Modify model parameters in Spark LLMs and zhipuai LLMs (#8078)

Co-authored-by: Charlie.Wei <[email protected]>

api/core/model_runtime/model_providers/spark/llm/_client.py

@@ -19,27 +19,25 @@ def __init__(self, model: str, app_id: str, api_key: str, api_secret: str, api_d
         endpoint = 'chat'
         if api_domain:
             domain = api_domain
-            if model == 'spark-v3':
-                endpoint = 'multimodal'
 
         model_api_configs = {
-            'spark-1.5': {
+            'spark-lite': {
                 'version': 'v1.1',
                 'chat_domain': 'general'
             },
-            'spark-2': {
-                'version': 'v2.1',
-                'chat_domain': 'generalv2'
-            },
-            'spark-3': {
+            'spark-pro': {
                 'version': 'v3.1',
                 'chat_domain': 'generalv3'
             },
-            'spark-3.5': {
+            'spark-pro-128k': {
+                'version': 'pro-128k',
+                'chat_domain': 'pro-128k'
+            },
+            'spark-max': {
                 'version': 'v3.5',
                 'chat_domain': 'generalv3.5'
             },
-            'spark-4': {
+            'spark-4.0-ultra': {
                 'version': 'v4.0',
                 'chat_domain': '4.0Ultra'
             }
@@ -48,7 +46,12 @@ def __init__(self, model: str, app_id: str, api_key: str, api_secret: str, api_d
         api_version = model_api_configs[model]['version']
 
         self.chat_domain = model_api_configs[model]['chat_domain']
-        self.api_base = f"wss://{domain}/{api_version}/{endpoint}"
+
+        if model == 'spark-pro-128k':
+            self.api_base = f"wss://{domain}/{endpoint}/{api_version}"
+        else:
+            self.api_base = f"wss://{domain}/{api_version}/{endpoint}"
+
         self.app_id = app_id
         self.ws_url = self.create_url(
             urlparse(self.api_base).netloc,

api/core/model_runtime/model_providers/spark/llm/_position.yaml

@@ -1,3 +1,8 @@
+- spark-4.0-ultra
+- spark-max
+- spark-pro-128k
+- spark-pro
+- spark-lite
 - spark-4
 - spark-3.5
 - spark-3

api/core/model_runtime/model_providers/spark/llm/spark-1.5.yaml

@@ -1,4 +1,5 @@
 model: spark-1.5
+deprecated: true
 label:
   en_US: Spark V1.5
 model_type: llm

api/core/model_runtime/model_providers/spark/llm/spark-3.5.yaml

@@ -1,4 +1,5 @@
 model: spark-3.5
+deprecated: true
 label:
   en_US: Spark V3.5
 model_type: llm

api/core/model_runtime/model_providers/spark/llm/spark-3.yaml

@@ -1,4 +1,5 @@
 model: spark-3
+deprecated: true
 label:
   en_US: Spark V3.0
 model_type: llm

api/core/model_runtime/model_providers/spark/llm/spark-4.0-ultra.yaml

@@ -0,0 +1,42 @@
+model: spark-4.0-ultra
+label:
+  en_US: Spark 4.0 Ultra
+model_type: llm
+model_properties:
+  mode: chat
+parameter_rules:
+  - name: temperature
+    use_template: temperature
+    default: 0.5
+    help:
+      zh_Hans: 核采样阈值。用于决定结果随机性，取值越高随机性越强即相同的问题得到的不同答案的可能性越高。
+      en_US: Kernel sampling threshold. Used to determine the randomness of the results. The higher the value, the stronger the randomness, that is, the higher the possibility of getting different answers to the same question.
+  - name: max_tokens
+    use_template: max_tokens
+    default: 4096
+    min: 1
+    max: 8192
+    help:
+      zh_Hans: 模型回答的tokens的最大长度。
+      en_US: Maximum length of tokens for the model response.
+  - name: top_k
+    label:
+      zh_Hans: 取样数量
+      en_US: Top k
+    type: int
+    default: 4
+    min: 1
+    max: 6
+    help:
+      zh_Hans: 从 k 个候选中随机选择一个（非等概率）。
+      en_US: Randomly select one from k candidates (non-equal probability).
+    required: false
+  - name: show_ref_label
+    label:
+      zh_Hans: 联网检索
+      en_US: web search
+    type: boolean
+    default: false
+    help:
+      zh_Hans: 该参数仅4.0 Ultra版本支持，当设置为true时，如果输入内容触发联网检索插件，会先返回检索信源列表，然后再返回星火回复结果，否则仅返回星火回复结果
+      en_US: The parameter is only supported in the 4.0 Ultra version. When set to true, if the input triggers the online search plugin, it will first return a list of search sources and then return the Spark response. Otherwise, it will only return the Spark response.

