From af53491cac22e38a8012edecdae5af69b1307e5f Mon Sep 17 00:00:00 2001 From: tzhouam Date: Tue, 10 Feb 2026 09:54:32 +0000 Subject: [PATCH] Update sampling parameters in openai_chat_completion_client_for_multimodal_generation.py for improved model performance, adjusting temperature, top_k, and max_tokens settings. Signed-off-by: tzhouam --- ...i_chat_completion_client_for_multimodal_generation.py | 9 +++++---- 1 file changed, 5 insertions(+), 4 deletions(-) diff --git a/examples/online_serving/qwen3_omni/openai_chat_completion_client_for_multimodal_generation.py b/examples/online_serving/qwen3_omni/openai_chat_completion_client_for_multimodal_generation.py index 79d72afbe..6e945509a 100644 --- a/examples/online_serving/qwen3_omni/openai_chat_completion_client_for_multimodal_generation.py +++ b/examples/online_serving/qwen3_omni/openai_chat_completion_client_for_multimodal_generation.py @@ -345,13 +345,14 @@ query_map = { def run_multimodal_generation(args) -> None: model_name = args.model thinker_sampling_params = { - "temperature": 0.4, # Deterministic + "temperature": 0.2, "top_p": 0.9, - "top_k": 1, - "max_tokens": 16384, + "top_k": -1, + "max_tokens": 1200, "repetition_penalty": 1.05, - "stop_token_ids": [151645], # Qwen EOS token <|im_end|> + "logit_bias": {}, "seed": SEED, + "stop_token_ids": [151645], # Qwen EOS token <|im_end|> } # Sampling parameters for Talker stage (codec generation)