jiazhizhong/higress
1
0
Fork 0
mirror of https://github.com/alibaba/higress.git synced 2026-06-09 12:47:28 +08:00
Code Issues Packages Projects Releases Wiki Activity

optimize model router&mapper (#1866)

Browse Source
This commit is contained in:
澄潭
2025-03-09 23:07:49 +08:00
committed by GitHub
parent 9e6bd6d2cc
commit a0c334a7cb
4 changed files with 35 additions and 28 deletions
Download Patch File Download Diff File Expand all files Collapse all files

5
plugins/wasm-cpp/extensions/model_mapper/plugin.cc
View File

@@ -44,8 +44,8 @@ static RegisterContextFactory register_ModelMapper(
namespace { namespace {
constexpr std::string_view SetDecoderBufferLimitKey = constexpr std::string_view SetDecoderBufferLimitKey =
"SetRequestBodyBufferLimit"; "set_decoder_buffer_limit";
constexpr std::string_view DefaultMaxBodyBytes = "10485760"; constexpr std::string_view DefaultMaxBodyBytes = "104857600";
} // namespace } // namespace
@@ -166,6 +166,7 @@ FilterHeadersStatus PluginRootContext::onHeader(
} }
removeRequestHeader(Wasm::Common::Http::Header::ContentLength); removeRequestHeader(Wasm::Common::Http::Header::ContentLength);
setFilterState(SetDecoderBufferLimitKey, DefaultMaxBodyBytes); setFilterState(SetDecoderBufferLimitKey, DefaultMaxBodyBytes);
LOG_INFO(absl::StrCat("SetRequestBodyBufferLimit: ", DefaultMaxBodyBytes));
return FilterHeadersStatus::StopIteration; return FilterHeadersStatus::StopIteration;
} }

6
plugins/wasm-cpp/extensions/model_router/README.md
View File

@@ -1,14 +1,14 @@
# 功能说明 ## 功能说明
`model-router`插件实现了基于LLM协议中的model参数路由的功能 `model-router`插件实现了基于LLM协议中的model参数路由的功能
# 配置字段 ## 配置字段
| 名称 | 数据类型 | 填写要求 | 默认值 | 描述 | | 名称 | 数据类型 | 填写要求 | 默认值 | 描述 |
| ----------- | --------------- | ----------------------- | ------ | ------------------------------------------- | | ----------- | --------------- | ----------------------- | ------ | ------------------------------------------- |
| `modelKey` | string | 选填 | model | 请求body中model参数的位置 | | `modelKey` | string | 选填 | model | 请求body中model参数的位置 |
| `addProviderHeader` | string | 选填 | - | 从model参数中解析出的provider名字放到哪个请求header中 | | `addProviderHeader` | string | 选填 | - | 从model参数中解析出的provider名字放到哪个请求header中 |
| `modelToHeader` | string | 选填 | - | 直接将model参数放到哪个请求header中 | | `modelToHeader` | string | 选填 | - | 直接将model参数放到哪个请求header中 |
| `enableOnPathSuffix` | array of string | 选填 | ["/v1/chat/completions"] | 只对这些特定路径后缀的请求生效 | | `enableOnPathSuffix` | array of string | 选填 | ["/v1/chat/completions"] | 只对这些特定路径后缀的请求生效，可以配置为 "*" 以匹配所有路径 |
## 运行属性 ## 运行属性