api/core/model_runtime/model_providers/spark/llm/spark-4.yaml

@@ -1,4 +1,5 @@
 model: spark-4
+deprecated: true
 label:
   en_US: Spark V4.0
 model_type: llm

api/core/model_runtime/model_providers/spark/llm/spark-lite.yaml

@@ -0,0 +1,33 @@
+model: spark-lite
+label:
+  en_US: Spark Lite
+model_type: llm
+model_properties:
+  mode: chat
+parameter_rules:
+  - name: temperature
+    use_template: temperature
+    default: 0.5
+    help:
+      zh_Hans: 核采样阈值。用于决定结果随机性，取值越高随机性越强即相同的问题得到的不同答案的可能性越高。
+      en_US: Kernel sampling threshold. Used to determine the randomness of the results. The higher the value, the stronger the randomness, that is, the higher the possibility of getting different answers to the same question.
+  - name: max_tokens
+    use_template: max_tokens
+    default: 4096
+    min: 1
+    max: 4096
+    help:
+      zh_Hans: 模型回答的tokens的最大长度。
+      en_US: Maximum length of tokens for the model response.
+  - name: top_k
+    label:
+      zh_Hans: 取样数量
+      en_US: Top k
+    type: int
+    default: 4
+    min: 1
+    max: 6
+    help:
+      zh_Hans: 从 k 个候选中随机选择一个（非等概率）。
+      en_US: Randomly select one from k candidates (non-equal probability).
+    required: false

api/core/model_runtime/model_providers/spark/llm/spark-max.yaml

@@ -0,0 +1,33 @@
+model: spark-max
+label:
+  en_US: Spark Max
+model_type: llm
+model_properties:
+  mode: chat
+parameter_rules:
+  - name: temperature
+    use_template: temperature
+    default: 0.5
+    help:
+      zh_Hans: 核采样阈值。用于决定结果随机性，取值越高随机性越强即相同的问题得到的不同答案的可能性越高。
+      en_US: Kernel sampling threshold. Used to determine the randomness of the results. The higher the value, the stronger the randomness, that is, the higher the possibility of getting different answers to the same question.
+  - name: max_tokens
+    use_template: max_tokens
+    default: 4096
+    min: 1
+    max: 8192
+    help:
+      zh_Hans: 模型回答的tokens的最大长度。
+      en_US: Maximum length of tokens for the model response.
+  - name: top_k
+    label:
+      zh_Hans: 取样数量
+      en_US: Top k
+    type: int
+    default: 4
+    min: 1
+    max: 6
+    help:
+      zh_Hans: 从 k 个候选中随机选择一个（非等概率）。
+      en_US: Randomly select one from k candidates (non-equal probability).
+    required: false

api/core/model_runtime/model_providers/spark/llm/spark-pro-128k.yaml

@@ -0,0 +1,33 @@
+model: spark-pro-128k
+label:
+  en_US: Spark Pro-128K
+model_type: llm
+model_properties:
+  mode: chat
+parameter_rules:
+  - name: temperature
+    use_template: temperature
+    default: 0.5
+    help:
+      zh_Hans: 核采样阈值。用于决定结果随机性，取值越高随机性越强即相同的问题得到的不同答案的可能性越高。
+      en_US: Kernel sampling threshold. Used to determine the randomness of the results. The higher the value, the stronger the randomness, that is, the higher the possibility of getting different answers to the same question.
+  - name: max_tokens
+    use_template: max_tokens
+    default: 4096
+    min: 1
+    max: 4096
+    help:
+      zh_Hans: 模型回答的tokens的最大长度。
+      en_US: Maximum length of tokens for the model response.
+  - name: top_k
+    label:
+      zh_Hans: 取样数量
+      en_US: Top k
+    type: int
+    default: 4
+    min: 1
+    max: 6
+    help:
+      zh_Hans: 从 k 个候选中随机选择一个（非等概率）。
+      en_US: Randomly select one from k candidates (non-equal probability).
+    required: false

