ExternalVendorCode/LocalAI
1
0
Fork 0
mirror of https://github.com/mudler/LocalAI.git synced 2025-05-02 08:43:10 +00:00
Code Issues Actions 13 Packages Projects Releases Wiki Activity

Transformers Backend: max_tokens adherence to OpenAI API (#2108)

Browse Source 
max token adherence to OpenAI API

improve adherence to OpenAI API when max tokens is omitted or equal to 0 in the request
This commit is contained in:
fakezeta 2024-04-23 18:42:17 +02:00 committed by GitHub
parent 0d8bf91699
commit 8e36fe9b6f
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 11 additions and 4 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
backend/python/transformers/transformers_server.py
View File

@ -159,6 +159,11 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
quantization_config=quantization, quantization_config=quantization,
device_map=device_map, device_map=device_map,
torch_dtype=compute) torch_dtype=compute)
if request.ContextSize > 0:
self.max_tokens = request.ContextSize
else:
self.max_tokens = self.model.config.max_position_embeddings
self.tokenizer = AutoTokenizer.from_pretrained(model_name, use_safetensors=True) self.tokenizer = AutoTokenizer.from_pretrained(model_name, use_safetensors=True)
self.XPU = False self.XPU = False
@ -217,10 +222,6 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
if request.TopK == 0: if request.TopK == 0:
request.TopK = 40 request.TopK = 40
max_tokens = 200
if request.Tokens > 0:
max_tokens = request.Tokens
prompt = request.Prompt prompt = request.Prompt
if not request.Prompt and request.UseTokenizerTemplate and request.Messages: if not request.Prompt and request.UseTokenizerTemplate and request.Messages:
prompt = self.tokenizer.apply_chat_template(request.Messages, tokenize=False, add_generation_prompt=True) prompt = self.tokenizer.apply_chat_template(request.Messages, tokenize=False, add_generation_prompt=True)
@ -232,6 +233,12 @@ class BackendServicer(backend_pb2_grpc.BackendServicer):
eos_token_id.append(self.tokenizer.convert_tokens_to_ids(word)) eos_token_id.append(self.tokenizer.convert_tokens_to_ids(word))
inputs = self.tokenizer(prompt, return_tensors="pt") inputs = self.tokenizer(prompt, return_tensors="pt")
if request.Tokens > 0:
max_tokens = request.Tokens
else:
max_tokens = self.max_tokens - inputs["input_ids"].size()[inputs["input_ids"].dim()-1]
if self.CUDA: if self.CUDA:
inputs = inputs.to("cuda") inputs = inputs.to("cuda")
if XPU and self.OV == False: if XPU and self.OV == False: