[Performance]: TRT-LLM 1.2.0.rc5 outperforms all later versions on Qwen3-0.6B

Proposal to improve performance

Does anyone know the reason? Our actual test shows that the Qwen3-0.6B model performs better with TRT-LLM 1.2.0.rc5 than with all subsequent versions.
Test Environment:
GPU: 1 × A30
Model: Qwen3-0.6B
Prompt length: ~100 tokens
Max output tokens: 25
Model Engine Export Commands:
bash
python3 convert_checkpoint.py --model_dir ${HF_MODEL_PATH}
--output_dir ${TEMP_IR_CHECKPOINT}
--dtype float16

trtllm-build --checkpoint_dir ${TEMP_IR_CHECKPOINT}
--output_dir ${ENGINE_PATH}
--gemm_plugin float16
--max_batch_size 32
--max_input_len 160
--max_beam_width 32
--max_num_tokens 9600
--tokens_per_block 256
--max_seq_len 256
start：
trtllm-serve serve /workspace/model/ --tokenizer /workspace/model/tokenizer --port 8000 --host 0.0.0.0 --tp_size 1 --kv_cache_free_gpu_memory_fraction 0.92 --backend trt --max_batch_size 32 --max_seq_len 8192
With all environments kept identical, the test results are as follows:
TRT-LLM 1.2.0.rc5 Test Results:
QPS AVG 延迟 (ms) P90 延迟 (ms) P95 延迟 (ms) P99 延迟 (ms)
45 45 65 75 80
80 45 65 75 80
115 50 70 85 90
140 50 80 90 95
170 60 90 100 105
TRT-LLM 1.3.0.rc8 Test Results:
QPS AVG 延迟 (ms) P90 延迟 (ms) P95 延迟 (ms) P99 延迟 (ms)
25 80 100 105 110
45 80 100 110 120
65 90 110 120 130
85 100 120 130 140
105 120 130 140 150

Report of performance regression

No response

Misc discussion on performance

No response

Your current environment (if you think it is necessary)

System Information:

• OS:
• Python version:
• CUDA version:
• GPU model(s):
• Driver version:
• TensorRT version:
• PyTorch version:
• TensorRT-LLM version:

Detailed output:

Paste the output of the above commands here

Before submitting a new issue...

• Make sure you already searched for relevant issues, and checked the documentation and examples for answers to frequently asked questions.