api/core/model_runtime/model_providers/spark/llm/spark-pro.yaml

@@ -0,0 +1,33 @@
+model: spark-pro
+label:
+  en_US: Spark Pro
+model_type: llm
+model_properties:
+  mode: chat
+parameter_rules:
+  - name: temperature
+    use_template: temperature
+    default: 0.5
+    help:
+      zh_Hans: 核采样阈值。用于决定结果随机性，取值越高随机性越强即相同的问题得到的不同答案的可能性越高。
+      en_US: Kernel sampling threshold. Used to determine the randomness of the results. The higher the value, the stronger the randomness, that is, the higher the possibility of getting different answers to the same question.
+  - name: max_tokens
+    use_template: max_tokens
+    default: 4096
+    min: 1
+    max: 8192
+    help:
+      zh_Hans: 模型回答的tokens的最大长度。
+      en_US: Maximum length of tokens for the model response.
+  - name: top_k
+    label:
+      zh_Hans: 取样数量
+      en_US: Top k
+    type: int
+    default: 4
+    min: 1
+    max: 6
+    help:
+      zh_Hans: 从 k 个候选中随机选择一个（非等概率）。
+      en_US: Randomly select one from k candidates (non-equal probability).
+    required: false

api/core/model_runtime/model_providers/zhipuai/llm/chatglm_turbo.yaml

@@ -19,15 +19,24 @@ parameter_rules:
     help:
       zh_Hans: 用温度取样的另一种方法，称为核取样取值范围是：(0.0, 1.0) 开区间，不能等于 0 或 1，默认值为 0.7 模型考虑具有 top_p 概率质量tokens的结果例如：0.1 意味着模型解码器只考虑从前 10% 的概率的候选集中取 tokens 建议您根据应用场景调整 top_p 或 temperature 参数，但不要同时调整两个参数。
       en_US: Another method of temperature sampling is called kernel sampling. The value range is (0.0, 1.0) open interval, which cannot be equal to 0 or 1. The default value is 0.7. The model considers the results with top_p probability mass tokens. For example 0.1 means The model decoder only considers tokens from the candidate set with the top 10% probability. It is recommended that you adjust the top_p or temperature parameters according to the application scenario, but do not adjust both parameters at the same time.
-  - name: incremental
+  - name: do_sample
     label:
-      zh_Hans: 增量返回
-      en_US: Incremental
+      zh_Hans: 采样策略
+      en_US: Sampling strategy
     type: boolean
     help:
-      zh_Hans: SSE接口调用时，用于控制每次返回内容方式是增量还是全量，不提供此参数时默认为增量返回，true 为增量返回，false 为全量返回。
-      en_US: When the SSE interface is called, it is used to control whether the content is returned incrementally or in full. If this parameter is not provided, the default is incremental return. true means incremental return, false means full return.
-    required: false
+      zh_Hans: do_sample 为 true 时启用采样策略，do_sample 为 false 时采样策略 temperature、top_p 将不生效。默认值为 true。
+      en_US: When `do_sample` is set to true, the sampling strategy is enabled. When `do_sample` is set to false, the sampling strategies such as `temperature` and `top_p` will not take effect. The default value is true.
+    default: true
+  - name: stream
+    label:
+      zh_Hans: 流处理
+      en_US: Event Stream
+    type: boolean
+    help:
+      zh_Hans: 使用同步调用时，此参数应当设置为 fasle 或者省略。表示模型生成完所有内容后一次性返回所有内容。默认值为 false。如果设置为 true，模型将通过标准 Event Stream ，逐块返回模型生成内容。Event Stream 结束时会返回一条data：[DONE]消息。注意：在模型流式输出生成内容的过程中，我们会分批对模型生成内容进行检测，当检测到违法及不良信息时，API会返回错误码（1301）。开发者识别到错误码（1301），应及时采取（清屏、重启对话）等措施删除生成内容，并确保不将含有违法及不良信息的内容传递给模型继续生成，避免其造成负面影响。
+      en_US: When using synchronous invocation, this parameter should be set to false or omitted. It indicates that the model will return all the generated content at once after the generation is complete. The default value is false. If set to true, the model will return the generated content in chunks via the standard Event Stream. A data：[DONE] message will be sent at the end of the Event Stream.Note：During the model's streaming output process, we will batch check the generated content. If illegal or harmful information is detected, the API will return an error code (1301). Developers who identify error code (1301) should promptly take actions such as clearing the screen or restarting the conversation to delete the generated content. They should also ensure that no illegal or harmful content is passed back to the model for continued generation to avoid negative impacts.
+    default: false
   - name: return_type
     label:
       zh_Hans: 回复类型

