire `Sequential`, while other backends are auxiliary or functional backends.
+If the input data range of all sub-backends is [1,1], then the range of `Sequential` is [1,1]. Otherwise, it is the union of the input ranges of sub-backends whose input ranges are not [1,1]. **Currently, only one sub-backend is allowed to satisfy `max() != 1 && max() != UINT32_MAX` at most**. In `Sequential`, currently only one sub-backend with non-trivial input range is allowed at most, and this sub-backend can be regarded as the main functional implementer of the entire `Sequential`, while other backends are auxiliary or functional backends.
+
+## Range
+Set a new maximum and minimum range [target_min, target_max].
+Use dynamic programming to determine whether all the integer 'target' in [target_min, target_max] can be represented as the sum
+of multiple numbers within `Range::backend`'s [min, max] range, and provide a result with the largest possible values. 
+### Initialization
+:::tip Parameters
+- **range** - Target [min, max]. for example, `range="1,10"`.
+- **Range::backend** - Proxy backend.
+:::
+Initialization fails if the new range cannot be represented by the range in the old range. For example, `target=10, [min, max]=[2,9]` => `10 = 8+2`. And the following situations will fail: `range=[10,10], [min, max]=[8,8]` => `10 != 8`.

docs/showcase/showcase.mdx

@@ -14,6 +14,7 @@ slug: /showcase
 |         [yolox]          | ![](../assets/yolox.svg)               | [TensorrtTensor](../backend-reference/torch.mdx#tensorrttensor)<br />[torchpipe.utils.cpp_extension.load] |      |
 |        [PP-OCRv2]        | ![](../assets/ppocr.svg)               | [MapReduce](../Inter-node/graphtraversal.mdx#mapreduce)<br />[Jump]                                       |      |
 | [tensorrt's native int8] |                                        | [TensorrtTensor](../backend-reference/torch.mdx#tensorrttensor)                                           |      |
+| [Llama] |                                        | [Llama](https://github.com/torchpipe/LLM.TensorRT.Serve)                                           |      |
 
 [resnet18]: https://github.com/torchpipe/torchpipe/tree/main/examples/resnet18
 [yolox]: https://github.com/torchpipe/torchpipe/tree/main/examples/yolox

docs/tools/quantization.mdx

@@ -11,7 +11,7 @@ The post-training quantization process of TensorRT mainly includes two steps:
 - Prepare quantization data, about 500 pieces. We will save the data in PyTorch format in the backend.
 - Launch the quantization process with the prepared data and generate the model.
 
-For more details, please refer to the [example](https://github.com/torchpipe/torchpipe//examples/int8/README_en.md).
+For more details, please refer to the [example](https://github.com/torchpipe/torchpipe/tree/develop/examples/int8).
 
 ## Training-based Quantization (based on pytorch_quantization)
 In the [official notebook](https://github.com/NVIDIA/TensorRT/blob/release/8.6/quickstart/quantization_tutorial/qat-ptq-workflow.ipynb), NVIDIA summarizes how to improve quantization accuracy through training-based quantization in PyTorch.

i18n/zh/docusaurus-plugin-content-docs/current/backend-reference/basic.mdx

@@ -17,6 +17,7 @@ displayed_sidebar: api
 | [`Identity`](#identity)                         | 无                      | `data/any`         | `result/any`       | `dict["result"]=dict["data"]` | 默认后端 |
 | [`Sequential`](#sequential)                     | `Sequential::backend`   | `data/any`         |                    |                               |          |
 | [`Jump`](../Inter-node/graphtraversal.mdx#jump) | `jump`                  |                    |                    | 跳转到其他（根）节点            |          |
+| [`Range`](#range)                     | `Range::backend`;`range`            | `data/any`              |                          |                               |                 |
 
 
 
@@ -91,4 +92,15 @@ jump="||node_name_1||node_name_2"
 
 :::caution
 被提升输入范围的后端的前向接口必须是线程安全的，并且状态不依赖于线程本身。[torch相关后端](./torch.mdx)不被支持
-::: -->
+::: -->
+
+## Range
+设定新的最大最小范围 [target_min, target_max]。
+使用动态规划来确定 [target_min, target_max] 范围内的所有整数 'target' 是否可以表示为 `Range::backend` 的 [min, max] 范围内的多个数字之和，并提供可能的最大值结果。
+### 初始化
+:::tip 参数
+- **range** - 目标 [min, max]。例如，`range="1,10"`。
+- **Range::backend** - 代理的后端。
+:::
+如果新的范围无法被旧的范围中的范围表示，则初始化失败。例如，`range=[10,10], [min, max]=[2,9]` => `10 = 8+2`。
+而以下情况会失败：`range=[10,10], [min, max]=[8,8]` => `10 != 8`。

i18n/zh/docusaurus-plugin-content-docs/current/tools/quantization.mdx

@@ -10,7 +10,7 @@ tensorrt 的训练后量化过程主要包含两步：
 - 准备量化数据，500份左右。这部分我们将进入后端的数据按照Pytorch的格式保存下来
 - 以准备好的数据启动量化过程，并生成模型。
 
-详细可参见[示例](https://github.com/torchpipe/torchpipe//examples/int8).
+详细可参见[示例](https://github.com/torchpipe/torchpipe/tree/develop/examples/int8).
 
 ## 训练时量化（基于pytorch_quantization）
 在[官方notebook](https://github.com/NVIDIA/TensorRT/blob/release/8.6/quickstart/quantization_tutorial/qat-ptq-workflow.ipynb)中，NVIDIA总结了如何在Pytorch中通过训练时量化提升量化精度。