@@ -71,111 +71,61 @@ def _validate_request(self, request: EmbeddingRequest) -> Optional[str]:
7171
7272 def _convert_to_internal_request (
7373 self ,
74- all_requests : List [ EmbeddingRequest ] ,
75- request_ids : List [ str ] ,
74+ request : EmbeddingRequest ,
75+ request_id : str ,
7676 ) -> tuple [EmbeddingReqInput , Union [EmbeddingRequest , List [EmbeddingRequest ]]]:
7777 """Convert OpenAI embedding request to internal format"""
78- prompts = [ request .input for request in all_requests ]
79-
80- # Handle single vs multiple requests
81- if len ( all_requests ) == 1 :
82- prompt = prompts [ 0 ]
83- if isinstance (prompt , str ):
84- # Single string input
78+ prompt = request .input
79+ if isinstance ( prompt , str ):
80+ # Single string input
81+ prompt_kwargs = { "text" : prompt }
82+ elif isinstance ( prompt , list ):
83+ if len ( prompt ) > 0 and isinstance (prompt [ 0 ] , str ):
84+ # List of strings
8585 prompt_kwargs = {"text" : prompt }
86- elif isinstance (prompt , list ):
87- if len (prompt ) > 0 and isinstance (prompt [0 ], str ):
88- # List of strings
89- prompt_kwargs = {"text" : prompt }
90- elif len (prompt ) > 0 and isinstance (
91- prompt [0 ], MultimodalEmbeddingInput
92- ):
93- # Handle multimodal embedding inputs
94- texts = []
95- images = []
96- for item in prompt :
97- # Use padding for text if None - this could be improved
98- texts .append (item .text if item .text is not None else "padding" )
99- images .append (item .image if item .image is not None else None )
100-
101- generate_prompts = []
102- # Check if we have a chat template for multimodal embeddings
103- # This would need to be passed in from the server configuration
104- chat_template_name = getattr (
105- self .tokenizer_manager , "chat_template_name" , None
106- )
107- if chat_template_name is not None :
108- convs = generate_embedding_convs (
109- texts , images , chat_template_name
110- )
111- for conv in convs :
112- generate_prompts .append (conv .get_prompt ())
113- else :
114- generate_prompts = texts
115-
116- if len (generate_prompts ) == 1 :
117- prompt_kwargs = {
118- "text" : generate_prompts [0 ],
119- "image_data" : images [0 ],
120- }
121- else :
122- prompt_kwargs = {
123- "text" : generate_prompts ,
124- "image_data" : images ,
125- }
86+ elif len (prompt ) > 0 and isinstance (prompt [0 ], MultimodalEmbeddingInput ):
87+ # Handle multimodal embedding inputs
88+ texts = []
89+ images = []
90+ for item in prompt :
91+ # Use padding for text if None - this could be improved
92+ texts .append (item .text if item .text is not None else "padding" )
93+ images .append (item .image if item .image is not None else None )
94+
95+ generate_prompts = []
96+ # Check if we have a chat template for multimodal embeddings
97+ # This would need to be passed in from the server configuration
98+ chat_template_name = getattr (
99+ self .tokenizer_manager , "chat_template_name" , None
100+ )
101+ if chat_template_name is not None :
102+ convs = generate_embedding_convs (texts , images , chat_template_name )
103+ for conv in convs :
104+ generate_prompts .append (conv .get_prompt ())
105+ else :
106+ generate_prompts = texts
107+
108+ if len (generate_prompts ) == 1 :
109+ prompt_kwargs = {
110+ "text" : generate_prompts [0 ],
111+ "image_data" : images [0 ],
112+ }
126113 else :
127- # List of integers (token IDs) or empty list
128- prompt_kwargs = {"input_ids" : prompt }
114+ prompt_kwargs = {
115+ "text" : generate_prompts ,
116+ "image_data" : images ,
117+ }
129118 else :
130- # Other types (should not happen but handle gracefully)
119+ # List of integers (token IDs) or empty list
131120 prompt_kwargs = {"input_ids" : prompt }
132- # Use the passed request_ids for single request
133- final_request_id = request_ids [0 ] if len (all_requests ) == 1 else request_ids
134121 else :
135- # Handle batch requests
136- if len (prompts ) > 0 :
137- # Validate that all prompts have the same type
138- first_prompt = prompts [0 ]
139- first_type = type (first_prompt )
140- for i , prompt in enumerate (prompts [1 :], 1 ):
141- if type (prompt ) != first_type :
142- raise AssertionError (
143- f"All prompts in batch must have the same type, but prompt at index { i }
144- )
145-
146- if isinstance (first_prompt , str ):
147- # Batch of strings
148- prompt_kwargs = {"text" : prompts }
149- elif isinstance (first_prompt , list ):
150- if len (first_prompt ) > 0 and isinstance (first_prompt [0 ], str ):
151- # Batch of lists of strings
152- prompt_kwargs = {"text" : prompts }
153- elif len (first_prompt ) > 0 and isinstance (
154- first_prompt [0 ], MultimodalEmbeddingInput
155- ):
156- # Handle multimodal batch requests
157- raise NotImplementedError (
158- "Multiple requests with multimodal inputs are not supported yet"
159- )
160- else :
161- # Batch of token ID lists
162- prompt_kwargs = {"input_ids" : prompts }
163- else :
164- # Other types
165- prompt_kwargs = {"input_ids" : prompts }
166- else :
167- prompt_kwargs = {"input_ids" : prompts }
168- # Use the passed request_ids for batch requests
169- final_request_id = request_ids
170-
122+ # Other types (should not happen but handle gracefully)
123+ prompt_kwargs = {"input_ids" : prompt }
171124 adapted_request = EmbeddingReqInput (
172- rid = final_request_id ,
173125 ** prompt_kwargs ,
174126 )
175127
176- return adapted_request , (
177- all_requests [0 ] if len (all_requests ) == 1 else all_requests
178- )
128+ return adapted_request , request
179129
180130 async def _handle_non_streaming_request (
181131 self ,
@@ -194,14 +144,10 @@ async def _handle_non_streaming_request(
194144 if not isinstance (ret , list ):
195145 ret = [ret ]
196146
197- response = self ._build_embedding_response (
198- ret , self .tokenizer_manager .model_path
199- )
147+ response = self ._build_embedding_response (ret )
200148 return response
201149
202- def _build_embedding_response (
203- self , ret : List [Dict [str , Any ]], model_path : str
204- ) -> EmbeddingResponse :
150+ def _build_embedding_response (self , ret : List [Dict [str , Any ]]) -> EmbeddingResponse :
205151 """Build the embedding response"""
206152 embedding_objects = []
207153 prompt_tokens = 0
@@ -219,7 +165,7 @@ def _build_embedding_response(
219165
220166 return EmbeddingResponse (
221167 data = embedding_objects ,
222- model = model_path ,
168+ model = self . tokenizer_manager . model_path ,
223169 usage = UsageInfo (
224170 prompt_tokens = prompt_tokens ,
225171 total_tokens = prompt_tokens ,
0 commit comments