add support for Deepseek R1 model and display CoT (#2118)

* feat: add support for R1 model and display CoT

* fix unpacking

* feat: customized font & font size

* auto hide tooltip when scoll down

* tooltip glass transparent css

* fix: Enhance API key validation in is_any_api_key function (#2113)

* support qwen2.5-max!

* update minior adjustment

---------

Co-authored-by: binary-husky <[email protected]>
Co-authored-by: Steven Moder <[email protected]>

Author: 3 people

config.py

@@ -13,6 +13,9 @@
 # [step 1-2]>> ( 接入通义 qwen-max ) 接入通义千问在线大模型，api-key获取地址 https://dashscope.console.aliyun.com/
 DASHSCOPE_API_KEY = "" # 阿里灵积云API_KEY
 
+# [step 1-3]>> ( 接入通义 deepseek-reasoner ) 深度求索(DeepSeek) API KEY，默认请求地址为"https://api.deepseek.com/v1/chat/completions"
+DEEPSEEK_API_KEY = ""
+
 # [step 2]>> 改为True应用代理，如果直接在海外服务器部署，此处不修改；如果使用本地或无地域限制的大模型时，此处也不需要修改
 USE_PROXY = False
 if USE_PROXY:
@@ -39,7 +42,8 @@
                     "gpt-4o", "gpt-4o-mini", "gpt-4-turbo", "gpt-4-turbo-2024-04-09",
                     "gpt-3.5-turbo-1106", "gpt-3.5-turbo-16k", "gpt-3.5-turbo", "azure-gpt-3.5",
                     "gpt-4", "gpt-4-32k", "azure-gpt-4", "glm-4", "glm-4v", "glm-3-turbo",
-                    "gemini-1.5-pro", "chatglm3", "chatglm4"
+                    "gemini-1.5-pro", "chatglm3", "chatglm4",
+                    "deepseek-chat", "deepseek-coder", "deepseek-reasoner"
                     ]
 
 EMBEDDING_MODEL = "text-embedding-3-small"
@@ -261,9 +265,6 @@
 # 零一万物(Yi Model) API KEY
 YIMODEL_API_KEY = ""
 
-# 深度求索(DeepSeek) API KEY，默认请求地址为"https://api.deepseek.com/v1/chat/completions"
-DEEPSEEK_API_KEY = ""
-
 
 # 紫东太初大模型 https://ai-maas.wair.ac.cn
 TAICHU_API_KEY = ""

request_llms/bridge_all.py

@@ -1090,18 +1090,18 @@ def decode(self, *args, **kwargs):
     except:
         logger.error(trimmed_format_exc())
 # -=-=-=-=-=-=- 幻方-深度求索大模型在线API -=-=-=-=-=-=-
-if "deepseek-chat" in AVAIL_LLM_MODELS or "deepseek-coder" in AVAIL_LLM_MODELS:
+if "deepseek-chat" in AVAIL_LLM_MODELS or "deepseek-coder" in AVAIL_LLM_MODELS or "deepseek-reasoner" in AVAIL_LLM_MODELS:
     try:
         deepseekapi_noui, deepseekapi_ui = get_predict_function(
             api_key_conf_name="DEEPSEEK_API_KEY", max_output_token=4096, disable_proxy=False
-            )
+        )
         model_info.update({
             "deepseek-chat":{
                 "fn_with_ui": deepseekapi_ui,
                 "fn_without_ui": deepseekapi_noui,
                 "endpoint": deepseekapi_endpoint,
                 "can_multi_thread": True,
-                "max_token": 32000,
+                "max_token": 64000,
                 "tokenizer": tokenizer_gpt35,
                 "token_cnt": get_token_num_gpt35,
             },
@@ -1114,6 +1114,16 @@ def decode(self, *args, **kwargs):
                 "tokenizer": tokenizer_gpt35,
                 "token_cnt": get_token_num_gpt35,
             },
+            "deepseek-reasoner":{
+                "fn_with_ui": deepseekapi_ui,
+                "fn_without_ui": deepseekapi_noui,
+                "endpoint": deepseekapi_endpoint,
+                "can_multi_thread": True,
+                "max_token": 64000,
+                "tokenizer": tokenizer_gpt35,
+                "token_cnt": get_token_num_gpt35,
+                "enable_reasoning": True
+            },
         })
     except:
         logger.error(trimmed_format_exc())