api/core/model_runtime/model_providers/zhipuai/llm/glm-4-0520.yaml

@@ -23,20 +23,29 @@ parameter_rules:
     help:
       zh_Hans: 用温度取样的另一种方法，称为核取样取值范围是：(0.0, 1.0) 开区间，不能等于 0 或 1，默认值为 0.7 模型考虑具有 top_p 概率质量tokens的结果例如：0.1 意味着模型解码器只考虑从前 10% 的概率的候选集中取 tokens 建议您根据应用场景调整 top_p 或 temperature 参数，但不要同时调整两个参数。
       en_US: Another method of temperature sampling is called kernel sampling. The value range is (0.0, 1.0) open interval, which cannot be equal to 0 or 1. The default value is 0.7. The model considers the results with top_p probability mass tokens. For example 0.1 means The model decoder only considers tokens from the candidate set with the top 10% probability. It is recommended that you adjust the top_p or temperature parameters according to the application scenario, but do not adjust both parameters at the same time.
-  - name: incremental
+  - name: do_sample
     label:
-      zh_Hans: 增量返回
-      en_US: Incremental
+      zh_Hans: 采样策略
+      en_US: Sampling strategy
     type: boolean
     help:
-      zh_Hans: SSE接口调用时，用于控制每次返回内容方式是增量还是全量，不提供此参数时默认为增量返回，true 为增量返回，false 为全量返回。
-      en_US: When the SSE interface is called, it is used to control whether the content is returned incrementally or in full. If this parameter is not provided, the default is incremental return. true means incremental return, false means full return.
-    required: false
+      zh_Hans: do_sample 为 true 时启用采样策略，do_sample 为 false 时采样策略 temperature、top_p 将不生效。默认值为 true。
+      en_US: When `do_sample` is set to true, the sampling strategy is enabled. When `do_sample` is set to false, the sampling strategies such as `temperature` and `top_p` will not take effect. The default value is true.
+    default: true
+  - name: stream
+    label:
+      zh_Hans: 流处理
+      en_US: Event Stream
+    type: boolean
+    help:
+      zh_Hans: 使用同步调用时，此参数应当设置为 fasle 或者省略。表示模型生成完所有内容后一次性返回所有内容。默认值为 false。如果设置为 true，模型将通过标准 Event Stream ，逐块返回模型生成内容。Event Stream 结束时会返回一条data：[DONE]消息。注意：在模型流式输出生成内容的过程中，我们会分批对模型生成内容进行检测，当检测到违法及不良信息时，API会返回错误码（1301）。开发者识别到错误码（1301），应及时采取（清屏、重启对话）等措施删除生成内容，并确保不将含有违法及不良信息的内容传递给模型继续生成，避免其造成负面影响。
+      en_US: When using synchronous invocation, this parameter should be set to false or omitted. It indicates that the model will return all the generated content at once after the generation is complete. The default value is false. If set to true, the model will return the generated content in chunks via the standard Event Stream. A data：[DONE] message will be sent at the end of the Event Stream.Note：During the model's streaming output process, we will batch check the generated content. If illegal or harmful information is detected, the API will return an error code (1301). Developers who identify error code (1301) should promptly take actions such as clearing the screen or restarting the conversation to delete the generated content. They should also ensure that no illegal or harmful content is passed back to the model for continued generation to avoid negative impacts.
+    default: false
   - name: max_tokens
     use_template: max_tokens
     default: 1024
     min: 1
-    max: 8192
+    max: 4095
 pricing:
   input: '0.1'
   output: '0.1'

