There is no vision part in Qwen3-VL trace, need help!

[junyyang-amd]

There is no vision part in Qwen3-VL trace, I don't know why, who can help me, thank you very much!

The launch server command I used is:
export SGLANG_PROFILE_WITH_STACK=1 export SGLANG_TORCH_PROFILER_DIR=./Qwen3-VL-30B-A3B-Instruct-FP8/trace export SGLANG_PROFILE_RECORD_SHAPES=1 python3 -m sglang.launch_server \ --model-path ./Qwen/Qwen3-VL-30B-A3B-Instruct-FP8 \ --served-model-name Qwen3-VL-30B-A3B-Instruct-FP8 \ --host 0.0.0.0 \ --port 9011 \ --tp-size 8 \ --ep-size 8 \ --trust-remote-code \ --chunked-prefill-size 32768 \ --mem-fraction-static 0.9 \ --disable-radix-cache \ --max-prefill-tokens 32768 \ --cuda-graph-max-bs 128 \ --mm-attention-backend aiter_attn \

The client command I used is:
python3 -m sglang.bench_serving \ --backend sglang-oai \ --model ${model} \ --port 9011 \ --dataset-name image \ --num-prompts 10 \ --image-count 1 \ --image-resolution 800x800 \ --max-concurrency 10 \ --random-input-len 1000 --random-output-len 2000 --random-range-ratio 1 \ --profile \

[llfl]

I also encountered a similar issue last week, not sure if it's the same one: In python/sglang/bench_serving.py:create_mm_data_row(), the function is supposed to insert the image tokens corresponding to the model into the text_prompt. However, for some reason, the function ultimately uses a use_raw_prompt flag that reverts the prompt back to its original form without the added image tokens.
This may help

diff --git a/python/sglang/bench_serving.py b/python/sglang/bench_serving.py
index 4b5da0445..57adefb78 100644
--- a/python/sglang/bench_serving.py
+++ b/python/sglang/bench_serving.py
@@ -1387,9 +1387,9 @@ def create_mm_data_row(
     vision_prompt_len = prompt_len - text_prompt_len
 
     use_raw_prompt = backend in [
-        "sglang",
-        "sglang-oai",
-        "sglang-oai-chat",
+        # "sglang",
+        # "sglang-oai",
+        # "sglang-oai-chat",
         "vllm",
         "vllm-chat",
         "lmdeploy",