[Bug]: Memory leak: Qwen3 30B A3B pytorch backend

System Info

• CPU architecture: amd64
• CPU memory size: 80Gb
• GPU properties
  • GPU name: NVIDIA H200
  • GPU memory size: 71424MiB (4g.71gb MIG)
  • Driver Version: 580.126.20
  • CUDA Version: 13.1
• TensorRT-LLM docker image used: nvcr.io/nvidia/tensorrt-llm/release:1.3.0rc12

Who can help?

No response

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder (such as GLUE/SQuAD, ...)
• My own task or dataset (give details below)

Reproduction

Steps to reproduce behaviour:

cat <<'EOF' > config.yaml
max_seq_len: 32768
max_num_tokens: 16384
max_batch_size: 64

num_postprocess_workers: 4

stream_interval: 5

enable_chunked_prefill: true

return_perf_metrics: true
enable_iter_perf_stats: true
enable_iter_req_stats: true

cuda_graph_config:
  enable_padding: true
  max_batch_size: 32
EOF

2. trtllm-serve Qwen/Qwen3-30B-A3B-Instruct-2507-FP8 --config=config.yaml
3. for i in $(seq 1 100000); do curl -s -X POST http://localhost:8080/v1/chat/completions -H "Content-Type: application/json" -d '{"messages":[{"role":"system","content":"You are a helpful assistant."},{"role":"user","content":"Say '\''ok'\''"}],"model":"dsp.optimized.qwen3-30b-a3b","max_tokens":512,"temperature":1e-05}'; echo; done

Expected behavior

Container memory (RAM) usage shouldn't keep increasing as requests are served.

actual behavior

Container memory usage keeps growing as requests are served till it's OOMKilled.

additional notes

• The memory usage remains flat when there is no traffic.
• This happens with 1.3.0rc13 as well.

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.