request_llms/oai_std_model_template.py

@@ -36,10 +36,11 @@ def get_full_error(chunk, stream_response):
 
 def decode_chunk(chunk):
     """
-    用于解读"content"和"finish_reason"的内容
+    用于解读"content"和"finish_reason"的内容（如果支持思维链也会返回"reasoning_content"内容）
     """
     chunk = chunk.decode()
     respose = ""
+    reasoning_content = ""
     finish_reason = "False"
     try:
         chunk = json.loads(chunk[6:])
@@ -57,14 +58,20 @@ def decode_chunk(chunk):
         return respose, finish_reason
 
     try:
-        respose = chunk["choices"][0]["delta"]["content"]
+        if chunk["choices"][0]["delta"]["content"] is not None:
+            respose = chunk["choices"][0]["delta"]["content"]
+    except:
+        pass
+    try:
+        if chunk["choices"][0]["delta"]["reasoning_content"] is not None:
+            reasoning_content = chunk["choices"][0]["delta"]["reasoning_content"]
     except:
         pass
     try:
         finish_reason = chunk["choices"][0]["finish_reason"]
     except:
         pass
-    return respose, finish_reason
+    return respose, reasoning_content, finish_reason
 
 
 def generate_message(input, model, key, history, max_output_token, system_prompt, temperature):
@@ -149,6 +156,7 @@ def predict_no_ui_long_connection(
         observe_window = None：
             用于负责跨越线程传递已经输出的部分，大部分时候仅仅为了fancy的视觉效果，留空即可。observe_window[0]：观测窗。observe_window[1]：看门狗
         """
+        from .bridge_all import model_info
         watch_dog_patience = 5  # 看门狗的耐心，设置5秒不准咬人(咬的也不是人
         if len(APIKEY) == 0:
             raise RuntimeError(f"APIKEY为空,请检查配置文件的{APIKEY}")
@@ -163,29 +171,21 @@ def predict_no_ui_long_connection(
             system_prompt=sys_prompt,
             temperature=llm_kwargs["temperature"],
         )
+
+        reasoning = model_info[llm_kwargs['llm_model']].get('enable_reasoning', False)
+
         retry = 0
         while True:
             try:
-                from .bridge_all import model_info
-
                 endpoint = model_info[llm_kwargs["llm_model"]]["endpoint"]
-                if not disable_proxy:
-                    response = requests.post(
-                        endpoint,
-                        headers=headers,
-                        proxies=proxies,
-                        json=playload,
-                        stream=True,
-                        timeout=TIMEOUT_SECONDS,
-                    )
-                else:
-                    response = requests.post(
-                        endpoint,
-                        headers=headers,
-                        json=playload,
-                        stream=True,
-                        timeout=TIMEOUT_SECONDS,
-                    )
+                response = requests.post(
+                    endpoint,
+                    headers=headers,
+                    proxies=None if disable_proxy else proxies,
+                    json=playload,
+                    stream=True,
+                    timeout=TIMEOUT_SECONDS,
+                )
                 break
             except:
                 retry += 1
@@ -194,10 +194,13 @@ def predict_no_ui_long_connection(
                     raise TimeoutError
                 if MAX_RETRY != 0:
                     logger.error(f"请求超时，正在重试 ({retry}/{MAX_RETRY}) ……")
-
-        stream_response = response.iter_lines()
+        
         result = ""
         finish_reason = ""
+        if reasoning:
+            resoning_buffer = ""
+        
+        stream_response = response.iter_lines()
         while True:
             try:
                 chunk = next(stream_response)
@@ -207,9 +210,9 @@ def predict_no_ui_long_connection(
                 break
             except requests.exceptions.ConnectionError:
                 chunk = next(stream_response)  # 失败了，重试一次？再失败就没办法了。
-            response_text, finish_reason = decode_chunk(chunk)
+            response_text, reasoning_content, finish_reason = decode_chunk(chunk)
             # 返回的数据流第一次为空，继续等待
-            if response_text == "" and finish_reason != "False":
+            if response_text == "" and (reasoning == False or reasoning_content == "") and finish_reason != "False":
                 continue
             if response_text == "API_ERROR" and (
                 finish_reason != "False" or finish_reason != "stop"
@@ -227,6 +230,8 @@ def predict_no_ui_long_connection(
                             print(f"[response] {result}")
                         break
                     result += response_text
+                    if reasoning:
+                        resoning_buffer += reasoning_content
                     if observe_window is not None:
                         # 观测窗，把已经获取的数据显示出去
                         if len(observe_window) >= 1:
@@ -241,6 +246,10 @@ def predict_no_ui_long_connection(
                     error_msg = chunk_decoded
                     logger.error(error_msg)
                     raise RuntimeError("Json解析不合常规")
+        if reasoning:
+            # reasoning 的部分加上框 (>)
+            return '\n'.join(map(lambda x: '> ' + x, resoning_buffer.split('\n'))) + \
+                   '\n\n' + result
         return result
 
