feat(large_language_model): Adds plugin-based token counting configuration option (#17706)

Signed-off-by: -LAN- <laipz8200@outlook.com>
Co-authored-by: Yeuoly <admin@srmxy.cn>

api/.env.example

@@ -326,6 +326,7 @@ UPLOAD_AUDIO_FILE_SIZE_LIMIT=50
 MULTIMODAL_SEND_FORMAT=base64
 PROMPT_GENERATION_MAX_TOKENS=512
 CODE_GENERATION_MAX_TOKENS=1024
+PLUGIN_BASED_TOKEN_COUNTING_ENABLED=false
 
 # Mail configuration, support: resend, smtp
 MAIL_TYPE=

api/configs/feature/__init__.py

@@ -442,7 +442,7 @@ class LoggingConfig(BaseSettings):
 
 class ModelLoadBalanceConfig(BaseSettings):
     """
-    Configuration for model load balancing
+    Configuration for model load balancing and token counting
     """
 
     MODEL_LB_ENABLED: bool = Field(
@@ -450,6 +450,11 @@ class ModelLoadBalanceConfig(BaseSettings):
         default=False,
     )
 
+    PLUGIN_BASED_TOKEN_COUNTING_ENABLED: bool = Field(
+        description="Enable or disable plugin based token counting. If disabled, token counting will return 0.",
+        default=False,
+    )
+
 
 class BillingConfig(BaseSettings):
     """

api/core/app/apps/agent_chat/app_runner.py

@@ -53,20 +53,6 @@ class AgentChatAppRunner(AppRunner):
         query = application_generate_entity.query
         files = application_generate_entity.files
 
-        # Pre-calculate the number of tokens of the prompt messages,
-        # and return the rest number of tokens by model context token size limit and max token size limit.
-        # If the rest number of tokens is not enough, raise exception.
-        # Include: prompt template, inputs, query(optional), files(optional)
-        # Not Include: memory, external data, dataset context
-        self.get_pre_calculate_rest_tokens(
-            app_record=app_record,
-            model_config=application_generate_entity.model_conf,
-            prompt_template_entity=app_config.prompt_template,
-            inputs=dict(inputs),
-            files=list(files),
-            query=query,
-        )
-
         memory = None
         if application_generate_entity.conversation_id:
             # get memory of conversation (read-only)

api/core/app/apps/chat/app_runner.py

@@ -61,20 +61,6 @@ class ChatAppRunner(AppRunner):
         )
         image_detail_config = image_detail_config or ImagePromptMessageContent.DETAIL.LOW
 
-        # Pre-calculate the number of tokens of the prompt messages,
-        # and return the rest number of tokens by model context token size limit and max token size limit.
-        # If the rest number of tokens is not enough, raise exception.
-        # Include: prompt template, inputs, query(optional), files(optional)
-        # Not Include: memory, external data, dataset context
-        self.get_pre_calculate_rest_tokens(
-            app_record=app_record,
-            model_config=application_generate_entity.model_conf,
-            prompt_template_entity=app_config.prompt_template,
-            inputs=inputs,
-            files=files,
-            query=query,
-        )
-
         memory = None
         if application_generate_entity.conversation_id:
             # get memory of conversation (read-only)

api/core/app/apps/completion/app_runner.py

@@ -54,20 +54,6 @@ class CompletionAppRunner(AppRunner):
         )
         image_detail_config = image_detail_config or ImagePromptMessageContent.DETAIL.LOW
 
-        # Pre-calculate the number of tokens of the prompt messages,
-        # and return the rest number of tokens by model context token size limit and max token size limit.
-        # If the rest number of tokens is not enough, raise exception.
-        # Include: prompt template, inputs, query(optional), files(optional)
-        # Not Include: memory, external data, dataset context
-        self.get_pre_calculate_rest_tokens(
-            app_record=app_record,
-            model_config=application_generate_entity.model_conf,
-            prompt_template_entity=app_config.prompt_template,
-            inputs=inputs,
-            files=files,
-            query=query,
-        )
-
         # organize all inputs and template to prompt messages
         # Include: prompt template, inputs, query(optional), files(optional)
         prompt_messages, stop = self.organize_prompt_messages(