api/core/model_runtime/model_providers/zhipuai/llm/glm-4-air.yaml

@@ -23,20 +23,29 @@ parameter_rules:
     help:
       zh_Hans: 用温度取样的另一种方法，称为核取样取值范围是：(0.0, 1.0) 开区间，不能等于 0 或 1，默认值为 0.7 模型考虑具有 top_p 概率质量tokens的结果例如：0.1 意味着模型解码器只考虑从前 10% 的概率的候选集中取 tokens 建议您根据应用场景调整 top_p 或 temperature 参数，但不要同时调整两个参数。
       en_US: Another method of temperature sampling is called kernel sampling. The value range is (0.0, 1.0) open interval, which cannot be equal to 0 or 1. The default value is 0.7. The model considers the results with top_p probability mass tokens. For example 0.1 means The model decoder only considers tokens from the candidate set with the top 10% probability. It is recommended that you adjust the top_p or temperature parameters according to the application scenario, but do not adjust both parameters at the same time.
-  - name: incremental
+  - name: do_sample
     label:
-      zh_Hans: 增量返回
-      en_US: Incremental
+      zh_Hans: 采样策略
+      en_US: Sampling strategy
     type: boolean
     help:
-      zh_Hans: SSE接口调用时，用于控制每次返回内容方式是增量还是全量，不提供此参数时默认为增量返回，true 为增量返回，false 为全量返回。
-      en_US: When the SSE interface is called, it is used to control whether the content is returned incrementally or in full. If this parameter is not provided, the default is incremental return. true means incremental return, false means full return.
-    required: false
+      zh_Hans: do_sample 为 true 时启用采样策略，do_sample 为 false 时采样策略 temperature、top_p 将不生效。默认值为 true。
+      en_US: When `do_sample` is set to true, the sampling strategy is enabled. When `do_sample` is set to false, the sampling strategies such as `temperature` and `top_p` will not take effect. The default value is true.
+    default: true
+  - name: stream
+    label:
+      zh_Hans: 流处理
+      en_US: Event Stream
+    type: boolean
+    help:
+      zh_Hans: 使用同步调用时，此参数应当设置为 fasle 或者省略。表示模型生成完所有内容后一次性返回所有内容。默认值为 false。如果设置为 true，模型将通过标准 Event Stream ，逐块返回模型生成内容。Event Stream 结束时会返回一条data：[DONE]消息。注意：在模型流式输出生成内容的过程中，我们会分批对模型生成内容进行检测，当检测到违法及不良信息时，API会返回错误码（1301）。开发者识别到错误码（1301），应及时采取（清屏、重启对话）等措施删除生成内容，并确保不将含有违法及不良信息的内容传递给模型继续生成，避免其造成负面影响。
+      en_US: When using synchronous invocation, this parameter should be set to false or omitted. It indicates that the model will return all the generated content at once after the generation is complete. The default value is false. If set to true, the model will return the generated content in chunks via the standard Event Stream. A data：[DONE] message will be sent at the end of the Event Stream.Note：During the model's streaming output process, we will batch check the generated content. If illegal or harmful information is detected, the API will return an error code (1301). Developers who identify error code (1301) should promptly take actions such as clearing the screen or restarting the conversation to delete the generated content. They should also ensure that no illegal or harmful content is passed back to the model for continued generation to avoid negative impacts.
+    default: false
   - name: max_tokens
     use_template: max_tokens
     default: 1024
     min: 1
-    max: 8192
+    max: 4095
 pricing:
   input: '0.001'
   output: '0.001'