     def predict(
@@ -262,6 +271,7 @@ def predict(
         chatbot 为WebUI中显示的对话列表，修改它，然后yeild出去，可以直接修改对话界面内容
         additional_fn代表点击的哪个按钮，按钮见functional.py
         """
+        from .bridge_all import model_info
         if len(APIKEY) == 0:
             raise RuntimeError(f"APIKEY为空,请检查配置文件的{APIKEY}")
         if inputs == "":
@@ -298,32 +308,23 @@ def predict(
             system_prompt=system_prompt,
             temperature=llm_kwargs["temperature"],
         )
+        
+        reasoning = model_info[llm_kwargs['llm_model']].get('enable_reasoning', False)
 
         history.append(inputs)
         history.append("")
         retry = 0
         while True:
             try:
-                from .bridge_all import model_info
-
                 endpoint = model_info[llm_kwargs["llm_model"]]["endpoint"]
-                if not disable_proxy:
-                    response = requests.post(
-                        endpoint,
-                        headers=headers,
-                        proxies=proxies,
-                        json=playload,
-                        stream=True,
-                        timeout=TIMEOUT_SECONDS,
-                    )
-                else:
-                    response = requests.post(
-                        endpoint,
-                        headers=headers,
-                        json=playload,
-                        stream=True,
-                        timeout=TIMEOUT_SECONDS,
-                    )
+                response = requests.post(
+                    endpoint,
+                    headers=headers,
+                    proxies=None if disable_proxy else proxies,
+                    json=playload,
+                    stream=True,
+                    timeout=TIMEOUT_SECONDS,
+                )
                 break
             except:
                 retry += 1
@@ -338,6 +339,8 @@ def predict(
                     raise TimeoutError
 
         gpt_replying_buffer = ""
+        if reasoning:
+            gpt_reasoning_buffer = ""
 
         stream_response = response.iter_lines()
         while True:
@@ -347,9 +350,9 @@ def predict(
                 break
             except requests.exceptions.ConnectionError:
                 chunk = next(stream_response)  # 失败了，重试一次？再失败就没办法了。
-            response_text, finish_reason = decode_chunk(chunk)
+            response_text, reasoning_content, finish_reason = decode_chunk(chunk)
             # 返回的数据流第一次为空，继续等待
-            if response_text == "" and finish_reason != "False":
+            if response_text == "" and (reasoning == False or reasoning_content == "") and finish_reason != "False":
                 status_text = f"finish_reason: {finish_reason}"
                 yield from update_ui(
                     chatbot=chatbot, history=history, msg=status_text
@@ -379,9 +382,14 @@ def predict(
                         logger.info(f"[response] {gpt_replying_buffer}")
                         break
                     status_text = f"finish_reason: {finish_reason}"
-                    gpt_replying_buffer += response_text
-                    # 如果这里抛出异常，一般是文本过长，详情见get_full_error的输出
-                    history[-1] = gpt_replying_buffer
+                    if reasoning:
+                        gpt_replying_buffer += response_text
+                        gpt_reasoning_buffer += reasoning_content
+                        history[-1] = '\n'.join(map(lambda x: '> ' + x, gpt_reasoning_buffer.split('\n'))) + '\n\n' + gpt_replying_buffer
+                    else:
+                        gpt_replying_buffer += response_text
+                        # 如果这里抛出异常，一般是文本过长，详情见get_full_error的输出
+                        history[-1] = gpt_replying_buffer
                     chatbot[-1] = (history[-2], history[-1])
                     yield from update_ui(
                         chatbot=chatbot, history=history, msg=status_text