api/core/model_runtime/docs/en_US/customizable_model_scale_out.md

@@ -192,7 +192,7 @@ def get_num_tokens(self, model: str, credentials: dict, prompt_messages: list[Pr
 ```
 
 
-Sometimes, you might not want to return 0 directly. In such cases, you can use `self._get_num_tokens_by_gpt2(text: str)` to get pre-computed tokens. This method is provided by the `AIModel` base class, and it uses GPT2's Tokenizer for calculation. However, it should be noted that this is only a substitute and may not be fully accurate.
+Sometimes, you might not want to return 0 directly. In such cases, you can use `self._get_num_tokens_by_gpt2(text: str)` to get pre-computed tokens and ensure environment variable `PLUGIN_BASED_TOKEN_COUNTING_ENABLED` is set to `true`, This method is provided by the `AIModel` base class, and it uses GPT2's Tokenizer for calculation. However, it should be noted that this is only a substitute and may not be fully accurate.
 
 - Model Credentials Validation
 

api/core/model_runtime/docs/zh_Hans/customizable_model_scale_out.md

@@ -179,7 +179,7 @@ provider_credential_schema:
     """
   ```
 
-  有时候，也许你不需要直接返回0，所以你可以使用`self._get_num_tokens_by_gpt2(text: str)`来获取预计算的tokens，这个方法位于`AIModel`基类中，它会使用GPT2的Tokenizer进行计算，但是只能作为替代方法，并不完全准确。
+  有时候，也许你不需要直接返回0，所以你可以使用`self._get_num_tokens_by_gpt2(text: str)`来获取预计算的tokens，并确保环境变量`PLUGIN_BASED_TOKEN_COUNTING_ENABLED`设置为`true`，这个方法位于`AIModel`基类中，它会使用GPT2的Tokenizer进行计算，但是只能作为替代方法，并不完全准确。
 
 - 模型凭据校验
 

api/core/model_runtime/model_providers/__base/large_language_model.py

@@ -295,18 +295,20 @@ class LargeLanguageModel(AIModel):
         :param tools: tools for tool calling
         :return:
         """
-        plugin_model_manager = PluginModelManager()
+        if dify_config.PLUGIN_BASED_TOKEN_COUNTING_ENABLED:
-        return plugin_model_manager.get_llm_num_tokens(
+            plugin_model_manager = PluginModelManager()
-            tenant_id=self.tenant_id,
+            return plugin_model_manager.get_llm_num_tokens(
-            user_id="unknown",
+                tenant_id=self.tenant_id,
-            plugin_id=self.plugin_id,
+                user_id="unknown",
-            provider=self.provider_name,
+                plugin_id=self.plugin_id,
-            model_type=self.model_type.value,
+                provider=self.provider_name,
-            model=model,
+                model_type=self.model_type.value,
-            credentials=credentials,
+                model=model,
-            prompt_messages=prompt_messages,
+                credentials=credentials,
-            tools=tools,
+                prompt_messages=prompt_messages,
-        )
+                tools=tools,
+            )
+        return 0
 
     def _calc_response_usage(
         self, model: str, credentials: dict, prompt_tokens: int, completion_tokens: int

docker/.env.example

@@ -616,6 +616,11 @@ PROMPT_GENERATION_MAX_TOKENS=512
 # Default: 1024 tokens.
 CODE_GENERATION_MAX_TOKENS=1024
 
+# Enable or disable plugin based token counting. If disabled, token counting will return 0.
+# This can improve performance by skipping token counting operations.
+# Default: false (disabled).
+PLUGIN_BASED_TOKEN_COUNTING_ENABLED=false
+
 # ------------------------------
 # Multi-modal Configuration
 # ------------------------------

docker/docker-compose.yaml

@@ -276,6 +276,7 @@ x-shared-env: &shared-api-worker-env
   SCARF_NO_ANALYTICS: ${SCARF_NO_ANALYTICS:-true}
   PROMPT_GENERATION_MAX_TOKENS: ${PROMPT_GENERATION_MAX_TOKENS:-512}
   CODE_GENERATION_MAX_TOKENS: ${CODE_GENERATION_MAX_TOKENS:-1024}
+  PLUGIN_BASED_TOKEN_COUNTING_ENABLED: ${PLUGIN_BASED_TOKEN_COUNTING_ENABLED:-false}
   MULTIMODAL_SEND_FORMAT: ${MULTIMODAL_SEND_FORMAT:-base64}
   UPLOAD_IMAGE_FILE_SIZE_LIMIT: ${UPLOAD_IMAGE_FILE_SIZE_LIMIT:-10}
   UPLOAD_VIDEO_FILE_SIZE_LIMIT: ${UPLOAD_VIDEO_FILE_SIZE_LIMIT:-100}