problem with output_log_probs

[Alireza3242]

I run triton with tensorrtllm. But when i give long text to llm, triton returns a long array of zeros named output_log_probs in every token. If my text be longer than some number, the request not work truly.

can you add a option to config.pbtxt that prevent sending output_log_probs?

[hello-11]

Could you tell me more details about your questions?

[Alireza3242]

@hello-11
For example i send this request:

curl -X POST my_ip:8000/v2/models/ensemble/generate_stream -d '{"text_input": "hello!, How are you?", "max_tokens":1024, "temperature":0.1, "top_p":0.9, "top_k":1, "repetition_penalty":1.15, "stream":true}'

answer is:

data: {"context_logits":0.0,"cum_log_probs":0.0,"generation_logits":0.0,"model_name":"ensemble","model_version":"1","output_log_probs":[0.0,0.0,0.0,0.0,0.0,0.0,0.0],"sequence_end":false,"sequence_id":0,"sequence_start":false,"text_output":""}

data: {"context_logits":0.0,"cum_log_probs":0.0,"generation_logits":0.0,"model_name":"ensemble","model_version":"1","output_log_probs":[0.0,0.0,0.0,0.0,0.0,0.0,0.0],"sequence_end":false,"sequence_id":0,"sequence_start":false,"text_output":"\n"}

data: {"context_logits":0.0,"cum_log_probs":0.0,"generation_logits":0.0,"model_name":"ensemble","model_version":"1","output_log_probs":[0.0,0.0,0.0,0.0,0.0,0.0,0.0],"sequence_end":false,"sequence_id":0,"sequence_start":false,"text_output":""}

data: {"context_logits":0.0,"cum_log_probs":0.0,"generation_logits":0.0,"model_name":"ensemble","model_version":"1","output_log_probs":[0.0,0.0,0.0,0.0,0.0,0.0,0.0],"sequence_end":false,"sequence_id":0,"sequence_start":false,"text_output":"model"}

data: {"context_logits":0.0,"cum_log_probs":0.0,"generation_logits":0.0,"model_name":"ensemble","model_version":"1","output_log_probs":[0.0,0.0,0.0,0.0,0.0,0.0,0.0],"sequence_end":false,"sequence_id":0,"sequence_start":false,"text_output":"\n"}

data: {"context_logits":0.0,"cum_log_probs":0.0,"generation_logits":0.0,"model_name":"ensemble","model_version":"1","output_log_probs":[0.0,0.0,0.0,0.0,0.0,0.0,0.0],"sequence_end":false,"sequence_id":0,"sequence_start":false,"text_output":"I"}

data: {"context_logits":0.0,"cum_log_probs":0.0,"generation_logits":0.0,"model_name":"ensemble","model_version":"1","output_log_probs":[0.0,0.0,0.0,0.0,0.0,0.0,0.0],"sequence_end":false,"sequence_id":0,"sequence_start":false,"text_output":" am"}

data: {"context_logits":0.0,"cum_log_probs":0.0,"generation_logits":0.0,"model_name":"ensemble","model_version":"1","output_log_probs":[0.0,0.0,0.0,0.0,0.0,0.0,0.0],"sequence_end":false,"sequence_id":0,"sequence_start":false,"text_output":" doing"}

data: {"context_logits":0.0,"cum_log_probs":0.0,"generation_logits":0.0,"model_name":"ensemble","model_version":"1","output_log_probs":[0.0,0.0,0.0,0.0,0.0,0.0,0.0],"sequence_end":false,"sequence_id":0,"sequence_start":false,"text_output":" well"}

data: {"context_logits":0.0,"cum_log_probs":0.0,"generation_logits":0.0,"model_name":"ensemble","model_version":"1","output_log_probs":[0.0,0.0,0.0,0.0,0.0,0.0,0.0],"sequence_end":false,"sequence_id":0,"sequence_start":false,"text_output":","}

data: {"context_logits":0.0,"cum_log_probs":0.0,"generation_logits":0.0,"model_name":"ensemble","model_version":"1","output_log_probs":[0.0,0.0,0.0,0.0,0.0,0.0,0.0],"sequence_end":false,"sequence_id":0,"sequence_start":false,"text_output":" thank"}
...

I dont want to get output_log_probs.

see line: 375 in this file:
https://github.com/triton-inference-server/tensorrtllm_backend/blob/v0.14.0/all_models/inflight_batcher_llm/tensorrt_llm/1/model.py