BackendsTensorRT-LLM

Known Issues and Mitigations

View as Markdown

For general TensorRT-LLM features and configuration, see the Reference Guide.

KV Cache Exhaustion Causing Worker Deadlock (Disaggregated Serving)

Issue: In disaggregated serving mode, TensorRT-LLM workers can become stuck and unresponsive after sustained high-load traffic. Once in this state, workers require a pod/process restart to recover.

Symptoms:

  • Workers function normally initially but hang after heavy load testing
  • Inference requests get stuck and eventually timeout
  • Logs show warnings: num_fitting_reqs=0 and fitting_disagg_gen_init_requests is empty, may not have enough kvCache
  • Error logs may contain: asyncio.exceptions.InvalidStateError: invalid state

Root Cause: When max_tokens_in_buffer in the cache transceiver config is smaller than the maximum input sequence length (ISL) being processed, KV cache exhaustion can occur under heavy load. This causes context transfers to timeout, leaving workers stuck waiting for phantom transfers and entering an irrecoverable deadlock state.

Mitigation: Ensure max_tokens_in_buffer exceeds your maximum expected input sequence length. Update your engine configuration files (e.g., prefill.yaml and decode.yaml):

1cache_transceiver_config:
2  backend: DEFAULT
3  max_tokens_in_buffer: 65536  # Must exceed max ISL

For example, see examples/backends/trtllm/engine_configs/gpt-oss-120b/prefill.yaml.

Related Issue: #4327

Driver mismatch produces cryptic PyTorch errors

When the host NVIDIA driver is too old for the container’s CUDA version, PyTorch surfaces the failure as:

RuntimeError: The NVIDIA driver on your system is too old (found version 570). Please update your GPU driver by downloading and installing a new version from the URL: http://www.nvidia.com/Download/index.aspx

This is the symptom, not the cause — the cause is that the container image you pulled needs a newer driver than the host ships.

Fix:

  • Check the minimum driver for the tag you pulled in the Container / driver matrix.
  • Either upgrade the host driver, or pull a lower-CUDA variant (e.g. vllm-runtime:1.0.2 on driver 575+ instead of vllm-runtime:1.0.2-cuda13 on driver 580+).

The driver-mismatch error message itself is being improved — tracked as an engineering follow-up.