Issues encountered with integrating trtllm metrics

[varungup90]

• Config

cat <<'EOF' > /tmp/config.yaml
backend: pytorch
kv_cache_config:
  free_gpu_memory_fraction: 0.85
max_num_tokens: 8192
max_batch_size: 16
trust_remote_code: true
return_perf_metrics: true
enable_iter_perf_stats: true
enable_iter_req_stats: true
perf_metrics_max_requests: 1000
EOF

trtllm-serve serve /models/Qwen3-8B \
  --host 0.0.0.0 \
  --port 8000 \
  --backend pytorch \
  --extra_llm_api_options /tmp/config.yaml

• On doing curl for /metrics endpoint, output is always empty []

root@qwen-tensorrt-llm-5956796475-d89cn:/app/tensorrt_llm# curl localhost:8000/metrics
[]

• In metrics output, engine_type and modelname are undefiend

root@qwen-tensorrt-llm-5956796475-d89cn:/app/tensorrt_llm# curl localhost:8000/prometheus/metrics
# HELP e2e_request_latency_seconds Histogram of end to end request latency in seconds.
# TYPE e2e_request_latency_seconds histogram
e2e_request_latency_seconds_sum{engine_type="undefined",model_name="undefined"} 3.4878239994868636
e2e_request_latency_seconds_bucket{engine_type="undefined",le="0.3",model_name="undefined"} 0.0
e2e_request_latency_seconds_bucket{engine_type="undefined",le="0.5",model_name="undefined"} 0.0
e2e_request_latency_seconds_bucket{engine_type="undefined",le="0.8",model_name="undefined"} 0.0

• Limited number of metrics are provided.

e2e_request_latency_seconds_bucket

time_to_first_token_seconds_bucket

time_per_output_token_seconds_bucket

request_queue_time_seconds_bucket

request_success_total