42
plugins/wasm-cpp/extensions/model_router/README_EN.md
View File

@@ -1,31 +1,31 @@
## Function Description ## Feature Description
The `model-router` plugin implements the function of routing based on the model parameter in the LLM protocol. The `model-router` plugin implements routing functionality based on the model parameter in LLM protocols.
## Configuration Fields ## Configuration Fields
| Name | Data Type | Filling Requirement | Default Value | Description | | Name | Data Type | Requirement | Default Value | Description |
| ----------- | --------------- | ----------------------- | ------ | ------------------------------------------- | | ----------- | --------------- | ----------------------- | ------ | ------------------------------------------- |
| `modelKey` | string | Optional | model | The location of the model parameter in the request body | | `modelKey` | string | Optional | model | Location of the model parameter in the request body |
| `addProviderHeader` | string | Optional | - | Which request header to place the provider name parsed from the model parameter | | `addProviderHeader` | string | Optional | - | Which request header to add the provider name parsed from the model parameter |
| `modelToHeader` | string | Optional | - | Which request header to directly place the model parameter | | `modelToHeader` | string | Optional | - | Which request header to directly add the model parameter to |
| `enableOnPathSuffix` | array of string | Optional | ["/v1/chat/completions"] | Only effective for requests with these specific path suffixes | | `enableOnPathSuffix` | array of string | Optional | ["/v1/chat/completions"] | Only effective for requests with these specific path suffixes, can be configured as "*" to match all paths |
## Runtime Attributes ## Runtime Properties
Plugin execution phase: Authentication phase Plugin execution phase: Authentication phase
Plugin execution priority: 900 Plugin execution priority: 900
## Effect Description ## Effect Description
### Routing Based on the model Parameter ### Routing Based on Model Parameter
The following configuration is required: The following configuration is needed:
```yaml ```yaml
modelToHeader: x-higress-llm-model modelToHeader: x-higress-llm-model
``` ```
The plugin will extract the model parameter from the request and set it in the x-higress-llm-model request header, which can be used for subsequent routing. For example, the original LLM request body: The plugin extracts the model parameter from the request and sets it to the x-higress-llm-model request header for subsequent routing. For example, the original LLM request body is:
```json ```json
{ {
@@ -35,7 +35,7 @@ The plugin will extract the model parameter from the request and set it in the x
"stream": false, "stream": false,
"messages": [{ "messages": [{
"role": "user", "role": "user",
"content": "What is the GitHub address of the main repository for the higress project" "content": "What is the GitHub address of the Higress project's main repository?"
}], }],
"presence_penalty": 0, "presence_penalty": 0,
"temperature": 0.7, "temperature": 0.7,
@@ -43,21 +43,21 @@ The plugin will extract the model parameter from the request and set it in the x
} }
``` ```
After processing by this plugin, the following request header (which can be used for route matching) will be added: After processing by this plugin, the following request header will be added (can be used for route matching):
x-higress-llm-model: qwen-long x-higress-llm-model: qwen-long
### Extracting the provider Field from the model Parameter for Routing ### Extracting Provider Field from Model Parameter for Routing
> Note that this mode requires the client to specify the provider using a `/` separator in the model parameter. > Note that this mode requires the client to specify the provider in the model parameter using the `/` delimiter
The following configuration is required: The following configuration is needed:
```yaml ```yaml
addProviderHeader: x-higress-llm-provider addProviderHeader: x-higress-llm-provider
``` ```
The plugin will extract the provider part (if present) from the model parameter in the request and set it in the x-higress-llm-provider request header, which can be used for subsequent routing, and rewrite the model parameter to the model name part. For example, the original LLM request body: The plugin extracts the provider part (if any) from the model parameter in the request, sets it to the x-higress-llm-provider request header for subsequent routing, and rewrites the model parameter to only contain the model name part. For example, the original LLM request body is:
```json ```json
{ {
@@ -67,7 +67,7 @@ The plugin will extract the provider part (if present) from the model parameter
"stream": false, "stream": false,
"messages": [{ "messages": [{
"role": "user", "role": "user",
"content": "What is the GitHub address of the main repository for the higress project" "content": "What is the GitHub address of the Higress project's main repository?"
}], }],
"presence_penalty": 0, "presence_penalty": 0,
"temperature": 0.7, "temperature": 0.7,
@@ -75,7 +75,7 @@ The plugin will extract the provider part (if present) from the model parameter
} }
``` ```
After processing by this plugin, the following request header (which can be used for route matching) will be added: After processing by this plugin, the following request header will be added (can be used for route matching):
x-higress-llm-provider: dashscope x-higress-llm-provider: dashscope
@@ -89,7 +89,7 @@ The original LLM request body will be changed to:
"stream": false, "stream": false,
"messages": [{ "messages": [{
"role": "user", "role": "user",
"content": "What is the GitHub address of the main repository for the higress project" "content": "What is the GitHub address of the Higress project's main repository?"
}], }],
"presence_penalty": 0, "presence_penalty": 0,
"temperature": 0.7, "temperature": 0.7,

10
plugins/wasm-cpp/extensions/model_router/plugin.cc
View File

@@ -44,8 +44,8 @@ static RegisterContextFactory register_ModelRouter(
namespace { namespace {
constexpr std::string_view SetDecoderBufferLimitKey = constexpr std::string_view SetDecoderBufferLimitKey =
"SetRequestBodyBufferLimit"; "set_decoder_buffer_limit";
constexpr std::string_view DefaultMaxBodyBytes = "10485760"; constexpr std::string_view DefaultMaxBodyBytes = "104857600";
} // namespace } // namespace
@@ -137,6 +137,11 @@ FilterHeadersStatus PluginRootContext::onHeader(
} }
bool enable = false; bool enable = false;
for (const auto& enable_suffix : rule.enable_on_path_suffix_) { for (const auto& enable_suffix : rule.enable_on_path_suffix_) {
// Support wildcard "*" to enable for all paths
if (enable_suffix == "*") {
enable = true;
break;
}
if (absl::EndsWith({path.c_str(), uri_end}, enable_suffix)) { if (absl::EndsWith({path.c_str(), uri_end}, enable_suffix)) {
enable = true; enable = true;
break; break;
@@ -153,6 +158,7 @@ FilterHeadersStatus PluginRootContext::onHeader(
} }
removeRequestHeader(Wasm::Common::Http::Header::ContentLength); removeRequestHeader(Wasm::Common::Http::Header::ContentLength);
setFilterState(SetDecoderBufferLimitKey, DefaultMaxBodyBytes); setFilterState(SetDecoderBufferLimitKey, DefaultMaxBodyBytes);
LOG_INFO(absl::StrCat("SetRequestBodyBufferLimit: ", DefaultMaxBodyBytes));
return FilterHeadersStatus::StopIteration; return FilterHeadersStatus::StopIteration;
} }