imoneoi / openchat

OpenChat: Advancing Open-source Language Models with Imperfect Data
https://openchat.team
Apache License 2.0
5.26k stars 399 forks source link

OpenChat API Non-Deterministic Behavior #183

Closed SkanderHellal closed 8 months ago

SkanderHellal commented 9 months ago

Issue: The OpenChat API appears to exhibit non-deterministic behavior in RAG applications using the same user query.

Details When using the same user query, I get different answer results. In fact, I am not getting the same output texts with the same user query.

Versions vllm==0.3.2 ochat=3.5.1

Openchat configuration

model='openchat/openchat-3.5-1210', tokenizer='openchat/openchat-3.5-1210', tokenizer_mode=auto, revision=None, tokenizer_revision=None, trust_remote_code=False, dtype=torch.float16, max_seq_len=8192, download_dir=None, load_format=auto, tensor_parallel_size=2, disable_custom_all_reduce=True, quantization=None, enforce_eager=False, kv_cache_dtype=auto, device_config=cuda, seed=123

imoneoi commented 8 months ago

@SkanderHellal Thanks for your PR. Can you check if the issue still happens?

SkanderHellal commented 8 months ago

@imoneoi Thank you for your reply, now openchat has a deterministic behavior with seed fix. I close the issue.